diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md
index e3bee32f8eeac0b2db9e15430fd7c950c6fc777a..58665e9f2b6299ec3959ed6858ab01d459f64dd8 100644
--- a/doc/howto/dev/new_op_cn.md
+++ b/doc/howto/dev/new_op_cn.md
@@ -23,17 +23,20 @@
 - `framework::OperatorWithKernel`：继承自OperatorBase，Op有计算函数，称作有Kernel。
 - `class OpProtoAndCheckerMaker`：描述该Op的输入、输出、属性、注释,主要用于Python API接口生成
 
-依据是否包含kernel，将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下：
+依据是否包含kernel，可以将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下：
 
-  
- 内容            | 定义位置         
---------------  | :----------------------  
+
+ 内容            | 定义位置
+--------------  | :----------------------
 OpProtoMake定义  | `.cc`文件，Backward Op不需要定义OpProtoMake
-Op定义           | `.cc`文件 
-Kernel实现       | CPU、GPU共享Kernel在`.h`文件，否则，CPU可以在`.cc`文件，GPU可在`.cu`文件。 
-注册Op           | Op注册在`.cc`文件；Kernel注册CPU在`.cc`文件，GPU在`.cu`文件
-     
-     
+Op定义           | `.cc`文件
+Kernel实现       | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，GPU 实现在`.cu`文件中。
+注册Op           | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中
+
+
+实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。
+
+
 下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。
 
 
@@ -42,9 +45,11 @@ Kernel实现       | CPU、GPU共享Kernel在`.h`文件，否则，CPU可以在`
 
 ### 1. 定义ProtoMaker类
 
-矩阵乘的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。首先定义`ProtoMaker`来描述该Op的输入、输出及注释：
-    
-```
+矩阵乘法的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。
+
+首先定义`ProtoMaker`来描述该Op的输入、输出，并添加注释：
+
+```cpp
 class MulOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
@@ -59,20 +64,20 @@ The equation is: Out = X * Y
   }
 };
 ```
-   
-[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数包括2个：
+
+[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数含有2个参数：
 
    - `framework::OpProto` ： 前者存储Op的输入输出和参数属性，将用于Python API接口的生成。
    - `framework::OpAttrChecker` ：后者用于检查参数属性的合法性。
-   
-构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加该Op的注释，这些函数会将对应内容添加到`OpProto`中。
 
-在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，该命名尽可能的规范。
+构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加Op的注释。这些函数会将对应内容添加到`OpProto`中。
 
-   
-再举个[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)的例子：
-   
-```
+上面的代码在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，命名请遵守命名规范。
+
+
+再以[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)为例：
+
+```cpp
 template <typename AttrType>
 class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
@@ -87,17 +92,19 @@ The equation is: Out = scale*X
   }
 };
 ```
- 
- 在这个例子里，两处不同：
- 
-  - `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中。
-  - `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
-   
+
+这个例子有两处不同：
+
+- `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中，如果Op的某个输入不参与反向梯度的计算，请显示地调用`.NotInGradient()`进行设置。
+
+- `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
+
 
 ### 2. 定义Operator类
 
+下面的点实现了MulOp的定义：
 
-```c++
+```cpp
 class MulOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -121,33 +128,46 @@ class MulOp : public framework::OperatorWithKernel {
 ```
 
 [`MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L22)继承自`OperatorWithKernel`。`public`成员：
-	 
-```c++
+
+```cpp
 using framework::OperatorWithKernel::OperatorWithKernel;
 ```
 
 这句表示使用基类`OperatorWithKernel`的构造函数，也可写成：
-   
-```c++
+
+```cpp
 MulOp(const std::string &type, const framework::VariableNameMap &inputs,
       const framework::VariableNameMap &outputs,
       const framework::AttributeMap &attrs)
   : OperatorWithKernel(type, inputs, outputs, attrs) {}
-```	
-	
+```
+
 还需要重写`InferShape`接口。`InferShape`为const函数，不能修改Op的成员变量，参数为`const framework::InferShapeContext &ctx`，通过该参数可获取到输入输出以及属性。它的功能是：
 
   - 1). 做检查， 尽早报错：检查输入数据维度、类型等是否合法。
   - 2). 设置输出Tensor的形状。
 
-通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和要讲到的注册函数一起放在`.cc`中
+通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和下面将要介绍的注册函数一起放在`.cc`中
 
 ### 3. 定义OpKernel类
 
-```C++
-template <typename Place, typename T>
-class MulKernel : public framework::OpKernel {
- public:
+`MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数:
+
+- `typename  Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+
+- `typename T` : 表示数据类型，如`float`, `double`等。
+
+需要为`MulKernel`类重写`Compute`接口。
+- `Compute`接受一个输入参数：`const framework::ExecutionContext& context`。
+- 与`InferShapeContext`相比，`ExecutionContext`增加了设备类型，同样可获取到输入输出和属性参数。
+- `Compute`函数里实现`OpKernel`的具体计算逻辑。
+
+下面是 `MulKernel` `Compute`的实现：
+
+  ```cpp
+  template <typename Place, typename T>
+  class MulKernel : public framework::OpKernel {
+  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* X = context.Input<Tensor>("X");
     auto* Y = context.Input<Tensor>("Y");
@@ -157,141 +177,136 @@ class MulKernel : public framework::OpKernel {
         const_cast<platform::DeviceContext*>(context.device_context_);
     math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
   }
-};
-```
+  };
+  ```
 
-`MulKernel`继承自`framework::OpKernel`，带有模板参数:
+需要注意：**不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
 
-  - `typename  Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
-  
- - `typename T` : 表示数据类型，如`float`, `double`等。
-   
-`MulKernel`需要重写`Compute`接口，该接口参数为`const framework::ExecutionContext& context`, `ExecutionContext`相比`InferShapeContext`增加了设备类型，同样可获取到输入输出和属性参数，`Compute`函数里写具体实现时。
-   
-注意，不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel`，`OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 
+`MulOp`的CPU、GPU实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+
+为了使`OpKernel`的计算过程书写更加简单，并且CPU、GPU的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。
+
+
+到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。
+反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。
 
-为了使得`OpKernel`的计算过程书写较为简单，CPU、GPU的代码可以复用，我们通常借助Eigen unsupported Tensor模块来实现。关于在paddle中如何使用Eigen库，请参考对应的使用[文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)
-   
-到此前向Op实现完成，需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似，这里不再重复。但注意，反向Op没有`ProtoMaker`。
-   
 ### 4. 注册Operator
 
-在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
+- 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
+
+    ```cpp
+    namespace ops = paddle::operators;
+    REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
+    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+    REGISTER_OP_CPU_KERNEL(mul_grad,
+                  ops::MulGradKernel<paddle::platform::CPUPlace, float>);
+    ```
+
+   在上面的代码中：
+
+    - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`。
+    - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
+    - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
 
-```c++
-namespace ops = paddle::operators;
-REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
-REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(mul_grad,
-              ops::MulGradKernel<paddle::platform::CPUPlace, float>);
-```
-    
-  - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`，
-  - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
-  - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
-
-在 `.cu`文件中注册GPU Kernel。请注意，如果GPU Kernel的实现是基于Eigen unsupported模块，那么在 `.cu`的最前面请加上宏定义 `#define EIGEN_USE_GPU`
-   
-```c++
-// if use Eigen unsupported module before include head files
-#define EIGEN_USE_GPU
-
-namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(mul_grad,
-                       ops::MulGradKernel<paddle::platform::GPUPlace, float>);
-```
+
+- 在 `.cu`文件中注册GPU Kernel。
+    - 请注意，如果GPU Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下：
+
+    ```cpp
+    // if use Eigen unsupported module before include head files
+    #define EIGEN_USE_GPU
+
+    namespace ops = paddle::operators;
+    REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
+    REGISTER_OP_GPU_KERNEL(mul_grad,
+                           ops::MulGradKernel<paddle::platform::GPUPlace, float>);
+    ```
 
 ### 5. 编译
 
-在[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)文件中添加编译。
-   
-```
-op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
-```
-   
-下面命令可以编译：
-   
-```
-make mul_op
-```
+- 简单**无特殊依赖**的OP无需修改CMakeList.txt文件。[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt) 会自动将 `paddle/operators` 目录下新增的 `*_op.cc` 文件加入编译。
+- 较为复杂、**有额外依赖** 的operator仍需要修改[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)。如，`mul_op` 依赖 `math_function`，需要在`CMakeLists.txt`中添加如下内容：
+
+    ```
+    op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)		 +
+    ```
+
+- 运行下面命令可以进行编译：
+
+    ```
+    make mul_op
+    ```
 
 ## 绑定Python
 
-- 绑定Python 
- 
-    在 [`paddle/pybind/pybind.cc 
-`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc)文件中添加该类：
+- 绑定Python
+
+    在 [`paddle/pybind/pybind.cc
+`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) 使用`USE_OP`告知编译器需要链接的Op，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
 
     ```
     USE_OP(mul);
     ```
     如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`:
-    
+
     ```
     USE_CPU_ONLY_OP(gather);
     ```
-    
+
     如果OP不带Kernel，则使用`USE_NO_KENREL_OP`:
-    
+
     ```
     USE_NO_KENREL_OP(recurrent);
     ```
-    
-    使用`USE_OP`告知编译器需要链接该Op的目标文件，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
-    
-    
+
+
  - 生成库
 
-   在 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件添加类到`DEPS`中，使得该Op可以链接到生成的lib库中。
-   
-   ```
-   if(WITH_PYTHON)
-     cc_library(paddle_pybind SHARED
-     SRCS pybind.cc
-     DEPS pybind python backward
-     mul_op
-     minus_op)
-   endif(WITH_PYTHON)
-   ```
+   无需修改 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件，`paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。
 
 ## 实现单元测试
 
-单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单测](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
+单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
 
-### 前向Operator单测
+### 前向Operator单元测试
 
-前向Op单测继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`，具体单测流程在`OpTestMeta`里完成。需在`setUp`函数定义输入输出和属性参数，以及Python对比的输出值。
+前向Op单元测试继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`。各项更加具体的单元测试在`OpTestMeta`里完成。测试前向Operator，需要：
 
-```
-import unittest
-import numpy as np
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
+1. 在`setUp`函数定义输入、输出，以及相关的属性参数。
+2. 生成随机的输入数据。
+3. 在Python脚本中实现与前向operator相同的计算逻辑，得到输出值，与operator前向计算的输出进行对比。
 
-class TestMulOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
 
-    def setUp(self):
-        self.type = "mul"
-        self.inputs = {
-            'X': np.random.random((32, 84)).astype("float32"),
-            'Y': np.random.random((84, 100)).astype("float32")
-        }
-        self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
-```
-   首先需要`import`必要的包,下面详细解释其他值：
-   
-   - `self.type = "mul" ` : 定义类型，和注册的类型一致。
-   - `self.inputs` : 定义输入，类型为Numpy.array，并初始化。
-   - `self.outputs` : 定义输出，并得到Python结算结果。
+  ```python
+  import unittest
+  import numpy as np
+  from gradient_checker import GradientChecker, create_op
+  from op_test_util import OpTestMeta
 
- 
-### 反向Operator单测
+  class TestMulOp(unittest.TestCase):
+      __metaclass__ = OpTestMeta
 
-反向Op单测继承自`GradientChecker`，而`GradientChecker`集成自`unittest.TestCase`，所以反向单测函数需要`test_`开头。
+      def setUp(self):
+          self.type = "mul"
+          self.inputs = {
+              'X': np.random.random((32, 84)).astype("float32"),
+              'Y': np.random.random((84, 100)).astype("float32")
+          }
+          self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
+  ```
 
-```
+上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
+
+- `self.type = "mul" ` : 定义类型，与operator注册时注册的类型一致。
+- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
+- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
+
+
+### 反向Operator单元测试
+
+反向Op单元测试继承自`GradientChecker`，而`GradientChecker`继承自`unittest.TestCase`，因此，**反向单元测试函数需要以`test_`开头**。
+
+```python
 class TestMulGradOp(GradientChecker):
     def setUp(self):
         self.op = create_op("mul")
@@ -325,33 +340,34 @@ class TestMulGradOp(GradientChecker):
             no_grad_set={"Y"})
 ```
 
-下面解释一些关键的地方:
+下面解释代码中一些关键的地方:
 
-   - 调用`create_op("mul")`创建反向Op对应的前向Op。
-   - 调用`compare_grad`函数对比CPU、GPU计算结果。
-   - `test_normal`中调用`check_grad`检查梯度稳定性，这里采用数值法检测梯度正确性。
-      - 第一个参数`self.op` : 前向Op。
-      - 第二个参数`self.inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。
-      - 第三个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
-      - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
-   - `test_ignore_x`和`test_ignore_y`分支测试只需要计算一个输入梯度的情况。
+- 调用`create_op("mul")`创建反向Op对应的前向Op。
+- 调用`compare_grad`函数对比CPU、GPU计算结果。
+- `test_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。
+  - 第一个参数`self.op` : 前向Op。
+  - 第二个参数`self.inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。
+  - 第三个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
+  - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
+- `test_ignore_x`和`test_ignore_y`分支用来测试只需要计算一个输入梯度的情况。
 
 
-### 编译和执行 
+### 编译和执行单元测试
 
-单测完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)里添加编译：
+单元测试编写完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)中添加以下内容，将单元测试加入工程：
 
 ```
 py_test(test_mul_op SRCS test_mul_op.py)
 ```
 
-编译时需要打开`WITH_TESTING`, 即 `cmake paddle_dir -DWITH_TESTING=ON`，编译成功之后执行单测命令为：
+请注意，**不同于Op的编译测试，运行单元测试测时需要编译整个工程**，并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后，执行下面的命令来运行单元测试：
 
-```
+```bash
 make test ARGS="-R test_mul_op -V"
 ```
+
 或者:
 
-```
+```bash
 ctest -R test_mul_op
 ```
diff --git a/paddle/framework/ddim.cc b/paddle/framework/ddim.cc
index cfd3e8dfdec0e92620aef5cd246b4622b779ce19..85b7de79743bb0390d66b8999f2e8342a51d14a9 100644
--- a/paddle/framework/ddim.cc
+++ b/paddle/framework/ddim.cc
@@ -21,16 +21,16 @@ namespace framework {
 /// @cond HIDDEN
 
 template <int i>
-Dim<i> make_dim(const int* d) {
+Dim<i> make_dim(const int64_t* d) {
   return Dim<i>(*d, make_dim<i - 1>(d + 1));
 }
 
 template <>
-Dim<1> make_dim<1>(const int* d) {
+Dim<1> make_dim<1>(const int64_t* d) {
   return Dim<1>(*d);
 }
 
-void make_ddim(DDim& ddim, const int* dims, int n) {
+void make_ddim(DDim& ddim, const int64_t* dims, int n) {
   switch (n) {
     case 1:
       ddim = make_dim<1>(dims);
@@ -67,13 +67,13 @@ void make_ddim(DDim& ddim, const int* dims, int n) {
 
 /// @endcond
 
-DDim make_ddim(std::initializer_list<int> dims) {
+DDim make_ddim(std::initializer_list<int64_t> dims) {
   DDim result(make_dim(0));
   make_ddim(result, dims.begin(), dims.size());
   return result;
 }
 
-DDim make_ddim(const std::vector<int>& dims) {
+DDim make_ddim(const std::vector<int64_t>& dims) {
   DDim result(make_dim(0));
   make_ddim(result, &dims[0], dims.size());
   return result;
@@ -81,12 +81,12 @@ DDim make_ddim(const std::vector<int>& dims) {
 
 /// @cond HIDDEN
 // XXX For some reason, putting this in an anonymous namespace causes errors
-class DynamicMutableIndexer : public boost::static_visitor<int&> {
+class DynamicMutableIndexer : public boost::static_visitor<int64_t&> {
  public:
   explicit DynamicMutableIndexer(int idx) : idx_(idx) {}
 
   template <int D>
-  int& operator()(Dim<D>& dim) const {
+  int64_t& operator()(Dim<D>& dim) const {
     return dim[idx_];
   }
 
@@ -94,12 +94,12 @@ class DynamicMutableIndexer : public boost::static_visitor<int&> {
   int idx_;
 };
 
-class DynamicConstIndexer : public boost::static_visitor<int> {
+class DynamicConstIndexer : public boost::static_visitor<int64_t> {
  public:
   explicit DynamicConstIndexer(int idx) : idx_(idx) {}
 
   template <int D>
-  int operator()(const Dim<D>& dim) const {
+  int64_t operator()(const Dim<D>& dim) const {
     return dim[idx_];
   }
 
@@ -109,22 +109,22 @@ class DynamicConstIndexer : public boost::static_visitor<int> {
 
 /// @endcond
 
-int& DDim::operator[](int idx) {
+int64_t& DDim::operator[](int idx) {
   return boost::apply_visitor(DynamicMutableIndexer(idx), var);
 }
 
-int DDim::operator[](int idx) const {
+int64_t DDim::operator[](int idx) const {
   return boost::apply_visitor(DynamicConstIndexer(idx), var);
 }
 
-ssize_t DDim::size() const { return arity(*this); }
+int64_t DDim::size() const { return arity(*this); }
 
 bool DDim::operator==(DDim d) const {
   if (var.which() != d.getVar().which()) {
     return false;
   } else {
-    std::vector<int> v1 = vectorize(*this);
-    std::vector<int> v2 = vectorize(d);
+    std::vector<int64_t> v1 = vectorize(*this);
+    std::vector<int64_t> v2 = vectorize(d);
 
     for (unsigned int i = 0; i < v1.size(); i++) {
       if (v1[i] != v2[i]) {
@@ -139,10 +139,10 @@ bool DDim::operator==(DDim d) const {
 bool DDim::operator!=(DDim d) const { return !(*this == d); }
 
 DDim DDim::operator+(DDim d) const {
-  std::vector<int> v1 = vectorize(*this);
-  std::vector<int> v2 = vectorize(d);
+  std::vector<int64_t> v1 = vectorize(*this);
+  std::vector<int64_t> v2 = vectorize(d);
 
-  std::vector<int> v3;
+  std::vector<int64_t> v3;
 
   assert(v1.size() == v2.size());
 
@@ -154,10 +154,10 @@ DDim DDim::operator+(DDim d) const {
 }
 
 DDim DDim::operator*(DDim d) const {
-  std::vector<int> v1 = vectorize(*this);
-  std::vector<int> v2 = vectorize(d);
+  std::vector<int64_t> v1 = vectorize(*this);
+  std::vector<int64_t> v2 = vectorize(d);
 
-  std::vector<int> v3;
+  std::vector<int64_t> v3;
 
   assert(v1.size() == v2.size());
 
@@ -168,15 +168,15 @@ DDim DDim::operator*(DDim d) const {
   return make_ddim(v3);
 }
 
-int get(const DDim& ddim, int idx) { return ddim[idx]; }
+int64_t get(const DDim& ddim, int idx) { return ddim[idx]; }
 
 void set(DDim& ddim, int idx, int value) { ddim[idx] = value; }
 
 /// @cond HIDDEN
 struct VectorizeVisitor : public boost::static_visitor<> {
-  std::vector<int>& vector;
+  std::vector<int64_t>& vector;
 
-  explicit VectorizeVisitor(std::vector<int>& v) : vector(v) {}
+  explicit VectorizeVisitor(std::vector<int64_t>& v) : vector(v) {}
 
   template <typename T>
   void operator()(const T& t) {
@@ -188,31 +188,31 @@ struct VectorizeVisitor : public boost::static_visitor<> {
 };
 /// @endcond
 
-std::vector<int> vectorize(const DDim& ddim) {
-  std::vector<int> result;
+std::vector<int64_t> vectorize(const DDim& ddim) {
+  std::vector<int64_t> result;
   VectorizeVisitor visitor(result);
   boost::apply_visitor(visitor, ddim);
   return result;
 }
 
-struct ProductVisitor : public boost::static_visitor<ssize_t> {
+struct ProductVisitor : public boost::static_visitor<int64_t> {
   template <int D>
-  ssize_t operator()(const Dim<D>& dim) {
+  int64_t operator()(const Dim<D>& dim) {
     return product(dim);
   }
 };
 
-ssize_t product(const DDim& ddim) {
+int64_t product(const DDim& ddim) {
   ProductVisitor visitor;
   return boost::apply_visitor(visitor, ddim);
 }
 
 struct SliceVectorizeVisitor : public boost::static_visitor<> {
-  std::vector<int>& vector;
+  std::vector<int64_t>& vector;
   int begin;
   int end;
 
-  SliceVectorizeVisitor(std::vector<int>& v, int b, int e)
+  SliceVectorizeVisitor(std::vector<int64_t>& v, int b, int e)
       : vector(v), begin(b), end(e) {
     PADDLE_ENFORCE(begin < end,
                    "Begin index must be less than end index in ddim slice.");
@@ -240,7 +240,7 @@ struct SliceVectorizeVisitor : public boost::static_visitor<> {
 };
 
 DDim slice_ddim(const DDim& dim, int begin, int end) {
-  std::vector<int> vec;
+  std::vector<int64_t> vec;
   vec.reserve(end - begin);
   SliceVectorizeVisitor visitor(vec, begin, end);
   boost::apply_visitor(visitor, dim);
@@ -280,7 +280,7 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
   return os;
 }
 
-DDim::DDim(std::initializer_list<int> init_list) {
+DDim::DDim(std::initializer_list<int64_t> init_list) {
   *this = make_ddim(init_list);
 }
 }  // namespace framework
diff --git a/paddle/framework/ddim.h b/paddle/framework/ddim.h
index 95f294b62737be5c3eac39303148ac35da29fe7d..db30c523948b1d437615aa0e9bfecb5e25569296 100644
--- a/paddle/framework/ddim.h
+++ b/paddle/framework/ddim.h
@@ -40,7 +40,7 @@ struct DDim {
   template <int D>
   explicit DDim(const Dim<D>& in) : var(in) {}
 
-  /*implicit*/ DDim(std::initializer_list<int> init_list);
+  /*implicit*/ DDim(std::initializer_list<int64_t> init_list);
 
   template <int D>
   DDim& operator=(const Dim<D>& in) {
@@ -48,8 +48,8 @@ struct DDim {
     return *this;
   }
 
-  int& operator[](int idx);
-  int operator[](int idx) const;
+  int64_t& operator[](int idx);
+  int64_t operator[](int idx) const;
 
   template <typename Visitor>
   typename Visitor::result_type apply_visitor(Visitor& visitor) {
@@ -71,15 +71,15 @@ struct DDim {
 
   DDim operator*(DDim d) const;
 
-  ssize_t size() const;
+  int64_t size() const;
 };
 
 /**
- * \brief Make a DDim from std::vector<int>
+ * \brief Make a DDim from std::vector<int64_t>
  *
  * \param dims An vector of ints. Must be sized between [1, 9]
  */
-DDim make_ddim(const std::vector<int>& dims);
+DDim make_ddim(const std::vector<int64_t>& dims);
 
 /**
  * \brief Make a DDim from an initializer list
@@ -87,14 +87,14 @@ DDim make_ddim(const std::vector<int>& dims);
  * \param dims An initializer list of ints. Must be sized between [1, 9]
  *
  */
-DDim make_ddim(std::initializer_list<int> dims);
+DDim make_ddim(std::initializer_list<int64_t> dims);
 
-int get(const DDim& dim, int idx);
+int64_t get(const DDim& dim, int idx);
 void set(DDim& dim, int idx, int val);
 
-std::vector<int> vectorize(const DDim& ddim);
+std::vector<int64_t> vectorize(const DDim& ddim);
 
-ssize_t product(const DDim& ddim);
+int64_t product(const DDim& ddim);
 
 /**
  * \brief Slice a ddim
diff --git a/paddle/framework/ddim_test.cc b/paddle/framework/ddim_test.cc
index 9d18a2972ce62139430b240b4599854b14290a32..756232b1b56a49d2c91cc2cac950ca508c54fb3f 100644
--- a/paddle/framework/ddim_test.cc
+++ b/paddle/framework/ddim_test.cc
@@ -12,7 +12,7 @@ TEST(DDim, Equality) {
   EXPECT_EQ(ddim[2], 5);
 
   // construct a DDim from a vector
-  std::vector<int> vec({9, 1, 5});
+  std::vector<int64_t> vec({9, 1, 5});
   paddle::framework::DDim vddim = paddle::framework::make_ddim(vec);
   EXPECT_EQ(ddim[0], 9);
   EXPECT_EQ(ddim[1], 1);
@@ -25,7 +25,7 @@ TEST(DDim, Equality) {
   EXPECT_EQ(paddle::framework::get(ddim, 0), 6);
 
   // vectorize a DDim
-  std::vector<int> res_vec = paddle::framework::vectorize(vddim);
+  std::vector<int64_t> res_vec = paddle::framework::vectorize(vddim);
   EXPECT_EQ(res_vec[0], 9);
   EXPECT_EQ(res_vec[1], 1);
   EXPECT_EQ(res_vec[2], 5);
diff --git a/paddle/framework/dim.h b/paddle/framework/dim.h
index 883fdc55eb929ebc51e8ae05938e9d07374406ce..04d4b0e604e6f73ad94e0ca79d6b69f663bd4076 100644
--- a/paddle/framework/dim.h
+++ b/paddle/framework/dim.h
@@ -17,13 +17,13 @@ struct Dim {
   static constexpr int dimensions = i;
 
   template <typename... Args>
-  HOSTDEVICE Dim(int _head, Args... _tail) : head(_head), tail(_tail...) {
+  HOSTDEVICE Dim(int64_t _head, Args... _tail) : head(_head), tail(_tail...) {
     static_assert(sizeof...(_tail) == i - 1,
                   "Dim initialized with the wrong number of parameters");
   }
 
   HOSTDEVICE
-  Dim(int _head, const Dim<i - 1>& _tail) : head(_head), tail(_tail) {}
+  Dim(int64_t _head, const Dim<i - 1>& _tail) : head(_head), tail(_tail) {}
 
   HOSTDEVICE
   Dim() : head(0), tail() {}
@@ -31,12 +31,12 @@ struct Dim {
   /** Construct a Dim from a linear index and size.  Uses Fortran order
    * indexing. */
   HOSTDEVICE
-  Dim(int idx, const Dim<i>& size)
+  Dim(int64_t idx, const Dim<i>& size)
       : head(idx % size.head), tail(idx / size.head, size.tail) {}
 
   /** Construct a Dim with each dimension set to the given index */
   HOSTDEVICE
-  Dim(int idx) : head(idx), tail(idx) {}
+  Dim(int64_t idx) : head(idx), tail(idx) {}
 
   HOSTDEVICE
   bool operator==(const Dim<i>& o) const {
@@ -47,13 +47,13 @@ struct Dim {
   bool operator!=(const Dim<i>& o) const { return !(*this == o); }
 
   HOSTDEVICE
-  int& operator[](int idx);
+  int64_t& operator[](int idx);
   HOSTDEVICE
-  int operator[](int idx) const;
+  int64_t operator[](int idx) const;
 
   HOST std::string to_string() const;
 
-  int head;
+  int64_t head;
   Dim<i - 1> tail;
 };
 
@@ -63,7 +63,7 @@ struct Dim<1> {
   static constexpr int dimensions = 1;
 
   HOSTDEVICE
-  Dim(int _head) : head(_head) {}
+  Dim(int64_t _head) : head(_head) {}
 
   HOSTDEVICE
   Dim() : head(0) {}
@@ -86,11 +86,11 @@ struct Dim<1> {
   bool operator!=(const Dim<1>& o) const { return !(*this == o); }
 
   HOSTDEVICE
-  int& operator[](int idx);
+  int64_t& operator[](int idx);
   HOSTDEVICE
-  int operator[](int idx) const;
+  int64_t operator[](int idx) const;
 
-  int head;
+  int64_t head;
 };
 
 namespace {
@@ -100,12 +100,12 @@ template <int i>
 struct DimGetter {
   // Return a copy if Dim is const
   template <typename D>
-  HOSTDEVICE static int impl(const D& d) {
+  HOSTDEVICE static int64_t impl(const D& d) {
     return DimGetter<i - 1>::impl(d.tail);
   }
   // Return a reference if Dim is mutable
   template <typename D>
-  HOSTDEVICE static int& impl(D& d) {
+  HOSTDEVICE static int64_t& impl(D& d) {
     return DimGetter<i - 1>::impl(d.tail);
   }
 };
@@ -115,18 +115,18 @@ template <>
 struct DimGetter<0> {
   // Return a copy if Dim is const
   template <typename D>
-  HOSTDEVICE static int impl(const D& d) {
+  HOSTDEVICE static int64_t impl(const D& d) {
     return d.head;
   }
   // Return a reference if Dim is mutable
   template <typename D>
-  HOSTDEVICE static int& impl(D& d) {
+  HOSTDEVICE static int64_t& impl(D& d) {
     return d.head;
   }
 };
 
 template <int D>
-HOSTDEVICE int& indexer(Dim<D>& dim, int idx) {
+HOSTDEVICE int64_t& indexer(Dim<D>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx < 0) {
     throw std::invalid_argument("Tried to access a negative dimension");
@@ -141,7 +141,7 @@ HOSTDEVICE int& indexer(Dim<D>& dim, int idx) {
 }
 
 template <>
-HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) {
+HOSTDEVICE int64_t& indexer<1>(Dim<1>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx != 0) {
     throw std::invalid_argument("Invalid index");
@@ -153,7 +153,7 @@ HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) {
 }
 
 template <int D>
-HOSTDEVICE int indexer(const Dim<D>& dim, int idx) {
+HOSTDEVICE int64_t indexer(const Dim<D>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx < 0) {
     throw std::invalid_argument("Tried to access a negative dimension");
@@ -168,7 +168,7 @@ HOSTDEVICE int indexer(const Dim<D>& dim, int idx) {
 }
 
 template <>
-HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) {
+HOSTDEVICE int64_t indexer<1>(const Dim<1>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx != 0) {
     throw std::invalid_argument("Invalid index");
@@ -182,73 +182,76 @@ HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) {
 }  // namespace
 // Static access to constant Dim
 template <int i, int l>
-HOSTDEVICE int get(const Dim<l>& d) {
+HOSTDEVICE int64_t get(const Dim<l>& d) {
   return DimGetter<i>::impl(d);
 }
 
 // Static access to mutable Dim
 template <int i, int l>
-HOSTDEVICE int& get(Dim<l>& d) {
+HOSTDEVICE int64_t& get(Dim<l>& d) {
   return DimGetter<i>::impl(d);
 }
 
 // Dynamic access to constant Dim
 template <int l>
-HOSTDEVICE int Dim<l>::operator[](int i) const {
+HOSTDEVICE int64_t Dim<l>::operator[](int i) const {
   return indexer(*this, i);
 }
 
 // Dynamic access to mutable Dim
 template <int l>
-HOSTDEVICE int& Dim<l>::operator[](int i) {
+HOSTDEVICE int64_t& Dim<l>::operator[](int i) {
   return indexer(*this, i);
 }
 
 // Dynamic access to constant Dim
-inline HOSTDEVICE int Dim<1>::operator[](int i) const {
+inline HOSTDEVICE int64_t Dim<1>::operator[](int i) const {
   return indexer(*this, i);
 }
 
 // Dynamic access to mutable Dim
-inline HOSTDEVICE int& Dim<1>::operator[](int i) { return indexer(*this, i); }
+inline HOSTDEVICE int64_t& Dim<1>::operator[](int i) {
+  return indexer(*this, i);
+}
 
 // Dynamic access to constant Dim
 // without std::enable_if will try to instantiate this on get<0>(d)
 template <int l>
-HOSTDEVICE typename std::enable_if<(l > 0), int>::type get(const Dim<l>& d,
-                                                           int i) {
+HOSTDEVICE typename std::enable_if<(l > 0), int64_t>::type get(const Dim<l>& d,
+                                                               int i) {
   return d[i];
 }
 
 // Dynamic access to mutable Dim
 template <int l>
-HOSTDEVICE typename std::enable_if<(l > 0), int&>::type get(Dim<l>& d, int i) {
+HOSTDEVICE typename std::enable_if<(l > 0), int64_t&>::type get(Dim<l>& d,
+                                                                int i) {
   return d[i];
 }
 
 // Dot product of two dims
 template <int i>
-HOSTDEVICE int linearize(const Dim<i>& a, const Dim<i>& b) {
+HOSTDEVICE int64_t linearize(const Dim<i>& a, const Dim<i>& b) {
   return a.head * b.head + linearize(a.tail, b.tail);
 }
 
 // Base case dot product of two Dims
 // Notice it is inline because it is no longer a template
 template <>
-HOSTDEVICE inline int linearize(const Dim<1>& a, const Dim<1>& b) {
+HOSTDEVICE inline int64_t linearize(const Dim<1>& a, const Dim<1>& b) {
   return a.head * b.head;
 }
 
 // Product of a Dim
 template <int i>
-HOSTDEVICE int product(const Dim<i>& a, int prod = 1) {
+HOSTDEVICE int64_t product(const Dim<i>& a, int prod = 1) {
   return prod * a.head * product(a.tail);
 }
 
 // Base case product of a Dim
 // Notice it is inline because it is no longer a template
 template <>
-HOSTDEVICE inline int product(const Dim<1>& a, int prod) {
+HOSTDEVICE inline int64_t product(const Dim<1>& a, int prod) {
   return prod * a.head;
 }
 
diff --git a/paddle/framework/dim_test.cu b/paddle/framework/dim_test.cu
index 3898d0a447aa502813b3cb5e86c29eebb814ff84..0a6a87669c900de6cb507dd48f0cfc871defe279 100644
--- a/paddle/framework/dim_test.cu
+++ b/paddle/framework/dim_test.cu
@@ -8,7 +8,7 @@ __global__ void test(paddle::framework::Dim<2>* o) {
   o[0] = paddle::framework::make_dim(5, 6);
 }
 
-__global__ void dyn_idx_gpu(int* o) {
+__global__ void dyn_idx_gpu(int64_t* o) {
   auto d = paddle::framework::make_dim(5, 6);
   o[0] = d[1];
 }
@@ -47,9 +47,9 @@ TEST(Dim, Equality) {
   EXPECT_EQ(b[1], 11);
 
   // dynamic access on GPU
-  thrust::device_vector<int> r(1);
+  thrust::device_vector<int64_t> r(1);
   dyn_idx_gpu<<<1, 1>>>(thrust::raw_pointer_cast(r.data()));
-  int res = r[0];
+  int64_t res = r[0];
   EXPECT_EQ(res, 6);
 
   // ex_prefix_mul
diff --git a/paddle/framework/eigen.h b/paddle/framework/eigen.h
index a4667cc51fadfc020d3211b7a82356db386fced1..2d8d9ae10c56e0632414a5bbc754d35bfa9ce6a5 100644
--- a/paddle/framework/eigen.h
+++ b/paddle/framework/eigen.h
@@ -28,7 +28,7 @@ struct EigenDim {
   static Type From(const DDim& dims) {
     PADDLE_ENFORCE(arity(dims) == D, "D must match arity(DDim)");
     Type ret;
-    for (int d = 0; d < arity(dims); d++) {
+    for (int64_t d = 0; d < arity(dims); d++) {
       ret[d] = dims[d];
     }
     return ret;
diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc
index 902c2655e9182d74a48ad13e17a39a3304d5fa57..9e3ca563c6765637f8471d142d32cec447f0b977 100644
--- a/paddle/framework/grad_op_builder_test.cc
+++ b/paddle/framework/grad_op_builder_test.cc
@@ -3,7 +3,7 @@
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 
-USE_OP(add_two);
+USE_OP(add);
 
 namespace paddle {
 namespace framework {
@@ -41,7 +41,7 @@ namespace f = paddle::framework;
 
 TEST(GradOpBuilder, AddTwo) {
   std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
-      "add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
+      "add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
   std::shared_ptr<f::OperatorBase> grad_add_op =
       f::OpRegistry::CreateGradOp(*add_op);
   EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);
diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h
index 7893e233b776425a61d9e3edd43d944a27743188..94f436294f350e2a39785a09959efb3b17bd00a5 100644
--- a/paddle/framework/tensor_impl.h
+++ b/paddle/framework/tensor_impl.h
@@ -58,7 +58,7 @@ inline T* Tensor::mutable_data(platform::Place place) {
                     "Tensor's numel must be larger than zero to call "
                     "Tensor::mutable_data. Call Tensor::set_dim first.");
   /* some versions of boost::variant don't have operator!= */
-  size_t size = product(dims_) * sizeof(T);
+  int64_t size = product(dims_) * sizeof(T);
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
     if (platform::is_cpu_place(place)) {
@@ -131,7 +131,7 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   PADDLE_ENFORCE_LT(begin_idx, end_idx,
                     "Begin index must be less than end index.");
   PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1.");
-  int base = product(dims_) / dims_[0];
+  size_t base = product(dims_) / dims_[0];
   Tensor dst;
   dst.holder_ = holder_;
   DDim dst_dims = dims_;
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index e5efcccb0e219a1c9df888cfec7f8902806676d4..8a0ff1eb535a542e106ceafca6713aefff2526d5 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -14,27 +14,31 @@ function(op_library TARGET)
     cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
             "${multiValueArgs}" ${ARGN})
 
-    foreach(src ${op_library_SRCS})
-        if (${src} MATCHES ".*\\.cu$")
-            list(APPEND cu_srcs ${src})
-        elseif(${src} MATCHES ".*\\.cc$")
-            list(APPEND cc_srcs ${src})
-        else()
-            message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+    list(LENGTH op_library_SRCS op_library_SRCS_len)
+    if (${op_library_SRCS_len} EQUAL 0)
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
+            list(APPEND cc_srcs ${TARGET}.cc)
         endif()
-    endforeach()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
+            list(APPEND cu_srcs ${TARGET}.cu)
+        endif()
+    else()
+        foreach(src ${op_library_SRCS})
+            if (${src} MATCHES ".*\\.cu$")
+                list(APPEND cu_srcs ${src})
+            elseif(${src} MATCHES ".*\\.cc$")
+                list(APPEND cc_srcs ${src})
+            else()
+                message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+            endif()
+        endforeach()
+    endif()
 
     list(LENGTH cc_srcs cc_srcs_len)
     if (${cc_srcs_len} EQUAL 0)
         message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
     endif()
 
-    list(LENGTH cu_srcs cu_srcs_len)
-    list(LENGTH op_library_DEPS dep_len)
-    if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0)
-        message(WARNING "The op library ${TARGET} not support GPU!")
-    endif()
-
     if (WITH_GPU)
         nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
@@ -46,22 +50,22 @@ endfunction()
 
 add_subdirectory(math)
 
-list(REMOVE_ITEM GENERAL_OPS
-     net_op
-     minus_op
-     mul_op
-     recurrent_op
-     scale_op)
-
-op_library(net_op SRCS net_op.cc)
-op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)
-op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
+set(DEPS_OPS
+    identity_op
+    minus_op
+    mul_op
+    recurrent_op
+    scale_op)
+op_library(identity_op DEPS scale_op)
+op_library(minus_op DEPS scale_op)
+op_library(mul_op DEPS math_function)
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc 
   DEPS framework_proto tensor operator net_op)
-op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
+op_library(scale_op DEPS net_op)
 
+list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
-    op_library(${src} SRCS ${src}.cc ${src}.cu)
+    op_library(${src})
 endforeach()
 
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc
index 8ab748ed71e9a5dc0ee0259a78a2b886870bec5b..8dbd47cf0dfbc265032a9966343eed5c7bd8692e 100644
--- a/paddle/operators/add_op.cc
+++ b/paddle/operators/add_op.cc
@@ -57,7 +57,6 @@ class AddOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker, add_two_grad, ops::AddOpGrad);
+REGISTER_OP(add, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad);
 
-REGISTER_OP_CPU_KERNEL(add_two,
-                       ops::AddKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(add, ops::AddKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/add_op.cu b/paddle/operators/add_op.cu
index cec5f558cbc161124620ad4241d6bd8a5324277c..d9c6d20a6c320b59e57ed25da3dd8b093833f8c7 100644
--- a/paddle/operators/add_op.cu
+++ b/paddle/operators/add_op.cu
@@ -12,10 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#define EIGEN_USE_GPU
-#include "paddle/framework/op_registry.h"
 #include "paddle/operators/add_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(add_two,
-                       ops::AddKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(add, ops::AddKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..c033af3b741ae26ad9d37b2164f87aa6e8651c6e
--- /dev/null
+++ b/paddle/operators/cos_sim_op.cc
@@ -0,0 +1,107 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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/operators/cos_sim_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class CosSimOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
+                      ctx.Input<Tensor>("Y")->dims(),
+                      "Dimensions of Input(X) and Input(Y) must be the same.");
+
+    auto dims = ctx.Input<Tensor>("X")->dims();
+    ctx.Output<Tensor>("Out")->Resize({dims[0], 1});
+    ctx.Output<Tensor>("XNorm")->Resize({dims[0], 1});
+    ctx.Output<Tensor>("YNorm")->Resize({dims[0], 1});
+  }
+};
+
+class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of cos_sim op.");
+    AddInput("Y", "The second input of cos_sim op.");
+    AddOutput("Out", "The output of cos_sim op.");
+    AddOutput("XNorm", "Row norm of the first input.").AsIntermediate();
+    AddOutput("YNorm", "Row norm of the second input.").AsIntermediate();
+
+    AddComment(R"DOC(
+Cosine Similarity Operator.
+
+The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y))
+)DOC");
+  }
+};
+
+class CosSimOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("XNorm"),
+                            "Input(XNorm) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("YNorm"),
+                            "Input(YNorm) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) must not be null.");
+
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    auto xnorm_dims = ctx.Input<Tensor>("XNorm")->dims();
+    auto ynorm_dims = ctx.Input<Tensor>("YNorm")->dims();
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    PADDLE_ENFORCE_EQ(x_dims, y_dims,
+                      "Dimensions of Input(X) and Input(Y) must be the same.");
+    PADDLE_ENFORCE_EQ(xnorm_dims[0], x_dims[0],
+                      "1st dimension of XNorm must equal that of Input(X).");
+    PADDLE_ENFORCE_EQ(xnorm_dims[1], 1, "2st dimension of XNorm must be one.");
+    PADDLE_ENFORCE_EQ(ynorm_dims[0], y_dims[0],
+                      "1st dimension of YNorm must equal that of Input(Y).");
+    PADDLE_ENFORCE_EQ(ynorm_dims[1], 1, "2st dimension of YNorm must be one.");
+    PADDLE_ENFORCE_EQ(out_dims[0], x_dims[0],
+                      "1st dimension of Out@GRAD must equal that of Input(X)");
+    PADDLE_ENFORCE_EQ(out_dims[1], 1, "1st dimension of Out@GRAD must be one.");
+
+    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    if (x_grad) x_grad->Resize(x_dims);
+    if (y_grad) y_grad->Resize(y_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(cos_sim, ops::CosSimOp, ops::CosSimOpMaker, cos_sim_grad,
+            ops::CosSimOpGrad);
+REGISTER_OP_CPU_KERNEL(cos_sim,
+                       ops::CosSimKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    cos_sim_grad, ops::CosSimGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/gather_op.cu b/paddle/operators/cos_sim_op.cu
similarity index 72%
rename from paddle/operators/gather_op.cu
rename to paddle/operators/cos_sim_op.cu
index 3f04a7b3f8142106917975cd1e0413fa1633a298..0cb8fd26de47a4a464db98664263544e3e503d63 100644
--- a/paddle/operators/gather_op.cu
+++ b/paddle/operators/cos_sim_op.cu
@@ -13,8 +13,10 @@
    limitations under the License. */
 
 #define EIGEN_USE_GPU
-#include "paddle/operators/gather_op.h"
+#include "paddle/operators/cos_sim_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(gather,
-                       ops::GatherOpKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(cos_sim,
+                       ops::CosSimKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    cos_sim_grad, ops::CosSimGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..9e2bcebe3b5432c157fac895a9bbab5164193dbb
--- /dev/null
+++ b/paddle/operators/cos_sim_op.h
@@ -0,0 +1,107 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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. */
+
+#pragma once
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class CosSimKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input_x = context.Input<Tensor>("X");
+    auto* input_y = context.Input<Tensor>("Y");
+    auto* output_z = context.Output<Tensor>("Out");
+    auto* output_x_norm = context.Output<Tensor>("XNorm");
+    auto* output_y_norm = context.Output<Tensor>("YNorm");
+
+    output_z->mutable_data<T>(context.GetPlace());
+    output_x_norm->mutable_data<T>(context.GetPlace());
+    output_y_norm->mutable_data<T>(context.GetPlace());
+
+    auto dims = input_x->dims();
+    int size = static_cast<int>(framework::product(dims));
+    auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
+    auto x = EigenMatrix<T>::From(*input_x, new_dims);
+    auto y = EigenMatrix<T>::From(*input_y, new_dims);
+    auto z = EigenVector<T>::Flatten(*output_z);
+    auto x_norm = EigenVector<T>::Flatten(*output_x_norm);
+    auto y_norm = EigenVector<T>::Flatten(*output_y_norm);
+
+    auto place = context.GetEigenDevice<Place>();
+    auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
+    x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
+    y_norm.device(place) = y.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
+    z.device(place) = xy / x_norm / y_norm;
+  }
+};
+
+template <typename Place, typename T>
+class CosSimGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input_x = context.Input<Tensor>("X");
+    auto* input_y = context.Input<Tensor>("Y");
+    auto* input_z = context.Input<Tensor>("Out");
+    auto* input_x_norm = context.Input<Tensor>("XNorm");
+    auto* input_y_norm = context.Input<Tensor>("YNorm");
+    auto* output_grad_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* output_grad_y = context.Output<Tensor>(framework::GradVarName("Y"));
+    auto* input_grad_z = context.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto dims = input_x->dims();
+    int size = static_cast<int>(framework::product(dims));
+    auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
+    auto x = EigenMatrix<T>::From(*input_x, new_dims);
+    auto y = EigenMatrix<T>::From(*input_y, new_dims);
+    auto z = EigenMatrix<T>::From(*input_z);
+    auto x_norm = EigenMatrix<T>::From(*input_x_norm);
+    auto y_norm = EigenMatrix<T>::From(*input_y_norm);
+    auto dz = EigenMatrix<T>::From(*input_grad_z);
+
+    Eigen::DSizes<int, 2> bcast(1, new_dims[1]);
+    auto z_bcast = z.broadcast(bcast);
+    auto dz_bcast = dz.broadcast(bcast);
+    auto place = context.GetEigenDevice<Place>();
+    auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast);
+    auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast);
+    auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast);
+    if (output_grad_x) {
+      output_grad_x->mutable_data<T>(context.GetPlace());
+      auto dx = EigenMatrix<T>::From(*output_grad_x, new_dims);
+      dx.device(place) =
+          dz_bcast * (y / norm_prod_bcast - z_bcast * x / x_snorm_bcast);
+    }
+    if (output_grad_y) {
+      output_grad_y->mutable_data<T>(context.GetPlace());
+      auto dy = EigenMatrix<T>::From(*output_grad_y, new_dims);
+      dy.device(place) =
+          dz_bcast * (x / norm_prod_bcast - z_bcast * y / y_snorm_bcast);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc
index 056447901d418355c01d499f7d92d0b59a39edfa..8bb61275badfccec49953015a47b87b0879153bf 100644
--- a/paddle/operators/gaussian_random_op.cc
+++ b/paddle/operators/gaussian_random_op.cc
@@ -31,8 +31,8 @@ class CPUGaussianRandomKernel : public framework::OpKernel {
     }
     engine.seed(seed);
     std::normal_distribution<T> dist(mean, std);
-    ssize_t size = framework::product(tensor->dims());
-    for (ssize_t i = 0; i < size; ++i) {
+    int64_t size = framework::product(tensor->dims());
+    for (int64_t i = 0; i < size; ++i) {
       data[i] = dist(engine);
     }
   }
@@ -46,9 +46,14 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   void InferShape(const framework::InferShapeContext& context) const override {
     auto* tensor = context.Output<framework::Tensor>("Out");
     auto dims = GetAttr<std::vector<int>>("dims");
+    std::vector<int64_t> temp;
+    temp.reserve(dims.size());
+    for (auto dim : dims) {
+      temp.push_back(static_cast<int64_t>(dim));
+    }
     PADDLE_ENFORCE(dims.size() > 0UL,
                    "dims can be one int or array. dims must be set.");
-    tensor->Resize(framework::make_ddim(dims));
+    tensor->Resize(framework::make_ddim(temp));
   }
 };
 
diff --git a/paddle/operators/identity_op.cc b/paddle/operators/identity_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..7d7a53baeddbab7119e603e6e0e2e288b091c8cd
--- /dev/null
+++ b/paddle/operators/identity_op.cc
@@ -0,0 +1,57 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   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/operators/net_op.h"
+#include "paddle/operators/scale_op.h"
+
+namespace paddle {
+namespace operators {
+
+// The identity operator is an alias of the scale operator. This is also an
+// example for creating the alias for an existing operator.
+template <typename AttrType>
+class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  IdentityOpMaker(framework::OpProto *proto,
+                  framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input tensor of identity operator.");
+    AddOutput("Out", "The output tensor of identity operator.");
+    AddComment(R"DOC(
+The identity operator is an alias of the scale operator
+with the attribute scale fixed to 1.0
+)DOC");
+  }
+};
+
+template <typename AttrType>
+class IdentityOp : public NetOp {
+ public:
+  IdentityOp(const std::string &type, const framework::VariableNameMap &inputs,
+             const framework::VariableNameMap &outputs,
+             const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    AppendOp(framework::OpRegistry::CreateOp(
+        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
+        {{"scale", static_cast<AttrType>(1)}}));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp<float>,
+                             ops::IdentityOpMaker<float>);
diff --git a/paddle/operators/rnn/recurrent_op_utils.cc b/paddle/operators/rnn/recurrent_op_utils.cc
index a9b65c30f25554e54e9fd7103f240946a93566e2..69e723b4019fe553426bafbf02b3334ea4acfcf1 100644
--- a/paddle/operators/rnn/recurrent_op_utils.cc
+++ b/paddle/operators/rnn/recurrent_op_utils.cc
@@ -61,7 +61,7 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
       PADDLE_ENFORCE(step_scope_var != nullptr, "%s not in scope",
                      outlinks[i].internal);
       f::DDim step_dims = step_scope_var->template GetMutable<Tensor>()->dims();
-      std::vector<int> dims_vec = vectorize(step_dims);
+      std::vector<int64_t> dims_vec = vectorize(step_dims);
       dims_vec.insert(dims_vec.begin(), seq_len);
       output->Resize(f::make_ddim(dims_vec));
     } else {
diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc
index c2e005444b640b22bb5111b9f4adc017ac752fba..0377f05b2c5f489f386ddb81f9468b4f71ca0f23 100644
--- a/paddle/operators/scale_op.cc
+++ b/paddle/operators/scale_op.cc
@@ -49,7 +49,7 @@ The equation is: Out = scale*X
   }
 };
 
-// IdentityOp's gradient is IdentityOp, too.
+// The gradients of a scale operator is just the scale operator itself.
 // Grad(Out=scale(X)) => Grad(X) = scale(Grad(Out))
 template <typename AttrType>
 class ScaleGradOp : public NetOp {
@@ -66,36 +66,6 @@ class ScaleGradOp : public NetOp {
   }
 };
 
-// IdentityOp is an alias of the ScaleOp. This is also an example for creating
-// an alias of an existing operator.
-template <typename AttrType>
-class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
- public:
-  IdentityOpMaker(framework::OpProto *proto,
-                  framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The input tensor of identity op.");
-    AddOutput("Out", "The output tensor of identity op.");
-    AddComment(R"DOC(
-The identity operator is just an alias of the scale operator with the
-attribute scale is fixed to 1.0.
-)DOC");
-  }
-};
-
-template <typename AttrType>
-class IdentityOp : public NetOp {
- public:
-  IdentityOp(const std::string &type, const framework::VariableNameMap &inputs,
-             const framework::VariableNameMap &outputs,
-             const framework::AttributeMap &attrs)
-      : NetOp(type, inputs, outputs, attrs) {
-    AppendOp(framework::OpRegistry::CreateOp(
-        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
-        {{"scale", static_cast<AttrType>(1)}}));
-  }
-};
-
 }  // namespace operators
 }  // namespace paddle
 
@@ -105,5 +75,3 @@ REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker<float>, scale_grad,
             ops::ScaleGradOp<float>);
 REGISTER_OP_CPU_KERNEL(scale,
                        ops::ScaleKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp<float>,
-                             ops::IdentityOpMaker<float>);
diff --git a/paddle/operators/scatter_op.cu b/paddle/operators/scatter_op.cu
deleted file mode 100644
index 6716b478833ff3adb6112cdb1ee25b7f1744ea1f..0000000000000000000000000000000000000000
--- a/paddle/operators/scatter_op.cu
+++ /dev/null
@@ -1,20 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-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. */
-
-#define EIGEN_USE_GPU
-#include "paddle/operators/scatter_op.h"
-
-namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(scatter,
-                       ops::ScatterOpKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc
index f3aea6df7c344dd78d0faf609b8198abf0fa2648..7edf1c3460c3433338ce789342a9109fb5b9d91c 100644
--- a/paddle/operators/softmax_op.cc
+++ b/paddle/operators/softmax_op.cc
@@ -23,9 +23,9 @@ class SoftmaxOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.Input<Tensor>("logits")->dims().size() == 2UL,
+    PADDLE_ENFORCE(ctx.Input<Tensor>("Logits")->dims().size() == 2UL,
                    "The input of softmax op must be a matrix.");
-    ctx.Output<Tensor>("softmax")->Resize(ctx.Input<Tensor>("logits")->dims());
+    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("Logits")->dims());
   }
 };
 
@@ -34,10 +34,10 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
   SoftmaxOpMaker(framework::OpProto *proto,
                  framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("logits",
+    AddInput("Logits",
              "The input tensor of softmax. "
              "2-D with shape [batch_size, input_feature_dimensions].");
-    AddOutput("softmax", "The normalized values with the same shape as X.");
+    AddOutput("Out", "The normalized values with the same shape as the input.");
     AddComment(R"DOC(
 The input of softmax operator is a 2-D tensor with shape N x K (N is the
 batch_size, K is the dimension of input feature). The output tensor has the
@@ -51,8 +51,8 @@ the other dimensions in the K-dimensional vector input. Then the ratio of the
 exponential of the given dimension and the sum of exponential values of all
 the other dimensions is the output of the softmax operator.
 
-For each row `i` and each column `j` in X, we have:
-    Y[i, j] = exp(X[i, j]) / sum_j(exp(X[i, j]))
+For each row `i` and each column `j` in the input: Logits, we have:
+    Out[i, j] = exp(Logits[i, j]) / sum_j(exp(Logits[i, j]))
 
 )DOC");
   }
@@ -64,17 +64,16 @@ class SoftmaxOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("softmax"),
-                            "Input(softmax) should be not null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("softmax")),
-                            "Input(softmax@GRAD) should be not null.");
-    PADDLE_ENFORCE_EQ(
-        ctx.Input<Tensor>("softmax")->dims(),
-        ctx.Input<Tensor>(framework::GradVarName("softmax"))->dims(),
-        "Input(softmax) and its gradients should have a same shape.");
-
-    ctx.Output<Tensor>(framework::GradVarName("logits"))
-        ->Resize(ctx.Input<Tensor>("logits")->dims());
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Out"),
+                            "Input(Out) should be not null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should be not null.");
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Out")->dims(),
+                      ctx.Input<Tensor>(framework::GradVarName("Out"))->dims(),
+                      "Input(Out) and its gradients should have a same shape.");
+
+    ctx.Output<Tensor>(framework::GradVarName("Logits"))
+        ->Resize(ctx.Input<Tensor>("Logits")->dims());
   }
 };
 
diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h
index 57e5bfad02492e7335d7b3aa89ffbf4719ab7c4c..2ef5915239fd88bee5bb397facee9c99390174ba 100644
--- a/paddle/operators/softmax_op.h
+++ b/paddle/operators/softmax_op.h
@@ -28,12 +28,12 @@ template <typename Place, typename T>
 class SoftmaxKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto X = context.Input<Tensor>("logits");
-    auto Y = context.Output<Tensor>("softmax");
+    auto X = context.Input<Tensor>("Logits");
+    auto Y = context.Output<Tensor>("Out");
     Y->mutable_data<T>(context.GetPlace());
 
     auto logits = EigenMatrix<T>::From(*X);
-    auto softmax = EigenMatrix<T>::From(*Y);
+    auto out = EigenMatrix<T>::From(*Y);
 
     const int kBatchDim = 0;
     const int kClassDim = 1;
@@ -51,11 +51,11 @@ class SoftmaxKernel : public framework::OpKernel {
                                .reshape(batch_by_one)
                                .broadcast(one_by_class));
 
-    softmax.device(context.GetEigenDevice<Place>()) = shifted_logits.exp();
+    out.device(context.GetEigenDevice<Place>()) = shifted_logits.exp();
 
-    softmax.device(context.GetEigenDevice<Place>()) =
-        (softmax *
-         softmax.sum(along_class)
+    out.device(context.GetEigenDevice<Place>()) =
+        (out *
+         out.sum(along_class)
              .inverse()
              .eval()
              .reshape(batch_by_one)
@@ -69,9 +69,9 @@ class SoftmaxGradKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& context) const override {
     std::shared_ptr<Tensor> scale_ = std::make_shared<Tensor>();
 
-    auto Y = context.Input<Tensor>("softmax");
-    auto dY = context.Input<Tensor>(framework::GradVarName("softmax"));
-    auto dX = context.Output<Tensor>(framework::GradVarName("logits"));
+    auto Y = context.Input<Tensor>("Out");
+    auto dY = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto dX = context.Output<Tensor>(framework::GradVarName("Logits"));
     dX->mutable_data<T>(context.GetPlace());
 
     const int batch_size = Y->dims()[0];
diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc
index 2d943c4508490d25a8747330c92c24c384bd0232..40cef8942a3648af5629e5a5db0f021ae3d6f1c1 100644
--- a/paddle/operators/uniform_random_op.cc
+++ b/paddle/operators/uniform_random_op.cc
@@ -35,8 +35,8 @@ class CPUUniformRandomKernel : public framework::OpKernel {
     std::uniform_real_distribution<T> dist(
         static_cast<T>(context.GetAttr<float>("min")),
         static_cast<T>(context.GetAttr<float>("max")));
-    ssize_t size = framework::product(tensor->dims());
-    for (ssize_t i = 0; i < size; ++i) {
+    int64_t size = framework::product(tensor->dims());
+    for (int64_t i = 0; i < size; ++i) {
       data[i] = dist(engine);
     }
   }
@@ -52,7 +52,12 @@ class UniformRandomOp : public framework::OperatorWithKernel {
                    "uniform_random's min must less then max");
     auto* tensor = ctx.Output<framework::Tensor>("Out");
     auto dims = GetAttr<std::vector<int>>("dims");
-    tensor->Resize(framework::make_ddim(dims));
+    std::vector<int64_t> temp;
+    temp.reserve(dims.size());
+    for (auto dim : dims) {
+      temp.push_back(static_cast<int64_t>(dim));
+    }
+    tensor->Resize(framework::make_ddim(temp));
   }
 };
 
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index 6896422617be0a3c73dc7b0d7cc1113075fa2f4b..6e637fa40fbd80fdfa0323a645c57c42d7ca502e 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -30,7 +30,7 @@ limitations under the License. */
 
 namespace py = pybind11;
 
-USE_OP(add_two);
+USE_OP(add);
 USE_OP(onehot_cross_entropy);
 USE_OP(sgd);
 USE_OP(mul);
@@ -46,6 +46,7 @@ USE_OP(lookup_table);
 USE_OP(scale);
 USE_NO_KERNEL_OP(identity);
 USE_OP(minus);
+USE_OP(cos_sim);
 USE_CPU_ONLY_OP(gather);
 USE_CPU_ONLY_OP(scatter);
 
@@ -76,7 +77,7 @@ PYBIND11_PLUGIN(core) {
       .def("get_dims",
            [](const Tensor &self) { return vectorize(self.dims()); })
       .def("set_dims",
-           [](Tensor &self, const std::vector<int> &dim) {
+           [](Tensor &self, const std::vector<int64_t> &dim) {
              self.Resize(make_ddim(dim));
            })
       .def("alloc_float",
diff --git a/paddle/pybind/tensor_py.h b/paddle/pybind/tensor_py.h
index 39ba60b4dc7ebe3f39a0aa4023b34540b340a841..95171acf729a513e5c92d1e0cba15cb12b38561a 100644
--- a/paddle/pybind/tensor_py.h
+++ b/paddle/pybind/tensor_py.h
@@ -85,7 +85,7 @@ void PyCPUTensorSetFromArray(
     framework::Tensor &self,
     py::array_t<T, py::array::c_style | py::array::forcecast> array,
     paddle::platform::CPUPlace &place) {
-  std::vector<int> dims;
+  std::vector<int64_t> dims;
   dims.reserve(array.ndim());
   for (size_t i = 0; i < array.ndim(); ++i) {
     dims.push_back((int)array.shape()[i]);
@@ -102,7 +102,7 @@ void PyCUDATensorSetFromArray(
     framework::Tensor &self,
     py::array_t<T, py::array::c_style | py::array::forcecast> array,
     paddle::platform::GPUPlace &place) {
-  std::vector<int> dims;
+  std::vector<int64_t> dims;
   dims.reserve(array.ndim());
   for (size_t i = 0; i < array.ndim(); ++i) {
     dims.push_back((int)array.shape()[i]);
diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py
index 0349407a851ebb48f69d7daef7a318cf348aad5d..db07bd329d06ba20eb380ee43aa384fb5bb1798b 100644
--- a/python/paddle/v2/framework/op.py
+++ b/python/paddle/v2/framework/op.py
@@ -4,7 +4,7 @@ import paddle.v2.framework.proto.framework_pb2 as framework_pb2
 
 def get_all_op_protos():
     """
-    Get all registered op proto from Paddle C++
+    Get all registered op proto from PaddlePaddle C++ end.
     :return: list of OpProto
     """
     protostrs = core.get_all_op_protos()
@@ -21,8 +21,8 @@ def is_str(s):
 
 class OpDescCreationMethod(object):
     """
-    A Functor object to convert user input(use key word args) to OpDesc based on
-    OpProto.
+    A Functor object converting the user's input(only keyword arguments are
+    supported) to OpDesc based on the OpProto.
 
     :param op_proto: The OpProto object.
     :type op_proto: op_proto_pb2.OpProto
@@ -30,17 +30,18 @@ class OpDescCreationMethod(object):
 
     def __init__(self, op_proto):
         if not isinstance(op_proto, framework_pb2.OpProto):
-            raise TypeError("Argument should be OpProto")
+            raise TypeError(
+                "Type of op_proto should be OpProto in PaddlePaddle.")
         self.__op_proto__ = op_proto
 
     def __call__(self, *args, **kwargs):
         """
-        Convert user input to OpDesc. Only key-word args are supported. 
+        Convert user's input to OpDesc. Only keyword arguments are supported.
         :return: OpDesc based on user input
         :rtype: op_desc_pb2.OpDesc
         """
         if len(args) != 0:
-            raise ValueError("Only keyword arguments is supported by Paddle")
+            raise ValueError("Only keyword arguments are supported.")
         op_desc = framework_pb2.OpDesc()
 
         for input_parameter in self.__op_proto__.inputs:
@@ -49,10 +50,11 @@ class OpDescCreationMethod(object):
                 input_arguments = [input_arguments]
 
             if not input_parameter.duplicable and len(input_arguments) > 1:
-                raise ValueError("Input %s only accepts one input, but give %d"
-                                 % (input_parameter.name, len(input_arguments)))
+                raise ValueError(
+                    "Input %s expects only one input, but %d are given." %
+                    (input_parameter.name, len(input_arguments)))
 
-            ipt = op_desc.inputs.add()
+                ipt = op_desc.inputs.add()
             ipt.parameter = input_parameter.name
             ipt.arguments.extend(input_arguments)
 
@@ -63,10 +65,10 @@ class OpDescCreationMethod(object):
 
             if not output_parameter.duplicable and len(output_arguments) > 1:
                 raise ValueError(
-                    "Output %s only accepts one output, but give %d" %
+                    "Output %s expects only one output, but %d are given." %
                     (output_parameter.name, len(output_arguments)))
 
-            out = op_desc.outputs.add()
+                out = op_desc.outputs.add()
             out.parameter = output_parameter.name
             out.arguments.extend(output_arguments)
 
@@ -100,10 +102,11 @@ class OpDescCreationMethod(object):
                         pair.first = p[0]
                         pair.second = p[1]
                 else:
-                    raise NotImplementedError("Not support attribute type " +
-                                              str(attr.type))
+                    raise NotImplementedError(
+                        "A not supported attribute type: %s." % (
+                            str(attr.type)))
 
-        return op_desc
+                    return op_desc
 
     @staticmethod
     def any_is_true(generator):
@@ -142,10 +145,10 @@ def create_op_creation_method(op_proto):
         outputs=[var.name for var in op_proto.outputs],
         attrs=[attr.name for attr in op_proto.attrs])
 
+    class OperatorFactory(object):
+        def __init__(self):
+            self.op_methods = dict()
 
-class OperatorFactory(object):
-    def __init__(self):
-        self.op_methods = dict()
         for op_proto in get_all_op_protos():
             method = create_op_creation_method(op_proto)
             self.op_methods[method.name] = method
@@ -153,14 +156,16 @@ class OperatorFactory(object):
     def __call__(self, *args, **kwargs):
         if 'type' in kwargs:
             if len(args) != 0:
-                raise ValueError("All Paddle argument should be key-word "
-                                 "argument except type")
-            t = kwargs.pop('type')
+                raise ValueError(
+                    ("All PaddlePaddle arguments should be keyword "
+                     "arguments except the argument \"type\"."))
+                t = kwargs.pop('type')
         else:
             if len(args) != 1:
-                raise ValueError("All Paddle argument should be key-word "
-                                 "argument except type")
-            t = args[0]
+                raise ValueError(
+                    ("All PaddlePaddle arguments should be keyword "
+                     "arguments except the argument \"type\"."))
+                t = args[0]
 
         return self.get_op_info(t).method(**kwargs)
 
@@ -169,7 +174,7 @@ class OperatorFactory(object):
 
     def get_op_info(self, t):
         if t not in self.op_methods:
-            raise ValueError("operator %s is not registered", t)
+            raise ValueError("The operator: %s is not registered." % t)
         return self.op_methods.get(t)
 
     def get_op_input_names(self, type):
diff --git a/python/paddle/v2/framework/tests/CMakeLists.txt b/python/paddle/v2/framework/tests/CMakeLists.txt
index 661ebd89648feec77367c278e5f045b8238e1dc1..e0f77d797390be0461f466726f63a70dd485a290 100644
--- a/python/paddle/v2/framework/tests/CMakeLists.txt
+++ b/python/paddle/v2/framework/tests/CMakeLists.txt
@@ -4,6 +4,7 @@ py_test(test_scope SRCS test_scope.py)
 
 py_test(test_tensor SRCS test_tensor.py)
 py_test(test_mul_op SRCS test_mul_op.py)
+py_test(test_cos_sim_op SRCS test_cos_sim_op.py)
 
 py_test(test_mean_op SRCS test_mean_op.py)
 
diff --git a/python/paddle/v2/framework/tests/gradient_checker.py b/python/paddle/v2/framework/tests/gradient_checker.py
index b8d7e4ea4329e50a086ae51451364a348ff8b360..fdb06b7988935ebbe53f72f4eba89d75ac2502d4 100644
--- a/python/paddle/v2/framework/tests/gradient_checker.py
+++ b/python/paddle/v2/framework/tests/gradient_checker.py
@@ -36,13 +36,13 @@ def get_numeric_gradient(op,
                          in_place=False):
     """
     Get Numeric Gradient for an operator's input.
-    
-    :param op: C++ operator instance, could be an network 
-    :param input_values: The input variables. Should be an dictionary, key is 
+
+    :param op: C++ operator instance, could be an network
+    :param input_values: The input variables. Should be an dictionary, key is
     variable name. Value is numpy array.
-    :param output_name: The final output variable name. 
+    :param output_name: The final output variable name.
     :param input_to_check: The input variable need to get gradient.
-    :param delta: The perturbation value for numeric gradient method. The 
+    :param delta: The perturbation value for numeric gradient method. The
     smaller delta is, the more accurate result will get. But if that delta is
      too small, it could occur numerical stability problem.
     :param local_scope: The local scope used for get_numeric_gradient.
@@ -229,9 +229,9 @@ class GradientChecker(unittest.TestCase):
         """Use relative error for the comparison.
 
         :param numeric_grads: the numerical graidents.
-        :type numeric_grads: a list of numpy.array 
+        :type numeric_grads: a list of numpy.array
         :param analytic_grads: the analytical graidents.
-        :type analytic_grads: a list of numpy.array 
+        :type analytic_grads: a list of numpy.array
         :param name: the names of gradients, used to print for debug.
         :type names: a list of string
         :param msg_prefix: string info, used to print for debug.
diff --git a/python/paddle/v2/framework/tests/mnist.py b/python/paddle/v2/framework/tests/mnist.py
index a68f302f9c344bf6d63e8d9b48836d69338c3d0b..f6f8f49b797fb6e5016a5e309f12f192d5096431 100644
--- a/python/paddle/v2/framework/tests/mnist.py
+++ b/python/paddle/v2/framework/tests/mnist.py
@@ -38,9 +38,9 @@ def feed_data(name, data):
     assert isinstance(data, numpy.ndarray)
     tensor = scope.find_var(name).get_tensor()
     tensor.set_dims(data.shape)
-    if data.dtype == numpy.dtype('int32'):
+    if data.dtype == numpy.dtype("int32"):
         tensor.alloc_int(place)
-    elif data.dtype == numpy.dtype('float32'):
+    elif data.dtype == numpy.dtype("float32"):
         tensor.alloc_float(place)
     else:
         raise ValueError("data type not supported")
@@ -74,22 +74,25 @@ def init_param(net, param_name, dims):
 # fc_layer
 def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None):
     """
-    Add a fc layer to net
+    The fully connected layer.
 
-    :param input: input variable name.
+    :param input: The name of input variable.
     :type input: str
-    :param size: fully connected layer size.
-    :param act: activation name
-    :param param: parameter attribute, used for initialize parameters.
-    :param bias: bias attribute. False will not have a bias.
-    :param name: the name of fc layer. If not set, model will generate a
-    readable name
-    :return: output variable name.
+    :param size: The size of fully connected layer.
+    :param act: The name of activation.
+    :param param: The attribute of learnable parameter which can be used to
+                  modify initialization mean and std of the parameter.
+    :param bias: The attribute of bias. If set False, this layer does not have
+                 a bias.
+    :param name: The name of this layer. If it is not set explictly, a name
+                 will be generated automatically.
+    :return: The name of the output variable.
     """
+
     if name is None:
-        name = 'fc_%d' % uniq_id()
+        name = "fc_%d" % uniq_id()
     if not isinstance(name, str):
-        raise ValueError("name should be string")
+        raise ValueError("The name of a layer should be a string.")
 
     input_dims = scope.find_var(input).get_tensor().get_dims()
 
@@ -123,7 +126,7 @@ def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None):
 
 
 def cross_entropy_layer(net, input, label):
-    cost_name = 'cross_entropy_%d' % uniq_id()
+    cost_name = "cross_entropy_%d" % uniq_id()
     cross_entropy_op = Operator(
         "onehot_cross_entropy", X=input, label=label, Y=cost_name)
     net.append_op(cross_entropy_op)
@@ -177,8 +180,8 @@ def error_rate(predict, label):
     return error_num / float(len(label))
 
 
-images = data_layer(name='pixel', dims=[BATCH_SIZE, 784])
-labels = data_layer(name='label', dims=[BATCH_SIZE])
+images = data_layer(name="pixel", dims=[BATCH_SIZE, 784])
+labels = data_layer(name="label", dims=[BATCH_SIZE])
 fc1 = fc_layer(net=forward_net, input=images, size=100, act="sigmoid")
 fc2 = fc_layer(net=forward_net, input=fc1, size=100, act="sigmoid")
 predict = fc_layer(net=forward_net, input=fc2, size=10, act="softmax")
diff --git a/python/paddle/v2/framework/tests/op_test_util.py b/python/paddle/v2/framework/tests/op_test_util.py
index 3bc05a0feccbbd3d5e7852d85bd3dc8edaccfd07..a4899355b53d62903b97999ebf9c2c7ecfc6c4cd 100644
--- a/python/paddle/v2/framework/tests/op_test_util.py
+++ b/python/paddle/v2/framework/tests/op_test_util.py
@@ -6,13 +6,13 @@ from paddle.v2.framework.op import Operator
 class OpTestMeta(type):
     """
     Operator Test ClassMeta.
-    
-    It injects `test_all` method into user's OperatorTest class, to make Python 
+
+    It injects `test_all` method into user's OperatorTest class, to make Python
     unittest module run that method.
-    
+
     The `test_all` read what value is stored in `self`. It use self's values to
     create and run a operator, and check whether that op is OK or not.
-    
+
     See `test_add_two_op` for example usage.
     """
 
diff --git a/python/paddle/v2/framework/tests/test_add_two_op.py b/python/paddle/v2/framework/tests/test_add_two_op.py
index 0def484eddb88604398ee10390d3f28058714a57..a578e74eca9a3c4327a4881f853028e2347c98ad 100644
--- a/python/paddle/v2/framework/tests/test_add_two_op.py
+++ b/python/paddle/v2/framework/tests/test_add_two_op.py
@@ -11,7 +11,7 @@ class TestAddOp(unittest.TestCase):
     __metaclass__ = OpTestMeta
 
     def setUp(self):
-        self.type = "add_two"
+        self.type = "add"
         self.inputs = {
             'X': numpy.random.random((102, 105)).astype("float32"),
             'Y': numpy.random.random((102, 105)).astype("float32")
diff --git a/python/paddle/v2/framework/tests/test_cos_sim_op.py b/python/paddle/v2/framework/tests/test_cos_sim_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..32013a7999a4be42e5974b9ac751d5d911730994
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_cos_sim_op.py
@@ -0,0 +1,60 @@
+import unittest
+import numpy as np
+from gradient_checker import GradientChecker, create_op
+from op_test_util import OpTestMeta
+
+
+class TestCosSimOp(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "cos_sim"
+        self.inputs = {
+            'X': np.random.random((32, 64)).astype("float32"),
+            'Y': np.random.random((32, 64)).astype("float32")
+        }
+        expect_x_norm = np.linalg.norm(self.inputs['X'], axis=1)
+        expect_y_norm = np.linalg.norm(self.inputs['Y'], axis=1)
+        expect_out = (self.inputs['X'] * self.inputs['Y']).sum(axis=1) / \
+            expect_x_norm / expect_y_norm
+        self.outputs = {
+            'XNorm': np.expand_dims(expect_x_norm, 1),
+            'YNorm': np.expand_dims(expect_y_norm, 1),
+            'Out': np.expand_dims(expect_out, 1)
+        }
+
+
+class TestCosSimGradOp(GradientChecker):
+    def setUp(self):
+        self.op = create_op("cos_sim")
+        self.inputs = {
+            'X': np.random.random((10, 5)).astype("float32"),
+            'Y': np.random.random((10, 5)).astype("float32")
+        }
+
+    def test_cpu_gpu_compare(self):
+        self.compare_grad(self.op, self.inputs)
+
+    def test_normal(self):
+        self.check_grad(
+            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.05)
+
+    def test_ignore_x(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["Y"],
+            "Out",
+            max_relative_error=0.05,
+            no_grad_set={"X"})
+
+    def test_ignore_y(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["X"],
+            "Out",
+            max_relative_error=0.05,
+            no_grad_set={"Y"})
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_gradient_checker.py b/python/paddle/v2/framework/tests/test_gradient_checker.py
index e0b315120862bea284e067070492dcdfbb661081..e6307bc2ecd33301bf974e2d15edb960a35fa9dd 100644
--- a/python/paddle/v2/framework/tests/test_gradient_checker.py
+++ b/python/paddle/v2/framework/tests/test_gradient_checker.py
@@ -7,11 +7,11 @@ from gradient_checker import get_numeric_gradient
 
 class GetNumericGradientTest(unittest.TestCase):
     def test_add_op(self):
-        add_op = Operator('add_two', X="X", Y="Y", Out="Z")
+        add_op = Operator("add", X="X", Y="Y", Out="Z")
         x = numpy.random.random((10, 1)).astype("float32")
         y = numpy.random.random((10, 1)).astype("float32")
 
-        arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X')
+        arr = get_numeric_gradient(add_op, {"X": x, "Y": y}, "Z", "X")
         self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4)
 
     def test_softmax_op(self):
@@ -28,16 +28,16 @@ class GetNumericGradientTest(unittest.TestCase):
                 dX[i, :] = Y[i, :] * (dY[i, :] - d)
             return dX
 
-        softmax_op = Operator("softmax", X="X", Y="Y")
+        softmax_op = Operator("softmax", Logits="Logits", Out="Out")
 
         X = numpy.random.random((2, 2)).astype("float32")
         Y = numpy.apply_along_axis(stable_softmax, 1, X)
         dY = numpy.ones(Y.shape)
         dX = label_softmax_grad(Y, dY)
 
-        arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X')
+        arr = get_numeric_gradient(softmax_op, {"Logits": X}, "Out", "Logits")
         numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2)
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_net.py b/python/paddle/v2/framework/tests/test_net.py
index 9339cf28dabc95b46b958777200fb1db9dcf284f..e4b7cd480cb36249bb64ba3cab9a4b220d812346 100644
--- a/python/paddle/v2/framework/tests/test_net.py
+++ b/python/paddle/v2/framework/tests/test_net.py
@@ -15,7 +15,7 @@ def fc(X, W, Y):
 class TestNet(unittest.TestCase):
     def test_net_all(self):
         net = core.Net.create()
-        op1 = Operator("add_two", X="X", Y="Y", Out="Out")
+        op1 = Operator("add", X="X", Y="Y", Out="Out")
         net.append_op(op1)
 
         net2 = core.Net.create()
@@ -26,7 +26,7 @@ class TestNet(unittest.TestCase):
 
         expected = '''
 Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}.
-    Op(add_two), inputs:{X[X], Y[Y]}, outputs:{Out[Out]}.
+    Op(add), inputs:{X[X], Y[Y]}, outputs:{Out[Out]}.
     Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
         Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
             Op(mul), inputs:{X[X], Y[W]}, outputs:{Out[pre_activation]}.
diff --git a/python/paddle/v2/framework/tests/test_operator.py b/python/paddle/v2/framework/tests/test_operator.py
index 1abc4eeb57bcedc81e34b0e156048ee4f5cfdc2d..040556322d79cbb594eb9af585a5b9920d7ab625 100644
--- a/python/paddle/v2/framework/tests/test_operator.py
+++ b/python/paddle/v2/framework/tests/test_operator.py
@@ -193,10 +193,10 @@ class TestOpDescCreationMethod(unittest.TestCase):
 
 class TestOpCreations(unittest.TestCase):
     def test_all(self):
-        add_op = op.Operator("add_two", X="a", Y="b", Out="z")
+        add_op = op.Operator("add", X="a", Y="b", Out="z")
         self.assertIsNotNone(add_op)
         # Invoke C++ DebugString()
-        self.assertEqual('Op(add_two), inputs:{X[a], Y[b]}, outputs:{Out[z]}.',
+        self.assertEqual('Op(add), inputs:{X[a], Y[b]}, outputs:{Out[z]}.',
                          str(add_op))
 
 
diff --git a/python/paddle/v2/framework/tests/test_recurrent_op.py b/python/paddle/v2/framework/tests/test_recurrent_op.py
index d6000ab9f9d5b969f96128b183f48d49000c8a5e..22e680fd783ec681e95326fb84db34570265cffc 100644
--- a/python/paddle/v2/framework/tests/test_recurrent_op.py
+++ b/python/paddle/v2/framework/tests/test_recurrent_op.py
@@ -146,7 +146,7 @@ class TestRecurrentOp(unittest.TestCase):
         stepnet = core.Net.create()
         x_fc_op = Operator("mul", X="x@alias", Y="W", Out="Wx")
         h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
-        sum_op = Operator("add_two", X="Wx", Y="Uh", Out="sum")
+        sum_op = Operator("add", X="Wx", Y="Uh", Out="sum")
         sig_op = Operator("sigmoid", X="sum", Y="h@alias")
 
         for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
diff --git a/python/paddle/v2/framework/tests/test_softmax_op.py b/python/paddle/v2/framework/tests/test_softmax_op.py
index d7279df1162cfce3469f5f01ac2e84ae02fa3422..63042e9bfddadaea437d517a560636a0396aae49 100644
--- a/python/paddle/v2/framework/tests/test_softmax_op.py
+++ b/python/paddle/v2/framework/tests/test_softmax_op.py
@@ -18,10 +18,9 @@ class TestSoftmaxOp(unittest.TestCase):
 
     def setUp(self):
         self.type = "softmax"
-        self.inputs = {"logits": np.random.random((10, 10)).astype("float32")}
+        self.inputs = {"Logits": np.random.random((10, 10)).astype("float32")}
         self.outputs = {
-            "softmax":
-            np.apply_along_axis(stable_softmax, 1, self.inputs["logits"])
+            "Out": np.apply_along_axis(stable_softmax, 1, self.inputs["Logits"])
         }
 
 
@@ -29,11 +28,11 @@ class TestSoftmaxGradOp(GradientChecker):
     def setUp(self):
         self.op = create_op("softmax")
         self.inputs = {
-            "logits": np.random.uniform(0.1, 1, [10, 10]).astype("float32")
+            "Logits": np.random.uniform(0.1, 1, [10, 10]).astype("float32")
         }
 
     def test_softmax_grad(self):
-        self.check_grad(self.op, self.inputs, ["logits"], "softmax")
+        self.check_grad(self.op, self.inputs, ["Logits"], "Out")
 
 
 if __name__ == "__main__":