diff --git a/paddle/cuda/include/hl_base.h b/paddle/cuda/include/hl_base.h
index 402302a5bfa877be33bf3bbd6ba05c8f3e48c3ba..77f5d82dbe2cad183491033736bac85961b6d320 100644
--- a/paddle/cuda/include/hl_base.h
+++ b/paddle/cuda/include/hl_base.h
@@ -229,6 +229,11 @@ extern __thread cudaStream_t default_stream;
 
 // __shfl has been deprecated as of CUDA 9.0.
 #if CUDA_VERSION < 9000
+template <typename T>
+__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) {
+  return __shfl_down(val, delta);
+}
+
 template <typename T>
 __forceinline__ __device__ T
 __shfl_sync(unsigned, T val, int src_line, int width) {
diff --git a/paddle/fluid/operators/row_conv_op.cu b/paddle/fluid/operators/row_conv_op.cu
index 79d08cf3d1edbc158ef551c94330e688c87e6c1e..082f761d37ebdecccfa69981856dea2daef818b8 100644
--- a/paddle/fluid/operators/row_conv_op.cu
+++ b/paddle/fluid/operators/row_conv_op.cu
@@ -189,6 +189,10 @@ __global__ void RowConvGradFilterImproved(const T *in, const T *dout,
   }
   __syncthreads();
 
+  // NOTE(zcd): temporary solution
+  unsigned mask = 0u;
+  CREATE_SHFL_MASK(mask, true);
+
   for (int i = 0; i < num_sequence; i++) {
     int start = static_cast<int>(batch_indices[i]);
     int end = static_cast<int>(batch_indices[i + 1]);
@@ -220,7 +224,7 @@ __global__ void RowConvGradFilterImproved(const T *in, const T *dout,
 
         for (int offset = 16; offset > 0;
              offset = offset / 2) {  // blockDim.x is 32.
-          val += platform::__shfl_down_sync(0, val, offset);
+          val += platform::__shfl_down_sync(mask, val, offset);
         }
         __syncthreads();
 
@@ -251,6 +255,10 @@ __global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence,
   T *sh_in = mem;
   T *sh_dout = &mem[block_x * block_y];
 
+  // NOTE(zcd): temporary solution
+  unsigned mask = 0u;
+  CREATE_SHFL_MASK(mask, true);
+
   for (int i = 0; i < num_sequence; i++) {
     int start = static_cast<int>(batch_indices[i]);
     int end = static_cast<int>(batch_indices[i + 1]);
@@ -276,7 +284,7 @@ __global__ void RowConvGradFilter(const T *in, const T *dout, int num_sequence,
 
         for (int offset = 16; offset > 0;
              offset = offset / 2) {  // blockDim.x is 32.
-          val += platform::__shfl_down_sync(0, val, offset);
+          val += platform::__shfl_down_sync(mask, val, offset);
         }
         __syncthreads();
 
diff --git a/paddle/function/RowConvOpGpu.cu b/paddle/function/RowConvOpGpu.cu
index 9d8a6d80bb2c95a9aa8eb3666fd637402c932559..f820ee9a9713ce17547aa03945dc3c291ef50a59 100644
--- a/paddle/function/RowConvOpGpu.cu
+++ b/paddle/function/RowConvOpGpu.cu
@@ -12,8 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "RowConvOp.h"
-#include "hl_base.h"
+#include "paddle/cuda/include/hl_base.h"
+#include "paddle/function/RowConvOp.h"
 
 namespace paddle {
 
@@ -94,7 +94,7 @@ __global__ void KeRowConv2(real* y,
 }
 
 template <>
-void RowConv<DEVICE_TYPE_GPU>(GpuMatrix& out,
+void RowConv<DEVICE_TYPE_GPU>(GpuMatrix& out,  // NOLINT
                               const GpuMatrix& in,
                               const GpuMatrix& filter,
                               const GpuIVector& seq) {
@@ -144,6 +144,10 @@ __global__ void KeRowConvBwWeight(real* dw,
   }
   __syncthreads();
 
+  // NOTE(zcd): temporary solution
+  unsigned mask = 0u;
+  CREATE_SHFL_MASK(mask, true);
+
   for (int i = 0; i < numSeq; ++i) {
     const int start = starts[i];
     const int end = starts[i + 1];
@@ -170,11 +174,10 @@ __global__ void KeRowConvBwWeight(real* dw,
         real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx + context - 1 - t];
         __syncthreads();
         // warp size and blockDim.x is 32.
-        val += __shfl_down(val, 16);
-        val += __shfl_down(val, 8);
-        val += __shfl_down(val, 4);
-        val += __shfl_down(val, 2);
-        val += __shfl_down(val, 1);
+
+        for (int offset = 16; offset > 0; offset /= 2)
+          val += __shfl_down_sync(mask, val, offset);
+
         __syncthreads();
         if (tidx == 0) {
           sh_dw[t][tidy] += val;
@@ -205,6 +208,10 @@ __global__ void KeRowConvBwWeight2(real* dw,
   __shared__ real sh_x[BLOCK_H][BLOCK_W];
   __shared__ real sh_dy[BLOCK_H][BLOCK_W];
 
+  // NOTE(zcd): temporary solution
+  unsigned mask = 0u;
+  CREATE_SHFL_MASK(mask, true);
+
   for (int i = 0; i < numSeq; ++i) {
     const int start = starts[i];
     const int end = starts[i + 1];
@@ -230,11 +237,9 @@ __global__ void KeRowConvBwWeight2(real* dw,
         real val = sh_x[tidy][tidx] * sh_dy[tidy][tidx];
         __syncthreads();
         // warp size and blockDim.x is 32.
-        val += __shfl_down(val, 16);
-        val += __shfl_down(val, 8);
-        val += __shfl_down(val, 4);
-        val += __shfl_down(val, 2);
-        val += __shfl_down(val, 1);
+        for (int offset = 16; offset > 0; offset /= 2)
+          val += __shfl_down_sync(mask, val, offset);
+
         __syncthreads();
 
         if (tidx == 0 && (gidx + tidy) < width) {
@@ -323,8 +328,8 @@ template <>
 void RowConvGrad<DEVICE_TYPE_GPU>(const GpuMatrix& outG,
                                   const GpuMatrix& in,
                                   const GpuMatrix& filter,
-                                  GpuMatrix& inG,
-                                  GpuMatrix& filterG,
+                                  GpuMatrix& inG,      // NOLINT
+                                  GpuMatrix& filterG,  // NOLINT
                                   const GpuIVector& seq) {
   const size_t numSeq = seq.getSize() - 1;
   const size_t contextLength = filter.getHeight();