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

[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
This commit is contained in:
Haocong WANG
2024-01-19 21:02:22 +08:00
committed by GitHub
parent 7e4eb4b800
commit bb63b9732c
17 changed files with 3015 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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