Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into op_transpose

.gitignore

@@ -22,6 +22,7 @@ cmake-build-*
 
 # generated while compiling
 python/paddle/v2/framework/core.so
+paddle/pybind/pybind.h
 CMakeFiles
 cmake_install.cmake
 paddle/.timestamp

CMakeLists.txt

@@ -67,6 +67,9 @@ endif()
 if(ANDROID)
     if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
         message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
+    elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
+        # TODO: support glog for Android api 16 ~ 19 in the future
+        message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
     endif()
 
     set(WITH_GPU OFF CACHE STRING

Dockerfile.android

@@ -6,13 +6,14 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
 
 # ENV variables
 ARG ANDROID_ABI
+ARG ANDROID_API
 
 ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
+ENV ANDROID_API=${ANDROID_API:-21}
 
 ENV HOME=/root \
     ANDROID_NDK_HOME=/opt/android-ndk-linux \
-    ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \
-    ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain
+    ANDROID_TOOLCHAINS_DIR=/opt/toolchains
 
 RUN apt-get update && \
     apt-get install -y \
@@ -42,14 +43,12 @@ RUN pip install --upgrade pip && \
     pip install pre-commit
 
 # Android NDK
-RUN mkdir /opt/android-ndk-tmp && \
+RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
+    mkdir -p /opt/android-ndk-tmp && \
     cd /opt/android-ndk-tmp && \
     wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
     unzip -q android-ndk-r14b-linux-x86_64.zip && \
     mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \
-    rm -rf /opt/android-ndk-tmp && \
-    rm -rf ${ANDROID_NDK_HOME}
+    rm -rf /opt/android-ndk-tmp
 
 CMD ["bash", "/paddle/paddle/scripts/docker/build_android.sh"]

cmake/cpplint.cmake

@@ -26,9 +26,9 @@ set(IGNORE_PATTERN
     .*ImportanceSampler.*
     .*cblas\\.h.*
     .*\\.pb\\.txt
-    .*LtrDataProvider.*
     .*MultiDataProvider.*
-    .*pb.*)
+    .*pb.*
+    .*pybind.h)
 
 # add_style_check_target
 #

cmake/external/gflags.cmake

@@ -56,3 +56,12 @@ SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
 ADD_DEPENDENCIES(gflags extern_gflags)
 
 LIST(APPEND external_project_dependencies gflags)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
+  IF(ANDROID)
+    INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
+  ENDIF()
+ENDIF()

cmake/external/glog.cmake

@@ -56,3 +56,12 @@ ADD_DEPENDENCIES(glog extern_glog gflags)
 LINK_LIBRARIES(glog gflags)
 
 LIST(APPEND external_project_dependencies glog)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
+  IF(ANDROID)
+    INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib)
+  ENDIF()
+ENDIF()

cmake/external/openblas.cmake

@@ -73,6 +73,26 @@ IF(NOT ${CBLAS_FOUND})
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )
+
+    IF(WITH_C_API)
+        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
+        # Because libopenblas.a is a symbolic link of another library, thus need to
+        # install the whole directory.
+        IF(ANDROID)
+            SET(TMP_INSTALL_DIR third_party/openblas/lib/${ANDROID_ABI})
+        ELSE()
+            SET(TMP_INSTALL_DIR third_party/openblas/lib)
+        ENDIF()
+        INSTALL(CODE "execute_process(
+            COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
+                    destination ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
+            )"
+        )
+        INSTALL(CODE "MESSAGE(STATUS \"Installing: \"
+                \"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
+            )"
+        )
+    ENDIF()
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")

cmake/external/protobuf.cmake

@@ -223,6 +223,15 @@ IF(NOT PROTOBUF_FOUND)
     SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
         CACHE FILEPATH "protoc library." FORCE)
 
+    IF(WITH_C_API)
+        INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
+        IF(ANDROID)
+            INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
+        ELSE()
+            INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib)
+        ENDIF()
+    ENDIF()
+
     IF(CMAKE_CROSSCOMPILING)
         PROMPT_PROTOBUF_LIB(protobuf_host extern_protobuf)
     ELSE()

cmake/external/zlib.cmake

@@ -49,3 +49,12 @@ ExternalProject_Add(
 )
 
 LIST(APPEND external_project_dependencies zlib)
+
+IF(WITH_C_API)
+  INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)
+  IF(ANDROID)
+    INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib/${ANDROID_ABI})
+  ELSE()
+    INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib)
+  ENDIF()
+ENDIF()

doc/design/ops/images/2_level_rnn.dot

+digraph G {
+
+  rnn [label="1-th level RNN" shape=box]
+
+  subgraph cluster0 {
+    label = "time step 0"
+
+    sent0 [label="sentence"]
+    sent1 [label="sentence"]
+
+    rnn1 [label="2-th level RNN" shape=box]
+
+    sent0 -> rnn1
+    sent1 -> rnn1
+  }
+
+  subgraph cluster1 {
+    label = "time step 1"
+
+    sent2 [label="sentence"]
+    sent3 [label="sentence"]
+
+    rnn2 [label="2-th level RNN" shape=box]
+
+    sent2 -> rnn2
+    sent3 -> rnn2
+  }
+
+  subgraph cluster2 {
+    label = "time step 2"
+
+    sent4 [label="sentence"]
+    sent5 [label="sentence"]
+
+    rnn3 [label="2-th level RNN" shape=box]
+
+    sent4 -> rnn3
+    sent5 -> rnn3
+  }
+
+
+  para0 [label="paragraph info 0"]
+  para1 [label="paragraph info 1"]
+  para2 [label="paragraph info 2"]
+
+  rnn1 -> para0
+  rnn2 -> para1
+  rnn3 -> para2
+
+  para0 -> rnn
+  para1 -> rnn
+  para2 -> rnn
+
+  chapter [label="chapter info"]
+  rnn -> chapter
+}

doc/design/ops/images/rnn.dot

+digraph G {
+  label = "simple RNN implementation" 
+
+  ranksep=2;
+
+  //graph [nodesep=1, ranksep=1];
+
+  node[nodesep=1]
+
+  subgraph cluster0 {
+    label = "global scope"
+    rankdir = TB
+    W
+    boot_memory
+    input
+    output
+  }
+
+  subgraph cluster1 {
+    label = "step-scope 0"
+    rankdir = TB
+    memory0[label="memory"]
+    prememory0[label="pre-memory"]
+    step_input0[label="step input"]
+    step_output0[label="step output"]
+  }
+
+  subgraph cluster2 {
+    label = "step-scope 1"
+    rankdir = TB
+    memory1[label="memory"]
+    prememory1[label="pre-memory"]
+    step_input1[label="step input"]
+    step_output1[label="step output"]
+  }
+
+  subgraph cluster3 {
+    label = "step-scope 2"
+    rankdir = TB
+    memory2[label="memory"]
+    prememory2[label="pre-memory"]
+    step_input2[label="step input"]
+    step_output2[label="step output"]
+  }
+
+  stepnet [shape=box]
+  stepnet0 [shape=box, style=dashed]
+  stepnet1 [shape=box, style=dashed]
+  stepnet2 [shape=box, style=dashed]
+
+
+  edge[color=blue]
+  boot_memory -> prememory0 [label="init" color="blue"]
+  memory0 -> prememory1  [label="copy/reference" color="blue"]
+  memory1 -> prememory2 [label="copy/reference" color="blue"]
+
+  edge[color=black]
+  W -> stepnet0[constraint=false, style=dashed]
+  W -> stepnet1[constraint=false, style=dashed]
+  W -> stepnet2[constraint=false, style=dashed]
+
+  memory0 -> stepnet0[style=dashed]
+  prememory0 -> stepnet0 -> step_output0[style=dashed]
+
+  memory1 -> stepnet1[style=dashed]
+  prememory1 -> stepnet1 -> step_output1[style=dashed]
+
+  memory2 -> stepnet2[style=dashed]
+  prememory2 -> stepnet2 -> step_output2[style=dashed]
+
+  input -> step_input0
+  input -> step_input1
+  input -> step_input2
+
+  step_input0 -> stepnet0 [style=dashed]
+  step_input1 -> stepnet1[style=dashed]
+  step_input2 -> stepnet2[style=dashed]
+
+  step_output0 -> output
+  step_output1 -> output
+  step_output2 -> output
+
+  stepnet0 -> stepnet[style=dashed]
+  stepnet1 -> stepnet[style=dashed]
+  stepnet2 -> stepnet[style=dashed]
+
+}

doc/design/ops/images/rnn_2level_data.dot

+digraph G {
+  chapter [label="chapter"]
+
+  subgraph cluster0 {
+    label = "paragraph 0"
+
+    top_rnn0[label="top rnn step 0" shape=box]
+
+    p0 [label="paragraph 0"]
+    p1 [label="paragraph 1"]
+  }
+
+  subgraph cluster1{
+    label = "paragraph 1"
+
+    top_rnn1[label="top rnn step 1" shape=box]
+
+    p2 [label="paragraph 0"]
+    p3 [label="paragraph 1"]
+  }
+
+  subgraph cluster_p0 {
+    label = "sentence 0"
+
+    low_rnn0 [label="low rnn step 0" shape=box]
+    s00 [label="sentence 0"]
+    s01 [label="sentence 1"]
+
+    low_rnn0 -> s00
+    low_rnn0 -> s01
+  }
+
+  subgraph cluster_p1 {
+    label = "sentence 1"
+    low_rnn1 [label="low rnn step 1" shape=box]
+    s10 [label="sentence 0"]
+    s11 [label="sentence 1"]
+    low_rnn1 -> s10
+    low_rnn1 -> s11
+  }
+
+  subgraph cluster_p2 {
+    label = "sentence 1"
+    low_rnn2 [label="low rnn step 0" shape=box]
+    s20 [label="sentence 0"]
+    s21 [label="sentence 1"]
+    low_rnn2 -> s20
+    low_rnn2 -> s21
+  }
+
+  subgraph cluster_p3 {
+    label = "sentence 1"
+    low_rnn3 [label="low rnn step 1" shape=box]
+    s30 [label="sentence 0"]
+    s31 [label="sentence 1"]
+    low_rnn3 -> s30
+    low_rnn3 -> s31
+  }
+
+
+  chapter -> top_rnn0
+  chapter -> top_rnn1
+
+  top_rnn0 -> p0
+  top_rnn0 -> p1
+  top_rnn1 -> p2
+  top_rnn1 -> p3
+
+
+  p0 -> low_rnn0
+  p1 -> low_rnn1
+  p2 -> low_rnn2
+  p3 -> low_rnn3
+
+}

doc/design/ops/rnn.md

+# RNNOp design
+
+This document is about an RNN operator which requires that instances in a mini-batch have the same length.  We will have a more flexible RNN operator.
+
+## RNN Algorithm Implementation
+
+<p aligh="center">
+<img src="./images/rnn.jpg"/>
+</p>
+
+The above diagram shows an RNN unrolled into a full network.
+
+There are several important concepts:
+
+- *step-net*: the sub-graph to run at each step,
+- *memory*, $h_t$, the state of the current step,
+- *ex-memory*, $h_{t-1}$, the state of the previous step,
+- *initial memory value*, the ex-memory of the first step.
+
+### Step-scope
+
+There could be local variables defined in step-nets.  PaddlePaddle runtime realizes these variables in *step-scopes* -- scopes created for each step.
+
+<p aligh="center">
+<img src="./images/rnn.png"/><br/>
+Figure 2 the RNN's data flow
+</p>
+
+Please be aware that all steps run the same step-net.  Each step
+
+1. creates the step-scope,
+2. realizes local variables, including step-outputs, in the step-scope, and
+3. runs the step-net, which could use these variables.
+
+The RNN operator will compose its output from step outputs in step scopes.
+
+### Memory and Ex-memory
+
+Let's give more details about memory and ex-memory via a simply example:
+
+$$
+h_t = U h_{t-1} + W x_t
+$$,
+
+where $h_t$ and $h_{t-1}$ are the memory and ex-memory of step $t$'s respectively.
+
+In the implementation, we can make an ex-memory variable either "refers to" the memory variable of the previous step,
+or copy the value of the previous memory value to the current ex-memory variable.
+
+### Usage in Python
+
+For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md).
+
+We can define an RNN's step-net using Block:
+
+```python
+import paddle as pd
+
+X = some_op() # x is some operator's output, and is a LoDTensor
+a = some_op()
+
+# declare parameters
+W = pd.Variable(shape=[20, 30])
+U = pd.Variable(shape=[20, 30])
+
+rnn = pd.create_rnn_op(output_num=1)
+with rnn.stepnet():
+    x = rnn.add_input(X)
+    # declare a memory (rnn's step)
+    h = rnn.add_memory(init=a)
+    # h.pre_state() means previous memory of rnn
+    new_state = pd.add_two( pd.matmul(W, x) + pd.matmul(U, h.pre_state()))
+    # update current memory
+    h.update(new_state)
+    # indicate that h variables in all step scopes should be merged
+    rnn.add_outputs(h)
+
+out = rnn()
+```
+
+Python API functions in above example:
+
+- `rnn.add_input` indicates the parameter is a variable that will be segmented into step-inputs.
+- `rnn.add_memory` creates a variable used as the memory.
+- `rnn.add_outputs` mark the variables that will be concatenated across steps into the RNN output.
+
+### Nested RNN and LoDTensor
+
+An RNN whose step-net includes other RNN operators is known as an *nested RNN*.
+
+For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences.
+
+The following figure illustrates the feeding of text into the lower level, one sentence each step, and the feeding of step outputs to the top level. The final top level output is about the whole text.
+
+<p aligh="center">
+<img src="./images/2_level_rnn.png"/>
+</p>
+
+```python
+import paddle as pd
+
+W = pd.Variable(shape=[20, 30])
+U = pd.Variable(shape=[20, 30])
+
+W0 = pd.Variable(shape=[20, 30])
+U0 = pd.Variable(shape=[20, 30])
+
+# a is output of some op
+a = some_op()
+
+# chapter_data is a set of 128-dim word vectors
+# the first level of LoD is sentence
+# the second level of LoD is chapter
+chapter_data = pd.Variable(shape=[None, 128], type=pd.lod_tensor, level=2)
+
+def lower_level_rnn(paragraph):
+    '''
+    x: the input
+    '''
+    rnn = pd.create_rnn_op(output_num=1)
+    with rnn.stepnet():
+        sentence = rnn.add_input(paragraph, level=0)
+        h = rnn.add_memory(shape=[20, 30])
+        h.update(
+            pd.matmul(W, sentence) + pd.matmul(U, h.pre_state()))
+        # get the last state as sentence's info
+        rnn.add_outputs(h)
+    return rnn
+
+top_level_rnn = pd.create_rnn_op(output_num=1)
+with top_level_rnn.stepnet():
+    paragraph_data = rnn.add_input(chapter_data, level=1)
+    low_rnn = lower_level_rnn(paragraph_data)
+    paragraph_out = low_rnn()
+
+    h = rnn.add_memory(init=a)
+    h.update(
+        pd.matmul(W0, paragraph_data) + pd.matmul(U0, h.pre_state()))
+    top_level_rnn.add_outputs(h)
+
+# just output the last step
+chapter_out = top_level_rnn(output_all_steps=False)
+```
+
+in above example, the construction of the `top_level_rnn` calls  `lower_level_rnn`.  The input is a LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences.
+
+By default, the `RNNOp` will concatenate the outputs from all the time steps,
+if the `output_all_steps` set to False, it will only output the final time step.
+
+
+<p align="center">
+<img src="images/rnn_2level_data.png"/>
+</p>

doc/howto/dev/new_op_cn.md

@@ -34,7 +34,7 @@ Kernel实现       | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU
 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。
+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。**系统会根据文件名自动构建op和其对应的Python扩展。**
 
 
 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。
@@ -224,45 +224,15 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
 ### 5. 编译
 
-- 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：
+运行下面命令可以进行编译：
 
-    ```
-    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
-    ```
+```
+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) 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);
-    ```
-
-
- - 生成库
-
-   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。
+系统会对新增的op自动绑定Python，并链接到生成的lib库中。
 
 ## 实现单元测试
 
@@ -354,11 +324,7 @@ class TestMulGradOp(GradientChecker):
 
 ### 编译和执行单元测试
 
-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)
-```
+`python/paddle/v2/framework/tests` 目录下新增的 `test_*.py` 单元测试会被自动加入工程进行编译。
 
 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：
 
@@ -371,3 +337,10 @@ make test ARGS="-R test_mul_op -V"
 ```bash
 ctest -R test_mul_op
 ```
+
+## 注意事项
+
+- 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，这将会导致编译出错。
+- 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。
+- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
+- 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。

doc/howto/dev/write_docs_cn.rst

@@ -5,15 +5,13 @@
 PaddlePaddle的文档包括英文文档 ``doc`` 和中文文档 ``doc_cn`` 两个部分。文档都是通过 `cmake`_ 驱动 `sphinx`_ 编译生成，生成后的文档分别存储在编译目录的 ``doc`` 和 ``doc_cn`` 两个子目录下。
 
 
-如何构建PaddlePaddle的文档
-==========================
+如何构建文档
+============
 
-PaddlePaddle的文档构建有直接构建和基于Docker构建两种方式，我们提供了一个构建脚本build_docs.sh来进行构建。
-PaddlePaddle文档需要准备的环境相对较复杂，所以我们推荐使用基于Docker来构建PaddlePaddle的文档。
+PaddlePaddle的文档构建有两种方式。
 
-
-使用Docker构建PaddlePaddle的文档
---------------------------------
+使用Docker构建
+--------------
 
 使用Docker构建PaddlePaddle的文档，需要在系统里先安装好Docker工具包。Docker安装请参考 `Docker的官网 <https://docs.docker.com/>`_ 。安装好Docker之后可以使用源码目录下的脚本构建文档，即
 
@@ -21,58 +19,46 @@ PaddlePaddle文档需要准备的环境相对较复杂，所以我们推荐使
 
     cd TO_YOUR_PADDLE_CLONE_PATH
     cd paddle/scripts/tools/build_docs
-    bash build_docs.sh with_docker
-
-编译完成后，会在当前目录生成两个子目录\:
-
-* doc 英文文档目录
-* doc_cn 中文文档目录
+    sh build_docs.sh
 
+编译完成之后，会在当前目录生成两个子目录\: doc(英文文档目录)和 doc_cn(中文文档目录)。
 打开浏览器访问对应目录下的index.html即可访问本地文档。
 
-
-
-直接构建PaddlePaddle的文档
---------------------------
-
-因为PaddlePaddle的v2 api文档生成过程依赖于py_paddle Python包，用户需要首先确认py_paddle包已经安装。
-
-..  code-block:: bash
-
-    python -c "import py_paddle"
-
-如果提示错误，那么用户需要在本地编译安装PaddlePaddle，请参考 `源码编译文档 <http://doc.paddlepaddle.org/develop/doc/getstarted/build_and_install/build_from_source_en.html>`_ 。
-注意，用户在首次编译安装PaddlePaddle时，请将WITH_DOC选项关闭。在编译安装正确之后，请再次确认py_paddle包已经安装，即可进行下一步操作。
+直接构建
+--------
 
 如果提示正确，可以执行以下命令编译生成文档，即
 
 ..  code-block:: bash
 
     cd TO_YOUR_PADDLE_CLONE_PATH
-    cd paddle/scripts/tools/build_docs
-    bash build_docs.sh local
-
-编译完成之后，会在当前目录生成两个子目录\:
-
-* doc 英文文档目录
-* doc_cn 中文文档目录
+    mkdir -p build
+    cd build
+    cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
+    make gen_proto_py
+    make paddle_docs paddle_docs_cn
 
+编译完成之后，会在当前目录生成两个子目录\: doc(英文文档目录)和 doc_cn(中文文档目录)。
 打开浏览器访问对应目录下的index.html即可访问本地文档。
 
 
-如何书写PaddlePaddle的文档
-==========================
+如何书写文档
+============
 
 PaddlePaddle文档使用 `sphinx`_ 自动生成，用户可以参考sphinx教程进行书写。
 
-如何更新www.paddlepaddle.org文档
-================================
+如何更新文档主题
+================
+
+PaddlePaddle文档主题在 `TO_YOUR_PADDLE_CLONE_PATH/doc_theme` 文件夹下，包含所有和前端网页设计相关的文件。
 
-开发者给PaddlePaddle代码增加的注释以PR的形式提交到github中，提交方式可参见 `贡献文档 <http://doc.paddlepaddle.org/develop/doc_cn/howto/dev/contribute_to_paddle_cn.html>`_ 。
+如何更新doc.paddlepaddle.org
+============================
+
+更新的文档以PR的形式提交到github中，提交方式参见 `贡献文档 <http://doc.paddlepaddle.org/develop/doc_cn/howto/dev/contribute_to_paddle_cn.html>`_ 。
 目前PaddlePaddle的develop分支的文档是自动触发更新的，用户可以分别查看最新的 `中文文档 <http://doc.paddlepaddle.org/develop/doc_cn/>`_ 和
 `英文文档 <http://doc.paddlepaddle.org/develop/doc/>`_ 。
 
 
-
 ..  _cmake: https://cmake.org/
 ..  _sphinx: http://www.sphinx-doc.org/en/1.4.8/

paddle/capi/CMakeLists.txt

@@ -64,9 +64,29 @@ link_paddle_exe(paddle_capi_shared)
 install(FILES ${CAPI_HEADERS} DESTINATION include/paddle)
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/config.h DESTINATION include/paddle)
 if(ANDROID)
+  execute_process(
+    COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1
+    OUTPUT_VARIABLE GIT_COMMITS_LIST
+    RESULT_VARIABLE GIT_COMMITS_LIST_RESULT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+  if(${GIT_COMMITS_LIST_RESULT})
+    set(GIT_COMMITS_LIST "No commits.")
+  endif()
   install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}
           DESTINATION lib/${ANDROID_ABI})
   install(TARGETS paddle_capi_shared DESTINATION lib/${ANDROID_ABI})
+  install(CODE "FILE(WRITE ${CMAKE_INSTALL_PREFIX}/lib/${ANDROID_ABI}/BUILD.txt
+          \"Compiler:\n\"
+          \"\\t${CMAKE_C_COMPILER}\\n\"
+          \"\\t${CMAKE_CXX_COMPILER}\\n\"
+          \"Compiler Flags:\\n\"
+          \"\\t${CMAKE_F_FLAGS}\\n\"
+          \"\\t${CMAKE_CXX_FLAGS}\\n\"
+          \"Android API: ${CMAKE_SYSTEM_VERSION}\\n\"
+          \"Lastest commit:\\n\"
+          \"\\t${GIT_COMMITS_LIST}\\n\"
+      )"
+  )
 else(ANDROID)
   install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library} DESTINATION lib)
   install(TARGETS paddle_capi_shared DESTINATION lib)

paddle/framework/CMakeLists.txt

@@ -9,6 +9,7 @@ cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
 cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor)
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor)
+nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
 cc_test(variable_test SRCS variable_test.cc)
 

paddle/framework/backward.md

@@ -2,11 +2,22 @@
 
 ## Motivation
 
-In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
+In Neural Network, many model is solved by the the backpropagation algorithm(known as BP) at present. Technically it caculates the gradient of the loss function, then distributed back through the networks. Follows the chain rule, so we need a module chains the gradient operators/expressions together with to construct the backward pass. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. 
 
-## Backward Operator Registry
+## Implementation
 
-A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients.
+In this design doc, we exported only one API for generating the backward pass.
+
+```c++
+std::unique_ptr<OperatorBase> Backward(const OperatorBase& forwardOp,
+    const std::unordered_set<std::string>& no_grad_vars);
+```
+
+The implementation behind it can be divided into two parts, **Backward Operator Creating** and **Backward Operator Building**.
+
+### Backward Operator Registry
+
+A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs, and output gradients and then calculate its input gradients.
 
 |                        | forward operator | backward operator 
 | ---------------------- | ---------------- |------------------------- |		
@@ -25,7 +36,7 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
 
 `mul_grad` is the type of backward operator, and `MulOpGrad` is its class name.
 
-## Backward Opeartor Creating
+### Backward Opeartor Creating
 
 Given a certain forward operator, we can get its corresponding backward operator by calling:
 
@@ -43,40 +54,47 @@ The function `BuildGradOp` will sequentially execute following processes:
 
 4. Building backward operator with `inputs`, `outputs` and forward operator's attributes.
 
-## Backward Network Building
-
-A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and put them together.
+### Backward Network Building
 
-In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. 
-
-given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`.
+A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and append them together one by one. There is some corner case need to process specially.
 
 1. Op 
 
-   when the input forward network is an Op, return its gradient Operator Immediately.
+   When the input forward network is an Op, return its gradient Operator Immediately. If all of its outputs are in no gradient set, then return a special `NOP`.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
+   In our design, the network itself is also a kind of operator(**NetOp**). So the operators contained by a big network may be some small network. When the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
+
+3. RnnOp
+
+   RnnOp is a nested stepnet operator.  Backward module need to recusively call `Backward` for every stepnet.
+
+4. Sharing Variables
+
+   **sharing variables**. As illustrated in the pictures, two operator's share the same variable name of W@GRAD, which will overwrite their sharing input variable. 
+
+<p align="center">
+<img src="./images/duplicate_op.png" width="50%" ><br/>
 
-   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable.  
+​	pic 1. Sharing variables in operators. 
 
-   <p align="center">
-   <img src="./images/duplicate_op.png" width="50%" ><br/>
+</p>
 
-   1. Shared variable in operators. 
+​	Sharing variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator to replace the overwrite links. 
 
-   </p>
+<p align="center">
+<img src="images/duplicate_op2.png" width="40%" ><br/>
 
-   Share variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator replace the overwrite links. 
+​	pic 2. Replace sharing variable's gradient with `Add` operator.
 
-   <p align="center">
-   <img src="images/duplicate_op2.png" width="50%" ><br/>
+</p>
 
-   2. Replace shared variable's gradient with `Add` operator.
+​	Because our framework finds variables accord to their names, we need to rename the output links. We add a suffix of number to represent its position in clockwise. 
 
-   </p>
+5. Part of Gradient is Zero.
 
+   In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator,  we need to fill a same shape gradient matrix in the position. In our implement, we insert a special `fillZeroLike` operator.
 
 
-​	Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.
+Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.

paddle/framework/lod_tensor.h

@@ -18,8 +18,10 @@
 #ifndef PADDLE_ONLY_CPU
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
+#include <thrust/system/cuda/experimental/pinned_allocator.h>
 #endif
 
+#include <glog/logging.h>
 #include "paddle/framework/ddim.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/enforce.h"
@@ -32,7 +34,8 @@ template <typename T>
 using Vector = std::vector<T>;
 #else
 template <typename T>
-using Vector = thrust::host_vector<T>;
+using Vector = thrust::host_vector<
+    T, thrust::system::cuda::experimental::pinned_allocator<T>>;
 #endif
 
 using LoD = std::vector<Vector<size_t>>;
@@ -48,18 +51,15 @@ bool operator==(const LoD& a, const LoD& b);
  * LoDTensor (Level of details Tensor)
  * see https://en.wikipedia.org/wiki/Level_of_details for reference.
  */
-class LoDTensor {
+class LoDTensor : public Tensor {
  public:
   LoDTensor() {}
-  LoDTensor(const LoD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
 
-  void set_lod(const LoD& lod) { lod_ = lod; }
-
-  void set_tensor(Tensor* tensor) { tensor_ = tensor; }
+  explicit LoDTensor(const LoD& lod) : lod_(lod) {}
 
-  Tensor& tensor() { return *tensor_; }
+  void set_lod(const LoD& lod) { lod_ = lod; }
 
-  LoD lod() { return lod_; }
+  LoD lod() const { return lod_; }
 
   /*
    * Get a element from LoD.
@@ -101,7 +101,6 @@ class LoDTensor {
 
  private:
   LoD lod_;
-  Tensor* tensor_;  // not owned
 };
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor_test.cc

@@ -36,69 +36,64 @@ class LoDTensorTester : public ::testing::Test {
 
     ASSERT_EQ(lod.size(), 3UL);
 
-    tensor.Resize({20 /*batch size*/, 128 /*dim*/});
+    lod_tensor_.Resize({20 /*batch size*/, 128 /*dim*/});
     // malloc memory
-    tensor.mutable_data<float>(place);
+    lod_tensor_.mutable_data<float>(place);
 
-    lod_tensor.set_lod(lod);
-    lod_tensor.set_tensor(&tensor);
+    lod_tensor_.set_lod(lod);
   }
 
  protected:
   platform::CPUPlace place;
-  Tensor tensor;
-  LoDTensor lod_tensor;
+  LoDTensor lod_tensor_;
 };
 
-TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
+TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor_.NumLevels(), 3UL); }
 
 TEST_F(LoDTensorTester, NumElements) {
-  ASSERT_EQ(lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(lod_tensor.NumElements(2), 8UL);
+  ASSERT_EQ(lod_tensor_.NumElements(0), 2UL);
+  ASSERT_EQ(lod_tensor_.NumElements(1), 4UL);
+  ASSERT_EQ(lod_tensor_.NumElements(2), 8UL);
 }
 
 TEST_F(LoDTensorTester, SliceLevels) {
   // slice 1 level
   for (size_t level = 0; level < 3UL; ++level) {
-    LoDTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.SliceLevels(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
-    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-              lod_tensor.tensor().data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    LoDTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.SliceLevels(level, level + 2);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
-    ASSERT_EQ(new_lod_tensor.NumElements(1), lod_tensor.NumElements(level + 1));
-    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-              lod_tensor.tensor().data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.NumElements(1),
+              lod_tensor_.NumElements(level + 1));
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
 }
 
 TEST_F(LoDTensorTester, SliceInLevel) {
   size_t level = 0;
-  LoDTensor new_lod_tensor = lod_tensor;
+  LoDTensor new_lod_tensor = lod_tensor_;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
   EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
   EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
   EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-            lod_tensor.tensor().data<float>());
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor;
+  new_lod_tensor = lod_tensor_;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-            lod_tensor.tensor().data<float>());
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 }
 
 }  // namespace framework

paddle/framework/lod_tensor_test.cu

+/*
+  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 <cuda.h>
+#include <cuda_runtime.h>
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/platform/assert.h"
+
+#include <gtest/gtest.h>
+
+__global__ void test(size_t* a, int size) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < size;
+       i += blockDim.x * gridDim.x) {
+    a[i] *= 2;
+  }
+}
+
+TEST(LoDTensor, LoDInGPU) {
+  paddle::framework::LoDTensor lod_tensor;
+  paddle::platform::GPUPlace place(0);
+
+  paddle::framework::LoD src_lod;
+  src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});
+
+  lod_tensor.Resize({14, 16});
+  lod_tensor.mutable_data<float>(place);
+
+  lod_tensor.set_lod(src_lod);
+  CHECK_EQ(lod_tensor.lod_element(0, 2), 4);
+  CHECK_EQ(lod_tensor.lod_element(0, 4), 8);
+
+  auto lod = lod_tensor.lod();
+
+  test<<<1, 8>>>(lod[0].data(), lod[0].size());
+  cudaDeviceSynchronize();
+
+  for (size_t i = 0; i < src_lod[0].size(); ++i) {
+    CHECK_EQ(lod[0].data()[i], src_lod[0].data()[i] * 2);
+  }
+}

paddle/framework/operator.cc

@@ -186,6 +186,48 @@ void OperatorBase::GenerateTemporaryNames() {
   }
 }
 
+template <>
+const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const {
+  auto* var = InputVar(name);
+  return var == nullptr ? nullptr : GetTensorFromVar(var);
+}
+
+template <>
+const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+    const std::string& name) const {
+  auto names = op().Inputs(name);
+  std::vector<const Tensor*> res;
+  res.reserve(names.size());
+  std::transform(names.begin(), names.end(), std::back_inserter(res),
+                 [&](const std::string& sub_name) {
+                   auto var = scope_.FindVar(sub_name);
+                   return var == nullptr ? nullptr : GetTensorFromVar(var);
+                 });
+  return res;
+}
+
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
+  auto* var = OutputVar(name);
+  return var == nullptr ? nullptr : const_cast<Tensor*>(GetTensorFromVar(var));
+}
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const {
+  auto names = op().Outputs(name);
+  std::vector<Tensor*> res;
+  res.reserve(names.size());
+  std::transform(names.begin(), names.end(), std::back_inserter(res),
+                 [&](const std::string& sub_name) {
+                   auto var = scope().FindVar(sub_name);
+                   return var == nullptr
+                              ? nullptr
+                              : const_cast<Tensor*>(GetTensorFromVar(var));
+                 });
+  return res;
+}
+
 void OpProtoAndCheckerMaker::Validate() {
   validated_ = true;
   CheckNoDuplicatedInOutAttrs();

paddle/framework/operator.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #include "op_info.h"
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/framework.pb.h"
+#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
@@ -326,11 +327,27 @@ class InferShapeContext {
     return res;
   }
 
+  const Tensor* GetTensorFromVar(const Variable* var) const {
+    if (var->IsType<LoDTensor>()) {
+      return &var->Get<LoDTensor>();
+    }
+    PADDLE_ENFORCE(var->IsType<Tensor>(),
+                   "The Input(%s) must be LoDTensor or Tensor.");
+    return &var->Get<Tensor>();
+  }
+
  private:
   const OperatorBase& op_;
   const Scope& scope_;
 };
 
+template <>
+const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const;
+
+template <>
+const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+    const std::string& name) const;
+
 template <typename T>
 struct EigenDeviceConverter;
 
@@ -363,9 +380,37 @@ class ExecutionContext : public InferShapeContext {
     return device_context_;
   }
 
+  // redefine Output function,
+  // use Variable::Get instead of Variable::GetMutable
+  template <typename T>
+  T* Output(const std::string& name) const {
+    auto var = OutputVar(name);
+    return var == nullptr ? nullptr : const_cast<T*>(&var->Get<T>());
+  }
+
+  // redefine MultiOutput function.
+  // use Variable::Get instead of Variable::GetMutable
+  template <typename T>
+  std::vector<T*> MultiOutput(const std::string& name) const {
+    auto names = op().Outputs(name);
+    std::vector<T*> res;
+    res.reserve(names.size());
+    std::transform(
+        names.begin(), names.end(), std::back_inserter(res),
+        [&](const std::string& sub_name) { return Output<T>(sub_name); });
+    return res;
+  }
+
   const platform::DeviceContext* device_context_;
 };
 
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const;
+
 class OpKernel {
  public:
   /**

paddle/framework/tensor.h

@@ -81,6 +81,9 @@ class Tensor {
   /*! Return the dimensions of the memory block. */
   inline const DDim& dims() const;
 
+  /*! Return the numel of the memory block. */
+  inline int64_t numel() const;
+
   /*! Resize the dimensions of the memory block. */
   inline Tensor& Resize(const DDim& dims);
 
@@ -162,6 +165,12 @@ class Tensor {
   /*! points to dimensions of memory block. */
   DDim dims_;
 
+  /**
+   * A cache of the number of elements in a tensor.
+   * Would be 0 for an uninitialized tensor.
+   */
+  int64_t numel_;
+
   /**
    * @brief   A PlaceHolder may be shared by more than one tensor.
    *

paddle/framework/tensor_impl.h

@@ -24,7 +24,7 @@ inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
       holder_, "Tenosr holds no memory. Call Tensor::mutable_data first.");
   PADDLE_ENFORCE_GE(
-      holder_->size(), product(dims_) * sizeof(T) + offset_,
+      holder_->size(), numel() * sizeof(T) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
       "first to re-allocate memory.\n"
       "or maybe the required data-type mismatches the data already stored.");
@@ -54,11 +54,11 @@ inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
 template <typename T>
 inline T* Tensor::mutable_data(platform::Place place) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  PADDLE_ENFORCE_GT(product(dims_), 0,
+  PADDLE_ENFORCE_GT(numel(), 0,
                     "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!= */
-  int64_t size = product(dims_) * sizeof(T);
+  int64_t size = numel() * sizeof(T);
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
     if (platform::is_cpu_place(place)) {
@@ -97,7 +97,7 @@ inline void Tensor::CopyFrom(const Tensor& src,
 
   auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
 
-  auto size = product(src.dims_) * sizeof(T);
+  auto size = src.numel() * sizeof(T);
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
@@ -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.");
-  size_t base = product(dims_) / dims_[0];
+  size_t base = numel() / dims_[0];
   Tensor dst;
   dst.holder_ = holder_;
   DDim dst_dims = dims_;
@@ -143,11 +143,14 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
 
 inline Tensor& Tensor::Resize(const DDim& dims) {
   dims_ = dims;
+  numel_ = product(dims_);
   return *this;
 }
 
 inline const DDim& Tensor::dims() const { return dims_; }
 
+inline int64_t Tensor::numel() const { return numel_; }
+
 template <typename T>
 inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
   Tensor res;

paddle/function/neon/NeonDepthwiseConv.h

@@ -594,7 +594,7 @@ struct StridePadding {
         float32x4_t s1 = vdupq_n_f32(0.f);
         for (int s = 0; s < step; s++) {
           float32x4_t s0 = vld1q_f32(input);
-          float32x4x2_t v = {s0, s1};
+          float32x4x2_t v = {{s0, s1}};
           vst2q_f32(inputPadding, v);
           input += 4;
           inputPadding += 8;

paddle/gserver/layers/DeConv3DLayer.cpp

@@ -53,27 +53,27 @@ bool DeConv3DLayer::init(const LayerMap &layerMap,
 
 size_t DeConv3DLayer::getSize() {
   CHECK_NE(inputLayers_.size(), 0UL);
-  outputH_.clear();
-  outputW_.clear();
-  outputD_.clear();
+  imgSizeW_.clear();
+  imgSizeH_.clear();
+  imgSizeD_.clear();
   N_.clear();
   NOut_.clear();
   size_t layerSize = 0;
   for (size_t i = 0; i < inputLayers_.size(); ++i) {
-    outputW_.push_back(
-        imageSize(imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true));
-    outputH_.push_back(imageSize(
-        imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
-    outputD_.push_back(imageSize(
-        imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
-    NOut_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
-    N_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]);
+    imgSizeW_.push_back(
+        imageSize(outputW_[i], filterSize_[i], padding_[i], stride_[i], true));
+    imgSizeH_.push_back(imageSize(
+        outputH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
+    imgSizeD_.push_back(imageSize(
+        outputD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
+    NOut_.push_back(imgSizeD_[i] * imgSizeH_[i] * imgSizeW_[i]);
+    N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
     CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize);
     layerSize += NOut_[i] * numFilters_;
   }
-  getOutput().setFrameHeight(outputH_[0]);
-  getOutput().setFrameWidth(outputW_[0]);
-  getOutput().setFrameDepth(outputD_[0]);
+  getOutput().setFrameHeight(imgSizeH_[0]);
+  getOutput().setFrameWidth(imgSizeW_[0]);
+  getOutput().setFrameDepth(imgSizeD_[0]);
   return layerSize;
 }
 
@@ -103,9 +103,9 @@ void DeConv3DLayer::forward(PassType passType) {
       }
       colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(),
                        numFilters_,
-                       outputD_[i],
-                       outputH_[i],
-                       outputW_[i],
+                       imgSizeD_[i],
+                       imgSizeH_[i],
+                       imgSizeW_[i],
                        filterSizeZ_[i],
                        filterSizeY_[i],
                        filterSize_[i],
@@ -144,9 +144,9 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
         colBuf_->vol2Col(
             getOutputGrad()->getData() + n * getOutputGrad()->getStride(),
             numFilters_,
-            outputD_[i],
-            outputH_[i],
-            outputW_[i],
+            imgSizeD_[i],
+            imgSizeH_[i],
+            imgSizeW_[i],
             filterSizeZ_[i],
             filterSizeY_[i],
             filterSize_[i],

paddle/gserver/layers/Layer.h

@@ -49,6 +49,12 @@ struct LayerState {
 };
 typedef std::shared_ptr<LayerState> LayerStatePtr;
 
+/// Paddle device ID, MKLDNN is -2, CPU is -1
+enum PADDLE_DEVICE_ID {
+  MKLDNN_DEVICE = -2,
+  CPU_DEVICE = -1,
+};
+
 /**
  * @brief Base class for layer.
  * Define necessary variables and functions for every layer.
@@ -59,11 +65,6 @@ protected:
   LayerConfig config_;
   /// whether to use GPU
   bool useGpu_;
-  /// Paddle device ID, MKLDNN is -2, CPU is -1
-  enum PADDLE_DEVICE_ID {
-    MKLDNN_DEVICE = -2,
-    CPU_DEVICE = -1,
-  };
   /// Device Id. MKLDNN is -2, CPU is -1, and GPU is 0, 1, 2 ...
   int deviceId_;
   /// Input layers

paddle/gserver/layers/MKLDNNConvLayer.cpp

+/* Copyright (c) 2017 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 "MKLDNNConvLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/utils/Logging.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_conv, MKLDNNConvLayer);
+
+bool MKLDNNConvLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+  CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet";
+  CHECK_EQ(inputLayers_.size(), parameters_.size());
+  CHECK(config_.shared_biases()) << "Only support shared biases yet";
+
+  oc_ = config_.num_filters();
+  const ConvConfig& conf = config_.inputs(0).conv_conf();
+  ic_ = conf.channels();
+  fw_ = conf.filter_size();
+  fh_ = conf.filter_size_y();
+  pw_ = conf.padding();
+  ph_ = conf.padding_y();
+  dw_ = conf.dilation();
+  dh_ = conf.dilation_y();
+  sw_ = conf.stride();
+  sh_ = conf.stride_y();
+  gp_ = conf.groups();
+  oh_ = conf.output_y();
+  ow_ = conf.output_x();
+  ih_ = conf.img_size_y();
+  iw_ = conf.img_size();
+  caffeMode_ = conf.caffe_mode();
+  CHECK(caffeMode_) << "Only support caffe mode yet";
+  CHECK(dh_ == 1 && dw_ == 1) << "Only support dilation 1 yet";
+  // check group setting
+  CHECK_EQ((oc_ / gp_) * gp_, oc_) << "group is indivisible for oc";
+  CHECK_EQ((ic_ / gp_) * gp_, ic_) << "group is indivisible for ic";
+
+  // create weight
+  size_t height = oc_ / gp_;
+  size_t width = ic_ * fh_ * fw_;
+  CHECK_EQ(parameters_[0]->getSize(), height * width);
+  weight_ =
+      std::unique_ptr<Weight>(new Weight(height, width, parameters_[0], 0));
+
+  // create biases
+  if (biasParameter_.get() != NULL) {
+    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_));
+  }
+  return true;
+}
+
+void MKLDNNConvLayer::convertWeightsFromPaddle() {
+  if (hasInitedWgt_) {
+    return;
+  }
+
+  CHECK(wgtVal_) << "should have been initialized";
+  // the paddle weight format is oihw or goihw
+  auto targetDim = wgtVal_->getDims();
+  auto srcFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw;
+  wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
+  hasInitedWgt_ = true;
+}
+
+void MKLDNNConvLayer::convertWeightsToPaddle() {
+  CHECK(wgtVal_) << "should have been initialized";
+  auto targetDim = wgtVal_->getDims();
+  auto dstFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw;
+  wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
+}
+
+void MKLDNNConvLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
+
+  // cal output sizes
+  // oc can not be changed
+  int fh = (fh_ - 1) * dh_ + 1;
+  int fw = (fw_ - 1) * dw_ + 1;
+  oh = outputSize(ih, fh, ph_, sh_, caffeMode_);
+  ow = outputSize(iw, fw, pw_, sw_, caffeMode_);
+
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc * oh * ow);
+
+  printSizeInfo();
+}
+
+void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  resetFwdPD(fwdPD_);
+
+  resetFwdBuffers(fwdPD_, in, wgt, bias, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNConvLayer::resetBwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  std::shared_ptr<conv_bwdWgt::primitive_desc> bwdWgtPD;
+  std::shared_ptr<conv_bwdData::primitive_desc> bwdDataPD;
+
+  resetBwdWgtPD(bwdWgtPD);
+
+  resetBwdDataPD(bwdDataPD);
+
+  resetBwdBuffers(bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNConvLayer::updateInputData() {
+  cpuInVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
+
+void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void MKLDNNConvLayer::loadConvSettings(memory::dims& wgt,
+                                       memory::dims& bias,
+                                       memory::dims& stride,
+                                       memory::dims& dilation,
+                                       memory::dims& padL,
+                                       memory::dims& padR) {
+  wgt = (gp_ == 1) ? memory::dims{oc_, ic_, fh_, fw_}
+                   : memory::dims{gp_, oc_ / gp_, ic_ / gp_, fh_, fw_};
+  bias = memory::dims{oc_};
+  stride = memory::dims{sh_, sw_};
+  padL = memory::dims{ph_, pw_};
+  padR = getPaddingR();
+  // note: mkldnn dilation start from 0
+  dilation = memory::dims{dh_ - 1, dw_ - 1};
+}
+
+void MKLDNNConvLayer::resetFwdPD(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd) {
+  // dims for conv
+  memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+
+  prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
+                                        : prop_kind::forward_training;
+  algorithm algo = algorithm::convolution_direct;
+  padding_kind padKind = padding_kind::zero;
+  conv_fwd::desc fwdDesc =
+      biases_ && biases_->getW()
+          ? conv_fwd::desc(pk,
+                           algo,
+                           MKLDNNMatrix::createMemoryDesc(inDims),
+                           MKLDNNMatrix::createMemoryDesc(wgtDims),
+                           MKLDNNMatrix::createMemoryDesc(biasDims),
+                           MKLDNNMatrix::createMemoryDesc(outDims),
+                           strides,
+                           dilations,
+                           padL,
+                           padR,
+                           padKind)
+          : conv_fwd::desc(pk,
+                           algo,
+                           MKLDNNMatrix::createMemoryDesc(inDims),
+                           MKLDNNMatrix::createMemoryDesc(wgtDims),
+                           MKLDNNMatrix::createMemoryDesc(outDims),
+                           strides,
+                           dilations,
+                           padL,
+                           padR,
+                           padKind);
+  pd.reset(new conv_fwd::primitive_desc(fwdDesc, engine_));
+}
+
+void MKLDNNConvLayer::resetFwdBuffers(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(pd);
+  resetInValue(pd, in);
+
+  resetWgtBiasValue(pd, wgt, bias);
+
+  resetOutValue(pd, out);
+}
+
+void MKLDNNConvLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (cvtInVal_) {
+    pipeline.push_back(*cvtInVal_);
+  }
+
+  if (bias) {
+    fwd_.reset(new conv_fwd(*pd, *in, *wgt, *bias, *out));
+  } else {
+    fwd_.reset(new conv_fwd(*pd, *in, *wgt, *out));
+  }
+  pipeline.push_back(*fwd_);
+
+  if (cvtOutVal_) {
+    pipeline.push_back(*cvtOutVal_);
+  }
+}
+
+void MKLDNNConvLayer::resetInValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& in) {
+  const MatrixPtr& inMat = inputLayers_[0]->getOutput().value;
+  in = MKLDNNMatrix::create(inMat, pd->src_primitive_desc());
+
+  // create buffer and reorder if input value do not match
+  cpuInVal_ = nullptr;
+  cvtInVal_ = nullptr;
+  if (inputIsOnlyMKLDNN()) {
+    MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
+    CHECK(dnnIn) << "Input should be MKLDNNMatrix";
+    if (dnnIn->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      CHECK_EQ(dnnIn->getFormat(), format::nc);
+      CHECK(ih_ == 1 && iw_ == 1) << "when input is nc format";
+      // create a new one with nchw format and same data
+      memory::dims inDims = memory::dims{bs_, ic_, 1, 1};
+      dnnIn = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
+      CHECK(dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc());
+    }
+    in = dnnIn;
+  } else {
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
+    memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+    cpuInVal_ = MKLDNNMatrix::create(cpuIn, inDims, format::nchw, engine_);
+    if (cpuInVal_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      // create new mkldnn matrix
+      in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
+      cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in);
+      CHECK(cvtInVal_) << "should not be emptry";
+    } else {
+      in = cpuInVal_;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtBiasValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias) {
+  wgt = MKLDNNMatrix::create(weight_->getW(), pd->weights_primitive_desc());
+  VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
+
+  bias = nullptr;
+  if (biases_ && biases_->getW()) {
+    bias = MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc());
+  }
+}
+
+void MKLDNNConvLayer::resetOutValue(
+    std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& out) {
+  out = MKLDNNMatrix::create(output_.value, pd->dst_primitive_desc());
+
+  // change original output value from cpu matrix to mkldnn matrix
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutVal_ = nullptr;
+  cpuOutVal_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
+    memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
+    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
+      CHECK(cvtOutVal_) << "should not be emptry";
+    } else {
+      // CPU output share the same data of MKLDNN output
+      cpuOut->setData(out->getData());
+      cpuOutVal_ = out;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetBwdWgtPD(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& pd) {
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+
+  // create backward weight using input, output and weight value memory desc
+  CHECK(inVal_) << "Should have input value";
+  CHECK(outVal_) << "Should have output value";
+  CHECK(wgtVal_) << "Should have weight value";
+  algorithm algo = algorithm::convolution_direct;
+  padding_kind padKind = padding_kind::zero;
+  auto bwdWgtDesc = biasVal_ != nullptr
+                        ? conv_bwdWgt::desc(algo,
+                                            inVal_->getMemoryDesc(),
+                                            wgtVal_->getMemoryDesc(),
+                                            biasVal_->getMemoryDesc(),
+                                            outVal_->getMemoryDesc(),
+                                            strides,
+                                            padL,
+                                            padR,
+                                            padKind)
+                        : conv_bwdWgt::desc(algo,
+                                            inVal_->getMemoryDesc(),
+                                            wgtVal_->getMemoryDesc(),
+                                            outVal_->getMemoryDesc(),
+                                            strides,
+                                            padL,
+                                            padR,
+                                            padKind);
+  pd.reset(new conv_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
+  CHECK(pd->src_primitive_desc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of in value should equal";
+  CHECK(pd->diff_dst_primitive_desc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad should equal the out value";
+  CHECK(pd->diff_weights_primitive_desc() == wgtVal_->getPrimitiveDesc())
+      << "primitive desc of weight grad should equal the weight value";
+}
+
+void MKLDNNConvLayer::resetBwdDataPD(
+    std::shared_ptr<conv_bwdData::primitive_desc>& pd) {
+  if (inputLayers_[0]->getOutput().grad == nullptr) {
+    return;
+  }
+
+  memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
+  loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
+  CHECK(inVal_) << "Should have input value";
+  CHECK(outVal_) << "Should have output value";
+  // create backward data using input and output value memory desc
+  // but using weight memory desc with any format
+  auto bwdDataDesc = conv_bwdData::desc(algorithm::convolution_direct,
+                                        inVal_->getMemoryDesc(),
+                                        MKLDNNMatrix::createMemoryDesc(wgtDims),
+                                        outVal_->getMemoryDesc(),
+                                        strides,
+                                        padL,
+                                        padR,
+                                        padding_kind::zero);
+  pd.reset(new conv_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
+  CHECK(pd->diff_src_primitive_desc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of in grad should equal the in value";
+  CHECK(pd->diff_dst_primitive_desc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad should equal";
+}
+
+void MKLDNNConvLayer::resetBwdBuffers(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(wgtPD);
+  resetOutGrad(wgtPD, out);
+
+  resetWgtBiasGrad(wgtPD, wgt, bias);
+
+  resetInGrad(dataPD, in);
+
+  resetWgtValBwdData(dataPD, wgtValBwdData_);
+}
+
+void MKLDNNConvLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (cvtOutGrad_) {
+    pipeline.push_back(*cvtOutGrad_);
+  }
+
+  // add bwdWgt handle
+  if (bias) {
+    bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt, *bias));
+  } else {
+    bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt));
+  }
+  pipeline.push_back(*bwdWgt_);
+
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  if (cvtWgtVal_) {
+    pipeline.push_back(*cvtWgtVal_);
+  }
+
+  // add bwdData handle
+  CHECK(wgtValBwdData_) << "Should have weight memory";
+  bwdData_.reset(new conv_bwdData(*dataPD, *out, *wgtValBwdData_, *in));
+  pipeline.push_back(*bwdData_);
+
+  if (cvtInGrad_) {
+    pipeline.push_back(*cvtInGrad_);
+  }
+}
+
+void MKLDNNConvLayer::resetOutGrad(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD, MKLDNNMatrixPtr& out) {
+  const MatrixPtr& outMat = output_.grad;
+  out = MKLDNNMatrix::create(outMat, wgtPD->diff_dst_primitive_desc());
+  CHECK(outVal_ != nullptr &&
+        out->getPrimitiveDesc() == outVal_->getPrimitiveDesc())
+      << "primitive desc of out grad and value should be equal";
+
+  // TODO(TJ): merge outgrad
+  // create reorder if has output grad does not match
+  cpuOutGrad_ = nullptr;
+  cvtOutGrad_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
+    // same PrimitiveDesc with cpuInVal_
+    CHECK(cpuOutVal_);
+    cpuOutGrad_ = MKLDNNMatrix::create(cpuOut, cpuOutVal_->getPrimitiveDesc());
+    if (cpuOutGrad_->getPrimitiveDesc() == out->getPrimitiveDesc()) {
+      outMat->setData(cpuOut->getData());
+      out = cpuOutGrad_;
+    } else {
+      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
+      CHECK(cvtOutGrad_);
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtBiasGrad(
+    std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias) {
+  wgt = MKLDNNMatrix::create(weight_->getWGrad(),
+                             wgtPD->diff_weights_primitive_desc());
+  CHECK(nullptr != wgtVal_ &&
+        wgt->getPrimitiveDesc() == wgtVal_->getPrimitiveDesc())
+      << "primitive desc of weight grad and value should be equal";
+  VLOG(MKLDNN_FMTS) << "weight grad format: " << wgt->getFormat();
+
+  if (biasVal_ == nullptr) {
+    return;
+  }
+  bias = MKLDNNMatrix::create(biases_->getWGrad(),
+                              wgtPD->diff_bias_primitive_desc());
+  CHECK(bias->getPrimitiveDesc() == biasVal_->getPrimitiveDesc())
+      << "primitive desc of bias grad should equal the bias value";
+}
+
+void MKLDNNConvLayer::resetInGrad(
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& in) {
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  in = MKLDNNMatrix::create(inputLayers_[0]->getOutput().grad,
+                            dataPD->diff_src_primitive_desc());
+  CHECK(nullptr != inVal_ &&
+        in->getPrimitiveDesc() == inVal_->getPrimitiveDesc())
+      << "primitive desc of input grad and value should be equal";
+
+  // create reorder if has output grad does not match
+  cpuInGrad_ = nullptr;
+  cvtInGrad_ = nullptr;
+  if (!inputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuIn = getInputGrad(0, CPU_DEVICE);
+    // same PrimitiveDesc with cpuInVal_
+    CHECK(cpuInVal_);
+    cpuInGrad_ = MKLDNNMatrix::create(cpuIn, cpuInVal_->getPrimitiveDesc());
+    if (cpuInGrad_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
+      const MatrixPtr& dnnIn = getInputGrad(0, MKLDNN_DEVICE);
+      in = MKLDNNMatrix::create(dnnIn, in->getPrimitiveDesc());
+      cvtInGrad_ = MKLDNNMatrix::createReorder(in, cpuInGrad_);
+      CHECK(cvtInGrad_);
+    } else {
+      in = cpuInGrad_;
+    }
+  }
+}
+
+void MKLDNNConvLayer::resetWgtValBwdData(
+    std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+    MKLDNNMatrixPtr& wgt) {
+  if (dataPD == nullptr) {
+    return;
+  }
+
+  // create new weight value for backward data, and create reorder if necessary
+  // since the primitive_desc would be different with wgtVal_
+  CHECK(wgtVal_) << "should have weight value";
+  if (dataPD->weights_primitive_desc() != wgtVal_->getPrimitiveDesc()) {
+    wgtValBwdData_ =
+        MKLDNNMatrix::create(nullptr, dataPD->weights_primitive_desc());
+    cvtWgtVal_ = MKLDNNMatrix::createReorder(wgtVal_, wgtValBwdData_);
+    CHECK(cvtWgtVal_);
+  } else {
+    wgtValBwdData_ = wgtVal_;
+  }
+  VLOG(MKLDNN_FMTS) << "weight value format for backward data"
+                    << wgtValBwdData_->getFormat();
+}
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNConvLayer.h

+/* Copyright (c) 2017 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 "MKLDNNLayer.h"
+#include "mkldnn.hpp"
+
+namespace paddle {
+typedef mkldnn::convolution_forward conv_fwd;
+typedef mkldnn::convolution_backward_weights conv_bwdWgt;
+typedef mkldnn::convolution_backward_data conv_bwdData;
+
+/**
+ * @brief A subclass of MKLDNNLayer conv layer.
+ *
+ * The config file api is mkldnn_conv
+ */
+class MKLDNNConvLayer : public MKLDNNLayer {
+protected:
+  // padding height and width
+  int ph_, pw_;
+  // stride height and width
+  int sh_, sw_;
+  // dilation height and width
+  int dh_, dw_;
+  // filter(kenerl) height and width
+  int fh_, fw_;
+  // group number
+  int gp_;
+
+  // in resetBwdData, the format of wgtValBwdData_ is different with wgtVal_
+  MKLDNNMatrixPtr wgtValBwdData_;
+  // convert handle from wgtVal_ to wgtValBwdData_
+  std::shared_ptr<mkldnn::reorder> cvtWgtVal_;
+
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<conv_fwd::primitive_desc> fwdPD_;
+
+  // MKLDNNMatrixPtr which should be created from CPU Device
+  MKLDNNMatrixPtr cpuInVal_;
+  MKLDNNMatrixPtr cpuInGrad_;
+  MKLDNNMatrixPtr cpuOutVal_;
+  MKLDNNMatrixPtr cpuOutGrad_;
+  // convert handle between CPU device and MKLDNN device
+  std::shared_ptr<mkldnn::reorder> cvtInVal_;
+  std::shared_ptr<mkldnn::reorder> cvtInGrad_;
+  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
+  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
+
+  // whether the weight has been init
+  bool hasInitedWgt_;
+
+  // true by default, which impact the calculation of output image size.
+  // details can refer to mathUtil.h
+  bool caffeMode_;
+
+  // weight and bias
+  std::unique_ptr<Weight> weight_;
+  std::unique_ptr<Weight> biases_;
+
+public:
+  explicit MKLDNNConvLayer(const LayerConfig& config)
+      : MKLDNNLayer(config), hasInitedWgt_(false), caffeMode_(true) {}
+
+  ~MKLDNNConvLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
+
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void updateInputData() override;
+
+  void updateWeights(const UpdateCallback& callback) override;
+
+  void convertWeightsFromPaddle() override;
+
+  void convertWeightsToPaddle() override;
+
+  void printSizeInfo() override {
+    MKLDNNLayer::printSizeInfo();
+    VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
+                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
+  }
+
+  void printValueFormatFlow() override {
+    if (cpuInVal_) {
+      VLOG(MKLDNN_FMTS) << cpuInVal_->getFormat() << " >>>";
+    }
+    MKLDNNLayer::printValueFormatFlow();
+    if (cpuOutVal_) {
+      VLOG(MKLDNN_FMTS) << " >>> " << cpuOutVal_->getFormat();
+    }
+  }
+
+  void printGradFormatFlow() override {
+    if (cpuInGrad_) {
+      VLOG(MKLDNN_FMTS) << cpuInGrad_->getFormat() << " <<<";
+    }
+    MKLDNNLayer::printGradFormatFlow();
+    if (cpuOutGrad_) {
+      VLOG(MKLDNN_FMTS) << " <<< " << cpuOutGrad_->getFormat();
+    }
+  }
+
+protected:
+  /**
+   * load the dims settings of this conv
+   */
+  void loadConvSettings(mkldnn::memory::dims& wgt,
+                        mkldnn::memory::dims& bias,
+                        mkldnn::memory::dims& stride,
+                        mkldnn::memory::dims& dilation,
+                        mkldnn::memory::dims& padL,
+                        mkldnn::memory::dims& padR);
+
+  /**
+   * reset the forward primitive descriptor.
+   */
+  void resetFwdPD(std::shared_ptr<conv_fwd::primitive_desc>& pd);
+  /**
+   * reset the MKLDNNMatrix buffers used in forward.
+   */
+  void resetFwdBuffers(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                       MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  /**
+   * reset the forward pipeline.
+   */
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * reset MKLDNNMatrix of input value
+   */
+  void resetInValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                    MKLDNNMatrixPtr& in);
+  /**
+   * reset MKLDNNMatrix of weight and bias value
+   */
+  void resetWgtBiasValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                         MKLDNNMatrixPtr& wgt,
+                         MKLDNNMatrixPtr& bias);
+  /**
+   * reset MKLDNNMatrix of output value
+   */
+  void resetOutValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
+                     MKLDNNMatrixPtr& out);
+
+  /**
+   * reset the backward weight primitive descriptor.
+   */
+  void resetBwdWgtPD(std::shared_ptr<conv_bwdWgt::primitive_desc>& pd);
+  /**
+   * reset the backward data primitive descriptor.
+   */
+  void resetBwdDataPD(std::shared_ptr<conv_bwdData::primitive_desc>& pd);
+  /**
+   * reset the MKLDNNMatrix buffers used in backward.
+   */
+  void resetBwdBuffers(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                       std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                       MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  /**
+   * reset the backward pipeline.
+   */
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                        std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * reset MKLDNNMatrix of output grad
+   */
+  void resetOutGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                    MKLDNNMatrixPtr& out);
+  /**
+   * reset MKLDNNMatrix of weight and bias grad
+   */
+  void resetWgtBiasGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias);
+  /**
+   * reset MKLDNNMatrix of input grad
+   */
+  void resetInGrad(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                   MKLDNNMatrixPtr& in);
+  /**
+   * reset MKLDNNMatrix of weight value for backward data
+   * since the primitive_desc would be different with wgtVal_
+   */
+  void resetWgtValBwdData(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
+                          MKLDNNMatrixPtr& wgt);
+
+  /**
+   * get padding_r according to
+   * https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+   * test_convolution_forward_common.hpp
+   * @note: mkldnn dilation start from 0 while paddle start from 1
+   */
+  mkldnn::memory::dims getPaddingR() const {
+    mkldnn::memory::dims padR = {ph_, pw_};
+    for (int i = 0; i < 2; ++i) {
+      if ((ih_ - ((fh_ - 1) * dh_ + 1) + ph_ + padR[0]) / sh_ + 1 != oh_) {
+        ++padR[0];
+      }
+      if ((iw_ - ((fw_ - 1) * dw_ + 1) + pw_ + padR[1]) / sw_ + 1 != ow_) {
+        ++padR[1];
+      }
+    }
+    return padR;
+  }
+};
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "MKLDNNFcLayer.h"
 #include "paddle/utils/Logging.h"
-#include "paddle/utils/Stat.h"
 
 using namespace mkldnn;  // NOLINT
 typedef memory::format format;
@@ -40,6 +39,8 @@ bool MKLDNNFcLayer::init(const LayerMap& layerMap,
   oc_ = getSize();
   oh_ = 1;
   ow_ = 1;
+  ih_ = 1;
+  iw_ = 1;
 
   // input size can not change in FC
   iLayerSize_ = inputLayers_[0]->getSize();
@@ -77,111 +78,86 @@ void MKLDNNFcLayer::convertWeightsToPaddle() {
   wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
 }
 
-void MKLDNNFcLayer::convertOutputToOtherDevice() {
-  copyOutputInfoToOtherDevice();
-  // find other cpu device and reorder output to cpu device
-  int cnt = 0;
-  for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-    if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
-      // fc cpu output value do not need convert
-      // just share point
-      outputOtherDevice_[i].value = output_.value;
-      ++cnt;
-    }
-  }
+void MKLDNNFcLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
 
-  if (cnt > 1) {
-    LOG(WARNING) << "should not have more than one CPU devie";
-  }
-}
-
-void MKLDNNFcLayer::reshape() {
-  const Argument& input = getInput(0, getPrev(0)->getDeviceId());
-  int batchSize = input.getBatchSize();
-  if (bs_ == batchSize) {
-    return;
-  }
-  bs_ = batchSize;
-  ih_ = input.getFrameHeight();
-  iw_ = input.getFrameWidth();
-  if (ih_ == 0) {
-    ih_ = 1;
-  }
-  if (iw_ == 0) {
-    iw_ = 1;
-  }
   CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
-  ic_ = iLayerSize_ / (ih_ * iw_);
-  CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
-  CHECK_EQ(size_t(oc_), getSize());
-  printSizeInfo();
+  ic = iLayerSize_ / (ih * iw);
+  CHECK_EQ(size_t(ic * ih * iw), iLayerSize_) << "not divisible";
+  CHECK_EQ(size_t(oc), getSize());
 
-  // reset output
-  output_.setFrameHeight(oh_);
-  output_.setFrameWidth(ow_);
-  resetOutput(bs_, oc_);
+  reshapeOutput(oh, ow);
+  resizeOutput(bs, oc);
 
-  // reset mkldnn forward
-  resetFwd();
-  needResetBwd_ = true;
-
-  convertWeightsFromPaddle();
+  printSizeInfo();
 }
 
-void MKLDNNFcLayer::resetFwd() {
+void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  pipeline.clear();
   bool hasBias = biases_ && biases_->getW();
-  const MatrixPtr& wgt = weight_->getW();
-  const MatrixPtr& bias = hasBias ? biases_->getW() : nullptr;
-  const MatrixPtr& out = output_.value;
+  const MatrixPtr& wgtVal = weight_->getW();
+  const MatrixPtr& biasVal = hasBias ? biases_->getW() : nullptr;
+  const MatrixPtr& outVal = output_.value;
 
   if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& in = getInputValue(0);
-    inVal_ = std::dynamic_pointer_cast<MKLDNNMatrix>(in);
-    CHECK(inVal_) << "Input should be MKLDNNMatrix";
+    const MatrixPtr& inVal = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(inVal);
+    CHECK(in) << "Input should be MKLDNNMatrix";
   } else {
     CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& in = getInputValue(0, CPU_DEVICE);
-    inVal_ = MKLDNNMatrix::create(
-        in, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
-  }
-  inVal_->downSpatial();
-  wgtVal_ = MKLDNNMatrix::create(
-      wgt, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
-  wgtVal_->downSpatial();
-  biasVal_ =
-      hasBias ? MKLDNNMatrix::create(bias, {oc_}, format::x, engine_) : nullptr;
-  outVal_ = MKLDNNMatrix::create(out, {bs_, oc_}, format::nc, engine_);
+    const MatrixPtr& inVal = getInputValue(0, CPU_DEVICE);
+    in = MKLDNNMatrix::create(
+        inVal, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+  }
+  in->downSpatial();
+  wgt = MKLDNNMatrix::create(
+      wgtVal, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
+  wgt->downSpatial();
+  bias = hasBias ? MKLDNNMatrix::create(biasVal, {oc_}, format::x, engine_)
+                 : nullptr;
+  out = MKLDNNMatrix::create(outVal, {bs_, oc_}, format::nc, engine_);
 
   // change original output value to mkldnn output value
-  output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+  output_.value = std::dynamic_pointer_cast<Matrix>(out);
   if (!outputIsOnlyMKLDNN()) {
-    convertOutputToOtherDevice();
+    // fc cpu output value do not need create convert
+    // just share point
+    getOutput(CPU_DEVICE).value->setData(output_.value->getData());
   }
 
   // create forward handle
   prop_kind pk = prop_kind::forward;
   fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
-                                                inVal_->getMemoryDesc(),
-                                                wgtVal_->getMemoryDesc(),
-                                                biasVal_->getMemoryDesc(),
-                                                outVal_->getMemoryDesc())
+                                                in->getMemoryDesc(),
+                                                wgt->getMemoryDesc(),
+                                                bias->getMemoryDesc(),
+                                                out->getMemoryDesc())
                                  : fc_fwd::desc(pk,
-                                                inVal_->getMemoryDesc(),
-                                                wgtVal_->getMemoryDesc(),
-                                                outVal_->getMemoryDesc());
+                                                in->getMemoryDesc(),
+                                                wgt->getMemoryDesc(),
+                                                out->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
   if (hasBias) {
-    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_));
+    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *bias, *out));
   } else {
-    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_));
+    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *out));
   }
   printValueFormatFlow();
 
-  pipelineFwd_.clear();
-  pipelineFwd_.push_back(*fwd_);
+  pipeline.push_back(*fwd_);
 }
 
-void MKLDNNFcLayer::resetBwd() {
+void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  pipeline.clear();
   if (!needResetBwd_) {
     return;
   }
@@ -190,8 +166,8 @@ void MKLDNNFcLayer::resetBwd() {
 
   /// backward weight
   CHECK(inVal_) << "Should have input value";
-  const MatrixPtr& wgt = weight_->getWGrad();
-  const MatrixPtr& bias = hasBias ? biases_->getWGrad() : nullptr;
+  const MatrixPtr& wgtGrad = weight_->getWGrad();
+  const MatrixPtr& biasGrad = hasBias ? biases_->getWGrad() : nullptr;
 
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
@@ -202,101 +178,66 @@ void MKLDNNFcLayer::resetBwd() {
   // for CPU device:
   // fc do not need to convert from cpu device since output is always nc format
   // only need create from cpu device
-  const MatrixPtr& out = getOutput(device).grad;
-  outGrad_ = MKLDNNMatrix::create(out, outVal_->getPrimitiveDesc());
-  wgtGrad_ = MKLDNNMatrix::create(wgt, wgtVal_->getPrimitiveDesc());
-  biasGrad_ = hasBias ? MKLDNNMatrix::create(bias, biasVal_->getPrimitiveDesc())
+  const MatrixPtr& outGrad = getOutput(device).grad;
+  out = MKLDNNMatrix::create(outGrad, outVal_->getPrimitiveDesc());
+  wgt = MKLDNNMatrix::create(wgtGrad, wgtVal_->getPrimitiveDesc());
+  bias = hasBias ? MKLDNNMatrix::create(biasGrad, biasVal_->getPrimitiveDesc())
                  : nullptr;
 
   // create memory primitive desc
   fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
                                       inVal_->getMemoryDesc(),
-                                      wgtGrad_->getMemoryDesc(),
-                                      outGrad_->getMemoryDesc());
+                                      wgt->getMemoryDesc(),
+                                      out->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
   fc_bwdWgt::desc bwdWgtDesc = hasBias
                                    ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgtGrad_->getMemoryDesc(),
-                                                     biasGrad_->getMemoryDesc(),
-                                                     outGrad_->getMemoryDesc())
+                                                     wgt->getMemoryDesc(),
+                                                     bias->getMemoryDesc(),
+                                                     out->getMemoryDesc())
                                    : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgtGrad_->getMemoryDesc(),
-                                                     outGrad_->getMemoryDesc());
+                                                     wgt->getMemoryDesc(),
+                                                     out->getMemoryDesc());
   fc_bwdWgt::primitive_desc bwdWgtPD =
       fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
 
   if (hasBias) {
-    bwdWgt_.reset(
-        new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_));
+    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt, *bias));
   } else {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_));
+    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt));
   }
-  pipelineBwd_.clear();
-  pipelineBwd_.push_back(*bwdWgt_);
+  pipeline.push_back(*bwdWgt_);
 
   /// backward data
-  device = inputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
-  const MatrixPtr& in = getInputGrad(0, device);
-  if (in == nullptr) {
+  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inGrad == nullptr) {
     return;
   }
-  if (getInput(0, device).getAllCount() > 1) {
-    // TODO(TJ): use outputMaps_ ways when merge outgrad done
+  if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) {
+    // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
   } else {
-    inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc());
+    in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
   }
 
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(inVal_->getMemoryDesc(),
-                                                  wgtGrad_->getMemoryDesc(),
-                                                  outGrad_->getMemoryDesc());
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
+      inVal_->getMemoryDesc(), wgt->getMemoryDesc(), out->getMemoryDesc());
   fc_bwdData::primitive_desc bwdDataPD =
       fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
 
   CHECK(wgtVal_) << "Should have weight memory";
-  bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_));
+  bwdData_.reset(new fc_bwdData(bwdDataPD, *out, *wgtVal_, *in));
   printGradFormatFlow();
-  pipelineBwd_.push_back(*bwdData_);
+  pipeline.push_back(*bwdData_);
 }
 
-void MKLDNNFcLayer::forward(PassType passType) {
-  Layer::forward(passType);
-  reshape();
-
-  {
-    REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
-    syncInputValue();
-
-    // just submit forward pipeline
-    stream_->submit(pipelineFwd_);
-  }
-
-  /* activation */ {
-    REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
-    forwardActivation();
-  }
+void MKLDNNFcLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
 }
 
-void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
-  /* Do derivation */ {
-    REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
-    backwardActivation();
-  }
-
-  {
-    REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
-    resetBwd();
-
-    syncOutputGrad();
-    // just sumbmit backward pipeline
-    stream_->submit(pipelineBwd_);
-  }
-
-  {
-    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
   weight_->getParameterPtr()->incUpdate(callback);
   if (biases_ && biases_->getWGrad()) {
     biases_->getParameterPtr()->incUpdate(callback);
   }
-  }
 }
 }  // namespace paddle

paddle/gserver/layers/MKLDNNFcLayer.h

@@ -45,35 +45,28 @@ public:
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
 
-  void convertWeightsFromPaddle() override;
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
 
-  void convertWeightsToPaddle() override;
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
 
-  void forward(PassType passType) override;
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
 
-  void backward(const UpdateCallback& callback) override;
+  void updateInputData() override;
 
-protected:
-  /**
-   * reshape the input image sizes
-   * and reset output buffer size
-   * and reset mkldnn forward
-   */
-  void reshape();
+  void updateWeights(const UpdateCallback& callback) override;
 
-  /**
-   * reset the forward primitve and memory
-   * only would be called when input size changes
-   */
-  void resetFwd();
-
-  /**
-   * reset the backward primitve and memory for mkldnn fc
-   * only would be called when needed
-   */
-  void resetBwd();
+  void convertWeightsFromPaddle() override;
 
-  void convertOutputToOtherDevice() override;
+  void convertWeightsToPaddle() override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/MKLDNNLayer.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include "MKLDNNBase.h"
 #include "mkldnn.hpp"
 #include "paddle/math/MKLDNNMatrix.h"
+#include "paddle/utils/Stat.h"
 
 DECLARE_bool(use_mkldnn);
 
@@ -33,6 +34,8 @@ typedef std::shared_ptr<MKLDNNLayer> MKLDNNLayerPtr;
  */
 class MKLDNNLayer : public Layer {
 protected:
+  // input value element count
+  size_t inputElemenCnt_;
   // batch size
   int bs_;
   // input image channel, height and width
@@ -52,7 +55,7 @@ protected:
   std::vector<mkldnn::primitive> pipelineFwd_;
   std::vector<mkldnn::primitive> pipelineBwd_;
 
-  // MKLDNNMatrixPtr
+  // MKLDNNMatrixPtr with internal format
   MKLDNNMatrixPtr inVal_;
   MKLDNNMatrixPtr inGrad_;
   MKLDNNMatrixPtr outVal_;
@@ -65,6 +68,7 @@ protected:
 public:
   explicit MKLDNNLayer(const LayerConfig& config)
       : Layer(config),
+        inputElemenCnt_(0),
         bs_(0),
         ic_(0),
         ih_(0),
@@ -95,12 +99,104 @@ public:
     if (!Layer::init(layerMap, parameterMap)) {
       return false;
     }
+    checkCPUOutputsNumber();
 
     stream_.reset(new MKLDNNStream());
     engine_ = CPUEngine::Instance().getEngine();
     return true;
   }
 
+  void forward(PassType passType) override {
+    passType_ = passType;
+
+    {
+      REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
+      CHECK(!inputLayers_.empty());
+      copySeqInfoToOutputs();
+      size_t elemenCnt = inputLayers_[0]->getOutput().value->getElementCnt();
+      if (inputElemenCnt_ != elemenCnt) {
+        // reset when input total sizes changed, not only the batchsize
+        inputElemenCnt_ = elemenCnt;
+        reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
+        resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
+        convertWeightsFromPaddle();
+        needResetBwd_ = true;
+      }
+
+      if (inputLayers_[0]->getType() == "data") {
+        updateInputData();
+      }
+
+      stream_->submit(pipelineFwd_);
+    }
+
+    /* activation */ {
+      REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
+      forwardActivation();
+    }
+  }
+
+  void backward(const UpdateCallback& callback) override {
+    /* Do derivation */ {
+      REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
+      backwardActivation();
+    }
+
+    {
+      REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
+      if (needResetBwd_) {
+        resetBwd(pipelineBwd_, inGrad_, wgtGrad_, biasGrad_, outGrad_);
+        needResetBwd_ = false;
+      }
+
+      stream_->submit(pipelineBwd_);
+    }
+
+    {
+      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+      updateWeights(callback);
+    }
+  }
+
+  /**
+   * reshape the input image sizes
+   * and reset output image and buffer size
+   * output channel can not be changed
+   */
+  virtual void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) = 0;
+
+  /**
+   * reset the mkldnn forward primitve and memory
+   * only would be called when input size changes
+   */
+  virtual void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out) = 0;
+
+  /**
+   * reset the mkldnn backward primitve and memory for mkldnn fc
+   * only would be called when needed
+   */
+  virtual void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out) = 0;
+
+  /**
+   * Update input value data when input layer is "data" type.
+   * Since the input value data address might be changed.
+   */
+  virtual void updateInputData() {}
+
+  /**
+   * Update weights and biases if necessary.
+   */
+  virtual void updateWeights(const UpdateCallback& callback) {}
+
   /**
    * convert weight from paddle format to mkldnn format
    * weight_ will be override
@@ -114,10 +210,38 @@ public:
   virtual void convertWeightsToPaddle() {}
 
   /**
-   * convert MKLDNN output to other device.
-   * only support CPU device yet
+   * add this interface as public for unit test
+   */
+  void addOutputArgument(int deviceId) { Layer::addOutputArgument(deviceId); }
+
+protected:
+  /**
+   * reshape the input image sizes and input batchsize
    */
-  virtual void convertOutputToOtherDevice() {}
+  virtual void reshapeInput(int& batchsize, int& height, int& width) {
+    const Argument& input = inputLayers_[0]->getOutput();
+    batchsize = input.getBatchSize();
+    int h = input.getFrameHeight();
+    int w = input.getFrameWidth();
+    if (h != 0) {
+      height = h;
+    }
+    if (w != 0) {
+      width = w;
+    }
+  }
+
+  /**
+   * reshape output image sizes
+   */
+  virtual void reshapeOutput(size_t height, size_t width) {
+    output_.setFrameHeight(height);
+    output_.setFrameWidth(width);
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].setFrameHeight(height);
+      outputOtherDevice_[i].setFrameWidth(width);
+    }
+  }
 
   /**
    * print info about sizes
@@ -133,8 +257,8 @@ public:
    */
   virtual void printValueFormatFlow() {
     if (inVal_ && outVal_) {
-      VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat()
-                        << " >>> " << outVal_->getFormat();
+      VLOG(MKLDNN_FMTS) << inVal_->getFormat() << " >>> "
+                        << outVal_->getFormat();
     }
   }
 
@@ -143,29 +267,12 @@ public:
    */
   virtual void printGradFormatFlow() {
     if (inGrad_ && outGrad_) {
-      VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat()
-                        << " <<< " << outGrad_->getFormat();
+      VLOG(MKLDNN_FMTS) << inGrad_->getFormat() << " <<< "
+                        << outGrad_->getFormat();
     }
   }
 
 protected:
-  /**
-   * copy image size and sequence info to other device
-   * @note: can not directly use Layer::copyOutputToOtherDevice since here only
-   *        copy base info and do not copy data value
-   */
-  void copyOutputInfoToOtherDevice() {
-    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
-      outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight());
-      outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth());
-      outputOtherDevice_[i].sequenceStartPositions =
-          output_.sequenceStartPositions;
-      outputOtherDevice_[i].subSequenceStartPositions =
-          output_.subSequenceStartPositions;
-      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
-    }
-  }
-
   /**
    * If input only has MKLDNN device.
    * Otherwise, only support the previous layer using CPU device.
@@ -193,37 +300,12 @@ protected:
     return outputOtherDevice_.size() == 0;
   }
 
-  /**
-   * Sync input value data
-   */
-  void syncInputValue() {
-    if (inputIsOnlyMKLDNN()) {
-      return;
-    }
-    real* iData = getInputValue(0, CPU_DEVICE)->getData();
-    // update input data
-    // since it might be changed if this is after data layer
-    inVal_->updateData(iData);
-  }
-
-  /**
-   * Sync output grad data
-   */
-  void syncOutputGrad() {
-    if (outputIsOnlyMKLDNN()) {
-      return;
-    }
-
-    // update diff
-    real* oDiff = getOutput(CPU_DEVICE).grad->getData();
-    outGrad_->updateData(oDiff);
-  }
-
   /**
    * Set deviceId of this layer.
    */
   void setDevice(int id) { deviceId_ = id; }
 
+private:
   /**
    * Set deviceId of the params used in this layer.
    */
@@ -247,6 +329,42 @@ protected:
       parameter->setDevice(id);
     }
   }
+
+  /**
+   * Check the cpu device number of outputOtherDevice_.
+   * should have only one at most.
+   */
+  void checkCPUOutputsNumber(int max = 1) {
+    int cnt = 0;
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      if (outputOtherDevice_[i].deviceId == CPU_DEVICE) {
+        ++cnt;
+      }
+    }
+    CHECK_LE(cnt, max) << "too much CPU devies";
+  }
+
+  /**
+   * copy SeqInfo from input layer to this output and other output devices.
+   * @note: do not use getInput(0) since it used this deviceId_,
+   *        use "inputLayers_[0]->getOutput()" instead.
+   */
+  void copySeqInfoToOutputs() {
+    if (inputLayers_.empty() || !needSequenceInfo_) {
+      return;
+    }
+    const Argument& input = inputLayers_[0]->getOutput();
+    output_.sequenceStartPositions = input.sequenceStartPositions;
+    output_.subSequenceStartPositions = input.subSequenceStartPositions;
+    output_.cpuSequenceDims = input.cpuSequenceDims;
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].sequenceStartPositions =
+          output_.sequenceStartPositions;
+      outputOtherDevice_[i].subSequenceStartPositions =
+          output_.subSequenceStartPositions;
+      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
+    }
+  }
 };
 
 }  // namespace paddle

paddle/gserver/layers/SwitchOrderLayer.cpp

@@ -83,8 +83,7 @@ void SwitchOrderLayer::forward(PassType passType) {
   setOutDims();
   resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]);
   if (heightAxis_.size() > 0) {
-    getOutputValue()->reshape(reshapeHeight_, reshapeWidth_);
-    getOutputGrad()->reshape(reshapeHeight_, reshapeWidth_);
+    resetOutput(reshapeHeight_, reshapeWidth_);
   }
 
   // switch NCHW to NHWC

paddle/gserver/tests/MKLDNNTester.cpp

@@ -63,8 +63,12 @@ void MKLDNNTester::reset(const TestConfig& dnn,
     initTestLayer(
         configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i]));
   }
-  dnnLayer_ = testLayers_[DNN];
   refLayer_ = testLayers_[REF];
+  dnnLayer_ = std::dynamic_pointer_cast<MKLDNNLayer>(testLayers_[DNN]);
+  CHECK(dnnLayer_);
+  // for comparison with Paddle reference results,
+  // need manually add cpu device output for test
+  dnnLayer_->addOutputArgument(CPU_DEVICE);
   EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size());
   EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
 
@@ -109,20 +113,22 @@ void MKLDNNTester::randomBotDatas() {
 
 void MKLDNNTester::randomTopDiffs() {
   refLayer_->getOutputGrad()->randomizeUniform();
-  dnnLayer_->getOutputGrad()->copyFrom(*(refLayer_->getOutputGrad()));
-  VLOG(lvl_) << "Random dom Backward Input, TopDiff: ";
+  dnnLayer_->getOutput(CPU_DEVICE)
+      .grad->copyFrom(*(refLayer_->getOutputGrad()));
+  VLOG(lvl_) << "Random Backward Input, TopDiff: ";
   printMatrix(refLayer_->getOutputGrad());
 }
 
 void MKLDNNTester::checkForward() {
-  printTopDatas();
-  double delta = compareMatrix(testLayers_[DNN]->getOutputValue(),
-                               testLayers_[REF]->getOutputValue());
   VLOG(MKLDNN_ALL) << "Check Forward";
+  printTopDatas();
+  double delta = compareMatrix(dnnLayer_->getOutput(-1).value,
+                               refLayer_->getOutputValue());
   EXPECT_LE(fabs(delta), eps_);
 }
 
 void MKLDNNTester::checkBackwardData() {
+  VLOG(MKLDNN_ALL) << "Check Backward Data";
   // TODO(TJ): uncomment me when batch norm ready
   // const bool isBN = dnnLayer_->getType() == "mkldnn_batch_norm";
   for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) {
@@ -144,14 +150,12 @@ void MKLDNNTester::checkBackwardData() {
 }
 
 void MKLDNNTester::checkBackwardWgts() {
+  VLOG(MKLDNN_ALL) << "Check Backward Weight";
   CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
   vector<VectorPtr> dnnWgts;  // used to temply save mkldnn weights
   saveWgt(parameters_[DNN], dnnWgts);
 
-  const MKLDNNLayerPtr dnnlayer =
-      std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
-  CHECK(dnnlayer);
-  dnnlayer->convertWeightsToPaddle();
+  dnnLayer_->convertWeightsToPaddle();
   for (size_t i = 0; i < parameters_[DNN].size(); ++i) {
     const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
     const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
@@ -189,8 +193,10 @@ void MKLDNNTester::restoreWgt(const vector<VectorPtr>& from,
 }
 
 // clear parameters grad
-void MKLDNNTester::clearWgtDiffs() {
+void MKLDNNTester::clearWgtDiffs(size_t id) {
+  CHECK_LE(id, parameters_.size());
   for (size_t n = 0; n < parameters_.size(); ++n) {
+    if (id == n || id == parameters_.size()) {
       for (size_t i = 0; i < parameters_[n].size(); ++i) {
         const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT);
         if (grad) {
@@ -198,30 +204,28 @@ void MKLDNNTester::clearWgtDiffs() {
         }
       }
     }
+  }
 }
 
-void MKLDNNTester::clearBotDiffs() {
-  // dnn and ref
+void MKLDNNTester::clearBotDiffs(size_t id) {
+  CHECK_LE(id, dataLayers_.size());
   for (size_t n = 0; n < dataLayers_.size(); ++n) {
-    // all inputs layers
+    if (id == n || id == dataLayers_.size()) {
+      // clear inputs layers of this specific layer
       for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
         dataLayers_[n][i]->getOutputGrad()->zeroMem();
       }
     }
-}
-
-void MKLDNNTester::clearBotDiffs(int n) {
-  CHECK_LT(n, NUM);
-  // all inputs layers
-  for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
-    dataLayers_[n][i]->getOutputGrad()->zeroMem();
   }
 }
 
-void MKLDNNTester::clearTopDatas() {
+void MKLDNNTester::clearTopDatas(size_t id) {
+  CHECK_LE(id, testLayers_.size());
   for (size_t i = 0; i < testLayers_.size(); ++i) {
+    if (id == i || id == testLayers_.size()) {
       testLayers_[i]->getOutputValue()->zeroMem();
     }
+  }
 }
 
 void MKLDNNTester::printTopDatas() {
@@ -300,16 +304,24 @@ void MKLDNNTester::runOnce() {
   checkForward();
 
   // test backward
+  // simple updater
+  UpdateCallback updateCallback = [](Parameter* para) {
+    auto& grad = para->getBuf(PARAMETER_GRADIENT);
+    auto& value = para->getBuf(PARAMETER_VALUE);
+    real lr = 1e-3;
+    value->add(*grad, lr);
+  };
   randomTopDiffs();
-  dnnLayer_->backward(nullptr);
-  refLayer_->backward(nullptr);
+  dnnLayer_->backward(updateCallback);
+  refLayer_->backward(updateCallback);
   checkBackwardData();
   checkBackwardWgts();
 
   // clear buffers
   // ref code will addto the diff, dnn code will writeto it
-  // and clearTopDatas() and clearWgtDiffs() should be coverd by test layers
+  // and clearTopDatas(REF) should be coverd by ref layers
   clearBotDiffs(REF);
+  clearWgtDiffs(REF);
 }
 
 void MKLDNNTester::run(const TestConfig& dnn,

paddle/gserver/tests/MKLDNNTester.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <vector>
 #include "LayerGradUtil.h"
 #include "paddle/gserver/layers/MKLDNNBase.h"
+#include "paddle/gserver/layers/MKLDNNLayer.h"
 
 namespace paddle {
 
@@ -40,7 +41,8 @@ protected:
   vector<LayerMap> layerMaps_;
   vector<vector<ParameterPtr>> parameters_;
   vector<LayerPtr> testLayers_;
-  LayerPtr dnnLayer_, refLayer_;
+  LayerPtr refLayer_;
+  MKLDNNLayerPtr dnnLayer_;
 
   /// run some iterations, all the result should pass
   size_t iter_;
@@ -88,10 +90,10 @@ private:
   void checkBackwardData();
   void checkBackwardWgts();
 
-  void clearWgtDiffs();
-  void clearBotDiffs();
-  void clearBotDiffs(int n);  // clear specific layer
-  void clearTopDatas();
+  // clear specific layer, clear all when id equals NUM
+  void clearWgtDiffs(size_t id = NUM);
+  void clearBotDiffs(size_t id = NUM);
+  void clearTopDatas(size_t id = NUM);
 
   void printTopDatas();
   void printMatrix(const MatrixPtr& m);

paddle/gserver/tests/test_LayerGrad.cpp

@@ -2302,26 +2302,27 @@ void test3DDeConvLayer(const string& type, bool trans, bool useGpu) {
   conv->set_stride(2);
   conv->set_stride_y(2);
   conv->set_stride_z(2);
-  conv->set_img_size(IMAGE_SIZE);
-  conv->set_img_size_y(IMAGE_SIZE_Y);
-  conv->set_img_size_z(IMAGE_SIZE_Z);
-  conv->set_output_x(imageSize(conv->img_size(),
+  conv->set_output_x(IMAGE_SIZE);
+  conv->set_output_y(IMAGE_SIZE_Y);
+  conv->set_output_z(IMAGE_SIZE_Z);
+
+  conv->set_img_size(imageSize(conv->output_x(),
                                conv->filter_size(),
                                conv->padding(),
                                conv->stride(),
                                true));
-  conv->set_output_y(imageSize(conv->img_size_y(),
+  conv->set_img_size_y(imageSize(conv->output_y(),
                                  conv->filter_size_y(),
                                  conv->padding_y(),
                                  conv->stride_y(),
                                  true));
-  conv->set_output_z(imageSize(conv->img_size_z(),
+  conv->set_img_size_z(imageSize(conv->output_z(),
                                  conv->filter_size_z(),
                                  conv->padding_z(),
                                  conv->stride_z(),
                                  true));
-  config.layerConfig.set_size(conv->output_x() * conv->output_y() *
-                              conv->output_z() * NUM_FILTERS);
+  config.layerConfig.set_size(conv->img_size() * conv->img_size_y() *
+                              conv->img_size_z() * NUM_FILTERS);
   conv->set_groups(1);
   conv->set_filter_channels(conv->channels() / conv->groups());
   config.inputDefs.push_back(

paddle/gserver/tests/test_MKLDNN.cpp

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <vector>
 #include "MKLDNNTester.h"
 #include "ModelConfig.pb.h"
+#include "paddle/math/MathUtils.h"
 
 using namespace paddle;  // NOLINT
 
@@ -63,6 +64,83 @@ TEST(MKLDNNLayer, FcLayer) {
   testFcLayer({/*bs*/ 15, /*ic*/ 3, /*oc*/ 6, /*ih*/ 16, /*iw*/ 16});
 }
 
+struct testConvDesc {
+  int bs, gp;
+  int ic, ih, iw;
+  int oc, oh, ow;
+  int fh, fw;
+  int ph, pw;
+  int sh, sw;
+  int dh, dw;
+};
+
+void testConvLayer(const testConvDesc& pm) {
+  const std::string compareTypes[] = {"mkldnn_conv", "exconv"};
+  TestConfig cfg;
+  cfg.layerConfig.set_type(compareTypes[0]);
+  cfg.layerConfig.set_num_filters(pm.oc);
+  cfg.layerConfig.set_size(pm.oc * pm.oh * pm.ow);
+  // cfg.layerConfig.set_partial_sum(1); // TODO: check it
+  cfg.layerConfig.set_shared_biases(true);
+  cfg.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw),
+       /* size of weight= */ size_t(pm.oc * pm.ic * pm.fh * pm.fw / pm.gp)});
+  LayerInputConfig* input = cfg.layerConfig.add_inputs();
+  ConvConfig* conv = input->mutable_conv_conf();
+  conv->set_groups(pm.gp);
+  conv->set_img_size(pm.iw);
+  conv->set_img_size_y(pm.ih);
+  conv->set_output_x(pm.ow);
+  conv->set_output_y(pm.oh);
+  conv->set_filter_size(pm.fw);
+  conv->set_filter_size_y(pm.fh);
+  conv->set_channels(pm.ic);
+  conv->set_padding(pm.pw);
+  conv->set_padding_y(pm.ph);
+  conv->set_stride(pm.sw);
+  conv->set_stride_y(pm.sh);
+  conv->set_dilation(pm.dw);
+  conv->set_dilation_y(pm.dh);
+  conv->set_caffe_mode(true);
+  conv->set_filter_channels(conv->channels() / conv->groups());
+  CHECK_EQ(conv->filter_channels() * pm.gp, conv->channels())
+      << "it is indivisible";
+
+  int fh = (pm.fh - 1) * pm.dh + 1;
+  int fw = (pm.fw - 1) * pm.dw + 1;
+  int ow = outputSize(pm.iw, fw, pm.pw, pm.sw, true);
+  int oh = outputSize(pm.ih, fh, pm.ph, pm.sh, true);
+  CHECK_EQ(ow, pm.ow) << "output size check failed";
+  CHECK_EQ(oh, pm.oh) << "output size check failed";
+
+  MKLDNNTester tester;
+  for (auto biasSize : {pm.oc, 0}) {
+    cfg.biasSize = biasSize;
+    TestConfig ref = cfg;
+    ref.layerConfig.set_type(compareTypes[1]);
+    for (auto bs : {pm.bs, 1}) {
+      tester.run(cfg, ref, bs, pm.ih, pm.iw);
+    }
+  }
+}
+
+TEST(MKLDNNLayer, ConvLayer) {
+  /* bs, gp, ic, ih, iw, oc, oh, ow, fh, fw, ph, pw, sh, sw, dh, dw */
+  testConvLayer({2, 1, 3, 32, 32, 16, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({2, 1, 8, 16, 16, 8, 16, 16, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({3, 1, 16, 32, 32, 3, 32, 32, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({8, 1, 16, 18, 18, 32, 18, 18, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({16, 1, 1, 42, 31, 32, 23, 11, 4, 5, 3, 2, 2, 3, 1, 1});
+  testConvLayer({2, 1, 8, 16, 16, 8, 8, 8, 3, 3, 1, 1, 2, 2, 1, 1});
+  testConvLayer({3, 1, 8, 13, 13, 8, 7, 7, 3, 3, 1, 1, 2, 2, 1, 1});
+  // with groups
+  testConvLayer({2, 2, 4, 5, 5, 8, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({2, 3, 3, 5, 5, 3, 5, 5, 3, 3, 1, 1, 1, 1, 1, 1});
+  testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1});
+}
+
 // TODO(TJ): add branch test
 
 int main(int argc, char** argv) {

paddle/math/MKLDNNMatrix.cpp

@@ -33,14 +33,12 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
     size_t width = cnts / dims[0];
     m = Matrix::create(height, width, false, false);
   }
-
   CHECK(m) << " Matrix should not be empty";
+
   CpuMatrixPtr cpuMatrix = std::dynamic_pointer_cast<CpuMatrix>(m);
   CHECK(cpuMatrix) << "Only support create from CPU matrix yet";
-
-  CHECK_EQ(cnts, m->getElementCnt()) << "Count size does not match";
-  return std::make_shared<MKLDNNMatrix>(
-      m->getData(), m->getHeight(), m->getWidth(), pd);
+  CHECK_EQ(cpuMatrix->getElementCnt(), cnts) << "Count size does not match";
+  return std::make_shared<MKLDNNMatrix>(cpuMatrix, pd);
 }
 
 MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
@@ -51,6 +49,27 @@ MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
   return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg));
 }
 
+std::shared_ptr<reorder> MKLDNNMatrix::createReorder(const MKLDNNMatrixPtr& src,
+                                                     const MKLDNNMatrixPtr& dst,
+                                                     bool checkData) {
+  if (src == dst || src->getPrimitiveDesc() == dst->getPrimitiveDesc()) {
+    return nullptr;
+  }
+
+  if (checkData && (src->getData() == dst->getData())) {
+    LOG(FATAL) << "can not create reorder with inplace data";
+    return nullptr;
+  }
+
+  memory::dims srcDims = src->getDims();
+  memory::dims dstDims = dst->getDims();
+  CHECK_EQ(srcDims.size(), dstDims.size());
+  for (size_t i = 0; i < srcDims.size(); ++i) {
+    CHECK_EQ(srcDims[i], dstDims[i]);
+  }
+  return std::make_shared<reorder>(*src, *dst);
+}
+
 void MKLDNNMatrix::reorderDataFrom(const MKLDNNMatrixPtr& m,
                                    memory::format srcFmt,
                                    memory::dims targetDim) {
@@ -138,7 +157,7 @@ void MKLDNNMatrix::downSpatial() {
       mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
       "could not create a memory primitive");
   reset(result);
-  set_data_handle(getData());
+  set_data_handle(data_);
 }
 
 }  // namespace paddle

paddle/math/MKLDNNMatrix.h

@@ -30,11 +30,10 @@ typedef std::shared_ptr<MKLDNNMatrix> MKLDNNMatrixPtr;
  */
 class MKLDNNMatrix : public CpuMatrix, public mkldnn::memory {
 public:
-  MKLDNNMatrix(real* data,
-               size_t height,
-               size_t width,
-               mkldnn::memory::primitive_desc pd)
-      : CpuMatrix(data, height, width, false), mkldnn::memory(pd, data) {}
+  MKLDNNMatrix(CpuMatrixPtr m, mkldnn::memory::primitive_desc pd)
+      : CpuMatrix(m->getData(), m->getHeight(), m->getWidth(), false),
+        mkldnn::memory(pd, m->getData()),
+        m_(m) {}
 
   ~MKLDNNMatrix() {}
 
@@ -53,6 +52,31 @@ public:
       mkldnn::engine& eg,
       mkldnn::memory::data_type dtype = mkldnn::memory::data_type::f32);
 
+  /**
+   * Create Memory descriptor.
+   * default with any format and f32 dtype
+   */
+  static mkldnn::memory::desc createMemoryDesc(
+      const mkldnn::memory::dims& dims,
+      const mkldnn::memory::format& fmt = mkldnn::memory::format::any,
+      const mkldnn::memory::data_type& dtype = mkldnn::memory::data_type::f32) {
+    return mkldnn::memory::desc(dims, dtype, fmt);
+  }
+
+  /**
+   * Create reorder primitive.
+   * Create a mkldnn::reorder handle for converting src MKLDNNMatrix to dst.
+   * checkData: for whether to check the data handle of src and dst is the same.
+   *            if true, means check it and do not want support inplace reorder;
+   *            otherwise do not check data which means the created reorder
+   *            maybe inplace buffer and do not guarantee the logical is correct
+   *            since not all format or conversion support inplace.
+   */
+  static std::shared_ptr<mkldnn::reorder> createReorder(
+      const MKLDNNMatrixPtr& src,
+      const MKLDNNMatrixPtr& dst,
+      bool checkData = true);
+
 public:
   /**
    * Reorder this MKLDNNMatrix from other format.
@@ -81,11 +105,29 @@ public:
   void downSpatial();
 
   /**
-   * Update the memory data handle.
+   * set the memory data handle.
    * Caution: This will not check the buffer size of the data,
    *          it should be coverd by user.
    */
-  void updateData(void* data) { set_data_handle(data); }
+  void setData(real* data) {
+    set_data_handle(data);
+    CpuMatrix::setData(data);
+    m_.reset();
+  }
+
+  /**
+   * override Matrix::getData
+   * check data before return
+   */
+  real* getData() override {
+    CHECK_EQ((void*)data_, get_data_handle());
+    return data_;
+  }
+
+  const real* getData() const override {
+    CHECK_EQ((void*)data_, get_data_handle());
+    return data_;
+  }
 
   /**
    * Get primitive descriptor.
@@ -143,6 +185,10 @@ protected:
                    memory::format srcFmt,
                    memory::format dstFmt,
                    memory::dims dm);
+
+private:
+  // save the CpuMatrixPtr in case the buffer released outside
+  CpuMatrixPtr m_;
 };
 
 }  // namespace paddle

paddle/operators/CMakeLists.txt

 file(GLOB GENERAL_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*_op.cc")
 string(REPLACE ".cc" "" GENERAL_OPS "${GENERAL_OPS}")
+set(pybind_file ${PADDLE_SOURCE_DIR}/paddle/pybind/pybind.h)
+file(WRITE ${pybind_file} "// Generated by the paddle/operator/CMakeLists.txt.  DO NOT EDIT!\n\n")
 function(op_library TARGET)
     # op_library is a function to create op library. The interface is same as
     # cc_library. But it handle split GPU/CPU code and link some common library
@@ -7,10 +9,11 @@ function(op_library TARGET)
     set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE)
     set(cc_srcs)
     set(cu_srcs)
-    set(op_common_deps operator op_registry)
+    set(op_common_deps operator op_registry math_function)
     set(options "")
     set(oneValueArgs "")
     set(multiValueArgs SRCS DEPS)
+    set(pybind_flag 0)
     cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
             "${multiValueArgs}" ${ARGN})
 
@@ -46,22 +49,40 @@ function(op_library TARGET)
         cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
     endif()
+
+    # net_op doesn't need pybind
+    if ("${TARGET}" STREQUAL "net_op")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_NO_KERNEL_OP
+    file(READ ${TARGET}.cc TARGET_CONTENT)
+    string(REGEX MATCH "OperatorWithKernel" regex_result "${TARGET_CONTENT}")
+    string(REPLACE "_op" "" TARGET "${TARGET}")
+    if (${pybind_flag} EQUAL 0 AND regex_result STREQUAL "")
+        file(APPEND ${pybind_file} "USE_NO_KERNEL_OP(${TARGET});\n")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_CPU_ONLY_OP
+    list(LENGTH cu_srcs cu_srcs_len)
+    if (${pybind_flag} EQUAL 0 AND ${cu_srcs_len} EQUAL 0)
+        file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n")
+        set(pybind_flag 1)
+    endif()
+
+    # pybind USE_OP
+    if (${pybind_flag} EQUAL 0)
+        file(APPEND ${pybind_file} "USE_OP(${TARGET});\n")
+    endif()
 endfunction()
 
 add_subdirectory(math)
 
 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)
+    recurrent_op)
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
-  DEPS framework_proto tensor operator net_op)
-op_library(scale_op DEPS net_op)
+  DEPS framework_proto tensor net_op)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})

paddle/operators/accuracy_op.cc

+/* 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/accuracy_op.h"
+
+namespace paddle {
+namespace operators {
+
+class AccuracyOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Inference"),
+                            "Input of Inference must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
+                            "Input of Inference must be initialized.");
+    auto *inference = ctx.Input<framework::Tensor>("Inference");
+    auto *label = ctx.Input<framework::Tensor>("Label");
+
+    PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label must be a vector");
+    PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0],
+                      "inference size must be the same as label size");
+
+    ctx.Output<framework::LoDTensor>("Accuracy")->Resize({1});
+  }
+};
+
+class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  AccuracyOpMaker(framework::OpProto *proto,
+                  framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    // TODO(typhoonzero): support both inference value and indices.
+    AddInput("Inference", "topk(indices) the network output");
+    AddInput("Label", "Label of the training data");
+    // TODO(typhoonzero): AddInput("Weight", ...
+    AddOutput("Accuracy", "The accuracy of current batch");
+
+    AddComment(
+        R"DOC(Accuracy. It will print accuracy rate for classification.
+The accuracy is:
+..  math::
+accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(accuracy, ops::AccuracyOp, ops::AccuracyOpMaker);
+REGISTER_OP_CPU_KERNEL(accuracy,
+                       ops::AccuracyKernel<paddle::platform::CPUPlace, float>);

paddle/operators/accuracy_op.cu

+/* 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/accuracy_op.h"
+
+namespace paddle {
+namespace operators {
+
+__global__ void AccuracySingleKernel(const int N, const int D, const int top_k,
+                                     const int* Xdata, const int* labelData,
+                                     float* accuracy) {
+  int correct = 0;
+  for (int row = 0; row < N; row++) {
+    const int label = labelData[row];
+    for (int col = 0; col < D; col++) {
+      const int pred = Xdata[row * D + col];
+      if (pred == label) {
+        ++correct;
+        break;
+      }
+    }
+  }
+  *accuracy = static_cast<float>(correct) / static_cast<float>(N);
+}
+
+template <typename T>
+class AccuracyOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+    auto* inference = ctx.Input<Tensor>("Inference");
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* accuracy = ctx.Output<Tensor>("Accuracy");
+    // FIXME(typhoonzero): only support indices currently
+    // if add support for output values, how to detect the data type?
+    const int* inference_data = inference->data<int>();
+    const int* label_data = label->data<int>();
+    float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
+
+    size_t num_samples = inference->dims()[0];
+    size_t infer_width = inference->dims()[1];
+    cudaMemset((void**)&accuracy_data, 0, sizeof(float));
+
+    if (num_samples == 0) {
+      return;
+    }
+
+    AccuracySingleKernel<<<1, 1>>>(num_samples, infer_width, 1, inference_data,
+                                   label_data, accuracy_data);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_GPU_KERNEL(accuracy,
+                       paddle::operators::AccuracyOpCUDAKernel<float>);

paddle/operators/accuracy_op.h

+/* 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 <algorithm>
+#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 T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class AccuracyKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* inference = ctx.Input<Tensor>("Inference");
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* accuracy = ctx.Output<Tensor>("Accuracy");
+
+    float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
+
+    const T* inference_data = inference->data<T>();
+    const T* label_data = label->data<T>();
+
+    size_t num_samples = inference->dims()[0];
+    size_t class_dim = inference->dims()[1];
+    *accuracy_data = 0.0f;
+
+    if (num_samples == 0) {
+      return;
+    }
+
+    int num_correct = 0;
+    // assume inference is already the topk of the output
+    for (size_t i = 0; i < num_samples; ++i) {
+      PADDLE_ENFORCE_GE(label_data[i], 0, "label must >= 0");
+      for (size_t j = 0; j < class_dim; ++j) {
+        if (inference_data[i * class_dim + j] == label_data[i]) {
+          ++num_correct;
+          break;
+        }
+      }
+    }
+
+    // FIXME(typhoonzero): we don't accumulate the accuracy for now.
+    *accuracy_data =
+        static_cast<float>(num_correct) / static_cast<float>(num_samples);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/add_op.cc

@@ -26,7 +26,8 @@ class AddOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
                       ctx.Input<Tensor>("Y")->dims(),
                       "Two input of Add Op's dimension must be same.");
-    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        ctx.Input<Tensor>("X")->dims());
   }
 };
 

paddle/operators/concat_op.cc

+/* 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/concat_op.h"
+#include <vector>
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+class ConcatOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
+    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    size_t n = ins.size();
+
+    PADDLE_ENFORCE_GT(n, 1, "Input tensors count should > 1.");
+
+    auto out_dims = ins[0]->dims();
+    size_t in_zero_dims_size = out_dims.size();
+    for (size_t i = 1; i < n; i++) {
+      for (size_t j = 0; j < in_zero_dims_size; j++) {
+        if (j == axis) {
+          out_dims[axis] += ins[i]->dims()[j];
+          continue;
+        }
+        PADDLE_ENFORCE_EQ(out_dims[j], ins[i]->dims()[j],
+                          "Input tensors should have the same "
+                          "elements except the specify axis.")
+      }
+    }
+    out->Resize(out_dims);
+  }
+};
+
+class ConcatOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ConcatOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input tensors of concat operator.").AsDuplicable();
+    AddOutput("Out", "the output tensor of concat operator.");
+    AddComment(R"DOC(
+            Join the input tensors along with the axis.
+            Examples:
+              Input[0] = [[1,2],[3,4]]
+              Input[1] = [[5,6]]
+              axis = 0
+              Output = [[1,2],
+                        [3,4],
+                        [5,6]]
+        )DOC");
+    AddAttr<int>("axis", "The axis which the inputs will be joined with.")
+        .SetDefault(0);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(concat, ops::ConcatOp, ops::ConcatOpMaker)
+REGISTER_OP_CPU_KERNEL(concat,
+                       ops::ConcatKernel<paddle::platform::CPUPlace, float>)

paddle/operators/concat_op.h

+/* 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 <vector>
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class ConcatKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
+    size_t n = ins.size();
+    size_t output_axis_dim = 0;
+    size_t before = 1, after = 1;
+    for (size_t i = 0; i < n; i++) {
+      output_axis_dim += ins[i]->dims()[axis];
+    }
+    auto& input_zero = ins[0];
+    for (int64_t i = 0; i < input_zero->dims().size(); i++) {
+      if (i == axis) {
+        continue;
+      }
+      if (i < axis) {
+        before *= input_zero->dims()[i];
+      } else {
+        after *= input_zero->dims()[i];
+      }
+    }
+    size_t output_offset = 0;
+    for (size_t i = 0; i < n; i++) {
+      auto& in = ins[i];
+      auto axis_dim = in->dims()[axis];
+      for (size_t j = 0; j < before; j++) {
+        size_t len = axis_dim * after * sizeof(T);
+        const T* src = in->data<T>() + axis_dim * after * j;
+        T* out_data = out->mutable_data<T>(platform::CPUPlace());
+        T* dest = out_data + output_offset + output_axis_dim * after * j;
+        memcpy(dest, src, len);
+      }
+      output_offset += axis_dim * after;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/cos_sim_op.cc

@@ -25,16 +25,30 @@ class CosSimOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    // notnull check
     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});
+
+    // shape check
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+
+    PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
+                      "Ranks of Input(X) and Input(Y) must be equal.");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Rank of Input(X) must not be less than 2.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
+                      framework::slice_ddim(y_dims, 1, y_dims.size()),
+                      "All dimensions except the 1st of Input(X) and Input(Y) "
+                      "must be equal.");
+    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
+                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
+                   " just 1 (which will be broadcasted to match Input(X)).");
+
+    // resize tensor
+    ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
+    ctx.Output<framework::LoDTensor>("XNorm")->Resize({x_dims[0], 1});
+    ctx.Output<framework::LoDTensor>("YNorm")->Resize({y_dims[0], 1});
   }
 };
 
@@ -42,16 +56,27 @@ 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.");
+    AddInput("X", "The 1st input of cos_sim op.");
+    AddInput("Y", "The 2nd 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();
+    AddOutput("XNorm",
+              "Norm of the first input, reduced along the 1st "
+              "dimension.")
+        .AsIntermediate();
+    AddOutput("YNorm",
+              "Norm of the second input, reduced along the 1st "
+              "dimension.")
+        .AsIntermediate();
 
     AddComment(R"DOC(
 Cosine Similarity Operator.
 
-The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y))
+The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)).
+
+Input(X) and Input(Y) must have the same shape, except that the 1st dimension
+of Input(Y) could be just 1 (different from Input(X)), which will be
+broadcasted to match the shape of Input(X) before computing their cosine
+similarity.
 )DOC");
   }
 };
@@ -62,34 +87,54 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    // notnull check
     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("Out"),
+                            "Input(Out) must not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                             "Input(Out@GRAD) must not be null.");
 
+    // shape check
     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"));
+    auto out_dims = ctx.Input<Tensor>("Out")->dims();
+    auto out_grad_dims =
+        ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Ranks of Input(X) and Input(Y) must be equal.");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Rank of Input(X) must not be less than 2.");
+    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
+                      framework::slice_ddim(y_dims, 1, y_dims.size()),
+                      "All dimensions except the 1st of Input(X) and Input(Y) "
+                      "must be equal.");
+    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
+                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
+                   " just 1 (which will be broadcasted to match Input(X)).");
+    auto target_xnorm_dims = framework::make_ddim({x_dims[0], 1});
+    auto target_ynorm_dims = framework::make_ddim({y_dims[0], 1});
+    PADDLE_ENFORCE_EQ(xnorm_dims, target_xnorm_dims,
+                      "Shape of Input(XNorm) must be [X.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(ynorm_dims, target_ynorm_dims,
+                      "Shape of Input(YNorm) must be [Y.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(out_dims, target_xnorm_dims,
+                      "Shape of Input(Out) must be [X.Dim(0), 1].");
+    PADDLE_ENFORCE_EQ(out_grad_dims, target_xnorm_dims,
+                      "Shape of Input(Out@Grad) must be [X.Dim(0), 1].");
+
+    // resize tensor
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);
   }

paddle/operators/cos_sim_op.h

@@ -31,30 +31,38 @@ 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");
+    // get Tensor
+    auto* in_x = context.Input<Tensor>("X");
+    auto* in_y = context.Input<Tensor>("Y");
+    auto* out_z = context.Output<Tensor>("Out");
+    auto* out_x_norm = context.Output<Tensor>("XNorm");
+    auto* out_y_norm = context.Output<Tensor>("YNorm");
+    out_z->mutable_data<T>(context.GetPlace());
+    out_x_norm->mutable_data<T>(context.GetPlace());
+    out_y_norm->mutable_data<T>(context.GetPlace());
 
-    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);
+    // convert Tensor to Eigen Tensor
+    int rows_x = in_x->dims()[0];
+    int rows_y = in_y->dims()[0];
+    auto x = EigenMatrix<T>::Reshape(*in_x, 1);
+    auto y = EigenMatrix<T>::Reshape(*in_y, 1);
+    auto z = EigenVector<T>::Flatten(*out_z);
+    auto x_norm = EigenVector<T>::Flatten(*out_x_norm);
+    auto y_norm = EigenVector<T>::Flatten(*out_y_norm);
 
+    // compute
     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();
+    auto row_along = Eigen::array<int, 1>({{1}});
+    x_norm.device(place) = x.square().sum(row_along).sqrt();
+    y_norm.device(place) = y.square().sum(row_along).sqrt();
+    if (rows_x == rows_y) {
+      auto xy = (x * y).sum(Eigen::array<int, 1>({1}));
       z.device(place) = xy / x_norm / y_norm;
+    } else {
+      Eigen::DSizes<int, 2> bcast(rows_x, 1);
+      auto xy = (x * y.broadcast(bcast)).sum(row_along);
+      z.device(place) = xy / x_norm / y_norm.broadcast(bcast);
+    }
   }
 };
 
@@ -62,43 +70,72 @@ 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"));
+    // get Tensor
+    auto* in_x = context.Input<Tensor>("X");
+    auto* in_y = context.Input<Tensor>("Y");
+    auto* in_z = context.Input<Tensor>("Out");
+    auto* in_x_norm = context.Input<Tensor>("XNorm");
+    auto* in_y_norm = context.Input<Tensor>("YNorm");
+    auto* out_grad_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* out_grad_y = context.Output<Tensor>(framework::GradVarName("Y"));
+    auto* in_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);
+    // convert Tensor to Eigen Tensor
+    auto x = EigenMatrix<T>::Reshape(*in_x, 1);
+    auto y = EigenMatrix<T>::Reshape(*in_y, 1);
+    auto z = EigenMatrix<T>::Reshape(*in_z, 1);
+    auto x_norm = EigenMatrix<T>::Reshape(*in_x_norm, 1);
+    auto y_norm = EigenMatrix<T>::Reshape(*in_y_norm, 1);
+    auto dz = EigenMatrix<T>::Reshape(*in_grad_z, 1);
 
-    Eigen::DSizes<int, 2> bcast(1, new_dims[1]);
-    auto z_bcast = z.broadcast(bcast);
-    auto dz_bcast = dz.broadcast(bcast);
+    // compute gradident
+    int rows_x = in_x->dims()[0];
+    int rows_y = in_y->dims()[0];
+    int cols = framework::product(in_x->dims()) / rows_x;
+    Eigen::DSizes<int, 2> bcast_cols(1, cols);
+    auto z_bcast = z.broadcast(bcast_cols);
+    auto dz_bcast = dz.broadcast(bcast_cols);
+    auto x_snorm_bcast = x_norm.square().eval().broadcast(bcast_cols);
     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 (rows_x == rows_y) {
+      auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_cols);
+      auto norm_prod_bcast = (x_norm * y_norm).eval().broadcast(bcast_cols);
+      // compute dx
+      if (out_grad_x) {
+        out_grad_x->mutable_data<T>(context.GetPlace());
+        auto dx = EigenMatrix<T>::Reshape(*out_grad_x, 1);
+        auto grad = y / norm_prod_bcast - z_bcast * x / x_snorm_bcast;
+        dx.device(place) = dz_bcast * grad;
+      }
+      // compute dy
+      if (out_grad_y) {
+        out_grad_y->mutable_data<T>(context.GetPlace());
+        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto grad = x / norm_prod_bcast - z_bcast * y / y_snorm_bcast;
+        dy.device(place) = dz_bcast * grad;
+      }
+    } else {
+      Eigen::DSizes<int, 2> bcast_rows(rows_x, 1);
+      Eigen::DSizes<int, 2> bcast_rows_cols(rows_x, cols);
+      auto y_bcast = y.broadcast(bcast_rows);
+      auto y_snorm_bcast = y_norm.square().eval().broadcast(bcast_rows_cols);
+      auto norm_prod_bcast = (x_norm * y_norm.eval().broadcast(bcast_rows))
+                                 .eval()
+                                 .broadcast(bcast_cols);
+      // compute dx
+      if (out_grad_x) {
+        out_grad_x->mutable_data<T>(context.GetPlace());
+        auto dx = EigenMatrix<T>::Reshape(*out_grad_x, 1);
+        auto grad = y_bcast / norm_prod_bcast - z_bcast * x / x_snorm_bcast;
+        dx.device(place) = dz_bcast * grad;
+      }
+      // compute dy
+      if (out_grad_y) {
+        out_grad_y->mutable_data<T>(context.GetPlace());
+        auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
+        auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
+        dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({0}));
       }
-    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);
     }
   }
 };

paddle/operators/elementwise_mul_op.cc

+/* 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/elementwise_mul_op.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+class ElementWiseMulOp : 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) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto y_dim = ctx.Input<Tensor>("Y")->dims();
+    PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
+                      "Rank of first input must >= rank of second input.")
+    ctx.Output<framework::LoDTensor>("Out")->Resize(x_dim);
+  }
+};
+
+class ElementWiseMulOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ElementWiseMulOpMaker(framework::OpProto *proto,
+                        framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of elementwise mul op");
+    AddInput("Y", "The second input of elementwise mul op");
+    AddAttr<int>("axis",
+                 R"DOC(
+When shape(Y) does not equal shape(X),Y will be broadcasted 
+to match the shape of X and axis should be dimension index Y in X
+        )DOC")
+        .SetDefault(-1)
+        .EqualGreaterThan(-1);
+
+    AddOutput("Out", "The output of elementwise mul op");
+    AddComment(R"DOC(
+Limited elementwise multiple operator.The equation is: Out = X ⊙ Y.
+1. The shape of Y should be same with X or
+2. Y's shape is a subset of X. 
+   Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
+   example:
+      shape(X) = (2, 3, 4, 5), shape(Y) = (,)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
+      shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+      shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
+)DOC");
+  }
+};
+
+class ElementWiseMulOpGrad : 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) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
+
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Rank of first input must >= rank of second input.")
+
+    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(elementwise_mul, ops::ElementWiseMulOp, ops::ElementWiseMulOpMaker,
+            elementwise_mul_grad, ops::ElementWiseMulOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_mul,
+    ops::ElementWiseMulKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    elementwise_mul_grad,
+    ops::ElementWiseMulGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/elementwise_mul_op.cu

+/* 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/elementwise_mul_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_GPU_KERNEL(
+    elementwise_mul,
+    ops::ElementWiseMulKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    elementwise_mul_grad,
+    ops::ElementWiseMulGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/elementwise_mul_op.h

+/* 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 <iostream>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+/*
+ * Out = X ⊙ Y
+ * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
+ *    pre=2, n=3*4, post=5
+ * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)
+ *    pre=2*3, n=4*5, post=1
+ */
+
+inline void get_mid_dims(const framework::DDim& x_dims,
+                         const framework::DDim& y_dims, const int axis,
+                         int& pre, int& n, int& post) {
+  pre = 1;
+  n = 1;
+  post = 1;
+  for (int i = 0; i < axis; ++i) {
+    pre *= x_dims[i];
+  }
+
+  for (int i = 0; i < y_dims.size(); ++i) {
+    PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
+                      "Broadcast dimension mismatch.");
+    n *= y_dims[i];
+  }
+
+  for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
+    post *= x_dims[i];
+  }
+}
+
+template <typename Place, typename T>
+class ElementWiseMulKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* z = ctx.Output<Tensor>("Out");
+    z->mutable_data<T>(ctx.GetPlace());
+
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto z_e = framework::EigenVector<T>::Flatten(*z);
+
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
+                      "Rank of first input must >= rank of second input.")
+
+    if (x_dims == y_dims || product(y_dims) == 1) {
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_e;
+      return;
+    }
+
+    int axis = ctx.Attr<int>("axis");
+    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+    PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
+                   "Axis should be in range [0, x_dims)");
+
+    int pre, n, post;
+    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+    if (post == 1) {
+      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
+                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
+      return;
+    } else {
+      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
+                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      z_e.device(ctx.GetEigenDevice<Place>()) = x_e * y_bcast;
+      return;
+    }
+  }
+};
+
+template <typename Place, typename T>
+class ElementWiseMulGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    using Tensor = framework::Tensor;
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* y = ctx.Input<Tensor>("Y");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dout_e = framework::EigenVector<T>::Flatten(*dout);
+
+    auto x_dims = x->dims();
+    auto y_dims = y->dims();
+
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    if (dx) {
+      dx->mutable_data<T>(ctx.GetPlace());
+    }
+    if (dy) {
+      dy->mutable_data<T>(ctx.GetPlace());
+    }
+
+    if (x_dims == y_dims || product(y_dims) == 1) {
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) = x_e * dout_e;
+      }
+      return;
+    }
+
+    int axis = ctx.Attr<int>("axis");
+    axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
+
+    int pre, n, post;
+    get_mid_dims(x_dims, y_dims, axis, pre, n, post);
+
+    // TODO(gongweibao): wrap reshape to a function.
+    if (post == 1) {
+      auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
+                           .broadcast(Eigen::DSizes<int, 2>(pre, 1))
+                           .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) =
+            (x_e * dout_e)
+                .reshape(Eigen::DSizes<int, 2>(pre, n))
+                .sum(Eigen::array<int, 1>{{0}});
+      }
+      return;
+    } else {
+      auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
+                           .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
+                           .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+      if (dx) {
+        auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+        dx_e.device(ctx.GetEigenDevice<Place>()) = dout_e * y_e_bcast;
+      }
+
+      if (dy) {
+        auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+        dy_e.device(ctx.GetEigenDevice<Place>()) =
+            (x_e * dout_e)
+                .reshape(Eigen::DSizes<int, 3>(pre, n, post))
+                .sum(Eigen::array<int, 2>{{0, 2}});
+      }
+      return;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/fill_zeros_like_op.cc

@@ -23,7 +23,7 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<framework::Tensor>("Dst")->Resize(
+    ctx.Output<framework::LoDTensor>("Dst")->Resize(
         ctx.Input<framework::Tensor>("Src")->dims());
   }
 };

paddle/operators/gather_op.cc

@@ -28,7 +28,7 @@ class GatherOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
     framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());
     output_dims[0] = batch_size;
-    ctx.Output<Tensor>("Out")->Resize(output_dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(output_dims);
   }
 };
 
@@ -38,7 +38,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto X_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto X_grad = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
     auto X = ctx.Input<Tensor>("X");
 
     X_grad->Resize(X->dims());

paddle/operators/gaussian_random_op.cc

@@ -31,7 +31,7 @@ class CPUGaussianRandomKernel : public framework::OpKernel {
     }
     engine.seed(seed);
     std::normal_distribution<T> dist(mean, std);
-    int64_t size = framework::product(tensor->dims());
+    int64_t size = tensor->numel();
     for (int64_t i = 0; i < size; ++i) {
       data[i] = dist(engine);
     }
@@ -44,7 +44,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& context) const override {
-    auto* tensor = context.Output<framework::Tensor>("Out");
+    auto* tensor = context.Output<framework::LoDTensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());

paddle/operators/gaussian_random_op.cu

@@ -50,8 +50,8 @@ class GPUGaussianRandomKernel : public framework::OpKernel {
     T mean = static_cast<T>(context.Attr<float>("mean"));
     T std = static_cast<T>(context.Attr<float>("std"));
     thrust::counting_iterator<unsigned int> index_sequence_begin(0);
-    ssize_t N = framework::product(tensor->dims());
-    thrust::transform(index_sequence_begin, index_sequence_begin + N,
+    int64_t size = tensor->numel();
+    thrust::transform(index_sequence_begin, index_sequence_begin + size,
                       thrust::device_ptr<T>(data),
                       GaussianGenerator<T>(mean, std, seed));
   }

paddle/operators/lookup_table_op.cc

@@ -25,7 +25,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
   void InferShape(const framework::InferShapeContext &context) const override {
     auto table_t = context.Input<Tensor>("W");
     auto ids_t = context.Input<Tensor>("Ids");
-    auto output_t = context.Output<Tensor>("Out");
+    auto output_t = context.Output<framework::LoDTensor>("Out");
 
     output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
   }
@@ -56,7 +56,8 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &context) const override {
     auto table = context.Input<Tensor>("W");
-    auto d_table = context.Output<Tensor>(framework::GradVarName("W"));
+    auto d_table =
+        context.Output<framework::LoDTensor>(framework::GradVarName("W"));
     d_table->Resize(table->dims());
   }
 };

paddle/operators/lookup_table_op.cu

@@ -70,7 +70,7 @@ class LookupTableCUDAKernel : public framework::OpKernel {
 
     size_t N = table_t->dims()[0];
     size_t D = table_t->dims()[1];
-    size_t K = product(ids_t->dims());
+    size_t K = ids_t->numel();
     auto ids = ids_t->data<int32_t>();
     auto table = table_t->data<T>();
     auto output = output_t->mutable_data<T>(context.GetPlace());
@@ -91,7 +91,7 @@ class LookupTableGradCUDAKernel : public framework::OpKernel {
 
     int N = d_table_t->dims()[0];
     int D = d_table_t->dims()[1];
-    int K = product(ids_t->dims());
+    int K = ids_t->numel();
     const int32_t* ids = ids_t->data<int32_t>();
     const T* d_output = d_output_t->data<T>();
     T* d_table = d_table_t->mutable_data<T>(context.GetPlace());

paddle/operators/lookup_table_op.h

@@ -35,7 +35,7 @@ class LookupTableKernel : public framework::OpKernel {
     auto ids = ids_t->data<int32_t>();
     auto table = table_t->data<T>();
     auto output = output_t->mutable_data<T>(context.GetPlace());
-    for (ssize_t i = 0; i < product(ids_t->dims()); ++i) {
+    for (int64_t i = 0; i < ids_t->numel(); ++i) {
       PADDLE_ENFORCE_LT(ids[i], N);
       PADDLE_ENFORCE_GE(ids[i], 0);
       memcpy(output + i * D, table + ids[i] * D, D * sizeof(T));
@@ -61,7 +61,7 @@ class LookupTableGradKernel : public framework::OpKernel {
     t.device(context.GetEigenDevice<platform::CPUPlace>()) =
         t.constant(static_cast<T>(0));
 
-    for (ssize_t i = 0; i < product(ids_t->dims()); ++i) {
+    for (int64_t i = 0; i < ids_t->numel(); ++i) {
       PADDLE_ENFORCE_LT(ids[i], N);
       PADDLE_ENFORCE_GE(ids[i], 0);
       for (int j = 0; j < D; ++j) {

paddle/operators/mean_op.cc

@@ -25,7 +25,7 @@ class MeanOp : public framework::OperatorWithKernel {
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
                             "Input of MeanOp must be initialized.");
-    ctx.Output<Tensor>("Out")->Resize({1});
+    ctx.Output<framework::LoDTensor>("Out")->Resize({1});
   }
 };
 
@@ -45,7 +45,7 @@ class MeanGradOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>(framework::GradVarName("X"))
+    ctx.Output<framework::LoDTensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };

paddle/operators/mean_op.h

@@ -49,12 +49,11 @@ class MeanGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto OG = context.Input<Tensor>(framework::GradVarName("Out"));
-    PADDLE_ENFORCE(framework::product(OG->dims()) == 1,
-                   "Mean Gradient should be scalar");
+    PADDLE_ENFORCE(OG->numel() == 1, "Mean Gradient should be scalar");
     auto IG = context.Output<Tensor>(framework::GradVarName("X"));
     IG->mutable_data<T>(context.GetPlace());
 
-    T ig_size = (T)framework::product(IG->dims());
+    T ig_size = static_cast<T>(IG->numel());
     Eigen::DSizes<int, 1> bcast(ig_size);
 
     EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =

paddle/operators/minus_op.cc

@@ -31,10 +31,9 @@ class MinusOp : public framework::OperatorWithKernel {
     auto *right_tensor = ctx.Input<framework::Tensor>("Y");
 
     PADDLE_ENFORCE_EQ(
-        framework::product(left_tensor->dims()),
-        framework::product(right_tensor->dims()),
+        left_tensor->numel(), right_tensor->numel(),
         "Minus operator must take two tensor with same num of elements");
-    ctx.Output<framework::Tensor>("Out")->Resize(left_tensor->dims());
+    ctx.Output<framework::LoDTensor>("Out")->Resize(left_tensor->dims());
   }
 };
 

paddle/operators/mul_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 using framework::Tensor;
+using framework::LoDTensor;
 
 class MulOp : public framework::OperatorWithKernel {
  public:
@@ -45,7 +46,8 @@ class MulOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         x_mat_dims[1], y_mat_dims[0],
         "First matrix's width must be equal with second matrix's height.");
-    ctx.Output<Tensor>("Out")->Resize({x_mat_dims[0], y_mat_dims[1]});
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        {x_mat_dims[0], y_mat_dims[1]});
   }
 };
 
@@ -94,8 +96,10 @@ class MulOpGrad : public framework::OperatorWithKernel {
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
-    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
+    auto *x_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *y_grad =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
 
     auto x_mat_dims =
         framework::flatten_to_2d(x_dims, Attr<int>("x_num_col_dims"));

paddle/operators/name_convention.md

+## Operator's Parameter Name Convention
+
+To make the operator document itself more clear, we recommend operator names obey the listing conventions.
+
+### OpProtoMaker names
+
+When defining an operator in Paddle, a corresponding [OpProtoMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L170) (TODO: OpProtoMaker Doc)need to be defined. All the Input/Output and Attributes will write into the [OpProto](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L61) , and will be used in client language to create operator. 
+
+- Input/Output.
+  - Input/Output names follow the **CamelCase**. e.g. `X`, `Y`, `Matrix`, `LastAxisInMatrix`. Input/Output much more like Variables, we prefer to meaningful English words. 
+  - If an operator's Input/Output are tensors in math, not match to any meaningful words, input name should starts from `X`. e.g. `X`, `Y`, and output name should starts from `Out`. e.g. `Out`. This rule intends making operators which have few inputs/outputs unified.
+
+- Attribute.
+  - Attribute name follows the **camelCase**. e.g. `x`, `y`, `axis`, `rowwiseMatrix`. Also, attribute name prefers to meaningful English words.
+
+- Comments.
+  - Input/Output/Attr comment follow the format of **(type,default value) usage**, corresponding to which type it can be and how it will be used in the operator. e.g.  Attribute in Accumulator`"gamma" `,`(float, default 1.0) Accumulation multiplier`.
+  - Operator comment format of` R"DOC(your comment here)DOC"`. You should explain the input/output of the operator first. If there is math calculation in this operator, you should write the equation in the comment. e.g. `Out = X + Y`. 
+
+- Order.
+  - Follow the order of Input/Output, then Attribute, then Comments. See the example in best practice.
+
+### Best Practice
+
+Here we give some examples to show how these rules will be used.
+
+- The operator has one input, one output. e.g.`relu`, inputs: `X`, outputs: `Out`. 
+
+- The operator has two input, one output. e.g. `rowwise_add`, inputs : `X`, `Y`, outputs : `Out`.
+
+- The operator contains attribute. e.g. `cosine`, inputs : `X`, `axis`, outputs : `Out`.
+
+  We give a full example of Accumulator Operator.
+
+```c++
+class AccumulateOpMaker : public framework::OpProtoAndCheckerMaker {
+public:
+  AccumulateOpMaker(framework::OpProto *proto,
+                            framework::OpAttrChecker *op_checker)
+    : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "(Tensor) The input tensor that has to be accumulated to the output tensor. 
+    If the output size is not the same as input size, 
+    the output tensor is first reshaped and initialized to zero, and only then, accumulation is done.");
+    AddOutput("Out", "(Tensor) Accumulated output tensor");
+    AddAttr<float>("gamma", "(float, default 1.0) Accumulation multiplier").SetDefault(1.0f);
+    AddComment(R"DOC(
+Accumulate operator accumulates the input tensor to the output tensor. If the
+output tensor already has the right size, we add to it; otherwise, we first
+initialize the output tensor to all zeros, and then do accumulation. Any
+further calls to the operator, given that no one else fiddles with the output
+in the interim, will do simple accumulations.
+Accumulation is done as shown:
+
+Out = 1*X + gamma*Out
+
+where X is the input tensor, Out is the output tensor and gamma is the multiplier
+argument.
+)DOC");
+  }
+};
+```

paddle/operators/cross_entropy_op.cc → paddle/operators/onehot_cross_entropy_op.cc

@@ -12,7 +12,7 @@ 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/cross_entropy_op.h"
+#include "paddle/operators/onehot_cross_entropy_op.h"
 
 namespace paddle {
 namespace operators {
@@ -29,7 +29,7 @@ class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(X->dims().size(), 2, "X's dimension must be 2.");
     PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1.");
     PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]);
-    ctx.Output<Tensor>("Y")->Resize({X->dims()[0]});
+    ctx.Output<framework::LoDTensor>("Y")->Resize({X->dims()[0], 1});
   }
 };
 
@@ -39,7 +39,7 @@ class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto dX = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
     auto X = ctx.Input<Tensor>("X");
 
     dX->Resize(X->dims());

paddle/operators/pad_op.cc

+/* 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/pad_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class PadOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto x_dim = ctx.Input<Tensor>("X")->dims();
+    auto paddings = Attr<std::vector<int>>("paddings");
+    PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()),
+                      "Size of paddings should be equal to 2 * dimension size "
+                      "of input tensor.");
+    std::vector<int64_t> out_dims(x_dim.size());
+    for (int i = 0; i < x_dim.size(); ++i) {
+      out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
+    }
+    ctx.Output<framework::LoDTensor>("Out")->Resize(
+        framework::make_ddim(out_dims));
+  }
+};
+
+class PadOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  PadOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "The input of pad op. "
+             "The input should be a k-D tensor(k > 0 and k < 7)");
+    AddOutput("Out",
+              "The output of pad op."
+              "A tensor with the same shape as X.")
+        .NotInGradient();
+    AddComment(R"DOC(
+Pad input into output, as specified by paddings and pad_value. The input should be a k-D tensor(k > 0 and k < 7). As an example:
+
+Given:
+
+X = [[1, 2],
+   [3, 4]]
+
+and 
+
+paddings = [0, 1, 1, 2]
+
+and
+ 
+pad_value = 0 
+
+then we get 
+
+Out = [[0, 1, 2, 0, 0]
+       [0, 3, 4, 0, 0]
+       [0, 0, 0, 0, 0]]
+)DOC");
+    AddAttr<std::vector<int>>(
+        "paddings",
+        "A list<int> to describes padding rules for each dimension."
+        " For 2-D image tensor, paddings=[0, 1, 2, 3] means"
+        " padding 0 row to top, 1 row to bottom, 2 columns to left"
+        " and 3 columns to right.Size of paddings should be equal to"
+        " 2 * dimension size of input tensor.");
+    AddAttr<float>("pad_value",
+                   "(float) default to 0; "
+                   "The value to fill padded areas.")
+        .SetDefault(0.0f);
+  }
+};
+
+class PadOpGrad : 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) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto *x_g = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    if (x_g != nullptr) {
+      x_g->Resize(x_dims);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(pad, ops::PadOp, ops::PadOpMaker, pad_grad, ops::PadOpGrad);
+REGISTER_OP_CPU_KERNEL(pad, ops::PadKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(pad_grad,
+                       ops::PadGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/pad_op.cu

+/* 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/pad_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(pad, ops::PadKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(pad_grad,
+                       ops::PadGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/pad_op.h

+/* 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, size_t D, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
+
+template <typename Place, typename T, size_t D>
+void PadFunction(const framework::ExecutionContext& context) {
+  auto pads = context.Attr<std::vector<int>>("paddings");
+  Eigen::array<std::pair<int, int>, D> paddings;
+  for (size_t i = 0; i < paddings.size(); ++i) {
+    paddings[i].first = pads[i * 2];
+    paddings[i].second = pads[i * 2 + 1];
+  }
+  T pad_value = context.Attr<T>("pad_value");
+
+  auto* x = context.Input<Tensor>("X");
+  auto* out = context.Output<Tensor>("Out");
+  out->mutable_data<T>(context.GetPlace());
+
+  auto x_tensor = EigenTensor<T, D>::From(*x);
+  auto out_tensor = EigenTensor<T, D>::From(*out);
+  auto place = context.GetEigenDevice<Place>();
+  out_tensor.device(place) = x_tensor.pad(paddings, pad_value);
+}
+
+template <typename Place, typename T>
+class PadKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    int rank = context.Input<Tensor>("X")->dims().size();
+    switch (rank) {
+      case 1:
+        PadFunction<Place, T, 1>(context);
+        break;
+      case 2:
+        PadFunction<Place, T, 2>(context);
+        break;
+      case 3:
+        PadFunction<Place, T, 3>(context);
+        break;
+      case 4:
+        PadFunction<Place, T, 4>(context);
+        break;
+      case 5:
+        PadFunction<Place, T, 5>(context);
+        break;
+      case 6:
+        PadFunction<Place, T, 6>(context);
+        break;
+      default:
+        PADDLE_THROW(
+            "PadOp only support tensors with no more than 6 dimensions.");
+    }
+  }
+};
+
+template <typename Place, typename T, size_t D>
+void PadGradFunction(const framework::ExecutionContext& context) {
+  auto pads = context.Attr<std::vector<int>>("paddings");
+  Eigen::array<std::pair<int, int>, D> paddings;
+  for (size_t i = 0; i < paddings.size(); ++i) {
+    paddings[i].first = -pads[i * 2];
+    paddings[i].second = -pads[i * 2 + 1];
+  }
+  auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
+  auto* d_x = context.Output<Tensor>(framework::GradVarName("X"));
+  if (d_x != nullptr) {
+    d_x->mutable_data<T>(context.GetPlace());
+    auto d_x_tensor = EigenTensor<T, D>::From(*d_x);
+    auto d_out_tensor = EigenTensor<T, D>::From(*d_out);
+    auto place = context.GetEigenDevice<Place>();
+    d_x_tensor.device(place) = d_out_tensor.pad(paddings, 0);
+  }
+}
+
+template <typename Place, typename T>
+class PadGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    size_t rank =
+        context.Input<Tensor>(framework::GradVarName("Out"))->dims().size();
+    switch (rank) {
+      case 1:
+        PadGradFunction<Place, T, 1>(context);
+        break;
+      case 2:
+        PadGradFunction<Place, T, 2>(context);
+        break;
+      case 3:
+        PadGradFunction<Place, T, 3>(context);
+        break;
+      case 4:
+        PadGradFunction<Place, T, 4>(context);
+        break;
+      case 5:
+        PadGradFunction<Place, T, 5>(context);
+        break;
+      case 6:
+        PadGradFunction<Place, T, 6>(context);
+        break;
+      default:
+        PADDLE_THROW(
+            "PadOp only support tensors with no more than 6 dimensions.");
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/recurrent_op.cc

@@ -26,10 +26,11 @@ namespace operators {
 using Scope = framework::Scope;
 using Variable = framework::Variable;
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 
 void RecurrentAlgorithm::InferShape(const Scope& scope) const {
   seq_len_ = scope.FindVar((arg_->inlinks[0]).external)
-                 ->GetMutable<Tensor>()
+                 ->GetMutable<LoDTensor>()
                  ->dims()[0];
   CreateScopes(scope);
   auto step_scopes = GetStepScopes(scope);
@@ -88,7 +89,7 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
         // the weight are located in parent scope
         for (auto& var_name : input.second) {
           if (!step_scope.FindVar(var_name)) {
-            step_scope.NewVar(var_name)->GetMutable<Tensor>();
+            step_scope.NewVar(var_name)->GetMutable<LoDTensor>();
           }
         }
       }
@@ -106,11 +107,12 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
 void RecurrentAlgorithm::InitMemories(Scope* step_scope,
                                       bool infer_shape_mode) const {
   for (auto& attr : arg_->memories) {
-    Tensor* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable<Tensor>();
+    auto* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable<LoDTensor>();
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
                    "memory [%s]'s boot variable [%s] not exists", attr.var,
                    attr.boot_var);
-    Tensor* boot_mem = step_scope->FindVar(attr.boot_var)->GetMutable<Tensor>();
+    auto* boot_mem =
+        step_scope->FindVar(attr.boot_var)->GetMutable<LoDTensor>();
     if (infer_shape_mode) {
       pre_mem->Resize(boot_mem->dims());
       PADDLE_ENFORCE_EQ(pre_mem->dims().size(), 2);
@@ -192,9 +194,9 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
                    "memory variable [%s] does not exists", attr.var);
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
                    "boot variable [%s] does not exists", attr.boot_var);
-    Tensor* mem_grad = step_scope->NewVar(attr.var)->GetMutable<Tensor>();
-    Tensor* boot_mem_grad =
-        step_scope->NewVar(attr.boot_var)->GetMutable<Tensor>();
+    auto* mem_grad = step_scope->NewVar(attr.var)->GetMutable<LoDTensor>();
+    auto* boot_mem_grad =
+        step_scope->NewVar(attr.boot_var)->GetMutable<LoDTensor>();
     if (infer_shape_mode) {
       boot_mem_grad->Resize(mem_grad->dims());
     } else {
@@ -205,7 +207,7 @@ void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
 
 void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const {
   seq_len_ = scope.FindVar((arg_->inlinks[0]).external)
-                 ->GetMutable<Tensor>()
+                 ->GetMutable<LoDTensor>()
                  ->dims()[0];
   auto step_scopes = GetStepScopes(scope);
   rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_,

paddle/operators/reshape_op.cc

+
+/* 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/reshape_op.h"
+
+namespace paddle {
+namespace operators {
+
+class ReshapeOp : public framework::OperatorWithKernel {
+ public:
+  ReshapeOp(const std::string &type, const framework::VariableNameMap &inputs,
+            const framework::VariableNameMap &outputs,
+            const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    // input check
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null");
+    auto shape = ctx.Attr<std::vector<int>>("shape");
+    PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
+    for (auto dim : shape) {
+      PADDLE_ENFORCE(dim > 0, "Each dimension of shape must be positive.");
+    }
+    // capacity check
+    int64_t capacity =
+        std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
+    auto *in = ctx.Input<framework::Tensor>("X");
+    int64_t in_size = framework::product(in->dims());
+    PADDLE_ENFORCE_EQ(capacity, in_size,
+                      "The size of Input(X) mismatches with Attr(shape).");
+    // resize output
+    std::vector<int64_t> shape_int64(shape.size(), 0);
+    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
+                   [](int a) { return static_cast<int64_t>(a); });
+    auto out_dims = framework::make_ddim(shape_int64);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(out_dims);
+  }
+};
+
+class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ReshapeOpMaker(framework::OpProto *proto,
+                 framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input tensor of reshape operator.");
+    AddOutput("Out", "The output tensor of reshape operator.");
+    AddAttr<std::vector<int>>("shape", "Target shape of reshape operator.");
+    AddComment(R"DOC(Reshape operator
+
+Reshape Input(X) into the shape specified by Attr(shape).
+
+An example:
+Given a 2-D tensor X with 2 rows and 2 columns
+
+    [[1, 2], [3, 4]]
+
+with target shape = [1, 4], the reshape operator will transform 
+the tensor X into a 1-D tensor:
+
+    [1, 2, 3, 4]
+
+)DOC");
+  }
+};
+
+class ReshapeGradOp : public framework::OperatorWithKernel {
+ public:
+  ReshapeGradOp(const std::string &type,
+                const framework::VariableNameMap &inputs,
+                const framework::VariableNameMap &outputs,
+                const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) shouldn't be null.");
+    auto dims = ctx.Input<framework::Tensor>("X")->dims();
+    auto *d_in = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    d_in->Resize(dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+
+REGISTER_OP(reshape, ops::ReshapeOp, ops::ReshapeOpMaker, reshape_grad,
+            ops::ReshapeGradOp);
+REGISTER_OP_CPU_KERNEL(reshape,
+                       ops::ReshapeKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    reshape_grad, ops::ReshapeGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/reshape_op.cu

+/* 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/reshape_op.h"
+
+REGISTER_OP_GPU_KERNEL(
+    reshape,
+    paddle::operators::ReshapeKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    reshape_grad,
+    paddle::operators::ReshapeGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/rowwise_add_op.cc

@@ -37,7 +37,7 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
         framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
         "The width of two operands must be same");
     PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1");
-    ctx.Output<Tensor>("Out")->Resize(x_dims);
+    ctx.Output<framework::LoDTensor>("Out")->Resize(x_dims);
   }
 };
 
@@ -76,8 +76,8 @@ class RowwiseAddGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         framework::slice_ddim(x_dims, num_col_dims, x_dims.size()), b_dims,
         "The width of two operands must be same");
-    auto *dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto *db = ctx.Output<Tensor>(framework::GradVarName("b"));
+    auto *dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *db = ctx.Output<framework::LoDTensor>(framework::GradVarName("b"));
     if (dx) dx->Resize(x_dims);
     if (db) db->Resize(b_dims);
   }

paddle/operators/scale_op.cc

@@ -28,7 +28,7 @@ class ScaleOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     auto *in = ctx.Input<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::Tensor>("Out");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
   }
 };

paddle/operators/sgd_op.cc

@@ -23,10 +23,11 @@ class SGDOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(
-        ctx.Input<Tensor>("param")->dims() == ctx.Input<Tensor>("grad")->dims(),
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("param")->dims(),
+                      ctx.Input<Tensor>("grad")->dims(),
                       "Two input of SGD Op's dimension must be same.");
-    ctx.Output<Tensor>("param_out")->Resize(ctx.Input<Tensor>("param")->dims());
+    ctx.Output<framework::LoDTensor>("param_out")
+        ->Resize(ctx.Input<Tensor>("param")->dims());
   }
 };