ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-05 22:22:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

CK Tile - small fix to hotloop scheduler & KPack value. (#1867)

Browse Source 
* Use SmemPack in HotLoop scheduler

* Additional debug print information

* Change KPack value.

Hardcode for now, as without AK1/BK1 there's no good way to determine
its value.

* Fix HotLoopScheduler MFMA instr parameters.
This commit is contained in:
Adam Osewski
2025-02-07 09:00:44 +01:00
committed by Sam Wu
parent ab5d027866
commit 0c15de6a83
5 changed files with 86 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp
View File

@@ -3,6 +3,9 @@
#pragma once
#include <string>
#include <sstream>
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
@@ -83,6 +86,56 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
return Policy::template GetSmemSize<Problem>();
}
CK_TILE_HOST static std::string Print()
{
constexpr index_t MPerXDL = BlockGemm::WarpGemm::kM;
constexpr index_t NPerXDL = BlockGemm::WarpGemm::kN;
constexpr index_t KPerXDL = BlockGemm::WarpGemm::WarpGemmAttribute::Impl::kK;
constexpr index_t WaveSize = 64;
constexpr index_t WaveNumM = BlockGemmShape::BlockWarps::at(I0{});
constexpr index_t WaveNumN = BlockGemmShape::BlockWarps::at(I1{});
// Below should be equal to AK1|BK1
constexpr index_t A_LDS_Read_Width = Policy::template GetSmemPackA<Problem>();
constexpr index_t B_LDS_Read_Width = Policy::template GetSmemPackB<Problem>();
constexpr index_t A_LDS_Write_Width = Policy::template GetSmemPackA<Problem>();
constexpr index_t B_LDS_Write_Width = Policy::template GetSmemPackB<Problem>();
constexpr index_t A_Buffer_Load_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * GetVectorSizeA());
constexpr index_t B_Buffer_Load_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * GetVectorSizeB());
constexpr index_t A_LDS_Write_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * A_LDS_Write_Width);
constexpr index_t B_LDS_Write_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * B_LDS_Write_Width);
constexpr index_t A_LDS_Read_Inst_Num =
WaveNumN * MPerBlock * KPerBlock / (BlockSize * A_LDS_Read_Width);
constexpr index_t B_LDS_Read_Inst_Num =
WaveNumM * MPerBlock * KPerBlock / (BlockSize * B_LDS_Read_Width);
constexpr index_t C_MFMA_Inst_Num = MPerBlock * NPerBlock * KPerBlock /
(BlockSize / WaveSize) / (MPerXDL * NPerXDL * KPerXDL);
auto str = std::stringstream{};
str << "A/B vector size: " << GetVectorSizeA() << ", " << GetVectorSizeB() << "\n"
<< "A/B LDS read/write width: " << A_LDS_Read_Width << ", " << B_LDS_Read_Width << "\n"
<< "A/B buffer load inst: " << A_Buffer_Load_Inst_Num << ", " << B_Buffer_Load_Inst_Num
<< "\n"
<< "A/B LDS write inst: " << A_LDS_Write_Inst_Num << ", " << B_LDS_Write_Inst_Num
<< "\n"
<< "A/B LDS read inst: " << A_LDS_Read_Inst_Num << ", " << B_LDS_Read_Inst_Num << "\n"
<< "C MFMA inst: " << C_MFMA_Inst_Num << "\n"
<< "KPack: " << BlockGemm::Traits::KPack << "\n"
<< "PrefetchStages: " << PrefetchStages << "\n";
return str.str();
}
template <GemmPipelineScheduler Scheduler>
struct PipelineImpl : public PipelineImplBase
{
@@ -95,29 +148,35 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
CK_TILE_DEVICE static constexpr auto HotLoopScheduler()
{
constexpr index_t MPerXDL = BlockGemmShape::WarpTile::at(I0{});
constexpr index_t NPerXDL = BlockGemmShape::WarpTile::at(I1{});
constexpr index_t KPerXDL = BlockGemmShape::WarpTile::at(I2{});
constexpr index_t MPerXDL = BlockGemm::WarpGemm::kM;
constexpr index_t NPerXDL = BlockGemm::WarpGemm::kN;
constexpr index_t KPerXDL = BlockGemm::WarpGemm::WarpGemmAttribute::Impl::kK;
constexpr index_t WaveSize = 64;
constexpr index_t WaveNumM = BlockGemmShape::BlockWarps::at(I0{});
constexpr index_t WaveNumN = BlockGemmShape::BlockWarps::at(I1{});
constexpr index_t A_LDS_Read_Width = KPerXDL;
constexpr index_t B_LDS_Read_Width = KPerXDL;
// Below should be equal to AK1|BK1
constexpr index_t A_LDS_Read_Width = Policy::template GetSmemPackA<Problem>();
constexpr index_t B_LDS_Read_Width = Policy::template GetSmemPackB<Problem>();
constexpr index_t A_LDS_Write_Width = Policy::template GetSmemPackA<Problem>();
constexpr index_t B_LDS_Write_Width = Policy::template GetSmemPackB<Problem>();
constexpr index_t A_Buffer_Load_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * GetVectorSizeA());
constexpr index_t B_Buffer_Load_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * GetVectorSizeB());
constexpr index_t A_LDS_Write_Inst_Num = MPerBlock * KPerBlock / (BlockSize * KPerXDL);
constexpr index_t B_LDS_Write_Inst_Num = NPerBlock * KPerBlock / (BlockSize * KPerXDL);
constexpr index_t A_LDS_Write_Inst_Num =
MPerBlock * KPerBlock / (BlockSize * A_LDS_Write_Width);
constexpr index_t B_LDS_Write_Inst_Num =
NPerBlock * KPerBlock / (BlockSize * B_LDS_Write_Width);
constexpr index_t A_LDS_Read_Inst_Num =
WaveNumN * MPerBlock * KPerBlock / (BlockSize * KPerXDL);
WaveNumN * MPerBlock * KPerBlock / (BlockSize * A_LDS_Read_Width);
constexpr index_t B_LDS_Read_Inst_Num =
WaveNumM * MPerBlock * KPerBlock / (BlockSize * KPerXDL);
WaveNumM * MPerBlock * KPerBlock / (BlockSize * B_LDS_Read_Width);
constexpr index_t C_MFMA_Inst_Num = MPerBlock * NPerBlock * KPerBlock /
(BlockSize / WaveSize) /

1
include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp
View File

@@ -185,7 +185,6 @@ struct UniversalGemmPipelineAgBgCrPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
{
using ADataType = remove_cvref_t<typename Problem::ADataType>;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;