ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-02 13:17:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix int4 moe 16x16x32 type perf issue

Browse Source
This commit is contained in:
mtgu0705
2025-03-19 07:15:57 +00:00
parent 1228d112ef
commit 045e7d65b3
3 changed files with 14 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
example/65_gemm_multiply_multiply/CMakeLists.txt
View File

@@ -12,6 +12,10 @@ foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_moe_gemm1_xdl_pk_i4 moe_gemm1_xdl_pk_i4.cpp)
add_example_executable(example_moe_gemm2_xdl_pk_i4 moe_gemm2_xdl_pk_i4.cpp)
set(EXAMPLE_COMPILE_OPTIONS)
list(APPEND EXAMPLE_COMPILE_OPTIONS -mllvm --amdgpu-enable-max-ilp-scheduling-strategy=1)
target_compile_options(example_moe_gemm1_xdl_pk_i4 PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
target_compile_options(example_moe_gemm2_xdl_pk_i4 PRIVATE ${EXAMPLE_COMPILE_OPTIONS})
set(target 1)
endif()
endforeach()

8
example/65_gemm_multiply_multiply/moe_gemm2_xdl_pk_i4.cpp
View File

@@ -125,10 +125,10 @@ using CDEElementOp = MulABScaleExpertWeight;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr ck::index_t MPerBlock = 128;
static constexpr ck::index_t BLOCKSIZE = 256;
static constexpr ck::index_t MXDLPerWave = 4;
static constexpr ck::index_t NXDLPerWave = 1;
static constexpr ck::index_t MXDLPerWave = 8;
static constexpr ck::index_t NXDLPerWave = 2;
static constexpr ck::index_t NPerBlock = 128;
static constexpr ck::index_t MNPerXDL = 32;
static constexpr ck::index_t MNPerXDL = 16;
static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType);
static constexpr ck::index_t CShuffleNLane = 32;
static constexpr ck::index_t CShuffleMLane = BLOCKSIZE / CShuffleNLane;
@@ -148,7 +148,7 @@ using DeviceOpInstance = ck::tensor_operation::device::Devic
MXDLPerWave, NXDLPerWave,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, AK1, AK1, 0,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, BK1, BK1, 0,
1, 1, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
2, 2, S<1, CShuffleMLane, 1, CShuffleNLane>, S<EVec, D0Vec, D1Vec, D2Vec>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, false, false, A0DataType>;
// clang-format on