fix the shape choose of vectorize for cuda

4 years ago · 068d905e1e
parent a0b60716f1
commit 068d905e1e
1 changed files with 1 additions and 1 deletions
--- a/paddle/fluid/operators/elementwise/elementwise_add_op.h
+++ b/paddle/fluid/operators/elementwise/elementwise_add_op.h
@ -348,7 +348,7 @@ class ElementwiseAddGradKernel : public ElemwiseGradKernel<T> {
      } else {
        size_t thread_nums = 1024;
        size_t block_nums = (width + thread_nums - 1) / thread_nums;
-        int vec_size = VectorizedSize<T>(dx_data);
+        int vec_size = VectorizedSize<T>(dout_data);
        if (vec_size == 4 && width % 4 == 0) {
          block_nums = (width / vec_size + thread_nums - 1) / thread_nums;
          VecMatrixReduceLongWidth<T,