diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h
index 06af88a993d19bb03ae468b468cbfef3b782d5f0..965c0df5325a39558e70842d04648a9223ca06f9 100644
--- a/paddle/operators/rowwise_add_op.h
+++ b/paddle/operators/rowwise_add_op.h
@@ -42,18 +42,18 @@ template <typename Place, typename T>
 class RowwiseAddGradKernel : public OpKernel {
  public:
   void Compute(const ExecutionContext& context) const override {
-    auto XGrad = context.Output<Tensor>(0);
-    auto bGrad = context.Output<Tensor>(1);
+    auto* XGrad = context.Output<Tensor>(0);
+    auto* bGrad = context.Output<Tensor>(1);
     XGrad->mutable_data<T>(context.GetPlace());
     bGrad->mutable_data<T>(context.GetPlace());
 
     // I, O, OG  => [X, b], [Out], [OutGrad]
     auto OutGrad = EigenMatrix<T>::From(*context.Input<Tensor>(3));
-    EigenMatrix<T>::From(*XGrad).device(*(context.GetEigenDevice<Place>())) =
+    EigenMatrix<T>::From(*XGrad).device(context.GetEigenDevice<Place>()) =
         OutGrad;
 
     // https://eigen.tuxfamily.org/dox/unsupported/TensorBase_8h_source.html
-    EigenVector<T>::Flatten(*bGrad).device(*(context.GetEigenDevice<Place>())) =
+    EigenVector<T>::Flatten(*bGrad).device(context.GetEigenDevice<Place>()) =
         OutGrad.cumsum(1);  // colwise add
   }
 };
diff --git a/python/paddle/v2/framework/tests/test_rowwise_add_op.py b/python/paddle/v2/framework/tests/test_rowwise_add_op.py
index f8521eb517057fbeb104b28af7da4fffe54f37de..e957dd6b3fe50632d889fc7935654756e97ec6e0 100644
--- a/python/paddle/v2/framework/tests/test_rowwise_add_op.py
+++ b/python/paddle/v2/framework/tests/test_rowwise_add_op.py
@@ -15,5 +15,7 @@ class TestRowwiseAddOp(unittest.TestCase):
         self.outputs = {'Out': np.add(self.inputs['X'], self.inputs['b'])}
 
 
+#TODO(dzh): rowwise_grad check
+
 if __name__ == '__main__':
     unittest.main()