ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 10:37:44 +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 d4dbf93119.

[ROCm/composable_kernel commit: 03b59f8c76]
This commit is contained in:
Illia Silin
2025-09-15 08:27:04 -07:00
committed by GitHub
parent d4dbf93119
commit 954db22b39
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;

3
example/09_convnd_fwd/CMakeLists.txt
View File

@@ -1,5 +1,4 @@
add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
add_example_executable(example_convnd_fwd_xdl_fp32_tf32 convnd_fwd_xdl_fp32_tf32.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
@@ -20,4 +19,4 @@ foreach(gpu IN LISTS GPU_TARGETS)
add_example_executable(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
set(target 1)
endif()
endforeach()
endforeach()

29
example/09_convnd_fwd/convnd_fwd_common.hpp
View File

@@ -27,14 +27,10 @@ void print_helper_msg()
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
template <typename DataType, typename GemmType = DataType>
template <typename DataType>
inline __host__ __device__ constexpr double get_rtol()
{
if constexpr(std::is_same_v<DataType, float> && std::is_same_v<GemmType, ck::tf32_t>)
{
return 5e-3;
}
else if constexpr(std::is_same_v<DataType, float>)
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
@@ -72,14 +68,10 @@ inline __host__ __device__ constexpr double get_rtol()
}
}
template <typename DataType, typename GemmType = DataType>
template <typename DataType>
inline __host__ __device__ constexpr double get_atol()
{
if constexpr(std::is_same_v<DataType, float> && std::is_same_v<GemmType, ck::tf32_t>)
{
return 1e-2;
}
else if constexpr(std::is_same_v<DataType, float>)
if constexpr(std::is_same_v<DataType, float>)
{
return 1e-3;
}
@@ -124,8 +116,7 @@ template <ck::index_t NDimSpatial,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp,
typename DeviceConvNDFwdInstance,
typename ComputeDataType = OutDataType>
typename DeviceConvNDFwdInstance>
bool run_grouped_conv_fwd(bool do_verification,
int init_method,
bool time_kernel,
@@ -237,11 +228,7 @@ bool run_grouped_conv_fwd(bool do_verification,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
0,
0,
0,
ComputeDataType>();
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in,
@@ -262,8 +249,8 @@ bool run_grouped_conv_fwd(bool do_verification,
return ck::utils::check_err(out_device,
out_host,
"Error: incorrect results!",
get_rtol<OutDataType, ComputeDataType>(),
get_atol<OutDataType, ComputeDataType>());
get_rtol<OutDataType>(),
get_atol<OutDataType>());
}
return true;

89
example/09_convnd_fwd/convnd_fwd_xdl_fp32_tf32.cpp
View File

@@ -1,89 +0,0 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#define EXAMPLE_WITH_COMPUTE_DATATYPE
using InDataType = float;
using WeiDataType = float;
using AccDataType = float;
using CShuffleDataType = float;
using OutDataType = float;
using ComputeDataType = ck::tf32_t;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial, typename InLayout, typename WeiLayout, typename OutLayout>
using DeviceGroupedConvNDFwdInstance =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
NDimSpatial,
InLayout, // ALayout
WeiLayout, // BLayout
ck::Tuple<>, // DsLayout
OutLayout, // ELayout
InDataType, // ADataType
WeiDataType, // BDataType
AccDataType, // AccDataType
CShuffleDataType, // CShuffleDataType
ck::Tuple<>, // DsDataType
OutDataType, // EDataType
InElementOp, // AElementwiseOperation
WeiElementOp, // BElementwiseOperation
OutElementOp, // CDEElementwiseOperation
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, // NumGemmKPrefetchStage
256, // BlockSize
128, // MPerBlock
192, // NPerBlock
16, // KPerBlock
4, // AK1
4, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
3, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
4, // ABlockTransferSrcScalarPerVector
4, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
4, // BBlockTransferSrcScalarPerVector
4, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 16, 1, 16>, // CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
4, // CDEBlockTransferScalarPerVector_NPerBlock
ComputeDataType, // AComputeDataType
ComputeDataType, // BComputeDataType
ck::LoopScheduler::Default, // LoopScheduler
1 // NumGroupsToMerge
>;
#include "run_convnd_fwd_example.inc"
int main(int argc, char* argv[]) { return run_convnd_fwd_example(argc, argv) ? 0 : 1; }
#undef EXAMPLE_WITH_COMPUTE_DATATYPE

4
example/09_convnd_fwd/convnd_fwd_xdl_fp8.cpp
View File

@@ -7,8 +7,6 @@
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#define EXAMPLE_WITH_COMPUTE_DATATYPE
using InDataType = ck::f8_t;
using WeiDataType = ck::f8_t;
using AccDataType = float;
@@ -89,5 +87,3 @@ int main(int argc, char* argv[])
}
return run_convnd_fwd_example(argc, argv) ? 0 : 1;
}
#undef EXAMPLE_WITH_COMPUTE_DATATYPE

27
example/09_convnd_fwd/run_convnd_fwd_example.inc
View File

@@ -3,11 +3,6 @@
#pragma once
// use macro to minimize code change
#ifndef EXAMPLE_WITH_COMPUTE_DATATYPE
using ComputeDataType = AccDataType;
#endif
bool run_convnd_fwd_example(int argc, char* argv[])
{
print_helper_msg();
@@ -70,17 +65,17 @@ bool run_convnd_fwd_example(int argc, char* argv[])
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdInstance<ndim_spatial_value, InLayout, WeiLayout, OutLayout>,
ComputeDataType>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
DeviceGroupedConvNDFwdInstance<ndim_spatial_value, InLayout, WeiLayout, OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
};
namespace ctc = ck::tensor_layout::convolution;