diff --git a/paddle/fluid/platform/cuda_device_function.h b/paddle/fluid/platform/cuda_device_function.h
index e81c385727be5c2ba3f02bfbd86168cb4650dfda..ecec4178f2d9937920e52eb74bf9068b84e741a0 100644
--- a/paddle/fluid/platform/cuda_device_function.h
+++ b/paddle/fluid/platform/cuda_device_function.h
@@ -63,6 +63,7 @@ __device__ T reduceSum(T val, int tid, int len) {
     val += platform::CudaShuffleDownSync(mask, val, offset);
 
   if (tid < warpSize) shm[tid] = 0;
+  __syncthreads();
 
   if (tid % warpSize == 0) {
     shm[tid / warpSize] = val;