ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add benchmark example

Browse Source
This commit is contained in:
Damien Lejeune
2026-02-06 14:55:13 +00:00
parent 804a9d488c
commit ec1e8ec58e
6 changed files with 231 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
include/ck_tile/ops/mhc/kernel/mhc_kernel_tile_v3.hpp
View File

@@ -22,9 +22,6 @@ namespace ck_tile {
template <typename Problem_,
typename Policy_ = MHCDefaultPolicy,
index_t kMTile_ = 64, // Batch tile size
index_t kNTile_ = 32, // Output dimension tile (can cover all 24 outputs)
index_t kKTile_ = 8, // K-tile for C dimension (must match BlockGemmShape::kK)
typename Activation_ = element_wise::Sigmoid>
struct MHCKernelV3
{
@@ -37,9 +34,10 @@ struct MHCKernelV3
using YDataType = ck_tile::remove_cvref_t<typename Problem::YDataType>;
using PhiDataType = ck_tile::remove_cvref_t<typename Problem::PhiDataType>;
static constexpr index_t kMTile = kMTile_; // Batch tile
static constexpr index_t kNTile = kNTile_; // Output tile
static constexpr index_t kKTile = kKTile_; // K tile for C dimension
// Automatically derive tile sizes from BlockGemmShape (single source of truth!)
static constexpr index_t kMTile = Problem::BlockGemmShape::kM; // Batch tile
static constexpr index_t kNTile = Problem::BlockGemmShape::kN; // Output tile
static constexpr index_t kKTile = Problem::BlockGemmShape::kK; // K tile for C dimension
static constexpr index_t kBlockSize = Problem::kBlockSize;

8
include/ck_tile/ops/mhc/pipeline/mhc_problem.hpp
View File

@@ -26,12 +26,12 @@ struct MHCProblem
using CDataType = ComputeDataType; // Output/accumulator matrix C
// BlockGemmShape with kM, kN, kK members for BlockGemm
// Use supported warp gemm configuration for float32: 32x32x8
// We'll use 2 warps in M and 1 warp in N to get 64x32 block
// 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 BlockGemmShape =
TileGemmShape<sequence<64, 32, 8>, // BlockTile (M, N, K)
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 (matches available float32 MFMA)
sequence<32, 32, 8>>; // WarpTile (32x32x8 is supported by MFMA)
// Layout types for BlockGemm
using ALayout = ck_tile::tensor_layout::gemm::RowMajor; // x is row-major [B, nC]