ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 00:40:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

update pipeline v1: add atomic IGLP schedule

Browse Source
This commit is contained in:
AMD-dteng
2025-07-29 22:48:00 +08:00
committed by Feng Shijie
parent f9e48148d2
commit a587701117
4 changed files with 429 additions and 811 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
example/ck_tile/18_flatmm/flatmm_basic.cpp
View File

@@ -155,7 +155,7 @@ float flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
using GemmPipelineProblem = using GemmPipelineProblem =
ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenFlatmmShape, Traits>; ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenFlatmmShape, Traits>;
using BaseGemmPipeline = ck_tile::BaseFlatmmPipelineAGmemBGmemCRegV0<GemmPipelineProblem>; using BaseGemmPipeline = ck_tile::BaseFlatmmPipelineAGmemBGmemCRegV1<GemmPipelineProblem>;
const ck_tile::index_t k_grain = args.k_batch * FlatmmConfig::K_Tile; const ck_tile::index_t k_grain = args.k_batch * FlatmmConfig::K_Tile;
const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * FlatmmConfig::K_Tile; const ck_tile::index_t K_split = (args.K + k_grain - 1) / k_grain * FlatmmConfig::K_Tile;
@@ -182,7 +182,7 @@ float flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
tail_number_v>; tail_number_v>;
using CodegenFlatmmPipeline = using CodegenFlatmmPipeline =
ck_tile::FlatmmPipelineAGmemBGmemCRegV0<CodegenPipelineProblem>; ck_tile::FlatmmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue< using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<ADataType, ck_tile::CShuffleEpilogueProblem<ADataType,

4
example/ck_tile/18_flatmm/flatmm_basic.hpp
View File

@@ -35,7 +35,7 @@ struct FlatmmConfig32
static constexpr bool TransposeC = false; static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false; static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 2; static constexpr int kBlockPerCu = 1;
static constexpr int TileParitionerGroupNum = 8; static constexpr int TileParitionerGroupNum = 8;
static constexpr int TileParitionerM01 = 4; static constexpr int TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default; static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;
@@ -73,7 +73,7 @@ struct FlatmmConfig16
static constexpr bool TransposeC = false; static constexpr bool TransposeC = false;
static constexpr bool UseStructuredSparsity = false; static constexpr bool UseStructuredSparsity = false;
static constexpr int kBlockPerCu = 2; static constexpr int kBlockPerCu = 1;
static constexpr int TileParitionerGroupNum = 8; static constexpr int TileParitionerGroupNum = 8;
static constexpr int TileParitionerM01 = 4; static constexpr int TileParitionerM01 = 4;
static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default; static constexpr auto Scheduler = ck_tile::GemmPipelineScheduler::Default;

5
example/ck_tile/18_flatmm/run_flatmm_example.inc
View File

@@ -55,8 +55,9 @@ int run_flatmm_example_with_layouts(int argc,
// TODO: add different init types // TODO: add different init types
if(init_method == 0) if(init_method == 0)
{ {
ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host); // ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host);
memset(a_host.data(), 0, 4); // ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
ck_tile::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_host);
ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host); ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_origin_host);
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_token_scale); ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_token_scale);
ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_channel_scale); ck_tile::FillUniformDistribution<AccDataType>{-1.f, 1.f}(per_channel_scale);

1227
include/ck_tile/ops/flatmm/pipeline/flatmm_pipeline_agmem_bgmem_creg_v1.hpp
View File

File diff suppressed because it is too large Load Diff