//
// Generated by LLVM NVPTX Back-End
//

.version 8.8
.target sm_89
.address_size 64

	// .globl	quant_key_per_thread_int8_kernel // -- Begin function quant_key_per_thread_int8_kernel
.extern .shared .align 16 .b8 global_smem[];
                                        // @quant_key_per_thread_int8_kernel
.visible .entry quant_key_per_thread_int8_kernel(
	.param .u64 .ptr .global .align 1 quant_key_per_thread_int8_kernel_param_0,
	.param .u64 .ptr .global .align 1 quant_key_per_thread_int8_kernel_param_1,
	.param .u64 .ptr .global .align 1 quant_key_per_thread_int8_kernel_param_2,
	.param .u32 quant_key_per_thread_int8_kernel_param_3,
	.param .u32 quant_key_per_thread_int8_kernel_param_4,
	.param .u32 quant_key_per_thread_int8_kernel_param_5,
	.param .u32 quant_key_per_thread_int8_kernel_param_6,
	.param .u32 quant_key_per_thread_int8_kernel_param_7,
	.param .u32 quant_key_per_thread_int8_kernel_param_8,
	.param .u32 quant_key_per_thread_int8_kernel_param_9,
	.param .u32 quant_key_per_thread_int8_kernel_param_10,
	.param .u32 quant_key_per_thread_int8_kernel_param_11,
	.param .u64 .ptr .global .align 1 quant_key_per_thread_int8_kernel_param_12,
	.param .u64 .ptr .global .align 1 quant_key_per_thread_int8_kernel_param_13
)
.reqntid 128
{
	.reg .pred 	%p<12>;
	.reg .b16 	%rs<25>;
	.reg .b32 	%r<144>;
	.reg .b64 	%rd<26>;
	.loc	1 49 0                          // quant_per_thread.py:49:0
$L__func_begin0:
	.loc	1 49 0                          // quant_per_thread.py:49:0

// %bb.0:
	ld.param.b64 	%rd6, [quant_key_per_thread_int8_kernel_param_0];
	ld.param.b64 	%rd7, [quant_key_per_thread_int8_kernel_param_1];
$L__tmp0:
	.loc	1 54 15                         // quant_per_thread.py:54:15
	mov.u32 	%r11, %ctaid.x;
	.loc	1 55 15                         // quant_per_thread.py:55:15
	and.b32 	%r12, %r11, 3;
	ld.param.b64 	%rd8, [quant_key_per_thread_int8_kernel_param_2];
	ld.param.b32 	%r13, [quant_key_per_thread_int8_kernel_param_3];
	.loc	1 56 13                         // quant_per_thread.py:56:13
	mov.u32 	%r14, %ctaid.y;
	.loc	1 57 13                         // quant_per_thread.py:57:13
	mov.u32 	%r15, %ctaid.z;
	ld.param.b32 	%r16, [quant_key_per_thread_int8_kernel_param_4];
	.loc	1 75 15                         // quant_per_thread.py:75:15
	shl.b32 	%r17, %r11, 4;
	ld.param.b32 	%r18, [quant_key_per_thread_int8_kernel_param_5];
	and.b32 	%r19, %r17, -64;
	ld.param.b32 	%r20, [quant_key_per_thread_int8_kernel_param_6];
	ld.param.b32 	%r21, [quant_key_per_thread_int8_kernel_param_7];
	.loc	1 75 31                         // quant_per_thread.py:75:31
	mov.u32 	%r22, %tid.x;
	shr.u32 	%r23, %r22, 1;
	ld.param.b32 	%r24, [quant_key_per_thread_int8_kernel_param_8];
	and.b32 	%r25, %r23, 56;
	ld.param.b32 	%r26, [quant_key_per_thread_int8_kernel_param_9];
	.loc	1 75 15                         // quant_per_thread.py:75:15
	or.b32 	%r27, %r19, %r25;
	ld.param.b32 	%r28, [quant_key_per_thread_int8_kernel_param_10];
	.loc	1 75 60                         // quant_per_thread.py:75:60
	shl.b32 	%r29, %r12, 1;
	ld.param.b32 	%r30, [quant_key_per_thread_int8_kernel_param_11];
	.loc	1 75 15                         // quant_per_thread.py:75:15
	or.b32 	%r31, %r27, %r29;
	.loc	1 76 15                         // quant_per_thread.py:76:15
	or.b32 	%r32, %r31, 1;
	.loc	1 79 27                         // quant_per_thread.py:79:27
	mul.lo.s32 	%r33, %r16, %r15;
	.loc	1 79 19                         // quant_per_thread.py:79:19
	mad.wide.s32 	%rd9, %r33, 2, %rd6;
	.loc	1 79 47                         // quant_per_thread.py:79:47
	mul.lo.s32 	%r34, %r18, %r14;
	.loc	1 79 19                         // quant_per_thread.py:79:19
	mad.wide.s32 	%rd10, %r34, 2, %rd9;
	.loc	1 79 67                         // quant_per_thread.py:79:67
	mul.lo.s32 	%r35, %r20, %r31;
	.loc	1 79 19                         // quant_per_thread.py:79:19
	mad.wide.s32 	%rd11, %r35, 2, %rd10;
	.loc	1 79 98                         // quant_per_thread.py:79:98
	shl.b32 	%r36, %r22, 2;
	and.b32 	%r37, %r36, 60;
	.loc	1 79 19                         // quant_per_thread.py:79:19
	cvt.u64.u32 	%rd12, %r37;
	mul.wide.u32 	%rd13, %r37, 2;
	add.s64 	%rd1, %rd11, %rd13;
	.loc	1 80 67                         // quant_per_thread.py:80:67
	mul.lo.s32 	%r38, %r20, %r32;
	.loc	1 80 19                         // quant_per_thread.py:80:19
	mad.wide.s32 	%rd14, %r38, 2, %rd10;
	add.s64 	%rd2, %rd14, %rd13;
	.loc	1 81 29                         // quant_per_thread.py:81:29
	mul.lo.s32 	%r39, %r21, %r15;
	.loc	1 81 20                         // quant_per_thread.py:81:20
	cvt.s64.s32 	%rd15, %r39;
	add.s64 	%rd16, %rd7, %rd15;
	.loc	1 81 49                         // quant_per_thread.py:81:49
	mul.lo.s32 	%r40, %r24, %r14;
	.loc	1 81 20                         // quant_per_thread.py:81:20
	cvt.s64.s32 	%rd17, %r40;
	add.s64 	%rd18, %rd16, %rd17;
	.loc	1 81 69                         // quant_per_thread.py:81:69
	mul.lo.s32 	%r41, %r26, %r31;
	.loc	1 81 20                         // quant_per_thread.py:81:20
	cvt.s64.s32 	%rd19, %r41;
	add.s64 	%rd20, %rd18, %rd19;
	add.s64 	%rd3, %rd20, %rd12;
	.loc	1 82 69                         // quant_per_thread.py:82:69
	mul.lo.s32 	%r42, %r26, %r32;
	.loc	1 82 20                         // quant_per_thread.py:82:20
	cvt.s64.s32 	%rd21, %r42;
	add.s64 	%rd22, %rd18, %rd21;
	add.s64 	%rd4, %rd22, %rd12;
	.loc	1 83 26                         // quant_per_thread.py:83:26
	mul.lo.s32 	%r43, %r28, %r15;
	.loc	1 83 18                         // quant_per_thread.py:83:18
	mad.wide.s32 	%rd23, %r43, 4, %rd8;
	.loc	1 83 46                         // quant_per_thread.py:83:46
	mul.lo.s32 	%r44, %r30, %r14;
	.loc	1 83 18                         // quant_per_thread.py:83:18
	mad.wide.s32 	%rd24, %r44, 4, %rd23;
	.loc	1 83 66                         // quant_per_thread.py:83:66
	and.b32 	%r45, %r11, 2147483644;
	.loc	1 83 18                         // quant_per_thread.py:83:18
	mad.wide.u32 	%rd25, %r45, 4, %rd24;
	mad.wide.u32 	%rd5, %r12, 4, %rd25;
	.loc	1 85 36                         // quant_per_thread.py:85:36
	setp.lt.s32 	%p1, %r31, %r13;
	.loc	1 85 10                         // quant_per_thread.py:85:10
	// begin inline asm
	.loc	1 85 10                         // quant_per_thread.py:85:10
	mov.u32 %r1, 0x0;
	.loc	1 86 10                         // quant_per_thread.py:86:10
	mov.u32 %r2, 0x0;
	@%p1 ld.global.v2.b32 { %r1, %r2 }, [ %rd1 + 0 ];

	// end inline asm
	mov.b32 	{%rs1, %rs2}, %r1;
	mov.b32 	{%rs3, %rs4}, %r2;
	.loc	1 86 36                         // quant_per_thread.py:86:36
	setp.lt.s32 	%p2, %r32, %r13;
	.loc	1 86 10                         // quant_per_thread.py:86:10
	// begin inline asm
	.loc	1 86 10                         // quant_per_thread.py:86:10
	mov.u32 %r3, 0x0;
	.loc	1 87 10                         // quant_per_thread.py:87:10
	mov.u32 %r4, 0x0;
	@%p2 ld.global.v2.b32 { %r3, %r4 }, [ %rd2 + 0 ];

	// end inline asm
	mov.b32 	{%rs5, %rs6}, %r3;
	mov.b32 	{%rs7, %rs8}, %r4;
	.loc	1 87 10                         // quant_per_thread.py:87:10
	cvt.f32.bf16 	%r46, %rs1;
	cvt.f32.bf16 	%r47, %rs2;
	cvt.f32.bf16 	%r48, %rs3;
	cvt.f32.bf16 	%r49, %rs4;
	.loc	1 88 10                         // quant_per_thread.py:88:10
	cvt.f32.bf16 	%r50, %rs5;
	cvt.f32.bf16 	%r51, %rs6;
	cvt.f32.bf16 	%r52, %rs7;
	cvt.f32.bf16 	%r53, %rs8;
	.loc	1 89 24                         // quant_per_thread.py:89:24
	abs.f32 	%r54, %r46;
	abs.f32 	%r55, %r47;
	abs.f32 	%r56, %r48;
	abs.f32 	%r57, %r49;
$L__tmp1:
	.loc	1 89 17                         // quant_per_thread.py:89:17
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r58, %r54, %r55;
	max.f32 	%r59, %r56, %r57;
	max.f32 	%r60, %r58, %r59;
$L__tmp2:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	shfl.sync.bfly.b32 	%r61, %r60, 16, 31, -1;
$L__tmp3:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r62, %r60, %r61;
$L__tmp4:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	shfl.sync.bfly.b32 	%r63, %r62, 8, 31, -1;
$L__tmp5:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r64, %r62, %r63;
$L__tmp6:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	shfl.sync.bfly.b32 	%r65, %r64, 4, 31, -1;
$L__tmp7:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r66, %r64, %r65;
$L__tmp8:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	shfl.sync.bfly.b32 	%r67, %r66, 2, 31, -1;
$L__tmp9:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r68, %r66, %r67;
$L__tmp10:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	shfl.sync.bfly.b32 	%r69, %r68, 1, 31, -1;
$L__tmp11:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r6, %r68, %r69;
$L__tmp12:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	shr.u32 	%r70, %r22, 3;
	and.b32 	%r71, %r70, 12;
	mov.b32 	%r72, global_smem;
	add.s32 	%r5, %r72, %r71;
	// begin inline asm
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	st.shared::cta.b32 [ %r5 + 0 ], %r6;

	// end inline asm
	bar.sync 	0;
	and.b32 	%r73, %r36, 12;
	add.s32 	%r74, %r72, %r73;
	ld.shared.b32 	%r75, [%r74];
	shfl.sync.bfly.b32 	%r76, %r75, 2, 31, -1;
$L__tmp13:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r77, %r75, %r76;
$L__tmp14:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 17 // standard.py:191:16
	shfl.sync.bfly.b32 	%r78, %r77, 1, 31, -1;
$L__tmp15:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r79, %r77, %r78;
$L__tmp16:
	.loc	1 89 44                         // quant_per_thread.py:89:44
	abs.f32 	%r80, %r50;
	abs.f32 	%r81, %r51;
	abs.f32 	%r82, %r52;
	abs.f32 	%r83, %r53;
$L__tmp17:
	.loc	1 89 37                         // quant_per_thread.py:89:37
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	bar.sync 	0;
$L__tmp18:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r84, %r80, %r81;
	max.f32 	%r85, %r82, %r83;
	max.f32 	%r86, %r84, %r85;
$L__tmp19:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	shfl.sync.bfly.b32 	%r87, %r86, 16, 31, -1;
$L__tmp20:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r88, %r86, %r87;
$L__tmp21:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	shfl.sync.bfly.b32 	%r89, %r88, 8, 31, -1;
$L__tmp22:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r90, %r88, %r89;
$L__tmp23:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	shfl.sync.bfly.b32 	%r91, %r90, 4, 31, -1;
$L__tmp24:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r92, %r90, %r91;
$L__tmp25:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	shfl.sync.bfly.b32 	%r93, %r92, 2, 31, -1;
$L__tmp26:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r94, %r92, %r93;
$L__tmp27:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	shfl.sync.bfly.b32 	%r95, %r94, 1, 31, -1;
$L__tmp28:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r7, %r94, %r95;
$L__tmp29:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	// begin inline asm
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	st.shared::cta.b32 [ %r5 + 0 ], %r7;

	// end inline asm
	bar.sync 	0;
	ld.shared.b32 	%r96, [%r74];
	shfl.sync.bfly.b32 	%r97, %r96, 2, 31, -1;
$L__tmp30:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r98, %r96, %r97;
$L__tmp31:
	.loc	2 191 16, function_name $L__info_string0, inlined_at 1 89 37 // standard.py:191:16
	shfl.sync.bfly.b32 	%r99, %r98, 1, 31, -1;
$L__tmp32:
	.loc	2 170 12, function_name $L__info_string0, inlined_at 2 191 16 // standard.py:170:12
	max.f32 	%r100, %r98, %r99;
$L__tmp33:
	.loc	1 89 13                         // quant_per_thread.py:89:13
	max.f32 	%r101, %r79, %r100;
	mov.b32 	%r102, 0f42FE0000;
	div.full.f32 	%r103, %r101, %r102;
	add.f32 	%r10, %r103, 0f33D6BF95;
	.loc	1 90 15                         // quant_per_thread.py:90:15
	div.full.f32 	%r104, %r46, %r10;
	div.full.f32 	%r105, %r47, %r10;
	div.full.f32 	%r106, %r48, %r10;
	div.full.f32 	%r107, %r49, %r10;
	.loc	1 91 15                         // quant_per_thread.py:91:15
	div.full.f32 	%r108, %r50, %r10;
	div.full.f32 	%r109, %r51, %r10;
	div.full.f32 	%r110, %r52, %r10;
	div.full.f32 	%r111, %r53, %r10;
	.loc	1 92 31                         // quant_per_thread.py:92:31
	setp.ge.f32 	%p4, %r104, 0f00000000;
	setp.ge.f32 	%p5, %r105, 0f00000000;
	setp.ge.f32 	%p6, %r106, 0f00000000;
	setp.ge.f32 	%p7, %r107, 0f00000000;
	.loc	1 92 22                         // quant_per_thread.py:92:22
	selp.f32 	%r112, 0f3F000000, 0fBF000000, %p7;
	selp.f32 	%r113, 0f3F000000, 0fBF000000, %p6;
	selp.f32 	%r114, 0f3F000000, 0fBF000000, %p5;
	selp.f32 	%r115, 0f3F000000, 0fBF000000, %p4;
	.loc	1 92 5                          // quant_per_thread.py:92:5
	add.f32 	%r116, %r104, %r115;
	add.f32 	%r117, %r105, %r114;
	add.f32 	%r118, %r106, %r113;
	add.f32 	%r119, %r107, %r112;
	.loc	1 94 15                         // quant_per_thread.py:94:15
	cvt.rzi.s16.f32 	%rs9, %r119;
	cvt.rzi.s16.f32 	%rs10, %r118;
	mov.b32 	%r120, {%rs10, %rs9};
	mov.b32 	{%rs11, %rs12}, %r120;
	cvt.u32.u16 	%r121, %rs12;
	cvt.u32.u16 	%r122, %rs11;
	prmt.b32 	%r123, %r122, %r121, 0x3340U;
	cvt.rzi.s16.f32 	%rs13, %r117;
	cvt.rzi.s16.f32 	%rs14, %r116;
	mov.b32 	%r124, {%rs14, %rs13};
	mov.b32 	{%rs15, %rs16}, %r124;
	cvt.u32.u16 	%r125, %rs16;
	cvt.u32.u16 	%r126, %rs15;
	prmt.b32 	%r127, %r126, %r125, 0x3340U;
	prmt.b32 	%r8, %r127, %r123, 0x5410U;
	.loc	1 93 31                         // quant_per_thread.py:93:31
	setp.ge.f32 	%p8, %r108, 0f00000000;
	setp.ge.f32 	%p9, %r109, 0f00000000;
	setp.ge.f32 	%p10, %r110, 0f00000000;
	setp.ge.f32 	%p11, %r111, 0f00000000;
	.loc	1 93 22                         // quant_per_thread.py:93:22
	selp.f32 	%r128, 0f3F000000, 0fBF000000, %p11;
	selp.f32 	%r129, 0f3F000000, 0fBF000000, %p10;
	selp.f32 	%r130, 0f3F000000, 0fBF000000, %p9;
	selp.f32 	%r131, 0f3F000000, 0fBF000000, %p8;
	.loc	1 93 5                          // quant_per_thread.py:93:5
	add.f32 	%r132, %r108, %r131;
	add.f32 	%r133, %r109, %r130;
	add.f32 	%r134, %r110, %r129;
	add.f32 	%r135, %r111, %r128;
	.loc	1 95 15                         // quant_per_thread.py:95:15
	cvt.rzi.s16.f32 	%rs17, %r135;
	cvt.rzi.s16.f32 	%rs18, %r134;
	mov.b32 	%r136, {%rs18, %rs17};
	mov.b32 	{%rs19, %rs20}, %r136;
	cvt.u32.u16 	%r137, %rs20;
	cvt.u32.u16 	%r138, %rs19;
	prmt.b32 	%r139, %r138, %r137, 0x3340U;
	cvt.rzi.s16.f32 	%rs21, %r133;
	cvt.rzi.s16.f32 	%rs22, %r132;
	mov.b32 	%r140, {%rs22, %rs21};
	mov.b32 	{%rs23, %rs24}, %r140;
	cvt.u32.u16 	%r141, %rs24;
	cvt.u32.u16 	%r142, %rs23;
	prmt.b32 	%r143, %r142, %r141, 0x3340U;
	prmt.b32 	%r9, %r143, %r139, 0x5410U;
	.loc	1 96 5                          // quant_per_thread.py:96:5
	// begin inline asm
	@%p1 st.global.b32 [ %rd3 + 0 ], { %r8 };

	// end inline asm
	.loc	1 97 5                          // quant_per_thread.py:97:5
	// begin inline asm
	@%p2 st.global.b32 [ %rd4 + 0 ], { %r9 };

	// end inline asm
	.loc	1 98 5                          // quant_per_thread.py:98:5
	setp.eq.b32 	%p3, %r22, 0;
	// begin inline asm
	@%p3 st.global.b32 [ %rd5 + 0 ], { %r10 };

	// end inline asm
	.loc	1 49 1                          // quant_per_thread.py:49:1
	ret;
$L__tmp34:
$L__func_end0:
                                        // -- End function
}
	.file	1 "D:\\Comfy\\App\\ComfyUI_windows_portable\\python_embeded\\Lib\\site-packages\\sageattention\\triton\\quant_per_thread.py"
	.file	2 "D:\\Comfy\\App\\ComfyUI_windows_portable\\python_embeded\\Lib\\site-packages\\triton\\language\\standard.py"
	.section	.debug_abbrev
	{
.b8 1                                   // Abbreviation Code
.b8 17                                  // DW_TAG_compile_unit
.b8 1                                   // DW_CHILDREN_yes
.b8 37                                  // DW_AT_producer
.b8 8                                   // DW_FORM_string
.b8 19                                  // DW_AT_language
.b8 5                                   // DW_FORM_data2
.b8 3                                   // DW_AT_name
.b8 8                                   // DW_FORM_string
.b8 16                                  // DW_AT_stmt_list
.b8 6                                   // DW_FORM_data4
.b8 27                                  // DW_AT_comp_dir
.b8 8                                   // DW_FORM_string
.b8 0                                   // EOM(1)
.b8 0                                   // EOM(2)
.b8 2                                   // Abbreviation Code
.b8 46                                  // DW_TAG_subprogram
.b8 0                                   // DW_CHILDREN_no
.b8 3                                   // DW_AT_name
.b8 8                                   // DW_FORM_string
.b8 32                                  // DW_AT_inline
.b8 11                                  // DW_FORM_data1
.b8 0                                   // EOM(1)
.b8 0                                   // EOM(2)
.b8 3                                   // Abbreviation Code
.b8 46                                  // DW_TAG_subprogram
.b8 1                                   // DW_CHILDREN_yes
.b8 17                                  // DW_AT_low_pc
.b8 1                                   // DW_FORM_addr
.b8 18                                  // DW_AT_high_pc
.b8 1                                   // DW_FORM_addr
.b8 49                                  // DW_AT_abstract_origin
.b8 19                                  // DW_FORM_ref4
.b8 0                                   // EOM(1)
.b8 0                                   // EOM(2)
.b8 4                                   // Abbreviation Code
.b8 29                                  // DW_TAG_inlined_subroutine
.b8 1                                   // DW_CHILDREN_yes
.b8 49                                  // DW_AT_abstract_origin
.b8 19                                  // DW_FORM_ref4
.b8 17                                  // DW_AT_low_pc
.b8 1                                   // DW_FORM_addr
.b8 18                                  // DW_AT_high_pc
.b8 1                                   // DW_FORM_addr
.b8 88                                  // DW_AT_call_file
.b8 11                                  // DW_FORM_data1
.b8 89                                  // DW_AT_call_line
.b8 11                                  // DW_FORM_data1
.b8 87                                  // DW_AT_call_column
.b8 11                                  // DW_FORM_data1
.b8 0                                   // EOM(1)
.b8 0                                   // EOM(2)
.b8 5                                   // Abbreviation Code
.b8 29                                  // DW_TAG_inlined_subroutine
.b8 0                                   // DW_CHILDREN_no
.b8 49                                  // DW_AT_abstract_origin
.b8 19                                  // DW_FORM_ref4
.b8 17                                  // DW_AT_low_pc
.b8 1                                   // DW_FORM_addr
.b8 18                                  // DW_AT_high_pc
.b8 1                                   // DW_FORM_addr
.b8 88                                  // DW_AT_call_file
.b8 11                                  // DW_FORM_data1
.b8 89                                  // DW_AT_call_line
.b8 11                                  // DW_FORM_data1
.b8 87                                  // DW_AT_call_column
.b8 11                                  // DW_FORM_data1
.b8 0                                   // EOM(1)
.b8 0                                   // EOM(2)
.b8 0                                   // EOM(3)
	}
	.section	.debug_info
	{
.b32 289                                // Length of Unit
.b8 2                                   // DWARF version number
.b8 0
.b32 .debug_abbrev                      // Offset Into Abbrev. Section
.b8 8                                   // Address Size (in bytes)
.b8 1                                   // Abbrev [1] 0xb:0x11a DW_TAG_compile_unit
.b8 116                                 // DW_AT_producer
.b8 114
.b8 105
.b8 116
.b8 111
.b8 110
.b8 0
.b8 2                                   // DW_AT_language
.b8 0
.b8 113                                 // DW_AT_name
.b8 117
.b8 97
.b8 110
.b8 116
.b8 95
.b8 112
.b8 101
.b8 114
.b8 95
.b8 116
.b8 104
.b8 114
.b8 101
.b8 97
.b8 100
.b8 46
.b8 112
.b8 121
.b8 0
.b32 .debug_line                        // DW_AT_stmt_list
.b8 68                                  // DW_AT_comp_dir
.b8 58
.b8 92
.b8 67
.b8 111
.b8 109
.b8 102
.b8 121
.b8 92
.b8 65
.b8 112
.b8 112
.b8 92
.b8 67
.b8 111
.b8 109
.b8 102
.b8 121
.b8 85
.b8 73
.b8 95
.b8 119
.b8 105
.b8 110
.b8 100
.b8 111
.b8 119
.b8 115
.b8 95
.b8 112
.b8 111
.b8 114
.b8 116
.b8 97
.b8 98
.b8 108
.b8 101
.b8 92
.b8 112
.b8 121
.b8 116
.b8 104
.b8 111
.b8 110
.b8 95
.b8 101
.b8 109
.b8 98
.b8 101
.b8 100
.b8 101
.b8 100
.b8 92
.b8 76
.b8 105
.b8 98
.b8 92
.b8 115
.b8 105
.b8 116
.b8 101
.b8 45
.b8 112
.b8 97
.b8 99
.b8 107
.b8 97
.b8 103
.b8 101
.b8 115
.b8 92
.b8 115
.b8 97
.b8 103
.b8 101
.b8 97
.b8 116
.b8 116
.b8 101
.b8 110
.b8 116
.b8 105
.b8 111
.b8 110
.b8 92
.b8 116
.b8 114
.b8 105
.b8 116
.b8 111
.b8 110
.b8 0
.b8 2                                   // Abbrev [2] 0x89:0x23 DW_TAG_subprogram
.b8 113                                 // DW_AT_name
.b8 117
.b8 97
.b8 110
.b8 116
.b8 95
.b8 107
.b8 101
.b8 121
.b8 95
.b8 112
.b8 101
.b8 114
.b8 95
.b8 116
.b8 104
.b8 114
.b8 101
.b8 97
.b8 100
.b8 95
.b8 105
.b8 110
.b8 116
.b8 56
.b8 95
.b8 107
.b8 101
.b8 114
.b8 110
.b8 101
.b8 108
.b8 0
.b8 1                                   // DW_AT_inline
.b8 3                                   // Abbrev [3] 0xac:0x78 DW_TAG_subprogram
.b64 $L__func_begin0                    // DW_AT_low_pc
.b64 $L__func_end0                      // DW_AT_high_pc
.b32 137                                // DW_AT_abstract_origin
.b8 4                                   // Abbrev [4] 0xc1:0x31 DW_TAG_inlined_subroutine
.b32 137                                // DW_AT_abstract_origin
.b64 $L__tmp1                           // DW_AT_low_pc
.b64 $L__tmp16                          // DW_AT_high_pc
.b8 1                                   // DW_AT_call_file
.b8 89                                  // DW_AT_call_line
.b8 17                                  // DW_AT_call_column
.b8 5                                   // Abbrev [5] 0xd9:0x18 DW_TAG_inlined_subroutine
.b32 137                                // DW_AT_abstract_origin
.b64 $L__tmp1                           // DW_AT_low_pc
.b64 $L__tmp16                          // DW_AT_high_pc
.b8 2                                   // DW_AT_call_file
.b8 191                                 // DW_AT_call_line
.b8 16                                  // DW_AT_call_column
.b8 0                                   // End Of Children Mark
.b8 4                                   // Abbrev [4] 0xf2:0x31 DW_TAG_inlined_subroutine
.b32 137                                // DW_AT_abstract_origin
.b64 $L__tmp17                          // DW_AT_low_pc
.b64 $L__tmp33                          // DW_AT_high_pc
.b8 1                                   // DW_AT_call_file
.b8 89                                  // DW_AT_call_line
.b8 37                                  // DW_AT_call_column
.b8 5                                   // Abbrev [5] 0x10a:0x18 DW_TAG_inlined_subroutine
.b32 137                                // DW_AT_abstract_origin
.b64 $L__tmp18                          // DW_AT_low_pc
.b64 $L__tmp33                          // DW_AT_high_pc
.b8 2                                   // DW_AT_call_file
.b8 191                                 // DW_AT_call_line
.b8 16                                  // DW_AT_call_column
.b8 0                                   // End Of Children Mark
.b8 0                                   // End Of Children Mark
.b8 0                                   // End Of Children Mark
	}
	.section	.debug_str
	{
$L__info_string0:
.b8 113                                 // string offset=0 ; quant_key_per_thread_int8_kernel
.b8 117
.b8 97
.b8 110
.b8 116
.b8 95
.b8 107
.b8 101
.b8 121
.b8 95
.b8 112
.b8 101
.b8 114
.b8 95
.b8 116
.b8 104
.b8 114
.b8 101
.b8 97
.b8 100
.b8 95
.b8 105
.b8 110
.b8 116
.b8 56
.b8 95
.b8 107
.b8 101
.b8 114
.b8 110
.b8 101
.b8 108
.b8 0
	}
	.section	.debug_macinfo	{	}