diff --git a/cmake/flags.cmake b/cmake/flags.cmake
index c31e62fc08b531a38a851b71a033e14277eff015..34fd348893058980964d723490d9cc220a157b5a 100644
--- a/cmake/flags.cmake
+++ b/cmake/flags.cmake
@@ -124,6 +124,7 @@ set(GPU_COMMON_FLAGS
     -Wno-error=literal-suffix
     -Wno-error=unused-local-typedefs
     -Wno-error=unused-function  # Warnings in Numpy Header.
+    -Wno-error=array-bounds # Warnings in Eigen::array
 )
 
 if (APPLE)
diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h
index 009ae8757f424ea4ec9864fbe1f4ab57069eed74..7bd1f7e759813b7309168a7d2aa253e3fc673b37 100644
--- a/paddle/operators/mul_op.h
+++ b/paddle/operators/mul_op.h
@@ -25,9 +25,10 @@ template <typename Place, typename T>
 class MulKernel : public framework::OpKernel {
 public:
   void Compute(const framework::KernelContext& context) const override {
-    Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair;
-    dim_pair[0].first = 1;
-    dim_pair[0].second = 0;
+    Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair = {
+        Eigen::IndexPair<Eigen::DenseIndex>(1, 0)};
+    // dim_pair[0].first = 1;
+    // dim_pair[0].second = 0;
 
     auto input0 = context.Input(0)->Get<framework::Tensor>();
     auto input1 = context.Input(1)->Get<framework::Tensor>();
diff --git a/paddle/operators/rowwise_add_op.h b/paddle/operators/rowwise_add_op.h
index ffe9378404e0220fc51593ac1db1c778ee9cbcbb..dc47fe7c847bd0c8c179ac0a5f44b8cc541b47cb 100644
--- a/paddle/operators/rowwise_add_op.h
+++ b/paddle/operators/rowwise_add_op.h
@@ -27,6 +27,7 @@ public:
     auto in0 = context.Input(0)->Get<framework::Tensor>();
     auto in1 = context.Input(1)->Get<framework::Tensor>();
     auto* out = context.Output(0)->GetMutable<framework::Tensor>();
+    out->mutable_data<T>(context.GetPlace());
 
     auto input = framework::EigenMatrix<T>::From(in0);
     auto bias = framework::EigenVector<T>::From(in1);
diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h
index 53c626a7929dd150582ba0e65ff5183fe1198b04..500c188dbfcf28ae52c2d5b06466539e115acc4a 100644
--- a/paddle/operators/softmax_op.h
+++ b/paddle/operators/softmax_op.h
@@ -27,6 +27,7 @@ public:
   void Compute(const framework::KernelContext& context) const override {
     auto input = context.Input(0)->Get<framework::Tensor>();
     auto* output = context.Output(0)->GetMutable<framework::Tensor>();
+    output->mutable_data<T>(context.GetPlace());
 
     auto logits = framework::EigenMatrix<T>::From(input);
     auto softmax = framework::EigenMatrix<T>::From(*output);
@@ -41,19 +42,21 @@ public:
     Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
     Eigen::DSizes<int, 2> one_by_class(1, num_classes);
 
-    auto shifted_logits = (logits - logits.maximum(along_class)
-                                        .eval()
-                                        .reshape(batch_by_one)
-                                        .broadcast(one_by_class));
+    auto shifted_logits = (logits -
+                           logits.maximum(along_class)
+                               .eval()
+                               .reshape(batch_by_one)
+                               .broadcast(one_by_class));
 
     softmax.device(*(context.GetEigenDevice<Place>())) = shifted_logits.exp();
 
     softmax.device(*(context.GetEigenDevice<Place>())) =
-        (softmax * softmax.sum(along_class)
-                       .inverse()
-                       .eval()
-                       .reshape(batch_by_one)
-                       .broadcast(one_by_class));
+        (softmax *
+         softmax.sum(along_class)
+             .inverse()
+             .eval()
+             .reshape(batch_by_one)
+             .broadcast(one_by_class));
   }
 };
 }  // namespace operators