Fix a bug in VecAutomaticAddPerBlock (#55929)

81511469 · niuliling123 · GitHub · 9f2d88e9 · 81511469
隐藏空白更改
内联并排

Showing with 1 addition and 1 deletion

paddle/phi/backends/gpu/gpu_primitives.h paddle/phi/backends/gpu/gpu_primitives.h +1 -1

未找到文件。
--- a/paddle/phi/backends/gpu/gpu_primitives.h
+++ b/paddle/phi/backends/gpu/gpu_primitives.h
@@ -711,7 +711,7 @@ __device__ __forceinline__ void VectorizedAtomicAddPerBlock(
  using NVT = typename VecAtomicAddHelper<T>::NVT;
  using NVVec2T = typename VecAtomicAddHelper<T>::NVVec2T;
  bool aligned_half2 =
-      (reinterpret_cast<std::uintptr_t>(out) % sizeof(NVT) == 0);
+      (reinterpret_cast<std::uintptr_t>(out) % sizeof(NVVec2T) == 0);

  if (aligned_half2) {
    for (i = tid * 2; i < loops; i += threads_per_block * 2) {