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

Dockerfile

@@ -10,13 +10,11 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
 ARG WITH_GPU
 ARG WITH_AVX
 ARG WITH_DOC
-ARG WITH_STYLE_CHECK
 
 ENV WOBOQ OFF
-ENV WITH_GPU=${WITH_GPU:-OFF}
+ENV WITH_GPU=${WITH_GPU:-ON}
 ENV WITH_AVX=${WITH_AVX:-ON}
 ENV WITH_DOC=${WITH_DOC:-OFF}
-ENV WITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
 
 ENV HOME /root
 # Add bash enhancements

cmake/external/mkldnn.cmake

@@ -51,7 +51,7 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     DEPENDS             ${MKLDNN_DEPENDS}
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
-    GIT_TAG             "v0.9"
+    GIT_TAG             "v0.10"
     PREFIX              ${MKLDNN_SOURCES_DIR}
     UPDATE_COMMAND      ""
     CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}

cmake/external/mklml.cmake

@@ -28,7 +28,7 @@ INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
 SET(MKLML_VER           "mklml_lnx_2018.0.20170720")
-SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.9/${MKLML_VER}.tgz")
+SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")
@@ -54,7 +54,8 @@ ExternalProject_Add(
     ${EXTERNAL_PROJECT_LOG_ARGS}
     PREFIX                ${MKLML_SOURCE_DIR}
     DOWNLOAD_DIR          ${MKLML_DOWNLOAD_DIR}
-    DOWNLOAD_COMMAND      wget --no-check-certificate -qO- ${MKLML_URL} | tar xz -C ${MKLML_DOWNLOAD_DIR}
+    DOWNLOAD_COMMAND      wget --no-check-certificate ${MKLML_URL} -c -q -O ${MKLML_VER}.tgz 
+                          && tar zxf ${MKLML_VER}.tgz
     DOWNLOAD_NO_PROGRESS  1
     UPDATE_COMMAND        ""
     CMAKE_ARGS            -DCMAKE_INSTALL_PREFIX=${MKLML_INSTALL_ROOT}

cmake/external/openblas.cmake

@@ -25,7 +25,12 @@ IF(NOT ${CBLAS_FOUND})
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         CACHE FILEPATH "openblas library." FORCE)
 
+    IF(APPLE)
+        SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
+        SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
+    ELSE()
         SET(COMMON_ARGS CC=${CMAKE_C_COMPILER} NO_SHARED=1 NO_LAPACK=1 libs)
+    ENDIF()
 
     IF(CMAKE_CROSSCOMPILING)
         IF(ANDROID)
@@ -40,11 +45,11 @@ IF(NOT ${CBLAS_FOUND})
             SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=${TARGET} ARM_SOFTFP_ABI=1 USE_THREAD=0)
         ELSEIF(RPI)
             # use hardfp
-            SET(OPENBLAS_COMMIT "v0.2.19")
+            SET(OPENBLAS_COMMIT "v0.2.20")
             SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 USE_THREAD=0)
         ENDIF()
     ELSE()
-        SET(OPENBLAS_COMMIT "v0.2.19")
+        SET(OPENBLAS_COMMIT "v0.2.20")
         SET(OPTIONAL_ARGS "")
         IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
             SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)

doc/about/index_cn.md

-关于PaddlePaddle
-================
-
-PaddlePaddle是一个最早由百度科学家和工程师共同研发的并行分布式深度学习平台，兼备易用性、高效性、灵活性和可扩展性，目前已被百度内部多个产品线广泛使用。
-PaddlePaddle目前已经开放源码, 但是远未完善，我们希望能在这个基础上不断的改进、扩展和延伸。
-同时我们希望广大开发者积极提供反馈和贡献源代码，建立一个活跃的开源社区。
-
-致谢
---------
-
-在此，特别感谢PaddlePaddle的[所有贡献者](https://github.com/PaddlePaddle/Paddle/graphs/contributors)。

doc/about/index_en.rst

-ABOUT
-=======
-
-PaddlPaddle is an easy-to-use, efficient, flexible and scalable deep learning platform,
-which is originally developed by Baidu scientists and engineers for the purpose of applying deep learning to many products at Baidu.
-
-PaddlePaddle is now open source but far from complete, which is intended to be built upon, improved, scaled, and extended.
-We hope to build an active open source community both by providing feedback and by actively contributing to the source code.
-
-
-Credits
---------
-
-We owe many thanks to `all contributors and developers <https://github.com/PaddlePaddle/Paddle/graphs/contributors>`_ of PaddlePaddle!

doc/api/v2/config/layer.rst

@@ -419,9 +419,14 @@ multi_binary_label_cross_entropy_cost
 ..  autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost
     :noindex:
 
-huber_cost
-----------
-..  autoclass:: paddle.v2.layer.huber_cost
+huber_regression_cost
+-------------------------
+..  autoclass:: paddle.v2.layer.huber_regression_cost
+    :noindex:
+
+huber_classification_cost
+-------------------------
+..  autoclass:: paddle.v2.layer.huber_classification_cost
     :noindex:
 
 lambda_cost

doc/getstarted/build_and_install/index_cn.rst

@@ -6,14 +6,12 @@
 安装流程
 ++++++++
 
-PaddlePaddle提供数个预编译的二进制来进行安装，包括Docker镜像，ubuntu的deb安装包等。我们推荐使用Docker镜像来部署环境，同时欢迎贡献更多的安装包。
+PaddlePaddle提供Docker镜像来部署环境。
 
 .. toctree::
    :maxdepth: 1
    
    docker_install_cn.rst 
-   ubuntu_install_cn.rst
-
 
 
 编译流程

doc/getstarted/build_and_install/index_en.rst

@@ -8,14 +8,13 @@ Install PaddlePaddle
     :maxdepth: 1
 
     docker_install_en.rst
-    ubuntu_install_en.rst
 
 Build from Source
 -----------------
 
 ..  warning::
 
-    Please use :code:`deb` package or :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code.
+    Please use :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code.
 
 ..  toctree::
     :maxdepth: 1

doc/getstarted/build_and_install/ubuntu_install_cn.rst

-Ubuntu部署PaddlePaddle
-===================================
-
-PaddlePaddle提供了ubuntu 14.04 deb安装包。
-
-安装
-------
-
-安装包的下载地址是\: https://github.com/PaddlePaddle/Paddle/releases
-
-它包含四个版本\:
-
-* cpu版本: 支持主流x86处理器平台, 使用了avx指令集。
-
-* cpu-noavx版本：支持主流x86处理器平台，没有使用avx指令集。
-
-* gpu版本：支持主流x86处理器平台，支持nvidia cuda平台，使用了avx指令集。
-
-* gpu-noavx版本：支持主流x86处理器平台，支持nvidia cuda平台，没有使用avx指令集。
-
-下载完相关安装包后，执行:
-
-..  code-block:: shell
-
-    sudo apt-get install gdebi
-    gdebi paddle-*-cpu.deb
-
-或者:
-
-..  code-block:: shell
-
-    dpkg -i paddle-*-cpu.deb
-    apt-get install -f
-
-
-在 :code:`dpkg -i` 的时候如果报一些依赖未找到的错误是正常的，
-在 :code:`apt-get install -f` 里会继续安装 PaddlePaddle。
-
-安装完成后，可以使用命令 :code:`paddle version` 查看安装后的paddle 版本:
-
-..  code-block:: shell
-
-    PaddlePaddle 0.8.0b1, compiled with
-        with_avx: ON
-        with_gpu: OFF
-        with_double: OFF
-        with_python: ON
-        with_rdma: OFF
-        with_timer: OFF
-        with_predict_sdk:
-
-
-可能遇到的问题
---------------
-
-libcudart.so/libcudnn.so找不到
-++++++++++++++++++++++++++++++
-
-安装完成后，运行 :code:`paddle train` 报错\:
-
-..  code-block:: shell
-
-      0831 12:36:04.151525  1085 hl_dso_loader.cc:70] Check failed: nullptr != *dso_handle For Gpu version of PaddlePaddle, it couldn't find CUDA library: libcudart.so Please make sure you already specify its path.Note: for training data on Cpu using Gpu version of PaddlePaddle,you must specify libcudart.so via LD_LIBRARY_PATH.
-
-原因是未设置cuda运行时环境变量。 如果使用GPU版本的PaddlePaddle，请安装CUDA 7.5 和CUDNN 5到本地环境中，并设置：
-
-..  code-block:: shell
-
-    export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/lib:$LD_LIBRARY_PATH
-    export PATH=/usr/local/cuda/bin:$PATH
-

doc/getstarted/build_and_install/ubuntu_install_en.rst

-Debian Package installation guide
-=================================
-
-PaddlePaddle supports :code:`deb` pacakge. The installation of this :code:`deb` package is tested in ubuntu 14.04, but it should be support other debian based linux, too.
-
-There are four versions of debian package, :code:`cpu`, :code:`gpu`, :code:`cpu-noavx`, :code:`gpu-noavx`. And :code:`noavx` version is used to support CPU which does not contain :code:`AVX` instructions. The download url of :code:`deb` package is \: https://github.com/baidu/Paddle/releases/
-
-
-After downloading PaddlePaddle deb packages, you can use :code:`gdebi` install.
-
-..	code-block:: bash
-
-	gdebi paddle-*.deb
-
-If :code:`gdebi` is not installed, you can use :code:`sudo apt-get install gdebi` to install it.
-
-Or you can use following commands to install PaddlePaddle.
-
-..	code-block:: bash
-
-	dpkg -i paddle-*.deb
-	apt-get install -f
-
-And if you use GPU version deb package, you need to install CUDA toolkit and cuDNN, and set related environment variables(such as LD_LIBRARY_PATH) first. It is normal when `dpkg -i` get errors. `apt-get install -f` will continue install paddle, and install dependences. 
-

doc/howto/dev/build_cn.md

+# 编译PaddlePaddle和运行单元测试
+
+## 需要的软硬件
+
+为了开发PaddlePaddle，我们需要
+
+1. 一台电脑，可以装的是 Linux, BSD, Windows 或者 MacOS 操作系统，以及
+1. Docker。
+
+不需要依赖其他任何软件了。即便是 Python 和 GCC 都不需要，因为我们会把所有编译工具都安装进一个 Docker image 里。
+
+## 总体流程
+
+1. 获取源码
+
+   ```bash
+   git clone https://github.com/paddlepaddle/paddle
+   ```
+
+2. 安装开发工具到 Docker image 里
+
+   ```bash
+   cd paddle; docker build -t paddle:dev .
+   ```
+
+   请注意这个命令结尾处的 `.`；它表示 `docker build` 应该读取当前目录下的 [`Dockerfile`文件](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile)，按照其内容创建一个名为 `paddle:dev` 的 Docker image，并且把各种开发工具安装进去。
+
+3. 编译
+
+   以下命令启动一个 Docker container 来执行 `paddle:dev` 这个 Docker image，同时把当前目录（源码树根目录）映射为 container 里的 `/paddle` 目录，并且运行 `Dockerfile` 描述的默认入口程序 [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `cmake` 和 `make` 来编译 `/paddle` 里的源码，结果输出到 `/paddle/build`，也就是本地的源码树根目录里的 `build` 子目录。
+
+   ```bash
+   docker run --rm -v $PWD:/paddle paddle:dev
+   ```
+
+   上述命令编译出一个 CUDA-enabled 版本。如果我们只需要编译一个只支持 CPU 的版本，可以用
+
+   ```bash
+   docker run --rm -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev
+   ```
+
+4. 运行单元测试
+
+   用本机的第一个 GPU 来运行包括 GPU 单元测试在内的所有单元测试：
+
+   ```bash
+   NV_GPU=0 nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
+   ```
+
+   如果编译的时候我们用了 `WITH_GPU=OFF` 选项，那么编译过程只会产生 CPU-based 单元测试，那么我们也就不需要 nvidia-docker 来运行单元测试了。我们只需要：
+
+   ```bash
+   docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
+   ```
+
+   有时候我们只想运行一个特定的单元测试，比如 `memory_test`，我们可以
+
+   ```bash
+   nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test"
+   ```
+
+5. 清理
+
+   有时候我们会希望清理掉已经下载的第三方依赖以及已经编译的二进制文件。此时只需要：
+
+   ```bash
+   rm -rf build
+   ```
+
+## 为什么要 Docker 呀？
+
+- 什么是 Docker?
+
+  如果您没有听说 Docker，可以把它想象为一个类似 virtualenv 的系统，但是虚拟的不仅仅是 Python 的运行环境。
+
+- Docker 还是虚拟机？
+
+  有人用虚拟机来类比 Docker。需要强调的是：Docker 不会虚拟任何硬件，Docker container 里运行的编译工具实际上都是在本机的 CPU 和操作系统上直接运行的，性能和把编译工具安装在本机运行一样。
+
+- 为什么用 Docker?
+
+  把工具和配置都安装在一个 Docker image 里可以标准化编译环境。这样如果遇到问题，其他人可以复现问题以便帮助。
+
+  另外，对于习惯使用Windows和MacOS的开发者来说，使用Docker就不用配置交叉编译环境了。
+
+- 我可以选择不用Docker吗？
+
+  当然可以。大家可以用把开发工具安装进入 Docker image 一样的方式，把这些工具安装到本机。这篇文档介绍基于 Docker 的开发流程，是因为这个流程比其他方法都更简便。
+
+- 学习 Docker 有多难？
+
+  理解 Docker 并不难，大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具，以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。
+
+- 我可以用 IDE 吗？
+
+  当然可以，因为源码就在本机上。IDE 默认调用 make 之类的程序来编译源码，我们只需要配置 IDE 来调用 Docker 命令编译源码即可。
+
+  很多 PaddlePaddle 开发者使用 Emacs。他们在自己的 `~/.emacs` 配置文件里加两行
+
+  ```emacs
+  (global-set-key "\C-cc" 'compile)
+  (setq compile-command
+   "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev")
+  ```
+
+  就可以按 `Ctrl-C` 和 `c` 键来启动编译了。
+
+- 可以并行编译吗？
+
+  是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。
+
+## 可能碰到的问题
+
+- Docker 需要 sudo
+
+  如果用自己的电脑开发，自然也就有管理员权限（sudo）了。如果用公用的电脑开发，需要请管理员安装和配置好 Docker。此外，PaddlePaddle 项目在努力开始支持其他不需要 sudo 的集装箱技术，比如 rkt。
+
+- 在 Windows/MacOS 上编译很慢
+
+  Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存，以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)。
+
+- 磁盘不够
+
+  本文中的例子里，`docker run` 命令里都用了 `--rm` 参数，这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果，是没有名字的 images，也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。

doc/howto/dev/build_en.md

+# Build PaddlePaddle from Source Code and Run Unit Test
+
+## What Developers Need
+
+To contribute to PaddlePaddle, you need
+
+1. A computer -- Linux, BSD, Windows, MacOS, and
+1. Docker.
+
+Nothing else.  Not even Python and GCC, because you can install all build tools into a Docker image.  We run all the tools by running this image.
+
+## General Process
+
+1. Retrieve source code.
+
+   ```bash
+   git clone https://github.com/paddlepaddle/paddle
+   ```
+
+2. Install build tools into a Docker image.
+
+   ```bash
+   cd paddle; docker build -t paddle:dev .
+   ```
+
+   Please be aware of the `.` at the end of the command, which refers to the [`./Dockerfile` file](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile).  `docker build` follows instructions in this file to create a Docker image named `paddle:dev`, and installs building tools into it.
+
+3. Build from source.
+
+   This following command starts a Docker container that executes the Docker image `paddle:dev`, mapping the current directory to `/paddle/` in the container, and runs the default entry-point [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh) as specified in the Dockefile.  `build.sh` invokes `cmake` and `make` to build PaddlePaddle source code, which had been mapped to `/paddle`, and writes outputs to `/paddle/build`, which maps to `build` in the current source directory on the computer.
+
+   ```bash
+   docker run -v $PWD:/paddle paddle:dev
+   ```
+
+   Above command builds a CUDA-enabled version.  If we want to build a CPU-only version, we can type
+
+   ```bash
+   docker run -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev
+   ```
+
+4. Run unit tests.
+
+   To run all unit tests using the first GPU of a node:
+
+   ```bash
+   NV_GPU=0 nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
+   ```
+
+   If we used `WITH_GPU=OFF` at build time, it generates only CPU-based unit tests, and we don't need nvidia-docker to run them.  We can just run
+
+   ```bash
+   docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
+   ```
+
+   Sometimes we want to run a specific unit test, say `memory_test`, we can run
+
+   ```bash
+   nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test"
+   ```
+
+5. Clean Build.
+
+   Sometimes, we might want to clean all thirt-party dependents and built binaries.  To do so, just
+
+   ```bash
+   rm -rf build
+   ```
+
+## Docker, Or Not?
+
+- What is Docker?
+
+  If you haven't heard of it, consider it something like Python's virtualenv.
+
+- Docker or virtual machine?
+
+  Some people compare Docker with VMs, but Docker doesn't virtualize any hardware nor running a guest OS, which means there is no compromise on the performance.
+
+- Why Docker?
+
+  Using a Docker image of build tools standardizes the building environment, which makes it easier for others to reproduce your problems and to help.
+
+  Also, some build tools don't run on Windows or Mac or BSD, but Docker runs almost everywhere, so developers can use whatever computer they want.
+
+- Can I choose not to use Docker?
+
+  Sure, you don't have to install build tools into a Docker image; instead, you can install them in your local computer.  This document exists because Docker would make the development way easier.
+
+- How difficult is it to learn Docker?
+
+    It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools.  Not even to mention the time saved when other people trying to reproduce the issue you have.
+
+- Can I use my favorite IDE?
+
+  Yes, of course.  The source code resides on your local computer, and you can edit it using whatever editor you like.
+
+  Many PaddlePaddle developers are using Emacs.  They add the following few lines into their `~/.emacs` configure file:
+
+  ```emacs
+  (global-set-key "\C-cc" 'compile)
+  (setq compile-command
+   "docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev")
+  ```
+
+  so they could type `Ctrl-C` and `c` to build PaddlePaddle from source.
+
+- Does Docker do parallel building?
+
+  Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores.
+
+## Some Gotchas
+
+- Docker requires sudo
+
+  An owner of a computer has the administrative privilege, a.k.a., sudo, and Docker requires this privilege to work properly.  If you use a shared computer for development, please ask the administrator to install and configure Docker.  We will do our best to support rkt, another container technology that doesn't require sudo.
+
+- Docker on Windows/MacOS builds slowly
+
+  On Windows and MacOS, Docker containers run in a Linux VM.  You might want to give this VM some more memory and CPUs so to make the building efficient.  Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details.
+
+- Not enough disk space
+
+  Examples in this article uses option `--rm` with the `docker run` command.  This option ensures that stopped containers do not exist on hard disks.  We can use `docker ps -a` to list all containers, including stopped.  Sometimes `docker build` generates some intermediate dangling images, which also take disk space.  To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/).

doc/howto/dev/new_op_cn.md

+# 如何写新的Operator
+
+ - [概念简介](#概念简介)
+ - [实现C++类](#实现C++类)
+   - [定义ProtoMaker类](#定义ProtoMaker类)
+   - [定义Operator类](#定义Operator类)
+   - [定义OpKernel类](#定义OpKernel类)
+   - [注册Operator](#注册Operator)
+   - [编译](#编译)
+ - [绑定Python](#绑定Python)
+ - [实现单元测试](#实现单元测试)
+   - [前向Operator单测](#前向Operator单测)
+   - [反向Operator单测](#反向Operator单测)
+   - [编译和执行](#编译和执行)
+
+
+## 概念简介
+
+简单介绍需要用到基类，详细介绍请参考设计文档。
+
+- `framework::OperatorBase`: Operator(简写，Op)基类。
+- `framework::OpKernel`: Op计算函数的基类，称作Kernel。
+- `framework::OperatorWithKernel`：继承自OperatorBase，Op有计算函数，称作有Kernel。
+- `class OpProtoAndCheckerMaker`：描述该Op的输入、输出、属性、注释,主要用于Python API接口生成
+
+依据是否包含kernel，将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下：
+
+  
+ 内容            | 定义位置         
+--------------  | :----------------------  
+OpProtoMake定义  | `.cc`文件，Backward Op不需要定义OpProtoMake
+Op定义           | `.cc`文件 
+Kernel实现       | CPU、GPU共享Kernel在`.h`文件，否则，CPU可以在`.cc`文件，GPU可在`.cu`文件。 
+注册Op           | Op注册在`.cc`文件；Kernel注册CPU在`.cc`文件，GPU在`.cu`文件
+     
+     
+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。
+
+
+## 实现C++类
+
+
+### 1. 定义ProtoMaker类
+
+矩阵乘的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。首先定义`ProtoMaker`来描述该Op的输入、输出及注释：
+    
+```
+class MulOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of mul op");
+    AddInput("Y", "The second input of mul op");
+    AddOutput("Out", "The output of mul op");
+    AddComment(R"DOC(
+Two Element Mul Operator.
+The equation is: Out = X * Y
+)DOC");
+  }
+};
+```
+   
+[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数包括2个：
+
+   - `framework::OpProto` ： 前者存储Op的输入输出和参数属性，将用于Python API接口的生成。
+   - `framework::OpAttrChecker` ：后者用于检查参数属性的合法性。
+   
+构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加该Op的注释，这些函数会将对应内容添加到`OpProto`中。
+
+在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，该命名尽可能的规范。
+
+   
+再举个[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)的例子：
+   
+```
+template <typename AttrType>
+class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input tensor of scale operator.").NotInGradient();
+    AddOutput("Out", "The output tensor of scale operator.").NotInGradient();
+    AddComment(R"DOC(Scale operator
+The equation is: Out = scale*X
+)DOC");
+    AddAttr<AttrType>("scale", "scale of scale operator.").SetDefault(1.0);
+  }
+};
+```
+ 
+ 在这个例子里，两处不同：
+ 
+  - `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中。
+  - `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
+   
+
+### 2. 定义Operator类
+
+
+```c++
+class MulOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto dim0 = ctx.Input<Tensor>("X")->dims();
+    auto dim1 = ctx.Input<Tensor>("Y")->dims();
+    PADDLE_ENFORCE_EQ(dim0.size(), 2,
+                      "input X(%s) should be a tensor with 2 dims, a matrix",
+                      ctx.op_.Input("X"));
+    PADDLE_ENFORCE_EQ(dim1.size(), 2,
+                      "input Y(%s) should be a tensor with 2 dims, a matrix",
+                      ctx.op_.Input("Y"));
+    PADDLE_ENFORCE_EQ(
+        dim0[1], dim1[0],
+        "First matrix's width must be equal with second matrix's height.");
+    ctx.Output<Tensor>("Out")->Resize({dim0[0], dim1[1]});
+  }
+};
+```
+
+[`MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L22)继承自`OperatorWithKernel`。`public`成员：
+	 
+```c++
+using framework::OperatorWithKernel::OperatorWithKernel;
+```
+
+这句表示使用基类`OperatorWithKernel`的构造函数，也可写成：
+   
+```c++
+MulOp(const std::string &type, const framework::VariableNameMap &inputs,
+      const framework::VariableNameMap &outputs,
+      const framework::AttributeMap &attrs)
+  : OperatorWithKernel(type, inputs, outputs, attrs) {}
+```	
+	
+还需要重写`InferShape`接口。`InferShape`为const函数，不能修改Op的成员变量，参数为`const framework::InferShapeContext &ctx`，通过该参数可获取到输入输出以及属性。它的功能是：
+
+  - 1). 做检查， 尽早报错：检查输入数据维度、类型等是否合法。
+  - 2). 设置输出Tensor的形状。
+
+通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和要讲到的注册函数一起放在`.cc`中
+
+### 3. 定义OpKernel类
+
+```C++
+template <typename Place, typename T>
+class MulKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* X = context.Input<Tensor>("X");
+    auto* Y = context.Input<Tensor>("Y");
+    auto* Z = context.Output<Tensor>("Out");
+    Z->mutable_data<T>(context.GetPlace());
+    auto* device_context =
+        const_cast<platform::DeviceContext*>(context.device_context_);
+    math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
+  }
+};
+```
+
+`MulKernel`继承自`framework::OpKernel`，带有模板参数:
+
+  - `typename  Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+  
+ - `typename T` : 表示数据类型，如`float`, `double`等。
+   
+`MulKernel`需要重写`Compute`接口，该接口参数为`const framework::ExecutionContext& context`, `ExecutionContext`相比`InferShapeContext`增加了设备类型，同样可获取到输入输出和属性参数，`Compute`函数里写具体实现时。
+   
+注意，不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel`，`OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 
+   
+到此前向Op实现完成，需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似，这里不再重复。但注意，反向Op没有`ProtoMaker`。
+   
+### 4. 注册Operator
+
+在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
+
+```c++
+namespace ops = paddle::operators;
+REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
+REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(mul_grad,
+              ops::MulGradKernel<paddle::platform::CPUPlace, float>);
+```
+    
+  - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`，
+  - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
+  - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
+
+在 `.cu`文件中注册GPU Kernel。
+   
+```c++
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(mul_grad,
+                       ops::MulGradKernel<paddle::platform::GPUPlace, float>);
+```
+
+### 5. 编译
+
+在[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)文件中添加编译。
+   
+```
+op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
+```
+   
+下面命令可以编译：
+   
+```
+make mul_op
+```
+
+## 绑定Python
+
+- 绑定Python 
+ 
+    在 [`paddle/pybind/pybind.cc 
+`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc)文件中添加该类：
+
+    ```
+    USE_OP(mul);
+    ```
+    如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`:
+    
+    ```
+    USE_CPU_ONLY_OP(gather);
+    ```
+    
+    使用`USE_OP`告知编译器需要链接该Op的目标文件，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
+    
+    
+ - 生成库
+
+   在 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件添加类到`DEPS`中，使得该Op可以链接到生成的lib库中。
+   
+   ```
+   if(WITH_PYTHON)
+     cc_library(paddle_pybind SHARED
+     SRCS pybind.cc
+     DEPS pybind python backward
+     mul_op
+     minus_op)
+   endif(WITH_PYTHON)
+   ```
+
+## 实现单元测试
+
+单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单测](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
+
+### 前向Operator单测
+
+前向Op单测继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`，具体单测流程在`OpTestMeta`里完成。需在`setUp`函数定义输入输出和属性参数，以及Python对比的输出值。
+
+```
+import unittest
+import numpy as np
+from gradient_checker import GradientChecker, create_op
+from op_test_util import OpTestMeta
+
+class TestMulOp(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "mul"
+        self.inputs = {
+            'X': np.random.random((32, 84)).astype("float32"),
+            'Y': np.random.random((84, 100)).astype("float32")
+        }
+        self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
+```
+   首先需要`import`必要的包,下面详细解释其他值：
+   
+   - `self.type = "mul" ` : 定义类型，和注册的类型一致。
+   - `self.inputs` : 定义输入，类型为Numpy.array，并初始化。
+   - `self.outputs` : 定义输出，并得到Python结算结果。
+
+ 
+### 反向Operator单测
+
+反向Op单测继承自`GradientChecker`，而`GradientChecker`集成自`unittest.TestCase`，所以反向单测函数需要`test_`开头。
+
+ ```
+ class MulGradOpTest(GradientChecker):
+    def test_mul(self):
+        op = create_op("mul")
+        inputs = {
+            'X': np.random.random((32, 84)).astype("float32"),
+            'Y': np.random.random((84, 100)).astype("float32")
+        }
+        self.compare_grad(op, inputs)      
+        # mul op will enlarge the relative error
+        self.check_grad(
+            op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5)
+ ```
+
+   - 调用`create_op("mul")`创建反向Op对应的前向Op。
+   - 定义输入`inputs`。
+   - 调用`compare_grad`函数对比CPU、GPU计算结果。
+   - 调用`check_grad`检查梯度稳定性，这里采用数值法检测梯度正确性。
+      - 第一个参数`op` : 前向op。
+      - 第二个参数`inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。
+      - 第三个参数`set(["X", "Y"])` : 指定对输入变量`X`、`Y`做梯度检测。
+      - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
+
+
+### 编译和执行 
+
+单测完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)里添加编译：
+
+```
+py_test(test_mul_op SRCS test_mul_op.py)
+```
+
+编译时需要打开`WITH_TESTING`, 即 `cmake paddle_dir -DWITH_TESTING=ON`，编译成功之后执行单测命令为：
+
+```
+make test ARGS="-R test_mul_op -V"
+```
+或者:
+
+```
+ctest -R test_mul_op
+```

doc/howto/index_cn.rst

@@ -19,6 +19,7 @@
 ..  toctree::
   :maxdepth: 1
 
+  dev/build_cn.rst
   dev/write_docs_cn.rst
   dev/contribute_to_paddle_cn.md
 

doc/howto/index_en.rst

@@ -18,6 +18,7 @@ Development
 ..  toctree::
   :maxdepth: 1
 
+  dev/build_en.rst
   dev/new_layer_en.rst
   dev/contribute_to_paddle_en.md
 

doc/index_en.rst

@@ -7,4 +7,3 @@ PaddlePaddle Documentation
   getstarted/index_en.rst
   howto/index_en.rst
   api/index_en.rst
-  about/index_en.rst

paddle/cuda/include/hl_cnn.h

@@ -173,6 +173,96 @@ extern void hl_avgpool_backward(const int frameCnt,
                                 real* backGrad,
                                 const int outStride);
 
+extern void hl_maxpool3D_forward(const int frameCnt,
+                                 const real* inputData,
+                                 const int channels,
+                                 const int depth,
+                                 const int height,
+                                 const int width,
+                                 const int pooledD,
+                                 const int pooledH,
+                                 const int pooledW,
+                                 const int sizeZ,
+                                 const int sizeY,
+                                 const int sizeX,
+                                 const int strideD,
+                                 const int strideH,
+                                 const int strideW,
+                                 const int paddingD,
+                                 const int paddingH,
+                                 const int paddingW,
+                                 real* tgtData,
+                                 real* maxPoolIdxData,
+                                 const int tgtStride);
+
+extern void hl_maxpool3D_backward(const int frameCnt,
+                                  const real* outGrad,
+                                  const int channels,
+                                  const int depth,
+                                  const int height,
+                                  const int width,
+                                  const int pooledD,
+                                  const int pooledH,
+                                  const int pooledW,
+                                  const int sizeZ,
+                                  const int sizeY,
+                                  const int sizeX,
+                                  const int strideD,
+                                  const int strideH,
+                                  const int strideW,
+                                  const int paddingD,
+                                  const int paddingH,
+                                  const int paddingW,
+                                  real scaleA,
+                                  real scaleB,
+                                  real* targetGrad,
+                                  real* maxPoolIdxData,
+                                  const int outStride);
+
+extern void hl_avgpool3D_forward(const int frameCnt,
+                                 const real* inputData,
+                                 const int channels,
+                                 const int depth,
+                                 const int height,
+                                 const int width,
+                                 const int pooledD,
+                                 const int pooledH,
+                                 const int pooledW,
+                                 const int sizeZ,
+                                 const int sizeY,
+                                 const int sizeX,
+                                 const int strideD,
+                                 const int strideH,
+                                 const int strideW,
+                                 const int paddingD,
+                                 const int paddingH,
+                                 const int paddingW,
+                                 real* tgtData,
+                                 const int tgtStride);
+
+extern void hl_avgpool3D_backward(const int frameCnt,
+                                  const real* outGrad,
+                                  const int channels,
+                                  const int depth,
+                                  const int height,
+                                  const int width,
+                                  const int pooledD,
+                                  const int pooledH,
+                                  const int pooledW,
+                                  const int sizeZ,
+                                  const int sizeY,
+                                  const int sizeX,
+                                  const int strideD,
+                                  const int strideH,
+                                  const int strideW,
+                                  int paddingD,
+                                  int paddingH,
+                                  int paddingW,
+                                  real scaleA,
+                                  real scaleB,
+                                  real* backGrad,
+                                  const int outStride);
+
 /**
  * @brief   Bilinear interpolation forward.
  *
@@ -275,4 +365,4 @@ extern void hl_maxout_backward(real* inGrad,
                                size_t featLen,
                                size_t groups);
 
-#endif /* HL_CNN_H_ */
+#endif  // HL_CNN_H_

paddle/cuda/include/hl_matrix.h

@@ -224,4 +224,80 @@ extern void hl_matrix_collect_shared_bias(real* B_d,
 extern void hl_matrix_rotate(
     real* mat, real* matRot, int dimM, int dimN, bool clockWise);
 
+/**
+ * @brief  Matrix vol2Col: Convert 3D volume into col matrix
+ *
+ * @param[in]   matSrc     input matrix.
+ * @param[in]   channel    channel of matSrc.
+ * @param[in]   depth      depth of matSrc.
+ * @param[in]   height     height of matSrc.
+ * @param[in]   width      width of matSrc.
+ * @param[in]   filterD    depth of filter.
+ * @param[in]   filterH    height of filter.
+ * @param[in]   filterW    width of filter.
+ * @param[in]   strideD    stride in the depth.
+ * @param[in]   strideH    stride in the height.
+ * @param[in]   strideW    stride in the width.
+ * @param[in]   paddingD   padding in the depth.
+ * @param[in]   paddingH   padding in the height.
+ * @param[in]   paddingW   padding in the width.
+ * @param[out]   dataDst     output matrix.
+ *
+ */
+extern void hl_matrix_vol2Col(const real* dataSrc,
+                              int channels,
+                              int depth,
+                              int height,
+                              int width,
+                              int filterD,
+                              int filterH,
+                              int filterW,
+                              int strideD,
+                              int strideH,
+                              int strideW,
+                              int paddingD,
+                              int paddingH,
+                              int paddingW,
+                              real* dataDst);
+
+/**
+ * @brief  Matrix col2Vol: Convert col matrix into 3D volume
+ *
+ * @param[out]  matDst     output matrix.
+ * @param[in]   channel    channel of matDst.
+ * @param[in]   depth      depth of matDst.
+ * @param[in]   height     height of matDst.
+ * @param[in]   width      width of matDst.
+ * @param[in]   filterD    depth of filter.
+ * @param[in]   filterH    height of filter.
+ * @param[in]   filterW    width of filter.
+ * @param[in]   strideD    stride in the depth.
+ * @param[in]   strideH    stride in the height.
+ * @param[in]   strideW    stride in the width.
+ * @param[in]   paddingD   padding in the depth.
+ * @param[in]   paddingH   padding in the height.
+ * @param[in]   paddingW   padding in the width.
+ * @param[in]   matSrc     input matrix.
+ * @param[in]   beta       input
+ * @param[in]   alpha      input
+ *
+ */
+extern void hl_matrix_col2Vol(real* dataDst,
+                              int channels,
+                              int depth,
+                              int height,
+                              int width,
+                              int filterD,
+                              int filterH,
+                              int filterW,
+                              int strideD,
+                              int strideH,
+                              int strideW,
+                              int paddingD,
+                              int paddingH,
+                              int paddingW,
+                              const real* dataSrc,
+                              real alpha,
+                              real beta);
+
 #endif /* HL_MATRIX_H_ */

paddle/cuda/include/stub/hl_cnn_stub.h

@@ -87,6 +87,96 @@ inline void hl_avgpool_backward(const int frameCnt,
                                 real* backGrad,
                                 const int outStride) {}
 
+inline void hl_maxpool3D_forward(const int frameCnt,
+                                 const real* inputData,
+                                 const int channels,
+                                 const int depth,
+                                 const int height,
+                                 const int width,
+                                 const int pooledD,
+                                 const int pooledH,
+                                 const int pooledW,
+                                 const int sizeZ,
+                                 const int sizeY,
+                                 const int sizeX,
+                                 const int strideD,
+                                 const int strideH,
+                                 const int strideW,
+                                 const int paddingD,
+                                 const int paddingH,
+                                 const int paddingW,
+                                 real* tgtData,
+                                 real* maxPoolIdxData,
+                                 const int tgtStride) {}
+
+inline void hl_maxpool3D_backward(const int frameCnt,
+                                  const real* outGrad,
+                                  const int channels,
+                                  const int depth,
+                                  const int height,
+                                  const int width,
+                                  const int pooledD,
+                                  const int pooledH,
+                                  const int pooledW,
+                                  const int sizeZ,
+                                  const int sizeY,
+                                  const int sizeX,
+                                  const int strideD,
+                                  const int strideH,
+                                  const int strideW,
+                                  const int paddingD,
+                                  const int paddingH,
+                                  const int paddingW,
+                                  real scaleA,
+                                  real scaleB,
+                                  real* targetGrad,
+                                  real* maxPoolIdxData,
+                                  const int outStride) {}
+
+inline void hl_avgpool3D_forward(const int frameCnt,
+                                 const real* inputData,
+                                 const int channels,
+                                 const int depth,
+                                 const int height,
+                                 const int width,
+                                 const int pooledD,
+                                 const int pooledH,
+                                 const int pooledW,
+                                 const int sizeZ,
+                                 const int sizeY,
+                                 const int sizeX,
+                                 const int strideD,
+                                 const int strideH,
+                                 const int strideW,
+                                 const int paddingD,
+                                 const int paddingH,
+                                 const int paddingW,
+                                 real* tgtData,
+                                 const int tgtStride) {}
+
+inline void hl_avgpool3D_backward(const int frameCnt,
+                                  const real* outGrad,
+                                  const int channels,
+                                  const int depth,
+                                  const int height,
+                                  const int width,
+                                  const int pooledD,
+                                  const int pooledH,
+                                  const int pooledW,
+                                  const int sizeZ,
+                                  const int sizeY,
+                                  const int sizeX,
+                                  const int strideD,
+                                  const int strideH,
+                                  const int strideW,
+                                  const int paddingD,
+                                  const int paddingH,
+                                  const int paddingW,
+                                  real scaleA,
+                                  real scaleB,
+                                  real* backGrad,
+                                  const int outStride) {}
+
 inline void hl_bilinear_forward(const real* inData,
                                 const size_t inImgH,
                                 const size_t inImgW,

paddle/cuda/include/stub/hl_matrix_stub.h

@@ -99,4 +99,38 @@ inline void hl_matrix_collect_shared_bias(real* B_d,
 inline void hl_matrix_rotate(
     real* mat, real* matRot, int dimM, int dimN, bool clockWise) {}
 
+inline void hl_matrix_vol2Col(const real* dataSrc,
+                              int channels,
+                              int depth,
+                              int height,
+                              int width,
+                              int filterD,
+                              int filterH,
+                              int filterW,
+                              int strideD,
+                              int strideH,
+                              int strideW,
+                              int paddingD,
+                              int paddingH,
+                              int paddingW,
+                              real* dataDst) {}
+
+inline void hl_matrix_col2Vol(real* dataDst,
+                              int channels,
+                              int depth,
+                              int height,
+                              int width,
+                              int filterD,
+                              int filterH,
+                              int filterW,
+                              int strideD,
+                              int strideH,
+                              int strideW,
+                              int paddingD,
+                              int paddingH,
+                              int paddingW,
+                              const real* dataSrc,
+                              real alpha,
+                              real beta) {}
+
 #endif  // HL_MATRIX_STUB_H_

paddle/cuda/src/hl_cuda_matrix.cu

@@ -592,3 +592,204 @@ void hl_matrix_rotate(
       mat, matRot, dimM, dimN, clockWise);
   CHECK_SYNC("hl_matrix_rotate failed");
 }
+
+__global__ void keMatrixVol2Col(int num_kernels,
+                                const real* dataSrc,
+                                real* dataDst,
+                                int depth,
+                                int height,
+                                int width,
+                                int filterD,
+                                int filterH,
+                                int filterW,
+                                int strideD,
+                                int strideH,
+                                int strideW,
+                                int paddingD,
+                                int paddingH,
+                                int paddingW,
+                                int depth_col,
+                                int height_col,
+                                int width_col) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels;
+       index += blockDim.x * gridDim.x) {
+    int w_out = index % width_col;
+    int h_out = (index / width_col) % height_col;
+    int d_out = (index / width_col / height_col) % depth_col;
+    int channel_in = index / width_col / height_col / depth_col;
+    int channel_out = channel_in * filterD * filterH * filterW;
+    int w_in = w_out * strideW - paddingW;
+    int h_in = h_out * strideH - paddingH;
+    int d_in = d_out * strideD - paddingD;
+
+    dataDst +=
+        ((channel_out * depth_col + d_out) * height_col + h_out) * width_col +
+        w_out;
+    dataSrc += ((channel_in * depth + d_in) * height + h_in) * width + w_in;
+    for (int k = 0; k < filterD; ++k) {
+      for (int i = 0; i < filterH; ++i) {
+        for (int j = 0; j < filterW; ++j) {
+          int d = d_in + k;
+          int h = h_in + i;
+          int w = w_in + j;
+          *dataDst = (d >= 0 && d < depth && h >= 0 && h < height && w >= 0 &&
+                      w < width)
+                         ? dataSrc[(k * height + i) * width + j]
+                         : 0;
+          dataDst += depth_col * height_col * width_col;
+        }
+      }
+    }
+  }
+}
+
+void hl_matrix_vol2Col(const real* dataSrc,
+                       int channels,
+                       int depth,
+                       int height,
+                       int width,
+                       int filterD,
+                       int filterH,
+                       int filterW,
+                       int strideD,
+                       int strideH,
+                       int strideW,
+                       int paddingD,
+                       int paddingH,
+                       int paddingW,
+                       real* dataDst) {
+  int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1;
+  int height_col = (height + 2 * paddingH - filterH) / strideH + 1;
+  int width_col = (width + 2 * paddingW - filterW) / strideW + 1;
+  int num_kernels = channels * depth_col * height_col * width_col;
+
+  const int threads = 512;
+  const int blocks = DIVUP(num_kernels, threads);
+
+  keMatrixVol2Col<<<blocks, threads, 0, STREAM_DEFAULT>>>(num_kernels,
+                                                          dataSrc,
+                                                          dataDst,
+                                                          depth,
+                                                          height,
+                                                          width,
+                                                          filterD,
+                                                          filterH,
+                                                          filterW,
+                                                          strideD,
+                                                          strideH,
+                                                          strideW,
+                                                          paddingD,
+                                                          paddingH,
+                                                          paddingW,
+                                                          depth_col,
+                                                          height_col,
+                                                          width_col);
+  CHECK_SYNC("hl_matrix_vol2Col failed");
+}
+
+__global__ void keMatrixCol2Vol(int num_kernels,
+                                real* dataDst,
+                                const real* dataSrc,
+                                int depth,
+                                int height,
+                                int width,
+                                int filterD,
+                                int filterH,
+                                int filterW,
+                                int strideD,
+                                int strideH,
+                                int strideW,
+                                int paddingD,
+                                int paddingH,
+                                int paddingW,
+                                int depth_col,
+                                int height_col,
+                                int width_col,
+                                real alpha,
+                                real beta) {
+  for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < num_kernels;
+       index += blockDim.x * gridDim.x) {
+    real srcVal = 0;
+    real dstVal = dataDst[index];
+    int w = index % width + paddingW;
+    int h = (index / width) % height + paddingH;
+    int d = (index / width / height) % depth + paddingD;
+    int c = index / width / height / depth;
+    // compute the start and end of the output
+    int w_col_start = (w < filterW) ? 0 : (w - filterW) / strideW + 1;
+    int w_col_end = min(w / strideW + 1, width_col);
+    int h_col_start = (h < filterH) ? 0 : (h - filterH) / strideH + 1;
+    int h_col_end = min(h / strideH + 1, height_col);
+    int d_col_start = (d < filterD) ? 0 : (d - filterD) / strideD + 1;
+    int d_col_end = min(d / strideD + 1, depth_col);
+
+    int offset = (c * filterD * filterW * filterH + d * filterW * filterH +
+                  h * filterW + w) *
+                 depth_col * height_col * width_col;
+
+    int coeff_d_col =
+        (1 - strideD * filterW * filterH * depth_col) * height_col * width_col;
+    int coeff_h_col =
+        (1 - strideH * filterW * depth_col * height_col) * width_col;
+    int coeff_w_col = (1 - strideW * depth_col * height_col * width_col);
+
+    for (int d_col = d_col_start; d_col < d_col_end; ++d_col) {
+      for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
+        for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
+          srcVal += dataSrc[offset + d_col * coeff_d_col + h_col * coeff_h_col +
+                            w_col * coeff_w_col];
+        }
+      }
+    }
+    dataDst[index] = alpha * srcVal + beta * dstVal;
+  }
+}
+
+void hl_matrix_col2Vol(real* dataDst,
+                       int channels,
+                       int depth,
+                       int height,
+                       int width,
+                       int filterD,
+                       int filterH,
+                       int filterW,
+                       int strideD,
+                       int strideH,
+                       int strideW,
+                       int paddingD,
+                       int paddingH,
+                       int paddingW,
+                       const real* dataSrc,
+                       real alpha,
+                       real beta) {
+  int depth_col = (depth + 2 * paddingD - filterD) / strideD + 1;
+  int height_col = (height + 2 * paddingH - filterH) / strideH + 1;
+  int width_col = (width + 2 * paddingW - filterW) / strideW + 1;
+  int num_kernels = channels * depth * height * width;
+
+  const int threads = 512;
+  const int blocks = DIVUP(num_kernels, threads);
+
+  keMatrixCol2Vol<<<blocks, threads, 0, STREAM_DEFAULT>>>(num_kernels,
+                                                          dataDst,
+                                                          dataSrc,
+                                                          depth,
+                                                          height,
+                                                          width,
+                                                          filterD,
+                                                          filterH,
+                                                          filterW,
+                                                          strideD,
+                                                          strideH,
+                                                          strideW,
+                                                          paddingD,
+                                                          paddingH,
+                                                          paddingW,
+                                                          depth_col,
+                                                          height_col,
+                                                          width_col,
+                                                          alpha,
+                                                          beta);
+
+  CHECK_SYNC("hl_matrix_col2Vol failed");
+}

paddle/framework/backward.cc

@@ -124,6 +124,9 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
     std::list<Pos> insert_position;
     for (auto& dup_output_op : dup_output_ops) {
       const std::string& name = dup_output_op.first;
+      // duplicate @Empty@ don't need to be added
+      if (name == kEmptyVarName) continue;
+
       auto& dup_op = dup_output_op.second;
       // no duplicate output
       if (dup_op.size() == 1) continue;
@@ -209,7 +212,7 @@ std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars) {
   std::unordered_set<std::string> no_grad_names;
-  no_grad_names.reserve(no_grad_vars.size());
+  no_grad_names.reserve(no_grad_vars.size() + 1);
 
   no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
 

paddle/framework/backward.md

-## Operator/expression 's Backward
+# Operator/expression 's Backward
 
-### Motivation
+## Motivation
 
-In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation lineage, the operator/ expression's Backward feature will generate the backward pass respect to forward pass.
+In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
   
-### Implement : gradient operator registry
+## Backward Operator Registry
 
-|                        | forward operator | backward operator                |
-| ---------------------- | ---------------- | -------------------------------- |
+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 
+| ---------------------- | ---------------- |------------------------- |		
 | **Operator::inputs_**  | Inputs       | Inputs, Outputs, OutputGradients |	
 | **Operator::outputs_** | Outputs          | InputGradients            |
 
-Inputs/Outputs means the input/output of the operator,  InputGradients/OutputGradients is the gradient respect to forward opeartor. Forward operator and Backward operator are isomorphic, save their corresponding needs into member attribute.
+ In most cases, there is a one-to-one correspondence between forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
+
+For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro:
+
+```cpp
+REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
+```
+
+`mul` is the operator's type. `MulOp` and `MulOpMaker` are the operator class and the operator maker class respectively.
+
+`mul_grad` is the type of backward operator, and `MulOpGrad` is its class name.
+
+## Backward Opeartor Creating
+
+Given a certain forward operator, we can get its corresponding backward opeartor by calling:
+
+```cpp
+OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op);
+``` 
+
+The function `BuildGradOp` will sequentially execute following processes:
+
+1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`.
 
-We use a global hash map record the gradient operators available, follow the philosophy  of minimum core, make operator pluggable unit. Each gradient is an operator and it needs to regist itself. 
+2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these are not necessary for gradient computing.
 
-grad_op_builder(fengjiayi)
+3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`.
 
-### Implement : Backward network
+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.
+
+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`.
 
-1. bla bla bla (yuyang)
+1. Op 
+
+   when the input forward network is a Op, return its gradient Operator Immediately.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively and ensure them done. During the process, we need to collect the `OutputGradients` name.
+   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp.
+
+   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable.  
+
+   <p align="center">
+   <img src="./images/duplicate_op.png" width="70%" ><br/>
+
+   1. shared variable in two operators. 
+
+   </p>
+
+   Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. 
+
+   <p align="center">
+   <img src="images/duplicate_op2.png" width="90%" ><br/>
 
-   We share variable in the same scope, as a result, duplicate operator `OutputGradients` will overwirte then duplicate variable.  
+   2. replace shared variable gradient with `Add` Operator
 
-   ![./images/duplicate_op]()
+   </p>
 
-    Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator instead. 
 
-![./images/duplicate_op2]()
 
-​	Then collect the sub graph OutputGradients/InputGradients as the NetOp's and return it.
+​	Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.

paddle/framework/lod_tensor.cc

@@ -19,25 +19,24 @@
 namespace paddle {
 namespace framework {
 
-LODTensor::LOD LODTensor::LOD::SliceLevels(size_t level_begin,
-                                           size_t level_end) const {
+LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) {
   LOD new_lod;
   new_lod.reserve(level_end - level_begin);
   for (size_t i = level_begin; i < level_end; i++) {
-    new_lod.emplace_back(at(i));
+    new_lod.emplace_back(in.at(i));
   }
   return new_lod;
 }
 
-LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
-                                            size_t elem_end) const {
+LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+                 size_t elem_end) {
   // slice the lod.
   LOD new_lod;
-  new_lod.reserve(size() - level);
-  auto start = this->at(level)[elem_begin];
-  auto end = this->at(level)[elem_end];
+  new_lod.reserve(in.size() - level);
+  auto start = in.at(level)[elem_begin];
+  auto end = in.at(level)[elem_end];
 
-  for (auto it = this->begin() + level; it != this->end(); it++) {
+  for (auto it = in.begin() + level; it != in.end(); it++) {
     auto it_begin = std::find(it->begin(), it->end(), start);
     auto it_end = std::find(it_begin, it->end(), end);
     PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
@@ -49,11 +48,11 @@ LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
                    [start](int v) { return v - start; });
     PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD");
   }
-  PADDLE_ENFORCE_LE(new_lod.size(), this->size());
+  PADDLE_ENFORCE_LE(new_lod.size(), in.size());
   return new_lod;
 }
 
-bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
+bool operator==(const LOD& a, const LOD& b) {
   if (a.size() != b.size()) {
     return false;
   }
@@ -70,9 +69,27 @@ bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
       }
     }
   }
-
   return true;
 }
 
+void LODTensor::SliceLevels(size_t level_begin, size_t level_end) {
+  auto new_lod = framework::SliceLevels(lod_, level_begin, level_end);
+  lod_ = new_lod;
+}
+
+void LODTensor::SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) {
+  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
+                 NumLevels());
+  PADDLE_ENFORCE(elem_begin < NumElements(level),
+                 "element begin [%d] out of range [%d]", elem_begin,
+                 NumElements(level));
+  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
+                 "element end [%d] out of range [%d]", elem_end,
+                 NumElements(level));
+
+  auto new_lod = framework::SliceInLevel(lod_, level, elem_begin, elem_end);
+  lod_ = new_lod;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor.h

@@ -15,7 +15,7 @@
 #pragma once
 
 #include <memory>
-#if !defined(PADDLE_ONLY_CPU)
+#ifndef PADDLE_ONLY_CPU
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
 #endif
@@ -27,33 +27,39 @@
 namespace paddle {
 namespace framework {
 
+#ifdef PADDLE_ONLY_CPU
+template <typename T>
+using Vector = std::vector<T>;
+#else
+template <typename T>
+using Vector = thrust::host_vector<T>;
+#endif
+
+using LOD = std::vector<Vector<size_t>>;
+
+LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end);
+
+LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+                 size_t elem_end);
+
+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 : public Tensor {
- public:
-// Level save offsets of each unit.
-#ifdef PADDLE_ONLY_CPU
-  template <typename T>
-  using Vector = std::vector<T>;
-#else
-  template <typename T>
-  using Vector = thrust::host_vector<T>;
-#endif
-  // LoD stores offsets of each level of units, the largest units level first,
-  // then the smaller units level. Each Level stores the offsets of units in
-  // Tesor.
-  class LOD : public std::vector<Vector<size_t>> {
+class LODTensor {
  public:
-    LOD SliceLevels(size_t level_begin, size_t level_end) const;
-    LOD SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) const;
-  };
-
   LODTensor() {}
-  explicit LODTensor(const LOD &lod) : lod_(lod) {}
+  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; }
+
+  Tensor& tensor() { return *tensor_; }
 
-  virtual Tensor *Clone() const { return new LODTensor(lod_); }
+  LOD lod() { return lod_; }
 
   /*
    * Get a element from LOD.
@@ -79,71 +85,23 @@ class LODTensor : public Tensor {
     PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
                    NumLevels());
     // the last offset is the end of last element
-    return lod_[level].size() - 1;
+    return (lod_)[level].size() - 1;
   }
 
   /*
-   * Slice of levels[level_begin:level_end], with tensor shared.
+   * Slice of levels[level_begin:level_end]
    */
-  template <typename T>
-  LODTensor SliceLevels(size_t level_begin, size_t level_end) const;
+  void SliceLevels(size_t level_begin, size_t level_end);
 
   /*
-   * Slice of elements of a level, [elem_begin: elem_end], with tensor shared.
+   * Slice of elements of a level, [elem_begin: elem_end]
    * @note: low performance in slice lod_.
    */
-  template <typename T>
-  LODTensor SliceInLevel(size_t level, size_t elem_begin,
-                         size_t elem_end) const;
-
-  /*
-   * Copy other's lod_'s content, free to mutate.
-   */
-  void CopyLOD(const LODTensor &other) { lod_ = other.lod_; }
-  /*
-   * Determine whether LODTensor has a valid LOD info.
-   */
-  const LOD &lod() const { return lod_; }
-  LOD *mutable_lod() { return &lod_; }
-
-  virtual ~LODTensor() {}
+  void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end);
 
  private:
   LOD lod_;
+  Tensor* tensor_;  // not owned
 };
-
-bool operator==(const LODTensor::LOD &a, const LODTensor::LOD &b);
-
-template <typename T>
-LODTensor LODTensor::SliceLevels(size_t level_begin, size_t level_end) const {
-  auto new_lod = lod_.SliceLevels(level_begin, level_end);
-  // slice levels just need to update LOD info, each level will contains the
-  // whole tensor_, so no need to modify tensor_.
-  LODTensor new_tensor(new_lod);
-  new_tensor.ShareDataWith<T>(*this);
-  return new_tensor;
-}
-
-template <typename T>
-LODTensor LODTensor::SliceInLevel(size_t level, size_t elem_begin,
-                                  size_t elem_end) const {
-  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
-                 NumLevels());
-  PADDLE_ENFORCE(elem_begin < NumElements(level),
-                 "element begin [%d] out of range [%d]", elem_begin,
-                 NumElements(level));
-  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
-                 "element end [%d] out of range [%d]", elem_end,
-                 NumElements(level));
-
-  auto new_lod = lod_.SliceInLevel(level, elem_begin, elem_end);
-
-  // slice elements just need to update LOD info, because offsets are not
-  // changed, so the original tensor_ can be reused.
-  LODTensor new_tensor(new_lod);
-  new_tensor.ShareDataWith<T>(*this);
-  return new_tensor;
-}
-
 }  // namespace framework
 }  // namespace paddle

paddle/framework/lod_tensor_test.cc

@@ -24,13 +24,12 @@ namespace framework {
 class LODTensorTester : public ::testing::Test {
  public:
   virtual void SetUp() override {
-    lod_tensor.reset(new LODTensor);
     // tensor's batch_size: 30
     // 3 levels
     // 0 10 20
     // 0 5 10 15 20
     // 0 2 5 7 10 12 15 20
-    LODTensor::LOD lod;
+    LOD lod;
     lod.push_back(std::vector<size_t>{0, 10, 20});
     lod.push_back(std::vector<size_t>{0, 5, 10, 15, 20});
     lod.push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
@@ -41,75 +40,65 @@ class LODTensorTester : public ::testing::Test {
     // malloc memory
     tensor.mutable_data<float>(place);
 
-    lod_tensor.reset(new LODTensor(lod));
-    lod_tensor->Resize({20 /*batch size*/, 128 /*dim*/});
-
-    lod_tensor->ShareDataWith<float>(tensor);
-    // lod_tensor->ShareDataWith<Tensor>(tensor);
+    lod_tensor.set_lod(lod);
+    lod_tensor.set_tensor(&tensor);
   }
 
  protected:
-  std::unique_ptr<LODTensor> lod_tensor;
   platform::CPUPlace place;
   Tensor 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) {
-    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 1);
+    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(0UL), lod_tensor->NumElements(level));
-    // ASSERT_EQ(new_lod_tensor, *lod_tensor);
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+              lod_tensor.tensor().data<float>());
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 2);
+    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.data<float>(), lod_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.tensor().data<float>(),
+              lod_tensor.tensor().data<float>());
   }
 }
 
 TEST_F(LODTensorTester, SliceInLevel) {
   size_t level = 0;
-  auto new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
+  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.data<float>(), lod_tensor->data<float>());
+  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+            lod_tensor.tensor().data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
+  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.data<float>(), lod_tensor->data<float>());
-}
-
-TEST_F(LODTensorTester, ShareLOD) {
-  LODTensor new_lod_tensor;
-  new_lod_tensor.CopyLOD(*lod_tensor);
-  ASSERT_EQ(new_lod_tensor.lod(), lod_tensor->lod());
-}
-
-TEST_F(LODTensorTester, CopyLOD) {
-  LODTensor new_lod_tensor;
-  new_lod_tensor.CopyLOD(*lod_tensor);
-  bool equals = std::equal(lod_tensor->lod().begin(), lod_tensor->lod().end(),
-                           new_lod_tensor.lod().begin());
-  ASSERT_TRUE(equals);
+  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
+            lod_tensor.tensor().data<float>());
 }
 
 }  // namespace framework

paddle/framework/tensor_impl.h

@@ -117,6 +117,8 @@ inline void Tensor::CopyFrom(const Tensor& src,
     memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
                  boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
   }
+  PADDLE_ENFORCE(cudaStreamSynchronize(0),
+                 "cudaStreamSynchronize failed in Tensor CopyFrom");
 
 #endif
 }

paddle/function/CMakeLists.txt

@@ -21,6 +21,8 @@ if(USE_NNPACK)
   endif()
 endif()
 
+list(APPEND cpp_files neon/NeonDepthwiseConv.cpp)
+
 add_library(paddle_function STATIC ${cpp_files} ${cu_objs})
 add_dependencies(paddle_function ${external_project_dependencies})
 add_dependencies(paddle_function paddle_proto)
@@ -42,11 +44,11 @@ if(WITH_GPU)
     add_simple_unittest(RowConvOpTest)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
-    add_simple_unittest(DepthwiseConvOpTest)
 endif()
 
 add_simple_unittest(Im2ColTest)
 add_simple_unittest(GemmConvOpTest)
+add_simple_unittest(DepthwiseConvOpTest)
 endif()
 
 add_style_check_target(paddle_function ${h_files})

paddle/function/DepthwiseConvOpTest.cpp

@@ -34,4 +34,13 @@ TEST(DepthwiseConv, BackwardFilter) {
 }
 #endif
 
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+TEST(DepthwiseConv, Forward) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
+      "GemmConv-CPU", "NeonDepthwiseConv-CPU", forward);
+}
+
+#endif
+
 }  // namespace paddle

paddle/function/Im2Col.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include "TensorShape.h"
 #include "TensorType.h"
+#include "neon/neon_util.h"
 
 namespace paddle {
 
@@ -93,4 +94,95 @@ public:
                   int paddingWidth);
 };
 
+template <class T>
+struct Padding {
+  static void run(const T* src,
+                  T* dest,
+                  int channels,
+                  int inputHeight,
+                  int inputWidth,
+                  int paddingHeight,
+                  int paddingWidth) {
+    const int destWidth = inputWidth + 2 * paddingWidth;
+    for (int c = 0; c < channels; c++) {
+      if (paddingHeight > 0) {
+        memset(dest, 0, destWidth * paddingHeight * sizeof(T));
+        dest += destWidth * paddingHeight;
+      }
+
+      for (int i = 0; i < inputHeight; i++) {
+        // padding head
+        for (int j = 0; j < paddingWidth; j++) {
+          *dest++ = T(0);
+        }
+
+        memcpy(dest, src, inputWidth * sizeof(T));
+        dest += inputWidth;
+        src += inputWidth;
+
+        // padding tail
+        for (int j = 0; j < paddingWidth; j++) {
+          *dest++ = T(0);
+        }
+      }
+
+      if (paddingHeight > 0) {
+        memset(dest, 0, destWidth * paddingHeight * sizeof(T));
+        dest += destWidth * paddingHeight;
+      }
+    }
+  }
+};
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+template <>
+struct Padding<float> {
+  static void run(const float* src,
+                  float* dest,
+                  int channels,
+                  int inputHeight,
+                  int inputWidth,
+                  int paddingHeight,
+                  int paddingWidth) {
+    const int destWidth = inputWidth + 2 * paddingWidth;
+    for (int c = 0; c < channels; c++) {
+      if (paddingHeight > 0) {
+        memset(dest, 0, destWidth * paddingHeight * sizeof(float));
+        dest += destWidth * paddingHeight;
+      }
+
+      for (int i = 0; i < inputHeight; i++) {
+        // padding head
+        for (int j = 0; j < paddingWidth; j++) {
+          *dest++ = float(0);
+        }
+
+        int step = inputWidth >> 2;
+        int remain = inputWidth & 3;
+        for (int s = 0; s < step; s++) {
+          float32x4_t s0 = vld1q_f32(src);
+          vst1q_f32(dest, s0);
+          src += 4;
+          dest += 4;
+        }
+        for (int r = 0; r < remain; r++) {
+          *dest++ = *src++;
+        }
+
+        // padding tail
+        for (int j = 0; j < paddingWidth; j++) {
+          *dest++ = float(0);
+        }
+      }
+
+      if (paddingHeight > 0) {
+        memset(dest, 0, destWidth * paddingHeight * sizeof(float));
+        dest += destWidth * paddingHeight;
+      }
+    }
+  }
+};
+
+#endif
+
 }  // namespace paddle

paddle/function/neon/neon_util.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
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+#include <arm_neon.h>
+
+namespace paddle {
+
+namespace neon {
+
+inline float32x4_t vld1q_f32_aligned(const float* p) {
+  return vld1q_f32(
+      (const float*)__builtin_assume_aligned(p, sizeof(float32x4_t)));
+}
+
+#ifndef __aarch64__
+inline float32_t vaddvq_f32(float32x4_t a) {
+  float32x2_t v = vadd_f32(vget_high_f32(a), vget_low_f32(a));
+  return vget_lane_f32(vpadd_f32(v, v), 0);
+}
+
+inline float32x4_t vmlaq_laneq_f32(float32x4_t a,
+                                   float32x4_t b,
+                                   float32x4_t v,
+                                   const int lane) {
+  return vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane));
+}
+#endif
+
+}  // namespace neon
+}  // namespace paddle
+
+#endif

paddle/gserver/layers/Conv3DLayer.cpp

+/* Copyright (c) 2016 Baidu, Inc. 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 "Conv3DLayer.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+REGISTER_LAYER(conv3d, Conv3DLayer);
+
+bool Conv3DLayer::init(const LayerMap &layerMap,
+                       const ParameterMap &parameterMap) {
+  if (!ConvBaseLayer::init(layerMap, parameterMap)) return false;
+  int index = 0;
+  for (auto &inputConfig : config_.inputs()) {
+    const ConvConfig &conf = inputConfig.conv_conf();
+    M_.push_back(numFilters_ / conf.groups());
+    K_.push_back(filterPixels_[index] * filterChannels_[index]);
+
+    // create a new weight
+    size_t height, width;
+    width = filterPixels_[index] * filterChannels_[index];
+    height = numFilters_;
+    CHECK_EQ(parameters_[index]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[index]);
+    weights_.emplace_back(w);
+    ++index;
+  }
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, numFilters_, biasParameter_));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
+    }
+  }
+  return true;
+}
+
+size_t Conv3DLayer::getSize() {
+  CHECK_NE(inputLayers_.size(), 0UL);
+  outputH_.clear();
+  outputW_.clear();
+  outputD_.clear();
+  N_.clear();
+  size_t layerSize = 0;
+  for (size_t i = 0; i < inputLayers_.size(); ++i) {
+    outputW_.push_back(outputSize(
+        imgSizeW_[i], filterSize_[i], padding_[i], stride_[i], true));
+    outputH_.push_back(outputSize(
+        imgSizeH_[i], filterSizeY_[i], paddingY_[i], strideY_[i], true));
+    outputD_.push_back(outputSize(
+        imgSizeD_[i], filterSizeZ_[i], paddingZ_[i], strideZ_[i], true));
+
+    N_.push_back(outputD_[i] * outputH_[i] * outputW_[i]);
+    CHECK(layerSize == 0 || N_[i] * size_t(numFilters_) == layerSize);
+    layerSize += N_[i] * numFilters_;
+  }
+  getOutput().setFrameHeight(outputH_[0]);
+  getOutput().setFrameWidth(outputW_[0]);
+  getOutput().setFrameDepth(outputD_[0]);
+  return layerSize;
+}
+
+void Conv3DLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  int outWidth = getSize();
+  resetOutput(batchSize, outWidth);
+
+  for (size_t i = 0; i != inputLayers_.size(); ++i) {
+    REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
+    const MatrixPtr &inMat = getInputValue(i);
+    const MatrixPtr &outMat = getOutputValue();
+    int M = M_[i];
+    int N = N_[i];
+    int K = K_[i];
+    Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+    MatrixPtr wMat = weights_[i]->getW();
+    for (int n = 0; n < batchSize; ++n) {
+      colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(),
+                       channels_[i],
+                       imgSizeD_[i],
+                       imgSizeH_[i],
+                       imgSizeW_[i],
+                       filterSizeZ_[i],
+                       filterSizeY_[i],
+                       filterSize_[i],
+                       strideZ_[i],
+                       strideY_[i],
+                       stride_[i],
+                       paddingZ_[i],
+                       paddingY_[i],
+                       padding_[i]);
+
+      real *outData = outMat->getData() + n * outMat->getStride();
+      MatrixPtr outMatSub =
+          Matrix::create(outData, groups_[i] * M, N, false, useGpu_);
+      for (int g = 0; g < groups_[i]; g++) {
+        MatrixPtr wMatSub = wMat->subMatrix(g * M, M);
+        MatrixPtr in = colBuf_->subMatrix(g * K, K);
+        MatrixPtr out = outMatSub->subMatrix(g * M, M);
+        out->mul(*wMatSub, *in, 1.0, 1.0);
+      }
+    }
+  }
+  if (nullptr != this->biasParameter_) {
+    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
+    this->addBias();
+  }
+  forwardActivation();
+}
+
+void Conv3DLayer::backward(const UpdateCallback &callback) {
+  backwardActivation();
+
+  if (biases_ && biases_->getWGrad()) {
+    bpropBiases();
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+
+  for (size_t i = 0; i != inputLayers_.size(); ++i) {
+    REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
+    if (weights_[i]->getWGrad()) {
+      bpropWeights(i);
+    }
+    if (getInputGrad(i)) {
+      bpropData(i);
+    }
+    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+    weights_[i]->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void Conv3DLayer::bpropWeights(int i) {
+  int M = M_[i];
+  int N = N_[i];
+  int K = K_[i];
+  const MatrixPtr &inMat = getInputValue(i);
+  Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+  MatrixPtr wGradMat = weights_[i]->getWGrad();
+  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  for (int n = 0; n < batchSize; ++n) {
+    colBuf_->vol2Col(inMat->getData() + n * inMat->getStride(),
+                     channels_[i],
+                     imgSizeD_[i],
+                     imgSizeH_[i],
+                     imgSizeW_[i],
+                     filterSizeZ_[i],
+                     filterSizeY_[i],
+                     filterSize_[i],
+                     strideZ_[i],
+                     strideY_[i],
+                     stride_[i],
+                     paddingZ_[i],
+                     paddingY_[i],
+                     padding_[i]);
+
+    real *outGradData =
+        getOutputGrad()->getData() + n * getOutputGrad()->getStride();
+    MatrixPtr outGradSub =
+        Matrix::create(outGradData, groups_[i] * M, N, false, useGpu_);
+    for (int g = 0; g < groups_[i]; ++g) {
+      MatrixPtr inMatSub = colBuf_->subMatrix(g * K, K);
+      MatrixPtr outG = outGradSub->subMatrix(g * M, M);
+      MatrixPtr wGradSub = wGradMat->subMatrix(g * M, M);
+      wGradSub->mul(*outG, *(inMatSub->getTranspose()), 1.0, 1.0);
+    }
+  }
+}
+
+void Conv3DLayer::bpropData(int i) {
+  int M = M_[i];
+  int N = N_[i];
+  int K = K_[i];
+  Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+  MatrixPtr wMat = weights_[i]->getW();
+  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  for (int n = 0; n < batchSize; ++n) {
+    real *outGradData =
+        getOutputGrad()->getData() + n * getOutputGrad()->getStride();
+    real *preGradData =
+        getInputGrad(i)->getData() + n * getInputGrad(i)->getStride();
+    MatrixPtr outGradSub =
+        Matrix::create(outGradData, M * groups_[i], N, false, useGpu_);
+    for (int g = 0; g < groups_[i]; ++g) {
+      MatrixPtr wMatSub = wMat->subMatrix(g * M, M);
+      MatrixPtr outG = outGradSub->subMatrix(g * M, M);
+      MatrixPtr inGradMatSub = colBuf_->subMatrix(g * K, K);
+      inGradMatSub->mul(*(wMatSub->getTranspose()), *outG, 1.0, 0.0);
+    }
+    colBuf_->col2Vol(preGradData,
+                     channels_[i],
+                     imgSizeD_[i],
+                     imgSizeH_[i],
+                     imgSizeW_[i],
+                     filterSizeZ_[i],
+                     filterSizeY_[i],
+                     filterSize_[i],
+                     strideZ_[i],
+                     strideY_[i],
+                     stride_[i],
+                     paddingZ_[i],
+                     paddingY_[i],
+                     padding_[i],
+                     1.0,
+                     1.0);
+  }
+}
+
+void Conv3DLayer::bpropBiases() {
+  MatrixPtr outGradMat = getOutputGrad();
+  if (this->sharedBiases_) {
+    biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f);
+  } else {
+    biases_->getWGrad()->collectBias(*outGradMat, 1.0f);
+  }
+}
+
+void Conv3DLayer::addBias() {
+  MatrixPtr outMat = getOutputValue();
+  if (this->sharedBiases_) {
+    outMat->addSharedBias(*(biases_->getW()), 1.0f);
+  } else {
+    outMat->addBias(*(biases_->getW()), 1.0f);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/Conv3DLayer.h

+/* Copyright (c) 2016 Baidu, Inc. 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 "ConvBaseLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief A subclass of convolution layer.
+ * This layer expands input and use matrix multiplication to
+ * calculate convolution operation.
+ */
+class Conv3DLayer : public ConvBaseLayer {
+public:
+  explicit Conv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
+  ~Conv3DLayer() {}
+
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+  void addBias();
+  void backward(const UpdateCallback& callback);
+  void bpropBiases();
+  void bpropData(int i);
+  void bpropWeights(int i);
+  size_t getSize();
+
+protected:
+  // Figure out the dimensions for individual gemms.
+  IntV M_;  /// numFilters_ / filter_group_;
+  IntV N_;  /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_
+  IntV K_;  /// outputD_ * outputH_ * outputW_
+  MatrixPtr colBuf_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/ConvBaseLayer.cpp

@@ -38,7 +38,6 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
     strideY_.push_back(conf.stride_y());
     dilationY_.push_back(conf.dilation_y());
     filterSizeY_.push_back(conf.filter_size_y());
-    filterPixels_.push_back(filterSize_.back() * filterSizeY_.back());
     channels_.push_back(conf.channels());
     imgSizeH_.push_back(conf.has_img_size_y() ? conf.img_size_y()
                                               : conf.img_size());
@@ -47,31 +46,20 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
     filterChannels_.push_back(conf.filter_channels());
     outputH_.push_back(conf.has_output_y() ? conf.output_y() : conf.output_x());
     outputW_.push_back(conf.output_x());
+
+    paddingZ_.push_back(conf.padding_z());
+    strideZ_.push_back(conf.stride_z());
+    filterSizeZ_.push_back(conf.filter_size_z());
+    imgSizeD_.push_back(conf.img_size_z());
+    outputD_.push_back(conf.output_z());
+    filterPixels_.push_back(filterSize_.back() * filterSizeY_.back() *
+                            filterSizeZ_.back());
   }
 
   CHECK(inputLayers_.size() == parameters_.size());
-  for (size_t i = 0; i < inputLayers_.size(); i++) {
-    size_t height, width;
-    height = filterPixels_[i] * filterChannels_[i];
-    width = (!isDeconv_) ? numFilters_ : channels_[i];
-
-    // create a new weight
-    CHECK_EQ(parameters_[i]->getSize(), width * height);
-    Weight* w = new Weight(height, width, parameters_[i]);
-    weights_.emplace_back(w);
-  }
 
-  /* initialize the biases_ */
-  if (biasParameter_.get()) {
-    if (sharedBiases_) {
-      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
-    } else {
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
-    }
-  }
+  // create new weights_ in derived class
+  // create new biases_ in derived class
 
   // default caffe model
   caffeMode_ = true;

paddle/gserver/layers/ConvBaseLayer.h

@@ -62,6 +62,13 @@ protected:
   IntV outputH_;
   /// The spatial dimensions of output feature map width.
   IntV outputW_;
+
+  IntV outputD_;
+  IntV imgSizeD_;
+  IntV filterSizeZ_;
+  IntV strideZ_;
+  IntV paddingZ_;
+
   /// Group size, refer to grouped convolution in
   /// Alex Krizhevsky's paper: when group=2, the first half of the
   /// filters are only connected to the first half of the input channels,

paddle/gserver/layers/CostLayer.cpp

@@ -572,12 +572,7 @@ void MultiBinaryLabelCrossEntropy::backwardImp(Matrix& output,
   }
 }
 
-//
-// Huber loss for robust 2-classes classification
-//
-REGISTER_LAYER(huber, HuberTwoClass);
-
-bool HuberTwoClass::init(const LayerMap& layerMap,
+bool HuberCost::init(const LayerMap& layerMap,
                      const ParameterMap& parameterMap) {
   CostLayer::init(layerMap, parameterMap);
   if (useGpu_) {
@@ -589,7 +584,7 @@ bool HuberTwoClass::init(const LayerMap& layerMap,
   return true;
 }
 
-void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) {
+void HuberCost::forwardImp(Matrix& output, Argument& label, Matrix& cost) {
   if (useGpu_) {
     for (size_t i = 0; i < inputLayers_.size(); i++) {
       tmpCpuInput_[i].resizeAndCopyFrom(
@@ -597,61 +592,123 @@ void HuberTwoClass::forwardImp(Matrix& output, Argument& label, Matrix& cost) {
     }
     hl_stream_synchronize(HPPL_STREAM_DEFAULT);
   }
-  forwardImpIn(output, label, cost);
 }
 
-void HuberTwoClass::forwardImpIn(Matrix& output,
+//
+// Huber loss for robust regression.
+//
+REGISTER_LAYER(huber_regression, HuberRegressionLoss);
+
+bool HuberRegressionLoss::init(const LayerMap& layerMap,
+                               const ParameterMap& parameterMap) {
+  HuberCost::init(layerMap, parameterMap);
+  delta_ = config_.delta();
+  return true;
+}
+
+void HuberRegressionLoss::forwardImp(Matrix& output,
                                      Argument& label,
                                      Matrix& target) {
+  HuberCost::forwardImp(output, label, target);
   size_t numSamples = target.getHeight();
-  CHECK_EQ((*label.ids).getSize(), numSamples);
+  size_t dim = output.getWidth();
+  CHECK(label.value);
+  CHECK_EQ((*label.value).getHeight(), numSamples);
   CHECK_EQ(output.getHeight(), numSamples);
-  CHECK_EQ(output.getWidth(), (size_t)1);
+  CHECK_EQ(dim, (*label.value).getWidth());
   CHECK_EQ(target.getWidth(), (size_t)1);
 
   real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData();
-  int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData();
-  std::vector<real> cost(numSamples);
+  real* lbl =
+      useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData();
+  std::vector<real> cost(numSamples, 0);
   for (size_t i = 0; i < numSamples; ++i) {
-    int y = 2 * lbl[i] - 1;
-    if (out[i] * y < -1)
-      cost[i] = -4 * out[i] * y;
-    else if (out[i] * y < 1)
-      cost[i] = (1 - out[i] * y) * (1 - out[i] * y);
+    for (size_t j = 0; j < dim; ++j) {
+      int index = i * dim + j;
+      real a = std::abs(lbl[index] - out[index]);
+      if (a <= delta_)
+        cost[i] += a * a / 2;
       else
-      cost[i] = 0;
+        cost[i] += delta_ * (a - delta_ / 2);
+    }
   }
   target.copyFrom(cost.data(), numSamples);
 }
 
-void HuberTwoClass::backwardImp(Matrix& outputValue,
+void HuberRegressionLoss::backwardImp(Matrix& output,
                                       Argument& label,
-                                Matrix& outputGrad) {
-  if (useGpu_) {
-    backwardImpIn(
-        *tmpCpuInput_[0].value, tmpCpuInput_[1], *tmpCpuInput_[0].grad);
-    outputGrad.copyFrom(*tmpCpuInput_[0].grad);
-  } else {
-    backwardImpIn(outputValue, label, outputGrad);
+                                      Matrix& outputG) {
+  size_t numSamples = output.getHeight();
+  size_t dim = output.getWidth();
+  real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData();
+  real* lbl =
+      useGpu_ ? tmpCpuInput_[1].value->getData() : (*label.value).getData();
+  real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData();
+  for (size_t i = 0; i < numSamples; ++i) {
+    for (size_t j = 0; j < dim; ++j) {
+      int index = i * dim + j;
+      real a = lbl[index] - out[index];
+      if (std::abs(a) <= delta_)
+        grad[index] += -a;
+      else
+        grad[index] += a > 0 ? -delta_ : delta_;
+    }
   }
+  if (useGpu_) outputG.copyFrom(grad, numSamples * dim);
 }
 
-void HuberTwoClass::backwardImpIn(Matrix& output,
+//
+// Huber loss for robust 2-classes classification
+//
+REGISTER_LAYER(huber_classification, HuberTwoClassification);
+
+bool HuberTwoClassification::init(const LayerMap& layerMap,
+                                  const ParameterMap& parameterMap) {
+  return HuberCost::init(layerMap, parameterMap);
+}
+
+void HuberTwoClassification::forwardImp(Matrix& output,
+                                        Argument& label,
+                                        Matrix& target) {
+  HuberCost::forwardImp(output, label, target);
+  size_t numSamples = target.getHeight();
+  CHECK(label.ids);
+  CHECK_EQ((*label.ids).getSize(), numSamples);
+  CHECK_EQ(output.getHeight(), numSamples);
+  CHECK_EQ(output.getWidth(), (size_t)1);
+  CHECK_EQ(target.getWidth(), (size_t)1);
+
+  real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData();
+  int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData();
+  std::vector<real> cost(numSamples, 0);
+  for (size_t i = 0; i < numSamples; ++i) {
+    int y = 2 * lbl[i] - 1;
+    real a = out[i] * y;
+    if (a < -1)
+      cost[i] = -4 * a;
+    else if (a < 1)
+      cost[i] = (1 - a) * (1 - a);
+  }
+  target.copyFrom(cost.data(), numSamples);
+}
+
+void HuberTwoClassification::backwardImp(Matrix& output,
                                          Argument& label,
                                          Matrix& outputG) {
   size_t numSamples = output.getHeight();
-  real* out = output.getData();
-  real* grad = outputG.getData();
-  int* lbl = (*label.ids).getData();
+  real* out = useGpu_ ? tmpCpuInput_[0].value->getData() : output.getData();
+  int* lbl = useGpu_ ? tmpCpuInput_[1].ids->getData() : (*label.ids).getData();
+  real* grad = useGpu_ ? tmpCpuInput_[0].grad->getData() : outputG.getData();
   for (size_t i = 0; i < numSamples; ++i) {
     int y = 2 * lbl[i] - 1;
-    if (y * out[i] < -1)
+    real a = out[i] * y;
+    if (a < -1)
       grad[i] += -4 * y;
-    else if (y * out[i] < 1)
-      grad[i] += -2 * (1 - y * out[i]) * y;
+    else if (a < 1)
+      grad[i] += -2 * (1 - a) * y;
   }
+  if (useGpu_) outputG.copyFrom(grad, numSamples);
 }
-
 /**
  * This cost layer compute the sum of its input as loss.
  * \f[

paddle/gserver/layers/CostLayer.h

@@ -304,37 +304,70 @@ public:
                    Matrix& outputGrad) override;
 };
 
-/**
- * Huber loss for robust 2-classes classification.
- *
- * For label={0, 1}, let y=2*label-1. Given output f, the loss is:
- * \f[
- * Loss =
- * \left\{\begin{matrix}
- *  4 * y * f     &   \textit{if}  \ \  y* f < -1 \\
- *  (1 - y * f)^2 &  \textit{if}   \ \  -1 < y * f < 1  \\
- *  0             &                    \textit{otherwise}
- * \end{matrix}\right.
- * \f]
+/*
+ * A base layer for HuberRegressionLoss and HuberTwoClassification.
  */
-class HuberTwoClass : public CostLayer {
+class HuberCost : public CostLayer {
+public:
   std::vector<Argument> tmpCpuInput_;
 
-public:
-  explicit HuberTwoClass(const LayerConfig& config) : CostLayer(config) {}
+  explicit HuberCost(const LayerConfig& config) : CostLayer(config) {}
 
   bool init(const LayerMap& layerMap,
             const ParameterMap& parameterMap) override;
 
   void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
-  void forwardImpIn(Matrix& output, Argument& label, Matrix& cost);
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override {}
+};
+
+/**
+ * Huber loss for robust regression.
+ *
+ * Given output f(x), label y and delta, the loss is:
+ * Loss = 0.5 * (1 - y * f)^2, if abs(y - f) <= delta \\
+ * Loss = delta * abs(y - f) - 0.5 * delta^2, otherwise
+ */
+class HuberRegressionLoss : public HuberCost {
+public:
+  explicit HuberRegressionLoss(const LayerConfig& config) : HuberCost(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
 
   void backwardImp(Matrix& outputValue,
                    Argument& label,
                    Matrix& outputGrad) override;
 
-  void backwardImpIn(Matrix& outputValue, Argument& label, Matrix& outputGrad);
+protected:
+  real delta_;
+};
+
+/**
+ * Huber loss for robust 2-classes classification.
+ *
+ * For label={0, 1}, let y=2*label-1. Given output f(x), the loss is:
+ * Loss = 4 * y * f, if y* f < -1 \\
+ * Loss = (1 - y * f)^2, if -1 < y * f < 1  \\
+ * Loss = 0, otherwise
+ */
+class HuberTwoClassification : public HuberCost {
+public:
+  explicit HuberTwoClassification(const LayerConfig& config)
+      : HuberCost(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forwardImp(Matrix& output, Argument& label, Matrix& cost) override;
+
+  void backwardImp(Matrix& outputValue,
+                   Argument& label,
+                   Matrix& outputGrad) override;
 };
 
 typedef std::shared_ptr<CostLayer> CostLayerPtr;

paddle/gserver/layers/CrossEntropyOverBeam.cpp

+/* 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 "CrossEntropyOverBeam.h"
+
+namespace paddle {
+
+void CostForOneSequence::calValidExpandStep() {
+  validExpansionCount_ = 0;
+  goldAsExtraPath_ = true;
+
+  for (size_t i = 0; i < beams_->expansionCount; ++i) {
+    real gold = static_cast<real>(beams_->gold[i]);
+    if (i) {
+      real* start = beams_->candidateIds[i - 1]->getData();
+      goldRowIds_[i] = std::count_if(
+          start,
+          start + goldRowIds_[i - 1] * beamSize_ + goldColIds_[i - 1],
+          [](const real& val) { return val != -1.; });
+    } else {
+      goldRowIds_[i] = 0;
+    }
+
+    real* start =
+        beams_->candidateIds[i]->getData() + goldRowIds_[i] * beamSize_;
+    real* findEnd = std::find(start, start + beamSize_, gold);
+    validExpansionCount_++;
+
+    if (start + beamSize_ == findEnd) return;
+    goldColIds_[i] = findEnd - start;
+  }
+  if (goldColIds_[beams_->expansionCount - 1] != -1) goldAsExtraPath_ = false;
+}
+
+size_t CostForOneSequence::initLastExpansion() {
+  int beamId = validExpansionCount_ - 1;
+  const MatrixPtr candidates = beams_->candidateIds[beamId];
+  size_t height = candidates->getHeight();
+
+  /* initialization the last expansion. */
+  size_t pathCount = std::count_if(candidates->getData(),
+                                   candidates->getData() + height * beamSize_,
+                                   [](const real& val) { return val != -1; });
+  /*
+   * if the gold sequence falls off the beam during search, add the gold
+   * sequence as the last path into the all expanded candidates.
+   */
+  if (goldAsExtraPath_) goldIdsInFinalExpansion_ = pathCount++;
+
+  pathRowIdsInEachBeam_.clear();
+  pathRowIdsInEachBeam_.resize(validExpansionCount_,
+                               std::vector<int>(pathCount, 0));
+  parentIdsInBeam_.clear();
+  parentIdsInBeam_.resize(pathCount, 0);
+
+  if (goldAsExtraPath_) {
+    /* add gold sequence into the total expansion. */
+    pathRowIdsInEachBeam_[beamId].back() =
+        beams_->gold[beamId] +
+        getSeqStartPos(beamId, goldRowIds_[validExpansionCount_ - 1]);
+    parentIdsInBeam_.back() = goldRowIds_[validExpansionCount_ - 1];
+  } else {
+    size_t goldOffset = goldRowIds_[beamId] * beamSize_ + goldColIds_[beamId];
+    goldIdsInFinalExpansion_ =
+        std::count_if(candidates->getData(),
+                      candidates->getData() + goldOffset,
+                      [](const real& val) { return val != -1.; });
+  }
+
+  /*
+   * TODO(caoying): fix this, store the indices of selected candidate
+   * paths into Argument.ids
+   */
+  real* ids = candidates->getData();
+  size_t curIdx = 0;
+  for (size_t i = 0; i < height; ++i) {
+    int basePos = getSeqStartPos(beamId, i);
+    for (size_t j = 0; j < beamSize_; ++j) {
+      int id = ids[i * beamSize_ + j];
+      if (id == -1) continue;
+      pathRowIdsInEachBeam_[beamId][curIdx] = id + basePos;
+      parentIdsInBeam_[curIdx++] = i;
+    }
+  }
+  return pathCount;
+}
+
+void CostForOneSequence::constructTotalExpansion() {
+  /*
+   * construct the entire expanded beam by begining with the last search
+   * in which gold falls off the beam.
+   */
+  size_t totalPathCount = initLastExpansion();
+
+  for (int beamId = validExpansionCount_ - 2; beamId >= 0; --beamId) {
+    const MatrixPtr candidates = beams_->candidateIds[beamId];
+    real* ids = candidates->getData();
+
+    int lastParentIdInBeam = -1;
+    int basePos = -1;
+    for (size_t i = 0;
+         i < (goldAsExtraPath_ ? totalPathCount - 1 : totalPathCount);
+         ++i) {
+      int id = ids[parentIdsInBeam_[i]];
+      int parentRowId = std::div(parentIdsInBeam_[i], beamSize_).quot;
+      if (parentIdsInBeam_[i] != lastParentIdInBeam)
+        basePos = getSeqStartPos(beamId, parentRowId);
+
+      pathRowIdsInEachBeam_[beamId][i] = id + basePos;
+      lastParentIdInBeam = parentIdsInBeam_[i];
+      parentIdsInBeam_[i] = parentRowId;
+
+      if (goldAsExtraPath_)
+        pathRowIdsInEachBeam_[beamId][totalPathCount - 1] =
+            beams_->gold[beamId] + getSeqStartPos(beamId, goldRowIds_[beamId]);
+    }
+  }
+}
+
+real CostForOneSequence::globallyNormalizedScore() {
+  expandedPathScores_.resize(validExpansionCount_);
+
+  Matrix::resizeOrCreate(
+      softmaxOut_, 1, pathRowIdsInEachBeam_[0].size(), false, false);
+  softmaxOut_->zeroMem();
+  MatrixPtr tmp = Matrix::create(
+      softmaxOut_->getData(), softmaxOut_->getWidth(), 1, false, false);
+
+  for (size_t i = 0; i < validExpansionCount_; ++i) {
+    Matrix::resizeOrCreate(expandedPathScores_[i],
+                           pathRowIdsInEachBeam_[i].size(),
+                           1,
+                           false,
+                           false);
+    expandedPathScores_[i]->zeroMem();
+
+    IVectorPtr rowIds = IVector::create(pathRowIdsInEachBeam_[i].data(),
+                                        pathRowIdsInEachBeam_[i].size(),
+                                        false);
+    expandedPathScores_[i]->selectRows(*(beams_->scores[i]), *rowIds);
+    tmp->add(*expandedPathScores_[i]);
+  }
+
+  softmaxOut_->softmax(*softmaxOut_);
+  return -std::log(softmaxOut_->getData()[goldIdsInFinalExpansion_]);
+}
+
+real CostForOneSequence::forward() {
+  calValidExpandStep();
+  constructTotalExpansion();
+  return globallyNormalizedScore();
+}
+
+void CostForOneSequence::backward() {
+  /*
+   * when softmax layer is the output layer, and it is combined with
+   * cross-entropy as cost. The derivate with regard to softmax's input
+   * is simply:
+   *
+   * grad_i = softmax_out_i - target_i,
+   *
+   * and here hard label is used.
+   */
+  softmaxOut_->getData()[goldIdsInFinalExpansion_] -= 1.;
+
+  MatrixPtr tmp = Matrix::create(
+      softmaxOut_->getData(), softmaxOut_->getWidth(), 1, false, false);
+
+  for (size_t i = 0; i < validExpansionCount_; ++i) {
+    IVectorPtr rowIds = IVector::create(pathRowIdsInEachBeam_[i].data(),
+                                        pathRowIdsInEachBeam_[i].size(),
+                                        false);
+    /*
+      beams_->scoreGrad[i] has been intialized outside this class, this
+      class only keeps a pointer pointing to the original input gradients,
+      so here does not need to allocate or initalize the memory.
+    */
+    tmp->addToRows(*beams_->scoreGrad[i], *rowIds);
+  }
+}
+
+REGISTER_LAYER(cross_entropy_over_beam, CrossEntropyOverBeam);
+
+bool CrossEntropyOverBeam::init(const LayerMap& layerMap,
+                                const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(0U, inputLayers_.size() % 3) << "Error input number.";
+
+  beamExpanCount_ = inputLayers_.size() / 3;
+
+  candidateScores_.resize(beamExpanCount_);
+  candidateScoreGrad_.resize(beamExpanCount_);
+
+  candidateInBeam_.resize(beamExpanCount_);
+  goldSequence_.resize(beamExpanCount_);
+  gradToInputs_.resize(beamExpanCount_);
+
+  setNeedSequenceInfo(false);
+  return true;
+}
+
+void CrossEntropyOverBeam::checkInputs() {
+  batchSize_ = 0;
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    const Argument& scores = getInput(i * 3);
+    const Argument& selCandidates = getInput(i * 3 + 1);
+    const Argument& goldSeq = getInput(i * 3 + 2);
+
+    if (i) {
+      CHECK(scores.hasSubseq()) << "input " << i << " "
+                                << inputLayers_[i * 3]->getName()
+                                << " should be a nested sequence";
+      CHECK_EQ(getInputValue(i * 3 + 1)->getWidth(), beamSize_);
+      CHECK_EQ(batchSize_, static_cast<size_t>(scores.getNumSequences()));
+      CHECK_EQ(scores.getNumSubSequences(), selCandidates.getBatchSize());
+    } else {
+      CHECK(scores.hasSeq()) << "input " << i << " "
+                             << inputLayers_[i]->getName()
+                             << " should be a sequence";
+      batchSize_ = scores.getNumSequences();
+      beamSize_ = getInputValue(i * 3 + 1)->getWidth();
+      CHECK_EQ(batchSize_, static_cast<size_t>(selCandidates.getBatchSize()));
+    }
+    CHECK_EQ(1U, scores.value->getWidth());
+    CHECK_EQ(batchSize_, static_cast<size_t>(goldSeq.getBatchSize()));
+  }
+}
+
+void CrossEntropyOverBeam::copyInputsToCpu() {
+  auto copyValue = [](const MatrixPtr& src, MatrixPtr& trg) {
+    if (dynamic_cast<GpuMatrix*>(src.get())) {
+      Matrix::resizeOrCreate(
+          trg, src->getHeight(), src->getWidth(), false, false);
+      trg->copyFrom(*src);
+    } else {
+      trg = std::move(src);
+    }
+  };
+
+  auto copyIds = [](const IVectorPtr& src, IVectorPtr& trg) {
+    if (dynamic_cast<GpuIVector*>(src.get())) {
+      IVector::resizeOrCreate(trg, src->getSize(), false);
+      trg->copyFrom(*src);
+    } else {
+      trg = std::move(src);
+    }
+  };
+
+  beamSplitPos_.clear();
+  beamSplitPos_.resize(batchSize_, std::vector<int>(beamExpanCount_, 0));
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    copyValue(getInputValue(i * 3), candidateScores_[i]);
+    copyValue(getInputValue(i * 3 + 1), candidateInBeam_[i]);
+    copyIds(getInput(i * 3 + 2).ids, goldSequence_[i]);
+
+    if (i) {
+      ICpuGpuVectorPtr seqInfo = getInput(i * 3).sequenceStartPositions;
+      const int* seqStarts = seqInfo->getMutableData(false);
+      ICpuGpuVectorPtr subSeqInfo = getInput(i * 3).subSequenceStartPositions;
+      const int* subSeqStarts = subSeqInfo->getMutableData(false);
+
+      size_t seqId = 1;
+      for (size_t subSeqId = 0; subSeqId < subSeqInfo->getSize() - 1;
+           ++subSeqId) {
+        CHECK_LT(seqId, seqInfo->getSize());
+        if (subSeqStarts[subSeqId] == seqStarts[seqId]) {
+          beamSplitPos_[seqId][i] = beamSplitPos_[seqId - 1][i];
+          seqId++;
+        }
+        beamSplitPos_[seqId - 1][i]++;
+      }
+    } else {
+      for (size_t j = 0; j < batchSize_; ++j) beamSplitPos_[j][i] = j + 1;
+    }
+  }
+}
+
+void CrossEntropyOverBeam::splitBatchBeams() {
+  beamCosts_.resize(batchSize_);
+  beamPerSeq_.resize(batchSize_, BeamExpansion(beamExpanCount_));
+
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    int* seqStarts =
+        getInput(i * 3).sequenceStartPositions->getMutableData(false);
+
+    int* subSeqStarts = nullptr;
+    int maxLen = 0;
+    if (i) {
+      subSeqStarts =
+          getInput(i * 3).subSequenceStartPositions->getMutableData(false);
+      maxLen = getInput(i * 3).subSequenceStartPositions->getSize() - 1;
+    } else {
+      maxLen = getInput(i).sequenceStartPositions->getSize() - 1;
+    }
+
+    for (size_t j = 0; j < batchSize_; ++j) {
+      beamPerSeq_[j].scores[i] =
+          Matrix::create(candidateScores_[i]->getData() + seqStarts[j],
+                         seqStarts[j + 1] - seqStarts[j],
+                         1,
+                         false,
+                         false);
+      beamPerSeq_[j].scoreGrad[i] =
+          Matrix::create(candidateScoreGrad_[i]->getData() + seqStarts[j],
+                         seqStarts[j + 1] - seqStarts[j],
+                         1,
+                         false,
+                         false);
+
+      int offset = j ? beamSplitPos_[j - 1][i] : 0;
+      int height = beamSplitPos_[j][i] - (j ? beamSplitPos_[j - 1][i] : 0);
+      CHECK_GE(maxLen, offset + height);
+      beamPerSeq_[j].seqInfo[i] = IVector::create(
+          (i ? subSeqStarts : seqStarts) + offset, height + 1, false);
+
+      beamPerSeq_[j].candidateIds[i] =
+          Matrix::create(candidateInBeam_[i]->getData() + offset * beamSize_,
+                         height,
+                         beamSize_,
+                         false,
+                         false);
+      beamPerSeq_[j].gold[i] = goldSequence_[i]->getData()[j];
+
+      CHECK_LE(beamPerSeq_[j].gold[i], seqStarts[j + 1] - seqStarts[j]);
+    }
+  }
+}
+
+void CrossEntropyOverBeam::resizeOutput() {
+  Matrix::resizeOrCreate(output_.value, batchSize_, 1, false, false);
+  output_.value->zeroMem();
+
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    MatrixPtr inGrad = getInputGrad(i * 3);
+    if (dynamic_cast<GpuMatrix*>(inGrad.get())) {
+      Matrix::resizeOrCreate(candidateScoreGrad_[i],
+                             inGrad->getHeight(),
+                             inGrad->getWidth(),
+                             false,
+                             false);
+    } else {
+      candidateScoreGrad_[i] = std::move(inGrad);
+    }
+    candidateScoreGrad_[i]->zeroMem();
+  }
+}
+
+void CrossEntropyOverBeam::copyGradToGpu(size_t copyCount) {
+  for (size_t i = 0; i < beamExpanCount_; ++i) {
+    if (dynamic_cast<GpuMatrix*>(getInputGrad(i * 3).get()))
+      getInputGrad(i * 3)->copyFrom(*candidateScoreGrad_[i]);
+
+    if (i == copyCount - 1) break;
+  }
+}
+
+void CrossEntropyOverBeam::forward(PassType passType) {
+  Layer::forward(passType);
+
+  checkInputs();
+  copyInputsToCpu();
+
+  resizeOutput();
+  splitBatchBeams();
+
+  MatrixPtr outputValue = getOutputValue();
+  for (size_t i = 0; i < batchSize_; ++i) {
+    BeamExpansionPtr ptr = std::make_shared<BeamExpansion>(beamPerSeq_[i]);
+    beamCosts_[i].setData(std::move(ptr), beamSize_);
+    outputValue->getData()[i] = beamCosts_[i].forward();
+  }
+}
+
+void CrossEntropyOverBeam::backward(const UpdateCallback& callback) {
+  for (size_t i = 0; i < batchSize_; ++i) {
+    beamCosts_[i].backward();
+    copyGradToGpu(beamCosts_[i].getValidExpansionCount());
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/CrossEntropyOverBeam.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 "CrossEntropyOverBeam.h"
+#include "Layer.h"
+
+namespace paddle {
+
+/* This struct stores the beams in all search steps for a single sequence. */
+struct BeamExpansion {
+  std::vector<MatrixPtr> scores;
+  std::vector<IVectorPtr> seqInfo;
+
+  std::vector<MatrixPtr> candidateIds;
+  std::vector<int> gold;
+
+  std::vector<MatrixPtr> scoreGrad;
+
+  size_t expansionCount;
+
+  explicit BeamExpansion(int n) {
+    expansionCount = n;
+    scores.resize(expansionCount);
+    seqInfo.resize(expansionCount);
+    candidateIds.resize(expansionCount);
+    scoreGrad.resize(expansionCount);
+
+    gold.resize(expansionCount);
+  }
+};
+typedef std::shared_ptr<BeamExpansion> BeamExpansionPtr;
+
+class CostForOneSequence {
+public:
+  CostForOneSequence()
+      : beamSize_(0), validExpansionCount_(0), goldAsExtraPath_(false) {}
+  void setData(const BeamExpansionPtr bPtr, size_t beamSize) {
+    beams_ = bPtr;
+    beamSize_ = beamSize;
+
+    expandedPathScores_.clear();
+    expandedPathScores_.resize(beams_->expansionCount);
+
+    goldRowIds_.clear();
+    goldRowIds_.resize(beams_->expansionCount, 0);
+    goldColIds_.clear();
+    goldColIds_.resize(beams_->expansionCount, -1);
+  }
+  size_t getValidExpansionCount() { return validExpansionCount_; }
+
+  real forward();
+  void backward();
+
+private:
+  void calValidExpandStep();
+  void constructTotalExpansion();
+  size_t initLastExpansion();
+  real globallyNormalizedScore();
+
+  int getSeqStartPos(size_t beamId, size_t rowId) {
+    CHECK_GT(beams_->seqInfo[beamId]->getSize() - 1, rowId);
+    int* starts = beams_->seqInfo[beamId]->getData();
+    return starts[rowId] - starts[0];
+  }
+
+  size_t beamSize_;
+  size_t validExpansionCount_;
+  bool goldAsExtraPath_;
+  std::vector<int> goldRowIds_;
+  std::vector<int> goldColIds_;
+
+  BeamExpansionPtr beams_;
+  std::vector<std::vector<int>> pathRowIdsInEachBeam_;
+  std::vector<int> parentIdsInBeam_;
+  size_t goldIdsInFinalExpansion_;
+
+  std::vector<MatrixPtr> expandedPathScores_;
+
+  MatrixPtr softmaxOut_;
+};
+
+class CrossEntropyOverBeam : public Layer {
+public:
+  explicit CrossEntropyOverBeam(const LayerConfig& config) : Layer(config) {}
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
+
+private:
+  void checkInputs();
+  void copyInputsToCpu();
+  void resizeOutput();
+  void copyGradToGpu(size_t copyCount);
+  void splitBatchBeams();
+
+  size_t beamExpanCount_;
+  size_t batchSize_;
+  size_t beamSize_;
+
+  /*
+   * the process of constructing beams is not friendly to GPU, currently, this
+   * layer only runs on CPU, if any of its inputs is on GPU memory, then copy
+   * it to CPU memory.
+   */
+  std::vector<MatrixPtr> candidateScores_;
+  std::vector<MatrixPtr> candidateScoreGrad_;
+  std::vector<MatrixPtr> candidateInBeam_;
+  std::vector<MatrixPtr> gradToInputs_;
+  std::vector<IVectorPtr> goldSequence_;
+  std::vector<std::vector<int>> beamSplitPos_;
+
+  /*
+   * split entire bath of beams into beam per sequnence and store the result
+   * into this member.
+   */
+  std::vector<BeamExpansion> beamPerSeq_;
+  /* beamCosts_ is used to propagate error in one sequence. */
+  std::vector<CostForOneSequence> beamCosts_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/CudnnConvBaseLayer.cpp

@@ -46,8 +46,26 @@ bool CudnnConvBaseLayer::init(const LayerMap &layerMap,
     projConf_.emplace_back(conf);
     projections_.emplace_back(
         Projection::create(*projConf_[i], parameters_[i], useGpu_));
+
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[i] * filterChannels_[i];
+    width = (!isDeconv_) ? numFilters_ : channels_[i];
+    CHECK_EQ(parameters_[i]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[i]);
+    weights_.emplace_back(w);
   }
 
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
+    }
+  }
   if (biases_.get() && sharedBiases_) {
     hl_create_tensor_descriptor(&biasDesc_);
     hl_create_tensor_descriptor(&outputDesc_);

paddle/gserver/layers/DeConv3DLayer.cpp

+/* Copyright (c) 2016 Baidu, Inc. 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 "DeConv3DLayer.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+REGISTER_LAYER(deconv3d, DeConv3DLayer);
+
+bool DeConv3DLayer::init(const LayerMap &layerMap,
+                         const ParameterMap &parameterMap) {
+  if (!ConvBaseLayer::init(layerMap, parameterMap)) return false;
+  // for Deconv, the dimension of Kernel is
+  // channel * output * depth * height * weigth
+  // Matrix storage format: (output * depth * height * weigth) x  channel
+  for (int index = 0; index < config_.inputs().size(); ++index) {
+    M_.push_back(filterChannels_[index]);
+    K_.push_back(filterPixels_[index] * (numFilters_ / groups_[index]));
+
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[index] * numFilters_;
+    width = filterChannels_[index];
+    CHECK_EQ(parameters_[index]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[index]);
+    weights_.emplace_back(w);
+  }
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, numFilters_, biasParameter_));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_));
+    }
+  }
+  return true;
+}
+
+size_t DeConv3DLayer::getSize() {
+  CHECK_NE(inputLayers_.size(), 0UL);
+  outputH_.clear();
+  outputW_.clear();
+  outputD_.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]);
+    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]);
+  return layerSize;
+}
+
+void DeConv3DLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  int outWidth = getSize();
+  resetOutput(batchSize, outWidth);
+  const MatrixPtr outMat = getOutputValue();
+
+  for (size_t i = 0; i != inputLayers_.size(); ++i) {
+    REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
+    const MatrixPtr &inMat = getInputValue(i);
+    int M = M_[i];
+    int N = N_[i];
+    int K = K_[i];
+    MatrixPtr wMat = weights_[i]->getW();
+    Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+    for (int n = 0; n < batchSize; ++n) {
+      real *inData = inMat->getData() + n * inMat->getStride();
+      for (int g = 0; g < groups_[i]; ++g) {
+        MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_);
+        MatrixPtr wMatSub = wMat->subMatrix(g * K, K);
+        MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K);
+        colBufDataSub->mul(*wMatSub, *inMatSub, 1.0, 0.0);
+        inData += M * N;
+      }
+      colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(),
+                       numFilters_,
+                       outputD_[i],
+                       outputH_[i],
+                       outputW_[i],
+                       filterSizeZ_[i],
+                       filterSizeY_[i],
+                       filterSize_[i],
+                       strideZ_[i],
+                       strideY_[i],
+                       stride_[i],
+                       paddingZ_[i],
+                       paddingY_[i],
+                       padding_[i],
+                       1.0,
+                       1.0);
+    }
+  }
+  if (nullptr != this->biasParameter_) {
+    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
+    this->addBias();
+  }
+  forwardActivation();
+}
+
+void DeConv3DLayer::backward(const UpdateCallback &callback) {
+  backwardActivation();
+  int batchSize = getOutputGrad()->getHeight();
+  if (biases_ && biases_->getWGrad()) {
+    bpropBiases();
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+  for (size_t i = 0; i < inputLayers_.size(); ++i) {
+    if (weights_[i]->getWGrad() || this->needGradient_) {
+      int M = M_[i];
+      int N = N_[i];
+      int K = K_[i];
+      REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str());
+      Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
+      const MatrixPtr &inMat = getInputValue(i);
+      for (int n = 0; n < batchSize; ++n) {
+        colBuf_->vol2Col(
+            getOutputGrad()->getData() + n * getOutputGrad()->getStride(),
+            numFilters_,
+            outputD_[i],
+            outputH_[i],
+            outputW_[i],
+            filterSizeZ_[i],
+            filterSizeY_[i],
+            filterSize_[i],
+            strideZ_[i],
+            strideY_[i],
+            stride_[i],
+            paddingZ_[i],
+            paddingY_[i],
+            padding_[i]);
+        if (weights_[i]->getWGrad()) {
+          real *inData = inMat->getData() + n * inMat->getStride();
+          for (int g = 0; g < groups_[i]; ++g) {
+            MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K);
+            MatrixPtr wGradMatSub =
+                weights_[i]->getWGrad()->subMatrix(g * K, K);
+            MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_);
+            wGradMatSub->mul(
+                *colBufDataSub, *(inMatSub->getTranspose()), 1.0, 1.0);
+            inData += M * N;
+          }
+        }
+        if (getInputGrad(i)) {
+          real *preGrad =
+              getInputGrad(i)->getData() + n * getInputGrad(i)->getStride();
+          for (int g = 0; g < groups_[i]; ++g) {
+            MatrixPtr w = weights_[i]->getW()->subMatrix(g * K, K);
+            MatrixPtr outGradMat = colBuf_->subMatrix(g * K, K);
+            MatrixPtr inGradMatSub =
+                Matrix::create(preGrad, M, N, false, useGpu_);
+            inGradMatSub->mul(*(w->getTranspose()), *outGradMat, 1.0, 1.0);
+            preGrad += M * N;
+          }
+        }
+      }
+      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+      weights_[i]->getParameterPtr()->incUpdate(callback);
+    }
+  }
+}
+void DeConv3DLayer::bpropWeights(int i) {}
+void DeConv3DLayer::bpropData(int i) {}
+
+void DeConv3DLayer::bpropBiases() {
+  const MatrixPtr &outGradMat = getOutputGrad();
+
+  if (this->sharedBiases_) {
+    biases_->getWGrad()->collectSharedBias(*outGradMat, 1.0f);
+  } else {
+    biases_->getWGrad()->collectBias(*outGradMat, 1.0f);
+  }
+}
+
+void DeConv3DLayer::addBias() {
+  MatrixPtr outMat = getOutputValue();
+  if (this->sharedBiases_) {
+    outMat->addSharedBias(*(biases_->getW()), 1.0f);
+  } else {
+    outMat->addBias(*(biases_->getW()), 1.0f);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/DeConv3DLayer.h

+/* Copyright (c) 2016 Baidu, Inc. 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 "ConvBaseLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief A subclass of deconvolution3D layer.
+ * This layer expands input and use matrix multiplication to
+ * calculate deconvolution3D operation.
+ */
+class DeConv3DLayer : public ConvBaseLayer {
+public:
+  explicit DeConv3DLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
+  ~DeConv3DLayer() {}
+  bool init(const LayerMap& layerMap, const ParameterMap& parameterMap);
+
+  void forward(PassType passType);
+  void addBias();
+  void backward(const UpdateCallback& callback);
+  void bpropBiases();
+  void bpropData(int i);
+  void bpropWeights(int i);
+  size_t getSize();
+
+protected:
+  // Figure out the dimensions for individual gemms.
+  IntV M_;  /// numFilters_ / filter_group_;
+  IntV N_;  /// channels_ * filterSizeZ_ * filterSize_ * filterSizeY_
+  IntV K_;  /// outputD_ * outputH_ * outputW_
+  IntV NOut_;
+  MatrixPtr colBuf_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/ExpandConvBaseLayer.cpp

@@ -22,12 +22,31 @@ bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
   /* Initialize the basic convolutional parent class */
   ConvBaseLayer::init(layerMap, parameterMap);
 
+  int index = 0;
   for (auto &inputConfig : config_.inputs()) {
     const ConvConfig &conf = inputConfig.conv_conf();
     /* Consistent caffe mode for multiple input */
     caffeMode_ = conf.caffe_mode();
-  }
 
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[index] * filterChannels_[index];
+    width = (!isDeconv_) ? numFilters_ : channels_[index];
+    CHECK_EQ(parameters_[index]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[index]);
+    weights_.emplace_back(w);
+    index++;
+  }
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
+    }
+  }
   getOutputSize();
 
   return true;

paddle/gserver/layers/ExpandConvLayer.cpp

@@ -29,6 +29,10 @@ namespace paddle {
 REGISTER_LAYER(exconv, ExpandConvLayer);
 REGISTER_LAYER(exconvt, ExpandConvLayer);
 
+inline bool isDepthwiseConv(int channels, int groups) {
+  return channels == groups;
+}
+
 bool ExpandConvLayer::init(const LayerMap &layerMap,
                            const ParameterMap &parameterMap) {
   /* Initialize the basic convolutional parent class */
@@ -47,14 +51,27 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
     std::vector<size_t> paddings = {(size_t)paddingY_[i], (size_t)padding_[i]};
     std::vector<size_t> strides = {(size_t)strideY_[i], (size_t)stride_[i]};
 
-    if (useGpu_ && (size_t)groups_[i] == (size_t)channels_[i] && !isDeconv_) {
-      convType = "DepthwiseConv";
-      convGradInputType = "DepthwiseConvGradInput";
-      convGradFilterType = "DepthwiseConvGradFilter";
-    } else {
+    // Convolution Layer uses the GemmConv function by default.
     convType = "GemmConv";
     convGradInputType = "GemmConvGradInput";
     convGradFilterType = "GemmConvGradFilter";
+
+    // If depth wise convolution and useGpu == true
+    if (useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) {
+      convType = "DepthwiseConv";
+      convGradInputType = "DepthwiseConvGradInput";
+      convGradFilterType = "DepthwiseConvGradFilter";
+    }
+
+    // If depth wise convolution and useGpu == false and ARM-NEON
+    if (!useGpu_ && isDepthwiseConv(channels_[i], groups_[i]) && !isDeconv_) {
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+      if ((filterSize_[i] == filterSizeY_[i]) &&
+          (filterSize_[i] == 3 || filterSize_[i] == 4) &&
+          (stride_[i] == strideY_[i]) && (stride_[i] == 1 || stride_[i] == 2)) {
+        convType = "NeonDepthwiseConv";
+      }
+#endif
     }
 
     if (FLAGS_use_nnpack && !isDeconv_) {

paddle/gserver/layers/Layer.cpp

@@ -41,7 +41,7 @@ namespace paddle {
 Layer::Layer(const LayerConfig& config, bool useGpu)
     : config_(config),
       useGpu_(useGpu),
-      deviceId_(-1),
+      deviceId_(CPU_DEVICE),
       needSequenceInfo_(true) {}
 
 bool Layer::init(const LayerMap& layerMap, const ParameterMap& parameterMap) {

paddle/gserver/layers/Layer.h

@@ -59,7 +59,12 @@ protected:
   LayerConfig config_;
   /// whether to use GPU
   bool useGpu_;
-  /// Device Id. CPU is -1, and GPU is 0, 1, 2 ...
+  /// 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
   std::vector<LayerPtr> inputLayers_;
@@ -77,6 +82,7 @@ protected:
   Argument output_;
   /// Several outputs stored on different devices, used in 'parallel_nn' case,
   /// and record them by deviceId_.
+  /// Also used in 'use_mkldnn' case.
   std::vector<Argument> outputOtherDevice_;
   /// If there are several outputs, map them by each name.
   std::map<std::string, Argument*> outputMap_;
@@ -172,6 +178,13 @@ protected:
     return inputLayer.getOutput(deviceId_);
   }
 
+  /**
+   * Get the argument of input layer with deviceId.
+   */
+  const Argument& getInput(size_t inputIndex, int deviceId) const {
+    return inputLayers_[inputIndex]->getOutput(deviceId);
+  }
+
   /**
    * Get the forward-input value.
    */
@@ -186,6 +199,13 @@ protected:
     return inputLayer.getOutput(deviceId_).value;
   }
 
+  /**
+   * Get the forward-input value with deviceId.
+   */
+  const MatrixPtr& getInputValue(int inputIndex, int deviceId) {
+    return inputLayers_[inputIndex]->getOutput(deviceId).value;
+  }
+
   /**
    * Get the forward-input grad.
    */
@@ -200,6 +220,13 @@ protected:
     return inputLayer.getOutput(deviceId_).grad;
   }
 
+  /**
+   * Get the forward-input grad.
+   */
+  const MatrixPtr& getInputGrad(int inputIndex, int deviceId) {
+    return inputLayers_[inputIndex]->getOutput(deviceId).grad;
+  }
+
   /**
    * Get the forward-input label.
    */

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -61,43 +61,42 @@ void MKLDNNFcLayer::convertWeightsFromPaddle() {
     return;
   }
 
-  // TODO(TJ): dst format should get from wgtVal_
-  int dstFmt = PARAM_FORMAT_MKLDNN_OI;
-  int srcFmt = weight_->getParameterPtr()->getHeaderFormat();
-  if (srcFmt == dstFmt) {
-    return;
-  }
-
-  // The weight_ is transposed from initial paddle weight
-  MatrixPtr paddleWgt = Matrix::create(
-      weight_->getW()->getData(), iLayerSize_, oc_, false, false);
-
-  // TODO(TJ): remove this print when do not need differ weights
-  std::ostringstream ostr;
-  paddleWgt->print(ostr);
-  VLOG(MKLDNN_ALL) << "Initial Weight from paddle: " << std::endl << ostr.str();
-
-  // The mkldnn weight is transposed from initial paddle matrix
-  MatrixPtr paddleWgtT;
-  paddleWgt->transpose(paddleWgtT, true);
-  weight_->getW()->copyFrom(*paddleWgtT);
-  weight_->getParameterPtr()->setHeaderFormat(dstFmt);
+  CHECK(wgtVal_) << "should have been initialized";
+  bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
+  auto targetDim = wgtVal_->getDims();
+  auto srcFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
+  wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
   hasInitedWgt_ = true;
 }
 
 void MKLDNNFcLayer::convertWeightsToPaddle() {
-  MatrixPtr dnnWgt = weight_->getW();
-  MatrixPtr paddleWgt;
-  dnnWgt->transpose(paddleWgt, true);
-
-  // copy paddle weight and override on weight_
-  MatrixPtr dnnWgtT = Matrix::create(
-      dnnWgt->getData(), dnnWgt->getWidth(), dnnWgt->getHeight(), false, false);
-  dnnWgtT->copyFrom(*paddleWgt);
+  CHECK(wgtVal_) << "should have been initialized";
+  bool hasNoSpatial_ = ih_ == 1 && iw_ == 1;
+  auto targetDim = wgtVal_->getDims();
+  auto dstFmt = hasNoSpatial_ ? memory::format::io : memory::format::ihwo;
+  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;
+    }
+  }
+
+  if (cnt > 1) {
+    LOG(WARNING) << "should not have more than one CPU devie";
+  }
 }
 
 void MKLDNNFcLayer::reshape() {
-  const Argument& input = getInput(0);
+  const Argument& input = getInput(0, getPrev(0)->getDeviceId());
   int batchSize = input.getBatchSize();
   if (bs_ == batchSize) {
     return;
@@ -111,10 +110,6 @@ void MKLDNNFcLayer::reshape() {
   if (iw_ == 0) {
     iw_ = 1;
   }
-  hasSpatial_ = true;
-  if (ih_ == 1 && iw_ == 1) {
-    hasSpatial_ = false;
-  }
   CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
   ic_ = iLayerSize_ / (ih_ * iw_);
   CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
@@ -135,37 +130,53 @@ void MKLDNNFcLayer::reshape() {
 
 void MKLDNNFcLayer::resetFwd() {
   bool hasBias = biases_ && biases_->getW();
-  real* iData = getInputValue(0)->getData();
-  real* oData = getOutputValue()->getData();
-  real* wData = weight_->getW()->getData();
-  real* bData = hasBias ? biases_->getW()->getData() : NULL;
-
-  // TODO(TJ): below create should be covered in MkldnnMatrix
-  // create memory desc
-  memory::desc iMD = hasSpatial_ ? createMD({bs_, ic_, ih_, iw_}, format::nchw)
-                                 : createMD({bs_, ic_}, format::nc);
-  memory::desc wMD = hasSpatial_ ? createMD({oc_, ic_, ih_, iw_}, format::oihw)
-                                 : createMD({oc_, ic_}, format::oi);
-  memory::desc bMD = bData != NULL ? createMD({oc_}, format::x)
-                                   : createMD({}, format::format_undef);
-  memory::desc oMD = createMD({bs_, oc_}, format::nc);
-
-  // create memory primitive desc and memory self
-  inVal_.reset(new memory(memory::primitive_desc(iMD, engine_), iData));
-  wgtVal_.reset(new memory(memory::primitive_desc(wMD, engine_), wData));
-  outVal_.reset(new memory(memory::primitive_desc(oMD, engine_), oData));
+  const MatrixPtr& wgt = weight_->getW();
+  const MatrixPtr& bias = hasBias ? biases_->getW() : nullptr;
+  const MatrixPtr& out = output_.value;
+
+  if (inputIsOnlyMKLDNN()) {
+    const MatrixPtr& in = getInputValue(0);
+    inVal_ = std::dynamic_pointer_cast<MKLDNNMatrix>(in);
+    CHECK(inVal_) << "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_);
+
+  // change original output value to mkldnn output value
+  output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+  if (!outputIsOnlyMKLDNN()) {
+    convertOutputToOtherDevice();
+  }
 
+  // create forward handle
   prop_kind pk = prop_kind::forward;
-  fc_fwd::desc fwdDesc = bData != NULL ? fc_fwd::desc(pk, iMD, wMD, bMD, oMD)
-                                       : fc_fwd::desc(pk, iMD, wMD, oMD);
+  fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
+                                                inVal_->getMemoryDesc(),
+                                                wgtVal_->getMemoryDesc(),
+                                                biasVal_->getMemoryDesc(),
+                                                outVal_->getMemoryDesc())
+                                 : fc_fwd::desc(pk,
+                                                inVal_->getMemoryDesc(),
+                                                wgtVal_->getMemoryDesc(),
+                                                outVal_->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-
-  if (bData != NULL) {
-    biasVal_.reset(new memory(memory::primitive_desc(bMD, engine_), bData));
+  if (hasBias) {
     fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_));
   } else {
     fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_));
   }
+  printValueFormatFlow();
+
   pipelineFwd_.clear();
   pipelineFwd_.push_back(*fwd_);
 }
@@ -175,45 +186,46 @@ void MKLDNNFcLayer::resetBwd() {
     return;
   }
   needResetBwd_ = false;
-
   bool hasBias = biases_ && biases_->getWGrad();
-  real* iData = getInputValue(0)->getData();
-  real* iDiff = getInputGrad(0) != nullptr ? getInputGrad(0)->getData() : NULL;
-  real* oDiff = getOutputGrad()->getData();
-  real* wDiff = weight_->getWGrad()->getData();
-  real* bDiff = hasBias ? biases_->getWGrad()->getData() : NULL;
 
   /// backward weight
-  // create memory desc for backward memory
-  memory::desc iMD = hasSpatial_ ? createMD({bs_, ic_, ih_, iw_}, format::nchw)
-                                 : createMD({bs_, ic_}, format::nc);
-  memory::desc wMD = hasSpatial_ ? createMD({oc_, ic_, ih_, iw_}, format::oihw)
-                                 : createMD({oc_, ic_}, format::oi);
-  memory::desc oMD = createMD({bs_, oc_}, format::nc);
-  memory::desc bMD = bDiff != NULL ? createMD({oc_}, format::x)
-                                   : createMD({}, format::format_undef);
-
-  if (inVal_) {
-    // update data
-    inVal_->set_data_handle(iData);
-  } else {
-    inVal_.reset(new memory(memory::primitive_desc(iMD, engine_), iData));
-  }
-
-  // create memory primitive desc and memory self
-  wgtGrad_.reset(new memory(memory::primitive_desc(wMD, engine_), wDiff));
-  outGrad_.reset(new memory(memory::primitive_desc(oMD, engine_), oDiff));
-
-  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward, iMD, wMD, oMD);
+  CHECK(inVal_) << "Should have input value";
+  const MatrixPtr& wgt = weight_->getWGrad();
+  const MatrixPtr& bias = hasBias ? biases_->getWGrad() : nullptr;
+
+  // TODO(TJ): merge outgrad
+  int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
+  // for MKLDNN device:
+  // can not directly cast outputgrad to mkldnnmatrix,
+  // since each layer can not write the inputgrad to mkldnn inputgrad.
+  // So just create from matrix with outputvalue format.
+  // 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())
+                      : nullptr;
+
+  // create memory primitive desc
+  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
+                                      inVal_->getMemoryDesc(),
+                                      wgtGrad_->getMemoryDesc(),
+                                      outGrad_->getMemoryDesc());
   fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  fc_bwdWgt::desc bwdWgtDesc = bDiff != NULL
-                                   ? fc_bwdWgt::desc(iMD, wMD, bMD, oMD)
-                                   : fc_bwdWgt::desc(iMD, wMD, oMD);
+  fc_bwdWgt::desc bwdWgtDesc = hasBias
+                                   ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                     wgtGrad_->getMemoryDesc(),
+                                                     biasGrad_->getMemoryDesc(),
+                                                     outGrad_->getMemoryDesc())
+                                   : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                     wgtGrad_->getMemoryDesc(),
+                                                     outGrad_->getMemoryDesc());
   fc_bwdWgt::primitive_desc bwdWgtPD =
       fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
 
-  if (bDiff != NULL) {
-    biasGrad_.reset(new memory(memory::primitive_desc(bMD, engine_), bDiff));
+  if (hasBias) {
     bwdWgt_.reset(
         new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_));
   } else {
@@ -223,15 +235,26 @@ void MKLDNNFcLayer::resetBwd() {
   pipelineBwd_.push_back(*bwdWgt_);
 
   /// backward data
-  if (iDiff == NULL) {
+  device = inputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
+  const MatrixPtr& in = getInputGrad(0, device);
+  if (in == nullptr) {
     return;
   }
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(iMD, wMD, oMD);
+  if (getInput(0, device).getAllCount() > 1) {
+    // TODO(TJ): use outputMaps_ ways when merge outgrad done
+  } else {
+    inGrad_ = MKLDNNMatrix::create(in, inVal_->getPrimitiveDesc());
+  }
+
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(inVal_->getMemoryDesc(),
+                                                  wgtGrad_->getMemoryDesc(),
+                                                  outGrad_->getMemoryDesc());
   fc_bwdData::primitive_desc bwdDataPD =
       fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
-  inGrad_.reset(new memory(memory::primitive_desc(iMD, engine_), iDiff));
+
   CHECK(wgtVal_) << "Should have weight memory";
   bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_));
+  printGradFormatFlow();
   pipelineBwd_.push_back(*bwdData_);
 }
 
@@ -241,11 +264,7 @@ void MKLDNNFcLayer::forward(PassType passType) {
 
   {
     REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
-
-    // update input data
-    // since it might be changed if this is after data layer
-    real* iData = getInputValue(0)->getData();
-    inVal_->set_data_handle(iData);
+    syncInputValue();
 
     // just submit forward pipeline
     stream_->submit(pipelineFwd_);
@@ -267,10 +286,7 @@ void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
     REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
     resetBwd();
 
-    // update diff
-    real* oDiff = getOutputGrad()->getData();
-    outGrad_->set_data_handle(oDiff);
-
+    syncOutputGrad();
     // just sumbmit backward pipeline
     stream_->submit(pipelineBwd_);
   }

paddle/gserver/layers/MKLDNNFcLayer.h

@@ -32,16 +32,13 @@ protected:
   // if has already init the weight
   bool hasInitedWgt_;
 
-  // if input layer has image size info (ih>1 && iw>1)
-  bool hasSpatial_;
-
   // fc weight and bias
   std::unique_ptr<Weight> weight_;
   std::unique_ptr<Weight> biases_;
 
 public:
   explicit MKLDNNFcLayer(const LayerConfig& config)
-      : MKLDNNLayer(config), hasInitedWgt_(false), hasSpatial_(true) {}
+      : MKLDNNLayer(config), hasInitedWgt_(false) {}
 
   ~MKLDNNFcLayer() {}
 
@@ -75,6 +72,8 @@ protected:
    * only would be called when needed
    */
   void resetBwd();
+
+  void convertOutputToOtherDevice() override;
 };
 
 }  // namespace paddle

paddle/gserver/layers/Pool3DLayer.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 "Layer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/math/Matrix.h"
+
+namespace paddle {
+
+/**
+ * @brief Basic parent layer of pooling
+ * Pools the input within regions
+ */
+class Pool3DLayer : public Layer {
+public:
+  explicit Pool3DLayer(const LayerConfig& config) : Layer(config) {}
+  ~Pool3DLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback) override;
+  size_t getSize();
+
+protected:
+  int channels_;
+  int sizeX_, sizeY_, sizeZ_;
+  int strideW_, strideH_, strideD_;
+  int paddingW_, paddingH_, paddingD_;
+  int imgSizeW_, imgSizeH_, imgSizeD_;
+  int outputW_, outputH_, outputD_;
+  std::string poolType_;
+  MatrixPtr maxPoolIdx_;
+};
+}  // namespace paddle

paddle/gserver/layers/PrintLayer.cpp

@@ -48,7 +48,16 @@ public:
                  << inputLayers_.size() << ") at " << getName();
     }
     s << format.substr(pos);
-    LOG(INFO) << s.str();
+
+    const std::string delimiter("\n");
+    std::string content = s.str();
+    std::string::size_type foundPos = 0;
+    std::string::size_type prevPos = 0;
+    while ((foundPos = content.find(delimiter, prevPos)) != std::string::npos) {
+      LOG(INFO) << content.substr(prevPos, foundPos - prevPos);
+      prevPos = foundPos + delimiter.size();
+    }
+    LOG(INFO) << content.substr(prevPos);
   }
 
   void backward(const UpdateCallback& callback) override {}

paddle/gserver/tests/CMakeLists.txt

@@ -34,6 +34,13 @@ add_unittest_without_exec(test_CRFLayerGrad
 add_test(NAME test_CRFLayerGrad
     COMMAND test_CRFLayerGrad)
 
+################ test_CrossEntropyOverBeam ####################
+add_unittest_without_exec(test_CrossEntropyOverBeam
+    test_CrossEntropyOverBeamGrad.cpp
+    LayerGradUtil.cpp)
+add_test(NAME test_CrossEntropyOverBeam
+    COMMAND test_CrossEntropyOverBeam)
+
 ################ test_SeqSliceLayerGrad ####################
 add_unittest_without_exec(test_SeqSliceLayerGrad
     test_SeqSliceLayerGrad.cpp

paddle/math/Allocator.h

@@ -48,7 +48,13 @@ public:
    */
   virtual void* alloc(size_t size) {
     void* ptr;
+#ifdef PADDLE_USE_MKLDNN
+    // refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
+    // memory alignment
+    CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0);
+#else
     CHECK_EQ(posix_memalign(&ptr, 32ul, size), 0);
+#endif
     CHECK(ptr) << "Fail to allocate CPU memory: size=" << size;
     return ptr;
   }

paddle/math/CMakeLists.txt

@@ -14,6 +14,17 @@
 #
 file(GLOB MATH_HEADERS . *.h)
 file(GLOB MATH_SOURCES . *.cpp)
+
+if(NOT WITH_MKLDNN)
+    set(DNN_HEADER "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.h")
+    set(DNN_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/MKLDNNMatrix.cpp")
+    list(REMOVE_ITEM MATH_HEADERS "${DNN_HEADER}")
+    list(REMOVE_ITEM MATH_SOURCES "${DNN_SOURCE}")
+    message(STATUS "Skip compiling with MKLDNNMatrix")
+else()
+    message(STATUS "Compile with MKLDNNMatrix")
+endif()
+
 set(MATH_SOURCES
     "${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu"
     "${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu"

paddle/parameter/Parameter.h

@@ -281,7 +281,11 @@ public:
   /**
    * @brief Set the format in header.
    */
-  void setHeaderFormat(int32_t fmt) { headerFormat_ = fmt; }
+  void setHeaderFormat(int32_t fmt) {
+    CHECK(isHeaderFormatSupported(fmt)) << "Unsupported format version: "
+                                        << fmt;
+    headerFormat_ = fmt;
+  }
 
   /**
    * @brief  Parameter Update Hook.