Added missing synchronization to avoid WAR hazards between tiles. (#386)

2021-12-20 11:34:08 -05:00 · 2021-12-20 11:34:08 -05:00 · 288af365db
commit 288af365db
parent 0dc3ba60b3
1 changed files with 3 additions and 0 deletions
--- a/include/cutlass/gemm/kernel/gemm_grouped.h
+++ b/include/cutlass/gemm/kernel/gemm_grouped.h
@ -546,6 +546,9 @@ public:
      // Compute threadblock-scoped matrix multiply-add
      int gemm_k_iterations = (problem_size.k() + Mma::Shape::kK - 1) / Mma::Shape::kK;

+      // Wait for all threads to finish their epilogue phases from the previous tile.
+      __syncthreads();
+      
      // Compute threadblock-scoped matrix multiply-add
      mma(
        gemm_k_iterations,