diff --git a/paddle/fluid/operators/batch_norm_op.cu b/paddle/fluid/operators/batch_norm_op.cu
index 95d7f23b2c0ac6e46cf85bef4340eb4180dc3dba..4b4fcd487ceda989491b1917bd54e9556a899b37 100644
--- a/paddle/fluid/operators/batch_norm_op.cu
+++ b/paddle/fluid/operators/batch_norm_op.cu
@@ -418,7 +418,7 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
           }
         } else {
           if (d_x) {
-            BNBackwardData<T, block, framework::DataLayout::kNCHW><<<
+            BNBackwardData<T, block, framework::DataLayout::kNHWC><<<
                 grid2, block, 0, dev_ctx.stream()>>>(
                 d_y->data<T>(), scale->data<BatchNormParamType<T>>(),
                 saved_mean_data, x->data<T>(), saved_var_data, C, N, H * W * D,