diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt
index c9ba478a099f4f9da93caa7741cfb53763568c12..79c00fd039b34bf2090865d1d4e797c24c3479e1 100644
--- a/paddle/fluid/framework/CMakeLists.txt
+++ b/paddle/fluid/framework/CMakeLists.txt
@@ -40,6 +40,7 @@ proto_library(async_executor_proto SRCS data_feed.proto)
 cc_library(ddim SRCS ddim.cc DEPS eigen3 boost enforce)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
+cc_test(unroll_array_ops_test SRCS unroll_array_ops_test.cc)
 cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context)
 cc_test(data_type_test SRCS data_type_test.cc DEPS data_type place tensor)
 if(WITH_GPU)
diff --git a/paddle/fluid/framework/array.h b/paddle/fluid/framework/array.h
index aa0abc22a6bc9fd32a7dab2096223eef140acef9..b53082986882c80a85826f10d5766525f72c0a97 100644
--- a/paddle/fluid/framework/array.h
+++ b/paddle/fluid/framework/array.h
@@ -26,11 +26,12 @@ class Array {
  public:
   static constexpr size_t kSize = N;
 
-  HOSTDEVICE inline Array() = default;
+  HOSTDEVICE inline Array() {}
 
   template <typename... Args>
   HOSTDEVICE inline explicit Array(const T &val, Args... args) {
-    UnrollVarArgsAssign<T, N>::Run(data_, val, args...);
+    static_assert(N == sizeof...(Args) + 1, "Invalid argument");
+    UnrollVarArgsAssign<T>::Run(data_, val, args...);
   }
 
   HOSTDEVICE inline void Fill(const T &val) {
@@ -41,10 +42,29 @@ class Array {
 
   HOSTDEVICE inline T *GetMutable() { return data_; }
 
-  HOSTDEVICE inline T &operator[](size_t index) { return data_[index]; }
+  HOSTDEVICE inline T &operator[](size_t i) { return *advance(data_, i); }
 
-  HOSTDEVICE inline const T &operator[](size_t index) const {
-    return data_[index];
+  // Writing "return data_[i]" would cause compilation warning/error:
+  // "array subscript is above array bound" in Python 35 CI.
+  // It seems that it is a false warning of GCC if we do not check the bounds
+  // of array index. But for better performance, we do not check in operator[]
+  // like what is in STL. If users want to check the bounds, use at() instead
+  HOSTDEVICE inline const T &operator[](size_t i) const {
+    return *advance(data_, i);
+  }
+
+  HOSTDEVICE inline T &at(size_t i) {
+#ifndef __CUDA_ARCH__
+    PADDLE_ENFORCE_LT(i, N, "Array index out of bounds");
+#endif
+    return (*this)[i];
+  }
+
+  HOSTDEVICE inline const T &at(size_t i) const {
+#ifndef __CUDA_ARCH__
+    PADDLE_ENFORCE_LT(i, N, "Array index out of bounds");
+#endif
+    return (*this)[i];
   }
 
   HOSTDEVICE constexpr size_t size() const { return N; }
@@ -58,6 +78,11 @@ class Array {
   }
 
  private:
+  template <typename U>
+  HOSTDEVICE static inline U *advance(U *ptr, size_t i) {
+    return ptr + i;
+  }
+
   T data_[N];
 };
 
@@ -66,7 +91,7 @@ class Array<T, 0> {
  public:
   static constexpr size_t kSize = 0;
 
-  HOSTDEVICE inline Array() = default;
+  HOSTDEVICE inline Array() {}
 
   HOSTDEVICE inline void Fill(const T &val) {}
 
@@ -75,18 +100,28 @@ class Array<T, 0> {
   // Add constexpr to GetMutable() cause warning in MAC
   HOSTDEVICE inline T *GetMutable() { return nullptr; }
 
-  HOSTDEVICE inline T &operator[](size_t index) {
-#ifndef __CUDA_ARCH__
+  HOSTDEVICE inline T &operator[](size_t) {
+#ifdef __CUDA_ARCH__
+    static T obj();
+    return obj;
+#else
     PADDLE_THROW("Array<T, 0> has no element");
 #endif
   }
 
-  HOSTDEVICE inline const T &operator[](size_t index) const {
-#ifndef __CUDA_ARCH__
+  HOSTDEVICE inline const T &operator[](size_t) const {
+#ifdef __CUDA_ARCH__
+    static const T obj();
+    return obj;
+#else
     PADDLE_THROW("Array<T, 0> has no element");
 #endif
   }
 
+  HOSTDEVICE inline T &at(size_t i) { return (*this)[i]; }
+
+  HOSTDEVICE inline const T &at(size_t i) const { return (*this)[i]; }
+
   HOSTDEVICE constexpr size_t size() const { return 0; }
 
   HOSTDEVICE constexpr bool operator==(const Array<T, 0> &other) const {
diff --git a/paddle/fluid/framework/ddim.h b/paddle/fluid/framework/ddim.h
index 1fd3badbb27ba4e9f4cafa8025061ce08483d273..28cb8171f623fd1a02438b5ffacd485d14ea8875 100644
--- a/paddle/fluid/framework/ddim.h
+++ b/paddle/fluid/framework/ddim.h
@@ -60,9 +60,7 @@ class DDim {
 
   DDim() : rank_(1) { dim_[0] = 0; }
 
-  DDim(const DDim& ddim) : dim_(), rank_(ddim.rank_) {
-    dynamic_dim_assign(ddim.dim_.Get(), dim_.GetMutable(), rank_);
-  }
+  DDim(const DDim& ddim) { CopyFrom(ddim); }
 
   DDim(const int* d, int n) : rank_(n) {
     dynamic_dim_assign(d, dim_.GetMutable(), n);
@@ -80,10 +78,12 @@ class DDim {
   /*implicit*/ DDim(std::initializer_list<int64_t> init_list)
       : DDim(init_list.begin(), init_list.size()) {}
 
+  inline DDim& operator=(const DDim& ddim) { return CopyFrom(ddim); }
+
   template <int D>
-  inline DDim& operator=(const Dim<D>& in) {
+  inline DDim& operator=(const Dim<D>& dim) {
     rank_ = D;
-    UnsafeCast<D>() = in;
+    UnsafeCast<D>() = dim;
     return *this;
   }
 
diff --git a/paddle/fluid/framework/unroll_array_ops.h b/paddle/fluid/framework/unroll_array_ops.h
index fb0a89530f61ac3ef92298fd62260bec3d964bc6..731da74eff4d22da6730e589a1af919514f1c4b7 100644
--- a/paddle/fluid/framework/unroll_array_ops.h
+++ b/paddle/fluid/framework/unroll_array_ops.h
@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #pragma once
+#include <cstddef>
 #include <type_traits>
 #include "paddle/fluid/platform/hostdevice.h"
 
@@ -52,21 +53,30 @@ struct UnrollAssign<kStart, kEnd, true> {
 };
 
 template <typename T, size_t kStart, size_t kEnd, bool kStop>
-struct UnrollVarArgsAssign {
+struct UnrollVarArgsAssignImpl {
   template <typename... Args>
   HOSTDEVICE inline static void Run(T *d, T val, Args... args) {
     static_assert(sizeof...(args) + 1 == kEnd - kStart, "Wrong argument");
     d[kStart] = val;
-    UnrollVarArgsAssign<T, kStart + 1, kEnd, kStart + 1 == kEnd>::Run(d,
-                                                                      args...);
+    UnrollVarArgsAssignImpl<T, kStart + 1, kEnd, kStart + 1 == kEnd>::Run(
+        d, args...);
   }
 };
 
 template <typename T, size_t kStart, size_t kEnd>
-struct UnrollVarArgsAssign<T, kStart, kEnd, true> {
+struct UnrollVarArgsAssignImpl<T, kStart, kEnd, true> {
   HOSTDEVICE inline static void Run(T *d) {}
 };
 
+template <typename T>
+struct UnrollVarArgsAssign {
+  template <typename... Args>
+  HOSTDEVICE inline static void Run(T *d, Args... args) {
+    UnrollVarArgsAssignImpl<T, 0, sizeof...(Args), sizeof...(Args) == 0>::Run(
+        d, args...);
+  }
+};
+
 template <size_t kStart, size_t kEnd, bool kStop>
 struct UnrollCompare {
   template <typename T>
@@ -150,8 +160,8 @@ using UnrollFillConstant = detail::UnrollFillConstant<0, N, N == 0>;
 template <size_t N>
 using UnrollAssign = detail::UnrollAssign<0, N, N == 0>;
 
-template <typename T, size_t N>
-using UnrollVarArgsAssign = detail::UnrollVarArgsAssign<T, 0, N, N == 0>;
+template <typename T>
+using UnrollVarArgsAssign = detail::UnrollVarArgsAssign<T>;
 
 template <size_t N>
 using UnrollCompare = detail::UnrollCompare<0, N, N == 0>;
diff --git a/paddle/fluid/framework/unroll_array_ops_test.cc b/paddle/fluid/framework/unroll_array_ops_test.cc
new file mode 100644
index 0000000000000000000000000000000000000000..51433c83c801765d8df10590abdd319ba60e4873
--- /dev/null
+++ b/paddle/fluid/framework/unroll_array_ops_test.cc
@@ -0,0 +1,108 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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 "paddle/fluid/framework/unroll_array_ops.h"
+#include <gtest/gtest.h>
+#include <algorithm>
+#include <array>
+#include <cstdint>
+
+namespace paddle {
+namespace framework {
+
+template <typename T>
+bool CheckEquality(const T* p, size_t n, T val) {
+  return std::all_of(p, p + n, [val](const T& v) { return v == val; });
+}
+
+template <int D1, int D2>
+bool FillConstantTestMain() {
+  static_assert(D1 >= D2, "");
+  std::array<int, D1> arr;
+  arr.fill(0);
+
+  UnrollFillConstant<D2>::Run(arr.data(), 1);
+  return CheckEquality(arr.data(), D2, 1) &&
+         CheckEquality(arr.data() + D2, arr.size() - D2, 0);
+}
+
+TEST(unroll_ops, fill_constant) {
+  EXPECT_TRUE((FillConstantTestMain<9, 0>()));
+  EXPECT_TRUE((FillConstantTestMain<9, 1>()));
+  EXPECT_TRUE((FillConstantTestMain<9, 4>()));
+  EXPECT_TRUE((FillConstantTestMain<9, 9>()));
+}
+
+TEST(unroll_ops, assign) {
+  const int a[] = {1, 2, 3, 4, 5};
+  int b[] = {0, 0, 0, 0, 0};
+  UnrollAssign<3>::Run(a, b);
+  EXPECT_EQ(b[0], 1);
+  EXPECT_EQ(b[1], 2);
+  EXPECT_EQ(b[2], 3);
+  EXPECT_EQ(b[3], 0);
+  EXPECT_EQ(b[4], 0);
+}
+
+TEST(unroll_ops, var_args_assign) {
+  int a[] = {0, 0, 0};
+  UnrollVarArgsAssign<int>::Run(a, 1, 2);
+  EXPECT_EQ(a[0], 1);
+  EXPECT_EQ(a[1], 2);
+  EXPECT_EQ(a[2], 0);
+}
+
+TEST(unroll_ops, compare) {
+  int a[] = {1, 2, 3};
+  int b[] = {1, 2, 4};
+  EXPECT_TRUE(UnrollCompare<2>::Run(a, b));
+  EXPECT_FALSE(UnrollCompare<3>::Run(a, b));
+
+  b[0] = -1;
+  EXPECT_TRUE(UnrollCompare<0>::Run(a, b));
+  EXPECT_FALSE(UnrollCompare<1>::Run(a, b));
+}
+
+TEST(unroll_ops, add) {
+  int a[] = {2, 3, 4};
+  int b[] = {5, 10, 102};
+  int c[] = {0, 0, 0};
+  UnrollAdd<2>::Run(a, b, c);
+  EXPECT_EQ(a[0] + b[0], c[0]);
+  EXPECT_EQ(a[1] + b[1], c[1]);
+  EXPECT_EQ(c[2], 0);
+}
+
+TEST(unroll_ops, mul) {
+  int a[] = {2, 3, 4};
+  int b[] = {5, 10, 102};
+  int c[] = {0, 0, 0};
+  UnrollMul<2>::Run(a, b, c);
+  EXPECT_EQ(a[0] * b[0], c[0]);
+  EXPECT_EQ(a[1] * b[1], c[1]);
+  EXPECT_EQ(c[2], 0);
+}
+
+TEST(unroll_ops, product) {
+  int a[] = {2, 3, 4};
+  int b[] = {5, 10, 102};
+
+  EXPECT_EQ(UnrollProduct<3>::Run(a), a[0] * a[1] * a[2]);
+
+  EXPECT_EQ(UnrollProduct<3>::Run(a, b),
+            a[0] * b[0] + a[1] * b[1] + a[2] * b[2]);
+}
+
+}  // namespace framework
+}  // namespace paddle