wu-arch/kernels
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

sgemm_impl: Disable wmma fast store

Browse Source 
Doesn't seem to have a big impact on tcore util.
This commit is contained in:
Hansung Kim
2024-11-11 14:06:15 -08:00
parent 7d7cb5f60a
commit 5ef4c8023e
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/regression/sgemm_tcore/sgemm_impl.hpp
View File

@@ -108,7 +108,7 @@ static_assert(WMITER * WNITER * TCM * TCN * NUM_WARPS * CORES_PER_CLUSTER ==
// scheme and instead do a fast coalesced GMEM writes for move out. This // scheme and instead do a fast coalesced GMEM writes for move out. This
// doesn't necessarily mean breaking correctness; it means that the final // doesn't necessarily mean breaking correctness; it means that the final
// result matrix will be stored in a swizzled form in the global memory. // result matrix will be stored in a swizzled form in the global memory.
#define WMMA_STORE_FAST 1 #define WMMA_STORE_FAST 0
#define GEMMINI_DMA 1 #define GEMMINI_DMA 1
#define GEMMINI_DMA_FAST 1 #define GEMMINI_DMA_FAST 1