Fix reduce_blocks_into_lanes race condition (#1798)

Fuzzkatt · web-flow · commit a7de60e57f05 · 2024-04-26T15:29:23.000+09:00
* move __sync_threads() outside if branch

* add clarifying comment
diff --git a/csrc/type_shim.h b/csrc/type_shim.h
@@ -362,8 +362,9 @@ __device__ __forceinline__ T reduce_block_into_lanes
     if(tid < lanes)
       x[tid] = final; // EpilogueOp
     // Make sure the smem result is visible to all warps.
-    __syncthreads();
   }
+  __syncthreads();
+  // Avoid potential write before read race when reduce_block_into_lanes is called back to back
 
   return final;
 }

Original file line number	Diff line number	Diff line change
`@@ -362,8 +362,9 @@ __device__ __forceinline__ T reduce_block_into_lanes`
`362`	`362`	`if(tid < lanes)`
`363`	`363`	`x[tid] = final; // EpilogueOp`
`364`	`364`	`// Make sure the smem result is visible to all warps.`
`365`		`- __syncthreads();`
`366`	`365`	`}`
	`366`	`+ __syncthreads();`
	`367`	`+ // Avoid potential write before read race when reduce_block_into_lanes is called back to back`
`367`	`368`
`368`	`369`	`return final;`
`369`	`370`	`}`