ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Support Wave32 in CK_TILE - Part 1 (#2594)

Browse Source 
* Support wave32/wave64 in CK_TILE - Part 1

* remove blocksize in kernel launch

* fix build error

* fix clang format

* fix clang format 2

* fix clang format 3

* fix fmha build error

* fix fmha build 2

* fix fmha build 3

* fix build error 4

* address review comment

* update change log

* replace KernelBlockSize with kBlockSize

* fix CI fail

* fix clang format

* address review comment and rebase code.

* fix universal test fail

---------

Co-authored-by: Lin, Qun <Quentin.Lin+amdeng@amd.com>
Co-authored-by: Thomas Ning <Thomas.Ning@amd.com>
This commit is contained in:
linqunAMD
2025-08-19 01:08:31 +08:00
committed by GitHub
parent 26d3300930
commit 9fcc1ee9fd
113 changed files with 610 additions and 531 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
example/ck_tile/03_gemm/gemm_basic.cpp
View File

@@ -65,7 +65,6 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
ck_tile::tuple<>,
CLayout,
ck_tile::element_wise::PassThrough,
CodegenPipelineProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
M_Warp,
@@ -81,8 +80,8 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
using Kernel = ck_tile::GemmKernel<TilePartitioner, CodegenGemmPipeline, GemmEpilogue>;
auto kargs = Kernel::MakeKernelArgs(args);
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 grids = Kernel::GridSize(args.M, args.N, args.k_batch);
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -100,10 +99,8 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
<< std::endl;
}
float ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
float ave_time = ck_tile::launch_kernel(
s, ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
return ave_time;
};

17
example/ck_tile/03_gemm/gemm_splitk_two_stage_reduce.cpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2025 Advanced Micro Devices, Inc. All rights reserved.
#include <hip/hip_runtime.h>
@@ -208,7 +208,6 @@ float gemm_stage1(const GemmSplitKHostArgs& args, const ck_tile::stream_config&
DsLayout,
ELayout,
CDEElementWise,
UniversalGemmProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
@@ -232,7 +231,7 @@ float gemm_stage1(const GemmSplitKHostArgs& args, const ck_tile::stream_config&
{
grids = Kernel::GridSize(args.M, args.N, args.k_batch);
}
constexpr dim3 blocks = Kernel::BlockSize();
const dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -279,15 +278,13 @@ float gemm_stage1(const GemmSplitKHostArgs& args, const ck_tile::stream_config&
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};
@@ -373,7 +370,7 @@ float reduce_stage2(const GemmSplitKHostArgs& args, const ck_tile::stream_config
float ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
ck_tile::make_kernel<kBlockPerCu>(
Kernel{},
kGridSize,
kBlockSize,

13
example/ck_tile/03_gemm/gemm_weight_preshuffle.cpp
View File

@@ -103,7 +103,6 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
DsLayout,
ELayout,
CDEElementWise,
UniversalGemmProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
@@ -126,7 +125,7 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{
grids = Kernel::GridSize(args.M, args.N, args.k_batch);
}
constexpr dim3 blocks = Kernel::BlockSize();
dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -172,15 +171,13 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};

13
example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -103,7 +103,6 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
DsLayout,
ELayout,
CDEElementWise,
UniversalGemmProblem::kBlockSize,
TilePartitioner::MPerBlock,
TilePartitioner::NPerBlock,
GemmConfig::M_Warp,
@@ -127,7 +126,7 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
{
grids = Kernel::GridSize(args.M, args.N, args.k_batch);
}
constexpr dim3 blocks = Kernel::BlockSize();
dim3 blocks = Kernel::BlockSize();
if(!Kernel::IsSupportedArgument(kargs))
{
@@ -173,15 +172,13 @@ float gemm(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
ave_time = ck_tile::launch_kernel_time_mask(
s,
run_flush_cache,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
else
{
ave_time =
ck_tile::launch_kernel(s,
ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
Kernel{}, grids, blocks, 0, kargs));
ave_time = ck_tile::launch_kernel(
s,
ck_tile::make_kernel<GemmConfig::kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
return ave_time;
};