From 5ef4c8023e8d2a86c46878ddcedee5620bd11ae1 Mon Sep 17 00:00:00 2001 From: Hansung Kim Date: Mon, 11 Nov 2024 14:06:15 -0800 Subject: [PATCH] sgemm_impl: Disable wmma fast store Doesn't seem to have a big impact on tcore util. --- tests/regression/sgemm_tcore/sgemm_impl.hpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/regression/sgemm_tcore/sgemm_impl.hpp b/tests/regression/sgemm_tcore/sgemm_impl.hpp index 989c5df9..4d121a0f 100644 --- a/tests/regression/sgemm_tcore/sgemm_impl.hpp +++ b/tests/regression/sgemm_tcore/sgemm_impl.hpp @@ -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 // doesn't necessarily mean breaking correctness; it means that the final // 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_FAST 1