ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-19 22:39:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

Revert "feature:tf32:add initial conv3d fwd kernel support (#2763)" (#2848)

Browse Source 
This reverts commit c51102144f.
This commit is contained in:
Illia Silin
2025-09-15 08:27:04 -07:00
committed by GitHub
parent c51102144f
commit 03b59f8c76
44 changed files with 175 additions and 1085 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
example/01_gemm/CMakeLists.txt
View File

@@ -87,9 +87,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
add_example_executable(example_gemm_xdl_lds_direct_load_fp32 gemm_xdl_lds_direct_load_fp32.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_lds_direct_load_fp32)
add_example_executable(example_gemm_xdl_lds_direct_load_fp32_tf32 gemm_xdl_lds_direct_load_fp32_tf32.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_lds_direct_load_fp32_tf32)
add_example_executable(example_gemm_xdl_lds_direct_load_fp16 gemm_xdl_lds_direct_load_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_lds_direct_load_fp16)

16
example/01_gemm/common.hpp
View File

@@ -310,14 +310,10 @@ bool parse_cmd_args<ProblemSizeSplitK>(int argc,
return true;
}
template <typename DataType, typename ComputeDataType = DataType>
template <typename DataType>
inline __host__ __device__ constexpr double get_rtol()
{
if constexpr(std::is_same_v<DataType, float> && std::is_same_v<ComputeDataType, ck::tf32_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, float>)
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
@@ -355,14 +351,10 @@ inline __host__ __device__ constexpr double get_rtol()
}
}
template <typename DataType, typename ComputeDataType = DataType>
template <typename DataType>
inline __host__ __device__ constexpr double get_atol()
{
if constexpr(std::is_same_v<DataType, float> && std::is_same_v<ComputeDataType, ck::tf32_t>)
{
return 1e-3;
}
else if constexpr(std::is_same_v<DataType, float>)
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}

85
example/01_gemm/gemm_xdl_lds_direct_load_fp32_tf32.cpp
View File

@@ -1,85 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include "common.hpp"
#define USING_DIRECT_LOADS 1
#if USING_DIRECT_LOADS
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_lds_direct_load.hpp"
#else
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
#endif
#define EXAMPLE_WITH_COMPUTE_DATATYPE
using F32 = float;
using ADataType = F32;
using BDataType = F32;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = F32;
using ComputeDataType = ck::tf32_t;
using ALayout = Row;
using BLayout = Col;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
#if USING_DIRECT_LOADS
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad
// ######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer|
// ######| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockLds|
// ######| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| LoopScheduler | pipeline ver | gemm type |
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block|
// ######| XDL| XDL| Per| Per| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraM| ThreadCluster| SrcAccessOrder| SrcVectorDim| Scalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| | | PerVector| | Lengths_K0_N_K1| | | PerVector| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 128, 128, 32,
8, 8, 32, 32, 2, 2, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1, S<4, 8, 8>, S<1, 0, 2>, 2, 1, 1,
1, 1, S<1, 8, 1, 8>, 4, ck::LoopScheduler::Default, ck::PipelineVersion::v4, ComputeDataType>;
// clang-format on
#else
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// ######| ALayout| BLayout| CLayout| AData| BData| CData| AccData| CShuffle| A| B| C| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// ######| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// ######| | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CElementOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 8, 1, 8>, 4>;
// clang-format on
#endif
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp,
ComputeDataType,
ComputeDataType>;
using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
#include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }
#undef EXAMPLE_WITH_COMPUTE_DATATYPE

13
example/01_gemm/run_gemm_example.inc
View File

@@ -4,11 +4,6 @@
#pragma once
#include "ck/library/utility/validation_common.hpp"
// use macro to minimize code change
#ifndef EXAMPLE_WITH_COMPUTE_DATATYPE
using ComputeDataType = AccDataType;
#endif
template <typename ProblemType>
bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
{
@@ -223,8 +218,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
pass &= ck::utils::check_err(c_m_n_device_result,
c_m_n_host_result,
"Error: Incorrect results!",
get_rtol<CDataType, ComputeDataType>(),
get_atol<CDataType, ComputeDataType>());
get_rtol<CDataType>(),
get_atol<CDataType>());
#endif
}
@@ -254,8 +249,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
pass &= ck::utils::check_err(c_m_n_device_result,
c_m_n_device_ref_result,
"Error: Incorrect results!",
get_rtol<CDataType, ComputeDataType>(),
get_atol<CDataType, ComputeDataType>());
get_rtol<CDataType>(),
get_atol<CDataType>());
}
return pass == true;