syntax = "proto2";
package NV.Profiler.Messages;

import "CpuStacktrace/CpuStacktrace.proto";
import "Nvtx/Nvtx.proto";
import "Profiler/ProfilerMetricOptions.proto";
import "Profiler/ProfilerResultsCommon.proto";
import "Profiler/ProfilerStringTable.proto";
import "ProfilerReport/ProfilerReportCommon.proto";
import "ProfilerSection/ProfilerSection.proto";
import "RuleSystem/RuleResults.proto";

// enable arenas for better allocation performance
option cc_enable_arenas = true;

message DeviceProperty
{
    required string Key     = 1;
    required string Value   = 2;
}

message DeviceProperties
{
    repeated DeviceProperty Properties = 1;
}

message DeviceAttribute
{
    required string             Name     = 1;
    required ProfileMetricValue Value    = 2;
}

message DeviceAttributes
{
    required uint32          ID          = 1;
    optional string          Name        = 2;
    repeated DeviceAttribute Attributes  = 3;
}

// has to match with SystemInfo.proto!
enum PlatformType
{
    Unknown = 0;
    Windows = 1;
    Linux   = 2;
    Android = 3;
    OSX     = 4;
    QNX     = 5;
    Hos     = 6;
    Integrity = 7;
    LinuxPPC  = 8;
}

enum IPAddressType
{
    IPV4    = 1;
    IPV6    = 2;
}

message IPAddress
{
    required string        Address = 1;
    optional IPAddressType Type    = 2;
}

message SystemInfo
{
    optional string         OSName          = 1;
    optional string         Build           = 2;
    optional string         Processor       = 3;
    optional string         Architecture    = 4;
    optional string         ComputerName    = 5;
    optional PlatformType   Platform        = 6;
    repeated IPAddress      IPAddresses     = 7;
}

message VersionInfo
{
    required string         Provider        = 1;
    required string         Version         = 2;
}

message FilterOptions
{
    optional string KernelName              = 1;
    optional string KernelNameBase          = 2;
    optional string KernelId                = 3;
    optional uint32 SkipCount               = 4;
    optional uint32 SkipBeforeMatchCount    = 5;
    optional uint32 CaptureCount            = 6;
}

message PmSamplingOptions
{
    optional uint64 Interval                = 1;
    optional uint64 BufferSize              = 2;
    optional uint64 MaxPasses               = 3;
}

message WarpSamplingOptions
{
    optional bool   IntervalAuto            = 1;
    optional uint64 Interval                = 2;
    optional uint64 MaxPasses               = 3;
    optional uint64 BufferSize              = 4;
}

message OtherOptions
{
    optional bool   ApplyRules  = 1;
    optional string Metrics     = 2;
}

message ProfilerSettings
{
    optional bool                   EnableNvtx                  = 1;
    optional bool                   DisableProfilingStartStop   = 2;
    optional bool                   EnableProfilingFromStart    = 3;
    optional string                 ActivityType                = 4;
    optional FilterOptions          FilterOptions               = 5;
    optional OtherOptions           OtherOptions                = 6;
    optional string                 EnabledSections             = 7;
    optional string                 ClockControlMode            = 8;
    optional WarpSamplingOptions    WarpSamplingOptions         = 9;
    optional bool                   ImportSource                = 10;
    optional string                 ReplayMode                  = 11;
    optional bool                   EnableCpuStack              = 12;
    optional PmSamplingOptions      PmSamplingOptions           = 13;
    optional string                 CpuStackTypes               = 14;
}

message ReportSessionDetails
{
    optional uint32             ProcessID           = 1;

    // Represented as a Unix time since Jan 1 1970
    optional uint64             CreationTime        = 2;

    optional SystemInfo         HostSystemInfo      = 3;
    optional SystemInfo         TargetSystemInfo    = 4;
    optional DeviceProperties   DeviceProperties    = 5;
    repeated DeviceAttributes   DeviceAttributes    = 6;

    // User can modify these to give themselves some hints about the report.
    optional string             Comments            = 7;

    repeated VersionInfo        VersionInfo         = 8;

    optional ExecutableSettings ExecutableSettings  = 9;
    optional ProfilerSettings   ProfilerSettings    = 10;
    optional string             ProfilerCmdline     = 11;
}

message CommentID
{
    required string SectionID = 1;

    // In the future, this could also include something like:
    // optional string TableID = 2;
    // optional uint32 TableRow = 3;
    // optional uint32 TableCol = 4;
}

message Comment
{
    required CommentID ID = 1;
    required string    DisplayName = 2;
    required string    Text = 3;
}

message ProfileSeriesInfoMessage
{
    required uint32         SeriesID        = 1;
    // Is set iff the current result is not the result for the default settings
    optional string         CombinationStr  = 2;
}

message ProfilerClockStatusMessage
{
    required bool         RequestedLock     = 1;
    required bool         ExecutedLocked    = 2;
}

// Result representing a single kernel launch.
message ProfileResult
{
    required uint32                         ThreadID            = 1;
    required uint64                         APICallID           = 2;
    required uint64                         ProgramHandle       = 3;
    required uint64                         KernelID            = 4;
    required string                         KernelMangledName   = 5;
    required string                         KernelFunctionName  = 6;
    required string                         KernelDemangledName = 7;
    required uint32                         WorkDimensions      = 8;

    // x,y,z descriptions
    // work offset
    required Uint64x3                       GlobalWorkOffset    = 9;
    // CUDA grid dims
    required Uint64x3                       GridSize            = 10;
    // CUDA block dims
    required Uint64x3                       BlockSize           = 11;

    optional string                         Comments            = 12;

    repeated ProfileMetricResult            MetricResults       = 13;

    // Represented as a Unix time since Jan 1 1970
    optional uint64                         CreationTime        = 14;

    required SourceData                     Source              = 15;

    // This field is only set to optional in order to be backwards compatible with
    // older report files. Maybe we can switch that once we rename file name endings.
    optional APIType                        Api                 = 16;

    // Note: Optional for now; will become required once everything is in place
    // Hint: Stored here so that any report can be opened with any version of the tool, independent
    //       of the actual configured template for data collection. I assume this is very common that
    //       old reports would have different sections than newer reports. Same is true if multiple
    //       people share their reports.
    repeated NV.Profiler.ProfilerSection    Sections            = 17;

    repeated Comment                        SectionComments     = 18;

    // Note: Optional for now, similar to Sections
    repeated NV.RuleSystem.RuleResult       RuleResults         = 19;

    // set to true if the target was not (fully) supported
    optional bool                           UnsupportedDevice   = 20;

    optional NV.Nvtx.NvtxState              Nvtx                = 21;

    // the CUDA context ID of the kernel
    optional uint32                         ContextID           = 22;

    // the CUDA stream ID of the kernel
    optional uint32                         StreamID            = 23;

    optional ProfileSeriesInfoMessage       SeriesInfo          = 24;

    optional ProfilerClockStatusMessage     ClockStatus         = 25;

    // the CUDA context's NVTX name
    optional string                         ContextName         = 26;

    // the CUDA stream's NVTX name
    optional string                         StreamName          = 27;

    // the stack frames of the host/cpu code for this CUDA launch
    optional NV.CpuStacktrace.CpuStacktraceStackMessage CpuStack = 28;

    optional CollectionScope                CollectionScope      = 29;

    // set to true if launch was on a Green Context
    optional bool                           UsesGreenContext     = 30;

    // Demangled name renamed by the user
    optional string                         RenamedDemangledName = 31;

    // the Python call stack of the host/cpu code for this CUDA launch
    optional NV.CpuStacktrace.CpuStacktraceStackMessage PythonStack = 32;

    // The PerfWorks PartitionDeviceInfos used to scale GC metrics correctly
    optional bytes                          PartitionDeviceInfoProfiler = 33;
    optional bytes                          PartitionDeviceInfoSampler  = 34;

}

// Result representing a replayed range including multiple API calls and kernel launches.
message RangeResult
{
    // thread and api call ID of the start marker
    required uint32                         ThreadID            = 1;
    required uint64                         APICallID           = 2;

    optional string                         Comments            = 3;

    repeated ProfileMetricResult            MetricResults       = 4;

    // Represented as a Unix time since Jan 1 1970
    optional uint64                         CreationTime        = 5;

    optional APIType                        Api                 = 6;

    repeated NV.Profiler.ProfilerSection    Sections            = 7;

    repeated Comment                        SectionComments     = 8;

    // set to true if the target was not (fully) supported
    optional bool                           UnsupportedDevice   = 9;

    optional NV.Nvtx.NvtxState              Nvtx                = 10;

    // the CUDA context ID of the kernel
    optional uint32                         ContextID           = 11;

    optional ProfilerClockStatusMessage     ClockStatus         = 12;

    // the CUDA context's NVTX name
    optional string                         ContextName         = 13;

    // the stack frames of the host/cpu code for this range (start marker)
    optional NV.CpuStacktrace.CpuStacktraceStackMessage CpuStack = 14;

    // store source data references. A range can have more than one associated SourceData.
    repeated uint64                         References           = 15;

    // set to true if launch was on a Green Context
    optional bool                           UsesGreenContext     = 16;

    // the Python call stack of the host/cpu code for this range (start marker)
    optional NV.CpuStacktrace.CpuStacktraceStackMessage PythonStack = 17;

    // The PerfWorks PartitionDeviceInfos used to scale GC metrics correctly
    optional bytes                          PartitionDeviceInfoProfiler = 18;
    optional bytes                          PartitionDeviceInfoSampler  = 19;

}

// Result representing a command list with one or more launched functions.
message CmdlistResult
{
    // thread and api call ID of the start marker
    required uint32                         ThreadID             = 1;
    required uint64                         APICallID            = 2;

    optional string                         Comments             = 3;

    repeated ProfileMetricResult            MetricResults        = 4;

    // Represented as a Unix time since Jan 1 1970
    optional uint64                         CreationTime         = 5;

    optional APIType                        Api                  = 6;

    repeated NV.Profiler.ProfilerSection    Sections             = 7;

    repeated Comment                        SectionComments      = 8;

    // set to true if the target was not (fully) supported
    optional bool                           UnsupportedDevice    = 9;

    optional NV.Nvtx.NvtxState              Nvtx                 = 10;

    // the CUDA context ID of the cmdlist
    optional uint32                         ContextID            = 11;

    // the CUDA stream ID of the cmdlist
    optional uint32                         StreamID             = 12;

    optional ProfilerClockStatusMessage     ClockStatus          = 13;

    // the CUDA context's NVTX name
    optional string                         ContextName          = 14;

    // the CUDA stream's NVTX name
    optional string                         StreamName           = 15;

    // the stack frames of the host/cpu code for this cmdlist (start marker)
    optional NV.CpuStacktrace.CpuStacktraceStackMessage CpuStack = 16;

    // store source data references. A cmdlist can have more than one associated SourceData.
    repeated uint64                         References           = 17;

    // display name which may contain the original API call name
    optional string                         DisplayName          = 18;

    optional uint32                         WorkDimensions       = 19;

    message CblQmdLaunchInfo
    {
        optional Uint64x3 GridSize   = 1;
        optional Uint64x3 BlockSize  = 2;
    }

    repeated CblQmdLaunchInfo               LaunchInfos          = 20;

    // Note: Optional for now, similar to Sections
    repeated NV.RuleSystem.RuleResult       RuleResults          = 21;

    // the Python call stack of the host/cpu code for this cmdlist (start marker)
    optional NV.CpuStacktrace.CpuStacktraceStackMessage PythonStack = 22;

    // The PerfWorks PartitionDeviceInfos used to scale GC metrics correctly
    optional bytes                          PartitionDeviceInfoProfiler = 23;
    optional bytes                          PartitionDeviceInfoSampler  = 24;

}

// Result representing a CUDA Graph
message GraphResult
{
    // thread and api call ID of the start marker
    required uint32                         ThreadID            = 1;
    required uint64                         APICallID           = 2;

    optional string                         Comments            = 3;

    repeated ProfileMetricResult            MetricResults       = 4;

    // Represented as a Unix time since Jan 1 1970
    optional uint64                         CreationTime        = 5;

    optional APIType                        Api                 = 6;

    repeated NV.Profiler.ProfilerSection    Sections            = 7;

    repeated Comment                        SectionComments     = 8;

    // set to true if the target was not (fully) supported
    optional bool                           UnsupportedDevice   = 9;

    optional NV.Nvtx.NvtxState              Nvtx                = 10;

    // the CUDA context ID of the graph
    optional uint32                         ContextID           = 11;

    // the CUDA stream ID of the graph
    optional uint32                         StreamID            = 12;

    optional ProfilerClockStatusMessage     ClockStatus         = 13;

    // the CUDA context's NVTX name
    optional string                         ContextName         = 14;

    // the CUDA stream's NVTX name
    optional string                         StreamName          = 15;

    // the stack frames of the host/cpu code for this graph (start marker)
    optional NV.CpuStacktrace.CpuStacktraceStackMessage CpuStack = 16;

    // store source data references. A graph can have more than one associated SourceData.
    repeated uint64                         References           = 17;

    // set to true if launch was on a Green Context
    optional bool                           UsesGreenContext     = 18;

    // the Python call stack of the host/cpu code for this graph (start marker)
    optional NV.CpuStacktrace.CpuStacktraceStackMessage PythonStack = 19;

    // The PerfWorks PartitionDeviceInfos used to scale GC metrics correctly
    optional bytes                          PartitionDeviceInfoProfiler = 20;
    optional bytes                          PartitionDeviceInfoSampler  = 21;

}

message ProcessInfo
{
    required uint32                 ProcessID       = 1;
    optional string                 Hostname        = 2;
    optional string                 ProcessName     = 3;
}

message BlockHeader
{
    // The data for the following two fields is encoded sequentially after this header.
    // They are not *in* the header, since a large protobuf would error out. (100000 results or so)

    // Contains the number of profile sources in this block.
    // Payload is of type NV.Profiler.SourceData
    optional uint32                 NumSources      = 1;

    // Contains the number of profile results in this block.
    // Payload is of type NV.Profiler.Messages.ProfileResult
    optional uint32                 NumResults      = 2;

    optional ReportSessionDetails   SessionDetails  = 3;

    // The master string table to resolve strings stored as IDs
    optional ProfilerStringTable    StringTable     = 4;

    // Size of the following payload block in bytes
    optional uint32                 PayloadSize     = 5;

    // Process this block is associated with
    // If not present, the block is considered to be associated
    // with the default process, identified by ReportSessionDetails
    optional ProcessInfo            Process         = 6;

    // Contains the number of range results in this block.
    // Payload is of type NV.Profiler.Messages.RangeResult
    optional uint32                 NumRangeResults = 7;

    // Contains the number of cmdlist results in this block.
    // Payload is of type NV.Profiler.Messages.CmdlistResult
    optional uint32                 NumCmdlistResults = 8;

    // Contains the number of CUDA graph results in this block.
    // Payload is of type NV.Profiler.Messages.GraphResult
    optional uint32                 NumGraphResults = 9;
}

message FileHeader
{
    required uint32                 Version         = 1;
}

message BaselineColor
{
    enum BaselineColorType
    {
        BaselineColorTypeUnknown     = 0;
        BaselineColorTypeArgb        = 1;
    }

    required BaselineColorType ColorType     = 1 [default = BaselineColorTypeUnknown];
    optional uint32            ColorValue    = 2;
}

message ExtendedBaselineData
{
    optional bool          IsVisible      = 1;
    optional BaselineColor Color          = 2;
    optional string        Name           = 3;
    optional string        ProcessName    = 5;
    optional string        ReportMoniker  = 6;

    // deleted "kernel name" field
    reserved 4;
}

message BaselineResult
{
    oneof ValueType
    {
        ProfileResult       ProfileResult = 1;
        RangeResult         RangeResult = 2;
        CmdlistResult       CmdlistResult = 3;
        GraphResult         GraphResult   = 4;
    }
}

message BaselinePersistenceData
{
    // deprecated in favor of more versatile BaselineResults entry
    repeated ProfileResult        DeprecatedBaselineResults      = 1;
    repeated ProfilerStringTable  StringTables                   = 2;
    repeated ExtendedBaselineData ExtendedBaselineData           = 3;
    repeated BaselineResult       BaselineResults                = 4;
}