CUTLASS 2.3.0 final.

2020-09-25 10:34:46 -07:00
parent c53f3339bb
commit 37a8f9e598
8 changed files with 15 additions and 15 deletions
@@ -512,7 +512,7 @@ public:

    // Broadcast the warp_id computed by lane 0 to ensure dependent code
    // is compiled as warp-uniform.
-    int warp_idx = __shfl_sync(0x1f, threadIdx.x / 32, 0);
+    int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);

    int lane_idx = threadIdx.x % 32;