ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 21:21:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Congma/ck tile/aquant mem pipeline (#3346)

Browse Source 
* [CK TILE GEMM QUANT] Fix the bug in HotLoopTail of memory pipeline
This commit is contained in:
Cong Ma
2025-12-05 09:35:27 -07:00
committed by GitHub
parent 7541d9b5b0
commit ed080f5a56
3 changed files with 127 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp
View File

@@ -36,17 +36,13 @@ struct BaseGemmPipelineAgBgCrMem
// TODO: Is this 32K value gfx9 arch specific?
static constexpr index_t MinMemInFlyBytes = 32768;
static constexpr index_t WgpPerCU =
(4 * get_warp_size() / BlockSize) >= 1 ? 4 * get_warp_size() / BlockSize : 1;
static constexpr index_t WgpPerCU = ck_tile::max(4 * get_warp_size() / BlockSize, 1);
static constexpr index_t FullMemBandPrefetchStages =
integer_divide_ceil(MinMemInFlyBytes / WgpPerCU,
(MPerBlock * sizeof(ADataType) / APackedSize +
NPerBlock * sizeof(BDataType) / BPackedSize) *
KPerBlock);
static constexpr index_t PrefetchStages =
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
: 2;
static constexpr index_t PrefetchStages = ck_tile::clamp(FullMemBandPrefetchStages, 2, 8);
static constexpr index_t LocalPrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;