ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 18:42:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK TILE] Implement cschuflle algorithm (#1842)

Browse Source 
* [CK TILE] Implement cschuflle algorithm

* Rebase

* Vector store size fixes

* fixes

* Fixes

* fixes

* fmha fix

* fixes

* fixes of fixes

[ROCm/composable_kernel commit: 25e2e0f04a]
This commit is contained in:
Bartłomiej Kocot
2025-01-30 11:57:39 +01:00
committed by GitHub
parent e4d8548dc5
commit 4f2c699f90
18 changed files with 408 additions and 371 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
example/ck_tile/03_gemm/gemm_basic.cpp
View File

@@ -20,10 +20,6 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool kTilePermute = false;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
constexpr int kBlockPerCu = 1;
// This part comes from the Codegen
@@ -39,11 +35,6 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
// Whether doing the CShuffle (transpose before the global memory), depending on the output
// layout.
constexpr bool CShuffleEpilogue =
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
@@ -51,26 +42,24 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
using TilePartitioner = ck_tile::GemmTile2DPartitioner<CodegenGemmShape>;
using GemmEpilogue = std::conditional_t<
CShuffleEpilogue,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
kPadM,
kPadN,
kTilePermute,
kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC>>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;

18
example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
@@ -60,9 +60,6 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<GemmShape>;
using GemmEpilogue = ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using GemmUniversalTraits = ck_tile::
TileGemmUniversalTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout, TransposeC>;
@@ -95,6 +92,19 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
using GemmPipeline =
GEMM_PIPELINE<UniversalGemmProblem, ck_tile::UniversalGemmPipelineAgBgCrPolicy>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
GemmPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
UniversalGemmProblem::TransposeC>>;
using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);

42
example/ck_tile/16_batched_gemm/batched_gemm.cpp
View File

@@ -19,12 +19,9 @@ template <typename ALayout, typename BLayout, typename CLayout>
float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stream_config& s)
{
// The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr bool kTilePermute = false;
// The rank and permutation will also be generate out by the CodeGen part.
constexpr ck_tile::index_t kOutputRank = 2;
constexpr bool kPadM = false;
constexpr bool kPadN = false;
constexpr bool kPadK = false;
constexpr int kBlockPerCu = 1;
@@ -41,11 +38,6 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 8;
// Whether doing the CShuffle (transpose before the global memory), depending on the output
// layout.
constexpr bool CShuffleEpilogue =
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>;
using CodegenGemmShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
@@ -53,26 +45,24 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
using TilePartitioner = ck_tile::GemmTile2DPartitioner<CodegenGemmShape>;
using GemmEpilogue = std::conditional_t<
CShuffleEpilogue,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
kPadM,
kPadN,
kTilePermute,
kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<
ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadM, kPadN>>>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using GemmEpilogue = ck_tile::CShuffleEpilogue<
ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
CLayout,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
N_Warp,
M_Warp_Tile,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC>>;
// ToDo: Will add the codegen part to test different pipeline policies in GEMM.
// Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
using Kernel = ck_tile::BatchedGemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;

46
example/ck_tile/17_grouped_gemm/grouped_gemm.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
@@ -20,12 +20,9 @@ namespace {
struct GroupedGemmKernelParam
{
static const bool kPadM = false;
static const bool kPadN = false;
static const bool kPadK = false;
static const bool kTilePermute = false;
static const ck_tile::index_t kOutputRank = 2;
static const bool kPadM = false;
static const bool kPadN = false;
static const bool kPadK = false;
static const int kBlockPerCu = 1;
static const ck_tile::index_t M_Tile = 128;
@@ -54,24 +51,6 @@ using CodegenGemmShape =
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
template <typename CLayout>
using GemmEpilogue = std::conditional_t<
std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>,
ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
CDataType,
GroupedGemmKernelParam::kPadM,
GroupedGemmKernelParam::kPadN,
GroupedGemmKernelParam::kTilePermute,
GroupedGemmKernelParam::kOutputRank,
1,
0,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock>>,
ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<AccDataType,
CDataType,
GroupedGemmKernelParam::kPadM,
GroupedGemmKernelParam::kPadN>>>;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmTraits = ck_tile::TileGemmTraits<GroupedGemmKernelParam::kPadM,
GroupedGemmKernelParam::kPadN,
@@ -92,10 +71,25 @@ template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>>;
template <typename ALayout, typename BLayout, typename CLayout>
using GemmEpilogue = ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<
AccDataType,
CDataType,
CLayout,
CodegenPipelineProblem<ALayout, BLayout, CLayout>::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GroupedGemmKernelParam::M_Warp,
GroupedGemmKernelParam::N_Warp,
GroupedGemmKernelParam::M_Warp_Tile,
GroupedGemmKernelParam::N_Warp_Tile,
GroupedGemmKernelParam::K_Warp_Tile,
CodegenPipelineProblem<ALayout, BLayout, CLayout>::TransposeC>>;
template <typename ALayout, typename BLayout, typename CLayout>
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner,
CodegenGemmPipeline<ALayout, BLayout, CLayout>,
GemmEpilogue<CLayout>>;
GemmEpilogue<ALayout, BLayout, CLayout>>;
}; // namespace
std::size_t get_workspace_size(const std::vector<grouped_gemm_kargs>& gemm_descs)