[CK_TILE] Fix for comp pipeline v4 (#3307)

* Fix for gemm_pipeline_ag_bg_cr_comp_v4

* Update hotloop condition

Co-authored-by: Bartłomiej Kocot <barkocot@amd.com>

* fix formating

---------

Co-authored-by: Bartłomiej Kocot <barkocot@amd.com>
This commit is contained in:
jakpiase
2025-12-02 11:38:06 +01:00
committed by GitHub
parent 59265d5eb2
commit 5d67d82a0b

View File

@@ -23,7 +23,8 @@ struct BaseGemmPipelineAgBgCrCompV4
CK_TILE_HOST_DEVICE static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
constexpr index_t HotLoopGlobalReads = 2;
return num_loop >= (HotLoopGlobalReads + PrefetchStages);
}
CK_TILE_HOST_DEVICE static constexpr TailNumber GetBlockLoopTailNum(index_t num_loop)