ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 20:40:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

MX GEMM - Add FP8 GEMM Tests for Different Layouts (#2152)

Browse Source 
* Add gemm_mx_fp8_bf8 example with row-major B

* Add more overloads of MX MFMA instructions

* Add MK_KN (RRR) tests

* Add KM_NK (CCR) tests

* Add more problem sizes to Large tests

* Add test_gemm_mx to the list of regression tests

[ROCm/composable_kernel commit: 79b0bfeb41]
This commit is contained in:
Andriy Roshchenko
2025-05-01 11:55:48 -06:00
committed by GitHub
parent 426ed45111
commit ac743ddf0b
15 changed files with 642 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
example/67_gemm_microscaling/CMakeLists.txt
View File

@@ -6,3 +6,6 @@ add_example_dependencies(example_gemm_mx example_gemm_mx_fp8)
add_example_executable(example_gemm_mx_bf8 gemm_mx_bf8.cpp)
add_example_dependencies(example_gemm_mx example_gemm_mx_bf8)
add_example_executable(example_gemm_mx_fp8_bf8 gemm_mx_fp8_bf8.cpp)
add_example_dependencies(example_gemm_mx example_gemm_mx_fp8_bf8)

2
example/67_gemm_microscaling/gemm_mx_common.hpp
View File

@@ -235,7 +235,7 @@ bool run_mx_gemm(const ProblemSizeSplitK& problem_size, const ExecutionConfig& c
break;
case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<BDataType>{-2.0, 2.0});
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-2.0, 2.0});
a_m_k_scale.GenerateTensorValue(GeneratorTensor_3<XDataType>{powf(2.0f, -125.0f), 1.0f});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-2.0, 2.0});

97
example/67_gemm_microscaling/gemm_mx_fp8_bf8.cpp Normal file
View File

@@ -0,0 +1,97 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "gemm_mx_common.hpp"
using ADataType = ck::f8_t;
using BDataType = ck::bf8_t;
using XDataType = ck::e8m0_bexp_t;
using CDataType = ck::bhalf_t;
using AccDataType = float;
using CShuffleDataType = CDataType;
using ALayout = Row;
using BLayout = Row;
using CLayout = Row;
using AElementOp = PassThrough; // elementwise transformation for A matrix
using BElementOp = PassThrough; // elementwise transformation for B matrix
using CElementOp = PassThrough; // elementwise transformation for C matrix
constexpr ck::index_t ScaleBlockSize = 32; // scaling block size
constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
constexpr auto BlkGemmPSched = ck::BlockGemmPipelineScheduler::Intrawave;
constexpr auto BlkGemmPVer = ck::BlockGemmPipelineVersion::v1;
using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMX_Xdl_CShuffleV3<
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
ADataType, // ADataType
XDataType, // AScaleDataType
BDataType, // BDataType
XDataType, // BScaleDataType
CDataType, // CDataType
AccDataType, // GemmAccDataType
CShuffleDataType, // CShuffleDataType
AElementOp, // AElementwiseOperation
BElementOp, // BElementwiseOperation
CElementOp, // CElementwiseOperation
GemmSpec, // GemmSpec
ScaleBlockSize, // ScaleBlockSize: Scaling block size
256, // BlockSize: Thread block size
256, // MPerBlock
256, // NPerBlock
128, // KPerBlock
16, // AK1
8, // BK1
16, // MPerXDL
16, // NPerXDL
8, // MXdlPerWave
8, // NXdlPerWave
S<8, 32, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_AK1
false, // ABlockLdsExtraM
S<16, 16, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<0, 2, 1>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1>, // BBlockTransferSrcAccessOrder
1, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_BK1
false, // BBlockLdsExtraN
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
BlkGemmPSched, // BlkGemmPipeSched
BlkGemmPVer, // BlkGemmPipelineVer
ADataType, // ComputeTypeA
BDataType // ComputeTypeB
>;
int main(int argc, char* argv[])
{
return run_mx_gemm_example<DeviceOpInstance,
ADataType,
BDataType,
XDataType,
CDataType,
ALayout,
BLayout,
CLayout,
AElementOp,
BElementOp,
CElementOp,
AccDataType,
CShuffleDataType,
ScaleBlockSize>(argc, argv)
? 0
: -1;
}