ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-05 06:01:23 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK-Tile] Enable vectorized reads on all layouts & improve perf. (#1835)

Browse Source 
* Refactor universal gemm policy.

* Adapt example to refactor changes.

* Introduce static encoding pattern

* Adding shuffled encoding patterns.

* Fix err in reverse tuple.

* Add transpose_tile2d

* Small refactoring + doc

* Enable reading on contiguous dimension in all layouts.

* Transpose A/B register tile if needed for comp v3 pipeline.

* Take contiguous dim size when calculating dram vector load size.

* A/B smem pack size taken from WarpGemm attributes

* Update B LDS layout and setup tile distribution pattern at class level.

* Fix static assert.

* Fix errors in examples.

* Formatting & fix IsTranspose

* Fix VectorSize & refactor.

* Add error loging messages.

* Fix VecLoadSize and TranspseC for mem pipeline.

* Update unit-tests & disable mem pipeline.

* Clang format

* Update include/ck_tile/core/tensor/tile_window.hpp

Co-authored-by: jakpiase <jakub.piasecki@amd.com>

* Fix compilation and reviewers comments.

* Refactor unit-test. Fallback to non-universal gemm.

Need to use GemmPipelineAGmemBGmemCRegV1 for now,
since GemmKernel is now supporting also non-K major vector reads.

---------

Co-authored-by: jakpiase <jakub.piasecki@amd.com>
This commit is contained in:
Adam Osewski
2025-01-27 16:37:19 +01:00
committed by GitHub
parent 64d5c4d6cb
commit 39dc25a9b8
31 changed files with 1393 additions and 688 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
example/ck_tile/17_grouped_gemm/grouped_gemm.cpp
View File

@@ -88,12 +88,9 @@ using CodegenPipelineProblem =
CodegenGemmShape,
CodegenGemmTraits<ALayout, BLayout, CLayout>>;
using CodegenGemmPolicy = ck_tile::UniversalGemmPipelineAgBgCrPolicy;
template <typename ALayout, typename BLayout, typename CLayout>
using CodegenGemmPipeline =
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>,
CodegenGemmPolicy>;
ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem<ALayout, BLayout, CLayout>>;
template <typename ALayout, typename BLayout, typename CLayout>
using Kernel = ck_tile::GroupedGemmKernel<TilePartitioner,

2
example/ck_tile/17_grouped_gemm/grouped_gemm.hpp
View File

@@ -41,7 +41,7 @@ auto create_args(int argc, char* argv[])
.insert("stride_Bs", "", "Tensor B strides - it is empty by default.")
.insert("stride_Cs", "", "Tensor C strides - it is empty by default.")
.insert("a_layout", "R", "A tensor data layout - Row by default.")
.insert("b_layout", "R", "B tensor data layout - Row by default.")
.insert("b_layout", "C", "B tensor data layout - Row by default.")
.insert("c_layout", "R", "C tensor data layout - Row by default.")
.insert("validate", "1", "0. No validation, 1. Validation on CPU.")
.insert("warmup", "10", "number of iterations before benchmark the kernel.")

17
example/ck_tile/17_grouped_gemm/run_grouped_gemm_example.inc
View File

@@ -135,12 +135,9 @@ int run_grouped_gemm_example_with_layouts(int argc,
const ck_tile::index_t N = Ns[i];
const ck_tile::index_t K = Ks[i];
stride_As[i] =
ck_tile::get_default_stride(M, N, stride_As[i], is_row_major(a_layout));
stride_Bs[i] =
ck_tile::get_default_stride(K, N, stride_Bs[i], is_row_major(b_layout));
stride_Cs[i] =
ck_tile::get_default_stride(M, N, stride_Cs[i], is_row_major(CLayout{}));
stride_As[i] = ck_tile::get_default_stride(M, N, stride_As[i], is_row_major(a_layout));
stride_Bs[i] = ck_tile::get_default_stride(K, N, stride_Bs[i], is_row_major(b_layout));
stride_Cs[i] = ck_tile::get_default_stride(M, N, stride_Cs[i], is_row_major(CLayout{}));
a_m_k_tensors.push_back(ck_tile::HostTensor<ADataType>(
ck_tile::host_tensor_descriptor(M, K, stride_As[i], is_row_major(a_layout))));
@@ -229,10 +226,10 @@ int run_grouped_gemm_example(int argc, char* argv[])
{
return run_grouped_gemm_example_with_layouts(argc, argv, Row{}, Col{}, Row{});
}
else if(a_layout == "R" && b_layout == "R")
{
return run_grouped_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
}
// else if(a_layout == "R" && b_layout == "R")
// {
// return run_grouped_gemm_example_with_layouts(argc, argv, Row{}, Row{}, Row{});
// }
else
{
throw std::runtime_error("Unsupported data layout configuration for A,B and C tensors!");