ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-16 19:09:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

[GEMM] Optimization for MI200/300. (#1135)

Browse Source 
* Optimize GEMM on MI200/300:
1. Add new blockwise gemm pipeline
2. Add irregular splitk intances

* clang format + typo fix

* Fix a bug

[ROCm/composable_kernel commit: bb63b9732c]
This commit is contained in:
Haocong WANG
2024-01-19 21:02:22 +08:00
committed by GitHub
parent a69d7546b1
commit ec7e5b1331
17 changed files with 3015 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
example/01_gemm/CMakeLists.txt
View File

@@ -19,6 +19,9 @@ add_custom_target(example_gemm_xdl)
add_example_executable(example_gemm_xdl_fp16 gemm_xdl_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16)
add_example_executable(example_gemm_xdl_fp16_v2 gemm_xdl_fp16_v2.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v2)
add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)

51
example/01_gemm/gemm_xdl_fp16_v2.cpp Normal file
View File

@@ -0,0 +1,51 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v2.hpp"
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using AccDataType = float;
using CShuffleDataType = ck::half_t;
using CDataType = ck::half_t;
using F16 = ck::half_t;
using F32 = float;
using ALayout = Row;
using BLayout = Row;
using CLayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV2<
ALayout, BLayout, CLayout,
F16, F16, F16, F32, F16,
PassThrough, PassThrough, PassThrough, GemmDefault,
2, 256,
256, 256,
32, 8, 4,
32, 32,
4, 4,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>,
1, 8, 4, 0,
1, 1, S<1, 32, 1, 8>, 8,
ck::LoopScheduler::Default, ck::PipelineVersion::v1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
#include "run_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

2
example/35_splitK_gemm/run_splitK_gemm_example.inc
View File

@@ -157,7 +157,7 @@ bool run_splitK_gemm(const ProblemSize& problem_size, const ExecutionConfig& con
if(config.time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel, 1});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =

2
example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
View File

@@ -42,7 +42,7 @@ using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::KPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle
// clang-format off