Replace 0x1f with 0xffffffff in __shfl_sync (#1097)

This fixes compatibility with H100 and resolves #1094
2023-09-18 19:58:19 -04:00
parent 6407bcdf0a
commit 8783c41851
3 changed files with 3 additions and 3 deletions
@@ -653,7 +653,7 @@ struct B2bGemm {

    // 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;

    // Construct iterators to accumulator scale/bias vector