merge develop

README.md

@@ -2,8 +2,8 @@
 
 
 [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
-[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://doc.paddlepaddle.org/develop/doc/)
-[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://doc.paddlepaddle.org/develop/doc_cn/)
+[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
+[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
 [![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop)
 [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
 [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
@@ -36,7 +36,7 @@ Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddl
     examples:
 
       - Optimized math operations through SSE/AVX intrinsics, BLAS libraries
-      (e.g. MKL, ATLAS, cuBLAS) or customized CPU/GPU kernels.
+      (e.g. MKL, OpenBLAS, cuBLAS) or customized CPU/GPU kernels.
       - Highly optimized recurrent networks which can handle **variable-length**
       sequence without padding.
       - Optimized local and distributed training for models with high dimensional

benchmark/paddle/image/run_mkldnn_train.sh → benchmark/paddle/image/run_mkl_train.sh

@@ -28,6 +28,10 @@ function train() {
     --test_period=100 \
     --config_args=$args \
     2>&1 | tee ${log} 
+
+  avg_time=`tail ${log} -n 1 | awk -F ' ' '{print $8}' | sed 's/avg=//'`
+  fps=`awk 'BEGIN{printf "%.2f",('$bs' / '$avg_time' * 1000)}'`
+  echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
 }
 
 if [ ! -f "train.list" ]; then

benchmark/paddle/image/run_openblas_infer.sh

+set -e
+
+function clock_to_seconds() {
+  hours=`echo $1 | awk -F ':' '{print $1}'`
+  mins=`echo $1 | awk -F ':' '{print $2}'`
+  secs=`echo $1 | awk -F ':' '{print $3}'`
+  echo `awk 'BEGIN{printf "%.2f",('$secs' + '$mins' * 60 + '$hours' * 3600)}'`
+}
+
+function infer() {
+  unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
+  topology=$1
+  layer_num=$2
+  bs=$3
+  thread=`nproc`
+  if [ $thread -gt $bs ]; then
+    thread=$bs
+  fi
+  log="logs/infer-${topology}-${layer_num}-${thread}openblas-${bs}.log"
+
+  models_in="models/${topology}-${layer_num}/pass-00000/"
+  if [ ! -d $models_in ]; then
+    echo "./run_mkl_infer.sh to save the model first"
+    exit 0
+  fi
+  log_period=$((256 / bs))
+  paddle train --job=test \
+    --config="${topology}.py" \
+    --use_gpu=False \
+    --trainer_count=$thread \
+    --log_period=$log_period \
+    --config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True" \
+    --init_model_path=$models_in \
+    2>&1 | tee ${log}
+
+  # calculate the last 5 logs period time of 1280 samples,
+  # the time before are burning time.
+  start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
+  end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
+  start_sec=`clock_to_seconds $start`
+  end_sec=`clock_to_seconds $end`
+  fps=`awk 'BEGIN{printf "%.2f",(1280 / ('$end_sec' - '$start_sec'))}'`
+  echo "Last 1280 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
+  echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
+}
+
+if [ ! -f "train.list" ]; then
+  echo " " > train.list
+fi
+if [ ! -f "test.list" ]; then
+  echo " " > test.list
+fi
+if [ ! -d "logs" ]; then
+  mkdir logs
+fi
+
+# inference benchmark
+for batchsize in 1 2 4 8 16; do
+  infer googlenet v1 $batchsize
+  infer resnet 50 $batchsize
+  infer vgg 19 $batchsize
+done

benchmark/paddle/image/run_openblas_train.sh

+set -e
+
+function train() {
+  unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
+  topology=$1
+  layer_num=$2
+  bs=$3
+  thread=`nproc`
+  # each trainer_count use only 1 core to avoid conflict
+  log="logs/train-${topology}-${layer_num}-${thread}openblas-${bs}.log"
+  args="batch_size=${bs},layer_num=${layer_num}"
+  config="${topology}.py"
+  paddle train --job=time \
+    --config=$config \
+    --use_gpu=False \
+    --trainer_count=$thread \
+    --log_period=10 \
+    --test_period=100 \
+    --config_args=$args \
+    2>&1 | tee ${log} 
+
+  avg_time=`tail ${log} -n 1 | awk -F ' ' '{print $8}' | sed 's/avg=//'`
+  fps=`awk 'BEGIN{printf "%.2f",('$bs' / '$avg_time' * 1000)}'`
+  echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
+}
+
+if [ ! -f "train.list" ]; then
+  echo " " > train.list
+fi
+if [ ! -d "logs" ]; then
+  mkdir logs
+fi
+
+# training benchmark
+for batchsize in 64 128 256; do
+  train vgg 19 $batchsize
+  train resnet 50 $batchsize
+  train googlenet v1 $batchsize
+done

cmake/cblas.cmake

@@ -3,7 +3,7 @@
 # It will search MKLML, atlas, OpenBlas, reference-cblas in order.
 #
 # If any cblas implementation found, the following variable will be set.
-#    CBLAS_PROVIDER  # one of MKLML, ATLAS, OPENBLAS, REFERENCE
+#    CBLAS_PROVIDER  # one of MKLML, OPENBLAS, REFERENCE
 #    CBLAS_INC_DIR   # the include directory for cblas.
 #    CBLAS_LIBS      # a list of libraries should be linked by paddle.
 #                    # Each library should be full path to object file.
@@ -25,42 +25,6 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
   return()
 endif()
 
-## Then find atlas.
-set(ATLAS_ROOT $ENV{ATLAS_ROOT} CACHE PATH "Folder contains Atlas")
-set(ATLAS_INCLUDE_SEARCH_PATHS
-        ${ATLAS_ROOT}/include
-        /usr/include
-        /usr/include/atlas)
-set(ATLAS_LIB_SEARCH_PATHS
-        ${ATLAS_ROOT}/lib
-        /usr/lib
-        /usr/lib/blas/atlas
-        /usr/lib/atlas
-        /usr/lib/atlas-base   # special for ubuntu 14.04.
-    )
-find_path(ATLAS_INC_DIR NAMES cblas.h
-  PATHS ${ATLAS_INCLUDE_SEARCH_PATHS})
-find_path(ATLAS_CLAPACK_INC_DIR NAMES clapack.h
-  PATHS ${ATLAS_INCLUDE_SEARCH_PATHS})
-find_library(ATLAS_CBLAS_LIB NAMES cblas libcblas.so.3
-  PATHS ${ATLAS_LIB_SEARCH_PATHS})
-find_library(ATLAS_CLAPACK_LIB NAMES lapack_atlas liblapack_atlas.so.3
-  PATHS ${ATLAS_LIB_SEARCH_PATHS})
-
-if(ATLAS_CLAPACK_INC_DIR AND ATLAS_INC_DIR AND ATLAS_CBLAS_LIB AND ATLAS_CLAPACK_LIB)
-  set(CBLAS_FOUND ON)
-  set(CBLAS_PROVIDER ATLAS)
-  set(CBLAS_INC_DIR ${ATLAS_INC_DIR} ${ATLAS_CLAPACK_INC_DIR})
-  set(CBLAS_LIBRARIES ${ATLAS_CLAPACK_LIB} ${ATLAS_CBLAS_LIB})
-
-  add_definitions(-DPADDLE_USE_ATLAS)
-  add_definitions(-DLAPACK_FOUND)
-
-  message(STATUS "Found ATLAS (include: ${ATLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
-  message(STATUS "Found lapack in ATLAS (include: ${ATLAS_CLAPACK_INC_DIR})")
-  return()
-endif()
-
 ## Then find openblas.
 set(OPENBLAS_ROOT $ENV{OPENBLAS_ROOT} CACHE PATH "Folder contains Openblas")
 set(OPENBLAS_INCLUDE_SEARCH_PATHS

doc/api/index_cn.rst

@@ -7,3 +7,4 @@ API
     模型配置 <v2/model_configs.rst>
     数据访问 <v2/data.rst>
     训练与应用 <v2/run_logic.rst>
+    v2/fluid.rst

doc/api/v2/fluid/layers.rst

@@ -300,3 +300,7 @@ conv2d_transpose
 ..  autofunction:: paddle.v2.fluid.layers.conv2d_transpose
     :noindex:
 
+sequence_expand
+---------
+..  autofunction:: paddle.v2.fluid.layers.sequence_expand
+    :noindex:

doc/design/executor.md

 # Executor Design Doc
 
 ## Motivation
+In [fluid](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/fluid.md), we encourage the user to use deep learning programming paradigms to describe the training process. When the user-written Python program is executed, it will first create a protobuf message
+[`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and is conceptually like an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree).
 
-We use executor to do the runtime evaluation of a `ProgramDesc`.
+The executor runs the `ProgramDesc` like an interpreter. `ProgramDesc` contains the intrinsics (operators in this case) and variables which will be used, executor explicitly executes the stored precompiled code.
 
 ## Overview
 
-An executor takes a `ProgramDesc`, a `block_id` and a `Scope`.  The `ProgramDesc` is a list of blocks and each block contains the protobuf definition of all the parameters and operators. The `block_id` specifies the entrance block. And the `Scope` is the container of all the variable instance, which is persistent throughout different runs.
+An executor takes a `ProgramDesc`, a `block_id` and a `Scope`.  The `ProgramDesc` is a list of blocks and each block contains the protobuf definition of all the parameters and operators in the block. The `block_id` specifies the entrance block. And the `Scope` is the container of all the variable instances, which is persistent throughout different runs.
 
-### What does executor do?
+## Executor
 
-It evaluates all the operators in the `block_id`th block of a `ProgramDesc`.
+The `Executor` explicitly executes all the intrinsics (operators here) in the `block_id`th block of a `ProgramDesc`. Essentially, it instantiates Variables and Operators, then runs all the operators in sequence one-by-one.
+It is very similar to how a push stack frame works when entering a block, following which it cleans up all the temporary variables when a mini-batch is finished. It does not however, have the stack frame pop process.
 
-### What does executor NOT do?
+### The interface
+```c++
+  Executor(places);
+```
+A executor does not own any computing resources, a user can only construct an executor using the specified places.
 
-It does not do runtime optimization, meaning intelligently parse the dependency of each op a choose which one to be run and in which order they should be run.
+### Running an Executor
 
-It does not do graph partitioning, meaning dividing the `ProgramDesc` into several small pieces and executing them on different devices.
-
-## Implementation
-
-`Executor` evaluates a `ProgramDesc`. Essentially, it instantiates Variables and Operators, then run all the operators in sequence. [[code]](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc)
+```
+  void Run(ProgramDesc, Scope, block_id, create_local_scope);
+```
+An `Executor` only provides a unified way to execute `ProgramDesc`. `ProgramDesc` is the target that will be executed, the `Scope` specifies the variable container, the `block_id` indicates the entrance block and `create_local_scope` is a boolean that states whether it will destroy the temporary variables after the execution is finished.

doc/design/fluid.md

+# Design Doc: PaddlePaddle Fluid
+
+## Why Fluid
+
+When Baidu developed PaddlePaddle in 2013, the only well-known open source deep learning system at the time was Caffe.  However, when PaddlePaddle was open-sourced in 2016, many other choices were available. There was a challenge -- what is the need for open sourcing yet another deep learning framework?
+
+Fluid is the answer.  Fluid is similar to PyTorch and TensorFlow Eager Execution, which describes the "process" of training or inference using the concept of a model.  In fact in PyTorch, TensorFlow Eager Execution and Fluid, there is no  concept of a model at all. The details are covered in the sections below. Fluid is currently more extreme in the above mentioned idea than PyTorch and Eager Execution, and we are trying to push Fluid towards the directions of a compiler and a new programming language for deep learning.
+
+## The Evolution of Deep Learning Systems
+
+Deep learning infrastructure is one of the fastest evolving technologies. Within four years, there have already been three generations of technologies invented.
+
+| Existed since | model as sequence of layers | model as graph of operators | No model |
+|--|--|--|--|
+| 2013 | Caffe, Theano, Torch, PaddlePaddle | | |
+| 2015 | | TensorFlow, MxNet, Caffe2, ONNX, n-graph | |
+| 2016 | | | PyTorch, TensorFlow Eager Execution, PaddlePaddle Fluid |
+
+From the above table, we see that the deep learning technology is evolving towards getting rid of the concept of a model.  To understand the reasons behind this direction, a comparison of the *programming paradigms* or the ways to program deep learning applications using these systems, would be helpful. The following section goes over these.
+
+## Deep Learning Programming Paradigms
+
+With the systems listed as the first or second generation, e.g., Caffe or TensorFlow, an AI application training program looks like the following:
+
+```python
+x = layer.data("image")
+l = layer.data("label")
+f = layer.fc(x, W)
+s = layer.softmax(f)
+c = layer.mse(l, s)
+
+for i in xrange(1000): # train for 1000 iterations
+    m = read_minibatch()
+    forward({input=x, data=m}, minimize=c)
+    backward(...)
+
+print W # print the trained model parameters.
+```
+
+The above program includes two parts:
+
+1. The first part describes the model, and
+2. The second part describes the training process (or inference process) for the model.
+
+This paradigm has a well-known problem that limits the productivity of programmers. If the programmer made a mistake in configuring the model, the error messages wouldn't show up until the second part is executed and `forward` and `backward` propagations are performed. This makes it difficult for the programmer to debug and locate a mistake that is located blocks away from the actual error prompt.
+
+This problem of being hard to debug and re-iterate fast on a program is the primary reason that programmers, in general,  prefer PyTorch over the older systems.  Using PyTorch, we would write the above program as following:
+
+```python
+W = tensor(...)
+
+for i in xrange(1000): # train for 1000 iterations
+    m = read_minibatch()
+    x = m["image"]
+    l = m["label"]
+    f = layer.fc(x, W)
+    s = layer.softmax(f)
+    c = layer.mse(l, s)
+    backward()
+
+print W # print the trained model parameters.
+```
+
+We can see that the main difference is the moving the model configuration part (the first step) into the training loop.  This change would allow the mistakes in model configuration to be reported where they actually appear in the programming block.  This change also represents the model better, or its forward pass, by keeping the configuration process in the training loop.
+
+## Describe Arbitrary Models for the Future
+
+Describing the process instead of the model also brings Fluid, the flexibility to define different non-standard models that haven't been invented yet.
+
+As we write out the program for the process, we can write an RNN as a loop, instead of an RNN as a layer or as an operator.  A PyTorch example would look like the following:
+
+```python
+for i in xrange(1000):
+    m = read_minibatch()
+    x = m["sentence"]
+    for t in xrange x.len():
+        h[t] = the_step(x[t])
+```        
+
+With Fluid, the training loop and the RNN in the above program are not really Python loops, but just a "loop structure" provided by Fluid and implemented in C++ as the following:
+
+```python
+train_loop = layers.While(cond)
+with train_loop.block():
+  m = read_minibatch()
+  x = m["sentence"]
+  rnn = layers.While(...)
+  with rnn.block():
+    h[t] = the_step(input[t])
+```    
+
+An actual Fluid example is described  [here](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44).
+
+From the example, the Fluid programs look very similar to their PyTorch equivalent programs, except that Fluid's loop structure, wrapped with Python's `with` statement, could run much faster than just a Python loop.
+
+We have more examples of the [`if-then-else`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/if_else_op.md) structure of Fluid.
+
+## Turing Completeness
+
+In computability theory, a system of data-manipulation rules, such as a programming language, is said to be Turing complete if it can be used to simulate any Turing machine.  For a programming language, if it provides if-then-else and loop, it is Turing complete.  From the above examples, Fluid seems to be Turing complete; however, it is noteworthy to notice that there  is a slight difference between the `if-then-else` of Fluid and that of a programming language. The difference being that the former runs both of its branches and splits the input mini-batch into two -- one for the True condition and another for the False condition. This hasn't been researched in depth if this is equivalent to the `if-then-else` in programming languages that makes them Turing-complete.  Based on a conversation with [Yuang Yu](https://research.google.com/pubs/104812.html), it seems to be the case but this needs to be looked into in-depth.
+
+## The Execution of a Fluid Program
+
+There are two ways to execute a Fluid program.  When a program is executed, it creates a protobuf message [`ProgramDesc`](https://github.com/PaddlePaddle/Paddle/blob/a91efdde6910ce92a78e3aa7157412c4c88d9ee8/paddle/framework/framework.proto#L145) that describes the process and is conceptually like an [abstract syntax tree](https://en.wikipedia.org/wiki/Abstract_syntax_tree).
+
+There is a C++ class [`Executor`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h), which runs a `ProgramDesc`, similar to how an interpreter runs a Python program.
+
+Fluid is moving towards the direction of a compiler, which is explain in more detail later in this article.
+
+## Backward Compatibility of Fluid
+
+Given all the advantages from the removal of the concept of a *model*, hardware manufacturers might still prefer the existence of the concept of a model, so it would be easier for them to support multiple frameworks all at once and could run a trained model during inference.  For example, Nervana, a startup company acquired by Intel, has been working on an XPU that reads the models in the format known as [n-graph](https://github.com/NervanaSystems/ngraph).  Similarly, [Movidius](https://www.movidius.com/) is producing a mobile deep learning chip that reads and runs graphs of operators.  The well-known [ONNX](https://github.com/onnx/onnx) is also a file format of graphs of operators.
+
+For Fluid, we can write a converter that extracts the parts in the `ProgramDesc` protobuf message, converts them into a graph of operators, and exports the graph into the ONNX or n-graph format.
+
+## Towards a Deep Learning Language and the Compiler
+
+We can change the `if-then-else` and loop structure a little bit in the above Fluid example programs, to make it into a new programming language, different than Python.
+
+Even if we do not invent a new language, as long as we get the `ProgramDesc` message filled in, we can write a transpiler, which translates each invocation to an operator, into a C++ call to a kernel function of that operator. For example, a transpiler that weaves the CUDA kernels outputs an NVIDIA-friendly C++ program, which can be built using `nvcc`.  Another transpiler could generate MKL-friendly code that should be built using `icc` from Intel.  More interestingly, we can translate a Fluid program into its distributed version of two `ProgramDesc` messages, one for running on the trainer process, and the other one for the parameter server.  For more details of the last example, the [concurrent programming design](concurrent_programming.md) document would be a good pointer.  The following figure explains the proposed two-stage process:
+
+![](fluid-compiler.png)

doc/design/mkl/mkl_packed.md

+# Intel® MKL Packed on PaddlePaddle: Design Doc
+
+
+## Contents
+
+- [Overview](#overview)
+- [Key Points](#key-points) 
+   - [Background](#background)
+   - [Solution](#solution)
+- [Actions](#actions)
+    - [CMake](#cmake)
+	- [Layers](#layers)
+	- [Unit Tests](#unit-tests)
+	- [Python API](#python-api)
+	- [Benchmarking](#benchmarking)
+
+
+## Overview
+我们计划将 Intel® MKL 中引入的 GEMM Packed APIs\[[1](#references)\] 集成到 PaddlePaddle 中，充分发挥英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
+现阶段的优化主要针对 Recurrent Neural Network（以下简称RNN）相关层（包括`RecurrentLayer`, `GatedRecurrentLayer`和`LstmLayer`）， 以及 PaddlePaddle V1 API。
+
+## Key Points
+
+### Background
+目前PaddlePaddle采用了 Intel® MKL库的[cblas_?gemm](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm)函数，这个函数本身会在计算前将原数据转换为更适合英特尔平台的内部格式。
+
+1. 转换耗时 \
+这一数据格式的转换操作（Packing），在问题本身的计算量比较小的时候，显得相对来说较为耗时。例如在DeepSpeech2 \[[2](#references)\] 的Vanilla RNN部分中，矩阵大小是`batch_size * 2048`。
+2. 转换冗余 \
+由于在现有的某些情况下（例如RNN），多次调用 cblas_?gemm 会使用相同的原数据，因此，每次调用时对原数据的重复Packing便成为了冗余。
+
+为了最大程度减少多次调用 cblas_?gemm 在Packing上的耗时，Intel® MKL 引入了以下四个API:
+   * [cblas_?gemm_alloc](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-alloc)
+   * [cblas_?gemm_pack](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-pack)
+   * [cblas_?gemm_compute](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-compute)
+   * [cblas_?gemm_free](https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-free)
+
+通过使用这些API，我们可以先完成对原数据的Packing操作，再把已转换为Packed格式的数据传递给那些复用同一数据的gemm_compute函数，从而避免了Packing冗余。
+
+### Solution
+在RNN的情况下，同一次前向、后向（forward/backward）过程中所有时间步（time step）共享同一个权重（weight）。当只做推断（inference）时，各次前向之间也都使用了相同的权重，没有必要在每次前向中每个时间步的计算时对权重进行重复的Packing操作。
+
+我们通过使用新引入的GEMM Packed APIs，在层初始化的时候，先完成对权重的Packing操作，然后在前向，后向时复用已经转换过的权重，并在每次权重更新后，对新的权重进行转换用于下次迭代。
+
+* 优化前，对于序列长度（sequence length）为`T`的网络模型（model）, `N`次迭代执行的转换次数为：
+  - `inference`： `N * T`  
+  - `training`： `2 * N * T`
+* 优化后，对于同样设置的网络模型，其转换次数减少至：
+  - `inference`： `1`    
+  - `training`： `2 * N`
+
+## Actions
+
+添加的相关文件和目录结构如下：
+
+```txt
+PaddlePaddle/Paddle
+├── ...
+└── paddle/
+    ├── ...
+    └── gserver/
+        ├── ...
+        ├── layers/
+        │   ├── ...
+        │   ├── MKLPackedRecurrentLayer.*
+        |   ├── MKLPackedGatedRecurrentLayer.*
+        |   ├── MKLPackedLstmLayer.*
+        |   └── MKLPackedGemm.h
+        └── tests/
+            ├── ...
+            └── test_MKLPacked.cpp
+```
+
+### CMake
+在对应的`CMakeLists.txt`中根据`WITH_MKL`是否打开，来决定是否开启MKL Packed相关功能。
+
+### Layers
+所有的`MKLPacked*Layer`都继承于PaddlePaddle的基类`Layer`, 并添加头文件 `MKLPackedGemm.h`，该文件对相关GEMM Packed APIs做了封装。
+
+### Unit Tests
+我们会添加`test_MKLPacked.cpp`用于MKL Packed优化后layer的测试。
+对于每一个新加的RNN layer，我们会对比如下2个方面：
+1. 对比优化后layer自身，sequence mode（`rnn_use_batch=false`）与batch mode(`rnn_use_batch=true`)的结果。
+2. 对比优化后layer与相对应的PaddlePaddle原有layer, 在batch mode下的结果。
+
+### Python API
+计划在`paddle/utils.Flags`中添加`use_mkl_packed`的flag，用于选择是否使用相关功能，并且当编译时`WITH_MKL=ON`的情况下，默认设置为`true`。
+
+同时，在`python/paddle/trainer/config_parser.py`中对应的layer处，添加`use_mkl_packed`这个选择，方便用户在Python端选择是否启用这个功能。
+
+具体实现方式比如：
+
+```python
+use_mkl_packed = bool(int(g_command_config_args.get("use_mkl_packed", 0)))
+if use_mkl_packed:
+    self.layer_type = mkl_packed_*
+```
+
+所有相关的`layer_type`会以*mkl_packed_*开头，这些会在`MKLPacked*Layer`注册layer的时候保证，以示区分。 
+
+
+### Benchmarking
+会添加相应的脚本用于测试和对比在使用MKL Packed recurrent layers 前后的网络性能。
+
+## References 
+1. [Introducing the new Packed APIs for GEMM](https://software.intel.com/en-us/articles/introducing-the-new-packed-apis-for-gemm)
+2. [DeepSpeech2 on PaddlePaddle](https://github.com/PaddlePaddle/DeepSpeech#deepspeech2-on-paddlepaddle)
+

doc/design/mkldnn/README.MD → doc/design/mkl/mkldnn.md

@@ -208,4 +208,3 @@ if use_mkldnn
 但是在PaddlePaddle中，无论是重构前的layer还是重构后的op，都不会想要知道next layer/op的信息。
 4. MKL-DNN的高性能格式与PaddlePaddle原有的`NCHW`不同(PaddlePaddle中的cuDNN部分使用的也是`NCHW`，所以不存在这个问题)。
 所以需要引入一个转换方法，并且只需要在必要的时候转换这种格式，才能更好的发挥MKL-DNN的性能。
-

doc/design/paddle_nccl.md

+# Design Doc: NCCL support in Paddle Fluid
+
+## Abstract
+
+This Design Doc refers to the NCCL feature in  paddle.  We propose an approach to support NCCL library both on a single machine and multiple machines. We wrapper the NCCL primitives `Broadcast`, `Allreduce`, `Reduce` as operators to utilize Multi-GPU powers in one script.
+
+
+## Motivation
+
+[NCCL](https://developer.nvidia.com/nccl) is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. With NCCL library, we can easily accelerate the training in parallel. 
+
+- Pros
+1. easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
+1. high performance in NVIDIA GPUs.
+1. MPI like primitives, which have low learning cost for users.
+
+- Cons
+1. Only design for NVIDIA GPUs, not a general multi-device solution.
+1. Although NCCL1 is opensourced under BSD license, but NCCL2 is not opensourced anymore.
+
+At the beginning of training, the framework needs to distribute the same parameters to every GPU, and merge the gradients at any time user interests.
+
+As a result, during training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the operator with correct place information.
+
+Besides, it needs interfaces to synchronize model update with each different GPU Cards. 
+
+## Implementation
+
+As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created.
+
+### Transpiler
+
+To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the transpiler compiles the user defined operation graph into sub-graphs to be executed on different devices.
+
+1. The user-defined model will be a single device program
+
+2. Broadcast/Reduce operators between GPUs will be inserted into the program, even for the multi-node, may insert the `Send`, `Recv` operator.
+
+   *Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, [Send, Recv](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter) in multiple machines*
+
+   <img src="images/multigpu_before_convert.png" width="300"/>
+
+After compiling, the graph as shows
+
+<img src="images/multigpu_allreduce.png" width="1000"/>
+
+Operators are added to the sub-graphs. Every GPU assigned a role of `rank0`, `rank1` etc. 
+
+- **Broadcast**. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from`rank0` GPU.
+- **AllReduce**. AllReduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based  communicating method, avoid of the bottle neck in a single GPU.
+
+Need to notice that AllReduce operator force GPUs synchronized at that point. The whole training process in asynchronous or synchronous mode depends on the AllReduce point in the graph.
+
+As it shown in the picture, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
+
+- **AllReduce**
+  Need to note that our AllReduce operator is a ring-base AllReduce implementation. If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, NCCL2 built-in AllReduce will only utilize the communicating resource during synchronization, then update the gradient will be a subsequent phase. In fact, we can amortize the update gradient time cost into the communicating phase. The process is
+1. Every parameter has its root card. That card will responsible for aggregating the gradients from GPUs.
+2. The whole model's parameter will be hashed to different root card, ensure the load balance between GPUs.
+3. Logically neighberhood card will start send parameter to the next one. After one round, the parameter main card will aggregate the full gradients.
+4. Then the root card will optimize the parameter.
+5. This parameter card will send its optimized result to its neighberhood, then the neighberhood will send parameter to its next one.
+6. Finish the sychronization round.
+
+The total time cost will be 2 * (n-1) * per-parameter-send-time, we reach the goal of amortize the upgrade time into communicating phase.

doc/design/support_new_device.md

+# Design Doc: Supporting new Device/Library
+
+## Background
+
+Deep learning has a high demand for computing resources. New high-performance devices and computing libraries are appearing very frequently. Deep learning frameworks have to integrate these high-performance devices and computing libraries flexibly and efficiently.
+
+On one hand, hardware and computing libraries usually do not have a one-to-one correspondence. For example,Intel CPUs support Eigen and MKL computing libraries while Nvidia GPUs support Eigen and cuDNN computing libraries. We have to implement operator specific kernels for each computing library.
+
+On the other hand, users usually do not want to care about the low-level hardware and computing libraries when writing a neural network configuration. In Fluid, `Layer` is exposed in `Python`, and `Operator` is exposed in `C++`. Both `Layer` and `Operator` are hardware independent.
+
+So, how to support a new Device/Library in Fluid becomes a challenge.
+
+
+## Basic: Integrate A New Device/Library
+
+For a general overview of fluid, please refer to the [overview doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/read_source.md).
+
+There are mainly three parts that we have to consider while integrating a new device/library:
+
+- Place and DeviceContext: indicates the device id and manages hardware resources
+
+- Memory and Tensor: malloc/free data on certain device
+
+- Math Functor and OpKernel: implement computing unit on certain devices/libraries
+
+### Place and DeviceContext
+
+
+#### Place
+Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent different devices and computing libraries. There are inheritance relationships between different kinds of `Place`.
+
+```
+        |   CPUPlace   --> MKLDNNPlace
+Place --|   CUDAPlace  --> CUDNNPlace
+        |   FPGAPlace
+```
+
+And `Place` is defined as follows:
+
+```
+typedef boost::variant<CUDAPlace, CPUPlace, FPGAPlace> Place;
+```
+
+#### DeviceContext
+
+Fluid uses class [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L30) to manage the resources in different hardwares, such as CUDA stream in `CDUADeviceContext`. There are also inheritance relationships between different kinds of `DeviceContext`.
+
+
+```
+                /->  CPUDeviceContext   --> MKLDeviceContext
+DeviceContext ---->  CUDADeviceContext  --> CUDNNDeviceContext
+                \->  FPGADeviceContext
+```
+
+An example of Nvidia GPU is as follows:
+
+- DeviceContext
+
+
+```
+class DeviceContext {
+  virtual Place GetPlace() const = 0;
+};  
+```
+
+
+- CUDADeviceContext
+
+
+```
+class CUDADeviceContext : public DeviceContext {
+  Place GetPlace() const override { return place_; }
+private:
+  CUDAPlace place_;
+  cudaStream_t stream_; 
+  cublasHandle_t cublas_handle_;
+  std::unique_ptr<Eigen::GpuDevice> eigen_device_;  // binds with stream_
+};
+```
+
+- CUDNNDeviceContext
+
+```
+class CUDNNDeviceContext : public CUDADeviceContext {
+  private:
+    cudnnHandle_t cudnn_handle_;
+};
+```
+
+
+### Memory and Tensor
+
+
+#### memory module
+
+Fluid provides the following [memory interfaces](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/memory/memory.h#L36):
+
+```
+template <typename Place>
+void* Alloc(Place place, size_t size);
+
+template <typename Place>
+void Free(Place place, void* ptr);
+
+template <typename Place>
+size_t Used(Place place);
+```
+
+To implementing these interfaces, we have to implement MemoryAllocator for different Devices
+
+
+#### Tensor
+
+[Tensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h#L36) holds data with some shape in a specific Place.
+
+```cpp
+class Tensor {
+ public:
+  /*! Return a pointer to mutable memory block. */
+  template <typename T>
+  inline T* data();
+
+  /**
+   * @brief   Return a pointer to mutable memory block.
+   * @note    If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(platform::Place place);
+
+  /**
+   * @brief     Return a pointer to mutable memory block.
+   *
+   * @param[in] dims    The dimensions of the memory block.
+   * @param[in] place   The place of the memory block.
+   *
+   * @note      If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(DDim dims, platform::Place place);
+
+  /*! Resize the dimensions of the memory block. */
+  inline Tensor& Resize(const DDim& dims);
+
+  /*! Return the dimensions of the memory block. */
+  inline const DDim& dims() const;
+
+ private:
+  /*! holds the memory block if allocated. */
+  std::shared_ptr<Placeholder> holder_;
+
+  /*! points to dimensions of memory block. */
+  DDim dim_;
+};
+```
+
+`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configure its shape, and then call `mutuable_data` to allocate the actual memory.
+
+```cpp
+paddle::framework::Tensor t;
+paddle::platform::CPUPlace place;
+// set size first
+t.Resize({2, 3});
+// allocate memory on CPU later
+t.mutable_data(place);
+```
+
+
+
+### Math Functor and OpKernel
+
+Fluid implements computing units based on different DeviceContexts. Some computing units are shared between operators. This common part will be put in operators/math directory as basic Functors.
+
+Let's take [MaxOutFunctor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/math/maxouting.h#L27) as an example:
+
+The interface is defined in header file.
+
+```
+template <typename DeviceContext, typename T>
+class MaxOutFunctor {
+ public:
+  void operator()(const DeviceContext& context, const framework::Tensor& input,
+                  framework::Tensor* output, int groups);
+};
+```
+
+CPU implemention is in .cc file
+
+```
+template <typename T>
+class MaxOutFunctor<platform::CPUDeviceContext, T> {
+  public:
+  void operator()(const platform::CPUDeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* output,
+                  int groups) {
+                  ...
+                  }
+};
+```
+
+CUDA implemention is in .cu file
+
+```
+template <typename T>
+class MaxOutFunctor<platform::CUDADeviceContext, T> {
+ public:
+  void operator()(const platform::CUDADeviceContext& context,
+                  const framework::Tensor& input, framework::Tensor* output,
+                  int groups) {
+                  ...
+                  }
+};                  
+```
+
+
+We get computing handle from a concrete DeviceContext, and make compution on tensors.
+
+The implemention of `OpKernel` is similar to math functors, the extra thing we need to do is to register the OpKernel in a global map.
+
+Fluid provides different register interfaces in op_registry.h
+
+
+Let's take [Crop](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/crop_op.cc#L134) operator as an example:
+
+In .cc file:
+
+```
+REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel<float>);
+REGISTER_OP_CPU_KERNEL(
+    crop_grad, ops::CropGradKernel<paddle::platform::CPUDeviceContext, float>);
+```
+
+In .cu file:
+
+```
+REGISTER_OP_CUDA_KERNEL(crop, ops::CropKernel<float>);
+REGISTER_OP_CUDA_KERNEL(
+    crop_grad, ops::CropGradKernel<paddle::platform::CUDADeviceContext, float>);
+```
+
+
+## Advanced topics: How to switch between different Device/Library
+
+Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
+
+
+We will discuss how to implement an efficient OpKernel switch policy. 
+
+- TBD

doc/faq/build_and_install/index_cn.rst

@@ -14,7 +14,7 @@
 
     $ export CUDA_SO="$(\ls usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
     $ export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-    $ docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddlepaddle:latest-gpu
+    $ docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu
 
 更多关于Docker的安装与使用, 请参考 `PaddlePaddle Docker 文档 <http://www.paddlepaddle.org/doc_cn/build_and_install/install/docker_install.html>`_ 。
 

doc/getstarted/build_and_install/docker_install_cn.rst

@@ -114,7 +114,7 @@ PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Note
 
   .. code-block:: bash
 
-     nvidia-docker run -it -v $PWD:/work paddledev/paddle:latest-gpu /bin/bash
+     nvidia-docker run -it -v $PWD:/work paddlepaddle/paddle:latest-gpu /bin/bash
 
 **注: 如果没有安装nvidia-docker，可以尝试以下的方法，将CUDA库和Linux设备挂载到Docker容器内：**
 
@@ -122,7 +122,7 @@ PaddlePaddle Book是为用户和开发者制作的一个交互式的Jupyter Note
 
      export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
      export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-     docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:latest-gpu
+     docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu
 
 **关于AVX：**
 

doc/getstarted/build_and_install/docker_install_en.rst

@@ -122,7 +122,7 @@ GPU driver installed before move on.
 
   .. code-block:: bash
 
-     nvidia-docker run -it -v $PWD:/work paddledev/paddle:latest-gpu /bin/bash
+     nvidia-docker run -it -v $PWD:/work paddlepaddle/paddle:latest-gpu /bin/bash
 
 **NOTE: If you don't have nvidia-docker installed, try the following method to mount CUDA libs and devices into the container.**
 
@@ -130,7 +130,7 @@ GPU driver installed before move on.
 
      export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"
      export DEVICES=$(\ls /dev/nvidia* | xargs -I{} echo '--device {}:{}')
-     docker run ${CUDA_SO} ${DEVICES} -it paddledev/paddle:latest-gpu
+     docker run ${CUDA_SO} ${DEVICES} -it paddlepaddle/paddle:latest-gpu
 
 **About AVX:**
 

doc/getstarted/concepts/src/infer.py

+import paddle.v2 as paddle
+import numpy as np
+
+paddle.init(use_gpu=False)
+x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(2))
+y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
+
+# loading the model which generated by training
+with open('params_pass_90.tar', 'r') as f:
+    parameters = paddle.parameters.Parameters.from_tar(f)
+
+# Input multiple sets of data，Output the infer result in a array.
+i = [[[1, 2]], [[3, 4]], [[5, 6]]]
+print paddle.infer(output_layer=y_predict, parameters=parameters, input=i)
+# Will print:
+# [[ -3.24491572]
+#  [ -6.94668722]
+#  [-10.64845848]]

doc/getstarted/concepts/src/train.py

@@ -26,6 +26,11 @@ def event_handler(event):
         if event.batch_id % 1 == 0:
             print "Pass %d, Batch %d, Cost %f" % (event.pass_id, event.batch_id,
                                                   event.cost)
+    # product model every 10 pass
+    if isinstance(event, paddle.event.EndPass):
+        if event.pass_id % 10 == 0:
+            with open('params_pass_%d.tar' % event.pass_id, 'w') as f:
+                trainer.save_parameter_to_tar(f)
 
 
 # define training dataset reader

doc/getstarted/concepts/use_concepts_cn.rst

@@ -147,4 +147,9 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和
 ..  literalinclude:: src/train.py
     :linenos:
 
+使用以上训练好的模型进行预测，取其中一个模型params_pass_90.tar，输入需要预测的向量组，然后打印输出：
+
+..  literalinclude:: src/infer.py
+    :linenos:
+
 有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 <http://book.paddlepaddle.org/index.html>`_。

doc/howto/dev/contribute_to_paddle_cn.md

@@ -76,18 +76,18 @@ no changes added to commit (use "git add" and/or "git commit -a")
 
 ## 构建和测试
 
-编译 PaddlePaddle 的源码以及生成文档需要多种开发工具。为了方便大家，我们的标准开发流程是把这些工具都装进一个Docker image，称为*开发镜像*，通常名字是 `paddle:dev`。然后所有用 `cmake && make` 的地方（比如IDE配置里）都用 `docker run paddle:dev`来代替。
+编译 PaddlePaddle 的源码以及生成文档需要多种开发工具。为了方便大家，我们的标准开发流程是把这些工具都装进一个Docker image，称为*开发镜像*，通常名字是 `paddle:latest-dev` 或者 `paddle:[version tag]-dev` 如 `paddle:0.11.0-dev`。然后所有用 `cmake && make` 的地方（比如IDE配置里）都用 `docker run paddle:latest-dev`来代替。
 
 如要build这个开发镜像，在源码目录树的根目录中运行：
 
 ```bash
-➜  docker build -t paddle:dev .
+➜  docker build -t paddle:latest-dev .
 ```
 
 随后可以用这个开发镜像开始build PaddlePaddle的源码。比如如果要build一个不依赖GPU，但是支持AVX指令集，并且包括unit tests的PaddlePaddle，可以：
 
 ```bash
-➜  docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=ON" paddle:dev
+➜  docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" paddle:latest-dev
 ```
 
 这个过程除了编译PaddlePaddle为 `./build/libpaddle.so`，并且输出一个 `./build/paddle.deb`文件之外，还会输出一个 `build/Dockerfile`。我们只需要运行下面命令把编译好的PaddlePaddle打包成一个*生产镜像*（`paddle:prod`）：
@@ -99,7 +99,7 @@ no changes added to commit (use "git add" and/or "git commit -a")
 如果要运行所有的单元测试，可以用如下命令：
 
 ```bash
-➜  docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest"
+➜  docker run -it -v $(pwd):/paddle paddle:latest-dev bash -c "cd /paddle/build && ctest"
 ```
 
 关于构建和测试的更多信息，请参见[这篇文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)。

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)
+ - [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)
+ - [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)
+   - [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)
+ - [注意事项](#注意事项)
 
 
 ## 概念简介
@@ -30,8 +31,8 @@
 --------------  | :----------------------
 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实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，GPU 实现在`.cu`文件中。
-注册Op           | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中
+Kernel实现       | CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，CUDA 实现在`.cu`文件中。
+注册Op           | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，CUDA实现在`.cu`文件中
 
 
 实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
@@ -43,7 +44,7 @@ Kernel实现       | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU
 ## 实现C++类
 
 
-### 1. 定义ProtoMaker类
+### 定义ProtoMaker类
 
 矩阵乘法的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。
 
@@ -100,7 +101,7 @@ The equation is: Out = scale*X
 - `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
+### 定义Operator类
 
 下面的点实现了MulOp的定义：
 
@@ -149,11 +150,11 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
 通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和下面将要介绍的注册函数一起放在`.cc`中
 
-### 3. 定义OpKernel类
+### 定义OpKernel类
 
 `MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数:
 
-- `typename  Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+- `typename DeviceContext`: 表示设备类型，不同设备(CPU、CUDA)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。
 
@@ -165,7 +166,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 下面是 `MulKernel` `Compute`的实现：
 
   ```cpp
-  template <typename Place, typename T>
+  template <typename DeviceContext, typename T>
   class MulKernel : public framework::OpKernel {
   public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -173,33 +174,32 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
     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);
+    auto& device_context = context.template device_context<DeviceContext>();
+    math::matmul<DeviceContext, T>(*X, false, *Y, false, 1, Z, 0, device_context);
   }
   };
   ```
 
-需要注意：**不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
+需要注意：**不同设备(CPU、CUDA)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
 
-`MulOp`的CPU、GPU实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+`MulOp`的CPU、CUDA实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
 
-为了使`OpKernel`的计算过程书写更加简单，并且CPU、GPU的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。
+为了使`OpKernel`的计算过程书写更加简单，并且CPU、CUDA的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。
 
 
 到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。
 反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。
 
-### 4. 注册Operator
+### 注册Operator
 
 - 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
 
     ```cpp
     namespace ops = paddle::operators;
     REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
-    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>);
     REGISTER_OP_CPU_KERNEL(mul_grad,
-                  ops::MulGradKernel<paddle::platform::CPUPlace, float>);
+                  ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>);
     ```
 
    在上面的代码中：
@@ -209,20 +209,20 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
     - `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::MulGradKernel`类。
 
 
-- 在 `.cu`文件中注册GPU Kernel。
-    - 请注意，如果GPU Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下：
+- 在 `.cu`文件中注册CUDA Kernel。
+    - 请注意，如果CUDA Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下：
 
     ```cpp
     // if use Eigen unsupported module before include head files
-    // #define EIGEN_USE_GPU
+    #define EIGEN_USE_GPU
 
     namespace ops = paddle::operators;
-    REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
-    REGISTER_OP_GPU_KERNEL(mul_grad,
-                           ops::MulGradKernel<paddle::platform::GPUPlace, float>);
+    REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel<paddle::platform::CUDADeviceContext, float>);
+    REGISTER_OP_CUDA_KERNEL(mul_grad,
+                           ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
     ```
 
-### 5. 编译
+### 编译
 
 运行下面命令可以进行编译：
 
@@ -236,71 +236,57 @@ make mul_op
 
 ## 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`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
+单测包括对比前向Op不同设备(CPU、CUDA)的实现、对比反向OP不同设备(CPU、CUDA)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
 
-### 前向Operator单元测试
+### 前向Operator单测
 
-前向Op单元测试继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`。各项更加具体的单元测试在`OpTestMeta`里完成。测试前向Operator，需要：
+Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp`里完成。测试Operator，需要：
 
 1. 在`setUp`函数定义输入、输出，以及相关的属性参数。
 2. 生成随机的输入数据。
 3. 在Python脚本中实现与前向operator相同的计算逻辑，得到输出值，与operator前向计算的输出进行对比。
+4. 反向计算已经自动集成进测试框架，直接调用相应接口即可。
 
 
   ```python
   import unittest
   import numpy as np
-  from gradient_checker import GradientChecker, create_op
-  from op_test_util import OpTestMeta
+  from op_test import OpTest
 
-  class TestMulOp(unittest.TestCase):
-      __metaclass__ = OpTestMeta
 
+  class TestMulOp(OpTest):
       def setUp(self):
-          self.type = "mul"
+          self.op_type = "mul"
           self.inputs = {
               'X': np.random.random((32, 84)).astype("float32"),
               'Y': np.random.random((84, 100)).astype("float32")
           }
           self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
-  ```
-
-上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
-
-- `self.type = "mul" ` : 定义类型，与operator注册时注册的类型一致。
-- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
-- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
 
+      def test_check_output(self):
+          self.check_output()
 
-### 反向Operator单元测试
+      def test_check_grad_normal(self):
+          self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
 
-反向Op单元测试继承自`GradientChecker`，而`GradientChecker`继承自`unittest.TestCase`，因此，**反向单元测试函数需要以`test_`开头**。
+      def test_check_grad_ingore_x(self):
+          self.check_grad(
+              ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
 
-```python
-class TestMulGradOp(GradientChecker):
-    def setUp(self):
-        self.op = create_op("mul")
-        self.inputs = {
-            'X': np.random.random((32, 84)).astype("float32"),
-            'Y': np.random.random((84, 100)).astype("float32")
-        }
-
-    def test_check_grad_normal(self):
-        # mul op will enlarge the relative error
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
+      def test_check_grad_ingore_y(self):
+          self.check_grad(
+              ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
+  ```
 
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
+上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
 
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
-```
+- `self.op_type = "mul" ` : 定义类型，与operator注册时注册的类型一致。
+- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
+- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
 
-下面解释代码中一些关键的地方:
+### 反向operator单测
 
-- 调用`create_op("mul")`创建反向Op对应的前向Op。
+而反向测试中：
 - `test_check_grad_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。
   - 第一个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
   - 第二个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`。
@@ -308,7 +294,7 @@ class TestMulGradOp(GradientChecker):
 - `test_check_grad_ingore_x`和`test_check_grad_ingore_y`分支用来测试只需要计算一个输入梯度的情况。
 
 
-### 编译和执行单元测试
+### 编译和执行
 
 `python/paddle/v2/framework/tests` 目录下新增的 `test_*.py` 单元测试会被自动加入工程进行编译。
 
@@ -328,5 +314,5 @@ ctest -R test_mul_op
 
 - 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，这将会导致编译出错。
 - 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。
-- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
+- 如果Op没有实现CUDA Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
 - 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。

doc/howto/dev/new_op_en.md

 # How to write a new operator
 
  - [Background](#background)
- - [Implementing C++ Types](#implementing-c++-types)
-   - [Defining ProtoMaker](#defining-protoMaker)
+ - [Implementing C++ Types](#implementing-c-types)
+   - [Defining ProtoMaker](#defining-protomaker)
    - [Defining Operator](#defining-operator)
    - [Registering Operator](#registering-operator)
    - [Compilation](#compilation)
@@ -28,8 +28,8 @@ An operator can be differentiated by whether in has kernel methods. An operator
 --------------  | :----------------------
 OpProtoMake definition  | `.cc`files, Backward Op does not need an OpProtoMake interface.
 Op definition           | `.cc` files
-Kernel implementation       | The kernel methods shared between CPU and GPU are defined in `.h` files. CPU-specific kernels live in `.cc` files, while GPU-specific kernels are implemented in `.cu`files.
-Registering the Op           | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the GPU implementation.
+Kernel implementation       | The kernel methods shared between CPU and CUDA are defined in `.h` files. CPU-specific kernels live in `.cc` files, while CUDA-specific kernels are implemented in `.cu`files.
+Registering the Op           | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation.
 
 
 New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions. **
@@ -41,7 +41,7 @@ Let's take matrix multiplication operator, [MulOp](https://github.com/PaddlePadd
 ## Implementing C++ Types
 
 
-### 1. Defining Class ProtoMaker
+### Defining ProtoMaker
 
 Matrix Multiplication can be written as $Out = X * Y$, meaning that the operation consists of two inputs and pne output.
 
@@ -98,7 +98,7 @@ There are two changes in this example:
 - `AddAttr<AttrType>("scale", "...").SetDefault(1.0);`  adds `scale`constant as an attribute, and sets the default value to 1.0.
 
 
-### 2. Defining Operator
+### Defining Operator
 
 The following code defines the interface for MulOp:
 
@@ -147,11 +147,11 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
 Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, which also include the registration methods introduced later.
 
-### 3. Defining OpKernel
+### Defining OpKernel
 
 `MulKernel` inherits `framework::OpKernel`, which includes the following templates:
 
-- `typename  Place` denotes device type. When different devices, namely the CPU and the GPU, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
+- `typename  DeviceContext` denotes device context type. When different devices, namely the CPUDeviceContext and the CUDADeviceContext, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
 
 - `typename T` denotes data type, such as `float` or `double`.
 
@@ -163,7 +163,7 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
 `MulKernel`'s implementation of `Compute` is as follows:
 
   ```cpp
-  template <typename Place, typename T>
+  template <typename DeviceContext, typename T>
   class MulKernel : public framework::OpKernel {
   public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -171,16 +171,15 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
     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);
+    auto& device_context = context.template device_context<DeviceContext>();
+    math::matmul<DeviceContext, T>(*X, false, *Y, false, 1, Z, 0, device_context);
   }
   };
   ```
 
-Note that **different devices (CPU, GPU)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.**
+Note that **different devices (CPU, CUDA)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.**
 
-`MulOp`'s CPU and GPU share the same `Kernel`. A non-sharing  `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
+`MulOp`'s CPU and CUDA share the same `Kernel`. A non-sharing  `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
 
 To ease the writing of `OpKernel` compute, and for reusing code cross-device, [`Eigen-unsupported Tensor`](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md?fileviewer=file-view-default) module is used to implement `Compute` interface. To learn about how the Eigen library is used in PaddlePaddle, please see [usage document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md).
 
@@ -189,16 +188,16 @@ This concludes the forward implementation of an operator. Next its operation and
 
 The definition of its corresponding backward operator, if applicable, is similar to that of an forward operator. **Note that a backward operator does not include a `ProtoMaker`**.
 
-### 4. Registering Operator
+### Registering Operator
 
 - In `.cc` files, register forward and backward operator classes and the CPU kernel.
 
     ```cpp
     namespace ops = paddle::operators;
     REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
-    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>);
     REGISTER_OP_CPU_KERNEL(mul_grad,
-                  ops::MulGradKernel<paddle::platform::CPUPlace, float>);
+                  ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>);
     ```
 
    In that code block,
@@ -208,20 +207,20 @@ The definition of its corresponding backward operator, if applicable, is similar
     - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`.
 
 
-- Registering GPU Kernel in `.cu` files
-    - Note that if GPU Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as
+- Registering CUDA Kernel in `.cu` files
+    - Note that if CUDA Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as
 
     ```cpp
     // if use Eigen unsupported module before include head files
     #define EIGEN_USE_GPU
 
     namespace ops = paddle::operators;
-    REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
-    REGISTER_OP_GPU_KERNEL(mul_grad,
-                           ops::MulGradKernel<paddle::platform::GPUPlace, float>);
+    REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel<paddle::platform::CUDADeviceContext, float>);
+    REGISTER_OP_CUDA_KERNEL(mul_grad,
+                           ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
     ```
 
-### 5. Compilation
+### Compilation
 
 Run the following commands to compile.
 
@@ -253,62 +252,51 @@ A forward operator unit test inherits `unittest.TestCase` and defines metaclass
 
 2. Generating random input data.
 
-3. Implementing the same computation logic in a Python script:
+3. Implementing the same computation logic in a Python script.
+
+4. Call check gradient function to check the backward operator.
 
   ```python
   import unittest
   import numpy as np
-  from gradient_checker import GradientChecker, create_op
-  from op_test_util import OpTestMeta
+  from op_test import OpTest
 
-  class TestMulOp(unittest.TestCase):
-      __metaclass__ = OpTestMeta
 
+  class TestMulOp(OpTest):
       def setUp(self):
-          self.type = "mul"
+          self.op_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'])}
+
+      def test_check_output(self):
+          self.check_output()
+          
+      def test_check_grad_normal(self):
+          self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
+
+      def test_check_grad_ingore_x(self):
+          self.check_grad(
+              ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
+
+      def test_check_grad_ingore_y(self):
+          self.check_grad(
+              ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
   ```
 Get its output, and compare it with the forward operator's own output.
 
 The code above first loads required packages. In addition, we have
 
-- `self.type = "mul" ` defines the type that is identical to what the operator's registered type.
+- `self.op_type = "mul" ` defines the type that is identical to what the operator's registered type.
 - `self.inputs` defines input, with type `numpy.array` and initializes it.
 - `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script.
 
 ### Testing Backward Operators
 
-A backward operator unit test inherits `GradientChecker`, which inherits `unittest.TestCase`. As a result, **a backward operator unit test needs to be have the prefix `test_`**.
-
-```python
-class TestMulGradOp(GradientChecker):
-    def setUp(self):
-        self.op = create_op("mul")
-        self.inputs = {
-            'X': np.random.random((32, 84)).astype("float32"),
-            'Y': np.random.random((84, 100)).astype("float32")
-        }
-
-    def test_check_grad_normal(self):
-        # mul op will enlarge the relative error
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
-
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
-
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
-```
-
-Some key points in the code above include:
+Some key points in checking gradient above include:
 
-- `create_op("mul")` creates the backward operator's corresponding forward operator.
 - `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods.
   - The first variable `["X", "Y"]` appoints `X` and `Y` to be scale tested.
   - The second variable `"Out"` points to the network's final output target `Out`.
@@ -338,5 +326,5 @@ ctest -R test_mul_op
 
 - Every `*_op.h` (if applicable), `*_op.cc`, and `*_op.cu` (if applicable) must be created for a unique Op. Compiling will fail if multiple operators are included per file.
 - The type with which an operator is registered needs to be identical to the Op's name. Registering `REGISTER_OP(B, ...)` in `A_op.cc` will cause unit testing failures.
-- If the operator does not implement a GPU kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail.
+- If the operator does not implement a CUDA kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail.
 - If multiple operators rely on some shared methods, a file NOT named `*_op.*` can be created to store them, such as `gather.h`.

doc/howto/index_cn.rst

@@ -9,9 +9,6 @@
 
   usage/cmd_parameter/index_cn.rst
   usage/cluster/cluster_train_cn.md
-  usage/k8s/k8s_basis_cn.md
-  usage/k8s/k8s_cn.md
-  usage/k8s/k8s_distributed_cn.md
 
 开发标准
 --------

doc/howto/index_en.rst

@@ -9,8 +9,6 @@ Usage
 
   usage/cmd_parameter/index_en.rst
   usage/cluster/cluster_train_en.md
-  usage/k8s/k8s_en.md
-  usage/k8s/k8s_aws_en.md
 
 Development
 ------------

doc/howto/read_source.md

+# PaddlePaddle Fluid Source Code Overview
+
+Examples: https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/v2/fluid/tests/book
+
+Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework
+
+Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators
+
+Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory
+
+Platform: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform
+
+# Compile Time
+
+The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto).
+
+```python
+x = fluid.layers.data(name='x', shape=[13], dtype='float32')
+y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+y_predict = fluid.layers.fc(input=x, size=1, act=None)
+cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+avg_cost = fluid.layers.mean(x=cost)
+
+sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
+sgd_optimizer.minimize(avg_cost)
+```
+
+- Variables: `x`,  `y`, `y_predict`, `cost` and `avg_cost`. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/framework.py#L93)
+- Layers: `fluid.layers.data`, `fluid.layers.fc` and `fluid.layers.mean` are layers. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/layers.py)
+  - Every Layer has one or more operators and variables/parameters
+    - All the operators are defined at [`paddle/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators). Other worth-looking files:
+      - Base class: [`paddle/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h)
+      - Operator Registration: [`paddle/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h) 
+      - Operator Lookup: [`paddle/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_info.h)
+- Optimizer: `fluid.optimizer.SGD`. It does the following
+  - Add backward operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/backward.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/backward.cc)]
+  - Add optimizer operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/optimizer.py), [C++](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/optimizer)]
+
+# Run Time
+
+The following **evaluates** the NN. Instantiates all the variables, operators.
+
+```python
+place = fluid.CPUPlace()
+feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
+exe = fluid.Executor(place)
+
+# Allocate memory. Initialize Parameter.
+exe.run(fluid.default_startup_program())
+
+# Allocate memory. Do computation.
+exe.run(fluid.default_main_program(),
+        feed=feeder.feed(data),
+        fetch_list=[avg_cost])
+```
+
+- Place: `place`. one of CPU, GPU or FPGA. [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h)
+  - The device handle are at [paddle/platform/device_context.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h)
+- Executor: `fluid.Executor(place)`. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/executor.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc)]
+  - Feeds the data: `feed=feeder.feed(data)`
+  - Evaluates all the operators
+  - Fetches the result: `fetch_list=[avg_cost]`
+- Other worth looking files:
+  - Scope: [paddle/framework/scope.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/scope.h). Where all the variables live
+    - Variable: [paddle/framework/variable.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h). Where all the data (most likely tensors) live
+      - Tensor: [paddle/framework/tensor.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h). Where we allocate memory through [`paddle/memory/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory)

doc/howto/usage/cluster/cluster_train_cn.md

 # PaddlePaddle分布式训练
 
-* [概述](#概述)
-* [环境准备](#环境准备)
-* [启动参数说明](#启动参数说明)
-  * [启动参数服务器](#启动参数服务器)
-  * [启动计算节点](#启动计算节点)
-  * [准备数据集](#准备数据集)
-  * [准备训练程序](#准备训练程序)
-* [使用分布式计算平台或工具](#使用分布式计算平台或工具)
-  * [使用Fabric启动集群作业](#使用fabric启动集群作业)
-     * [准备一个Linux集群](#准备一个linux集群)
-     * [启动集群作业](#启动集群作业)
-     * [终止集群作业](#终止集群作业)
-     * [检查集群训练结果](#检查集群训练结果)
-     * [检查模型输出](#检查模型输出)
-  * [在OpenMPI集群中提交训练作业](#在openmpi集群中提交训练作业)
-     * [准备OpenMPI集群](#准备OpenMPI集群)
-     * [启动集群作业](#启动集群作业-1)
-  * [在Kubernetes集群中提交训练作业](#在kubernetes集群中提交训练作业)
 
 ## 概述
+
 本文将介绍如何使用PaddlePaddle在不同的集群框架下完成分布式训练。分布式训练架构如下图所示：
 
 <img src="https://user-images.githubusercontent.com/13348433/31772175-5f419eca-b511-11e7-9db7-5231fe3d9ccb.png" width="500">
@@ -32,10 +15,11 @@
 
 在使用同步SGD训练神经网络时，PaddlePaddle使用同步屏障（barrier），使梯度的提交和参数的更新按照顺序方式执行。在异步SGD中，则并不会等待所有trainer提交梯度才更新参数，这样极大地提高了计算的并行性：参数服务器之间不相互依赖，并行地接收梯度和更新参数，参数服务器也不会等待计算节点全部都提交梯度之后才开始下一步，计算节点之间也不会相互依赖，并行地执行模型的训练。可以看出，虽然异步SGD方式会提高参数更新并行度, 但是并不能保证参数同步更新，在任意时间某一台参数服务器上保存的参数可能比另一台要更新，与同步SGD相比，梯度会有噪声。
 
+
 ## 环境准备
 
 1. 准备您的计算集群。计算集群通常由一组（几台到几千台规模）的Linux服务器组成。服务器之间可以通过局域网（LAN）联通，每台服务器具有集群中唯一的IP地址（或者可被DNS解析的主机名）。集群中的每台计算机通常被成为一个“节点”。
-1. 我们需要在集群的所有节点上安装 PaddlePaddle。 如果要启用GPU，还需要在节点上安装对应的GPU驱动以及CUDA。PaddlePaddle的安装可以参考[build_and_install](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/getstarted/build_and_install)的多种安装方式。我们推荐使用[Docker](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)安装方式来快速安装PaddlePaddle。
+1. 我们需要在集群的所有节点上安装 PaddlePaddle。 如果要启用GPU，还需要在节点上安装对应的GPU驱动以及CUDA。PaddlePaddle的安装可以参考[build_and_install](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/index_cn.html)的多种安装方式。我们推荐使用[Docker](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/build_and_install/docker_install_cn.html)安装方式来快速安装PaddlePaddle。
 
 安装完成之后，执行下面的命令可以查看已经安装的版本（docker安装方式可以进入docker容器执行：`docker run -it paddlepaddle/paddle:[tag] /bin/bash`）：
 ```bash
@@ -63,12 +47,12 @@ $ paddle pserver --port=7164 --ports_num=1 --ports_num_for_sparse=1 --num_gradie
 $ stdbuf -oL /usr/bin/nohup paddle pserver --port=7164 --ports_num=1 --ports_num_for_sparse=1 --num_gradient_servers=1 &> pserver.log
 ```
 
-| 参数  | 是否必选 | 默认值 | 说明 |
-| ------------- | ------------- | ------------- | ------------- |
-| port  | 必选 | 7164 | pserver监听的起始端口，根据ports_num决定<br>总端口个数，从起始端口监听多个端口用于通信  |
-| ports_num  | 必选 | 1 | 监听的端口个数  |
-| ports_num_for_sparse  | 必选 | 1 | 用于稀疏类型参数通信的端口个数  |
-| num_gradient_servers  | 必选 | 1 | 当前训练任务pserver总数 |
+参数说明
+
+- port：**必选，默认7164**，pserver监听的起始端口，根据ports_num决定总端口个数，从起始端口监听多个端口用于通信
+- ports_num：**必选，默认1**，监听的端口个数
+- ports_num_for_sparse：**必选，默认1**，用于稀疏类型参数通信的端口个数
+- num_gradient_servers：**必选，默认1**，当前训练任务pserver总数
 
 ### 启动计算节点
 执行以下命令启动使用python编写的trainer程序（文件名为任意文件名，如train.py）
@@ -105,16 +89,16 @@ paddle.init(
         pservers="127.0.0.1")
 ```
 
-| 参数  | 是否必选 | 默认 | 说明 |
-| ------------- | ------------- | ------------- | ------------- |
-| use_gpu  | 可选 | False | 是否启用GPU训练 |
-| trainer_count  | 必选 | 1 | 当前训练任务trainer总个数 |
-| port  | 必选 | 7164 | 连接到pserver的端口  |
-| ports_num  | 必选 | 1 | 连接到pserver的端口个数  |
-| ports_num_for_sparse  | 必选 | 1 | 和pserver之间用于稀疏类型参数通信的端口个数  |
-| num_gradient_servers  | 必选 | 1 | 当前训练任务pserver总数 |
-| trainer_id  | 必选 | 0 | 每个trainer的唯一ID，从0开始的整数 |
-| pservers  | 必选 | 127.0.0.1 | 当前训练任务启动的pserver的IP列表，多个IP使用“,”隔开 |
+参数说明
+
+- use_gpu： **可选，默认False**，是否启用GPU训练
+- trainer_count：**必选，默认1**，当前训练任务trainer总个数
+- port：**必选，默认7164**，连接到pserver的端口
+- ports_num：**必选，默认1**，连接到pserver的端口个数
+- ports_num_for_sparse：**必选，默认1**，和pserver之间用于稀疏类型参数通信的端口个数
+- num_gradient_servers：**必选，默认1**，当前训练任务pserver总数
+- trainer_id：**必选，默认0**，每个trainer的唯一ID，从0开始的整数
+- pservers：**必选，默认127.0.0.1**，当前训练任务启动的pserver的IP列表，多个IP使用“,”隔开
 
 
 ### 准备数据集
@@ -171,7 +155,7 @@ test.txt-00002
 
 - `my_lib.py`：会被`train.py`调用的一些用户定义的库函数，比如PIL库等。
 - `word_dict.pickle`：在`train.py`中会使用到的字典数据文件。
-- `train.py`：训练程序，代码参考[api_train_v2_cluster.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/prepare.py)。***注意：*** 对于本样例代码，在使用不同的分布式计算平台时，您可能需要修改`train.py`开头的部分（如下），以便获得训练数据的位置和获取环境变量配置：
+- `train.py`：训练程序，代码参考[api_train_v2_cluster.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/api_train_v2_cluster.py)。***注意：*** 对于本样例代码，在使用不同的分布式计算平台时，您可能需要修改`train.py`开头的部分（如下），以便获得训练数据的位置和获取环境变量配置：
 
   ```python
   cluster_train_file = "./train_data_dir/train/train.txt"
@@ -195,91 +179,10 @@ PaddlePaddle可以使用多种分布式计算平台构建分布式计算任务
 
 在使用分布式计算平台进行训练时，任务被调度在集群中时，分布式计算平台通常会通过API或者环境变量提供任务运行需要的参数，比如节点的ID、IP和任务节点个数等。
 
-### 使用Fabric启动集群作业
-
-#### 准备一个Linux集群
-可以在`paddle/scripts/cluster_train_v2/fabric/docker_cluster`目录下，执行`kubectl -f ssh_servers.yaml`启动一个测试集群，并使用`kubectl get po -o wide`获得这些节点的IP地址。
-
-#### 启动集群作业
-
-`paddle.py` 提供了自动化脚本来启动不同节点中的所有 PaddlePaddle 集群进程。默认情况下，所有命令行选项可以设置为 `paddle.py` 命令选项并且 `paddle.py` 将透明、自动地将这些选项应用到 PaddlePaddle 底层进程。
-
-`paddle.py` 为方便作业启动提供了两个独特的命令选项。
-
--  `job_dispatch_package`  设为本地 `workspace` 目录，它将被分发到 `conf.py` 中设置的所有节点。它有助于帮助频繁修改和访问工作区文件的用户减少负担，否则频繁的多节点工作空间部署可能会很麻烦。
--  `job_workspace`  设为已部署的工作空间目录，`paddle.py` 将跳过分发阶段直接启动所有节点的集群作业。它可以帮助减少分发延迟。
-
-`cluster_train/run.sh` 提供了命令样例来运行 `doc/howto/usage/cluster/src/word2vec` 集群任务，只需用您定义的目录修改 `job_dispatch_package` 和 `job_workspace`，然后：
-```
-sh run.sh
-```
-
-集群作业将会在几秒后启动。
-
-#### 终止集群作业
-`paddle.py`能获取`Ctrl + C` SIGINT 信号来自动终止它启动的所有进程。只需中断 `paddle.py` 任务来终止集群作业。如果程序崩溃你也可以手动终止。
-
-#### 检查集群训练结果
-详细信息请检查 $workspace/log 里的日志，每一个节点都有相同的日志结构。
-
-`paddle_trainer.INFO`
-提供几乎所有训练的内部输出日志，与本地训练相同。这里检验运行时间模型的收敛。
-
-`paddle_pserver2.INFO`
-提供 pserver 运行日志，有助于诊断分布式错误。
-
-`server.log`
-提供 parameter server 进程的 stderr 和 stdout。训练失败时可以检查错误日志。
-
-`train.log`
-提供训练过程的 stderr 和 stdout。训练失败时可以检查错误日志。
-
-#### 检查模型输出
-运行完成后，模型文件将被写入节点 0 的 `output` 目录中。
-工作空间中的 `nodefile` 表示当前集群作业的节点 ID。
-
-### 在OpenMPI集群中提交训练作业
-
-#### 准备OpenMPI集群
-
-执行下面的命令以启动3个节点的OpenMPI集群和一个"head"节点：
-
-```bash
-paddle/scripts/cluster_train_v2/openmpi/docker_cluster
-kubectl create -f head.yaml
-kubectl create -f mpi-nodes.yaml
-```
-
-然后可以从head节点ssh无密码登录到OpenMPI的每个节点上。
-
-#### 启动集群作业
-
-您可以按照下面的步骤在OpenMPI集群中提交paddle训练任务：
-
-```bash
-# 获得head和node节点的IP地址
-kubectl get po -o wide
-# 将node节点的IP地址保存到machines文件中
-kubectl get po -o wide | grep nodes | awk '{print $6}' > machines
-# 拷贝必要的文件到head节点
-scp -i ssh/id_rsa.mpi.pub machines prepare.py train.py start_mpi_train.sh tutorial@[headIP]:~
-# ssh 登录到head节点
-ssh -i ssh/id_rsa.mpi.pub tutorial@[headIP]
-# --------------- 以下操作均在head节点中执行 ---------------
-# 准备训练数据
-python prepare.py
-# 拷贝训练程序和字典文件到每台MPI节点
-cat machines | xargs -i scp word_dict.pickle train.py start_mpi_train.sh machines {}:/home/tutorial
-# 创建日志目录
-mpirun -hostfile machines -n 3 mkdir /home/tutorial/logs
-# 拷贝训练数据到各自的节点
-scp train.txt-00000 test.txt-00000 [node1IP]:/home/tutorial
-scp train.txt-00001 test.txt-00001 [node2IP]:/home/tutorial
-scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial
-# 启动训练任务
-mpirun -hostfile machines -n 3  /home/tutorial/start_mpi_train.sh
-```
-
-### 在Kubernetes集群中提交训练作业
+## 在不同集群中运行
 
-此部分的使用方法可以参考[here](../k8s/k8s_distributed_cn.md)。
+  - [fabric](fabric_cn.md)
+  - [openmpi](openmpi_cn.md)
+  - [kubernetes](k8s_cn.md)
+  - [kubernetes distributed](k8s_distributed_cn.md)
+  - [kubernetes on AWS](k8s_aws_cn.md)

doc/howto/usage/cluster/cluster_train_en.md

 # PaddlePaddle Distributed Training
 
-* [Introduction](#introduction)
-* [Preparations](#preparations)
-* [Command-line arguments](#command-line-arguments)
-   * [Starting parameter server](#starting-parameter-server)
-   * [Starting trainer](#starting-trainer)
-   * [Prepare Training Dataset](#prepare-training-dataset)
-   * [Prepare Training program](#prepare-training-program)
-* [Use cluster platforms or cluster management tools](#use-cluster-platforms-or-cluster-management-tools)
-   * [Cluster Training Using Fabric](#cluster-training-using-fabric)
-      * [Prepare a Linux cluster](#prepare-a-linux-cluster)
-      * [Launching Cluster Job](#launching-cluster-job)
-      * [Kill Cluster Job](#kill-cluster-job)
-      * [Check Cluster Training Result](#check-cluster-training-result)
-      * [Check Model Output](#check-model-output)
-   * [Cluster Training Using OpenMPI](#cluster-training-using-openmpi)
-      * [Prepare an OpenMPI cluster](#prepare-an-openmpi-cluster)
-      * [Launching Cluster Job](#launching-cluster-job-1)
-   * [Cluster Training Using Kubernetes](#cluster-training-using-kubernetes)
-
 ## Introduction
 
 In this article, we'll explain how to run distributed training jobs with PaddlePaddle on different types of clusters. The diagram below shows the main architecture of a distributed trainning job:
@@ -35,7 +16,7 @@ When training with synchronize SGD, PaddlePaddle uses an internal "synchronize b
 
 ## Preparations
 1. Prepare your computer cluster. It's normally a bunch of Linux servers connected by LAN. Each server will be assigned a unique IP address. The computers in the cluster can be called "nodes".
-2. Install PaddlePaddle on every node. If you are going to take advantage of GPU cards, you'll also need to install proper driver and CUDA libraries. To install PaddlePaddle please read [this build and install](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/getstarted/build_and_install) document. We strongly recommend using [Docker installation](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_en.rst).
+2. Install PaddlePaddle on every node. If you are going to take advantage of GPU cards, you'll also need to install proper driver and CUDA libraries. To install PaddlePaddle please read [this build and install](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html) document. We strongly recommend using [Docker installation](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/docker_install_en.html).
 
 After installation, you can check the version by typing the below command (run a docker container  if using docker: `docker run -it paddlepaddle/paddle:[tag] /bin/bash`):
 
@@ -67,12 +48,12 @@ If you wish to run parameter servers in background, and save a log file, you can
 $ stdbuf -oL /usr/bin/nohup paddle pserver --port=7164 --ports_num=1 --ports_num_for_sparse=1 --num_gradient_servers=1 &> pserver.log
 ```
 
-| param  | required | default | description |
-| ------------- | ------------- | ------------- | ------------- |
-| port  | required | 7164 | port which parameter server will listen on. If ports_num greater than 1, parameter server will listen on multiple ports for more network throughput |
-| ports_num  | required | 1 | total number of ports will listen on  |
-| ports_num_for_sparse  | required | 1 | number of ports which serves sparse parameter update  |
-| num_gradient_servers  | required | 1 | total number of gradient servers |
+Parameter Description
+
+- port: **required, default 7164**, port which parameter server will listen on. If ports_num greater than 1, parameter server will listen on multiple ports for more network throughput.
+- ports_num: **required, default 1**, total number of ports will listen on.
+- ports_num_for_sparse: **required, default 1**, number of ports which serves sparse parameter update.
+- num_gradient_servers: **required, default 1**, total number of gradient servers.
 
 ### Starting trainer
 Type the command below to start the trainer(name the file whatever you want, like "train.py")
@@ -111,16 +92,16 @@ paddle.init(
         pservers="127.0.0.1")
 ```
 
-| param  | required | default | description |
-| ------------- | ------------- | ------------- | ------------- |
-| use_gpu  | optional | False | set to "True" to enable GPU training |
-| trainer_count  | required | 1 | total count of trainers in the training job |
-| port  | required | 7164 | port to connect to parameter server  |
-| ports_num  | required | 1 | number of ports for communication |
-| ports_num_for_sparse  | required | 1 | number of ports for sparse type caculation |
-| num_gradient_servers  | required | 1 | total number of gradient server |
-| trainer_id  | required | 0 | ID for every trainer, start from 0 |
-| pservers  | required | 127.0.0.1 | list of IPs of parameter servers, separated by "," |
+Parameter Description
+
+- use_gpu: **optional, default False**, set to "True" to enable GPU training.
+- trainer_count: **required, default 1**, total count of trainers in the training job.
+- port: **required, default 7164**, port to connect to parameter server.
+- ports_num: **required, default 1**, number of ports for communication.
+- ports_num_for_sparse: **required, default 1**, number of ports for sparse type caculation.
+- num_gradient_servers: **required, default 1**, total number of gradient server.
+- trainer_id: **required, default 0**, ID for every trainer, start from 0.
+- pservers: **required, default 127.0.0.1**, list of IPs of parameter servers, separated by ",".
 
 ### Prepare Training Dataset
 
@@ -178,7 +159,7 @@ Your workspace may looks like:
 
 - `my_lib.py`: user defined libraries, like PIL libs. This is optional.
 - `word_dict.pickle`: dict file for training word embeding.
-- `train.py`: training program. Sample code: [api_train_v2_cluster.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/prepare.py). ***NOTE:*** You may need to modify the head part of `train.py` when using different cluster platform to retrive configuration environment variables:
+- `train.py`: training program. Sample code: [api_train_v2_cluster.py](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/src/word2vec/api_train_v2_cluster.py). ***NOTE:*** You may need to modify the head part of `train.py` when using different cluster platform to retrive configuration environment variables:
 
   ```python
   cluster_train_file = "./train_data_dir/train/train.txt"
@@ -202,92 +183,10 @@ We'll introduce cluster job management on these platforms. The examples can be f
 
 These cluster platforms provide API or environment variables for training processes, when the job is dispatched to different nodes. Like node ID, IP or total number of nodes etc.
 
-### Cluster Training Using Fabric
-
-#### Prepare a Linux cluster
-
-Run `kubectl -f ssh_servers.yaml` under the directory:  `paddle/scripts/cluster_train_v2/fabric/docker_cluster` will launch a demo cluster. Run `kubectl get po -o wide` to get IP addresses of these nodes.
-
-#### Launching Cluster Job
-`paddle.py` provides automatical scripts to start all PaddlePaddle cluster processes in different nodes. By default, all command line options can be set as `paddle.py` command options and `paddle.py` will transparently and automatically set these options to PaddlePaddle lower level processes.
-
-`paddle.py`provides two distinguished command option for easy job launching.
-
-- `job_dispatch_package` set it with local `workspace` directory, it will be dispatched to all nodes which is set in `conf.py`. It could be helpful for frequently manipulating workspace files. otherwise, frequent multi-nodes workspace deployment is very annoying.
-- `job_workspace`  set it with already deployed workspace directory, `paddle.py` will skip dispatch stage to directly launch cluster job with all nodes. It could help to reduce heavy
-dispatch latency.
-
-`cluster_train/run.sh` provides command line sample to run `demo/recommendation` cluster job, just modify `job_dispatch_package` and `job_workspace` with your defined directory, then:
-```
-sh run.sh
-```
-
-The cluster Job will start in several seconds.
-
-#### Kill Cluster Job
-`paddle.py` can capture `Ctrl + C` SIGINT signal to automatically kill all processes launched by it. So just stop `paddle.py` to kill cluster job. You should manually kill the job if the program crashed.
-
-#### Check Cluster Training Result
-Check log in $workspace/log for details, each node owns same log structure.
-
-`paddle_trainer.INFO`
-It provides almost all internal output log for training,  same as local training. Check runtime model convergence here.
-
-`paddle_pserver2.INFO`
-It provides parameter server running log, which could help to diagnose distributed error.
-
-`server.log`
-It provides stderr and stdout of parameter server process. Check error log if training crashes.
-
-`train.log`
-It provides stderr and stdout of trainer process. Check error log if training crashes.
-
-#### Check Model Output
-After one pass finished, model files will be written in `output` directory in node 0.
-`nodefile` in workspace indicates the node id of current cluster job.
-
-### Cluster Training Using OpenMPI
-
-#### Prepare an OpenMPI cluster
-
-Run the following command to start a 3-node MPI cluster and one "head" node.
-
-```bash
-cd paddle/scripts/cluster_train_v2/openmpi/docker_cluster
-kubectl create -f head.yaml
-kubectl create -f mpi-nodes.yaml
-```
-
-Then you can log in to every OpenMPI node using ssh without input any passwords.
-
-#### Launching Cluster Job
-
-Follow the steps to launch a PaddlePaddle training job in OpenMPI cluster:\
-
-```bash
-# find out node IP addresses
-kubectl get po -o wide
-# generate a "machines" file containing node IP addresses
-kubectl get po -o wide | grep nodes | awk '{print $6}' > machines
-# copy necessary files onto "head" node
-scp -i ssh/id_rsa.mpi.pub machines prepare.py train.py start_mpi_train.sh tutorial@[headIP]:~
-# login to head node using ssh
-ssh -i ssh/id_rsa.mpi.pub tutorial@[headIP]
-# --------------- in head node ---------------
-# prepare training data
-python prepare.py
-# copy training data and dict file to MPI nodes
-cat machines | xargs -i scp word_dict.pickle train.py start_mpi_train.sh machines {}:/home/tutorial
-# creat a directory for storing log files
-mpirun -hostfile machines -n 3 mkdir /home/tutorial/logs
-# copy training data to every node
-scp train.txt-00000 test.txt-00000 [node1IP]:/home/tutorial
-scp train.txt-00001 test.txt-00001 [node2IP]:/home/tutorial
-scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial
-# start the job
-mpirun -hostfile machines -n 3  /home/tutorial/start_mpi_train.sh
-```
-
-### Cluster Training Using Kubernetes
+## Use different clusters
 
-The details can be found [here](../k8s/k8s_cn.md)
+  - [fabric](fabric_en.md)
+  - [openmpi](openmpi_en.md)
+  - [kubernetes](k8s_en.md)
+  - kubernetes distributed
+  - [kubernetes on AWS](k8s_aws_en.md)

doc/howto/usage/cluster/fabric_cn.md

+# 使用fabric启动集群训练
+
+## 准备一个Linux集群
+可以在`paddle/scripts/cluster_train_v2/fabric/docker_cluster`目录下，执行`kubectl -f ssh_servers.yaml`启动一个测试集群，并使用`kubectl get po -o wide`获得这些节点的IP地址。
+
+## 启动集群作业
+
+`paddle.py` 提供了自动化脚本来启动不同节点中的所有 PaddlePaddle 集群进程。默认情况下，所有命令行选项可以设置为 `paddle.py` 命令选项并且 `paddle.py` 将透明、自动地将这些选项应用到 PaddlePaddle 底层进程。
+
+`paddle.py` 为方便作业启动提供了两个独特的命令选项。
+
+-  `job_dispatch_package`  设为本地 `workspace` 目录，它将被分发到 `conf.py` 中设置的所有节点。它有助于帮助频繁修改和访问工作区文件的用户减少负担，否则频繁的多节点工作空间部署可能会很麻烦。
+-  `job_workspace`  设为已部署的工作空间目录，`paddle.py` 将跳过分发阶段直接启动所有节点的集群作业。它可以帮助减少分发延迟。
+
+`cluster_train/run.sh` 提供了命令样例来运行 `doc/howto/usage/cluster/src/word2vec` 集群任务，只需用您定义的目录修改 `job_dispatch_package` 和 `job_workspace`，然后：
+```
+sh run.sh
+```
+
+集群作业将会在几秒后启动。
+
+## 终止集群作业
+`paddle.py`能获取`Ctrl + C` SIGINT 信号来自动终止它启动的所有进程。只需中断 `paddle.py` 任务来终止集群作业。如果程序崩溃你也可以手动终止。
+
+## 检查集群训练结果
+详细信息请检查 $workspace/log 里的日志，每一个节点都有相同的日志结构。
+
+`paddle_trainer.INFO`
+提供几乎所有训练的内部输出日志，与本地训练相同。这里检验运行时间模型的收敛。
+
+`paddle_pserver2.INFO`
+提供 pserver 运行日志，有助于诊断分布式错误。
+
+`server.log`
+提供 parameter server 进程的 stderr 和 stdout。训练失败时可以检查错误日志。
+
+`train.log`
+提供训练过程的 stderr 和 stdout。训练失败时可以检查错误日志。
+
+## 检查模型输出
+运行完成后，模型文件将被写入节点 0 的 `output` 目录中。
+工作空间中的 `nodefile` 表示当前集群作业的节点 ID。

doc/howto/usage/cluster/fabric_en.md

+# Cluster Training Using Fabric
+
+## Prepare a Linux cluster
+
+Run `kubectl -f ssh_servers.yaml` under the directory:  `paddle/scripts/cluster_train_v2/fabric/docker_cluster` will launch a demo cluster. Run `kubectl get po -o wide` to get IP addresses of these nodes.
+
+## Launching Cluster Job
+`paddle.py` provides automatical scripts to start all PaddlePaddle cluster processes in different nodes. By default, all command line options can be set as `paddle.py` command options and `paddle.py` will transparently and automatically set these options to PaddlePaddle lower level processes.
+
+`paddle.py`provides two distinguished command option for easy job launching.
+
+- `job_dispatch_package` set it with local `workspace` directory, it will be dispatched to all nodes which is set in `conf.py`. It could be helpful for frequently manipulating workspace files. otherwise, frequent multi-nodes workspace deployment is very annoying.
+- `job_workspace`  set it with already deployed workspace directory, `paddle.py` will skip dispatch stage to directly launch cluster job with all nodes. It could help to reduce heavy
+dispatch latency.
+
+`cluster_train/run.sh` provides command line sample to run `demo/recommendation` cluster job, just modify `job_dispatch_package` and `job_workspace` with your defined directory, then:
+```
+sh run.sh
+```
+
+The cluster Job will start in several seconds.
+
+## Kill Cluster Job
+`paddle.py` can capture `Ctrl + C` SIGINT signal to automatically kill all processes launched by it. So just stop `paddle.py` to kill cluster job. You should manually kill the job if the program crashed.
+
+## Check Cluster Training Result
+Check log in $workspace/log for details, each node owns same log structure.
+
+`paddle_trainer.INFO`
+It provides almost all internal output log for training,  same as local training. Check runtime model convergence here.
+
+`paddle_pserver2.INFO`
+It provides parameter server running log, which could help to diagnose distributed error.
+
+`server.log`
+It provides stderr and stdout of parameter server process. Check error log if training crashes.
+
+`train.log`
+It provides stderr and stdout of trainer process. Check error log if training crashes.
+
+## Check Model Output
+After one pass finished, model files will be written in `output` directory in node 0.
+`nodefile` in workspace indicates the node id of current cluster job.

doc/howto/usage/cluster/k8s_aws_cn.md

+k8s_aws_en.md
\ No newline at end of file

doc/howto/usage/k8s/k8s_distributed_cn.md → doc/howto/usage/cluster/k8s_distributed_cn.md

 # Kubernetes分布式训练
 
-前一篇文章介绍了如何在Kubernetes集群上启动一个单机PaddlePaddle训练作业 (Job)。在这篇文章里，我们介绍如何在Kubernetes集群上进行分布式PaddlePaddle训练作业。关于PaddlePaddle的分布式训练，文章 [Cluster Training](https://github.com/baidu/Paddle/blob/develop/doc/cluster/opensource/cluster_train.md)介绍了一种通过SSH远程分发任务，进行分布式训练的方法，与此不同的是，本文将介绍在Kubernetes容器管理平台上快速构建PaddlePaddle容器集群，进行分布式训练的方案。
+前一篇文章介绍了如何在Kubernetes集群上启动一个单机PaddlePaddle训练作业 (Job)。在这篇文章里，我们介绍如何在Kubernetes集群上进行分布式PaddlePaddle训练作业。关于PaddlePaddle的分布式训练，文章 [Cluster Training](http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cluster/cluster_train_cn.html)介绍了一种通过SSH远程分发任务，进行分布式训练的方法，与此不同的是，本文将介绍在Kubernetes容器管理平台上快速构建PaddlePaddle容器集群，进行分布式训练的方案。
 
 有关Kubernetes相关概念以及如何搭建和配置Kubernetes集群，可以参考[k8s_basis](./k8s_basis_cn.md)。
 
@@ -28,7 +28,7 @@ PaddlePaddle镜像需要提供`paddle pserver`与`paddle train`进程的运行
 - 拷贝训练文件到容器内
 - 生成`paddle pserver`与`paddle train`进程的启动参数，并且启动训练
 
-因为官方镜像 `paddledev/paddle:cpu-latest` 内已经包含PaddlePaddle的执行程序但是还没上述功能，所以我们可以在这个基础上，添加启动脚本，制作新镜像来完成以上的工作。参考镜像的[*Dockerfile*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/k8s/src/k8s_train/Dockerfile)。
+因为官方镜像 `paddledev/paddle:cpu-latest` 内已经包含PaddlePaddle的执行程序但是还没上述功能，所以我们可以在这个基础上，添加启动脚本，制作新镜像来完成以上的工作。参考镜像的[*Dockerfile*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/src/k8s_train/Dockerfile)。
 
 ```bash
 $ cd doc/howto/usage/k8s/src/k8s_train
@@ -149,20 +149,19 @@ spec:
 
 文件中，`metadata`下的`name`表示这个job的名字。`parallelism，completions`字段表示这个job会同时开启3个PaddlePaddle节点，成功训练且退出的pod数目为3时，这个job才算成功结束。然后申明一个存储卷`jobpath`，代表宿主机目录`/home/work/mfs`，在对容器的描述`containers`字段中，将此目录挂载为容器的`/home/jobpath`目录，这样容器的`/home/jobpath`目录就成为了共享存储，放在这个目录里的文件其实是保存到了MFS上。
 
-`env`字段表示容器的环境变量，我们将`paddle`运行的一些参数通过这种方式传递到容器内。
+`env`字段表示容器的环境变量，我们将`paddle`运行的一些参数通过这种方式传递到容器内：
 
-环境变量 | 说明
---- | ---
-JOB_PATH | 共享存储挂在的路径
-JOB_NAME | Job的名字
-TRAIN_CONFIG_DIR | 本次训练文件所在目录，与JOB_PATH,JOB_NAME组合可以找到本次训练需要的文件路径
-CONF_PADDLE_NIC | `paddle pserver`进程需要的`--nics`参数，即网卡名
-CONF_PADDLE_PORT | `paddle paserver`的`--port`参数
-CONF_PADDLE_PORTS_NUM | 稠密更新的端口数量，即`--ports_num`参数
-CONF_PADDLE_PORTS_NUM_SPARSE | 稀疏更新的端口数量，即`--ports_num_for_sparse`参数
-CONF_PADDLE_GRADIENT_NUM | 训练节点数量，即`--num_gradient_servers参数`
 
-这些参数的具体描述，读者可以查看[这里](http://www.paddlepaddle.org/doc/ui/cmd_argument/detail_introduction.html#parameter-server-and-distributed-communication)。
+- JOB_PATH：共享存储挂在的路径
+- JOB_NAME：Job的名字
+- TRAIN_CONFIG_DIR：本次训练文件所在目录，与JOB_PATH,JOB_NAME组合可以找到本次训练需要的文件路径
+- CONF_PADDLE_NIC：`paddle pserver`进程需要的`--nics`参数，即网卡名
+- CONF_PADDLE_PORT：`paddle paserver`的`--port`参数
+- CONF_PADDLE_PORTS_NUM：稠密更新的端口数量，即`--ports_num`参数
+- CONF_PADDLE_PORTS_NUM_SPARSE：稀疏更新的端口数量，即`--ports_num_for_sparse`参数
+- CONF_PADDLE_GRADIENT_NUM：训练节点数量，即`--num_gradient_servers参数`
+
+这些参数的具体描述，读者可以查看[这里](http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cmd_parameter/detail_introduction_cn.html)。
 
 编写完YAML文件后，可以使用Kubernetes的命令行工具创建job。
 

doc/howto/usage/cluster/openmpi_cn.md

+# 在OpenMPI集群中提交训练作业
+
+## 准备OpenMPI集群
+
+执行下面的命令以启动3个节点的OpenMPI集群和一个"head"节点：
+
+```bash
+paddle/scripts/cluster_train_v2/openmpi/docker_cluster
+kubectl create -f head.yaml
+kubectl create -f mpi-nodes.yaml
+```
+
+然后可以从head节点ssh无密码登录到OpenMPI的每个节点上。
+
+## 启动集群作业
+
+您可以按照下面的步骤在OpenMPI集群中提交paddle训练任务：
+
+```bash
+# 获得head和node节点的IP地址
+kubectl get po -o wide
+# 将node节点的IP地址保存到machines文件中
+kubectl get po -o wide | grep nodes | awk '{print $6}' > machines
+# 拷贝必要的文件到head节点
+scp -i ssh/id_rsa.mpi.pub machines prepare.py train.py start_mpi_train.sh tutorial@[headIP]:~
+# ssh 登录到head节点
+ssh -i ssh/id_rsa.mpi.pub tutorial@[headIP]
+# --------------- 以下操作均在head节点中执行 ---------------
+# 准备训练数据
+python prepare.py
+# 拷贝训练程序和字典文件到每台MPI节点
+cat machines | xargs -i scp word_dict.pickle train.py start_mpi_train.sh machines {}:/home/tutorial
+# 创建日志目录
+mpirun -hostfile machines -n 3 mkdir /home/tutorial/logs
+# 拷贝训练数据到各自的节点
+scp train.txt-00000 test.txt-00000 [node1IP]:/home/tutorial
+scp train.txt-00001 test.txt-00001 [node2IP]:/home/tutorial
+scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial
+# 启动训练任务
+mpirun -hostfile machines -n 3  /home/tutorial/start_mpi_train.sh
+```

doc/howto/usage/cluster/openmpi_en.md

+# Cluster Training Using OpenMPI
+
+## Prepare an OpenMPI cluster
+
+Run the following command to start a 3-node MPI cluster and one "head" node.
+
+```bash
+cd paddle/scripts/cluster_train_v2/openmpi/docker_cluster
+kubectl create -f head.yaml
+kubectl create -f mpi-nodes.yaml
+```
+
+Then you can log in to every OpenMPI node using ssh without input any passwords.
+
+## Launching Cluster Job
+
+Follow the steps to launch a PaddlePaddle training job in OpenMPI cluster:\
+
+```bash
+# find out node IP addresses
+kubectl get po -o wide
+# generate a "machines" file containing node IP addresses
+kubectl get po -o wide | grep nodes | awk '{print $6}' > machines
+# copy necessary files onto "head" node
+scp -i ssh/id_rsa.mpi.pub machines prepare.py train.py start_mpi_train.sh tutorial@[headIP]:~
+# login to head node using ssh
+ssh -i ssh/id_rsa.mpi.pub tutorial@[headIP]
+# --------------- in head node ---------------
+# prepare training data
+python prepare.py
+# copy training data and dict file to MPI nodes
+cat machines | xargs -i scp word_dict.pickle train.py start_mpi_train.sh machines {}:/home/tutorial
+# creat a directory for storing log files
+mpirun -hostfile machines -n 3 mkdir /home/tutorial/logs
+# copy training data to every node
+scp train.txt-00000 test.txt-00000 [node1IP]:/home/tutorial
+scp train.txt-00001 test.txt-00001 [node2IP]:/home/tutorial
+scp train.txt-00002 test.txt-00002 [node3IP]:/home/tutorial
+# start the job
+mpirun -hostfile machines -n 3  /home/tutorial/start_mpi_train.sh
+```

doc/howto/usage/k8s/k8s_basis_cn.md

-# Kubernetes 简介
-
-[*Kubernetes*](http://kubernetes.io/)是Google开源的容器集群管理系统，其提供应用部署、维护、扩展机制等功能，利用Kubernetes能方便地管理跨机器运行容器化的应用。Kubernetes可以在物理机或虚拟机上运行，且支持部署到[AWS](http://kubernetes.io/docs/getting-started-guides/aws)，[Azure](http://kubernetes.io/docs/getting-started-guides/azure/)，[GCE](http://kubernetes.io/docs/getting-started-guides/gce)等多种公有云环境。介绍分布式训练之前，需要对[Kubernetes](http://kubernetes.io/)有一个基本的认识，下面先简要介绍一下本文用到的几个Kubernetes概念。
-
-- [*Node*](http://kubernetes.io/docs/admin/node/) 表示一个Kubernetes集群中的一个工作节点，这个节点可以是物理机或者虚拟机，Kubernetes集群就是由node节点与master节点组成的。
-
-- [*Pod*](http://kubernetes.io/docs/user-guide/pods/) 是一组(一个或多个)容器，pod是Kubernetes的最小调度单元，一个pod中的所有容器会被调度到同一个node上。Pod中的容器共享NET，PID，IPC，UTS等Linux namespace。由于容器之间共享NET namespace，所以它们使用同一个IP地址，可以通过*localhost*互相通信。不同pod之间可以通过IP地址访问。
-
-- [*Job*](http://kubernetes.io/docs/user-guide/jobs/) 描述Kubernetes上运行的作业，一次作业称为一个job，通常每个job包括一个或者多个pods，job启动后会创建这些pod并开始执行一个程序，等待这个程序执行成功并返回0则成功退出，如果执行失败，也可以配置不同的重试机制。
-
-- [*Volume*](http://kubernetes.io/docs/user-guide/volumes/) 存储卷，是pod内的容器都可以访问的共享目录，也是容器与node之间共享文件的方式，因为容器内的文件都是暂时存在的，当容器因为各种原因被销毁时，其内部的文件也会随之消失。通过volume，就可以将这些文件持久化存储。Kubernetes支持多种volume，例如hostPath(宿主机目录)，gcePersistentDisk，awsElasticBlockStore等。
-
-- [*Namespaces*](https://kubernetes.io/docs/user-guide/namespaces/) 命名空间，在kubernetes中创建的所有资源对象(例如上文的pod，job)等都属于一个命名空间，在同一个命名空间中，资源对象的名字是唯一的，不同空间的资源名可以重复，命名空间主要为了对象进行逻辑上的分组便于管理。本文只使用了默认命名空间。
-
-- [*PersistentVolume*](https://kubernetes.io/docs/user-guide/persistent-volumes/): 和[*PersistentVolumeClaim*](https://kubernetes.io/docs/user-guide/persistent-volumes/#persistentvolumeclaims)结合，将外部的存储服务在Kubernetes中描述成为统一的资源形式，便于存储资源管理和Pod引用。
-
-## 部署Kubernetes集群
-
-Kubernetes提供了多种集群部署的方案，本文档内不重复介绍。这里给出集中常见的部署方法：
-
-- [*minikube*](https://kubernetes.io/docs/getting-started-guides/minikube/): 快速在本地启动一个单机的kubernetes服务器，便于本地验证和测试。
-- [*kubeadm*](http://kubernetes.io/docs/getting-started-guides/kubeadm/): 在不同操作系统，不同主机(Bare-Metal, AWS, GCE)条件下，快速部署集群。
-- [*AWS EC2*](https://kubernetes.io/docs/getting-started-guides/aws/): 在aws上快速部署集群。
-- [*Bare-Metal*](https://kubernetes.io/docs/getting-started-guides/centos/centos_manual_config/): 在物理机上手动部署。
-
-可以参考[这个表格](https://kubernetes.io/docs/getting-started-guides/#table-of-solutions)选择适合您的场景的合适方案。
-
-## 选择存储方案
-
-容器不会保留在运行时生成的数据，job或者应用程序在容器中运行时生成的数据会在容器销毁时消失。为了完成分布式机器学习训练任务，需要有一个外部的存储服务来保存训练所需数据和训练输出。
-常见的可选存储服务包括：
-
-- [*NFS*](https://github.com/kubernetes/kubernetes/tree/master/examples/volumes/nfs): 可以将磁盘上某个目录共享给网络中其他机器访问。部署和配置比较简单，可以用于小量数据的验证。不提供分布式存储，高可用，冗余等功能。NFS的部署方法可以参考[这里](http://www.tecmint.com/how-to-setup-nfs-server-in-linux/)。
-- [*GlusterFS*](http://gluster.readthedocs.io/en/latest/Quick-Start-Guide/Quickstart/): 网络分布式文件系统，可以在Kubernetes中按照[这个](https://github.com/kubernetes/kubernetes/tree/master/examples/volumes/glusterfs)例子使用。
-- [*Ceph*](http://docs.ceph.com/docs/master/): 分布式文件系统，支持rbd，POSIX API接口(ceph fs)和对象存储API，参考[这里](https://kubernetes.io/docs/user-guide/volumes/#rbd)。
-- [*MooseFS*](https://moosefs.com/documentation.html): 一个分布式的存储系统。需要先挂载到服务器Node上再通过kubernetes hostPath Volume挂载到容器中。
-
-## 配置kubectl
-
-### 安装kubectl
-```
-# OS X
-curl -LO https://storage.googleapis.com/kubernetes-release/release/$(curl -s https://storage.googleapis.com/kubernetes-release/release/stable.txt)/bin/darwin/amd64/kubectl
-
-# Linux
-curl -LO https://storage.googleapis.com/kubernetes-release/release/$(curl -s https://storage.googleapis.com/kubernetes-release/release/stable.txt)/bin/linux/amd64/kubectl
-
-# Windows
-curl -LO https://storage.googleapis.com/kubernetes-release/release/$(curl -s https://storage.googleapis.com/kubernetes-release/release/stable.txt)/bin/windows/amd64/kubectl.exe
-```
-
-### 配置kubectl访问你的kubernetes集群
-
-编辑`~/.kube/config`这个配置文件，修改`Master-IP`的地址。如果使用SSL认证，则需要配置`certificate-authority`和`users`中的用户证书。如果是使用非SSL方式访问（比如通过8080端口），也可以去掉这些证书的配置。
-```
-apiVersion: v1
-clusters:
-- cluster:
-    certificate-authority: /path/to/ca.crt
-    server: https://[Master-IP]:443
-  name: minikube
-contexts:
-- context:
-    cluster: minikube
-    user: minikube
-  name: minikube
-current-context: minikube
-kind: Config
-preferences: {}
-users:
-- name: minikube
-  user:
-    client-certificate: /path/to/apiserver.crt
-    client-key: /Users/wuyi/.minikube/apiserver.key
-```

doc/mobile/cross_compiling_for_ios_cn.md

@@ -18,11 +18,11 @@ PaddlePaddle为交叉编译提供了工具链配置文档[cmake/cross_compiling/
 
 - `CMAKE_SYSTEM_NAME`，CMake编译的目标平台，必须设置为`iOS`。在设置`CMAKE_SYSTEM_NAME=iOS`后，PaddlePaddle的CMake系统会自动编译所有的第三方依赖库，并且强制设置一些PaddlePaddle参数的值（`WITH_C_API=ON`、`WITH_GPU=OFF`、`WITH_AVX=OFF`、`WITH_PYTHON=OFF`、`WITH_RDMA=OFF`）。
 - `WITH_C_API`，是否编译C-API预测库，必须设置为ON。在iOS平台上只支持使用C-API来预测。
-- `WITH_SWIG_PY`，必须设置为ON。在iOS平台上不支持通过swig调用来训练或者预测。
+- `WITH_SWIG_PY`，必须设置为`OFF`。在iOS平台上不支持通过swig调用来训练或者预测。
 
 iOS平台可选配置参数：
 
-- `IOS_PLATFORM`，可设置为`OS/SIMULATOR`，默认值为`OS`。
+- `IOS_PLATFORM`，可设置为`OS`（默认值）或`SIMULATOR`。
   - `OS`，构建目标为`arm`架构的iPhone或者iPad等物理设备。
   - `SIMULATOR`，构建目标为`x86`架构的模拟器平台。
 - `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示，默认编译所有架构：

doc/mobile/cross_compiling_for_ios_en.md

+# PaddlePaddle Compiling Guide for iOS
+
+This tutorial will walk you through cross compiling the PaddlePaddle library for iOS from the source in MacOS.
+
+## Preparation
+
+Apple provides Xcode for cross-compiling and IDE for iOS development. Download from App store or [here](https://developer.apple.com/cn/xcode/). To verify your installation, run command as follows
+
+```bash
+$ xcodebuild -version
+Xcode 9.0
+Build version 9A235
+```
+
+## Cross-compiling configurations
+
+PaddlePaddle provides cross-compiling toolchain configuration documentation [cmake/cross_compiling/ios.cmake](https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/cross_compiling/ios.cmake), which has some default settings for frequently used compilers.
+
+There are some mandatory environment variables need to be set before cross compiling PaddlePaddle for iOS:
+
+- `CMAKE_SYSTEM_NAME`, CMake compiling target platform name, has to be `iOS`. PaddlePaddle CMake will compile all the third party dependencies and enforce some parameters (`WITH_C_API=ON`, `WITH_GPU=OFF`, `WITH_AVX=OFF`, `WITH_PYTHON=OFF`,`WITH_RDMA=OFF`) when this variable is set with value `iOS`.
+
+- `WITH_C_API`, Whether to compile inference C-API library, has to be `ON`, since C-API is the only supported interface for inferencing in iOS.
+- `WITH_SWIG_PY`, has to be `OFF`. It's not supported to inference or train via swig in iOS.
+
+Optional environment variables for iOS are:
+
+- `IOS_PLATFORM`, either `OS` (default) or `SIMULATOR`.
+  - `OS`, build targets ARM-based physical devices like iPhone or iPad.
+  - `SIMULATOR`, build targets x86 architecture simulators.
+- `IOS_ARCH`, target architecture. By default, all architecture types will be compiled. If you need to specify the architecture to compile for, please find valid values for different `IOS_PLATFORM` settings from the table below:
+
+    <table class="docutils">
+    <colgroup>
+      <col width="35%" />
+      <col width="65%" />
+    </colgroup>
+    <thead valign="bottom">
+      <tr class="row-odd">
+      <th class="head">IOS_PLATFORM</th>
+      <th class="head">IOS_ARCH</th>
+    </tr>
+    </thead>
+    <tbody valign="top">
+      <tr class="row-even">
+      <td>OS</td>
+      <td>armv7, armv7s, arm64 </td>
+    </tr>
+    <tr class="row-odd">
+      <td>SIMULATOR</td>
+      <td>i386, x86_64 </td>
+    </tr>
+    </tbody>
+    </table>
+
+- `IOS_DEPLOYMENT_TARGET`, minimum iOS version to deployment, `7.0` by default.
+- `IOS_ENABLE_BITCODE`, whether to enable [Bitcode](https://developer.apple.com/library/content/documentation/IDEs/Conceptual/AppDistributionGuide/AppThinning/AppThinning.html#//apple_ref/doc/uid/TP40012582-CH35-SW3), values can be `ON/OFF`, `ON` by default.
+- `IOS_USE_VECLIB_FOR_BLAS`, whether to use [vecLib](https://developer.apple.com/documentation/accelerate/veclib) framework for BLAS computing. values can be `ON/OFF`, `OFF` by default.
+- `IOS_DEVELOPMENT_ROOT`, the path to `Developer` directory, can be explicitly set with your `/path/to/platform/Developer`. If left blank, PaddlePaddle will automatically pick the Xcode corresponding `platform`'s `Developer` directory based on your `IOS_PLATFORM` value.
+- `IOS_SDK_ROOT`, the path to `SDK` root, can be explicitly set with your  `/path/to/platform/Developer/SDKs/SDK`. if left black, PaddlePaddle will pick the latest SDK in the directory of `IOS_DEVELOPMENT_ROOT`.
+
+other settings：
+
+- `USE_EIGEN_FOR_BLAS`, whether to use Eigen for matrix computing. effective when `IOS_USE_VECLIB_FOR_BLAS=OFF`. Values can be `ON/OFF`, `OFF` by default.
+- `HOST_C/CXX_COMPILER`, host C/C++ compiler. Uses value from environment variable `CC/CXX` by default or `cc/c++` if `CC/CXX` doesn't exist.
+
+some typical cmake configurations:
+
+```bash
+cmake -DCMAKE_SYSTEM_NAME=iOS \
+      -DIOS_PLATFORM=OS \
+      -DIOS_ARCH="armv7;arm64" \
+      -DIOS_ENABLE_BITCODE=ON \
+      -DIOS_USE_VECLIB_FOR_BLAS=ON \
+      -DCMAKE_INSTALL_PREFIX=your/path/to/install \
+      -DWITH_C_API=ON \
+      -DWITH_TESTING=OFF \
+      -DWITH_SWIG_PY=OFF \
+      ..
+```
+
+```bash
+cmake -DCMAKE_SYSTEM_NAME=iOS \
+      -DIOS_PLATFORM=SIMULATOR \
+      -DIOS_ARCH="x86_64" \
+      -DIOS_USE_VECLIB_FOR_BLAS=ON \
+      -DCMAKE_INSTALL_PREFIX=your/path/to/install \
+      -DWITH_C_API=ON \
+      -DWITH_TESTING=OFF \
+      -DWITH_SWIG_PY=OFF \
+      ..
+```
+
+You can set other compiling parameters for your own need. I.E. if you are trying to minimize the library size, set `CMAKE_BUILD_TYPE` with `MinSizeRel`; or if the performance is your concern, set `CMAKE_BUILD_TYPE` with `Release`. You can even manipulate the PaddlePaddle compiling procedure by manually set `CMAKE_C/CXX_FLAGS` values.
+
+**TIPS for a better performance**:
+
+- set `CMAKE_BUILD_TYPE` with `Release`
+- set `IOS_USE_VECLIB_FOR_BLAS` with `ON`
+
+## Compile and install
+
+After CMake, run following commands, PaddlePaddle will download the compile 3rd party dependencies, compile and install PaddlePaddle inference library.
+
+```
+$ make
+$ make install
+```
+
+Please Note: if you compiled PaddlePaddle in the source directory for other platforms, do remove `third_party` and `build` directory within the source with `rm -rf` to ensure that all the 3rd party libraries dependencies and PaddlePaddle is newly compiled with current CMake configuration.
+
+`your/path/to/install` directory will have following directories after `compile` and `install`:
+
+- `include`, contains all the C-API header files.
+- `lib`, contains PaddlePaddle C-API static library.
+- `third_party` contains all the 3rd party libraries.
+
+Please note: if PaddlePaddle library need to support both physical devices and simulators, you will need to compile correspondingly, then merge fat library with `lipo`.
+
+Now you will have PaddlePaddle library compiled and installed, the fat library can be used in deep learning related iOS APPs. Please refer to C-API documentation for usage guides.

doc/mobile/index_en.rst

@@ -5,4 +5,5 @@ MOBILE
   :maxdepth: 1
 
   cross_compiling_for_android_en.md
+  cross_compiling_for_ios_en.md
   cross_compiling_for_raspberry_en.md

paddle/capi/error.cpp

@@ -14,7 +14,7 @@ limitations under the License. */
 
 #include "error.h"
 
-const char* paddle_error_string(paddle_error err) {
+extern "C" const char* paddle_error_string(paddle_error err) {
   switch (err) {
     case kPD_NULLPTR:
       return "nullptr error";

paddle/capi/error.h

@@ -29,9 +29,17 @@ typedef enum {
   kPD_UNDEFINED_ERROR = -1,
 } paddle_error;
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 /**
  * Error string for Paddle API.
  */
 PD_API const char* paddle_error_string(paddle_error err);
 
+#ifdef __cplusplus
+}
+#endif
+
 #endif

paddle/framework/CMakeLists.txt

@@ -58,3 +58,6 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
         proto_desc)
 cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
 cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
+
+cc_library(init SRCS init.cc DEPS gflags executor place stringpiece)
+cc_test(init_test SRCS init_test.cc DEPS init)

paddle/framework/backward.cc

@@ -430,14 +430,14 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
     std::vector<std::unique_ptr<OpDescBind>> op_grads;
 
     if ((*it)->Type() == "recurrent" || (*it)->Type() == "while") {
-      int step_block_idx = (*it)->GetBlockAttr("step_block");
+      int step_block_idx = (*it)->GetBlockAttr("sub_block");
       BlockDescBind* backward_block = CreateStepBlock(
           program_desc, no_grad_vars, grad_to_var, step_block_idx);
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
     } else if ((*it)->Type() == "conditional_block") {
       BlockDescBind* backward_block =
           CreateStepBlock(program_desc, no_grad_vars, grad_to_var,
-                          (*it)->GetBlockAttr("block"));
+                          (*it)->GetBlockAttr("sub_block"));
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var, {backward_block});
     } else {
       op_grads = MakeOpGrad(*it, no_grad_vars, grad_to_var);

paddle/framework/ddim_test.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 <sstream>
 #include <vector>
 

paddle/framework/executor.cc

@@ -33,32 +33,12 @@ namespace framework {
 const std::string kFeedOpType = "feed";
 const std::string kFetchOpType = "fetch";
 
-Executor::Executor(const std::vector<platform::Place>& places) : own_(true) {
-  PADDLE_ENFORCE_GT(places.size(), 0);
-  device_contexts_.resize(places.size());
-  for (size_t i = 0; i < places.size(); i++) {
-    if (platform::is_cpu_place(places[i])) {
-      device_contexts_[i] = new platform::CPUDeviceContext(
-          boost::get<platform::CPUPlace>(places[i]));
-    } else if (platform::is_gpu_place(places[i])) {
-#ifdef PADDLE_WITH_CUDA
-      device_contexts_[i] = new platform::CUDADeviceContext(
-          boost::get<platform::GPUPlace>(places[i]));
-#else
-      PADDLE_THROW(
-          "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
-          "option");
-#endif
-    }
-  }
-}
+DeviceContextPool* DeviceContextPool::pool = nullptr;
 
-Executor::~Executor() {
-  if (own_) {
-    for (auto& device_context : device_contexts_) {
-      delete device_context;
-    }
-  }
+Executor::Executor(const std::vector<platform::Place>& places) {
+  DeviceContextPool& pool = DeviceContextPool::Get();
+  auto borrowed_contexts = pool.Borrow(places);
+  device_contexts_.swap(borrowed_contexts);
 }
 
 static void CreateTensor(Variable* var, VarDesc::VarType var_type) {
@@ -132,8 +112,5 @@ void Executor::Run(const ProgramDescBind& pdesc, Scope* scope, int block_id,
   }
 }
 
-Executor::Executor(const platform::DeviceContext& device)
-    : device_contexts_({&device}), own_(false) {}
-
 }  // namespace framework
 }  // namespace paddle

paddle/framework/executor.h

@@ -14,19 +14,98 @@ limitations under the License. */
 
 #pragma once
 
+#include <map>
+#include <unordered_map>
+
 #include "paddle/framework/op_info.h"
 #include "paddle/framework/program_desc.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
 
 namespace paddle {
 namespace framework {
 
+class DeviceContextPool {
+ public:
+  static DeviceContextPool& Get() {
+    PADDLE_ENFORCE_NOT_NULL(pool, "Need to Create DeviceContextPool first!");
+    return *pool;
+  }
+
+  static DeviceContextPool& Create(const std::vector<platform::Place>& places) {
+    if (pool == nullptr) {
+      pool = new DeviceContextPool(places);
+    }
+    return *pool;
+  }
+
+  std::vector<const platform::DeviceContext*> Borrow(
+      const std::vector<platform::Place>& places) {
+    PADDLE_ENFORCE_GT(places.size(), 0);
+    PADDLE_ENFORCE_LE(places.size(), device_contexts_.size());
+    std::vector<const platform::DeviceContext*> borrowed_contexts;
+    for (auto& place : places) {
+      auto range = device_contexts_.equal_range(place);
+      if (range.first == range.second) {
+        PADDLE_THROW(
+            "'Place' is not supported, Please re-compile with WITH_GPU "
+            "option");
+      }
+      // TODO(dzhwinter) : assign the first found device. Will enhanced later.
+      // device load balancer maybe useful here.
+      borrowed_contexts.emplace_back(range.first->second);
+    }
+    return borrowed_contexts;
+  }
+
+  explicit DeviceContextPool(const std::vector<platform::Place>& places) {
+    PADDLE_ENFORCE_GT(places.size(), 0);
+    for (size_t i = 0; i < places.size(); i++) {
+      if (platform::is_cpu_place(places[i])) {
+        device_contexts_.emplace(
+            places[i], new platform::CPUDeviceContext(
+                           boost::get<platform::CPUPlace>(places[i])));
+      } else if (platform::is_gpu_place(places[i])) {
+#ifdef PADDLE_WITH_CUDA
+        device_contexts_.emplace(
+            places[i], new platform::CUDADeviceContext(
+                           boost::get<platform::GPUPlace>(places[i])));
+#else
+        PADDLE_THROW(
+            "'GPUPlace' is not supported, Please re-compile with WITH_GPU "
+            "option");
+#endif
+      }
+    }
+  }
+
+  ~DeviceContextPool() {}
+
+ private:
+  static DeviceContextPool* pool;
+  struct Hash {
+    std::hash<int> hash_;
+    size_t operator()(const platform::Place& place) const {
+      return hash_(place.which());
+    }
+  };
+  std::unordered_multimap<const platform::Place, const platform::DeviceContext*,
+                          Hash>
+      device_contexts_;
+  DISABLE_COPY_AND_ASSIGN(DeviceContextPool);
+};
+
 class Executor {
  public:
+  // TODO(dzhwinter) : Do not rely on this function, it will be removed
+  explicit Executor(const platform::DeviceContext& device)
+      : Executor(std::vector<platform::Place>({device.GetPlace()})) {}
+
+  explicit Executor(const platform::Place& place)
+      : Executor(std::vector<platform::Place>({place})) {}
+
   explicit Executor(const std::vector<platform::Place>& places);
-  explicit Executor(const platform::DeviceContext& devices);
-  ~Executor();
 
   /* @Brief
    * Runtime evaluation of the given ProgramDesc under certain Scope
@@ -39,7 +118,6 @@ class Executor {
 
  private:
   std::vector<const platform::DeviceContext*> device_contexts_;
-  bool own_;
 };
 
 }  // namespace framework

paddle/framework/init.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 <algorithm>
+#include <string>
+
+#include "paddle/framework/executor.h"
+#include "paddle/framework/init.h"
+#include "paddle/platform/place.h"
+#include "paddle/string/piece.h"
+
+namespace paddle {
+namespace framework {
+
+std::once_flag gflags_init_flag;
+
+// TODO(qijun) move init gflags to init.cc
+void InitGflags(std::vector<std::string> &argv) {
+  std::call_once(gflags_init_flag, [&]() {
+    int argc = argv.size();
+    char **arr = new char *[argv.size()];
+    std::string line;
+    for (size_t i = 0; i < argv.size(); i++) {
+      arr[i] = &argv[i][0];
+      line += argv[i];
+      line += ' ';
+    }
+    google::ParseCommandLineFlags(&argc, &arr, true);
+    VLOG(1) << "Init commandline: " << line;
+  });
+}
+
+bool InitDevices(const std::vector<std::string> &devices) {
+  // device format
+  // CPU
+  // GPU:1
+  // TODO(dzhwinter) : add device format annotation for users.
+  std::vector<platform::Place> places;
+  for (auto &device : devices) {
+    auto p = string::Piece(device);
+    if (string::Find(p, ':', 0) == string::Piece::npos) {
+      places.emplace_back(platform::CPUPlace());
+    } else if (string::HasPrefix(p, "GPU")) {
+#ifdef PADDLE_WITH_CUDA
+      auto pos = string::RFind(p, ':', string::Piece::npos);
+      auto number = device.substr(pos + 1);
+      places.emplace_back(platform::GPUPlace(std::stoi(number)));
+#else
+      LOG(WARNING)
+          << "'GPU' is not supported, Please re-compile with WITH_GPU option";
+#endif
+    } else {
+      return false;
+    }
+  }
+
+  if (std::find_if(places.begin(), places.end(),
+                   [&](const platform::Place &place) {
+                     return platform::is_cpu_place(place);
+                   }) == places.end()) {
+    places.emplace_back(platform::CPUPlace());
+    LOG(WARNING) << "Not specified any device, use CPU by Default.";
+  }
+  DeviceContextPool::Create(places);
+  return true;
+  return true;
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/init.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 <mutex>
+
+#include "gflags/gflags.h"
+#include "glog/logging.h"
+
+namespace paddle {
+namespace framework {
+
+void InitGflags(std::vector<std::string> &argv);
+
+bool InitDevices(const std::vector<std::string> &devices);
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/init_test.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 "gtest/gtest.h"
+
+#include "paddle/framework/init.h"
+
+TEST(Init, InitDevices) {
+  using paddle::framework::InitDevices;
+  std::vector<std::string> ds1 = {"CPU"};
+  ASSERT_EQ(InitDevices(ds1), true);
+
+#ifdef PADDLE_WITH_CUDA
+  std::vector<std::string> ds2 = {"CPU", "GPU:0", "GPU:1"};
+  ASSERT_EQ(InitDevices(ds2), true);
+#endif
+}

paddle/function/GemmConvOp.cpp

@@ -126,6 +126,11 @@ public:
       inputData += inputChannels * inputHeight * inputWidth;
       outputData += outputChannels * outputHeight * outputWidth;
     }
+#ifdef PADDLE_MOBILE_INFERENCE
+    if (Device == DEVICE_TYPE_CPU) {
+      memory_.reset();
+    }
+#endif
   }
 };
 

paddle/gserver/layers/ROIPoolLayer.cpp

@@ -84,12 +84,15 @@ void ROIPoolLayer::forward(PassType passType) {
   size_t poolChannelOffset = pooledHeight_ * pooledWidth_;
 
   real* outputData = outputValue->getData();
-  Matrix::resizeOrCreate(maxIdxs_,
-                         numROIs,
-                         channels_ * pooledHeight_ * pooledWidth_,
-                         false,
-                         false);
-  real* argmaxData = maxIdxs_->getData();
+  real* argmaxData = nullptr;
+  if (passType != PASS_TEST) {
+    Matrix::resizeOrCreate(maxIdxs_,
+                           numROIs,
+                           channels_ * pooledHeight_ * pooledWidth_,
+                           false,
+                           false);
+    argmaxData = maxIdxs_->getData();
+  }
 
   for (size_t n = 0; n < numROIs; ++n) {
     // the first five elememts of each RoI should be:
@@ -128,14 +131,18 @@ void ROIPoolLayer::forward(PassType passType) {
           bool isEmpty = (hend <= hstart) || (wend <= wstart);
           size_t poolIndex = ph * pooledWidth_ + pw;
           outputData[poolIndex] = isEmpty ? 0 : -FLT_MAX;
-          argmaxData[poolIndex] = -1;
+          if (argmaxData) {
+            argmaxData[poolIndex] = -1;
+          }
 
           for (size_t h = hstart; h < hend; ++h) {
             for (size_t w = wstart; w < wend; ++w) {
               size_t index = h * width_ + w;
               if (batchData[index] > outputData[poolIndex]) {
                 outputData[poolIndex] = batchData[index];
-                argmaxData[poolIndex] = index;
+                if (argmaxData) {
+                  argmaxData[poolIndex] = index;
+                }
               }
             }
           }
@@ -143,7 +150,9 @@ void ROIPoolLayer::forward(PassType passType) {
       }
       batchData += channelOffset;
       outputData += poolChannelOffset;
-      argmaxData += poolChannelOffset;
+      if (argmaxData) {
+        argmaxData += poolChannelOffset;
+      }
     }
     bottomROIs += roiOffset;
   }

paddle/gserver/layers/SequenceToBatch.cpp

@@ -171,12 +171,31 @@ void SequenceToBatch::sequence2BatchCopy(Matrix &batch,
     hl_sequence2batch_copy(
         batchData, seqData, idxData, seqWidth, batchCount, seq2batch);
   } else {
-    for (int i = 0; i < batchCount; ++i) {
-      if (seq2batch) {
+    if (seq2batch) {
+#ifdef PADDLE_USE_MKLML
+      const int blockMemSize = 8 * 1024;
+      const int blockSize = blockMemSize / sizeof(real);
+#pragma omp parallel for collapse(2)
+      for (int i = 0; i < batchCount; ++i) {
+        for (int j = 0; j < seqWidth; j += blockSize) {
+          memcpy(batch.rowBuf(i) + j,
+                 sequence.rowBuf(idxData[i]) + j,
+                 (j + blockSize > seqWidth) ? (seqWidth - j) * sizeof(real)
+                                            : blockMemSize);
+        }
+      }
+#else
+      for (int i = 0; i < batchCount; ++i) {
         memcpy(batch.rowBuf(i),
                sequence.rowBuf(idxData[i]),
                seqWidth * sizeof(real));
-      } else {
+      }
+#endif
+    } else {
+#ifdef PADDLE_USE_MKLML
+#pragma omp parallel for
+#endif
+      for (int i = 0; i < batchCount; ++i) {
         memcpy(sequence.rowBuf(idxData[i]),
                batch.rowBuf(i),
                seqWidth * sizeof(real));

paddle/math/Matrix.cpp

@@ -28,6 +28,7 @@ limitations under the License. */
 #include "hl_top_k.h"
 #include "paddle/utils/Logging.h"
 
+#include "NEONFunctions.h"
 #include "paddle/function/GemmFunctor.h"
 #include "paddle/utils/ThreadLocal.h"
 
@@ -4165,16 +4166,36 @@ void CpuMatrix::print(std::ostream& os) const {
 void CpuMatrix::paramReluForward(Matrix& data, Matrix& W) {
   real* input = data.getData();
   real* w = W.getData();
+  real* output = data_;
   size_t numElements = data.getWidth();
   size_t numSamples = data.getHeight();
   size_t paraSize = W.getHeight() * W.getWidth();
   CHECK(!(numElements % paraSize));  // this check from ParameterReluLayer::init
+
   size_t partial_sum = numElements / paraSize;
+  if (paraSize == numElements) {
+    for (size_t n = 0; n < numSamples * numElements; ++n) {
+      output[n] = input[n] > 0 ? input[n] : input[n] * w[n % numElements];
+    }
+    return;
+  }
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+  for (size_t n = 0; n < numSamples; ++n) {
+    for (size_t i = 0; i < paraSize; i++) {
+      neon::prelu(
+          input + i * partial_sum, w[i], output + i * partial_sum, partial_sum);
+    }
+    input = input + numElements;
+    output = output + numElements;
+  }
+#else
   for (size_t n = 0, k = 0; n < numSamples; ++n) {
     for (size_t i = 0; i < numElements; ++i, ++k) {
-      data_[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
+      output[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
     }
   }
+#endif
 }
 
 void CpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) {

paddle/math/NEONFunctions.cpp

@@ -49,6 +49,46 @@ void relu(const float* a, float* b, int len) {
   }
 }
 
+// b[i] = a[i] > 0.0f ? a[i] : a[i] * w
+void prelu(const float* a, float w, float* b, int len) {
+  int offset = len % 16;
+  float32x4_t ma0, ma1, ma2, ma3;
+
+  float32x4_t zero = vdupq_n_f32(0.f);
+  float32x4_t vw = vdupq_n_f32(w);
+
+  for (int k = 0; k < len / 16; k++, a += 16, b += 16) {
+    ma0 = vld1q_f32(a);
+    ma1 = vld1q_f32(a + 4);
+    ma2 = vld1q_f32(a + 8);
+    ma3 = vld1q_f32(a + 12);
+
+    uint32x4_t flag0 = vcgtq_f32(ma0, zero);
+    uint32x4_t flag1 = vcgtq_f32(ma1, zero);
+    uint32x4_t flag2 = vcgtq_f32(ma2, zero);
+    uint32x4_t flag3 = vcgtq_f32(ma3, zero);
+
+    float32x4_t mul0 = vmulq_f32(ma0, vw);
+    float32x4_t mul1 = vmulq_f32(ma1, vw);
+    float32x4_t mul2 = vmulq_f32(ma2, vw);
+    float32x4_t mul3 = vmulq_f32(ma3, vw);
+
+    ma0 = vbslq_f32(flag0, ma0, mul0);
+    ma1 = vbslq_f32(flag1, ma1, mul1);
+    ma2 = vbslq_f32(flag2, ma2, mul2);
+    ma3 = vbslq_f32(flag3, ma3, mul3);
+
+    vst1q_f32(b, ma0);
+    vst1q_f32(b + 4, ma1);
+    vst1q_f32(b + 8, ma2);
+    vst1q_f32(b + 12, ma3);
+  }
+
+  for (int i = 0; i < offset; i++) {
+    b[i] = a[i] > 0.0f ? a[i] : a[i] * w;
+  }
+}
+
 }  // namespace neon
 }  // namespace paddle
 

paddle/math/NEONFunctions.h

@@ -18,6 +18,7 @@ namespace paddle {
 namespace neon {
 
 void relu(const float* a, float* b, int len);
+void prelu(const float* a, float w, float* b, int len);
 
 }  // namespace neon
 }  // namespace paddle

paddle/math/float16.h

@@ -79,7 +79,7 @@ public:
 #ifdef PADDLE_CUDA_FP16
   HOSTDEVICE inline explicit float16(const half& h) {
 #if CUDA_VERSION >= 9000
-    x = reinterpret_cast<__half_raw*>(&h)->x;
+    x = reinterpret_cast<__half_raw*>(const_cast<half*>(&h))->x;
 #else
     x = h.x;
 #endif  // CUDA_VERSION >= 9000
@@ -145,7 +145,7 @@ public:
 #ifdef PADDLE_CUDA_FP16
   HOSTDEVICE inline float16& operator=(const half& rhs) {
 #if CUDA_VERSION >= 9000
-    x = reinterpret_cast<__half_raw*>(&rhs)->x;
+    x = reinterpret_cast<__half_raw*>(const_cast<half*>(&rhs))->x;
 #else
     x = rhs.x;
 #endif

paddle/math/tests/test_matrixCompare.cpp

@@ -244,7 +244,7 @@ TEST(Matrix, unary) {
     LOG(WARNING) << "This version of PaddlePaddle was not built with LAPACK"
                  << "support so we cannot test matrix inverse. To test "
                  << "matrix inverse, please install LAPACKE "
-                 << "and MKL/Openblas/ATLAS, and re-build PaddlePaddle.";
+                 << "and MKL/Openblas, and re-build PaddlePaddle.";
 #endif
   }
 }

paddle/memory/detail/system_allocator.cc

@@ -19,6 +19,7 @@ limitations under the License. */
 
 #include <stdlib.h>    // for malloc and free
 #include <sys/mman.h>  // for mlock and munlock
+#include <algorithm>   // for std::max
 
 #include "gflags/gflags.h"
 
@@ -28,7 +29,7 @@ limitations under the License. */
 // of memory available to the system for paging.  So, by default, we
 // should set false to use_pinned_memory.
 DEFINE_bool(use_pinned_memory, true, "If set, allocate cpu pinned memory.");
-
+DECLARE_double(fraction_of_gpu_memory_to_use);
 namespace paddle {
 namespace memory {
 namespace detail {
@@ -77,45 +78,20 @@ void* GPUAllocator::Alloc(size_t& index, size_t size) {
   // CUDA documentation doesn't explain if cudaMalloc returns nullptr
   // if size is 0.  We just make sure it does.
   if (size <= 0) return nullptr;
-
-  size_t available = 0;
-  size_t capacity = 0;
-  paddle::platform::GpuMemoryUsage(available, capacity);
-
-  // Reserve memory for page tables, etc.
-  size_t reserving = 0.05 * capacity + paddle::platform::GpuMinChunkSize();
-  size_t usable = available > reserving ? available - reserving : 0;
-
-  // If remaining size no less than expected size, using general
-  // cudaMalloc to allocate GPU memory.
-  void* p = 0;
-  if (size <= usable) {
-    cudaError_t result = cudaMalloc(&p, size);
-    if (result == cudaSuccess) {
-      index = 0;
-      gpu_alloc_size_ += size;
-      return p;
-    }
-  }
-
-  // If remaining size less than expected size or cudaMalloc failed,
-  // cudaMallocHost will be considered as a fallback allocator.
-  //
-  // NOTE: here, we use GpuMaxAllocSize() as the maximum memory size
-  // of host fallback allocation. Allocates too much would reduce
-  // the amount of memory available to the underlying system for paging.
-  usable = paddle::platform::GpuMaxAllocSize() - fallback_alloc_size_;
-
-  if (size > usable) return nullptr;
-
-  cudaError_t result = cudaMallocHost(&p, size);
+  void* p;
+  cudaError_t result = cudaMalloc(&p, size);
   if (result == cudaSuccess) {
-    index = 1;
-    fallback_alloc_size_ += size;
+    index = 0;
+    gpu_alloc_size_ += size;
     return p;
+  } else {
+    LOG(WARNING)
+        << "Cannot malloc " << size / 1024.0 / 1024.0
+        << " MB GPU memory. Please shrink FLAGS_fraction_of_gpu_memory_to_use "
+           "environment variable to a lower value. Current value is "
+        << FLAGS_fraction_of_gpu_memory_to_use;
+    return nullptr;
   }
-
-  return nullptr;
 }
 
 void GPUAllocator::Free(void* p, size_t size, size_t index) {

paddle/operators/cast_op.cc

@@ -74,4 +74,5 @@ REGISTER_OP_WITH_KERNEL(cast, ops::CastOpGradMaker, ops::CastOpInferShape,
 REGISTER_OP_CPU_KERNEL(cast, ops::CastOpKernel<CPU, float>,
                        ops::CastOpKernel<CPU, double>,
                        ops::CastOpKernel<CPU, int>,
-                       ops::CastOpKernel<CPU, int64_t>);
+                       ops::CastOpKernel<CPU, int64_t>,
+                       ops::CastOpKernel<CPU, bool>);

paddle/operators/cast_op.cu

@@ -19,4 +19,5 @@ using CastOpKernel =
     paddle::operators::CastOpKernel<paddle::platform::CUDADeviceContext, T>;
 
 REGISTER_OP_CUDA_KERNEL(cast, CastOpKernel<float>, CastOpKernel<double>,
-                        CastOpKernel<int>, CastOpKernel<int64_t>);
+                        CastOpKernel<int>, CastOpKernel<int64_t>,
+                        CastOpKernel<bool>);

paddle/operators/chunk_eval_op.cc

@@ -32,6 +32,13 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
                    "Output(Recall) of ChunkEvalOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("F1-Score"),
                    "Output(F1-Score) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("NumInferChunks"),
+                   "Output(NumInferChunks) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("NumLabelChunks"),
+                   "Output(NumLabelChunks) of ChunkEvalOp should not be null.");
+    PADDLE_ENFORCE(
+        ctx->HasOutput("NumCorrectChunks"),
+        "Output(NumCorrectChunks) of ChunkEvalOp should not be null.");
 
     auto inference_dim = ctx->GetInputDim("Inference");
     auto label_dim = ctx->GetInputDim("Label");
@@ -42,6 +49,9 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("Precision", {1});
     ctx->SetOutputDim("Recall", {1});
     ctx->SetOutputDim("F1-Score", {1});
+    ctx->SetOutputDim("NumInferChunks", {1});
+    ctx->SetOutputDim("NumLabelChunks", {1});
+    ctx->SetOutputDim("NumCorrectChunks", {1});
   }
 
  protected:
@@ -70,6 +80,16 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
               "sensitivity) of chunks on the given mini-batch.");
     AddOutput("F1-Score",
               "(float). The evaluated F1-Score on the given mini-batch.");
+    AddOutput("NumInferChunks",
+              "(int64_t). The number of chunks in Inference on the given "
+              "mini-batch.");
+    AddOutput(
+        "NumLabelChunks",
+        "(int64_t). The number of chunks in Label on the given mini-batch.");
+    AddOutput(
+        "NumCorrectChunks",
+        "(int64_t). The number of chunks both in Inference and Label on the "
+        "given mini-batch.");
     AddAttr<int>("num_chunk_types",
                  "(int). The number of chunk type. See below for details.");
     AddAttr<std::string>(

paddle/operators/chunk_eval_op.h

@@ -111,9 +111,7 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
     std::vector<Segment> label_segments;
     std::vector<Segment> output_segments;
     std::set<int> excluded_chunk_types;
-    int64_t num_output_segments = 0;
-    int64_t num_label_segments = 0;
-    int64_t num_correct = 0;
+
     if (context.Attr<std::string>("chunk_scheme") == "IOB") {
       num_tag_types = 2;
       tag_begin = 0;
@@ -151,12 +149,24 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
     auto* precision = context.Output<Tensor>("Precision");
     auto* recall = context.Output<Tensor>("Recall");
     auto* f1 = context.Output<Tensor>("F1-Score");
+    auto* num_infer_chunks = context.Output<Tensor>("NumInferChunks");
+    auto* num_label_chunks = context.Output<Tensor>("NumLabelChunks");
+    auto* num_correct_chunks = context.Output<Tensor>("NumCorrectChunks");
 
     const int64_t* inference_data = inference->data<int64_t>();
     const int64_t* label_data = label->data<int64_t>();
     T* precision_data = precision->mutable_data<T>(context.GetPlace());
     T* racall_data = recall->mutable_data<T>(context.GetPlace());
     T* f1_data = f1->mutable_data<T>(context.GetPlace());
+    int64_t* num_infer_chunks_data =
+        num_infer_chunks->mutable_data<int64_t>(context.GetPlace());
+    int64_t* num_label_chunks_data =
+        num_label_chunks->mutable_data<int64_t>(context.GetPlace());
+    int64_t* num_correct_chunks_data =
+        num_correct_chunks->mutable_data<int64_t>(context.GetPlace());
+    *num_infer_chunks_data = 0;
+    *num_label_chunks_data = 0;
+    *num_correct_chunks_data = 0;
 
     auto lod = label->lod();
     PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
@@ -166,17 +176,23 @@ class ChunkEvalKernel : public framework::OpKernel<T> {
     for (int i = 0; i < num_sequences; ++i) {
       int seq_length = lod[0][i + 1] - lod[0][i];
       EvalOneSeq(inference_data + lod[0][i], label_data + lod[0][i], seq_length,
-                 output_segments, label_segments, num_output_segments,
-                 num_label_segments, num_correct, num_chunk_types,
-                 num_tag_types, other_chunk_type, tag_begin, tag_inside,
-                 tag_end, tag_single, excluded_chunk_types);
+                 output_segments, label_segments, *num_infer_chunks_data,
+                 *num_label_chunks_data, *num_correct_chunks_data,
+                 num_chunk_types, num_tag_types, other_chunk_type, tag_begin,
+                 tag_inside, tag_end, tag_single, excluded_chunk_types);
     }
-    *precision_data = !num_output_segments ? 0 : static_cast<T>(num_correct) /
-                                                     num_output_segments;
-    *racall_data = !num_label_segments ? 0 : static_cast<T>(num_correct) /
-                                                 num_label_segments;
-    *f1_data = !num_correct ? 0 : 2 * (*precision_data) * (*racall_data) /
-                                      ((*precision_data) + (*racall_data));
+    *precision_data = !(*num_infer_chunks_data)
+                          ? 0
+                          : static_cast<T>(*num_correct_chunks_data) /
+                                (*num_infer_chunks_data);
+    *racall_data = !(*num_label_chunks_data)
+                       ? 0
+                       : static_cast<T>(*num_correct_chunks_data) /
+                             (*num_label_chunks_data);
+    *f1_data = !(*num_correct_chunks_data)
+                   ? 0
+                   : 2 * (*precision_data) * (*racall_data) /
+                         ((*precision_data) + (*racall_data));
   }
 
   void EvalOneSeq(const int64_t* output, const int64_t* label, int length,

paddle/operators/conditional_block_op.cc

@@ -65,7 +65,7 @@ class ConditionalBlockOp : public ConditionalOp {
       scopes->front() = &scope.NewScope();
       auto &cur_scope = *scopes->front();
 
-      auto *block = Attr<framework::BlockDescBind *>("block");
+      auto *block = Attr<framework::BlockDescBind *>("sub_block");
       framework::Executor exec(dev_ctx);
       exec.Run(*block->Program(), &cur_scope, block->ID(), false);
     }
@@ -88,7 +88,7 @@ class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker {
               "unify the conditional block, rnn and while op, the type of "
               "scope is std::vector<Scope*>");
     AddAttr<framework::BlockDescBind *>(
-        "block", "The step block of conditional block operator");
+        "sub_block", "The step block of conditional block operator");
     AddComment(R"DOC(Conditional block operator
 
 Run the sub-block if X is not empty. Params is the other inputs and Out is the
@@ -117,7 +117,7 @@ class ConditionalBlockGradOp : public ConditionalOp {
       auto &scopes = scope_var->Get<std::vector<framework::Scope *>>();
       framework::Scope &cur_scope = *scopes[0];
 
-      auto *block = Attr<framework::BlockDescBind *>("block");
+      auto *block = Attr<framework::BlockDescBind *>("sub_block");
       framework::Executor exec(dev_ctx);
       exec.Run(*block->Program(), &cur_scope, block->ID(), false);
 
@@ -181,7 +181,7 @@ class ConditionalBlockGradMaker : public framework::SingleGradOpDescMaker {
     grad_op->SetInput("Scope", Output("Scope"));
     grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
     grad_op->SetOutput(framework::GradVarName("Params"), InputGrad("Params"));
-    grad_op->SetBlockAttr("block", *this->grad_block_[0]);
+    grad_op->SetBlockAttr("sub_block", *this->grad_block_[0]);
     return std::unique_ptr<framework::OpDescBind>(grad_op);
   }
 };

paddle/operators/conv_op.h

@@ -261,8 +261,12 @@ class GemmConvGradKernel : public framework::OpKernel<T> {
 
     if (input_grad) {
       input_grad->mutable_data<T>(context.GetPlace());
-      set_zero(dev_ctx, input_grad, static_cast<T>(0));
 
+      // if is_expand is false, the operation of set_zero is unnecessary,
+      // because math::matmul will reset input_grad.
+      if (is_expand) {
+        set_zero(dev_ctx, input_grad, static_cast<T>(0));
+      }
       math::Col2VolFunctor<DeviceContext, T> col2vol;
       math::Col2ImFunctor<math::ColFormat::kCFO, DeviceContext, T> col2im;
 

paddle/operators/conv_transpose_op.h

@@ -225,7 +225,6 @@ class GemmConvTransposeGradKernel : public framework::OpKernel<T> {
 
       if (input_grad) {
         input_grad->mutable_data<T>(context.GetPlace());
-        set_zero(dev_ctx, input_grad, static_cast<T>(0));
       }
       if (filter_grad) {  // filter size (m, c, k_h, k_w)
         filter_grad->mutable_data<T>(context.GetPlace());

paddle/operators/crop_op.cc

@@ -88,7 +88,8 @@ There are two ways to set shape:
 
 The input should be a k-D tensor(k > 0 and k < 7). As an example:
 
-Given:
+Case 1:
+Given
 
     X = [[0, 1, 2, 0, 0]
          [0, 3, 4, 0, 0]
@@ -107,6 +108,27 @@ we get:
     Out = [[1, 2],
            [3, 4]].
 
+
+Case 2:
+Given
+
+    X = [[0, 1, 2, 5, 0]
+         [0, 3, 4, 6, 0]
+         [0, 0, 0, 0, 0]],
+
+and
+
+    offsets = [0, 1],
+
+and
+
+    Y = [[0, 0, 0]
+         [0, 0, 0]],
+
+we get:
+
+    Out = [[1, 2, 5],
+           [3, 4, 6]].
 )DOC");
   }
 };

paddle/operators/fill_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/framework/data_type.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/detail/safe_ref.h"
+
+namespace paddle {
+namespace operators {
+
+struct FillOpVisitor {
+  FillOpVisitor(framework::LoDTensor *tensor, const std::vector<float> &value)
+      : tensor_(tensor), value_(value) {}
+
+  template <typename T>
+  void operator()() const {
+    platform::CPUPlace cpu;
+    auto *data = tensor_->mutable_data<T>(cpu);
+    std::transform(value_.data(), value_.data() + tensor_->numel(), data,
+                   [](float dat) { return static_cast<T>(dat); });
+  }
+
+  framework::LoDTensor *tensor_;
+  const std::vector<float> &value_;
+};
+
+class FillOp : public framework::OperatorBase {
+ public:
+  FillOp(const std::string &type, const framework::VariableNameMap &inputs,
+         const framework::VariableNameMap &outputs,
+         const framework::AttributeMap &attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void Run(const framework::Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {
+    auto &out =
+        detail::Ref(detail::Ref(scope.FindVar(Output("Out")),
+                                "Cannot find variable %s", Output("Out"))
+                        .GetMutable<framework::LoDTensor>());
+    out.Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+    auto dtype = static_cast<framework::DataType>(Attr<int>("dtype"));
+    platform::CPUPlace cpu;
+    auto force_cpu = Attr<bool>("force_cpu");
+    out.mutable_data(force_cpu ? cpu : dev_ctx.GetPlace(),
+                     framework::ToTypeIndex(dtype));
+
+    framework::LoDTensor tensor;
+
+    if (force_cpu || platform::is_cpu_place(dev_ctx.GetPlace())) {
+      tensor.ShareDataWith(out);
+    } else {
+      // Always make tensor in CPU memory.
+      tensor.Resize(out.dims());
+      tensor.mutable_data(cpu, framework::ToTypeIndex(dtype));
+    }
+
+    framework::VisitDataType(
+        dtype, FillOpVisitor(&tensor, Attr<std::vector<float>>("value")));
+
+    if (!force_cpu && platform::is_gpu_place(dev_ctx.GetPlace())) {
+      // Copy tensor to out
+      framework::CopyFrom(tensor, dev_ctx.GetPlace(), dev_ctx, &out);
+    }
+  }
+};
+
+class FillOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  FillOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddComment(R"DOC(Fill operator
+
+Fill an tensor with `value` and `shape`. The type of the tensor is specify by
+`dtype`.
+)DOC");
+    AddOutput("Out", "(LoDTensor) The output tensor.");
+    AddAttr<std::vector<float>>(
+        "value", "The float values of tensor, which are flatten in row major");
+    AddAttr<std::vector<int>>("shape", "The shape of output tensor");
+    AddAttr<int>("dtype", "The data type of output tensor, Default is float")
+        .SetDefault(framework::DataType::FP32);
+    AddAttr<bool>("force_cpu",
+                  "Whether the output tensor must be at CPU memory or not. "
+                  "Default is false.")
+        .SetDefault(false);
+  }
+};
+
+class FillOpInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext *context) const override {
+    context->SetOutputDim(
+        "Out",
+        framework::make_ddim(context->Attrs().Get<std::vector<int>>("shape")));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(fill, ops::FillOp, ops::FillOpInferShape, ops::FillOpMaker);

paddle/operators/math/math_function.cc

@@ -277,6 +277,14 @@ void set_constant_with_place<platform::CPUPlace>(
                            TensorSetConstantCPU(tensor, value));
 }
 
+template <>
+void set_constant_with_place<platform::MKLDNNPlace>(
+    const platform::DeviceContext& context, framework::Tensor* tensor,
+    float value) {
+  framework::VisitDataType(framework::ToDataType(tensor->type()),
+                           TensorSetConstantCPU(tensor, value));
+}
+
 struct TensorSetConstantWithPlace : public boost::static_visitor<void> {
   TensorSetConstantWithPlace(const platform::DeviceContext& context,
                              framework::Tensor* tensor, float value)

paddle/operators/math/math_function.cu

@@ -273,6 +273,13 @@ void set_constant_with_place<platform::GPUPlace>(
                            TensorSetConstantGPU(context, tensor, value));
 }
 
+template <>
+void set_constant_with_place<platform::CUDNNPlace>(
+    const platform::DeviceContext& context, framework::Tensor* tensor,
+    float value) {
+  set_constant_with_place<platform::GPUPlace>(context, tensor, value);
+}
+
 template struct RowwiseAdd<platform::CUDADeviceContext, float>;
 template struct RowwiseAdd<platform::CUDADeviceContext, double>;
 template struct ColwiseSum<platform::CUDADeviceContext, float>;