Files
rocm-systems/shared/amdgpu-windows-interop/pal/inc/gpuUtil/palCodeObjectTraceSource.h
T
Scott Todd fa772be675 Reapply amdgpu-windows-interop revert. (#1893)
## Overview and rationale

This reverts https://github.com/ROCm/rocm-systems/pull/1886, which...
* Re-applies https://github.com/ROCm/rocm-systems/pull/1866
* Reverts https://github.com/ROCm/rocm-systems/pull/1728

(So it restores the [`amdgpu-windows-interop/`](https://github.com/ROCm/rocm-systems/tree/develop/shared/amdgpu-windows-interop) folder back to the state from a few weeks ago)

The rationale for this change is at https://github.com/ROCm/rocm-systems/pull/1866:
> Last PAL update broke applications on gfx12 Windows.

## Cross-repository change details

That PR failed to build but was merged with this explanation:

> TheRock CI Windows build fails as expected with this revert.
> 
> References to these PAL members need to be stripped out in a patch on TheRock.
> 
> ```
> 11.3	C:\home\runner\_work\rocm-systems\rocm-systems\projects\clr\rocclr\device\pal\palubercapturemgr.cpp(152): error C2039: 'RegisterTraceStateChangeCallback': is not a member of 'GpuUtil::TraceSession'
> 11.4	C:\home\runner\_work\rocm-systems\rocm-systems\shared\amdgpu-windows-interop\pal\inc\gpuUtil\palTraceSession.h(372): note: see declaration of 'GpuUtil::TraceSession'
> 11.4	C:\home\runner\_work\rocm-systems\rocm-systems\projects\clr\rocclr\device\pal\palubercapturemgr.cpp(195): error C2039: 'UnregisterTraceStateChangeCallback': is not a member of 'GpuUtil::TraceSession'
> 11.4	C:\home\runner\_work\rocm-systems\rocm-systems\shared\amdgpu-windows-interop\pal\inc\gpuUtil\palTraceSession.h(372): note: see declaration of 'GpuUtil::TraceSession'
> ```

The patch in TheRock was updated in https://github.com/ROCm/TheRock/pull/2154. This rolls forward by updating the ref for TheRock.

That original PR could have been sequenced differently to avoid a build break - perhaps by
* Pointing to a branch in TheRock with the patch rebased
* Deleting the patch in the workflows here but holding a local copy of the path to be applied in workflows
* Landing the patch as a normal commit instead of carrying it at all

## Test plan

1. Watch TheRock CI here (https://github.com/ROCm/rocm-systems/actions/runs/19447202693/job/55644411119?pr=1893)
2. Build locally:
    
    ```bash
    # In rocm-systems
    git am --whitespace=nowarn D:\projects\TheRock\patches\amd-mainline\rocm-systems\0001-Revert-SWDEV-543498-Some-compute-Ubertrace-profiles-.patch
    git am --whitespace=nowarn D:\projects\TheRock\patches\amd-mainline\rocm-systems\0003-Use-is_versioned-true-consistently-in-both-Comgr-Loa.patch
    git am --whitespace=nowarn D:\projects\TheRock\patches\amd-mainline\rocm-systems\0006-Explicitly-load-libamdhip64.so.7.patch
    # Note: the build fails with the observed errors if patch 0001 is not applied!
    
    # In TheRock
    cmake -DCMAKE_BUILD_TYPE=Release \
      -DCMAKE_C_COMPILER=cl.exe -DCMAKE_CXX_COMPILER=cl.exe \
      -DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
      -DPython3_EXECUTABLE=d:/projects/TheRock/.venv/Scripts/python \
      -DTHEROCK_ROCM_SYSTEMS_SOURCE_DIR=d:/projects/TheRock/../rocm-systems \  # IMPORTANT
      -DTHEROCK_AMDGPU_FAMILIES=gfx110X-all \
      -DBUILD_TESTING=ON \
      -DTHEROCK_ENABLE_ALL=ON \
      -Damd-llvm_BUILD_TYPE=RelWithDebInfo \
      -S D:/projects/TheRock \
      -B D:/projects/TheRock/build \
      -G Ninja
    
    cmake --build D:/projects/TheRock/build --target hip-clr
    # [build] Build finished with exit code 0
    cmake --build D:/projects/TheRock/build --target ocl-clr+dist
    # [build] Build finished with exit code 0
    ```
2025-11-18 07:17:06 -08:00

213 regels
8.7 KiB
C++

/*
***********************************************************************************************************************
*
* Copyright (c) 2023-2025 Advanced Micro Devices, Inc. All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
**********************************************************************************************************************/
#pragma once
#include "palGpaSession.h"
#include "palGpuUtil.h"
#include "palTraceSession.h"
#include "palVector.h"
#include "palHashSet.h"
#include "palMutex.h"
namespace Pal
{
class IPlatform;
class IDevice;
class IShaderLibrary;
} // namespace Pal
namespace GpuUtil
{
class GpaSession;
} // namespace GpuUtil
namespace GpuUtil
{
namespace TraceChunk
{
/// "CodeObject" RDF chunk identifier & version
constexpr char CodeObjectChunkId[TextIdentifierSize] = "CodeObject";
constexpr Pal::uint32 CodeObjectChunkVersion = 2;
/// Header for the "CodeObject" RDF chunk
struct CodeObjectHeader
{
Pal::uint32 pciId; /// The ID of the GPU the trace was run on
Pal::ShaderHash codeObjectHash; /// Hash of the Code Object binary
};
/// "COLoadEvent" RDF chunk identifier & version
constexpr char CodeObjectLoadEventChunkId[TextIdentifierSize] = "COLoadEvent";
constexpr Pal::uint32 CodeObjectLoadEventChunkVersion = 3;
struct CodeObjectLoadEventHeader
{
Pal::uint32 count; /// Number of load events in this chunk
};
/// Describes whether a load event was into GPU memory or from.
enum class CodeObjectLoadEventType : Pal::uint32
{
LoadToGpuMemory = 0, /// Code Object was loaded into GPU memory
UnloadFromGpuMemory = 1 /// Code Object was unloaded from GPU memory
};
/// Describes one or more GPU load/unload(s) of a Code Object. Payload for "COLoadEvent" RDF chunk.
struct CodeObjectLoadEvent
{
Pal::uint32 pciId; /// The ID of the GPU the trace was run on
CodeObjectLoadEventType eventType; /// Type of loader event
Pal::uint64 baseAddress; /// Base address where the Code Object was loaded
Pal::ShaderHash codeObjectHash; /// Hash of the (un)loaded Code Object binary
Pal::uint64 timestamp; /// CPU timestamp of this event being triggered
};
/// "PsoCorrelation" RDF chunk identifier & version
constexpr char PsoCorrelationChunkId[TextIdentifierSize] = "PsoCorrelation";
constexpr Pal::uint32 PsoCorrelationChunkVersion = 3;
struct PsoCorrelationHeader
{
Pal::uint32 count; /// Number of PSO correlations in this chunk
};
/// Payload for the "PsoCorrelation" RDF chunks
struct PsoCorrelation
{
Pal::uint32 pciId; /// The ID of the GPU the trace was run on
Pal::uint64 apiPsoHash; /// Hash of the API-level Pipeline State Object
Pal::PipelineHash internalPipelineHash; /// Hash of all inputs to the pipeline compiler
char apiLevelObjectName[64]; /// Debug object name (null-terminated)
};
/// "COCorrelation" RDF chunk identifier & version
constexpr char CodeObjectCorrelationChunkId[TextIdentifierSize] = "COCorrelation";
constexpr uint32_t CodeObjectCorrelationChunkVersion = 4;
struct CodeObjectCorrelationHeader
{
Pal::uint32 count; /// Number of Code Object Correlations in this chunk
};
/// Payload for the "CodeObjectCorrelation" RDF chunks
struct CodeObjectCorrelation
{
Pal::PipelineHash internalPipelineHash; /// Hash of all inputs to the pipeline compiler
Pal::ShaderHash codeObjectHash; /// Hash of the Code Object binary in the CO Database
Pal::uint32 containsMetadata : 1; /// 1 if the code object contains metadata, 0 otherwise
Pal::uint32 reserved : 31; /// Bitflags reserved for future use
};
} // namespace TraceChunk
/// CodeObject Trace Source name & version
constexpr char CodeObjectTraceSourceName[] = "codeobject";
constexpr Pal::uint32 CodeObjectTraceSourceVersion = 3;
// =====================================================================================================================
class CodeObjectTraceSource : public ITraceSource
{
public:
CodeObjectTraceSource(Pal::IPlatform* pPlatform);
~CodeObjectTraceSource();
// ==== TraceSource Native Functions ========================================================================== //
Pal::Result RegisterPipeline(const Pal::IPipeline* pPipeline, const RegisterPipelineInfo& clientInfo);
Pal::Result UnregisterPipeline(const Pal::IPipeline* pPipeline);
Pal::Result RegisterLibrary(const Pal::IShaderLibrary* pLibrary, const RegisterLibraryInfo& clientInfo);
Pal::Result UnregisterLibrary(const Pal::IShaderLibrary* pLibrary);
Pal::Result RegisterElfBinary(const ElfBinaryInfo& elfBinaryInfo);
Pal::Result UnregisterElfBinary(const ElfBinaryInfo& elfBinaryInfo);
// ==== Base Class Overrides =================================================================================== //
virtual void OnConfigUpdated(DevDriver::StructuredValue* pJsonConfig) override { }
virtual Pal::uint64 QueryGpuWorkMask() const override { return 0; }
#if PAL_CLIENT_INTERFACE_MAJOR_VERSION >= 908
virtual void OnTraceAccepted(Pal::uint32 gpuIndex, Pal::ICmdBuffer* pCmdBuf) override { }
#else
virtual void OnTraceAccepted() override { }
#endif
virtual void OnTraceBegin(Pal::uint32 gpuIndex, Pal::ICmdBuffer* pCmdBuf) override { }
virtual void OnTraceEnd(Pal::uint32 gpuIndex, Pal::ICmdBuffer* pCmdBuf) override { }
#if PAL_CLIENT_INTERFACE_MAJOR_VERSION >= 939
virtual void OnPostambleEnd(
Pal::uint32 gpuIndex,
Pal::ICmdBuffer* pCmdBuf) override { }
#endif
virtual void OnTraceFinished() override;
virtual const char* GetName() const override { return CodeObjectTraceSourceName; }
virtual Pal::uint32 GetVersion() const override { return CodeObjectTraceSourceVersion; }
private:
Pal::Result RegisterSinglePipeline(const Pal::IPipeline* pPipeline, const RegisterPipelineInfo& clientInfo);
Pal::Result UnregisterSinglePipeline(const Pal::IPipeline* pPipeline);
Pal::Result AddCodeObjectLoadEvent(
const Pal::IShaderLibrary* pLibrary,
TraceChunk::CodeObjectLoadEventType eventType);
Pal::Result AddCodeObjectLoadEvent(
const Pal::IPipeline* pLibrary,
TraceChunk::CodeObjectLoadEventType eventType);
Pal::Result AddCodeObjectLoadEvent(
const ElfBinaryInfo& elfBinaryInfo,
TraceChunk::CodeObjectLoadEventType eventType);
Pal::Result WriteCodeObjectChunks();
Pal::Result WriteLoaderEventsChunk();
Pal::Result WritePsoCorrelationChunk();
Pal::Result WriteCoCorrelationChunk();
struct CodeObjectDatabaseRecord
{
Pal::uint32 recordSize;
Pal::ShaderHash codeObjectHash;
};
Pal::IPlatform* const m_pPlatform;
Util::RWLock m_registerPipelineLock;
Util::Vector<CodeObjectDatabaseRecord*, 1, Pal::IPlatform> m_codeObjectRecords;
Util::Vector<TraceChunk::CodeObjectLoadEvent, 1, Pal::IPlatform> m_loadEventRecords;
Util::Vector<TraceChunk::PsoCorrelation, 1, Pal::IPlatform> m_psoCorrelationRecords;
Util::Vector<TraceChunk::CodeObjectCorrelation, 1, Pal::IPlatform> m_coCorrelationRecords;
// API hashes -> internal pipeline hash (-> child code object hashes)
Util::HashSet<Pal::uint64, Pal::IPlatform, Util::JenkinsHashFunc> m_registeredApiHashes;
Util::HashSet<Pal::uint64, Pal::IPlatform, Util::JenkinsHashFunc> m_registeredPipelines;
Util::HashSet<Pal::uint64, Pal::IPlatform, Util::JenkinsHashFunc> m_registeredCoHashes;
};
} // namespace GpuUtil