ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-13 17:55:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update to GetNumPrefetchV() for kM0=64 path

Browse Source
This commit is contained in:
Qianfeng Zhang
2025-12-23 08:20:08 +00:00
parent 6c91b0c407
commit 489e2554ea
1 changed files with 12 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_whole_k_prefetch_default_policy.hpp
View File

@@ -32,9 +32,18 @@ struct BlockFmhaPipelineQRKSVSWholeKPrefetchDefaultPolicy
if constexpr(Problem::kUseTrLoad)
{
if constexpr(n0_loops >= 4 && k1_loops >= 6)
return 3;
return 2;
// kM0 is 64, kN0 is 128, prefetch all k_tiles
if constexpr(IsPreloadWholeNextIterationK<Problem>())
{
if constexpr(n0_loops >= 4 && k1_loops >= 6)
return 2;
return 2;
}
else // kM0 is 128, kN0 is 64, prefetch one k_tile
{
// kN0 == 64, try to prefetch more v_tiles
return 2;
};
}
else
{