Merge branch 'develop' into build_ios

.pre-commit-config.yaml

@@ -22,7 +22,7 @@
     -   id: clang-format-with-version-check
         name: clang-format
         description: Format files with ClangFormat.
-        entry: ./.clang_format.hook -i
+        entry: bash ./.clang_format.hook -i
         language: system
         files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
 -   repo: https://github.com/PaddlePaddle/pre-commit-golang

CMakeLists.txt

@@ -55,6 +55,7 @@ option(WITH_C_API       "Compile PaddlePaddle with C-API(Prediction)"   OFF)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
+option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 
 # CMAKE_BUILD_TYPE
 if(NOT CMAKE_BUILD_TYPE)

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/configure.cmake

@@ -28,6 +28,10 @@ if(NOT WITH_TIMER)
     add_definitions(-DPADDLE_DISABLE_TIMER)
 endif(NOT WITH_TIMER)
 
+if(USE_EIGEN_FOR_BLAS)
+    add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
+endif(USE_EIGEN_FOR_BLAS)
+
 if(NOT WITH_PROFILER)
     add_definitions(-DPADDLE_DISABLE_PROFILER)
 endif(NOT WITH_PROFILER)

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")

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

@@ -257,6 +257,11 @@ seq_concat
 ..  autoclass:: paddle.v2.layer.seq_concat
     :noindex:
 
+seq_slice
+---------
+..  autoclass:: paddle.v2.layer.seq_slice
+    :noindex:
+
 kmax_sequence_score
 -------------------
 ..  autoclass:: paddle.v2.layer.kmax_sequence_score
@@ -362,6 +367,11 @@ trans
 ..  autoclass:: paddle.v2.layer.trans
     :noindex:
 
+scale_shift
+-----------
+..  autoclass:: paddle.v2.layer.scale_shift
+    :noindex:
+
 Sampling Layers
 ===============
 
@@ -409,9 +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/design/cluster_train/README.md

@@ -54,17 +54,18 @@ The life cycle of a single task is illustrated below:
 <img src="src/paddle-task-states.png"/>
 
 1. When a new pass of training starts, all tasks will be placed in the todo queue.
-1. The master server will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion.
-1. The trainer will work on its tasks and tell the master server once a task is completed. The master server will dispatch a new task to that trainer.
-1. If a task timeout. the master server will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded.
+1. Upon trainer requests for new task, the master server will dispatch a task from todo queue to it, put the task in the pending queue and wait for completion.
+1. The trainer will work on its task and tell the master server once the task is completed and ask for new task. The master server will dispatch a new task to that trainer.
+1. If a task fails for any reason in trainer, or takes longer than a specific period of time,  the master server will move the task back to the todo queue. The timeout count for that task will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, then it will be discarded.
 1. The master server will move completed task to the done queue. When the todo queue is empty, the master server will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero.
 
 ### Trainer Process
 
 The trainer process will:
 
-- Receive tasks from the master.
-- Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
+- Request tasks from the master.
+- Work on the tasks
+- Upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
 
 ### Parameter Server Process
 
@@ -119,8 +120,8 @@ When the master is started by the Kubernetes, it executes the following steps at
 
 1. Grabs a unique *master* lock in etcd, which prevents concurrent master instantiations.
 1. Recovers the task queues from etcd if they already exist, otherwise, the master will create them.
-1. Watches the trainer prefix keys `/trainer/` on etcd to find the live trainers.
-1. Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update.
+1. Write its ip address to */master/addr* so that trainers can discover it.
+1. Listens to trainers' request of task, dispatch one upon request, and updates task queue using an etcd transaction to ensure lock is held during the update.
 
 When the master server process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes.
 
@@ -128,13 +129,11 @@ When the master server process is dead for any reason, Kubernetes will restart i
 
 When the trainer is started by the Kubernetes, it executes the following steps at startup:
 
-1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count.
-1. Generates a unique ID, and sets key `/trainer/<unique ID>` with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline.
-1. Waits for tasks from the master to start training.
+1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count */ps_desired*.
+1. Finds and watches */master/addr* to get master's address.
+1. Requests for tasks from the master to start training.
 
-If trainer's etcd lease expires, it will try set key `/trainer/<unique ID>` again so that the master server can discover the trainer again.
-
-When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from the TODO queue and go on training.
+When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from master and go on training.
 
 ### Parameter Server Process
 

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

go/master/client.go

@@ -63,13 +63,24 @@ func WithAddr(addr string) func(c *Client) error {
 // WithEtcd sets the client to use etcd for master discovery.
 func WithEtcd(endpoints []string, timeout time.Duration) func(*Client) error {
 	return func(c *Client) error {
-		cli, err := clientv3.New(clientv3.Config{
+		var cli *clientv3.Client
+		f := func() error {
+			var err error
+			cli, err = clientv3.New(clientv3.Config{
 				Endpoints:   endpoints,
 				DialTimeout: timeout,
 			})
-		if err != nil {
 			return err
 		}
+		for {
+			err := f()
+			if err != nil {
+				log.Warningln(err)
+			} else {
+				break
+			}
+			time.Sleep(time.Second)
+		}
 
 		ch := make(chan string, 1)
 		a, err := GetKey(cli, DefaultAddrPath, timeout)
@@ -101,9 +112,6 @@ func NewClient(opts ...func(*Client) error) (*Client, error) {
 		}
 	}
 	c.ch = make(chan record, c.bufSize)
-	// FIXME: connection is created asyncrosly in monitorMaster go routine,
-	//        ensure the connection is ready for use before calling c.addClient.
-	time.Sleep(time.Second)
 	return c, nil
 }
 

paddle/CMakeLists.txt

@@ -19,6 +19,7 @@ if(Boost_FOUND)
   add_subdirectory(platform)
   add_subdirectory(framework)
   add_subdirectory(operators)
+  add_subdirectory(pybind)
 endif()
 
 if(WITH_C_API)

paddle/capi/CMakeLists.txt

@@ -28,7 +28,6 @@ add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER}
 
 add_dependencies(paddle_capi paddle_proto)
 
-
 # combine all paddle static libraries together, into libpaddle_capi_whole.a
 # user should use PaddleCAPI as -lpaddle_capi_whole
 set(PADDLE_INFER_LIBS
@@ -47,7 +46,10 @@ cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_INFER_LIBS})
 
 # No shared library for iOS
 if(NOT IOS)
+  set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
+  # TODO: merge mkl into paddle_capi_shared
   add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
+  set_target_properties(paddle_capi_shared	PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
   target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
   link_paddle_exe(paddle_capi_shared)
 endif()

paddle/capi/export.map

+{
+	global:
+		paddle_*;
+	local:
+		*;
+};

paddle/cuda/include/hl_cuda_cudnn.h

@@ -214,7 +214,8 @@ extern void hl_conv_workspace(hl_tensor_descriptor input,
                               int* convBwdDataAlgo,
                               size_t* bwdDataLimitBytes,
                               int* convBwdFilterAlgo,
-                              size_t* bwdFilterLimitBytes);
+                              size_t* bwdFilterLimitBytes,
+                              bool useDilation);
 
 /**
  * @brief   destroy filter descriptor.
@@ -242,7 +243,9 @@ extern void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                              int padding_height,
                                              int padding_width,
                                              int stride_height,
-                                             int stride_width);
+                                             int stride_width,
+                                             int dilation_h = 1,
+                                             int dilation_w = 1);
 
 /**
  * @brief   reset convolution descriptor.
@@ -262,7 +265,9 @@ extern void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                             int padding_height,
                                             int padding_width,
                                             int stride_height,
-                                            int stride_width);
+                                            int stride_width,
+                                            int dilation_h = 1,
+                                            int dilation_w = 1);
 
 /**
  * @brief   destroy convolution descriptor.

paddle/cuda/include/stub/hl_cuda_cudnn_stub.h

@@ -78,7 +78,9 @@ inline void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                              int padding_height,
                                              int padding_width,
                                              int stride_height,
-                                             int stride_width) {}
+                                             int stride_width,
+                                             int dilation_h,
+                                             int dilation_w) {}
 
 inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                             hl_tensor_descriptor image,
@@ -86,7 +88,9 @@ inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                             int padding_height,
                                             int padding_width,
                                             int stride_height,
-                                            int stride_width) {}
+                                            int stride_width,
+                                            int dilation_h,
+                                            int dilation_w) {}
 
 inline void hl_destroy_convolution_descriptor(hl_convolution_descriptor conv) {}
 
@@ -99,7 +103,8 @@ inline void hl_conv_workspace(hl_tensor_descriptor input,
                               int* convBwdDataAlgo,
                               size_t* bwdDataLimitBytes,
                               int* convBwdFilterAlgo,
-                              size_t* bwdFilterLimitBytes) {}
+                              size_t* bwdFilterLimitBytes,
+                              bool useDilation) {}
 
 inline void hl_convolution_forward(hl_tensor_descriptor input,
                                    real* input_data,

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -201,7 +201,8 @@ void hl_conv_workspace(hl_tensor_descriptor input,
                        int* convBwdDataAlgo,
                        size_t* bwdDataLimitBytes,
                        int* convBwdFilterAlgo,
-                       size_t* bwdFilterLimitBytes) {
+                       size_t* bwdFilterLimitBytes,
+                       bool useDilation) {
 #if CUDNN_VERSION >= 4000
 
   CHECK_NOTNULL(input);
@@ -213,12 +214,32 @@ void hl_conv_workspace(hl_tensor_descriptor input,
   size_t memoryLimitBytes =
       (1LL << 20) * FLAGS_cudnn_conv_workspace_limit_in_mb;
 
+  // For dilation
+  int algo = 0;
+
   // cudnn convolution forward configuration
   cudnnTensorDescriptor_t fwd_src_desc = GET_TENSOR_DESCRIPTOR(input);
   cudnnTensorDescriptor_t fwd_dest_desc = GET_TENSOR_DESCRIPTOR(output);
   cudnnFilterDescriptor_t fwd_filter_desc = GET_FILTER_DESCRIPTOR(filter);
   cudnnConvolutionDescriptor_t fwd_conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv);
+  // cudnn convolution backward data configuration
+  cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter);
+  cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output);
+  cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input);
+  cudnnConvolutionDescriptor_t bwd_data_conv_desc =
+      GET_CONVOLUTION_DESCRIPTOR(conv);
+  // cudnn convolution backward filter configuration
+  cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input);
+  cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output);
+  cudnnConvolutionDescriptor_t bwd_filter_conv_desc =
+      GET_CONVOLUTION_DESCRIPTOR(conv);
+  cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter);
 
+  if (useDilation) {
+    convFwdAlgo = &algo;
+    convBwdDataAlgo = &algo;
+    convBwdFilterAlgo = &algo;
+  } else {
     CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm(
         t_resource.cudnn_handle,
         fwd_src_desc,
@@ -228,23 +249,6 @@ void hl_conv_workspace(hl_tensor_descriptor input,
         CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
         memoryLimitBytes,
         reinterpret_cast<cudnnConvolutionFwdAlgo_t*>(convFwdAlgo)));
-
-  CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize(
-      t_resource.cudnn_handle,
-      fwd_src_desc,
-      fwd_filter_desc,
-      fwd_conv_desc,
-      fwd_dest_desc,
-      static_cast<cudnnConvolutionFwdAlgo_t>(*convFwdAlgo),
-      fwdLimitBytes));
-
-  // cudnn convolution backward data configuration
-  cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter);
-  cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output);
-  cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input);
-  cudnnConvolutionDescriptor_t bwd_data_conv_desc =
-      GET_CONVOLUTION_DESCRIPTOR(conv);
-
     CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm(
         t_resource.cudnn_handle,
         bwd_data_filter_desc,
@@ -254,23 +258,6 @@ void hl_conv_workspace(hl_tensor_descriptor input,
         CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
         memoryLimitBytes,
         reinterpret_cast<cudnnConvolutionBwdDataAlgo_t*>(convBwdDataAlgo)));
-
-  CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
-      t_resource.cudnn_handle,
-      bwd_data_filter_desc,
-      bwd_data_diff_desc,
-      bwd_data_conv_desc,
-      bwd_data_grad_desc,
-      static_cast<cudnnConvolutionBwdDataAlgo_t>(*convBwdDataAlgo),
-      bwdDataLimitBytes));
-
-  // cudnn convolution backward filter configuration
-  cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input);
-  cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output);
-  cudnnConvolutionDescriptor_t bwd_filter_conv_desc =
-      GET_CONVOLUTION_DESCRIPTOR(conv);
-  cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter);
-
     CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
         t_resource.cudnn_handle,
         bwd_filter_src_desc,
@@ -280,6 +267,25 @@ void hl_conv_workspace(hl_tensor_descriptor input,
         CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
         memoryLimitBytes,
         reinterpret_cast<cudnnConvolutionBwdFilterAlgo_t*>(convBwdFilterAlgo)));
+  }
+
+  CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize(
+      t_resource.cudnn_handle,
+      fwd_src_desc,
+      fwd_filter_desc,
+      fwd_conv_desc,
+      fwd_dest_desc,
+      static_cast<cudnnConvolutionFwdAlgo_t>(*convFwdAlgo),
+      fwdLimitBytes));
+
+  CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
+      t_resource.cudnn_handle,
+      bwd_data_filter_desc,
+      bwd_data_diff_desc,
+      bwd_data_conv_desc,
+      bwd_data_grad_desc,
+      static_cast<cudnnConvolutionBwdDataAlgo_t>(*convBwdDataAlgo),
+      bwdDataLimitBytes));
 
   CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
       t_resource.cudnn_handle,
@@ -603,7 +609,9 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                       int padding_height,
                                       int padding_width,
                                       int stride_height,
-                                      int stride_width) {
+                                      int stride_width,
+                                      int dilation_h,
+                                      int dilation_w) {
   CHECK_NOTNULL(conv);
 
   cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)malloc(
@@ -625,18 +633,24 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                                        padding_width,
                                                        stride_height,
                                                        stride_width,
-                                                       1,
-                                                       1,
+                                                       dilation_h,
+                                                       dilation_w,
                                                        mode,
                                                        data_type));
 #else
+  if (dilation_h > 1 || dilation_w > 1) {
+    LOG(FATAL)
+        << "Current cuDNN version does't support for dilation convolution. "
+        << "The dilation convolution requires cuDNN >= v6.0.";
+  }
+
   CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc,
                                                        padding_height,
                                                        padding_width,
                                                        stride_height,
                                                        stride_width,
-                                                       1,
-                                                       1,
+                                                       dilation_h,
+                                                       dilation_w,
                                                        mode));
 #endif
 
@@ -659,7 +673,9 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                      int padding_height,
                                      int padding_width,
                                      int stride_height,
-                                     int stride_width) {
+                                     int stride_width,
+                                     int dilation_h,
+                                     int dilation_w) {
   CHECK_NOTNULL(conv);
   CHECK_NOTNULL(image);
   CHECK_NOTNULL(filter);
@@ -678,8 +694,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                                        padding_width,
                                                        stride_height,
                                                        stride_width,
-                                                       1,
-                                                       1,
+                                                       dilation_h,
+                                                       dilation_w,
                                                        mode,
                                                        data_type));
 #else
@@ -688,8 +704,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                                        padding_width,
                                                        stride_height,
                                                        stride_width,
-                                                       1,
-                                                       1,
+                                                       dilation_h,
+                                                       dilation_w,
                                                        mode));
 #endif
 

paddle/framework/CMakeLists.txt

@@ -18,8 +18,8 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope)
 proto_library(framework_proto SRCS framework.proto)
 
 cc_library(attribute SRCS attribute.cc DEPS framework_proto)
-
-cc_library(operator SRCS operator.cc DEPS framework_proto device_context tensor scope attribute)
+cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
+cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
 cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator)
@@ -39,21 +39,3 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
 
 cc_library(backward SRCS backward.cc DEPS net_op)
 cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context)
-
-if(WITH_PYTHON)
-cc_library(paddle_pybind SHARED
-    SRCS pybind.cc
-    DEPS pybind python backward
-    sgd_op
-    add_op
-    mul_op
-    rowwise_add_op
-    sigmoid_op
-    softmax_op
-    mean_op
-    cross_entropy_op
-    recurrent_op
-    uniform_random_op
-    gaussian_random_op
-    fill_zeros_like_op)
-endif(WITH_PYTHON)

paddle/framework/backward.cc

@@ -110,7 +110,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
                        dup_output_ops[out].emplace_back(local_op_id);
                        return false;
                      });
-      net->AddOp(std::move(bwd));
+      net->AppendOp(std::move(bwd));
     }
     // Get unique ID for this method.
     auto uid = uniq_id++;
@@ -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;
@@ -163,7 +166,8 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
 
         // If part of input gradient of that operator is not calculated, fill
         // zero variables to that input gradient.
-        net->AddOp(OpRegistry::CreateOp("fill_zeros_like", {{"Src", {prefix}}},
+        net->AppendOp(OpRegistry::CreateOp("fill_zeros_like",
+                                           {{"Src", {prefix}}},
                                            {{"Dst", {grad_input}}}, {}));
       }
       return false;
@@ -195,7 +199,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
     if (net->ops_.empty()) {  // Current no aux op is added to network
       return grad_op;
     }
-    net->AddOp(std::move(grad_op));
+    net->AppendOp(std::move(grad_op));
   }
   net->SetType("@GENERATED_BACKWARD@");
   net->CompleteAddOp();
@@ -208,7 +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/backward_test.cc

@@ -72,16 +72,16 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
 
 class FcOp : public operators::NetOp {
  public:
-  FcOp(const std::string &type, const VarNameMap &inputs,
-       const VarNameMap &outputs, const AttributeMap &attrs)
+  FcOp(const std::string &type, const VariableNameMap &inputs,
+       const VariableNameMap &outputs, const AttributeMap &attrs)
       : NetOp(type, inputs, outputs, attrs) {
-    AddOp(OpRegistry::CreateOp("mul",
+    AppendOp(OpRegistry::CreateOp("mul",
                                   {{"X", {Input("X")}}, {"Y", {Input("W")}}},
                                   {{"Out", {Output("mul_result")}}}, {}));
     auto input_b = Inputs("b");
     std::string before_act = "mul_result";
     if (input_b.size() != 0) {
-      AddOp(OpRegistry::CreateOp(
+      AppendOp(OpRegistry::CreateOp(
           "rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}},
           {{"Out", {Output("add_result")}}}, {}));
       before_act = "add_result";
@@ -92,7 +92,7 @@ class FcOp : public operators::NetOp {
       }
     }
 
-    AddOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
+    AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
                                   {{"Out", {Output("Out")}}}, {}));
     CompleteAddOp(false);
   }
@@ -234,13 +234,13 @@ TEST(Backward, net_fc_backward_not_have_b) {
 
 TEST(Backward, net_input_of_network_not_need_grad) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}},
       {{"mul_result", {"mul_tmp_0"}},
        {"add_result", {"add_tmp_0"}},
        {"Out", {"hidden0"}}},
       {}));
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}},
       {{"mul_result", {"mul_tmp_1"}},
        {"add_result", {"add_tmp_1"}},
@@ -273,9 +273,9 @@ TEST(Backward, net_input_of_network_not_need_grad) {
 
 TEST(Backward, net_shared_weight) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
+  net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
                                        {{"Out", {"out"}}}, {}));
-  net.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
+  net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
                                        {{"Out", {"FinalOut"}}}, {}));
   net.CompleteAddOp();
 
@@ -357,19 +357,19 @@ TEST(Backward, op_part_of_input_are_not_need) {
 
 TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}},
       {{"mul_result", {"mul_out1"}},
        {"add_result", {"add_out1"}},
        {"Out", {"out1"}}},
       {}));
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}},
       {{"mul_result", {"mul_out2"}},
        {"add_result", {"tmp_out2"}},
        {"Out", {"out2"}}},
       {}));
-  net.AddOp(f::OpRegistry::CreateOp(
+  net.AppendOp(f::OpRegistry::CreateOp(
       "fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}},
       {{"mul_result", {"mul_out3"}},
        {"add_result", {"tmp_out3"}},

paddle/framework/grad_op_builder.cc

@@ -20,13 +20,13 @@ namespace framework {
 enum class OpArgType { IN, OUT };
 
 static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
-                       bool is_grad, OperatorBase::VarNameMap* vars) {
+                       bool is_grad, VariableNameMap* vars) {
   const auto& src_inout =
       src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
   auto& dst_inout = *vars;
-  const OpProto* proto = OpRegistry::op_info_map().at(src_op->Type()).proto_;
+  auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto();
   const auto& src_arg_list =
-      src_type == OpArgType::IN ? proto->inputs() : proto->outputs();
+      src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
   for (const auto& arg : src_arg_list) {
     if (arg.not_in_gradient() && !is_grad) continue;
     const std::string src_name = arg.name();
@@ -40,26 +40,18 @@ static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
 }
 
 OperatorBase* BuildGradOp(const OperatorBase* op) {
-  auto it = OpRegistry::op_info_map().find(op->Type());
-  PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
-                 "'%s' has not been registered.", op->Type());
-  PADDLE_ENFORCE(it->second.proto_ != nullptr, "'%s' has no OpProto.",
-                 op->Type());
-  std::string grad_op_type = it->second.grad_op_type_;
-  PADDLE_ENFORCE(!grad_op_type.empty(), "'%s' has no gradient operator.",
-                 op->Type());
+  auto& info = OpInfoMap::Instance().Get(op->Type());
+  PADDLE_ENFORCE(info.HasGradientOp());
 
-  OperatorBase::VarNameMap inputs;
-  OperatorBase::VarNameMap outputs;
+  VariableNameMap inputs;
+  VariableNameMap outputs;
   TransOpArg(op, OpArgType::IN, false, &inputs);   // I
   TransOpArg(op, OpArgType::OUT, false, &inputs);  // O
   TransOpArg(op, OpArgType::OUT, true, &inputs);   // OG
   TransOpArg(op, OpArgType::IN, true, &outputs);   // IG
 
-  it = OpRegistry::op_info_map().find(grad_op_type);
-  PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
-                 "'%s' has not been registered.", grad_op_type);
-  return it->second.creator_(grad_op_type, inputs, outputs, op->Attrs());
+  auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_);
+  return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs());
 }
 
 }  // namespace framework

paddle/framework/op_info.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/op_info.h"
+
+namespace paddle {
+namespace framework {
+
+static OpInfoMap* g_op_info_map = nullptr;
+
+OpInfoMap& OpInfoMap::Instance() {
+  if (g_op_info_map == nullptr) {
+    g_op_info_map = new OpInfoMap();
+  }
+  return *g_op_info_map;
+}
+}  // namespace framework
+}  // namespace paddle

paddle/framework/op_info.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 <functional>
+#include <map>
+#include <string>
+#include <unordered_map>
+
+#include "paddle/framework/attribute.h"
+
+namespace paddle {
+namespace framework {
+class OperatorBase;
+using VariableNameMap = std::map<std::string, std::vector<std::string>>;
+
+using OpCreator = std::function<OperatorBase*(
+    const std::string& /*type*/, const VariableNameMap& /*inputs*/,
+    const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
+
+struct OpInfo {
+  OpCreator creator_;
+  std::string grad_op_type_;
+  OpProto* proto_;
+  OpAttrChecker* checker_;
+
+  bool HasOpProtoAndChecker() const {
+    return proto_ != nullptr && checker_ != nullptr;
+  }
+
+  const OpProto& Proto() const {
+    PADDLE_ENFORCE_NOT_NULL(proto_, "Operator Proto has not been registered");
+    PADDLE_ENFORCE(proto_->IsInitialized(),
+                   "Operator Proto must be initialized in op info");
+    return *proto_;
+  }
+
+  const OpAttrChecker& Checker() const {
+    PADDLE_ENFORCE_NOT_NULL(checker_,
+                            "Operator Checker has not been registered");
+    return *checker_;
+  }
+
+  const OpCreator& Creator() const {
+    PADDLE_ENFORCE_NOT_NULL(creator_,
+                            "Operator Creator has not been registered");
+    return creator_;
+  }
+
+  bool HasGradientOp() const { return !grad_op_type_.empty(); }
+};
+
+class OpInfoMap {
+ public:
+  static OpInfoMap& Instance();
+
+  OpInfoMap(const OpInfoMap& o) = delete;
+  OpInfoMap(OpInfoMap&& o) = delete;
+  OpInfoMap& operator=(const OpInfoMap& o) = delete;
+  OpInfoMap& operator=(OpInfoMap&& o) = delete;
+
+  bool Has(const std::string& op_type) const {
+    return map_.find(op_type) != map_.end();
+  }
+
+  void Insert(const std::string& type, const OpInfo& info) {
+    PADDLE_ENFORCE(!Has(type), "Operator %s has been registered", type);
+    map_.insert({type, info});
+  }
+
+  const OpInfo& Get(const std::string& type) const {
+    auto it = map_.find(type);
+    PADDLE_ENFORCE(it != map_.end(), "Operator %s are not found", type);
+    return it->second;
+  }
+
+  template <typename Callback>
+  void IterAllInfo(Callback callback) {
+    for (auto& it : map_) {
+      callback(it.first, it.second);
+    }
+  }
+
+ private:
+  OpInfoMap() = default;
+  std::unordered_map<std::string, const OpInfo> map_;
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/op_registry.cc

@@ -19,32 +19,18 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
-std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
-                                                   const VarNameMap& inputs,
-                                                   const VarNameMap& outputs,
-                                                   AttributeMap attrs) {
-  auto it = op_info_map().find(type);
-  PADDLE_ENFORCE(it != op_info_map().end(),
-                 "Operator '%s' has not been registered.", type);
-  it->second.checker_->Check(attrs);
-  auto op = it->second.creator_(type, inputs, outputs, attrs);
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
+    const std::string& type, const VariableNameMap& inputs,
+    const VariableNameMap& outputs, AttributeMap attrs) {
+  auto& info = OpInfoMap::Instance().Get(type);
+  info.Checker().Check(attrs);
+  auto op = info.Creator()(type, inputs, outputs, attrs);
   return std::unique_ptr<OperatorBase>(op);
 }
 
-std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
-  VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
-  VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
-  AttributeMap attrs;
-  for (auto& attr : op_desc.attrs()) {
-    attrs[attr.name()] = GetAttrValue(attr);
-  }
-
-  return CreateOp(op_desc.type(), inputs, outputs, attrs);
-}
-
-OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
+static VariableNameMap ConvertOpDescVarsToVarNameMap(
     const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
-  VarNameMap ret_val;
+  VariableNameMap ret_val;
   for (auto& var : op_desc_vars) {
     auto& var_names = ret_val[var.parameter()];
     auto& var_names_in_proto = var.arguments();
@@ -55,6 +41,17 @@ OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
   return ret_val;
 }
 
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
+  VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
+  VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
+  AttributeMap attrs;
+  for (auto& attr : op_desc.attrs()) {
+    attrs[attr.name()] = GetAttrValue(attr);
+  }
+
+  return CreateOp(op_desc.type(), inputs, outputs, attrs);
+}
+
 std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
   PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
   return std::unique_ptr<OperatorBase>(BuildGradOp(&op));

paddle/framework/op_registry.h

@@ -23,6 +23,7 @@ limitations under the License. */
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/grad_op_builder.h"
+#include "paddle/framework/op_info.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
 
@@ -30,28 +31,16 @@ namespace paddle {
 namespace framework {
 
 class OpRegistry {
-  using VarNameMap = OperatorBase::VarNameMap;
-  using OpCreator = std::function<OperatorBase*(
-      const std::string& /*type*/, const VarNameMap& /*inputs*/,
-      const VarNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
-
  public:
-  struct OpInfo {
-    OpCreator creator_;
-    std::string grad_op_type_;
-    OpProto* proto_;
-    OpAttrChecker* checker_;
-  };
-
   template <typename OpType, typename ProtoMakerType, typename GradOpType>
   static void RegisterOp(const std::string& op_type,
                          const std::string& grad_op_type) {
-    PADDLE_ENFORCE(op_info_map().count(op_type) == 0,
+    PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
                    "'%s' is registered more than once.", op_type);
     OpInfo op_info;
-    op_info.creator_ = [](const std::string& type, const VarNameMap& inputs,
-                          const VarNameMap& outputs,
-                          const AttributeMap& attrs) {
+    op_info.creator_ = [](
+        const std::string& type, const VariableNameMap& inputs,
+        const VariableNameMap& outputs, const AttributeMap& attrs) {
       return new OpType(type, inputs, outputs, attrs);
     };
     op_info.grad_op_type_ = grad_op_type;
@@ -70,7 +59,7 @@ class OpRegistry {
       op_info.proto_ = nullptr;
       op_info.checker_ = nullptr;
     }
-    op_info_map().insert(std::make_pair(op_type, op_info));
+    OpInfoMap::Instance().Insert(op_type, op_info);
     // register gradient op
     if (!grad_op_type.empty()) {
       RegisterOp<GradOpType, NOPMaker, NOP>(grad_op_type, "");
@@ -78,21 +67,13 @@ class OpRegistry {
   }
 
   static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
-                                                const VarNameMap& inputs,
-                                                const VarNameMap& outputs,
+                                                const VariableNameMap& inputs,
+                                                const VariableNameMap& outputs,
                                                 AttributeMap attrs);
 
   static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
 
-  static VarNameMap ConvertOpDescVarsToVarNameMap(
-      const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars);
-
   static std::unique_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
-
-  static std::unordered_map<std::string, const OpInfo>& op_info_map() {
-    static std::unordered_map<std::string, const OpInfo> op_info_map_;
-    return op_info_map_;
-  }
 };
 
 class Registrar {

paddle/framework/operator.cc

@@ -115,8 +115,8 @@ void OperatorBase::Rename(const std::string& old_name,
 }
 
 OperatorBase::OperatorBase(const std::string& type,
-                           const OperatorBase::VarNameMap& inputs,
-                           const OperatorBase::VarNameMap& outputs,
+                           const VariableNameMap& inputs,
+                           const VariableNameMap& outputs,
                            const AttributeMap& attrs)
     : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
   static std::atomic<size_t> gUniqId(0UL);
@@ -141,18 +141,10 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
     }
     return ret_val;
   }
-  auto it = OpRegistry::op_info_map().find(type_);
-  PADDLE_ENFORCE(
-      it != OpRegistry::op_info_map().end(),
-      "Operator %s not registered, cannot figure out intermediate outputs",
-      type_);
-  PADDLE_ENFORCE(
-      it->second.proto_ != nullptr,
-      "Operator %s has no OpProto, cannot figure out intermediate outputs",
-      type_);
+  auto& info = OpInfoMap::Instance().Get(Type());
 
   // get all OpProto::Var for outputs
-  for (auto& o : it->second.proto_->outputs()) {
+  for (auto& o : info.Proto().outputs()) {
     // ignore all intermediate output
     if (o.intermediate()) continue;
     auto out = outputs_.find(o.name());

paddle/framework/operator.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <unordered_map>
 #include <vector>
 
+#include "op_info.h"
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/scope.h"
@@ -62,10 +63,8 @@ class ExecutionContext;
  */
 class OperatorBase {
  public:
-  using VarNameMap = std::map<std::string, std::vector<std::string>>;
-
-  OperatorBase(const std::string& type, const VarNameMap& inputs,
-               const VarNameMap& outputs, const AttributeMap& attrs);
+  OperatorBase(const std::string& type, const VariableNameMap& inputs,
+               const VariableNameMap& outputs, const AttributeMap& attrs);
 
   virtual ~OperatorBase() {}
 
@@ -93,8 +92,8 @@ class OperatorBase {
   /// rename inputs outputs name
   void Rename(const std::string& old_name, const std::string& new_name);
 
-  const VarNameMap& Inputs() const { return inputs_; }
-  const VarNameMap& Outputs() const { return outputs_; }
+  const VariableNameMap& Inputs() const { return inputs_; }
+  const VariableNameMap& Outputs() const { return outputs_; }
   //! Get a input with argument's name described in `op_proto`
   const std::string& Input(const std::string& name) const;
   //! Get a input which has multiple variables.
@@ -122,30 +121,32 @@ class OperatorBase {
   // I (Inputs)opear
   // O (Outputs)
   // OG (Output Gradients)
-  VarNameMap inputs_;
+  VariableNameMap inputs_;
 
   // NOTE: in case of OpGrad, outputs_ contains
   // IG (Inputs Gradients)
-  VarNameMap outputs_;
+  VariableNameMap outputs_;
   AttributeMap attrs_;
 };
 
 // Macro for define a clone method.
 // If you are writing an kernel operator, `Clone` will be defined when you
 // register it. i.e. `Clone` method is not needed to define by yourself.
-#define DEFINE_OP_CLONE_METHOD(CLS)                       \
+#define DEFINE_OP_CLONE_METHOD(cls)                       \
   std::unique_ptr<OperatorBase> Clone() const final {     \
-    return std::unique_ptr<OperatorBase>(new CLS(*this)); \
+    return std::unique_ptr<OperatorBase>(new cls(*this)); \
   }
 
 // Macro for define a default constructor for Operator.
 // You can also use
 //   using PARENT_CLASS::PARENT_CLASS;
 // to use parent's constructor.
-#define DEFINE_OP_CONSTRUCTOR(CLS, PARENT_CLS)                                 \
-  CLS(const std::string& type, const VarNameMap& inputs,                       \
-      const VarNameMap& outputs, const paddle::framework::AttributeMap& attrs) \
-      : PARENT_CLS(type, inputs, outputs, attrs) {}
+#define DEFINE_OP_CONSTRUCTOR(cls, parent_cls)             \
+  cls(const std::string& type,                             \
+      const ::paddle::framework::VariableNameMap& inputs,  \
+      const ::paddle::framework::VariableNameMap& outputs, \
+      const paddle::framework::AttributeMap& attrs)        \
+      : parent_cls(type, inputs, outputs, attrs) {}
 
 class NOP : public OperatorBase {
  public:
@@ -389,8 +390,8 @@ class OperatorWithKernel : public OperatorBase {
   using OpKernelMap =
       std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;
 
-  OperatorWithKernel(const std::string& type, const VarNameMap& inputs,
-                     const VarNameMap& outputs, const AttributeMap& attrs)
+  OperatorWithKernel(const std::string& type, const VariableNameMap& inputs,
+                     const VariableNameMap& outputs, const AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
 
   void InferShape(const Scope& scope) const override {

paddle/framework/operator_test.cc

@@ -23,8 +23,8 @@ static int op_run_num = 0;
 
 class OpWithoutKernelTest : public OperatorBase {
  public:
-  OpWithoutKernelTest(const std::string& type, const VarNameMap& inputs,
-                      const VarNameMap& outputs, const AttributeMap& attrs)
+  OpWithoutKernelTest(const std::string& type, const VariableNameMap& inputs,
+                      const VariableNameMap& outputs, const AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs), x(1) {}
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
@@ -249,8 +249,9 @@ TEST(OpKernel, multi_inputs) {
 class OperatorClone : public paddle::framework::OperatorBase {
  public:
   DEFINE_OP_CLONE_METHOD(OperatorClone);
-  OperatorClone(const std::string& type, const VarNameMap& inputs,
-                const VarNameMap& outputs,
+  OperatorClone(const std::string& type,
+                const paddle::framework::VariableNameMap& inputs,
+                const paddle::framework::VariableNameMap& outputs,
                 const paddle::framework::AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
   void InferShape(const paddle::framework::Scope& scope) const override {}

paddle/framework/tensor.h

@@ -105,7 +105,10 @@ class Tensor {
   template <typename T>
   inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
 
-  platform::Place place() const { return holder_->place(); }
+  platform::Place place() const {
+    PADDLE_ENFORCE_NOT_NULL(holder_, "Tensor get place() must contains holder");
+    return holder_->place();
+  }
 
  private:
   template <typename T>

paddle/function/CMakeLists.txt

@@ -4,6 +4,10 @@ file(GLOB cpp_files . *Op.cpp)
 list(APPEND h_files Function.h)
 list(APPEND cpp_files Function.cpp)
 list(APPEND cpp_files BufferArg.cpp)
+list(APPEND cpp_files GemmFunctor.cpp)
+if(USE_EIGEN_FOR_BLAS)
+  list(APPEND cpp_files EigenGemm.cpp)
+endif(USE_EIGEN_FOR_BLAS)
 
 if(WITH_GPU)
     file(GLOB cu_files . *OpGpu.cu)
@@ -17,6 +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)
@@ -38,11 +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/DepthwiseConvOp.cpp

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "DepthwiseConvOp.h"
 #include "ConvOp.h"
-#include "GemmFunctor.h"
 
 namespace paddle {
 

paddle/function/DepthwiseConvOpGpu.cu

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "DepthwiseConvOp.h"
-#include "GemmFunctor.h"
 #include "paddle/math/BaseMatrix.h"
 
 namespace paddle {

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/EigenGemm.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 <glog/logging.h>
+#include "unsupported/Eigen/CXX11/Tensor"
+
+namespace paddle {
+
+template <class T>
+struct EigenBlasGemm {
+  typedef Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor, int>,
+                           Eigen::Aligned>
+      Matrix;
+
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+    Eigen::array<int, 2> sizeA;
+    if (transA) {
+      sizeA[0] = K;
+      sizeA[1] = M;
+      CHECK_EQ(M, lda);
+    } else {
+      sizeA[0] = M;
+      sizeA[1] = K;
+      CHECK_EQ(K, lda);
+    }
+    Eigen::array<int, 2> sizeB;
+    if (transB) {
+      sizeB[0] = N;
+      sizeB[1] = K;
+      CHECK_EQ(K, ldb);
+    } else {
+      sizeB[0] = K;
+      sizeB[1] = N;
+      CHECK_EQ(N, ldb);
+    }
+    Eigen::array<int, 2> sizeC;
+    sizeC[0] = M;
+    sizeC[1] = N;
+    CHECK_EQ(N, ldc);
+
+    const Matrix a(const_cast<T*>(A), sizeA);
+    const Matrix b(const_cast<T*>(B), sizeB);
+    Matrix c(C, sizeC);
+
+    typedef typename Eigen::Tensor<T, 2>::DimensionPair DimPair;
+    Eigen::array<DimPair, 1> dims;
+    dims[0] = DimPair(1, 0);
+    dims[0].first = transA ? 0 : 1;
+    dims[0].second = transB ? 1 : 0;
+
+    Eigen::DefaultDevice device;
+    if (alpha == T(1) && beta == T(0)) {
+      c.device(device) = a.contract(b, dims);
+    } else if (alpha == T(1) && beta == T(1)) {
+      c.device(device) += a.contract(b, dims);
+    } else {
+      c.device(device) = alpha * a.contract(b, dims) + beta * c;
+    }
+  }
+};
+
+#ifdef PADDLE_TYPE_DOUBLE
+template class EigenBlasGemm<double>;
+#else
+template class EigenBlasGemm<float>;
+#endif
+
+}  // namespace paddle

paddle/function/GemmConvOp.cpp

@@ -85,7 +85,6 @@ public:
     }
 
     Im2ColFunctor<kCFO, Device, real> im2col;
-    GemmFunctor<Device, real> gemm;
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -108,8 +107,8 @@ public:
         int M = outputChannels / groups_;
         int N = outputHeight * outputWidth;
         int K = inputChannels / groups_ * filterHeight * filterWidth;
-        gemm(CblasNoTrans,
-             CblasNoTrans,
+        BlasGemm<Device, real>::compute(false,
+                                        false,
                                         M,
                                         N,
                                         K,
@@ -188,8 +187,6 @@ public:
     }
 
     Col2ImFunctor<kCFO, Device, real> col2im;
-    GemmFunctor<Device, real> gemm;
-
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -205,8 +202,8 @@ public:
           colData = inputGrad + g * inputOffset;
           scale = 1.0f;
         }
-        gemm(CblasTrans,
-             CblasNoTrans,
+        BlasGemm<Device, real>::compute(true,
+                                        false,
                                         M,
                                         N,
                                         K,
@@ -299,7 +296,6 @@ public:
     }
 
     Im2ColFunctor<kCFO, Device, real> im2col;
-    GemmFunctor<Device, real> gemm;
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -321,8 +317,8 @@ public:
         int M = outputChannels / groups_;
         int K = outputHeight * outputWidth;
         int N = inputChannels / groups_ * filterHeight * filterWidth;
-        gemm(CblasNoTrans,
-             CblasTrans,
+        BlasGemm<Device, real>::compute(false,
+                                        true,
                                         M,
                                         N,
                                         K,

paddle/function/GemmFunctor.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 "GemmFunctor.h"
+#include "paddle/math/MathFunctions.h"
+
+namespace paddle {
+
+template <class T>
+struct BlasGemm<DEVICE_TYPE_CPU, T> {
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+#ifdef PADDLE_USE_EIGEN_FOR_BLAS
+    EigenBlasGemm<T>::compute(
+        transA, transB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);
+#else
+    gemm<T>(transA == false ? CblasNoTrans : CblasTrans,
+            transB == false ? CblasNoTrans : CblasTrans,
+            M,
+            N,
+            K,
+            alpha,
+            A,
+            lda,
+            B,
+            ldb,
+            beta,
+            C,
+            ldc);
+#endif
+  }
+};
+
+template <class T>
+struct BlasGemm<DEVICE_TYPE_GPU, T> {
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+    hl_matrix_mul((T*)A,
+                  transA == false ? HPPL_OP_N : HPPL_OP_T,
+                  (T*)B,
+                  transB == false ? HPPL_OP_N : HPPL_OP_T,
+                  C,
+                  M,
+                  N,
+                  K,
+                  alpha,
+                  beta,
+                  lda,
+                  ldb,
+                  ldc);
+  }
+};
+
+template struct BlasGemm<DEVICE_TYPE_CPU, real>;
+template struct BlasGemm<DEVICE_TYPE_GPU, real>;
+
+}  // namespace paddle

paddle/function/GemmFunctor.h

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/math/MathFunctions.h"
+#include "TensorType.h"
 
 namespace paddle {
 
@@ -24,10 +24,9 @@ namespace paddle {
 // of MatMulFunction, we need to consider the reconstruction of hl_matrix_mul
 // interface.
 template <DeviceType Device, class T>
-class GemmFunctor {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
+struct BlasGemm {
+  static void compute(const bool transA,
+                      const bool transB,
                       const int M,
                       const int N,
                       const int K,
@@ -41,11 +40,15 @@ public:
                       const int ldc);
 };
 
+// TODO(hedaoyuan): Since the definition of the real type in the Paddle
+// conflicts with the Eigen library, so compile the Eigen code can not
+// include the Paddle header file. And need an EigenBlasGemm template class
+// that does not contain the DeviceType parameter.
+// I will fix this problem and merge BlasGemm and EigenBlasGemm into one.
 template <class T>
-class GemmFunctor<DEVICE_TYPE_CPU, T> {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
+struct EigenBlasGemm {
+  static void compute(const bool transA,
+                      const bool transB,
                       const int M,
                       const int N,
                       const int K,
@@ -56,41 +59,7 @@ public:
                       const int ldb,
                       const T beta,
                       T* C,
-                  const int ldc) {
-    gemm<T>(transA, TransB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc);
-  }
-};
-
-template <class T>
-class GemmFunctor<DEVICE_TYPE_GPU, T> {
-public:
-  void operator()(const CBLAS_TRANSPOSE transA,
-                  const CBLAS_TRANSPOSE TransB,
-                  const int M,
-                  const int N,
-                  const int K,
-                  const T alpha,
-                  const T* A,
-                  const int lda,
-                  const T* B,
-                  const int ldb,
-                  const T beta,
-                  T* C,
-                  const int ldc) {
-    hl_matrix_mul((T*)A,
-                  transA == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T,
-                  (T*)B,
-                  TransB == CblasNoTrans ? HPPL_OP_N : HPPL_OP_T,
-                  C,
-                  M,
-                  N,
-                  K,
-                  alpha,
-                  beta,
-                  lda,
-                  ldb,
-                  ldc);
-  }
+                      const int ldc);
 };
 
 }  // 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/gradientmachines/NeuralNetwork.cpp

@@ -202,7 +202,7 @@ void NeuralNetwork::prefetch(const std::vector<Argument>& inArgs) {
         auto mat = dynamic_cast<SparsePrefetchRowCpuMatrix*>(
             para->getMat(PARAMETER_VALUE).get());
         para->clearGradient();
-        mat->clearIndices();
+        if (mat) mat->clearIndices();
       }
     }
   }

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -184,7 +184,7 @@ public:
   }
 
   void backward(const UpdateCallback& callback) override {
-    if (biases_) {
+    if (biases_ && biases_->getWGrad()) {
       backwardActivation();
       biases_->getWGrad()->collectBias(*getOutputGrad(), 1);
       biases_->getParameterPtr()->incUpdate(callback);
@@ -1012,11 +1012,6 @@ void RecurrentGradientMachine::generateSequence() {
                            /* width */ resultNum,
                            false,
                            /* useGpu */ false);
-    Matrix::resizeOrCreate(generator_.outArg.value,
-                           /* height */ maxGenWordCount,
-                           /* width */ 1,
-                           false,
-                           /* useGpu */ false);
   }
   ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
                                 numSequences + 1,
@@ -1026,7 +1021,7 @@ void RecurrentGradientMachine::generateSequence() {
   } else {
     oneWaySearch(numSequences);
   }
-  if (dataArgsSize_) createDataOutlink(batchMachineIdVec_);
+  if (dataArgsSize_) createDataOutlink();
 
   size_t size = generator_.ids.size();
   generator_.outArg.ids->resize(size);
@@ -1106,6 +1101,7 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) {
   }
 
   batchMachineIdVec_.clear();
+  batchMachineStartPos_.clear();
   int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
   starts[0] = 0;
   generator_.ids.clear();
@@ -1312,13 +1308,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
     finalPaths_[i].resize(minFinalPathsSize);
   }
 
-  batchMachineIdVec_.clear();
   generator_.ids.clear();
   int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
   starts[0] = 0;
   if (numResults > 1) {
-    real* probs = generator_.outArg.in->getData();
+    int idsProbSaveSize = 0;
+    for (auto inSeq : finalPaths_) {
+      for (auto path : inSeq) idsProbSaveSize += path.ids.size();
+      idsProbSaveSize += inSeq.size();
+    }
+    Matrix::resizeOrCreate(
+        generator_.outArg.value, idsProbSaveSize, 1, false, false);
     real* idsProb = generator_.outArg.value->getData();
+
+    real* probs = generator_.outArg.in->getData();
     size_t curPos = 0;
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
@@ -1333,24 +1336,16 @@ void RecurrentGradientMachine::fillGenOutputs() {
         curPos += genLen;
         idsProb[curPos++] = -1.0;
         probs[i * numResults + j] = path.logProb;
-
-        if (!j && dataArgsSize_) {
-          // in beam search, here only reserved the top 1 generated result
-          // for out_links that are not the generated word indices.
-          batchMachineIdVec_.insert(batchMachineIdVec_.end(),
-                                    path.machineIdVec.begin(),
-                                    path.machineIdVec.end());
-        }
       }
       starts[i + 1] = generator_.ids.size();
     }
   } else {
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       CHECK(!finalPaths_[i].empty());
-      generator_.ids.insert(generator_.ids.begin(),
-                            finalPaths_[i][0].ids.begin(),
-                            finalPaths_[i][0].ids.end());
-      starts[i + 1] = starts[i] + finalPaths_[i][0].ids.size();
+      Path& path = finalPaths_[i][0];
+      generator_.ids.insert(
+          generator_.ids.end(), path.ids.begin(), path.ids.end());
+      starts[i + 1] = starts[i] + path.ids.size();
     }
   }
 }
@@ -1364,25 +1359,76 @@ void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) {
   }
 }
 
-void RecurrentGradientMachine::createDataOutlink(
-    std::vector<int>& machineIdVec) {
-  size_t seqNum =
-      getBeamSize() > 1UL ? finalPaths_.size() : finalPaths_[0].size();
-  std::vector<int> starts(seqNum + 1, 0);
-  for (size_t i = 0; i < seqNum; ++i) {
-    size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size()
-                                        : finalPaths_[0][i].ids.size();
-    starts[i + 1] = starts[i] + seqLen;
+void RecurrentGradientMachine::createDataOutlinkSelRowsInfo(
+    bool isSeq, std::vector<Argument>& outArgs) {
+  batchMachineIdVec_.clear();
+
+  size_t seqIdx = 0;
+  for (size_t i = 0; i < finalPaths_.size(); ++i) {
+    for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
+      std::vector<int>& machineIdVec = finalPaths_[i][j].machineIdVec;
+      if (isSeq) {
+        for (size_t i = 0; i < machineIdVec.size(); ++i) {
+          size_t rowId = machineIdVec[i];
+          int* seqPos =
+              outArgs[i].sequenceStartPositions->getMutableData(false);
+          batchMachineIdVec_.push_back(seqPos[rowId]);
+        }
+      } else {
+        batchMachineIdVec_.insert(
+            batchMachineIdVec_.end(), machineIdVec.begin(), machineIdVec.end());
+      }
+      seqIdx++;
+    }
+  }
+}
+
+void RecurrentGradientMachine::createDataOutlinkCopySizeInfo(
+    bool isSeq, std::vector<Argument>& outArgs, std::vector<int>& copySize) {
+  size_t totalSeqNum = std::accumulate(
+      finalPaths_.begin(),
+      finalPaths_.end(),
+      0UL,
+      [](size_t a, const std::vector<Path>& b) { return a + b.size(); });
+  copySize.resize(totalSeqNum, 1);
+
+  batchMachineStartPos_.resize(totalSeqNum + 1, 0);
+  if (isSeq) {
+    ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions;
+    CHECK_EQ(static_cast<size_t>(inputSeqStartPos->getSize() - 1),
+             getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size());
+    int* starts = inputSeqStartPos->getMutableData(false);
+    int seqId = 0;
+    for (size_t i = 0; i < finalPaths_.size(); ++i) {
+      for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
+        copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i]
+                                            : starts[j + 1] - starts[j];
+        batchMachineStartPos_[seqId + 1] =
+            batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size();
+        seqId++;
+      }
+    }
+  } else {
+    for (size_t i = 0; i < finalPaths_[0].size(); ++i)
+      batchMachineStartPos_[i + 1] =
+          batchMachineStartPos_[i] + finalPaths_[0][i].ids.size();
   }
+}
 
+void RecurrentGradientMachine::createDataOutlink() {
   for (size_t i = 0; i < dataArgsSize_; i++) {
+    bool isSeq = dataArgsFrame_[i][0].hasSeq();
+    std::vector<int> copySize;
+    createDataOutlinkCopySizeInfo(isSeq, dataArgsFrame_[i], copySize);
+    createDataOutlinkSelRowsInfo(isSeq, dataArgsFrame_[i]);
+
     dataArgs_[i].concat(dataArgsFrame_[i],
-                        machineIdVec,
-                        starts,
+                        batchMachineIdVec_,
+                        batchMachineStartPos_,
+                        copySize,
                         useGpu_,
                         HPPL_STREAM_1,
                         PASS_TEST);
-
     auto dataAgent =
         dynamic_cast<DataLayer*>(outFrameLines_[i + 1].agentLayer.get());
     CHECK_NOTNULL(dataAgent);

paddle/gserver/gradientmachines/RecurrentGradientMachine.h

@@ -190,7 +190,7 @@ public:
     std::vector<int> ids;
 
     /**
-     * @brief idsProb, log probability of each generated words.
+     * @brief idsProb, log probability of each generated word.
      */
     std::vector<real> idsProb;
 
@@ -472,15 +472,43 @@ private:
   void copyDataOutlinkFrame(size_t machineCur);
 
   /*
-   * @brief In generation, if the layer group has more than 1 outlink, outlinks
-   * except the first one are data outlinks. This function creates the data
-   * outlinks.
-   * @note In beam search, only one generated sequence with the hightest log
-   * probabilites are retained.
-   * @param machineIdVec : select a row of output matrix in each frame
-   * that the generation process expanded.
+   * @brief In generation, if the layer group has more than 1 outlink, outlink
+   * except the first one is a data outlink. In RecurrentLayerGroup, each time
+   * step is a separate Network, outputs of a layer inside the
+   * RecurrentLayerGroup are stored in separate Arguments. If one layer is
+   * specified as an outlink of RecurrentLayerGroup. This function will
+   * collect outputs in each time step of each generated sequence which are
+   * dispersed in separate Arguments to form a new single Argument as output of
+   * RecurrentLayerGroup.
    */
-  void createDataOutlink(std::vector<int>& machineIdVec);
+  void createDataOutlink();
+
+  /*
+   * @brief decide to select how many rows from the Matrix stored the forward
+   * pass results from a start position.
+   *
+   * @param isSeq: a flag indicating whetehr the layer to be output of the
+   * RecurrentGradientMachine is a sequence or not
+   * @param outArgs: all of the the returned Arguments of the forward pass
+   * during the generation process.
+   * @param copySize: the returned result, number of rows to select from the
+   * Matrix stored the forward pass results from a start position.
+   */
+  void createDataOutlinkCopySizeInfo(bool isSeq,
+                                     std::vector<Argument>& outArgs,
+                                     std::vector<int>& copySize);
+
+  /*
+   * @brief decide index of the start row for each time step of a generated
+   * sequence in Matrix stored the entire beam search batch's forward pass
+   * results.
+   *
+   * @param isSeq: a flag indicating whether the layer to be output of the
+   * RecurrentGradientMachine is a sequence or not
+   * @param outArgs: all of the returned Arguments of the forward pass
+   * during the generation process.
+   */
+  void createDataOutlinkSelRowsInfo(bool isSeq, std::vector<Argument>& outArgs);
 
   /*
    * @brief used in beam search, connect previous frame to form recurrent link
@@ -543,6 +571,7 @@ private:
   std::vector<int> topIds_;
   std::vector<int> seqIds_;
   std::vector<int> batchMachineIdVec_;
+  std::vector<int> batchMachineStartPos_;
   std::vector<std::vector<Path>> finalPaths_;
   std::vector<real> minFinalPathLogProb_;
   BeamSearchControlCallbacks* beamSearchCtrlCallbacks_;

paddle/gserver/layers/ConvBaseLayer.cpp

@@ -32,9 +32,11 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
     const ConvConfig& conf = inputConfig.conv_conf();
     padding_.push_back(conf.padding());
     stride_.push_back(conf.stride());
+    dilation_.push_back(conf.dilation());
     filterSize_.push_back(conf.filter_size());
     paddingY_.push_back(conf.padding_y());
     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());
@@ -89,7 +91,11 @@ size_t ConvBaseLayer::calOutputSize() {
   size_t layerSize = 0;
 
   auto setLayerSize = [&](IntV& inH, IntV& inW, IntV& outH, IntV& outW) {
+    size_t filterSizeY;
+    size_t filterSize;
     for (size_t i = 0; i < inputLayers_.size(); i++) {
+      filterSizeY = (filterSizeY_[i] - 1) * dilationY_[i] + 1;
+      filterSize = (filterSize_[i] - 1) * dilation_[i] + 1;
       inH.push_back(inputLayers_[i]->getOutput().getFrameHeight());
       inW.push_back(inputLayers_[i]->getOutput().getFrameWidth());
       const ConvConfig& conf = config_.inputs(i).conv_conf();
@@ -98,17 +104,17 @@ size_t ConvBaseLayer::calOutputSize() {
           inH[i] = conf.has_output_y() ? conf.output_y() : conf.output_x();
         if (inW[i] == 0) inW[i] = conf.output_x();
         outH.push_back(imageSize(
-            inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_));
-        outW.push_back(imageSize(
-            inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_));
+            inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_));
+        outW.push_back(
+            imageSize(inW[i], filterSize, padding_[i], stride_[i], caffeMode_));
       } else {
         if (inH[i] == 0)
           inH[i] = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
         if (inW[i] == 0) inW[i] = conf.img_size();
         outH.push_back(outputSize(
-            inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_));
+            inH[i], filterSizeY, paddingY_[i], strideY_[i], caffeMode_));
         outW.push_back(outputSize(
-            inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_));
+            inW[i], filterSize, padding_[i], stride_[i], caffeMode_));
       }
       CHECK_EQ(outH[i], outH[0]);
       CHECK_EQ(outW[i], outW[0]);

paddle/gserver/layers/ConvBaseLayer.h

@@ -40,6 +40,10 @@ protected:
   IntV stride_;
   /// The y dimension of the stride.
   IntV strideY_;
+  /// The x dimension of the dilation.
+  IntV dilation_;
+  /// The y dimension of the dilation.
+  IntV dilationY_;
   /// The x dimension of a filter kernel.
   IntV filterSize_;
   /// The y dimension of a filter kernel.

paddle/gserver/layers/ConvBaseOperator.cpp

@@ -59,7 +59,8 @@ void ConvBaseOperator::allocConvWorkSpace() {
                     &bwdDataAlgo_,
                     &bwdDataLimitBytes_,
                     &bwdFilterAlgo_,
-                    &bwdFilterLimitBytes_);
+                    &bwdFilterLimitBytes_,
+                    /*useDilation*/ false);
 
   size_t maxWorkSpace = 0;
   maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);

paddle/gserver/layers/ConvBaseProjection.cpp

@@ -41,6 +41,11 @@ void ConvBaseProjection::getConvParams() {
   strideH_ = conf.stride_y();
   strideW_ = conf.stride();
 
+  dilationH_ = conf.dilation_y();
+  dilationW_ = conf.dilation();
+  CHECK_GT(dilationH_, 0);
+  CHECK_GT(dilationW_, 0);
+
   filterH_ = conf.filter_size_y();
   filterW_ = conf.filter_size();
 
@@ -77,7 +82,9 @@ void ConvBaseProjection::initCudnn() {
                                    paddingH_,
                                    paddingW_,
                                    strideH_,
-                                   strideW_);
+                                   strideW_,
+                                   dilationH_,
+                                   dilationW_);
 
   // initialize all to default algorithms
   fwdAlgo_ = 0;
@@ -131,7 +138,9 @@ void ConvBaseProjection::reshapeTensorDesc(int batchSize) {
                                   paddingH_,
                                   paddingW_,
                                   strideH_,
-                                  strideW_);
+                                  strideW_,
+                                  dilationH_,
+                                  dilationW_);
 }
 
 void ConvBaseProjection::reshape(int batchSize) {
@@ -140,6 +149,10 @@ void ConvBaseProjection::reshape(int batchSize) {
   CHECK_EQ(calInputSize(), in_->value->getWidth());
 
   reshapeTensorDesc(batchSize);
+  bool useDilation = false;
+  if (dilationH_ > 1 || dilationW_ > 1) {
+    useDilation = true;
+  }
   hl_conv_workspace(imageDesc_,
                     outputDesc_,
                     filterDesc_,
@@ -149,7 +162,8 @@ void ConvBaseProjection::reshape(int batchSize) {
                     &bwdDataAlgo_,
                     &bwdDataLimitBytes_,
                     &bwdFilterAlgo_,
-                    &bwdFilterLimitBytes_);
+                    &bwdFilterLimitBytes_,
+                    useDilation);
 
   size_t maxWorkSpace = 0;
   maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);

paddle/gserver/layers/ConvBaseProjection.h

@@ -63,6 +63,7 @@ protected:
   int configChannels_, configNumFilters_;
   int paddingH_, paddingW_;
   int strideH_, strideW_;
+  int dilationH_, dilationW_;
   int filterH_, filterW_;
   /// One group offset of input data.
   int inputOffset_;

paddle/gserver/layers/ConvProjection.cpp

@@ -25,12 +25,12 @@ size_t ConvProjection::calOutputSize() {
   if (imageH_ == 0) imageH_ = configImgH_;
   if (imageW_ == 0) imageW_ = configImgW_;
   outputH_ = outputSize(imageH_,
-                        filterH_,
+                        (filterH_ - 1) * dilationH_ + 1,
                         paddingH_,
                         strideH_,
                         /* caffeMode */ true);
   outputW_ = outputSize(imageW_,
-                        filterW_,
+                        (filterW_ - 1) * dilationW_ + 1,
                         paddingW_,
                         strideW_,
                         /* caffeMode */ true);

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 +304,68 @@ 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) {}
+};
+
+/**
+ * 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/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/KmaxSeqScoreLayer.cpp

@@ -80,13 +80,14 @@ void KmaxSeqScoreLayer::forward(PassType passType) {
       << "input of " << getName()
       << " must be a sequence or a nested sequence.";
   CHECK_EQ(input.value->getWidth(), 1UL)
-      << "input of " << getName()
-      << " is score over a sequence or a nested sequence, so its width "
-      << " must be 1.";
+      << "input of " << getName() << " are scores over a sequence or "
+      << "a nested sequence, so its width must be 1.";
 
   if (useGpu_) {
-    // this Layer runs only in CPU, if the model is runing on GPU,
-    // then copy the input to this layer from GPU to CPU.
+    /*
+     * currently, this Layer only runs in CPU, if the other part of the model is
+     * runing on GPU, then copy the input to this layer from GPU to CPU.
+     */
     Matrix::resizeOrCreate(scores_,
                            inputScore->getHeight(),
                            1,
@@ -97,6 +98,14 @@ void KmaxSeqScoreLayer::forward(PassType passType) {
     scores_ = inputScore;
   }
 
+  /*
+   * TODO(caoying)
+   * In PaddePaddle, currently all matrices are real number types,
+   * but output of this layer which is some selected indices of the give
+   * sequence are actually filled with int types so that storing int types
+   * information in a real number matrix is dangerous, since real numbers will
+   * be convered to int types.
+   */
   Matrix::resizeOrCreate(
       output_.value,
       input.hasSubseq() ? input.getNumSubSequences() : input.getNumSequences(),

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/MKLDNNLayer.h

@@ -18,9 +18,9 @@ limitations under the License. */
 #include "Layer.h"
 #include "MKLDNNBase.h"
 #include "mkldnn.hpp"
+#include "paddle/math/MKLDNNMatrix.h"
 
 DECLARE_bool(use_mkldnn);
-DECLARE_bool(use_mkldnn_wgt);
 
 namespace paddle {
 
@@ -52,15 +52,15 @@ protected:
   std::vector<mkldnn::primitive> pipelineFwd_;
   std::vector<mkldnn::primitive> pipelineBwd_;
 
-  // TODO(TJ): change below memory as MKLDNNMatrixPtr type
-  std::shared_ptr<mkldnn::memory> inVal_;
-  std::shared_ptr<mkldnn::memory> inGrad_;
-  std::shared_ptr<mkldnn::memory> outVal_;
-  std::shared_ptr<mkldnn::memory> outGrad_;
-  std::shared_ptr<mkldnn::memory> wgtVal_;
-  std::shared_ptr<mkldnn::memory> wgtGrad_;
-  std::shared_ptr<mkldnn::memory> biasVal_;
-  std::shared_ptr<mkldnn::memory> biasGrad_;
+  // MKLDNNMatrixPtr
+  MKLDNNMatrixPtr inVal_;
+  MKLDNNMatrixPtr inGrad_;
+  MKLDNNMatrixPtr outVal_;
+  MKLDNNMatrixPtr outGrad_;
+  MKLDNNMatrixPtr wgtVal_;
+  MKLDNNMatrixPtr wgtGrad_;
+  MKLDNNMatrixPtr biasVal_;
+  MKLDNNMatrixPtr biasGrad_;
 
 public:
   explicit MKLDNNLayer(const LayerConfig& config)
@@ -83,17 +83,21 @@ public:
 
   virtual bool init(const LayerMap& layerMap,
                     const ParameterMap& parameterMap) {
+    CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
+                            << "Please set WITH_MKLDNN=ON "
+                            << "and set use_mkldnn=True";
+    CHECK(!useGpu_) << "Do not support GPU yet";
+
+    // set device id before Layer::init
+    setDevice(MKLDNN_DEVICE);
+    // change param device to MKLDNN device
+    setParamsDevice(MKLDNN_DEVICE, parameterMap);
     if (!Layer::init(layerMap, parameterMap)) {
       return false;
     }
 
-    CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
-                            << "Please set WITH_MKLDNN=ON "
-                            << "and set use_mkldnn=True";
     stream_.reset(new MKLDNNStream());
     engine_ = CPUEngine::Instance().getEngine();
-
-    // TODO(TJ): deivecId
     return true;
   }
 
@@ -109,6 +113,12 @@ public:
    */
   virtual void convertWeightsToPaddle() {}
 
+  /**
+   * convert MKLDNN output to other device.
+   * only support CPU device yet
+   */
+  virtual void convertOutputToOtherDevice() {}
+
   /**
    * print info about sizes
    */
@@ -118,14 +128,124 @@ public:
                        << ", oh: " << oh_ << ", ow: " << ow_;
   }
 
-  // TODO(TJ): move to MkldnnMatrix
-  // create memory desc
-  inline mkldnn::memory::desc createMD(
-      mkldnn::memory::dims dims,
-      mkldnn::memory::format fmt,
-      mkldnn::memory::data_type type = mkldnn::memory::data_type::f32) {
-    // TODO(TJ): isFmtSuppoted(fmt)
-    return mkldnn::memory::desc(dims, type, fmt);
+  /**
+   * Print the mkldnn memory format flow of value
+   */
+  virtual void printValueFormatFlow() {
+    if (inVal_ && outVal_) {
+      VLOG(MKLDNN_FMTS) << "value format flow --- " << inVal_->getFormat()
+                        << " >>> " << outVal_->getFormat();
+    }
+  }
+
+  /**
+   * Print the mkldnn memory format flow of grad
+   */
+  virtual void printGradFormatFlow() {
+    if (inGrad_ && outGrad_) {
+      VLOG(MKLDNN_FMTS) << "grad format flow --- " << inGrad_->getFormat()
+                        << " <<< " << outGrad_->getFormat();
+    }
+  }
+
+protected:
+  /**
+   * copy image size and sequence info to other device
+   * @note: can not directly use Layer::copyOutputToOtherDevice since here only
+   *        copy base info and do not copy data value
+   */
+  void copyOutputInfoToOtherDevice() {
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      outputOtherDevice_[i].setFrameHeight(output_.getFrameHeight());
+      outputOtherDevice_[i].setFrameWidth(output_.getFrameWidth());
+      outputOtherDevice_[i].sequenceStartPositions =
+          output_.sequenceStartPositions;
+      outputOtherDevice_[i].subSequenceStartPositions =
+          output_.subSequenceStartPositions;
+      outputOtherDevice_[i].cpuSequenceDims = output_.cpuSequenceDims;
+    }
+  }
+
+  /**
+   * If input only has MKLDNN device.
+   * Otherwise, only support the previous layer using CPU device.
+   */
+  bool inputIsOnlyMKLDNN(int index = 0) {
+    int prevDevice = getPrev(index)->getDeviceId();
+    if (prevDevice == MKLDNN_DEVICE) {
+      return true;
+    } else {
+      // do not support GPU yet
+      CHECK_EQ(prevDevice, CPU_DEVICE) << "Only support CPU yet";
+      return false;
+    }
+  }
+
+  /**
+   * If output only has MKLDNN device.
+   * Otherwise, other devices should only using CPU device.
+   */
+  bool outputIsOnlyMKLDNN() {
+    for (size_t i = 0; i < outputOtherDevice_.size(); i++) {
+      CHECK_EQ(outputOtherDevice_[i].deviceId, CPU_DEVICE)
+          << "Only support other device is CPU yet";
+    }
+    return outputOtherDevice_.size() == 0;
+  }
+
+  /**
+   * Sync input value data
+   */
+  void syncInputValue() {
+    if (inputIsOnlyMKLDNN()) {
+      return;
+    }
+    real* iData = getInputValue(0, CPU_DEVICE)->getData();
+    // update input data
+    // since it might be changed if this is after data layer
+    inVal_->updateData(iData);
+  }
+
+  /**
+   * Sync output grad data
+   */
+  void syncOutputGrad() {
+    if (outputIsOnlyMKLDNN()) {
+      return;
+    }
+
+    // update diff
+    real* oDiff = getOutput(CPU_DEVICE).grad->getData();
+    outGrad_->updateData(oDiff);
+  }
+
+  /**
+   * Set deviceId of this layer.
+   */
+  void setDevice(int id) { deviceId_ = id; }
+
+  /**
+   * Set deviceId of the params used in this layer.
+   */
+  void setParamsDevice(int id, const ParameterMap& parameterMap) {
+    for (auto& inputConfig : config_.inputs()) {
+      if (inputConfig.has_input_parameter_name()) {
+        ParameterPtr parameter;
+        std::string name = inputConfig.input_parameter_name();
+        CHECK(mapGet(name, parameterMap, &parameter))
+            << "Cannot find input parameter " << name << " for layer "
+            << getName();
+        parameter->setDevice(id);
+      }
+    }
+    if (config_.has_bias_parameter_name()) {
+      ParameterPtr parameter;
+      std::string name = config_.bias_parameter_name();
+      CHECK(mapGet(name, parameterMap, &parameter))
+          << "Cannot find bias parameter " << name << " for layer "
+          << getName();
+      parameter->setDevice(id);
+    }
   }
 };
 

paddle/gserver/layers/ScaleShiftLayer.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 "Layer.h"
+
+namespace paddle {
+
+/**
+ * A layer applies a linear transformation to each element in each row of
+ * the input matrix. For each element, the layer first re-scale it and then
+ * adds a bias to it.
+ *
+ * \f[
+ *    y = wx + b
+ * \f]
+ *
+ * Here, w is the scale and b is the bias. Both w and b are trainable scalars.
+ *
+ */
+
+class ScaleShiftLayer : public Layer {
+protected:
+  std::unique_ptr<Weight> scale_;
+  std::unique_ptr<Weight> offset_;
+
+public:
+  explicit ScaleShiftLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(scale_shift, ScaleShiftLayer);
+
+bool ScaleShiftLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  Layer::init(layerMap, parameterMap);
+  CHECK_EQ(inputLayers_.size(), 1U);
+  scale_.reset(new Weight(1, 1, parameters_[0]));
+  if (biasParameter_.get() != NULL) {
+    offset_ = std::unique_ptr<Weight>(new Weight(1, 1, biasParameter_));
+  }
+  return true;
+}
+
+void ScaleShiftLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  MatrixPtr inV = getInputValue(0);
+  resetOutput(inV->getHeight(), inV->getWidth());
+  MatrixPtr outV = getOutputValue();
+  real scaleValue = scale_->getW()->getElement(0, 0);
+  outV->mulScalar(*inV, scaleValue);
+  if (offset_) {
+    real offsetValue = offset_->getW()->getElement(0, 0);
+    outV->add(offsetValue);
+  }
+}
+
+void ScaleShiftLayer::backward(const UpdateCallback& callback) {
+  MatrixPtr inV = getInputValue(0);
+  MatrixPtr inG = getInputGrad(0);
+  MatrixPtr outV = getOutputValue();
+  MatrixPtr outG = getOutputGrad();
+
+  /* Calculate the parameter gradient for the current layer */
+  if (scale_->getWGrad()) {
+    MatrixPtr rowSumMtx;
+    Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_);
+    // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ij} * c_{ij}
+    rowSumMtx->sumOfProducts(
+        /* b= */ *inV, /* c= */ *outG, /* scaleSum= */ 1, /* scaleDest= */ 0.);
+    // this_i = scaleDest * this_i + scaleSum * \sum_j b_{ji}
+    scale_->getWGrad()->sumCols(
+        /* b= */ *rowSumMtx, /* scaleSum= */ 1., /* scaleDest= */ 1.);
+    scale_->getParameterPtr()->incUpdate(callback);
+  }
+  if (offset_ && offset_->getWGrad()) {
+    MatrixPtr rowSumMtx;
+    Matrix::resizeOrCreate(rowSumMtx, outG->getHeight(), 1, false, useGpu_);
+    rowSumMtx->sumRows(*outG, 1., 0.);
+    offset_->getWGrad()->sumCols(*rowSumMtx, 1., 1.);
+    offset_->getParameterPtr()->incUpdate(callback);
+  }
+
+  /* Calculate the input layers error */
+  if (inG) {
+    real scaleValue = scale_->getW()->getElement(0, 0);
+    inG->add(*outG, scaleValue);
+  }
+}
+
+}  // namespace paddle

paddle/gserver/layers/SequenceSliceLayer.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 "Layer.h"
+#include "paddle/math/Matrix.h"
+#include "paddle/math/Vector.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+class SequenceSliceLayer : public Layer {
+public:
+  explicit SequenceSliceLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+
+private:
+  /*
+   * TODO(caoying)
+   * In PaddePaddle, currently all matrices are real number types,
+   * but the second and the (optional) third input which are some
+   * selected indices of the give sequence to trim the sequence, are actually
+   * filled with int types so that storing int types information in real number
+   * matrices is very dangerous, since real numbers will be convered to int
+   * types. If a user fills this matrix himself, invalid data may occor.
+   */
+
+  MatrixPtr startIdsOnCpu_;
+  MatrixPtr endIdsOnCpu_;
+
+  std::vector<int> selectedRows_;
+  IVectorPtr rowIndice_;
+  std::vector<std::vector<int>> inputSeqInfoVec_;
+  std::vector<int> outSubSeqStartPos_;
+  std::vector<int> outSeqStartPos_;
+
+  void checkInputs();
+  void copySliceIdsToCpu();
+  void calSelectedRows(const MatrixPtr starts, const MatrixPtr ends);
+};
+
+REGISTER_LAYER(seq_slice, SequenceSliceLayer);
+
+bool SequenceSliceLayer::init(const LayerMap& layerMap,
+                              const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+  CHECK_GE(inputLayers_.size(), 2U);
+  CHECK_LE(inputLayers_.size(), 3U);
+
+  setNeedSequenceInfo(false);
+  return true;
+}
+
+void SequenceSliceLayer::checkInputs() {
+  const Argument& inputSeq = getInput(0);
+  CHECK(inputSeq.hasSeq()) << "The first input of sequence slice layer "
+                           << "must be a sequence.";
+  const MatrixPtr indices1 = getInputValue(1);
+  CHECK_EQ(static_cast<size_t>(indices1->getHeight()),
+           inputSeq.hasSubseq() ? inputSeq.getNumSubSequences()
+                                : inputSeq.getNumSequences())
+      << "Height of the second input should be equal to number of sequence "
+      << "in the first input.";
+  if (inputLayers_.size() == 3) {
+    const MatrixPtr indices2 = getInputValue(2);
+    CHECK_EQ(indices2->getHeight(), indices1->getHeight())
+        << "start indices and end indices should have the same height.";
+    CHECK_EQ(indices2->getWidth(), indices1->getWidth())
+        << "start indices and end indices should have the same Width.";
+  }
+}
+
+void SequenceSliceLayer::copySliceIdsToCpu() {
+  const MatrixPtr indices1 = getInputValue(1);
+  if (inputLayers_.size() == 2U) {
+    if (config_.select_first()) {
+      Matrix::resizeOrCreate(startIdsOnCpu_,
+                             indices1->getHeight(),
+                             indices1->getWidth(),
+                             false /* trans */,
+                             false /* useGpu */);
+      startIdsOnCpu_->copyFrom(*indices1);
+      endIdsOnCpu_ = nullptr;
+    } else {
+      Matrix::resizeOrCreate(endIdsOnCpu_,
+                             indices1->getHeight(),
+                             indices1->getWidth(),
+                             false /* trans */,
+                             false /* useGpu */);
+      endIdsOnCpu_->copyFrom(*indices1);
+      startIdsOnCpu_ = nullptr;
+    }
+  } else if (inputLayers_.size() == 3U) {
+    Matrix::resizeOrCreate(startIdsOnCpu_,
+                           indices1->getHeight(),
+                           indices1->getWidth(),
+                           false /* trans */,
+                           false /* useGpu */);
+    startIdsOnCpu_->copyFrom(*indices1);
+
+    const MatrixPtr indices2 = getInputValue(2);
+    Matrix::resizeOrCreate(endIdsOnCpu_,
+                           indices2->getHeight(),
+                           indices2->getWidth(),
+                           false /* trans */,
+                           false /* useGpu */);
+    endIdsOnCpu_->copyFrom(*indices2);
+  }
+}
+
+void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts,
+                                         const MatrixPtr ends) {
+  CHECK(starts || ends) << "At least one of the start or end indices "
+                        << "should be given.";
+
+  bool hasSubseq = getInput(0).hasSubseq();
+
+  outSeqStartPos_.resize(1, 0);
+  outSubSeqStartPos_.resize(1, 0);
+  selectedRows_.clear();
+
+  size_t beamSize = starts ? starts->getWidth() : ends->getWidth();
+  size_t rowIdx = 0;
+  for (size_t i = 0; i < inputSeqInfoVec_.size(); ++i) {
+    for (size_t j = 0; j < inputSeqInfoVec_[i].size() - 1; ++j) {
+      for (size_t k = 0; k < beamSize; ++k) {
+        if (starts && starts->getElement(rowIdx, k) == -1.) break;
+        if (ends && ends->getElement(rowIdx, k) == -1.) break;
+
+        int begPos = inputSeqInfoVec_[i][j];
+        if (starts) begPos += starts->getElement(rowIdx, k);
+
+        int endPos = inputSeqInfoVec_[i][j + 1] - 1;
+        if (ends) endPos = inputSeqInfoVec_[i][j] + ends->getElement(rowIdx, k);
+
+        int seqLen = endPos - begPos + 1;
+        CHECK_GT(seqLen, 0U);
+        for (int m = begPos; m <= endPos; ++m) selectedRows_.push_back(m);
+        hasSubseq
+            ? outSubSeqStartPos_.push_back(outSubSeqStartPos_.back() + seqLen)
+            : outSeqStartPos_.push_back(outSeqStartPos_.back() + seqLen);
+      }
+      rowIdx++;
+    }
+    if (hasSubseq) outSeqStartPos_.push_back(outSubSeqStartPos_.back());
+  }
+
+  if (useGpu_) {
+    rowIndice_ = IVector::create(selectedRows_.size(), useGpu_);
+    rowIndice_->copyFrom(selectedRows_.data(), selectedRows_.size());
+  } else {
+    rowIndice_ =
+        IVector::create(selectedRows_.data(), selectedRows_.size(), useGpu_);
+  }
+
+  // create the sequence information for the output.
+  ICpuGpuVector::resizeOrCreate(
+      output_.sequenceStartPositions, outSeqStartPos_.size(), false);
+  output_.sequenceStartPositions->copyFrom(
+      outSeqStartPos_.data(), outSeqStartPos_.size(), false);
+
+  if (hasSubseq) {
+    ICpuGpuVector::resizeOrCreate(
+        output_.subSequenceStartPositions, outSubSeqStartPos_.size(), false);
+    output_.subSequenceStartPositions->copyFrom(
+        outSubSeqStartPos_.data(), outSubSeqStartPos_.size(), false);
+  }
+}
+
+void SequenceSliceLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  checkInputs();
+
+  const Argument& inputSeq = getInput(0);
+  inputSeqInfoVec_.clear();
+  Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions,
+                              inputSeq.subSequenceStartPositions,
+                              inputSeqInfoVec_);
+  if (!useGpu_) {
+    if (inputLayers_.size() == 2U) {
+      startIdsOnCpu_ = config_.select_first() ? getInputValue(1) : nullptr;
+      endIdsOnCpu_ = config_.select_first() ? nullptr : getInputValue(1);
+    } else if (inputLayers_.size() == 3U) {
+      startIdsOnCpu_ = getInputValue(1);
+      endIdsOnCpu_ = getInputValue(2);
+    }
+  } else {
+    copySliceIdsToCpu();
+  }
+
+  /*
+   * calculate the selected row indices in a batch, and build the output
+   * sequence information.
+   */
+  calSelectedRows(startIdsOnCpu_, endIdsOnCpu_);
+
+  resetOutput(selectedRows_.size(), getSize());
+
+  getOutputValue()->selectRows(*getInputValue(0), *rowIndice_);
+}
+
+void SequenceSliceLayer::backward(const UpdateCallback& callback) {
+  getOutputGrad()->addToRows(*getInputGrad(0), *rowIndice_);
+}
+
+}  // namespace paddle

paddle/gserver/layers/SubNestedSequenceLayer.cpp

@@ -52,23 +52,34 @@ private:
    *   ]
    *
    * ths output is saved to private member rowIndice_;
-   * [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,
-   *  16,17,18,19,20,21,22,23,24,25,26,27]
+   * [0,1,2,3,4,5,6,7,8,9,15,16,17,18,19,20,21,23,24,25,26,27]
    */
 
-  void calSelectedCols(const MatrixPtr selectedIndices,
+  void calSelectedRows(const MatrixPtr selectedIndices,
                        const std::vector<std::vector<int>>& inputSeqInfo);
 
-  // if the second input of this layer is on GPU memory, copy it to CPU memory.
+  /*
+   * TODO(caoying)
+   * In PaddePaddle, currently all matrices are real number types,
+   * but the second is some selected indices of the give sequence to trim
+   * the nested sequence, are actually filled with int types so that storing
+   * int types information in real number matrices is very dangerous, since
+   * real numbers will be convered to int types. If a user fills this matrix
+   * himself, invalid data may occor.
+   *
+   * if the second input of this layer is on GPU memory, copy it to CPU memory.
+   */
   MatrixPtr selIdsCpu_;
 
-  // reorganized sequenceStartPositions and subSequenceStartPositions
-  // into a 2d vector to facilitate the sequence selection process.
+  /*
+   * reorganize sequenceStartPositions and subSequenceStartPositions
+   * into a 2d vector to facilitate the sequence selection process.
+   */
   std::vector<std::vector<int>> inputSeqInfoVec_;
 
-  // the final selected row indices in a batch,
-  // rowIdx_ and selectedRows_ actually share a same memory.
+  /* store the final selected row indices in a batch */
   IVectorPtr rowIndice_;
+  /* rowIndice_ and selectedRows_ actually share a same memory. */
   std::vector<int> selectedRows_;
 };
 
@@ -83,7 +94,7 @@ bool SubNestedSequenceLayer::init(const LayerMap& layerMap,
   return true;
 }
 
-void SubNestedSequenceLayer::calSelectedCols(
+void SubNestedSequenceLayer::calSelectedRows(
     const MatrixPtr selectedIndices,
     const std::vector<std::vector<int>>& inputSeqInfo) {
   selectedRows_.clear();
@@ -160,7 +171,7 @@ void SubNestedSequenceLayer::forward(PassType passType) {
   Argument::reorganizeSeqInfo(inputSeq.sequenceStartPositions,
                               inputSeq.subSequenceStartPositions,
                               inputSeqInfoVec_);
-  calSelectedCols(selIdsCpu_, inputSeqInfoVec_);
+  calSelectedRows(selIdsCpu_, inputSeqInfoVec_);
 
   resetOutput(selectedRows_.size(), getSize());
   getOutputValue()->selectRows(*getInputValue(0), *rowIndice_);

paddle/gserver/tests/CMakeLists.txt

@@ -34,6 +34,12 @@ add_unittest_without_exec(test_CRFLayerGrad
 add_test(NAME test_CRFLayerGrad
     COMMAND test_CRFLayerGrad)
 
+################ test_SeqSliceLayerGrad ####################
+add_unittest_without_exec(test_SeqSliceLayerGrad
+    test_SeqSliceLayerGrad.cpp
+    LayerGradUtil.cpp)
+add_test(NAME test_SeqSliceLayerGrad
+    COMMAND test_SeqSliceLayerGrad)
 
 add_unittest_without_exec(test_ActivationGrad
     test_ActivationGrad.cpp

paddle/gserver/tests/test_LayerGrad.cpp

@@ -12,6 +12,9 @@ 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. */
 
+#ifndef PADDLE_ONLY_CPU
+#include <cudnn.h>
+#endif
 #include <gtest/gtest.h>
 #include <string>
 #include <vector>
@@ -189,10 +192,16 @@ TEST(Projection, scaling) {
 void testProjectionConv(size_t groups, bool isDeconv) {
   const int NUM_FILTERS = 18;
   const int FILTER_SIZE = 2;
-  const int FILTER_SIZE_Y = 4;
+  const int FILTER_SIZE_Y = 2;
   const int CHANNELS = 3;
   const int IMAGE_SIZE = 16;
 
+#if CUDNN_VERSION >= 6000
+  const int DILATION = 2;
+#else
+  const int DILATION = 1;
+#endif
+
   ProjectionConfig conf;
   if (isDeconv) {
     conf.set_type("convt");
@@ -209,6 +218,8 @@ void testProjectionConv(size_t groups, bool isDeconv) {
   conv->set_padding_y(1);
   conv->set_stride(2);
   conv->set_stride_y(2);
+  conv->set_dilation(DILATION);
+  conv->set_dilation_y(DILATION);
   conv->set_groups(groups);
   if (isDeconv) {
     conv->set_filter_channels(NUM_FILTERS / conv->groups());
@@ -217,12 +228,12 @@ void testProjectionConv(size_t groups, bool isDeconv) {
   }
   conv->set_img_size(IMAGE_SIZE);
   int output_x = outputSize(conv->img_size(),
-                            conv->filter_size(),
+                            (conv->filter_size() - 1) * DILATION + 1,
                             conv->padding(),
                             conv->stride(),
                             /* caffeMode */ true);
   int output_y = outputSize(conv->img_size(),
-                            conv->filter_size_y(),
+                            (conv->filter_size_y() - 1) * DILATION + 1,
                             conv->padding_y(),
                             conv->stride_y(),
                             /* caffeMode */ true);
@@ -424,27 +435,38 @@ void testConvLayer(const string& type, bool trans, bool useGpu) {
   config.layerConfig.set_partial_sum(1);
   config.layerConfig.set_shared_biases(true);
 
-  config.inputDefs.push_back({INPUT_DATA, "layer_0", 384, 288});
+  int dilation = 1;
+  if (type == "cudnn_conv") {
+#if CUDNN_VERSION >= 6000
+    dilation = 2;
+#else
+    dilation = 1;
+#endif
+  }
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 768, 192});
   LayerInputConfig* input = config.layerConfig.add_inputs();
   ConvConfig* conv = input->mutable_conv_conf();
   conv->set_filter_size(2);
-  conv->set_filter_size_y(3);
+  conv->set_filter_size_y(2);
   conv->set_channels(3);
   conv->set_padding(0);
   conv->set_padding_y(1);
   conv->set_stride(2);
   conv->set_stride_y(2);
+  conv->set_dilation(dilation);
+  conv->set_dilation_y(dilation);
   conv->set_groups(1);
   conv->set_filter_channels(conv->channels() / conv->groups());
   conv->set_img_size(16);
-  conv->set_img_size_y(8);
+  conv->set_img_size_y(16);
   conv->set_output_x(outputSize(conv->img_size(),
-                                conv->filter_size(),
+                                (conv->filter_size() - 1) * dilation + 1,
                                 conv->padding(),
                                 conv->stride(),
                                 /* caffeMode */ true));
   conv->set_output_y(outputSize(conv->img_size_y(),
-                                conv->filter_size_y(),
+                                (conv->filter_size_y() - 1) * dilation + 1,
                                 conv->padding_y(),
                                 conv->stride_y(),
                                 /* caffeMode */ true));
@@ -828,9 +850,27 @@ TEST(Layer, square_error_weighted) {
   }
 }
 
+TEST(Layer, huber_regression_loss) {
+  TestConfig config;
+  config.layerConfig.set_type("huber_regression");
+  config.biasSize = 0;
+
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 10, 0});
+  config.inputDefs.push_back({INPUT_DATA_TARGET, "layer_1", 10, 0});
+  config.layerConfig.add_inputs();
+  config.layerConfig.add_inputs();
+
+  for (auto useGpu : {false, true}) {
+    for (auto delta : {1, 3, 5}) {
+      config.layerConfig.set_delta(delta);
+      testLayerGrad(config, "huber_regression", 100, /* trans */ false, useGpu);
+    }
+  }
+}
+
 TEST(Layer, huber_two_class) {
   TestConfig config;
-  config.layerConfig.set_type("huber");
+  config.layerConfig.set_type("huber_classification");
   config.biasSize = 0;
 
   config.inputDefs.push_back({INPUT_DATA, "layer_0", 1, 0});
@@ -839,7 +879,7 @@ TEST(Layer, huber_two_class) {
   config.layerConfig.add_inputs();
 
   for (auto useGpu : {false, true}) {
-    testLayerGrad(config, "huber", 100, /* trans */ false, useGpu);
+    testLayerGrad(config, "huber_two_class", 100, /* trans */ false, useGpu);
   }
 }
 
@@ -2007,6 +2047,21 @@ TEST(Layer, RowL2NormLayer) {
   }
 }
 
+TEST(Layer, ScaleShiftLayer) {
+  const size_t batchSize = 16;
+  const size_t size = 32;
+  TestConfig config;
+  config.layerConfig.set_type("scale_shift");
+  config.layerConfig.set_size(size);
+  config.biasSize = 1;
+  config.inputDefs.push_back(
+      {INPUT_DATA, "input", /* dim= */ size, /* paraSize= */ 1});
+  config.layerConfig.add_inputs();
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "scale_shift", batchSize, false, useGpu, false);
+  }
+}
+
 int main(int argc, char** argv) {
   testing::InitGoogleTest(&argc, argv);
   initMain(argc, argv);

paddle/gserver/tests/test_NetworkCompare.cpp

@@ -269,7 +269,8 @@ TEST(Compare, img_conv2) {
   bool useGpu = FLAGS_use_gpu;
   double eps = FLAGS_checkgrad_eps;
   FLAGS_use_gpu = true;
-  FLAGS_checkgrad_eps = 1e-2;
+  // Sometimes, this unit test will fail with 1e-2
+  FLAGS_checkgrad_eps = 4e-2;
   compareNetwork(config_file_a, config_file_b);
   FLAGS_use_gpu = useGpu;
   FLAGS_checkgrad_eps = eps;

paddle/gserver/tests/test_SeqSliceLayerGrad.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 <gtest/gtest.h>
+#include "ModelConfig.pb.h"
+#include "paddle/gserver/layers/DataLayer.h"
+#include "paddle/trainer/Trainer.h"
+
+#include "LayerGradUtil.h"
+#include "paddle/testing/TestUtil.h"
+
+using namespace paddle;  // NOLINT
+using namespace std;     // NOLINT
+
+DECLARE_int32(gpu_id);
+DECLARE_bool(thread_local_rand_use_global_seed);
+
+const int MAX_SEQ_NUM = 17;
+const int MAX_SEQ_LEN = 23;
+const int MAX_BEAM_SIZE = 13;
+
+const size_t SEED = (size_t)(time(NULL));
+
+vector<real> randSampling(real range, int n) {
+  CHECK_GE(range, n);
+  vector<real> num(range);
+  iota(begin(num), end(num), 0.);
+  if (range == n) return num;
+
+  random_shuffle(begin(num), end(num));
+  num.resize(n);
+  sort(begin(num), end(num));
+  return num;
+}
+
+void genSeqInfo(vector<int>& seqStartPos, vector<int>& subSeqStartPos) {
+  seqStartPos.resize(1, 0);
+  subSeqStartPos.resize(1, 0);
+
+  srand(SEED);
+  int seqNum = 1 + (rand() % MAX_SEQ_NUM);
+  for (int i = 0; i < seqNum; ++i) {
+    int subSeqNum = 1 + (rand() % MAX_SEQ_NUM);
+    for (int j = 0; j < subSeqNum; ++j)
+      subSeqStartPos.push_back(subSeqStartPos.back() +
+                               (1 + (rand() % MAX_SEQ_LEN)));
+    seqStartPos.push_back(subSeqStartPos.back());
+  }
+}
+
+/*
+  generate start indices according to sequence start positions.
+ */
+void genStarts(vector<int>& seqStartPos,
+               vector<vector<real>>& starts,
+               size_t beamSize) {
+  starts.clear();
+  starts.resize(seqStartPos.size() - 1, vector<real>(beamSize, -1.));
+
+  for (size_t i = 0; i < seqStartPos.size() - 1; ++i) {
+    int seqLen = seqStartPos[i + 1] - seqStartPos[i];
+    vector<real> randStarts =
+        randSampling(seqLen, min(seqLen, static_cast<int>(beamSize)));
+    copy(begin(randStarts), end(randStarts), begin(starts[i]));
+  }
+}
+
+/*
+  generate end indices according to sequence start positions and start indices.
+ */
+void genEnds(vector<int>& seqStartPos,
+             vector<vector<real>>& starts,
+             vector<vector<real>>& ends,
+             size_t beamSize) {
+  CHECK_EQ(seqStartPos.size() - 1, starts.size());
+  ends.clear();
+  ends.resize(seqStartPos.size() - 1, vector<real>(beamSize, -1.));
+
+  for (size_t i = 0; i < starts.size(); ++i) {
+    for (size_t j = 0; j < starts[i].size(); ++j) {
+      int seqLen = seqStartPos[i + 1] - seqStartPos[i];
+      CHECK_GE(seqLen - 1, starts[i][j]);
+      if (starts[i][j] == -1.) break;
+      if (starts[i][j] == (seqLen - 1)) {
+        ends[i][j] = starts[i][j];
+      } else {
+        ends[i][j] = starts[i][j] + randSampling(seqLen - starts[i][j], 1)[0];
+      }
+    }
+  }
+}
+
+void genTestData(vector<int>& seqStartPos,
+                 vector<int>& subSeqStartPos,
+                 vector<vector<real>>& starts,
+                 vector<vector<real>>& ends,
+                 bool hasSubseq) {
+  size_t beamSize = 1 + (rand() % MAX_BEAM_SIZE);
+  genSeqInfo(seqStartPos, subSeqStartPos);
+
+  genStarts(hasSubseq ? subSeqStartPos : seqStartPos, starts, beamSize);
+  genEnds(hasSubseq ? subSeqStartPos : seqStartPos, starts, ends, beamSize);
+}
+
+template <typename T>
+void flatten2dVector(vector<vector<T>>& inVec, vector<T>& outVec) {
+  size_t totalSize{0};
+  for (auto const& items : inVec) totalSize += items.size();
+  outVec.reserve(totalSize);
+
+  for (auto& items : inVec)
+    move(items.begin(), items.end(), back_inserter(outVec));
+}
+
+void testSeqSliceLayer(bool hasSubseq,
+                       bool useGpu,
+                       vector<int>& seqStartPos,
+                       vector<int>& subSeqStartPos,
+                       vector<vector<real>>& starts,
+                       vector<vector<real>>& ends) {
+  // layer size is not crutial for this layer,
+  // so here use a small layer size in the unittest.
+  const size_t layerSize{4};
+  TestConfig config;
+  config.layerConfig.set_type("seq_slice");
+  config.layerConfig.set_size(layerSize);
+
+  // add the first input
+  MatrixPtr seqInputPtr =
+      Matrix::create(hasSubseq ? subSeqStartPos.back() : seqStartPos.back(),
+                     layerSize,
+                     false,
+                     false);
+  seqInputPtr->randomizeUniform();
+
+  if (hasSubseq) {
+    config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA,
+                                "seq_input",
+                                seqInputPtr,
+                                seqStartPos,
+                                subSeqStartPos});
+  } else {
+    config.inputDefs.push_back(
+        {INPUT_SELF_DEFINE_DATA, "seq_input", seqInputPtr, seqStartPos});
+  }
+  config.layerConfig.add_inputs();
+
+  // add start indices
+  if (starts.size()) {
+    vector<real> startsToVec;
+    flatten2dVector(starts, startsToVec);
+
+    MatrixPtr startMatrixPtr =
+        Matrix::create(starts.size(), starts[0].size(), false, false);
+    startMatrixPtr->copyFrom(startsToVec.data(), startsToVec.size());
+
+    config.inputDefs.push_back(
+        {INPUT_SELF_DEFINE_DATA, "starts", startMatrixPtr});
+    config.layerConfig.add_inputs();
+    config.layerConfig.set_select_first(true);
+  }
+
+  // add end indices
+  if (ends.size()) {
+    vector<real> endsToVec;
+    flatten2dVector(ends, endsToVec);
+
+    MatrixPtr endMatrixPtr =
+        Matrix::create(ends.size(), ends[0].size(), false, false);
+    endMatrixPtr->copyFrom(endsToVec.data(), endsToVec.size());
+
+    config.inputDefs.push_back({INPUT_SELF_DEFINE_DATA, "ends", endMatrixPtr});
+    config.layerConfig.add_inputs();
+    config.layerConfig.set_select_first(false);
+  }
+
+  testLayerGrad(config, "seq_slice", /*batchSize*/ 100, false, useGpu, false);
+}
+
+TEST(Layer, SeqSliceLayer) {
+  vector<int> seqStartPos;
+  vector<int> subSeqStartPos;
+  vector<vector<real>> starts;
+  vector<vector<real>> ends;
+
+  std::vector<bool> mode = {false};
+#ifndef PADDLE_ONLY_CPU
+  mode.push_back(true);
+#endif
+  genSeqInfo(seqStartPos, subSeqStartPos);
+  for (bool hasSubseq : {true, false}) {
+    LOG(INFO) << "hasSubSeq : " << hasSubseq;
+    genTestData(seqStartPos, subSeqStartPos, starts, ends, hasSubseq);
+    for (bool useGpu : mode) {
+      vector<vector<real>> tmp;
+      testSeqSliceLayer(
+          hasSubseq, useGpu, seqStartPos, subSeqStartPos, tmp, ends);
+      testSeqSliceLayer(
+          hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, tmp);
+      testSeqSliceLayer(
+          hasSubseq, useGpu, seqStartPos, subSeqStartPos, starts, ends);
+    }
+  }
+}
+
+int main(int argc, char** argv) {
+  initMain(argc, argv);
+  hl_start();
+  hl_init(FLAGS_gpu_id);
+  FLAGS_thread_local_rand_use_global_seed = true;
+  srand(1);
+  testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}

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/operators/CMakeLists.txt

@@ -42,9 +42,12 @@ function(op_library TARGET)
 endfunction()
 
 add_subdirectory(math)
+
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)
+op_library(gather_op SRCS gather_op.cc gather_op.cu)
 
 cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
+op_library(scatter_op SRCS scatter_op.cc scatter_op.cu)
 
 cc_library(net_op SRCS net_op.cc DEPS op_registry)
 cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
@@ -66,5 +69,7 @@ op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)
 
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
     DEPS framework_proto tensor op_registry operator net_op)
-op_library(uniform_random_op
-        SRCS uniform_random_op.cc uniform_random_op.cu)
+op_library(uniform_random_op SRCS uniform_random_op.cc uniform_random_op.cu)
+op_library(lookup_table_op SRCS lookup_table_op.cc lookup_table_op.cu)
+op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
+op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)

paddle/operators/fill_zeros_like_op.h

@@ -26,7 +26,7 @@ class FillZerosLikeKernel : public framework::OpKernel {
     auto* output = context.Output<framework::Tensor>("Dst");
     output->mutable_data<T>(context.GetPlace());
     auto t = framework::EigenVector<T>::Flatten(*output);
-    t.device(context.GetEigenDevice<Place>()) = t.constant(T(0));
+    t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
   }
 };