ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 01:10:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Improve parallelism in the benchmark

Browse Source
This commit is contained in:
Damien Lejeune
2026-02-09 13:16:21 +00:00
parent ec1e8ec58e
commit 8b59a1e192
2 changed files with 11 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
include/ck_tile/ops/mhc/pipeline/mhc_problem.hpp
View File

@@ -26,12 +26,14 @@ struct MHCProblem
using CDataType = ComputeDataType; // Output/accumulator matrix C
// BlockGemmShape with kM, kN, kK members for BlockGemm
// Using 32x32x8 warp tiles (supported by MFMA) with 2x1 warp layout for 64x32 block
// This gives better parallelism than 64x32 while using supported warp sizes
// Using 16x16x16 warp tiles with 1x1 warp layout for 16x16 block
// Minimal tile size to maximize block count: (1024/16) × (24/16) = 64 × 2 = 128 blocks
// This provides 8x better parallelism than original (128 blocks vs 16 blocks)
// Testing if overhead from many small blocks becomes a problem
using BlockGemmShape =
TileGemmShape<sequence<64, 32, 8>, // BlockTile (M, N, K) - keep original for now
sequence<2, 1, 1>, // BlockWarps (2 warps in M, 1 in N, 1 in K)
sequence<32, 32, 8>>; // WarpTile (32x32x8 is supported by MFMA)
TileGemmShape<sequence<16, 16, 16>, // BlockTile (M, N, K) - minimal tiles for max blocks
sequence<1, 1, 1>, // BlockWarps (1 warp per block)
sequence<16, 16, 16>>; // WarpTile (16x16x16 is supported by MFMA)
// Layout types for BlockGemm
using ALayout = ck_tile::tensor_layout::gemm::RowMajor; // x is row-major [B, nC]