ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK-Tile] Remove usage of tile partitioner's full gemm shape (#3204)

Browse Source 
gemm shape should be used from the pipeline instead (where it gets from a problem description struct)
This commit is contained in:
Max Podkorytov
2025-11-18 09:56:40 -08:00
committed by GitHub
parent ac70206b2c
commit a3a4eb12bd
9 changed files with 31 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
include/ck_tile/ops/flatmm/kernel/flatmm_kernel.hpp
View File

@@ -363,8 +363,8 @@ struct FlatmmKernel
template <class KernelArgs>
__device__ SplitKBatchOffset(const KernelArgs& kargs, const std::size_t k_id = blockIdx.z)
{
constexpr auto N1 = TilePartitioner::BlockGemmShape::WarpTile::at(number<1>{});
constexpr auto K1 = TilePartitioner::BlockGemmShape::WarpTile::at(number<2>{});
constexpr auto N1 = BlockGemmShape::WarpTile::at(number<1>{});
constexpr auto K1 = BlockGemmShape::WarpTile::at(number<2>{});
const index_t K_t = kargs.k_batch * K1;
const index_t KRead = (kargs.K + K_t - 1) / K_t * K1;

2
include/ck_tile/ops/flatmm/kernel/moe_flatmm_kernel.hpp
View File

@@ -369,7 +369,7 @@ struct MoeFlatmmKernel
template <class KernelArgs>
__device__ SplitKBatchOffset(const KernelArgs& kargs, const std::size_t k_id = blockIdx.z)
{
constexpr auto K1 = TilePartitioner::BlockGemmShape::WarpTile::at(number<2>{});
constexpr auto K1 = BlockGemmShape::WarpTile::at(number<2>{});
const index_t K_t = kargs.k_batch * K1;
const index_t KRead = (kargs.K + K_t - 1) / K_t * K1;