Use int32_t for kThreadGroupSize in warp_per_row (#1332)

Summary: Pull Request resolved: #1332 Use int32_t for kThreadGroupSize instead of size_t. Using size_t causes a slowdown in some cases. Reviewed By: jianyuh Differential Revision: D39568569 fbshipit-source-id: 7a1e56f96a0e9929fb9600156f22fe3c58a3c59e
pytorch · Sep 16, 2022 · bb74666 · bb74666
1 parent 54594d0
commit bb74666
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/fbgemm_gpu/codegen/embedding_backward_split_template.cu b/fbgemm_gpu/codegen/embedding_backward_split_template.cu
@@ -518,7 +518,7 @@ template <
     typename grad_t,
     typename cache_t,
     size_t kMaxVecsPerThread,
-    size_t kThreadGroupSize = kWarpSize>
+    int32_t kThreadGroupSize = kWarpSize>
 __global__
 __launch_bounds__(kBackwardMaxThreads)
 void