diff --git a/csrc/bgmv/bgmv_impl.cuh b/csrc/bgmv/bgmv_impl.cuh
index d2ad687..3c4b973 100644
--- a/csrc/bgmv/bgmv_impl.cuh
+++ b/csrc/bgmv/bgmv_impl.cuh
@@ -89,7 +89,9 @@ __global__ void bgmv_shrink_kernel(T* __restrict__ Y, const T* __restrict__ X,
     for (size_t offset = tx / 2; offset > 0; offset /= 2) {
       sum += __shfl_down_sync(0xffffffff, sum, offset);
     }
-    y_warpwise[threadIdx.y] = sum;
+    if (threadIdx.x == 0) {
+      y_warpwise[threadIdx.y] = sum;
+    }
     block.sync();
 #pragma unroll
     for (size_t i = 0; i < ty; ++i) {
@@ -117,10 +119,12 @@ __global__ void bgmv_shrink_kernel(T* __restrict__ Y, const T* __restrict__ X,
   for (size_t offset = tx / 2; offset > 0; offset /= 2) {
     sum += __shfl_down_sync(0xffffffff, sum, offset);
   }
-  y_warpwise[threadIdx.y] =
-      ((tile_idx - 1) * tile_size + threadIdx.y * tx * vec_size < feat_in)
-          ? sum
-          : 0.f;
+  if (threadIdx.x == 0) {
+    y_warpwise[threadIdx.y] =
+        ((tile_idx - 1) * tile_size + threadIdx.y * tx * vec_size < feat_in)
+            ? sum
+            : 0.f;
+  }
   block.sync();
 #pragma unroll
   for (size_t i = 0; i < ty; ++i) {