ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 04:19:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK Tile] Spatially local GEMM tile partitioner. (#1843)

Browse Source 
* Add spatially local tile partitioner

* Use 1D Grid size & create partitioner object.

* Docs & use 1D partitioner in example.

* Clang format.

* Change kernel grid size

Now: X is the # of output C-tiles,
     Y is the batch count
     Z is the splitK

* Formatting & more doc.

* Clang format.

* Fix batched gemm test. Use 1d partitioner.

* Move condition.

* FIx ctor.

* clang-format.

[ROCm/composable_kernel commit: ce448002ee]
This commit is contained in:
Adam Osewski
2025-01-31 00:10:16 +01:00
committed by GitHub
parent e4a3ea71dd
commit 7b47def8f8
10 changed files with 285 additions and 88 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
example/ck_tile/03_gemm/gemm_basic.cpp
View File

@@ -40,7 +40,7 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<CodegenGemmShape>;
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;

2
example/ck_tile/03_gemm/gemm_basic.hpp
View File

@@ -79,7 +79,7 @@ auto create_args(int argc, char* argv[])
.insert("n", "4096", "n dimension")
.insert("k", "2048", "k dimension")
.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 - Column by default")
.insert("c_layout", "R", "C tensor data layout - Row by default")
.insert("stride_a", "0", "Tensor A stride")
.insert("stride_b", "0", "Tensor B stride")

7
example/ck_tile/03_gemm/universal_gemm.cpp
View File

@@ -50,7 +50,9 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
constexpr bool TransposeC = false;
constexpr int kBlockPerCu = 1;
constexpr int kBlockPerCu = 1;
constexpr ck_tile::index_t TileParitionerGroupNum = 8;
constexpr ck_tile::index_t TileParitionerM01 = 4;
// ===============================================
@@ -58,7 +60,8 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<GemmShape>;
using TilePartitioner = ck_tile::
GemmSpatiallyLocalTilePartitioner<GemmShape, TileParitionerGroupNum, TileParitionerM01>;
using Traits = ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
using GemmUniversalTraits = ck_tile::

2
example/ck_tile/16_batched_gemm/batched_gemm.cpp
View File

@@ -43,7 +43,7 @@ float batched_gemm(const ck_tile::BatchedGemmHostArgs& args, const ck_tile::stre
ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
using TilePartitioner = ck_tile::GemmTile2DPartitioner<CodegenGemmShape>;
using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
using CodegenGemmTraits =
ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;