ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-15 19:44:39 +00:00
Code Issues Packages Projects Releases Wiki Activity

custom vector size

Browse Source
This commit is contained in:
Bartlomiej Kocot
2025-05-13 14:38:36 +00:00
parent aaca2e2b08
commit 976e32ccfa
5 changed files with 77 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
example/ck_tile/37_grouped_convolution/grouped_convolution_forward.cpp
View File

@@ -38,6 +38,10 @@ float grouped_conv_fwd_calc(const ck_tile::GroupedConvHostArgs& args,
constexpr ck_tile::index_t N_Warp_Tile = 32;
constexpr ck_tile::index_t K_Warp_Tile = 16;
constexpr ck_tile::index_t VectorSizeA = 8;
constexpr ck_tile::index_t VectorSizeB = 8;
constexpr ck_tile::index_t VectorSizeC = 8;
// Implicit GEMM Traits
using CodegenShape =
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
@@ -47,9 +51,16 @@ float grouped_conv_fwd_calc(const ck_tile::GroupedConvHostArgs& args,
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenShape>;
using CodegenTraits = ck_tile::GroupedConvImplicitGemmTraits;
using CodegenPipelineProblem = ck_tile::
GemmPipelineProblem<InDataType, WeiDataType, AccDataType, CodegenShape, CodegenTraits>;
using CodegenPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
using CodegenPipelineProblem = ck_tile::GemmPipelineProblem<InDataType,
WeiDataType,
AccDataType,
CodegenShape,
CodegenTraits,
InDataType,
true,
VectorSizeA,
VectorSizeB>;
using CodegenPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
const auto Run = [&](const auto memory_operation_) {
constexpr auto memory_operation = memory_operation_.value;
@@ -69,7 +80,9 @@ float grouped_conv_fwd_calc(const ck_tile::GroupedConvHostArgs& args,
N_Warp_Tile,
K_Warp_Tile,
CodegenPipelineProblem::TransposeC,
memory_operation>>;
memory_operation,
true,
VectorSizeC>>;
constexpr auto ConvSpec = ck_tile::ConvolutionForwardSpecialization::Default;
@@ -99,7 +112,10 @@ float grouped_conv_fwd_calc(const ck_tile::GroupedConvHostArgs& args,
<< "pipeline: " << CodegenPipeline::GetName() << '\n'
<< "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
<< ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
<< std::endl;
<< '\n'
<< "Vector size A: " << CodegenPipeline::GetVectorSizeA()
<< ", Vector size B: " << CodegenPipeline::GetVectorSizeB()
<< ", Vector size C: " << ConvEpilogue::GetVectorSizeC() << std::endl;
}
float ave_time = ck_tile::launch_kernel(