ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 10:09:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

MX GEMM - Parameterized Test Template (#2088)

Browse Source 
* Tests for MX FP8 GEMM

* Improve documentation


[ROCm/composable_kernel commit: 213b203a3c]
This commit is contained in:
Andriy Roshchenko
2025-04-16 19:56:00 -06:00
committed by GitHub
parent a96e96ab4e
commit 7972a39081
12 changed files with 948 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp
View File

@@ -22,6 +22,7 @@ namespace ck {
namespace tensor_operation {
namespace device {
// clang-format off
/**
* \brief WIP: Implements XDL CShuffle V3 GEMM for microscale-compliant data types
*
@@ -31,8 +32,8 @@ namespace device {
* Assumptions:
* - A and B data types are compliant with the OCP Microscaling Formats (MX) Specification
* - Each scale applies to ScaleBlockSize elements in K direction
* - A scale matrix is row-major
* - B scale matrix is column-major
* - A scale matrix is a row-major
* - B scale matrix is a column-major
* - Scale data types must have get_exponent_value() specialization, whereas lowest 8 bits of the
* exponent will be interpreted as conventional biased Float32 exponent (E8M0)
*
@@ -72,10 +73,10 @@ namespace device {
* for(int mw = m0; mw < m0 + MWaves * MPerXDL; mw += MPerXDL){
* for(int nw = n0; nw < n0 + NWaves * NPerXDL; nw += NPerXDL){
* for(int k0 = kb; k0 < kb + KPerBlock; k0 += mfma.num_input_blks*KPack){
* // MFMA accumulation for multirate instructions
* for(int k_pack = k0; k_pack < k0 + mfma.num_input_blks*KPack; k_pack += KPack){
* for(int k_mfma = k_pack; k_mfma < k_pack + KPack; k_mfma += mfma.k_per_blk){
* // MFMA instruction
* // MFMA accumulation
* for(int k_pack = k0; k_pack < k0 + mfma.num_input_blks*KPack; k_pack += KPerXdlops){
* // MFMA instruction
* for(int k_mfma = k_pack; k_mfma < k_pack + KPerXdlops; k_mfma += mfma.k_per_blk){
* for(int m = mw; m < mw + MPerXDL; m++){
* for(int n = nw; n < nw + NPerXDL; n++){
* for(int k = k_mfma; k < k_mfma + mfma.k_per_blk; k++){
@@ -96,6 +97,7 @@ namespace device {
* \endcode
*
*/
// clang-format on
template <typename ALayout,
typename BLayout,
typename CLayout,
@@ -104,7 +106,7 @@ template <typename ALayout,
typename BDataType,
typename BScaleDataType,
typename CDataType,
typename GemmAccDataType,
typename GemmAccDataType, // TODO: always float
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,

111
library/include/ck/library/tensor_operation_instance/gpu/gemm_mx.hpp Normal file
View File

@@ -0,0 +1,111 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_default_instances(
std::vector<std::unique_ptr<DeviceGemmMX<Row,
Col,
Row,
F8,
e8m0_bexp_t,
F8,
e8m0_bexp_t,
F16,
32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_default_instances(
std::vector<std::unique_ptr<DeviceGemmMX<Row,
Col,
Row,
F8,
e8m0_bexp_t,
F8,
e8m0_bexp_t,
BF16,
32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
template <typename ADataType,
typename AScaleDataType,
typename BDataType,
typename BScaleDataType,
typename CDataType,
index_t ScaleBlockSize,
typename ALayout,
typename BLayout,
typename CLayout>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmMX<ALayout,
BLayout,
CLayout,
ADataType,
AScaleDataType,
BDataType,
BScaleDataType,
CDataType,
ScaleBlockSize,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
{
using DeviceOp = DeviceGemmMX<ALayout,
BLayout,
CLayout,
ADataType,
AScaleDataType,
BDataType,
BScaleDataType,
CDataType,
ScaleBlockSize,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> && is_same_v<CLayout, Row>)
{
if constexpr(is_same_v<ADataType, F8> && is_same_v<BDataType, F8> &&
is_same_v<CDataType, F16>)
{
add_device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_default_instances(op_ptrs);
}
if constexpr(is_same_v<ADataType, F8> && is_same_v<BDataType, F8> &&
is_same_v<CDataType, BF16>)
{
add_device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_default_instances(op_ptrs);
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

13
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View File

@@ -60,6 +60,13 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
# Do not build MX instances if gfx950 targets are not on the target list
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx950" AND source MATCHES "_mx")
message("removing MX instance ${source} ")
list(REMOVE_ITEM ARGN "${source}")
endif()
endforeach()
# Do not build WMMA instances if gfx11 targets are not on the target list
foreach(source IN LISTS ARGN)
if(NOT INST_TARGETS MATCHES "gfx11" AND NOT INST_TARGETS MATCHES "gfx12" AND source MATCHES "_wmma")
@@ -100,6 +107,8 @@ function(add_instance_library INSTANCE_NAME)
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx950)
elseif(source MATCHES "mha")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
elseif(source MATCHES "_mx")
list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack- gfx908:xnack+ gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
endif()
#only build the fp8 gemm instances for gfx90a if the build argument is set, otherwise only build for gfx942/gfx950
if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
@@ -234,6 +243,10 @@ FOREACH(subdir_path ${dir_list})
if(("${cmake_instance}" MATCHES "ONLY XDL_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx9"))
message("Found only xdl instances, but gfx9 is not on the targets list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "ONLY MX_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx950"))
message("Found only MX instances, but gfx950 is not on the targets list. Skipping.")
set(add_inst 0)
endif()
if(("${cmake_instance}" MATCHES "ONLY WMMA_KERNELS") AND (NOT INST_TARGETS MATCHES "gfx11") AND (NOT INST_TARGETS MATCHES "gfx12"))
message("Found only wmma instances, but gfx11 is not on the targets list. Skipping.")

14
library/src/tensor_operation_instance/gpu/gemm_mx/CMakeLists.txt Normal file
View File

@@ -0,0 +1,14 @@
# ONLY MX_KERNELS
set(GEMM_MX_INSTANCES)
list(APPEND GEMM_MX_INSTANCES
device_gemm_mx_xdl_f8_f8_f16/device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_default_instance.cpp
device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_default_instance.cpp
)
set_source_files_properties(device_gemm_mx_xdl_f8_f8_f16/device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
set_source_files_properties(device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
add_instance_library(device_gemm_mx_instance ${GEMM_MX_INSTANCES})

63
library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn.hpp Normal file
View File

@@ -0,0 +1,63 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F8 = f8_t;
using F16 = half_t;
using BF16 = bhalf_t;
using F32 = float;
using E8M0 = ck::e8m0_bexp_t;
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
template <index_t... Is>
using S = Sequence<Is...>;
using PassThrough = element_wise::PassThrough;
static constexpr auto GemmDefault = GemmSpecialization::Default;
static constexpr auto GemmKPadding = GemmSpecialization::KPadding;
static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
static constexpr auto ScaleBlockSize = 32;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData|AScale|BData|BScale| CData| AccData| Cshuffle| A| B| C| GEMM| Scale Block| 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| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Data| Type| Data| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | Type| | Type| | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx950__) || defined(CK_USE_NATIVE_MX_SUPPORT)
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 128, 128, 16, 128, 16, 16, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 256, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 64, 16, 16, 512, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>
//Require verification
//DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

32
library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_bf16/device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_default_instance.cpp Normal file
View File

@@ -0,0 +1,32 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_default_instances(
std::vector<std::unique_ptr<DeviceGemmMX<Row,
Col,
Row,
F8,
E8M0,
F8,
E8M0,
BF16,
32,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_mx_xdl_f8_f8_bf16_mk_nk_mn_instances<Intrawave, GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

63
library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_f16/device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn.hpp Normal file
View File

@@ -0,0 +1,63 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_mx.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F8 = f8_t;
using F16 = half_t;
using BF16 = bhalf_t;
using F32 = float;
using E8M0 = ck::e8m0_bexp_t;
using Row = tensor_layout::gemm::RowMajor;
using Col = tensor_layout::gemm::ColumnMajor;
template <index_t... Is>
using S = Sequence<Is...>;
using PassThrough = element_wise::PassThrough;
static constexpr auto GemmDefault = GemmSpecialization::Default;
static constexpr auto GemmKPadding = GemmSpecialization::KPadding;
static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
static constexpr auto ScaleBlockSize = 32;
template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
using device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_instances = std::tuple<
// clang-format off
//#########################| ALayout| BLayout| CLayout|AData|AScale|BData|BScale| CData| AccData| Cshuffle| A| B| C| GEMM| Scale Block| 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| Block-wiseGemm| Block-wiseGemm|
//#########################| | | | Type| Data| Type| Data| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| 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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//#########################| | | | | Type| | Type| | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
#if defined(__gfx950__) || defined(CK_USE_NATIVE_MX_SUPPORT)
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 128, 128, 16, 128, 16, 16, 16, 16, 4, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 256, 16, 16, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, false, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 128, 128, 128, 16, 16, 32, 32, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 64, 16, 16, 512, 16, 16, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>
//Require verification
//DeviceGemmMX_Xdl_CShuffleV3< Row, Col, Row, F8, E8M0, F8, E8M0, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, ScaleBlockSize, 256, 256, 256, 128, 16, 16, 16, 16, 8, 8, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, 8, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
#endif
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

32
library/src/tensor_operation_instance/gpu/gemm_mx/device_gemm_mx_xdl_f8_f8_f16/device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_default_instance.cpp Normal file
View File

@@ -0,0 +1,32 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_default_instances(
std::vector<std::unique_ptr<DeviceGemmMX<Row,
Col,
Row,
F8,
E8M0,
F8,
E8M0,
F16,
32,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances, device_gemm_mx_xdl_f8_f8_f16_mk_nk_mn_instances<Intrawave, GemmDefault>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

1
test/CMakeLists.txt
View File

@@ -279,6 +279,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx942" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9
endif()
if(SUPPORTED_GPU_TARGETS MATCHES "gfx950")
add_subdirectory(mx_mfma_op)
add_subdirectory(gemm_mx)
endif()
add_subdirectory(position_embedding)
add_subdirectory(scatter_gather)

4
test/gemm_mx/CMakeLists.txt Normal file
View File

@@ -0,0 +1,4 @@
add_gtest_executable(test_gemm_mx test_gemm_mx.cpp)
if(result EQUAL 0)
target_link_libraries(test_gemm_mx PRIVATE utility device_gemm_mx_instance)
endif()

108
test/gemm_mx/test_gemm_mx.cpp Normal file
View File

@@ -0,0 +1,108 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <tuple>
#include "gtest/gtest.h"
#include "test_gemm_mx_util.hpp"
using E8M0 = ck::e8m0_bexp_t;
using F8 = ck::f8_t;
using BF8 = ck::bf8_t;
using F6 = ck::f6_t;
using BF6 = ck::bf6_t;
using F4 = ck::f4_t;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
namespace {
template <typename X, typename Y>
struct tuple_concat;
template <typename... Xs, typename... Ys>
struct tuple_concat<std::tuple<Xs...>, std::tuple<Ys...>>
{
using type = std::tuple<Xs..., Ys...>;
};
} // namespace
template <typename Tuple>
class TestGemmMX_MK_NK
: public ck::test::TestGemmMX<typename tuple_concat<std::tuple<Row, Col>, Tuple>::type>
{
};
// clang-format off
using KernelTypes_MK_NK = ::testing::Types<
#if defined(CK_ENABLE_FP8)
// ADataType, BDataType, CDataType, ScaleBlockSize
std::tuple< F8, F8, F16, ck::Number<32> >,
std::tuple< F8, F8, BF16, ck::Number<32> >
#endif
>;
// clang-format on
TYPED_TEST_SUITE(TestGemmMX_MK_NK, KernelTypes_MK_NK);
TYPED_TEST(TestGemmMX_MK_NK, SmallM)
{
std::vector<int> Ms{1, 2, 3, 4, 5, 6};
constexpr int N = 256;
constexpr int K = 512;
constexpr int StrideA = K;
constexpr int StrideB = K;
constexpr int StrideC = N;
for(int M : Ms)
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
TYPED_TEST(TestGemmMX_MK_NK, MidLargeM)
{
std::vector<int> Ms{127, 255, 312, 799, 1573};
constexpr int N = 256;
constexpr int K = 512;
constexpr int StrideA = K;
constexpr int StrideB = K;
constexpr int StrideC = N;
for(int M : Ms)
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
TYPED_TEST(TestGemmMX_MK_NK, Regular)
{
std::vector<int> Ms{3840};
constexpr int N = 512;
constexpr int K = 1024;
constexpr int StrideA = K;
constexpr int StrideB = K;
constexpr int StrideC = N;
for(int M : Ms)
this->Run(M, N, K, StrideA, StrideB, StrideC);
}
TYPED_TEST(TestGemmMX_MK_NK, Large)
{
std::vector<int> Ms{4096};
constexpr int N = 3840;
constexpr int K = 4096;
constexpr int StrideA = K;
constexpr int StrideB = K;
constexpr int StrideC = N;
for(int M : Ms)
this->Run(M, N, K, StrideA, StrideB, StrideC);
}

498
test/gemm_mx/test_gemm_mx_util.hpp Normal file
View File

@@ -0,0 +1,498 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <gtest/gtest.h>
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/utility/number.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/fill.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_mx.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_mx.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_mx_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
namespace ck {
namespace test {
namespace {
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
} // namespace
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
int ScaleBlockSize>
bool profile_gemm_mx_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideC,
int KBatch,
int n_warmup,
int n_iter,
uint64_t rotating = 0)
{
if(K % ScaleBlockSize != 0)
{
throw std::runtime_error("wrong! K must be multiple of ScaleBlockSize.");
};
using ScaleDataType = e8m0_bexp_t;
using AScaleLayout = Row;
using BScaleLayout = Col;
bool pass = true;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
auto f_get_default_stride =
[](ck::index_t row, ck::index_t col, ck::index_t stride, auto layout) {
if(stride == -1)
{
// give a chance if stride is -1, return a default packed stride
if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
{
return static_cast<ck::index_t>(col);
}
else
{
return static_cast<ck::index_t>(row);
}
}
else
return static_cast<ck::index_t>(stride);
};
auto Scale_Stride_AM = f_get_default_stride(M, K / ScaleBlockSize, -1, AScaleLayout{});
auto Scale_Stride_BN = f_get_default_stride(K / ScaleBlockSize, N, -1, BScaleLayout{});
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<ScaleDataType> a_m_k_scale(f_host_tensor_descriptor(
M, K / ScaleBlockSize, Scale_Stride_AM, AScaleLayout{})); // scales for A
Tensor<ScaleDataType> b_k_n_scale(f_host_tensor_descriptor(
K / ScaleBlockSize, N, Scale_Stride_BN, BScaleLayout{})); // scales for B
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
std::size_t total_gemm_needed =
a_m_k.GetElementSpaceSizeInBytes() + b_k_n.GetElementSpaceSizeInBytes() +
a_m_k_scale.GetElementSpaceSizeInBytes() + b_k_n_scale.GetElementSpaceSizeInBytes();
int rotating_count = std::max(
1,
std::min(n_iter,
static_cast<int>(std::ceil(static_cast<double>(rotating) / total_gemm_needed))));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "a_m_k_scale: " << a_m_k_scale.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "b_k_n_scale: " << b_k_n_scale.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
std::cout << "rotating count: " << rotating_count << std::endl;
switch(init_method)
{
case 0: // Initializations for development and debugging
ck::utils::FillConstant<ADataType>{ck::type_convert<ADataType>(1.0f)}(a_m_k);
ck::utils::FillConstant<ScaleDataType>{ck::type_convert<ScaleDataType>(2.0f)}(a_m_k_scale);
ck::utils::FillConstant<BDataType>{ck::type_convert<BDataType>(0.5f)}(b_k_n);
ck::utils::FillConstant<ScaleDataType>{ck::type_convert<ScaleDataType>(1.0f)}(b_k_n_scale);
if(do_log)
{
std::cout << "Init A = {1}" << std::endl;
std::cout << "Init A scale = {2.0}" << std::endl;
std::cout << "Init B = {0.5}" << std::endl;
std::cout << "Init B scale = {1.0}" << std::endl;
std::cout << "Expect C = {K}" << std::endl;
}
break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-4, 5}); // Z[-4,4]
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-4, 5}); // Z[-4,4]
a_m_k_scale.GenerateTensorValue(
GeneratorTensor_2<ScaleDataType>{125, 129}); // scales: {0.25, 0.5, 1, 2}
b_k_n_scale.GenerateTensorValue(
GeneratorTensor_2<ScaleDataType>{125, 129}); // scales: {0.25, 0.5, 1, 2}
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<BDataType>{-2.0, 2.0});
a_m_k_scale.GenerateTensorValue(
GeneratorTensor_3<ScaleDataType>{powf(2.0f, -125.0f), 1.0f}); // R[2^-125, 1]
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-2.0, 2.0});
b_k_n_scale.GenerateTensorValue(
GeneratorTensor_3<ScaleDataType>{powf(2.0f, -125.0f), 1.0f});
break;
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
if(do_log > 0)
std::cout << "Device memory allocation..." << std::endl;
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem a_scale_device_buf(sizeof(ScaleDataType) * a_m_k_scale.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem b_scale_device_buf(sizeof(ScaleDataType) * b_k_n_scale.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
if(do_log > 0)
std::cout << "Upload data to device..." << std::endl;
a_device_buf.ToDevice(a_m_k.mData.data());
a_scale_device_buf.ToDevice(a_m_k_scale.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
b_scale_device_buf.ToDevice(b_k_n_scale.mData.data());
if(do_log > 0)
std::cout << "Done." << std::endl;
using DeviceOp = ck::tensor_operation::device::DeviceGemmMX<ALayout,
BLayout,
CLayout,
ADataType,
ScaleDataType,
BDataType,
ScaleDataType,
CDataType,
ScaleBlockSize,
AElementOp,
BElementOp,
CElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// Run reference GEMM
if(do_verification)
{
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceMXGemm<ADataType,
BDataType,
CDataType,
float, // AccDataType
ScaleDataType,
AElementOp,
BElementOp,
CElementOp,
float, // ComputeTypeA
float // ComputeTypeB
>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k,
a_m_k_scale,
b_k_n,
b_k_n_scale,
c_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
}
std::string best_op_name;
std::optional<std::string> best_op_object_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
std::vector<int> kbatch_list = {1, 2, 4, 8, 16, 19, 32, 38}; // use these when KBatch <= 0
if(KBatch > 0)
{
kbatch_list = {KBatch};
}
for(std::size_t i = 0; i < kbatch_list.size(); i++)
{
auto kbatch_curr = kbatch_list[i];
auto argument_ptr = op_ptr->MakeArgumentPointer(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<ScaleDataType*>(a_scale_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<ScaleDataType*>(b_scale_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
Scale_Stride_AM,
StrideB,
Scale_Stride_BN,
StrideC,
kbatch_curr,
a_element_op,
b_element_op,
c_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
if(do_log)
{
if(init_method == 0)
{
auto expected = static_cast<float>(K);
auto computed = type_convert<float>(c_m_n_device_result(0, 12));
pass = pass & (std::abs(expected - computed) <= 0.0f);
std::cout << "\nExpected vs Computed: " << expected << " vs "
<< computed << ((pass) ? " (PASSED!)" : " (FAILED!)")
<< std::endl
<< std::endl;
}
else
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "a_scale : ", a_m_k_scale.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b_scale: ", b_k_n_scale.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
}
std::string op_name = op_ptr->GetTypeString();
std::optional<std::string> op_obj_name = op_ptr->GetObjectName();
float ave_time = invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr,
time_kernel,
0,
n_warmup,
n_iter,
rotating_count > 1,
rotating_count});
// Output size(M*N) * [dot product(2K) + product of scales(K/ScaleBlockSize) +
// scaling of partial sums(K/ScaleBlockSize)]
// FLOPS = 2 * M * N * K + 2 * M * N * K / ScaleBlockSize
std::size_t flop =
std::size_t(2) * M * N * K + std::size_t(2) * M * N * K / ScaleBlockSize;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N +
sizeof(ScaleDataType) * (M * K + K * N) / ScaleBlockSize;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
if(tflops > best_tflops && ave_time > 1e-10)
{
best_op_name = op_name;
best_op_object_name = op_obj_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem"
<< std::endl;
}
}
}
if constexpr(is_same<CDataType, float>::value)
{
std::cout << "Best Perf for datatype = f32";
}
else if constexpr(is_same<CDataType, half_t>::value)
{
std::cout << "Best Perf for datatype = f16";
}
else if constexpr(is_same<CDataType, bhalf_t>::value)
{
std::cout << "Best Perf for datatype = bf16";
}
else if constexpr(is_same<CDataType, int8_t>::value)
{
std::cout << "Best Perf for datatype = int8";
}
if constexpr(is_same<ALayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " ALayout = RowMajor";
}
else if constexpr(is_same<ALayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " ALayout = ColumnMajor";
}
if constexpr(is_same<BLayout, tensor_layout::gemm::RowMajor>::value)
{
std::cout << " BLayout = RowMajor";
}
else if constexpr(is_same<BLayout, tensor_layout::gemm::ColumnMajor>::value)
{
std::cout << " BLayout = ColumnMajor";
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
if(best_op_object_name)
std::cout << best_op_object_name.value() << std::endl;
return pass;
}
template <typename Tuple>
class TestGemmMX : public testing::Test
{
using Row = ck::tensor_layout::gemm::RowMajor;
using F32 = float;
using ScaleType = e8m0_bexp_t;
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CLayout = Row;
using ADataType = std::tuple_element_t<2, Tuple>;
using BDataType = std::tuple_element_t<3, Tuple>;
using CDataType = std::tuple_element_t<4, Tuple>;
using AccDataType = float;
public:
static constexpr index_t ScaleBlockSize = std::tuple_element_t<5, Tuple>{};
static constexpr bool verify_ = true;
static constexpr int init_method_ = 2; // decimal value initialization
static constexpr bool log_ = false;
static constexpr bool bench_ = false; // measure kernel performance
std::vector<int> k_batches_;
void SetUp() override { k_batches_ = {1}; }
void Run(const int M,
const int N,
const int K,
const int StrideA,
const int StrideB,
const int StrideC)
{
for(auto kb : k_batches_)
{
RunSingle(M, N, K, StrideA, StrideB, StrideC, kb);
}
}
void RunSingle(const int M,
const int N,
const int K,
const int StrideA,
const int StrideB,
const int StrideC,
int kbatch = 1,
int n_warmup = 1,
int n_iter = 10)
{
bool pass = ck::test::profile_gemm_mx_impl<ADataType,
BDataType,
CDataType,
ALayout,
BLayout,
CLayout,
ScaleBlockSize>(verify_,
init_method_,
log_,
bench_,
M,
N,
K,
StrideA,
StrideB,
StrideC,
kbatch,
n_warmup,
n_iter);
EXPECT_TRUE(pass);
}
};
} // namespace test
} // namespace ck