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

.travis.yml

@@ -4,7 +4,6 @@ cache:
     - $HOME/.ccache
     - $HOME/.cache/pip
     - $TRAVIS_BUILD_DIR/build/third_party
-    - $TRAVIS_BUILD_DIR/build_android/third_party
 sudo: required
 dist: trusty
 os:
@@ -12,7 +11,6 @@ os:
 env:
   - JOB=build_doc
   - JOB=check_style
-  - JOB=build_android
 addons:
   apt:
     packages:
@@ -23,7 +21,6 @@ addons:
       - python
       - python-pip
       - python2.7-dev
-      - python-numpy
       - python-wheel
       - libboost-dev
       - curl
@@ -37,8 +34,8 @@ before_install:
   - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
   # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
   # protobuf version.
-  - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
-  - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
+  - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
+  - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
   - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
   - go get -u github.com/alecthomas/gometalinter

CMakeLists.txt

@@ -65,8 +65,8 @@ if(NOT CMAKE_BUILD_TYPE)
 endif()
 
 if(ANDROID)
-    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
-        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 21")
+    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
+        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
     endif()
 
     set(WITH_GPU OFF CACHE STRING

Dockerfile.android

@@ -4,9 +4,15 @@ MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
 ARG UBUNTU_MIRROR
 RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
 
+# ENV variables
+ARG ANDROID_ABI
+
+ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
+
 ENV HOME=/root \
     ANDROID_NDK_HOME=/opt/android-ndk-linux \
-    ANDROID_STANDALONE_TOOLCHAIN=/opt/android-toolchain-gcc
+    ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \
+    ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain
 
 RUN apt-get update && \
     apt-get install -y \
@@ -15,12 +21,11 @@ RUN apt-get update && \
     apt-get clean -y
 
 # Install Go and glide
-RUN wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \
-    tar -C /usr/local -xzf go.tgz && \
+RUN wget -qO- go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
+    tar -xz -C /usr/local && \
     mkdir /root/gopath && \
     mkdir /root/gopath/bin && \
-    mkdir /root/gopath/src && \
-    rm go.tgz
+    mkdir /root/gopath/src
 ENV GOROOT=/usr/local/go GOPATH=/root/gopath
 # should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT.
 ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin
@@ -42,7 +47,8 @@ RUN mkdir /opt/android-ndk-tmp && \
     wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
     unzip -q android-ndk-r14b-linux-x86_64.zip && \
     mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-21 --install-dir=${ANDROID_STANDALONE_TOOLCHAIN} && \
+    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \
+    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \
     rm -rf /opt/android-ndk-tmp && \
     rm -rf ${ANDROID_NDK_HOME}
 

cmake/cross_compiling/android.cmake

@@ -20,6 +20,7 @@
 # The supported variables are listed belows:
 # 
 # ANDROID_STANDALONE_TOOLCHAIN
+# ANDROID_TOOLCHAIN
 # ANDROID_ABI
 # ANDROID_NATIVE_API_LEVEL
 # ANDROID_ARM_MODE
@@ -57,6 +58,10 @@ IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL)
     ENDIF()
 ENDIF()
 
+IF(NOT DEFINED ANDROID_TOOLCHAIN)
+    SET(ANDROID_TOOLCHAIN clang)
+ENDIF()
+
 IF(NOT DEFINED ANDROID_ABI)
     SET(ANDROID_ABI "armeabi-v7a")
 ENDIF()
@@ -82,6 +87,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
             "${CMAKE_VERSION}), when cross-compiling for Android.")
 
     IF(ANDROID_STANDALONE_TOOLCHAIN)
+        # Use standalone toolchain
         SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot")
 
         IF(NOT CMAKE_SYSTEM_VERSION)
@@ -96,26 +102,44 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
         ENDIF()
 
         # Toolchain
-        SET(ANDROID_TOOLCHAIN "gcc")
         SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN})
-        IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
-            SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
-            IF(ANDROID_ABI STREQUAL "armeabi")
-                SET(CMAKE_SYSTEM_PROCESSOR armv5te)
-            ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
-                SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
-            ENDIF()
-        ENDIF()
-        IF(ANDROID_ABI STREQUAL "arm64-v8a")
-            SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
-            SET(CMAKE_SYSTEM_PROCESSOR aarch64)
+    ELSE(ANDROID_NDK)
+        # TODO: use android ndk
+    ENDIF()
+
+    IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
+        SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
+        IF(ANDROID_ABI STREQUAL "armeabi")
+            SET(CMAKE_SYSTEM_PROCESSOR armv5te)
+            SET(ANDROID_CLANG_TRIPLE armv5te-none-linux-androideabi)
+        ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
+            SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
+            SET(ANDROID_CLANG_TRIPLE armv7-none-linux-androideabi)
         ENDIF()
-        SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
+    ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
+        SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
+        SET(CMAKE_SYSTEM_PROCESSOR aarch64)
+        SET(ANDROID_CLANG_TRIPLE aarch64-none-linux-android)
+    ELSE()
+        MESSAGE(FATAL_ERROR "Invalid Android ABI: ${ANDROID_ABI}.")
+    ENDIF()
+    SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
+
+    IF(ANDROID_TOOLCHAIN STREQUAL clang)
+        SET(ANDROID_C_COMPILER_NAME clang)
+        SET(ANDROID_CXX_COMPILER_NAME clang++)
+        SET(CMAKE_C_COMPILER_TARGET   ${ANDROID_CLANG_TRIPLE})
+        SET(CMAKE_CXX_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE})
+    ELSEIF(ANDROID_TOOLCHAIN STREQUAL gcc)
+        SET(ANDROID_C_COMPILER_NAME gcc)
+        SET(ANDROID_CXX_COMPILER_NAME g++)
+    ELSE()
+        MESSAGE(FATAL_ERROR "Invalid Android toolchain: ${ANDROID_TOOLCHAIN}")
     ENDIF()
 
     # C compiler
     IF(NOT CMAKE_C_COMPILER)
-        SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}gcc")
+        SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_C_COMPILER_NAME}")
     ELSE()
         GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
     ENDIF()
@@ -125,7 +149,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
 
     # CXX compiler
     IF(NOT CMAKE_CXX_COMPILER)
-        SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}g++")
+        SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_CXX_COMPILER_NAME}")
     ELSE()
         GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
     ENDIF()
@@ -137,7 +161,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
     SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
 
     # Toolchain and ABI specific flags.
-    SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections -finline-limit=64")
+    SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections")
     SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections")
 
     IF(ANDROID_ABI STREQUAL "armeabi")
@@ -145,8 +169,7 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
              -march=armv5te
              -mtune=xscale
              -msoft-float)
-    ENDIF()
-    IF(ANDROID_ABI STREQUAL "armeabi-v7a")
+    ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
         LIST(APPEND ANDROID_COMPILER_FLAGS
              -march=armv7-a
              -mfloat-abi=softfp)
@@ -156,6 +179,8 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
             LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16)
         ENDIF()
         LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8)
+    ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
+        LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
     ENDIF()
 
     IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
@@ -164,10 +189,18 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
         ELSE()
             LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb)
         ENDIF()
+        IF(ANDROID_TOOLCHAIN STREQUAL clang)
+            # Disable integrated-as for better compatibility.
+            LIST(APPEND ANDROID_COMPILER_FLAGS -fno-integrated-as)
+        ENDIF()
     ENDIF()
 
-    IF(ANDROID_ABI STREQUAL "arm64-v8a")
-        LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
+    IF(ANDROID_TOOLCHAIN STREQUAL clang)
+        # CMake automatically forwards all compiler flags to the linker,
+        # and clang doesn't like having -Wa flags being used for linking.
+        # To prevent CMake from doing this would require meddling with
+        # the CMAKE_<LANG>_COMPILE_OBJECT rules, which would get quite messy.
+        LIST(APPEND ANDROID_LINKER_FLAGS -Qunused-arguments)
     ENDIF()
 
     STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}")

cmake/external/openblas.cmake

@@ -12,6 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+IF(USE_EIGEN_FOR_BLAS)
+    return()
+ENDIF(USE_EIGEN_FOR_BLAS)
+
 INCLUDE(cblas)
 
 IF(NOT ${CBLAS_FOUND})

doc/design/functions_operators_layers.md

@@ -86,12 +86,13 @@ def layer.fc(X):
 
 We'd like to have Python bindings to operators in package `paddle.operator`, and Python compositions of operators in package `paddle.layer`.  So we have the following concepts in above illustrative example:
 
-```
+
 | C++ functions/functors | mul          | add          |             |          |
+|------------------------|--------------|--------------|-------------|----------|
 | C++ operator class     | mulOp        | addOp        | FCOp        |          |
 | Python binding         | operator.mul | operator.add | operator.fc |          |
 | Python function        |              |              |             | layer.fc |
-```
+
 
 This is how we differentiate layer and operators in PaddlePaddle:
 

doc/design/graph.md

-# Design Doc: Computations as Graphs
+# Design Doc: Computations as a Graph
 
 A primary goal of the refactorization of PaddlePaddle is a more flexible representation of deep learning computation, in particular, a graph of operators and variables, instead of sequences of layers as before.
 
@@ -8,6 +8,8 @@ This document explains that the construction of a graph as three steps:
 - construct the backward part
 - construct the optimization part
 
+## The Construction of a Graph
+
 Let us take the problem of image classification as a simple example.  The application program that trains the model looks like:
 
 ```python
@@ -25,7 +27,9 @@ The first four lines of above program build the forward part of the graph.
 
 ![](images/graph_construction_example_forward_only.png)
 
-In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x.  `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b.
+In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x.  `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b, and the initialization operators.
+
+Initialization operators are kind of "run-once" operators -- the `Run` method increments a class data member counter so to run at most once.  By doing so, a parameter wouldn't be initialized repeatedly, say, in every minibatch.
 
 In this example, all operators are created as `OpDesc` protobuf messages, and all variables are `VarDesc`.  These protobuf messages are saved in a `BlockDesc` protobuf message.
 
@@ -49,3 +53,18 @@ According to the chain rule of gradient computation, `ConstructBackwardGraph` wo
 For each parameter, like W and b created by `layer.fc`, marked as double circles in above graphs, `ConstructOptimizationGraph` creates an optimization operator to apply its gradient.  Here results in the complete graph:
 
 ![](images/graph_construction_example_all.png)
+
+## Block and Graph
+
+The word block and graph are interchangable in the desgin of PaddlePaddle.  A [Block[(https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions.  A graph of operators and variables is a representation of the block.
+
+A Block keeps operators in an array `BlockDesc::ops`
+
+```protobuf
+message BlockDesc {
+  repeated OpDesc ops = 1;
+  repeated VarDesc vars = 2;
+}
+```
+
+in the order that there appear in user programs, like the Python program at the beginning of this article.  We can imagine that in `ops`,  we have some forward operators, followed by some gradient operators, and then some optimization operators.

doc/design/images/graph_construction_example.dot

@@ -2,6 +2,8 @@ digraph ImageClassificationGraph {
         ///////// The forward part /////////
         FeedX [label="Feed", color=blue, shape=box];
         FeedY [label="Feed", color=blue, shape=box];
+        InitW [label="Init", color=blue, shape=diamond];
+        Initb [label="Init", color=blue, shape=diamond];
         FC [label="FC", color=blue, shape=box];
         MSE [label="MSE", color=blue, shape=box];
 
@@ -14,6 +16,8 @@ digraph ImageClassificationGraph {
 
         FeedX -> x -> FC -> y -> MSE -> cost [color=blue];
         FeedY -> l [color=blue];
+        InitW -> W [color=blue];
+        Initb -> b [color=blue];
         W -> FC [color=blue];
         b -> FC [color=blue];
         l -> MSE [color=blue];

doc/design/simple_op_design.md

@@ -147,7 +147,7 @@ class CosineOp {
 struct CosineOpProtoMaker : public OpProtoMaker {
 	CosineOpProtoMaker(OpProto* proto) : OpProtoMaker(proto) {
 		AddInput("input", "input of cosine op");
-		AddAttr("scale", "scale of cosine op", float).Default(1.0).LargerThan(0.0);
+		AddAttr("scale", "scale of cosine op", float).Default(1.0).GreaterThan(0.0);
 		AddType("cos");
 		AddComment("This is cos op");
 	}

doc/design/var_desc.md

+## Background
+PaddlePaddle divides the description of neural network computation graph into two stages: compile time and runtime.
+
+PaddlePaddle use proto message to describe compile time graph for
+
+1. Computation graph should be able to be saved to a file.
+1. In distributed training, the graph will be serialized and send to multiple workers.
+
+The computation graph is constructed by Data Node and Operation Node. The concept to represent them is in the table below.
+
+| |compile time|runtime|
+|---|---|---|
+|Data|VarDesc(proto)|Variable(cpp)|
+|Operation|OpDesc(proto)|Operator(cpp)|
+
+
+## Definition of VarDesc
+
+A VarDesc should have a name and value, in PaddlePaddle, the value will always be a tensor. Since we use LoDTensor most of the time. We add a LoDTesnorDesc to represent it.
+
+```proto
+message VarDesc {
+  required string name = 1;
+  optional LoDTensorDesc lod_tensor = 2;
+}
+```
+
+## Definition of LodTensorDesc
+
+```proto
+enum DataType {
+  BOOL = 0;
+  INT16 = 1;
+  INT32 = 2;
+  INT64 = 3;
+  FP16 = 4;
+  FP32 = 5;
+  FP64 = 6;
+}
+
+message LoDTensorDesc {
+  required DataType data_type = 1;
+  repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
+  optional int32 lod_level = 3 [default=0];
+}
+```
+
+## Definition of Variable in Python
+
+In Python API, layer will take Variable as Input, and return Variable as Output. There should be a class `Variable` in python to help create and manage Variable.
+
+```python
+image = Variable(dims=[-1, 640, 480])
+# fc1 and fc2 are both Variable
+fc1 = layer.fc(input=image, output_size=10)
+fc2 = layer.fc(input=fc1, output_size=20)
+```
+### what should class `Variable` Have
+1. `name`.a name of string type is used to mark the value of the Variable.
+1. `initializer`. Since our Tensor does not have value. we will always use some Operator to fullfill it when run. So we should have a initialize method to help add the init operator.
+1. `operator`. Variable should record which operator produce itself. The reaon is:
+  - we use pd.eval(targets=[var1, var2]) to run the related ops to get the value of var1 and var2. var.op is used to trace the dependency of the current variable.
+
+In PaddlePaddle, we use Block to describe Computation Graph, so in the code we will use Block but not Graph.
+
+```python
+import VarDesc
+import LoDTensorDesc
+import framework
+
+def AddInitialOperator(variable, initializer):
+	# add an initialize Operator to block to init this Variable
+
+class Variable(object):
+   def __init__(self, name, dims, type, initializer):
+      self._block = get_default_block()
+      self._name = name
+      self.op = None
+
+      tensor_desc = LoDTensorDesc(data_type=type, dims=dims)
+      _var_desc = VarDesc(name=name, lod_tensor=tensor_desc)
+      self._var = framework.CreateVar(_var_desc)
+      self._block.add_var(self)
+
+      # add initial op according to initializer
+      if initializer is not None:
+          AddInitialOperator(self, initializer)
+
+   def dims(self):
+      return self._var.dims()
+
+   def data_type(self):
+       return self._var.data_type()
+
+   def to_proto(self):
+       pass
+```
+
+Then we can use this Variable to create a fc layer in Python.
+
+```python
+import paddle as pd
+
+def flatten_size(X, num_flatten_dims):
+  prod = 1 # of last num_flatten_dims
+  for i in xrange(num_flatten_dims):
+    prod = prod * X.dims[-i-1]
+  return prod
+
+def layer.fc(X, output_size, num_flatten_dims):
+  W = Variable(pd.random_uniform(), type=FP32, dims=[flatten_size(X, num_flatten_dims), output_size])
+  b = Variable(pd.random_uniform(), type=FP32, dims=[output_size])
+  out = Variable(type=FP32)
+  y = operator.fc(X, W, b, output=out) # fc will put fc op input into out
+  pd.InferShape(y)
+  return out
+
+x = Variable(dims=[-1, 640, 480])
+y = layer.fc(x, output_size=100)
+z = layer.fc(y, output_size=200)
+
+paddle.eval(targets=[z], ...)
+print(z)
+```

doc/howto/dev/new_op_cn.md

@@ -45,7 +45,9 @@ Kernel实现       | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU
 
 ### 1. 定义ProtoMaker类
 
-矩阵乘的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。首先定义`ProtoMaker`来描述该Op的输入、输出及注释：
+矩阵乘法的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。
+
+首先定义`ProtoMaker`来描述该Op的输入、输出，并添加注释：
 
 ```cpp
 class MulOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -63,17 +65,17 @@ The equation is: Out = X * Y
 };
 ```
 
-[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数包括2个参数：
+[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`，构造函数含有2个参数：
 
    - `framework::OpProto` ： 前者存储Op的输入输出和参数属性，将用于Python API接口的生成。
    - `framework::OpAttrChecker` ：后者用于检查参数属性的合法性。
 
-构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加该Op的注释，这些函数会将对应内容添加到`OpProto`中。
+构造函数里通过`AddInput`添加输入参数，通过`AddOutput`添加输出参数，通过`AddComment`添加Op的注释。这些函数会将对应内容添加到`OpProto`中。
 
-在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，命名请遵守命名规范。
+上面的代码在`MulOp`中添加两个输入`X`和`Y`，添加了一个输出`Out`，并解释了各自含义，命名请遵守命名规范。
 
 
-再举个[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)的例子：
+再以[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)为例：
 
 ```cpp
 template <typename AttrType>
@@ -91,14 +93,16 @@ The equation is: Out = scale*X
 };
 ```
 
- 这个例子有两处不同：
+这个例子有两处不同：
+
+- `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中，如果Op的某个输入不参与反向梯度的计算，请显示地调用`.NotInGradient()`进行设置。
 
-  - `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中。
-  - `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
+- `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
 
 
 ### 2. 定义Operator类
 
+下面的点实现了MulOp的定义：
 
 ```cpp
 class MulOp : public framework::OperatorWithKernel {
@@ -143,14 +147,27 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
   - 1). 做检查， 尽早报错：检查输入数据维度、类型等是否合法。
   - 2). 设置输出Tensor的形状。
 
-通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和要讲到的注册函数一起放在`.cc`中
+通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和下面将要介绍的注册函数一起放在`.cc`中
 
 ### 3. 定义OpKernel类
 
-```cpp
-template <typename Place, typename T>
-class MulKernel : public framework::OpKernel {
- public:
+`MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数:
+
+- `typename  Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+
+- `typename T` : 表示数据类型，如`float`, `double`等。
+
+需要为`MulKernel`类重写`Compute`接口。
+- `Compute`接受一个输入参数：`const framework::ExecutionContext& context`。
+- 与`InferShapeContext`相比，`ExecutionContext`增加了设备类型，同样可获取到输入输出和属性参数。
+- `Compute`函数里实现`OpKernel`的具体计算逻辑。
+
+下面是 `MulKernel` `Compute`的实现：
+
+  ```cpp
+  template <typename Place, typename T>
+  class MulKernel : public framework::OpKernel {
+  public:
   void Compute(const framework::ExecutionContext& context) const override {
     auto* X = context.Input<Tensor>("X");
     auto* Y = context.Input<Tensor>("Y");
@@ -160,50 +177,50 @@ class MulKernel : public framework::OpKernel {
         const_cast<platform::DeviceContext*>(context.device_context_);
     math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
   }
-};
-```
+  };
+  ```
 
-`MulKernel`继承自`framework::OpKernel`，带有模板参数:
+需要注意：**不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
 
-  - `typename  Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
+`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)。
 
- - `typename T` : 表示数据类型，如`float`, `double`等。
+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)。
 
-`MulKernel`需要重写`Compute`接口，该接口参数为`const framework::ExecutionContext& context`, `ExecutionContext`相比`InferShapeContext`增加了设备类型，同样可获取到输入输出和属性参数，`Compute`函数里写具体实现时。
 
-注意，不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel`，`OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 
+到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。
+反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。
 
-为了使得`OpKernel`的计算过程书写较为简单，CPU、GPU的代码可以复用，我们通常借助Eigen unsupported Tensor模块来实现。关于在paddle中如何使用Eigen库，请参考对应的使用[文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)
+### 4. 注册Operator
 
-到此前向Op实现完成，需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似，这里不再重复。但注意，反向Op没有`ProtoMaker`。
+- 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
 
-### 4. 注册Operator
+    ```cpp
+    namespace ops = paddle::operators;
+    REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
+    REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
+    REGISTER_OP_CPU_KERNEL(mul_grad,
+                  ops::MulGradKernel<paddle::platform::CPUPlace, float>);
+    ```
 
-在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
+   在上面的代码中：
 
-```cpp
-namespace ops = paddle::operators;
-REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
-REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(mul_grad,
-              ops::MulGradKernel<paddle::platform::CPUPlace, float>);
-```
+    - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`。
+    - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
+    - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
 
-  - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`，
-  - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
-  - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
 
-在 `.cu`文件中注册GPU Kernel。请注意，如果GPU Kernel的实现是基于Eigen unsupported模块，那么在 `.cu`的最前面请加上宏定义 `#define EIGEN_USE_GPU`
+- 在 `.cu`文件中注册GPU Kernel。
+    - 请注意，如果GPU Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下：
 
-```cpp
-// if use Eigen unsupported module before include head files
-#define EIGEN_USE_GPU
+    ```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>);
-```
+    namespace ops = paddle::operators;
+    REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
+    REGISTER_OP_GPU_KERNEL(mul_grad,
+                           ops::MulGradKernel<paddle::platform::GPUPlace, float>);
+    ```
 
 ### 5. 编译
 
@@ -225,7 +242,7 @@ REGISTER_OP_GPU_KERNEL(mul_grad,
 - 绑定Python
 
     在 [`paddle/pybind/pybind.cc
-`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc)文件中添加该类：
+`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) 使用`USE_OP`告知编译器需要链接的Op，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
 
     ```
     USE_OP(mul);
@@ -242,50 +259,54 @@ REGISTER_OP_GPU_KERNEL(mul_grad,
     USE_NO_KENREL_OP(recurrent);
     ```
 
-    使用`USE_OP`告知编译器需要链接该Op的目标文件，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
-
 
  - 生成库
 
-   无需修改 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件，`paddle/operators` 目录下新增的 `*_op.cc` 文件会自动被添加链接到生成的lib库中。
+   无需修改 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件，`paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。
 
 ## 实现单元测试
 
-单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单测](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
+单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
 
 ### 前向Operator单元测试
 
-前向Op单测继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`，具体单测流程在`OpTestMeta`里完成。需在`setUp`函数定义输入输出和属性参数，以及Python对比的输出值。
+前向Op单元测试继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`。各项更加具体的单元测试在`OpTestMeta`里完成。测试前向Operator，需要：
 
-```python
-import unittest
-import numpy as np
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
+1. 在`setUp`函数定义输入、输出，以及相关的属性参数。
+2. 生成随机的输入数据。
+3. 在Python脚本中实现与前向operator相同的计算逻辑，得到输出值，与operator前向计算的输出进行对比。
 
-class TestMulOp(unittest.TestCase):
-    __metaclass__ = OpTestMeta
 
-    def setUp(self):
-        self.type = "mul"
-        self.inputs = {
-            'X': np.random.random((32, 84)).astype("float32"),
-            'Y': np.random.random((84, 100)).astype("float32")
-        }
-        self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
-```
-   首先需要`import`必要的包,下面详细解释其他值：
+  ```python
+  import unittest
+  import numpy as np
+  from gradient_checker import GradientChecker, create_op
+  from op_test_util import OpTestMeta
+
+  class TestMulOp(unittest.TestCase):
+      __metaclass__ = OpTestMeta
 
-   - `self.type = "mul" ` : 定义类型，和注册的类型一致。
-   - `self.inputs` : 定义输入，类型为Numpy.array，并初始化。
-   - `self.outputs` : 定义输出，并得到Python结算结果。
+      def setUp(self):
+          self.type = "mul"
+          self.inputs = {
+              'X': np.random.random((32, 84)).astype("float32"),
+              'Y': np.random.random((84, 100)).astype("float32")
+          }
+          self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
+  ```
+
+上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
+
+- `self.type = "mul" ` : 定义类型，与operator注册时注册的类型一致。
+- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
+- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
 
 
 ### 反向Operator单元测试
 
-反向Op单测继承自`GradientChecker`，而`GradientChecker`集成自`unittest.TestCase`，所以反向单测函数需要`test_`开头。
+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**。
 
-```cpp
+```python
 class TestMulGradOp(GradientChecker):
     def setUp(self):
         self.op = create_op("mul")
@@ -319,27 +340,27 @@ class TestMulGradOp(GradientChecker):
             no_grad_set={"Y"})
 ```
 
-下面解释一些关键的地方:
+下面解释代码中一些关键的地方:
 
-   - 调用`create_op("mul")`创建反向Op对应的前向Op。
-   - 调用`compare_grad`函数对比CPU、GPU计算结果。
-   - `test_normal`中调用`check_grad`检查梯度稳定性，这里采用数值法检测梯度正确性。
-      - 第一个参数`self.op` : 前向Op。
-      - 第二个参数`self.inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。
-      - 第三个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
-      - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
-   - `test_ignore_x`和`test_ignore_y`分支测试只需要计算一个输入梯度的情况。
+- 调用`create_op("mul")`创建反向Op对应的前向Op。
+- 调用`compare_grad`函数对比CPU、GPU计算结果。
+- `test_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。
+  - 第一个参数`self.op` : 前向Op。
+  - 第二个参数`self.inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。
+  - 第三个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
+  - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
+- `test_ignore_x`和`test_ignore_y`分支用来测试只需要计算一个输入梯度的情况。
 
 
 ### 编译和执行单元测试
 
-单测完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)里添加以下内容将单测加入工程中：
+单元测试编写完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)中添加以下内容，将单元测试加入工程：
 
 ```
 py_test(test_mul_op SRCS test_mul_op.py)
 ```
 
-请注意，**不同于Op的编译测试，运行单元测试测时需要编译整个工程**，并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后，执行下面的命令来运行单测：
+请注意，**不同于Op的编译测试，运行单元测试测时需要编译整个工程**，并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后，执行下面的命令来运行单元测试：
 
 ```bash
 make test ARGS="-R test_mul_op -V"

paddle/cuda/include/hl_cpu_gru.cuh

@@ -18,14 +18,6 @@ limitations under the License. */
 
 #ifndef __NVCC__
 
-#include "paddle/math/MathFunctions.h"
-
-#ifndef PADDLE_TYPE_DOUBLE
-#define     CBLAS_GEMM     paddle::gemm<float>
-#else
-#define     CBLAS_GEMM     paddle::gemm<double>
-#endif
-
 template<class OpResetOutput>
 void hl_naive_gru_forward_reset_output(OpResetOutput opResetOutput,
                                        real *gateValue,
@@ -210,51 +202,6 @@ inline void forward_final_output(OpFinalOutput opFinalOutput,
   }
 }
 
-template<class OpResetOutput, class OpFinalOutput>
-void hl_cpu_gru_forward(OpResetOutput opResetOutput,
-                        OpFinalOutput opFinalOutput,
-                        hl_gru_value value,
-                        int frameSize,
-                        int batchSize,
-                        hl_activation_mode_t active_node,
-                        hl_activation_mode_t active_gate) {
-  if (value.prevOutValue) {
-    CBLAS_GEMM(CblasNoTrans,
-               CblasNoTrans,
-               batchSize,
-               2 * frameSize,
-               frameSize,
-               1,
-               value.prevOutValue,
-               frameSize,
-               value.gateWeight,
-               frameSize * 2,
-               1,
-               value.gateValue,
-               frameSize * 3);
-  }
-
-  forward_reset_output(opResetOutput, value, frameSize, batchSize, active_gate);
-
-  if (value.prevOutValue) {
-    CBLAS_GEMM(CblasNoTrans,
-               CblasNoTrans,
-               batchSize,
-               frameSize,
-               frameSize,
-               1,
-               value.resetOutputValue,
-               frameSize,
-               value.stateWeight,
-               frameSize,
-               1,
-               value.gateValue + frameSize * 2,
-               frameSize * 3);
-  }
-
-  forward_final_output(opFinalOutput, value, frameSize, batchSize, active_node);
-}
-
 template<class OpStateGrad>
 void hl_naive_gru_backward_state_grad(OpStateGrad opStateGrad,
                                       real *gateValue,
@@ -525,86 +472,6 @@ inline void backward_reset_grad(OpResetGrad opResetGrad,
   }
 }
 
-template<class OpStateGrad, class OpResetGrad>
-void hl_cpu_gru_backward(OpStateGrad opStateGrad,
-                         OpResetGrad opResetGrad,
-                         hl_gru_value value,
-                         hl_gru_grad  grad,
-                         int frameSize,
-                         int batchSize,
-                         hl_activation_mode_t active_node,
-                         hl_activation_mode_t active_gate) {
-  backward_state_grad(opStateGrad, value, grad,
-    frameSize, batchSize, active_node);
-
-  if (value.prevOutValue && grad.prevOutGrad) {
-    CBLAS_GEMM(CblasNoTrans,
-               CblasTrans,
-               batchSize,
-               frameSize,
-               frameSize,
-               1,
-               grad.gateGrad + frameSize * 2,
-               frameSize * 3,
-               value.stateWeight,
-               frameSize,
-               0,
-               grad.resetOutputGrad,
-               frameSize);
-
-    if (grad.stateWeightGrad) {
-      CBLAS_GEMM(CblasTrans,
-                 CblasNoTrans,
-                 frameSize,
-                 frameSize,
-                 batchSize,
-                 1,
-                 value.resetOutputValue,
-                 frameSize,
-                 grad.gateGrad + frameSize * 2,
-                 frameSize * 3,
-                 1,
-                 grad.stateWeightGrad,
-                 frameSize);
-    }
-  }
-
-  backward_reset_grad(opResetGrad, value, grad,
-    frameSize, batchSize, active_gate);
-
-  if (grad.prevOutGrad && value.prevOutValue) {
-    CBLAS_GEMM(CblasNoTrans,
-               CblasTrans,
-               batchSize,
-               frameSize,
-               frameSize * 2,
-               1,
-               grad.gateGrad,
-               frameSize * 3,
-               value.gateWeight,
-               frameSize * 2,
-               1,
-               grad.prevOutGrad,
-               frameSize);
-
-    if (grad.gateWeightGrad) {
-      CBLAS_GEMM(CblasTrans,
-                 CblasNoTrans,
-                 frameSize,
-                 frameSize * 2,
-                 batchSize,
-                 1,
-                 value.prevOutValue,
-                 frameSize,
-                 grad.gateGrad,
-                 frameSize * 3,
-                 1,
-                 grad.gateWeightGrad,
-                 frameSize * 2);
-    }
-  }
-}
-
 #endif
 
 #endif  // HL_CPU_GRU_CUH_

paddle/framework/attribute.h

@@ -41,11 +41,23 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
 
 // check whether a value(attribute) fit a certain limit
 template <typename T>
-class LargerThanChecker {
+class GreaterThanChecker {
  public:
-  explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  explicit GreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
   void operator()(T& value) const {
-    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
+    PADDLE_ENFORCE(value > lower_bound_, "larger_than check fails.");
+  }
+
+ private:
+  T lower_bound_;
+};
+
+template <typename T>
+class EqualGreaterThanChecker {
+ public:
+  explicit EqualGreaterThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  void operator()(T& value) const {
+    PADDLE_ENFORCE_GE(value, lower_bound_, "equal_larger_than check fails.");
   }
 
  private:
@@ -110,8 +122,13 @@ class TypedAttrChecker {
     return *this;
   }
 
-  TypedAttrChecker& LargerThan(const T& lower_bound) {
-    value_checkers_.push_back(LargerThanChecker<T>(lower_bound));
+  TypedAttrChecker& GreaterThan(const T& lower_bound) {
+    value_checkers_.push_back(GreaterThanChecker<T>(lower_bound));
+    return *this;
+  }
+
+  TypedAttrChecker& EqualGreaterThan(const T& lower_bound) {
+    value_checkers_.push_back(EqualGreaterThanChecker<T>(lower_bound));
     return *this;
   }
 

paddle/framework/backward.md

@@ -2,20 +2,20 @@
 
 ## Motivation
 
-In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
-  
+In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the gradient operators/expressions together with the chain rule. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
+
 ## Backward Operator Registry
 
-A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs and output gradients and then calculate its input gradients.
+A backward network is built up with several backward operators. Backward operators take forward operators' inputs outputs, and output gradients and then calculate its input gradients.
 
 |                        | forward operator | backward operator 
 | ---------------------- | ---------------- |------------------------- |		
 | **Operator::inputs_**  | Inputs       | Inputs, Outputs, OutputGradients |	
 | **Operator::outputs_** | Outputs          | InputGradients            |
 
- In most cases, there is a one-to-one correspondence between forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
+ In most cases, there is a one-to-one correspondence between the forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
 
-For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro:
+For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro:
 
 ```cpp
 REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
@@ -27,17 +27,17 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
 
 ## Backward Opeartor Creating
 
-Given a certain forward operator, we can get its corresponding backward opeartor by calling:
+Given a certain forward operator, we can get its corresponding backward operator by calling:
 
 ```cpp
 OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op);
-``` 
+```
 
 The function `BuildGradOp` will sequentially execute following processes:
 
 1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`.
 
-2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these are not necessary for gradient computing.
+2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing.
 
 3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`.
 
@@ -49,31 +49,31 @@ A backward network is a series of backward operators. The main idea of building
 
 In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. 
 
-given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`.
+given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`, `InputGradients`.
 
 1. Op 
 
-   when the input forward network is a Op, return its gradient Operator Immediately.
+   when the input forward network is an Op, return its gradient Operator Immediately.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp.
+   when the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
 
-   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable.  
+   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwrite their shared input variable.  
 
    <p align="center">
-   <img src="./images/duplicate_op.png" width="70%" ><br/>
+   <img src="./images/duplicate_op.png" width="50%" ><br/>
 
-   1. shared variable in two operators. 
+   1. Shared variable in operators. 
 
    </p>
 
-   Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links. 
+   Share variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator replace the overwrite links. 
 
    <p align="center">
-   <img src="images/duplicate_op2.png" width="90%" ><br/>
+   <img src="images/duplicate_op2.png" width="50%" ><br/>
 
-   2. replace shared variable gradient with `Add` Operator
+   2. Replace shared variable's gradient with `Add` operator.
 
    </p>
 

paddle/framework/ddim.cc

@@ -283,5 +283,14 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
 DDim::DDim(std::initializer_list<int64_t> init_list) {
   *this = make_ddim(init_list);
 }
+
+DDim flatten_to_2d(const DDim& src, int num_col_dims) {
+  int rank = src.size();
+  return make_ddim({product(slice_ddim(src, 0, num_col_dims)),
+                    product(slice_ddim(src, num_col_dims, rank))});
+}
+
+DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); }
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/ddim.h

@@ -115,6 +115,12 @@ int arity(const DDim& ddim);
 
 std::ostream& operator<<(std::ostream&, const DDim&);
 
+// Reshape a tensor to a matrix. The matrix's first dimension(column length)
+// will be the product of tensor's first `num_col_dims` dimensions.
+DDim flatten_to_2d(const DDim& src, int num_col_dims);
+
+DDim flatten_to_1d(const DDim& src);
+
 }  // namespace framework
 }  // namespace paddle
 

paddle/framework/eigen.h

@@ -63,20 +63,35 @@ struct EigenTensor {
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
-struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {};
+struct EigenMatrix : public EigenTensor<T, 2, MajorType, IndexType> {
+  static typename EigenMatrix::Type Reshape(Tensor& tensor, int num_col_dims) {
+    int rank = tensor.dims_.size();
+    PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
+                   "`num_col_dims` must be between (0, rank_of_tensor).");
+    return EigenMatrix::From(tensor,
+                             flatten_to_2d(tensor.dims(), num_col_dims));
+  }
+
+  static typename EigenMatrix::ConstType Reshape(const Tensor& tensor,
+                                                 int num_col_dims) {
+    int rank = tensor.dims_.size();
+    PADDLE_ENFORCE(num_col_dims > 0 && num_col_dims < rank,
+                   "`num_col_dims` must be between (0, rank_of_tensor).");
+    return EigenMatrix::From(tensor,
+                             flatten_to_2d(tensor.dims(), num_col_dims));
+  }
+};
 
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 struct EigenVector : public EigenTensor<T, 1, MajorType, IndexType> {
   // Flatten reshapes a Tensor into an EigenVector.
   static typename EigenVector::Type Flatten(Tensor& tensor) {
-    return EigenVector::From(
-        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+    return EigenVector::From(tensor, {product(tensor.dims_)});
   }
 
   static typename EigenVector::ConstType Flatten(const Tensor& tensor) {
-    return EigenVector::From(
-        tensor, make_ddim({static_cast<int>(product(tensor.dims_))}));
+    return EigenVector::From(tensor, {product(tensor.dims_)});
   }
 };
 

paddle/framework/eigen_test.cc

@@ -108,5 +108,24 @@ TEST(Eigen, Matrix) {
   }
 }
 
+TEST(Eigen, MatrixReshape) {
+  Tensor t;
+  float* p = t.mutable_data<float>({2, 3, 6, 4}, platform::CPUPlace());
+  for (int i = 0; i < 2 * 3 * 6 * 4; ++i) {
+    p[i] = static_cast<float>(i);
+  }
+
+  EigenMatrix<float>::Type em = EigenMatrix<float>::Reshape(t, 2);
+
+  ASSERT_EQ(2 * 3, em.dimension(0));
+  ASSERT_EQ(6 * 4, em.dimension(1));
+
+  for (int i = 0; i < 2 * 3; i++) {
+    for (int j = 0; j < 6 * 4; j++) {
+      ASSERT_NEAR(i * 6 * 4 + j, em(i, j), 1e-6f);
+    }
+  }
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/framework.proto

@@ -87,3 +87,24 @@ message OpProto {
   repeated Attr attrs = 4;
   required string comment = 5;
 }
+
+enum DataType {
+  BOOL = 0;
+  INT16 = 1;
+  INT32 = 2;
+  INT64 = 3;
+  FP16 = 4;
+  FP32 = 5;
+  FP64 = 6;
+}
+
+message LoDTensorDesc {
+  required DataType data_type = 1;
+  repeated int32 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
+  optional int32 lod_level = 3 [ default = 0 ];
+}
+
+message VarDesc {
+  required string name = 1;
+  optional LoDTensorDesc lod_tensor = 2;
+}

paddle/framework/grad_op_builder_test.cc

@@ -3,7 +3,7 @@
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 
-USE_OP(add_two);
+USE_OP(add);
 
 namespace paddle {
 namespace framework {
@@ -41,7 +41,7 @@ namespace f = paddle::framework;
 
 TEST(GradOpBuilder, AddTwo) {
   std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
-      "add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
+      "add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
   std::shared_ptr<f::OperatorBase> grad_add_op =
       f::OpRegistry::CreateGradOp(*add_op);
   EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);

paddle/framework/lod_tensor.cc

@@ -19,8 +19,8 @@
 namespace paddle {
 namespace framework {
 
-LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) {
-  LOD new_lod;
+LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end) {
+  LoD new_lod;
   new_lod.reserve(level_end - level_begin);
   for (size_t i = level_begin; i < level_end; i++) {
     new_lod.emplace_back(in.at(i));
@@ -28,10 +28,10 @@ LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end) {
   return new_lod;
 }
 
-LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
                  size_t elem_end) {
   // slice the lod.
-  LOD new_lod;
+  LoD new_lod;
   new_lod.reserve(in.size() - level);
   auto start = in.at(level)[elem_begin];
   auto end = in.at(level)[elem_end];
@@ -46,13 +46,13 @@ LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
     std::transform(new_lod.back().begin(), new_lod.back().end(),
                    new_lod.back().begin(),
                    [start](int v) { return v - start; });
-    PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD");
+    PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LoD");
   }
   PADDLE_ENFORCE_LE(new_lod.size(), in.size());
   return new_lod;
 }
 
-bool operator==(const LOD& a, const LOD& b) {
+bool operator==(const LoD& a, const LoD& b) {
   if (a.size() != b.size()) {
     return false;
   }
@@ -72,12 +72,12 @@ bool operator==(const LOD& a, const LOD& b) {
   return true;
 }
 
-void LODTensor::SliceLevels(size_t level_begin, size_t level_end) {
+void LoDTensor::SliceLevels(size_t level_begin, size_t level_end) {
   auto new_lod = framework::SliceLevels(lod_, level_begin, level_end);
   lod_ = new_lod;
 }
 
-void LODTensor::SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) {
+void LoDTensor::SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) {
   PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
                  NumLevels());
   PADDLE_ENFORCE(elem_begin < NumElements(level),

paddle/framework/lod_tensor.h

@@ -35,34 +35,34 @@ template <typename T>
 using Vector = thrust::host_vector<T>;
 #endif
 
-using LOD = std::vector<Vector<size_t>>;
+using LoD = std::vector<Vector<size_t>>;
 
-LOD SliceLevels(const LOD& in, size_t level_begin, size_t level_end);
+LoD SliceLevels(const LoD& in, size_t level_begin, size_t level_end);
 
-LOD SliceInLevel(const LOD& in, size_t level, size_t elem_begin,
+LoD SliceInLevel(const LoD& in, size_t level, size_t elem_begin,
                  size_t elem_end);
 
-bool operator==(const LOD& a, const LOD& b);
+bool operator==(const LoD& a, const LoD& b);
 
 /*
- * LODTensor (Level of details Tensor)
+ * LoDTensor (Level of details Tensor)
  * see https://en.wikipedia.org/wiki/Level_of_details for reference.
  */
-class LODTensor {
+class LoDTensor {
  public:
-  LODTensor() {}
-  LODTensor(const LOD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
+  LoDTensor() {}
+  LoDTensor(const LoD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
 
-  void set_lod(const LOD& lod) { lod_ = lod; }
+  void set_lod(const LoD& lod) { lod_ = lod; }
 
   void set_tensor(Tensor* tensor) { tensor_ = tensor; }
 
   Tensor& tensor() { return *tensor_; }
 
-  LOD lod() { return lod_; }
+  LoD lod() { return lod_; }
 
   /*
-   * Get a element from LOD.
+   * Get a element from LoD.
    */
   size_t lod_element(size_t level, size_t elem) const {
     PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
@@ -74,7 +74,7 @@ class LODTensor {
   }
 
   /*
-   * Number of LODTensor's levels, each level has units of data, for example,
+   * Number of LoDTensor's levels, each level has units of data, for example,
    * in the sentence's view, article, paragraph, sentence are 3 levels.
    */
   size_t NumLevels() const { return lod_.size(); }
@@ -100,7 +100,7 @@ class LODTensor {
   void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end);
 
  private:
-  LOD lod_;
+  LoD lod_;
   Tensor* tensor_;  // not owned
 };
 }  // namespace framework

paddle/framework/lod_tensor.md

@@ -94,7 +94,7 @@ Let's go on slicing this slice.  Its <1,1>-slice is
 |||
 ```
 
-### The General Slicing Algorithm
+### The Slicing Algorithm
 
 The algorithm, with over-simplified data structure, is defined as
 
@@ -106,17 +106,41 @@ struct LoDTensor {
   float* tensor_;
 };
 
-LoDTensor Slice(const LoDTensor& lodt, int level, int sequence) {
+LoDTensor Slice(const LoDTensor& lodt, int level, int sequence);
+```
+
+Let us revisit the example above
 
-}
+```
+         3
+3           1  2
+3   2  4    1  2  3
+||| || |||| |  || |||
 ```
 
-### Slicing the Top Level
+Suppose that we want to retrieve the <1,2>-slice
 
-Please be aware that an RNN operator only slices the top level of a LoD Tensor to get the step inputs.
+```
+2
+2  3
+|| |||
+```
 
-```c++
-LoDTensor Slice(const LoDTensor& lodt, int sequence) {
+we will need to find out the starting position of this slice by summing over all leaf nodes in `LoD` to the left of the slice, i.e., 3 + 2 + 4 + 1 = 10.
+
+To avoid the traversal of the LoD tree at slcing time,  we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level.  For example, above LoD Tensor can be transformed into
+
+```
+        0
+0           9  10
+0   3  5    9  10 12
+||| || |||| |  || |||
+```
+
+We don't really need the 0 on top, so the LoD Tensor could be
 
-}
+```
+0           9  10
+0   3  5    9  10 12
+||| || |||| |  || |||
 ```

paddle/framework/lod_tensor_test.cc

@@ -21,7 +21,7 @@
 namespace paddle {
 namespace framework {
 
-class LODTensorTester : public ::testing::Test {
+class LoDTensorTester : public ::testing::Test {
  public:
   virtual void SetUp() override {
     // tensor's batch_size: 30
@@ -29,7 +29,7 @@ class LODTensorTester : public ::testing::Test {
     // 0 10 20
     // 0 5 10 15 20
     // 0 2 5 7 10 12 15 20
-    LOD lod;
+    LoD lod;
     lod.push_back(std::vector<size_t>{0, 10, 20});
     lod.push_back(std::vector<size_t>{0, 5, 10, 15, 20});
     lod.push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
@@ -47,21 +47,21 @@ class LODTensorTester : public ::testing::Test {
  protected:
   platform::CPUPlace place;
   Tensor tensor;
-  LODTensor lod_tensor;
+  LoDTensor lod_tensor;
 };
 
-TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
+TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
 
-TEST_F(LODTensorTester, NumElements) {
+TEST_F(LoDTensorTester, NumElements) {
   ASSERT_EQ(lod_tensor.NumElements(0), 2UL);
   ASSERT_EQ(lod_tensor.NumElements(1), 4UL);
   ASSERT_EQ(lod_tensor.NumElements(2), 8UL);
 }
 
-TEST_F(LODTensorTester, SliceLevels) {
+TEST_F(LoDTensorTester, SliceLevels) {
   // slice 1 level
   for (size_t level = 0; level < 3UL; ++level) {
-    LODTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor;
     new_lod_tensor.SliceLevels(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
     ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
@@ -70,7 +70,7 @@ TEST_F(LODTensorTester, SliceLevels) {
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    LODTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor;
     new_lod_tensor.SliceLevels(level, level + 2);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
     ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
@@ -80,9 +80,9 @@ TEST_F(LODTensorTester, SliceLevels) {
   }
 }
 
-TEST_F(LODTensorTester, SliceInLevel) {
+TEST_F(LoDTensorTester, SliceInLevel) {
   size_t level = 0;
-  LODTensor new_lod_tensor = lod_tensor;
+  LoDTensor new_lod_tensor = lod_tensor;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
   EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);

paddle/framework/op_registry_test.cc

@@ -21,7 +21,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     AddOutput("output", "output of cosine op");
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
-        .LargerThan(0.0);
+        .GreaterThan(0.0);
     AddComment("This is cos op");
   }
 };
@@ -80,7 +80,7 @@ TEST(OpRegistry, CreateOp) {
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
-  float scale_get = op->GetAttr<float>("scale");
+  float scale_get = op->Attr<float>("scale");
   ASSERT_EQ(scale_get, scale);
 }
 
@@ -121,7 +121,7 @@ TEST(OpRegistry, DefaultValue) {
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
-  ASSERT_EQ(op->GetAttr<float>("scale"), 1.0);
+  ASSERT_EQ(op->Attr<float>("scale"), 1.0);
 }
 
 TEST(OpRegistry, CustomChecker) {
@@ -172,6 +172,6 @@ TEST(OpRegistry, CustomChecker) {
   paddle::platform::CPUDeviceContext dev_ctx;
   paddle::framework::Scope scope;
   op->Run(scope, dev_ctx);
-  int test_attr = op->GetAttr<int>("test_attr");
+  int test_attr = op->Attr<int>("test_attr");
   ASSERT_EQ(test_attr, 4);
 }
\ No newline at end of file

paddle/framework/operator.h

@@ -69,7 +69,7 @@ class OperatorBase {
   virtual ~OperatorBase() {}
 
   template <typename T>
-  inline const T& GetAttr(const std::string& name) const {
+  inline const T& Attr(const std::string& name) const {
     PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
                    name);
     return boost::get<T>(attrs_.at(name));
@@ -238,8 +238,8 @@ class InferShapeContext {
   const Scope& scope() const { return scope_; }
 
   template <typename T>
-  inline const T& GetAttr(const std::string& name) const {
-    return op_.GetAttr<T>(name);
+  inline const T& Attr(const std::string& name) const {
+    return op_.Attr<T>(name);
   }
 
   size_t InputSize(const std::string& name) const {

paddle/framework/operator_test.cc

@@ -102,7 +102,7 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
     AddOutput("y", "output of test op");
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
-        .LargerThan(0.0);
+        .GreaterThan(0.0);
     AddComment("This is test op");
   }
 };
@@ -140,7 +140,7 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker
     AddOutput("ys", "outputs of test op").AsDuplicable();
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
-        .LargerThan(0.0);
+        .GreaterThan(0.0);
     AddComment("This is test op");
   }
 };

paddle/framework/tensor.h

@@ -43,6 +43,9 @@ class Tensor {
   template <typename T, size_t D, int MajorType, typename IndexType>
   friend struct EigenTensor;
 
+  template <typename T, int MajorType, typename IndexType>
+  friend struct EigenMatrix;
+
   template <typename T, int MajorType, typename IndexType>
   friend struct EigenVector;
 

paddle/framework/tensor_impl.h

@@ -148,5 +148,13 @@ inline Tensor& Tensor::Resize(const DDim& dims) {
 
 inline const DDim& Tensor::dims() const { return dims_; }
 
+template <typename T>
+inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
+  Tensor res;
+  res.ShareDataWith<T>(src);
+  res.Resize(flatten_to_2d(src.dims(), num_col_dims));
+  return res;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/tensor_test.cc

@@ -262,3 +262,16 @@ TEST(Tensor, CopyFrom) {
   }
 #endif
 }
+
+TEST(Tensor, ReshapeToMatrix) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  Tensor src;
+  int* src_ptr = src.mutable_data<int>({2, 3, 4, 9}, CPUPlace());
+  for (int i = 0; i < 2 * 3 * 4 * 9; ++i) {
+    src_ptr[i] = i;
+  }
+  Tensor res = ReshapeToMatrix<int>(src, 2);
+  ASSERT_EQ(res.dims()[0], 2 * 3);
+  ASSERT_EQ(res.dims()[1], 4 * 9);
+}
\ No newline at end of file

paddle/function/CMakeLists.txt

@@ -44,6 +44,7 @@ if(WITH_GPU)
     add_simple_unittest(RowConvOpTest)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
+    add_simple_unittest(SwitchOpTest)
 endif()
 
 add_simple_unittest(Im2ColTest)

paddle/function/EigenGemm.cpp

@@ -83,9 +83,9 @@ struct EigenBlasGemm {
 };
 
 #ifdef PADDLE_TYPE_DOUBLE
-template class EigenBlasGemm<double>;
+template struct EigenBlasGemm<double>;
 #else
-template class EigenBlasGemm<float>;
+template struct EigenBlasGemm<float>;
 #endif
 
 }  // namespace paddle

paddle/function/GruFunctor.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 "GemmFunctor.h"
+#include "hl_cpu_gru.cuh"
+
+namespace paddle {
+
+template <DeviceType Device, class T>
+struct GruFunctor {
+  template <class OpResetOutput, class OpFinalOutput>
+  static void compute(OpResetOutput opResetOutput,
+                      OpFinalOutput opFinalOutput,
+                      hl_gru_value value,
+                      int frameSize,
+                      int batchSize,
+                      hl_activation_mode_t active_node,
+                      hl_activation_mode_t active_gate) {
+#ifndef __NVCC__
+    if (value.prevOutValue) {
+      BlasGemm<Device, T>::compute(false,
+                                   false,
+                                   batchSize,
+                                   2 * frameSize,
+                                   frameSize,
+                                   1,
+                                   value.prevOutValue,
+                                   frameSize,
+                                   value.gateWeight,
+                                   frameSize * 2,
+                                   1,
+                                   value.gateValue,
+                                   frameSize * 3);
+    }
+
+    forward_reset_output(
+        opResetOutput, value, frameSize, batchSize, active_gate);
+
+    if (value.prevOutValue) {
+      BlasGemm<Device, T>::compute(false,
+                                   false,
+                                   batchSize,
+                                   frameSize,
+                                   frameSize,
+                                   1,
+                                   value.resetOutputValue,
+                                   frameSize,
+                                   value.stateWeight,
+                                   frameSize,
+                                   1,
+                                   value.gateValue + frameSize * 2,
+                                   frameSize * 3);
+    }
+
+    forward_final_output(
+        opFinalOutput, value, frameSize, batchSize, active_node);
+#endif
+  }
+};
+
+template <DeviceType Device, class T>
+struct GruGradFunctor {
+  template <class OpStateGrad, class OpResetGrad>
+  static void compute(OpStateGrad opStateGrad,
+                      OpResetGrad opResetGrad,
+                      hl_gru_value value,
+                      hl_gru_grad grad,
+                      int frameSize,
+                      int batchSize,
+                      hl_activation_mode_t active_node,
+                      hl_activation_mode_t active_gate) {
+#ifndef __NVCC__
+    backward_state_grad(
+        opStateGrad, value, grad, frameSize, batchSize, active_node);
+
+    if (value.prevOutValue && grad.prevOutGrad) {
+      BlasGemm<Device, T>::compute(false,
+                                   true,
+                                   batchSize,
+                                   frameSize,
+                                   frameSize,
+                                   1,
+                                   grad.gateGrad + frameSize * 2,
+                                   frameSize * 3,
+                                   value.stateWeight,
+                                   frameSize,
+                                   0,
+                                   grad.resetOutputGrad,
+                                   frameSize);
+
+      if (grad.stateWeightGrad) {
+        BlasGemm<Device, T>::compute(true,
+                                     false,
+                                     frameSize,
+                                     frameSize,
+                                     batchSize,
+                                     1,
+                                     value.resetOutputValue,
+                                     frameSize,
+                                     grad.gateGrad + frameSize * 2,
+                                     frameSize * 3,
+                                     1,
+                                     grad.stateWeightGrad,
+                                     frameSize);
+      }
+    }
+
+    backward_reset_grad(
+        opResetGrad, value, grad, frameSize, batchSize, active_gate);
+
+    if (grad.prevOutGrad && value.prevOutValue) {
+      BlasGemm<Device, T>::compute(false,
+                                   true,
+                                   batchSize,
+                                   frameSize,
+                                   frameSize * 2,
+                                   1,
+                                   grad.gateGrad,
+                                   frameSize * 3,
+                                   value.gateWeight,
+                                   frameSize * 2,
+                                   1,
+                                   grad.prevOutGrad,
+                                   frameSize);
+
+      if (grad.gateWeightGrad) {
+        BlasGemm<Device, T>::compute(true,
+                                     false,
+                                     frameSize,
+                                     frameSize * 2,
+                                     batchSize,
+                                     1,
+                                     value.prevOutValue,
+                                     frameSize,
+                                     grad.gateGrad,
+                                     frameSize * 3,
+                                     1,
+                                     grad.gateWeightGrad,
+                                     frameSize * 2);
+      }
+    }
+#endif
+  }
+};
+
+}  // namespace paddle

paddle/function/Im2Col.h

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

paddle/function/MulOp.cpp

@@ -13,18 +13,10 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "MulOp.h"
-/// todo(tianbing), delete it
-#include <iostream>
-#include "paddle/math/MathFunctions.h"
+#include "GemmFunctor.h"
 #include "paddle/math/SIMDFunctions.h"
 #include "paddle/utils/ThreadLocal.h"
 
-#ifndef PADDLE_TYPE_DOUBLE
-#define GEMM paddle::gemm<float>
-#else
-#define GEMM paddle::gemm<double>
-#endif
-
 namespace {
 inline void vecAddTo(real* a, const real* b, real scaleB, size_t len) {
   for (unsigned int i = 0; i < len; ++i) {
@@ -114,19 +106,20 @@ void MulOp<DEVICE_TYPE_CPU>(CpuMatrix& out,
                             real scaleT,
                             bool aTrans,
                             bool bTrans) {
-  GEMM(aTrans ? CblasTrans : CblasNoTrans,
-       bTrans ? CblasTrans : CblasNoTrans,
-       out.getHeight(),
-       out.getWidth(),
-       !aTrans ? a.getWidth() : a.getHeight(),
-       scaleAB,
-       a.getData(),
-       a.getStride(),
-       b.getData(),
-       b.getStride(),
-       scaleT,
-       out.getData(),
-       out.getStride());
+  BlasGemm<DEVICE_TYPE_CPU, real>::compute(
+      aTrans,
+      bTrans,
+      out.getHeight(),
+      out.getWidth(),
+      !aTrans ? a.getWidth() : a.getHeight(),
+      scaleAB,
+      a.getData(),
+      a.getStride(),
+      b.getData(),
+      b.getStride(),
+      scaleT,
+      out.getData(),
+      out.getStride());
 }
 
 /// dense matrix (+)= sparse matrix * dense matrix

paddle/function/SwitchOp.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "SwitchOp.h"
+#include "paddle/math/Vector.h"
+
+namespace paddle {
+
+template <>
+void NCHW2NHWC<DEVICE_TYPE_CPU>(real* outputs,
+                                const real* inputs,
+                                const int num,
+                                const int inC,
+                                const int inH,
+                                const int inW,
+                                const int argType) {
+  for (int n = 0; n < num; ++n) {
+    for (int c = 0; c < inC; ++c) {
+      for (int h = 0; h < inH; ++h) {
+        for (int w = 0; w < inW; ++w) {
+          if (argType == ADD_TO) {
+            outputs[((n * inH + h) * inW + w) * inC + c] += *(inputs++);
+          } else {
+            outputs[((n * inH + h) * inW + w) * inC + c] = *(inputs++);
+          }
+        }
+      }
+    }
+  }
+}
+
+template <>
+void NHWC2NCHW<DEVICE_TYPE_CPU>(real* outputs,
+                                const real* inputs,
+                                const int num,
+                                const int inH,
+                                const int inW,
+                                const int inC,
+                                const int argType) {
+  for (int n = 0; n < num; ++n) {
+    for (int h = 0; h < inH; ++h) {
+      for (int w = 0; w < inW; ++w) {
+        for (int c = 0; c < inC; ++c) {
+          if (argType == ADD_TO) {
+            outputs[((n * inC + c) * inH + h) * inW + w] += *(inputs++);
+          } else {
+            outputs[((n * inC + c) * inH + h) * inW + w] = *(inputs++);
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief  Switch dimension order of image input.
+ *         The input and output is a 4D tensor. Switch order
+ *         'batch_size,channels, height, width' to
+ *         order 'batch_size, height, width, channels'.
+ *
+ * Argument in this Function:
+ * \param inputs  input data with order 'batch_size,channels, height, width'.
+ * \param outputs output data with order 'batch_size, height, width, channels'.
+ */
+template <DeviceType Device>
+class NCHW2NHWCFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override {}
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+
+    size_t num = inputs[0].shape()[0];
+    size_t inC = inputs[0].shape()[1];
+    size_t inH = inputs[0].shape()[2];
+    size_t inW = inputs[0].shape()[3];
+    NCHW2NHWC<Device>(outputs[0].data<real>(),
+                      inputs[0].data<real>(),
+                      num,
+                      inC,
+                      inH,
+                      inW,
+                      outputs[0].getArgType());
+  }
+};
+
+/**
+ * \brief  Switch dimension order of image input.
+ *         The input and output is a 4D tensor. Switch order
+ *         'batch_size, height, width, channels' to
+ *         order 'batch_size, channels, height, width'.
+ *
+ * Argument in this Function:
+ * \param inputs  input data with order 'batch_size, height, width, channels'.
+ * \param outputs output data with order 'batch_size, channels, height, width'.
+ */
+template <DeviceType Device>
+class NHWC2NCHWFunc : public FunctionBase {
+public:
+  void init(const FuncConfig& config) override {}
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(1UL, inputs.size());
+    CHECK_EQ(1UL, outputs.size());
+
+    size_t num = inputs[0].shape()[0];
+    size_t inH = inputs[0].shape()[1];
+    size_t inW = inputs[0].shape()[2];
+    size_t inC = inputs[0].shape()[3];
+
+    NHWC2NCHW<Device>(outputs[0].data<real>(),
+                      inputs[0].data<real>(),
+                      num,
+                      inH,
+                      inW,
+                      inC,
+                      outputs[0].getArgType());
+  }
+};
+
+REGISTER_TYPED_FUNC(NCHW2NHWC, CPU, NCHW2NHWCFunc);
+REGISTER_TYPED_FUNC(NHWC2NCHW, CPU, NHWC2NCHWFunc);
+#ifndef PADDLE_ONLY_CPU
+REGISTER_TYPED_FUNC(NCHW2NHWC, GPU, NCHW2NHWCFunc);
+REGISTER_TYPED_FUNC(NHWC2NCHW, GPU, NHWC2NCHWFunc);
+#endif
+
+}  // namespace paddle

paddle/function/SwitchOp.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 "Function.h"
+
+namespace paddle {
+
+/**
+ * \brief  This funtion switch dimension order of image input.
+ *         The input and output is a 4D tensor. Switch order 'batch_size,
+ *channels, height, width' to
+ *         order 'batch_size, height, width, channels'.
+ *
+ * \param[out] outputs save results.
+ * \param[in]  inputs  input data.
+ * \param[in]  num     batch size of input data.
+ * \param[in]  inC     channel number of input data.
+ * \param[in]  inH     height of input data.
+ * \param[in]  inH     with of input data.
+ * \param[in]  argType     type of output argument.
+ */
+template <DeviceType Device>
+void NCHW2NHWC(real* outputs,
+               const real* inputs,
+               const int num,
+               const int inC,
+               const int inH,
+               const int inW,
+               const int argtype);
+
+/**
+ * \brief  This funtion switch dimension order of image input.
+ *         The input and output is a 4D tensor. Switch order 'batch_size,
+ *height, width, channels' to
+ *         order 'batch_size, channels, height, width'.
+ *
+ * \param[out] inGrad  gradients of previous layer.
+ * \param[in]  outGrad output gradients.
+ * \param[in]  num     batch size of input data.
+ * \param[in]  inH     height of input data.
+ * \param[in]  inW     with of input data.
+ * \param[in]  inC     channel number of input data.
+ * \param[in]  argType     type of output argument.
+ */
+template <DeviceType Device>
+void NHWC2NCHW(real* inGrad,
+               const real* outGrad,
+               const int num,
+               const int inH,
+               const int inW,
+               const int inC,
+               const int argType);
+}  // namespace paddle

paddle/function/SwitchOpGpu.cu

+/* Copyright (c) 2016 Paddle
+
+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 "SwitchOp.h"
+#include "hl_base.h"
+
+namespace paddle {
+
+__global__ void KeNCHW2NHWC(real* outputs,
+                            const real* inputs,
+                            int inC,
+                            int inH,
+                            int inW,
+                            int nthreads,
+                            int argType) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int w = idx % inW;
+    const int h = (idx / inW) % inH;
+    const int c = (idx / inW / inH) % inC;
+    const int n = idx / inW / inH / inC;
+
+    const int off = ((n * inH + h) * inW + w) * inC + c;
+    if (argType == ADD_TO) {
+      outputs[off] += inputs[idx];
+    } else {
+      outputs[off] = inputs[idx];
+    }
+  }
+}
+
+template <>
+void NCHW2NHWC<DEVICE_TYPE_GPU>(real* outputs,
+                                const real* inputs,
+                                const int num,
+                                const int inC,
+                                const int inH,
+                                const int inW,
+                                const int argType) {
+  size_t nth = num * inC * inH * inW;
+  int blockSize = 1024;
+  int gridSize = (nth + 1024 - 1) / 1024;
+  KeNCHW2NHWC<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      outputs, inputs, inC, inH, inW, nth, argType);
+  CHECK_SYNC("NCHW2NHWC");
+}
+
+__global__ void KeNHWC2NCHW(real* outputs,
+                            const real* inputs,
+                            int inH,
+                            int inW,
+                            int inC,
+                            int nthreads,
+                            int argType) {
+  const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+  if (idx < nthreads) {
+    const int c = idx % inC;
+    const int w = (idx / inC) % inW;
+    const int h = (idx / inC / inW) % inH;
+    const int n = idx / inW / inH / inC;
+
+    const int off = ((n * inC + c) * inH + h) * inW + w;
+    if (argType == ADD_TO) {
+      outputs[off] += inputs[idx];
+    } else {
+      outputs[off] = inputs[idx];
+    }
+  }
+}
+
+template <>
+void NHWC2NCHW<DEVICE_TYPE_GPU>(real* outputs,
+                                const real* inputs,
+                                const int num,
+                                const int inH,
+                                const int inW,
+                                const int inC,
+                                const int argType) {
+  int nth = num * inC * inH * inW;
+  int blockSize = 1024;
+  int gridSize = (nth + 1024 - 1) / 1024;
+  KeNHWC2NCHW<<<gridSize, blockSize, 0, STREAM_DEFAULT>>>(
+      outputs, inputs, inH, inW, inC, nth, argType);
+  CHECK_SYNC("NHWC2NCHW");
+}
+
+}  // namespace paddle

paddle/function/SwitchOpTest.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "FunctionTest.h"
+
+namespace paddle {
+
+TEST(Pad, real) {
+  for (size_t numSamples : {1, 4, 8, 16}) {
+    for (size_t channels : {1, 4, 8, 16}) {
+      for (size_t imgSizeH : {1, 4, 8, 16}) {
+        for (size_t imgSizeW : {1, 4, 8, 16}) {
+          VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
+                  << " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW;
+          for (bool test_grad : {true, false}) {
+            CpuGpuFuncCompare compare(test_grad ? "NHWC2NCHW" : "NCHW2NHWC",
+                                      FuncConfig());
+            TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
+            TensorShape outDims{numSamples, imgSizeH, imgSizeW, channels};
+            compare.addInputs(
+                BufferArg(VALUE_TYPE_FLOAT, test_grad ? outDims : inDims));
+            compare.addOutputs(BufferArg(
+                VALUE_TYPE_FLOAT, test_grad ? inDims : outDims, ASSIGN_TO));
+            compare.run();
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/function/neon/NeonDepthwiseConvTranspose.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "NeonDepthwiseConv.h"
+#include "paddle/function/ConvOp.h"
+
+namespace paddle {
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+template <DeviceType Device>
+class NeonDepthwiseConvTransposeFunction : public ConvFunctionBase {
+public:
+  void init(const FuncConfig& config) override {
+    ConvFunctionBase::init(config);
+  }
+
+  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    const TensorShape& input = inputs[0].shape();
+    const TensorShape& filter = inputs[1].shape();
+    const TensorShape& output = outputs[0].shape();
+    checkShape(input, filter, output);
+  }
+
+  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
+    CHECK_EQ(numInputs_, inputs.size());
+    CHECK_EQ(numOutputs_, outputs.size());
+    check(inputs, outputs);
+
+    const TensorShape& input = inputs[0].shape();
+    const TensorShape& filter = inputs[1].shape();
+    const TensorShape& output = outputs[0].shape();
+
+    int batchSize = input[0];
+    int inputChannels = input[1];
+    int inputHeight = input[2];
+    int inputWidth = input[3];
+    int filterHeight = getFilterHeight(filter);
+    int filterWidth = getFilterWidth(filter);
+    int outputChannels = output[1];
+    int outputHeight = output[2];
+    int outputWidth = output[3];
+    int filterMultiplier = outputChannels / groups_;
+    CHECK_EQ(inputChannels, groups_);
+
+    // only support strideH() == strideW() and filterHeight == filterWidth.
+    CHECK_EQ(strideH(), strideW());
+    CHECK_EQ(paddingH(), paddingW());
+    CHECK_EQ(filterHeight, filterWidth);
+
+    float* inputData = inputs[0].data<float>();
+    float* filterData = inputs[1].data<float>();
+    float* outputData = outputs[0].data<float>();
+
+    // padding the input, input -> inputPadding
+    float* inputPadding = inputData;
+    int padInputHeight =
+        (inputHeight - 1) * strideH() + 2 * filterHeight - 1 - 2 * paddingH();
+    int padInputWidth =
+        (inputWidth - 1) * strideW() + 2 * filterWidth - 1 - 2 * paddingW();
+
+    if (padInputHeight > inputHeight || padInputWidth > inputWidth) {
+      int newSize = batchSize * inputChannels * padInputHeight * padInputWidth;
+      resizeBuffer<Device>(newSize);
+      inputPadding = reinterpret_cast<float*>(memory_->getBuf());
+      if (strideH() == 1) {
+        neon::Padding<float>::run(inputData,
+                                  inputPadding,
+                                  batchSize * inputChannels,
+                                  inputHeight,
+                                  inputWidth,
+                                  padInputHeight,
+                                  padInputWidth);
+      } else if (strideH() == 2) {
+        neon::StridePadding::run(inputData,
+                                 inputPadding,
+                                 batchSize * inputChannels,
+                                 inputHeight,
+                                 inputWidth,
+                                 padInputHeight,
+                                 padInputWidth);
+      } else {
+        LOG(FATAL) << "Not supported";
+      }
+    }
+
+    std::function<void(
+        const float*, const float*, int, int, int, int, int, int, float*)>
+        DepthWiseConv;
+
+    if (filterWidth == 3) {
+      DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run;
+    } else if (filterWidth == 4) {
+      DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run;
+    } else {
+      LOG(FATAL) << "Not supported";
+    }
+
+    for (int i = 0; i < batchSize; i++) {
+      DepthWiseConv(inputPadding,
+                    filterData,
+                    padInputHeight,
+                    padInputWidth,
+                    outputChannels,
+                    outputHeight,
+                    outputWidth,
+                    filterMultiplier,
+                    outputData);
+      inputPadding += inputChannels * padInputHeight * padInputWidth;
+      outputData += outputChannels * outputHeight * outputWidth;
+    }
+  }
+};
+
+#ifndef PADDLE_TYPE_DOUBLE
+
+REGISTER_TYPED_FUNC(NeonDepthwiseConvTranspose,
+                    CPU,
+                    NeonDepthwiseConvTransposeFunction);
+
+#endif
+
+#endif
+
+}  // namespace paddle

paddle/function/neon/neon_util.h

@@ -33,12 +33,8 @@ inline float32_t vaddvq_f32(float32x4_t a) {
   return vget_lane_f32(vpadd_f32(v, v), 0);
 }
 
-inline float32x4_t vmlaq_laneq_f32(float32x4_t a,
-                                   float32x4_t b,
-                                   float32x4_t v,
-                                   const int lane) {
-  return vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane));
-}
+#define vmlaq_laneq_f32(a, b, v, lane) \
+  vmlaq_n_f32(a, b, vgetq_lane_f32(v, lane))
 #endif
 
 }  // namespace neon

paddle/gserver/layers/Conv3DLayer.cpp

@@ -83,8 +83,8 @@ void Conv3DLayer::forward(PassType passType) {
   int outWidth = getSize();
   resetOutput(batchSize, outWidth);
 
+  REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
   for (size_t i = 0; i != inputLayers_.size(); ++i) {
-    REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
     const MatrixPtr &inMat = getInputValue(i);
     const MatrixPtr &outMat = getOutputValue();
     int M = M_[i];
@@ -120,7 +120,6 @@ void Conv3DLayer::forward(PassType passType) {
     }
   }
   if (nullptr != this->biasParameter_) {
-    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
     this->addBias();
   }
   forwardActivation();
@@ -134,15 +133,14 @@ void Conv3DLayer::backward(const UpdateCallback &callback) {
     biases_->getParameterPtr()->incUpdate(callback);
   }
 
+  REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
   for (size_t i = 0; i != inputLayers_.size(); ++i) {
-    REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
     if (weights_[i]->getWGrad()) {
       bpropWeights(i);
     }
     if (getInputGrad(i)) {
       bpropData(i);
     }
-    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
     weights_[i]->getParameterPtr()->incUpdate(callback);
   }
 }

paddle/gserver/layers/DeConv3DLayer.cpp

@@ -84,8 +84,8 @@ void DeConv3DLayer::forward(PassType passType) {
   resetOutput(batchSize, outWidth);
   const MatrixPtr outMat = getOutputValue();
 
+  REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
   for (size_t i = 0; i != inputLayers_.size(); ++i) {
-    REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
     const MatrixPtr &inMat = getInputValue(i);
     int M = M_[i];
     int N = N_[i];
@@ -120,7 +120,6 @@ void DeConv3DLayer::forward(PassType passType) {
     }
   }
   if (nullptr != this->biasParameter_) {
-    REGISTER_TIMER_INFO("FwBiasTimer", getName().c_str());
     this->addBias();
   }
   forwardActivation();
@@ -133,12 +132,12 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
     bpropBiases();
     biases_->getParameterPtr()->incUpdate(callback);
   }
+  REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str());
   for (size_t i = 0; i < inputLayers_.size(); ++i) {
     if (weights_[i]->getWGrad() || this->needGradient_) {
       int M = M_[i];
       int N = N_[i];
       int K = K_[i];
-      REGISTER_TIMER_INFO("BwdDeConv3D", getName().c_str());
       Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
       const MatrixPtr &inMat = getInputValue(i);
       for (int n = 0; n < batchSize; ++n) {
@@ -182,7 +181,6 @@ void DeConv3DLayer::backward(const UpdateCallback &callback) {
           }
         }
       }
-      REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
       weights_[i]->getParameterPtr()->incUpdate(callback);
     }
   }

paddle/gserver/layers/GruCompute.cpp

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "GruCompute.h"
 #include "hl_recurrent_apply.cuh"
+#include "paddle/function/GruFunctor.h"
 #include "paddle/utils/Util.h"
 
 namespace paddle {
@@ -25,13 +26,13 @@ void GruCompute::init(LayerConfig &config) {
 
 template <>
 void GruCompute::forward<0>(hl_gru_value value, int frameSize, int batchSize) {
-  hl_cpu_gru_forward(hppl::forward::gru_resetOutput(),
-                     hppl::forward::gru_finalOutput(),
-                     value,
-                     frameSize,
-                     batchSize,
-                     activeNode_,
-                     activeGate_);
+  GruFunctor<DEVICE_TYPE_CPU, real>::compute(hppl::forward::gru_resetOutput(),
+                                             hppl::forward::gru_finalOutput(),
+                                             value,
+                                             frameSize,
+                                             batchSize,
+                                             activeNode_,
+                                             activeGate_);
 }
 
 template <>
@@ -39,14 +40,15 @@ void GruCompute::backward<0>(hl_gru_value value,
                              hl_gru_grad grad,
                              int frameSize,
                              int batchSize) {
-  hl_cpu_gru_backward(hppl::backward::gru_stateGrad(),
-                      hppl::backward::gru_resetGrad(),
-                      value,
-                      grad,
-                      frameSize,
-                      batchSize,
-                      activeNode_,
-                      activeGate_);
+  GruGradFunctor<DEVICE_TYPE_CPU, real>::compute(
+      hppl::backward::gru_stateGrad(),
+      hppl::backward::gru_resetGrad(),
+      value,
+      grad,
+      frameSize,
+      batchSize,
+      activeNode_,
+      activeGate_);
 }
 
 }  // namespace paddle

paddle/gserver/layers/SwitchOrderLayer.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "SwitchOrderLayer.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+REGISTER_LAYER(switch_order, SwitchOrderLayer);
+
+bool SwitchOrderLayer::init(const LayerMap& layerMap,
+                            const ParameterMap& parameterMap) {
+  /* Initialize the basic parent class */
+  Layer::init(layerMap, parameterMap);
+  auto& img_conf = config_.inputs(0).image_conf();
+  size_t inH =
+      img_conf.has_img_size_y() ? img_conf.img_size_y() : img_conf.img_size();
+  size_t inW = img_conf.img_size();
+  size_t inC = img_conf.channels();
+  inDims_ = TensorShape({0, inC, inH, inW});
+  outDims_ = TensorShape(4);
+
+  auto& reshape_conf = config_.reshape_conf();
+  for (int i = 0; i < reshape_conf.height_axis_size(); i++) {
+    heightAxis_.push_back(reshape_conf.height_axis(i));
+  }
+  for (int i = 0; i < reshape_conf.width_axis_size(); i++) {
+    widthAxis_.push_back(reshape_conf.width_axis(i));
+  }
+  createFunction(nchw2nhwc_, "NCHW2NHWC", FuncConfig());
+  createFunction(nhwc2nchw_, "NHWC2NCHW", FuncConfig());
+  return true;
+}
+
+void SwitchOrderLayer::setOutDims() {
+  outDims_.setDim(0, inDims_[0]);
+  outDims_.setDim(1, inDims_[2]);
+  outDims_.setDim(2, inDims_[3]);
+  outDims_.setDim(3, inDims_[1]);
+  reshapeHeight_ = 1;
+  for (size_t i = 0; i < heightAxis_.size(); i++) {
+    reshapeHeight_ *= outDims_[heightAxis_[i]];
+  }
+  output_.setFrameHeight(reshapeHeight_);
+  reshapeWidth_ = 1;
+  for (size_t i = 0; i < widthAxis_.size(); i++) {
+    reshapeWidth_ *= outDims_[widthAxis_[i]];
+  }
+  output_.setFrameWidth(reshapeWidth_);
+}
+
+void SwitchOrderLayer::setInDims() {
+  MatrixPtr input = inputLayers_[0]->getOutputValue();
+  size_t batchSize = input->getHeight();
+  inDims_.setDim(0, batchSize);
+
+  int h = inputLayers_[0]->getOutput().getFrameHeight();
+  if (h != 0) inDims_.setDim(2, h);
+  int w = inputLayers_[0]->getOutput().getFrameWidth();
+  if (w != 0) inDims_.setDim(3, w);
+  int totalCount = input->getElementCnt();
+  int channels = totalCount / (inDims_[0] * inDims_[2] * inDims_[3]);
+  if (channels != 0) inDims_.setDim(1, channels);
+}
+
+void SwitchOrderLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  setInDims();
+  setOutDims();
+  resetOutput(outDims_[0], outDims_[1] * outDims_[2] * outDims_[3]);
+  if (heightAxis_.size() > 0) {
+    getOutputValue()->reshape(reshapeHeight_, reshapeWidth_);
+    getOutputGrad()->reshape(reshapeHeight_, reshapeWidth_);
+  }
+
+  // switch NCHW to NHWC
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getInputValue(0), inDims_);
+  outputs.addArg(*getOutputValue(), outDims_);
+  nchw2nhwc_[0]->calc(inputs, outputs);
+  forwardActivation();
+}
+
+void SwitchOrderLayer::backward(const UpdateCallback& callback) {
+  (void)callback;
+  backwardActivation();
+
+  // switch NHWC to NCHW
+  BufferArgs inputs;
+  BufferArgs outputs;
+  inputs.addArg(*getOutputGrad(), outDims_);
+  outputs.addArg(*getInputGrad(0), inDims_, ADD_TO);
+  nhwc2nchw_[0]->calc(inputs, outputs);
+}
+}  // namespace paddle

paddle/operators/scatter_op.cu → paddle/gserver/layers/SwitchOrderLayer.h

@@ -12,9 +12,36 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#define EIGEN_USE_GPU
-#include "paddle/operators/scatter_op.h"
+#pragma once
 
-namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(scatter,
-                       ops::ScatterOpKernel<paddle::platform::GPUPlace, float>);
+#include "Layer.h"
+
+namespace paddle {
+
+/**
+ * \brief  This layer calculate softmax in image channel dimension.
+ */
+class SwitchOrderLayer : public Layer {
+public:
+  explicit SwitchOrderLayer(const LayerConfig& config) : Layer(config) {}
+
+  ~SwitchOrderLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+  void setInDims();
+  void setOutDims();
+
+protected:
+  std::vector<std::shared_ptr<FunctionBase>> nchw2nhwc_;
+  std::vector<std::shared_ptr<FunctionBase>> nhwc2nchw_;
+  TensorShape inDims_;
+  TensorShape outDims_;
+  std::vector<int> heightAxis_;
+  std::vector<int> widthAxis_;
+  size_t reshapeHeight_;
+  size_t reshapeWidth_;
+};
+}  // namespace paddle

paddle/gserver/tests/test_LayerGrad.cpp

@@ -2008,6 +2008,31 @@ TEST(Layer, CropLayer) {
   }
 }
 
+TEST(Layer, SwitchOrderLayer) {
+  TestConfig config;
+  // config input_0
+  config.inputDefs.push_back({INPUT_DATA, "layer_0", 1024, 0});
+  LayerInputConfig* input = config.layerConfig.add_inputs();
+  ImageConfig* img = input->mutable_image_conf();
+  img->set_channels(4);
+  img->set_img_size(16);
+  img->set_img_size_y(16);
+
+  ReshapeConfig* reshape = config.layerConfig.mutable_reshape_conf();
+  reshape->add_height_axis(0);
+  reshape->add_height_axis(1);
+  reshape->add_height_axis(2);
+  reshape->add_width_axis(3);
+
+  // config softmax layer
+  config.layerConfig.set_type("switch_order");
+  config.layerConfig.set_name("switchOrderLayer");
+
+  for (auto useGpu : {false, true}) {
+    testLayerGrad(config, "switch_order", 100, false, useGpu, true);
+  }
+}
+
 vector<real> randSampling(real range, int n) {
   CHECK_GE(range, n);
   vector<real> num(range);

paddle/math/MathFunctions.cpp

@@ -84,6 +84,7 @@ LAPACK_ROUTINE_EACH(DYNAMIC_LOAD_LAPACK_WRAP)
 
 namespace paddle {
 
+#ifndef PADDLE_USE_EIGEN_FOR_BLAS
 template <>
 void gemm<float>(const CBLAS_TRANSPOSE transA,
                  const CBLAS_TRANSPOSE transB,
@@ -143,6 +144,7 @@ void gemm<double>(const CBLAS_TRANSPOSE transA,
               C,
               ldc);
 }
+#endif
 
 template <>
 int getrf<float>(const CBLAS_ORDER order,
@@ -182,6 +184,7 @@ int getri<double>(const CBLAS_ORDER order,
   return dynload::PADDLE_DGETRI(order, N, A, lda, ipiv);
 }
 
+#ifndef PADDLE_USE_EIGEN_FOR_BLAS
 template <>
 void axpy<float>(const int n, const float alpha, const float* x, float* y) {
   cblas_saxpy(n, alpha, x, 1, y, 1);
@@ -201,6 +204,7 @@ template <>
 double dotProduct<double>(const int n, const double* x, const double* y) {
   return cblas_ddot(n, x, 1, y, 1);
 }
+#endif
 
 #if defined(PADDLE_USE_MKL) || defined(PADDLE_USE_MKLML)
 

paddle/math/MathFunctions.h

@@ -40,7 +40,14 @@ extern "C" {
 
 #ifndef LAPACK_FOUND
 extern "C" {
+#ifndef PADDLE_USE_EIGEN_FOR_BLAS
 #include <cblas.h>
+#else
+typedef enum CBLAS_ORDER {
+  CblasRowMajor = 101,
+  CblasColMajor = 102
+} CBLAS_ORDER;
+#endif
 int LAPACKE_sgetrf(
     int matrix_layout, int m, int n, float* a, int lda, int* ipiv);
 int LAPACKE_dgetrf(
@@ -56,6 +63,7 @@ int LAPACKE_dgetri(
 
 namespace paddle {
 
+#ifndef PADDLE_USE_EIGEN_FOR_BLAS
 template <class T>
 void gemm(const CBLAS_TRANSPOSE transA,
           const CBLAS_TRANSPOSE transB,
@@ -70,6 +78,7 @@ void gemm(const CBLAS_TRANSPOSE transA,
           const T beta,
           T* C,
           const int ldc);
+#endif
 
 template <class T>
 int getrf(const CBLAS_ORDER Order,
@@ -84,10 +93,21 @@ int getri(
     const CBLAS_ORDER Order, const int N, T* A, const int lda, const int* ipiv);
 
 template <class T>
-void axpy(const int n, const T alpha, const T* x, T* y);
+void axpy(const int n, const T alpha, const T* x, T* y) {
+  /// y = y + alpha * x
+  for (int i = 0; i < n; i++) {
+    y[i] = y[i] + alpha * x[i];
+  }
+}
 
 template <class T>
-T dotProduct(const int n, const T* x, const T* y);
+T dotProduct(const int n, const T* x, const T* y) {
+  T result = static_cast<T>(0);
+  for (int i = 0; i < n; i++) {
+    result += x[i] * y[i];
+  }
+  return result;
+}
 
 template <class T>
 void vExp(const int n, const T* a, T* r);

paddle/math/Matrix.cpp

@@ -28,6 +28,7 @@ limitations under the License. */
 #include "hl_top_k.h"
 #include "paddle/utils/Logging.h"
 
+#include "paddle/function/GemmFunctor.h"
 #include "paddle/utils/ThreadLocal.h"
 
 #include "SIMDFunctions.h"
@@ -2773,24 +2774,24 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
   CHECK(!isTransposed()) << "Not supported";
 
   size_t a_col, b_col, a_row, b_row;
-  CBLAS_TRANSPOSE a_trans, b_trans;
+  bool a_trans, b_trans;
   if (!a->isTransposed()) {
     a_col = a->getWidth();
     a_row = a->getHeight();
-    a_trans = CblasNoTrans;
+    a_trans = false;
   } else {
     a_col = a->getHeight();
     a_row = a->getWidth();
-    a_trans = CblasTrans;
+    a_trans = true;
   }
   if (!b->isTransposed()) {
     b_col = b->getWidth();
     b_row = b->getHeight();
-    b_trans = CblasNoTrans;
+    b_trans = false;
   } else {
     b_col = b->getHeight();
     b_row = b->getWidth();
-    b_trans = CblasTrans;
+    b_trans = true;
   }
 
   CHECK_EQ(a_col, b_row);
@@ -2807,7 +2808,7 @@ void CpuMatrix::mul(CpuMatrix* a, CpuMatrix* b, real scaleAB, real scaleT) {
   int lda = a->getStride();
   int ldb = b->getStride();
   int ldc = getStride();
-  gemm<real>(
+  BlasGemm<DEVICE_TYPE_CPU, real>::compute(
       a_trans, b_trans, M, N, K, scaleAB, A, lda, B, ldb, scaleT, C, ldc);
 }
 

paddle/math/Matrix.h

@@ -1616,6 +1616,10 @@ public:
 };
 
 class CpuMatrix : public Matrix {
+private:
+  MatrixPtr sftmaxSum_;
+  MatrixPtr sftmaxDot_;
+
 public:
   CpuMatrix(size_t height, size_t width, bool trans = false);
   CpuMatrix(real* data, size_t height, size_t width, bool trans = false)

paddle/operators/CMakeLists.txt

@@ -14,27 +14,31 @@ function(op_library TARGET)
     cmake_parse_arguments(op_library "${options}" "${oneValueArgs}"
             "${multiValueArgs}" ${ARGN})
 
-    foreach(src ${op_library_SRCS})
-        if (${src} MATCHES ".*\\.cu$")
-            list(APPEND cu_srcs ${src})
-        elseif(${src} MATCHES ".*\\.cc$")
-            list(APPEND cc_srcs ${src})
-        else()
-            message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+    list(LENGTH op_library_SRCS op_library_SRCS_len)
+    if (${op_library_SRCS_len} EQUAL 0)
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cc)
+            list(APPEND cc_srcs ${TARGET}.cc)
         endif()
-    endforeach()
+        if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu)
+            list(APPEND cu_srcs ${TARGET}.cu)
+        endif()
+    else()
+        foreach(src ${op_library_SRCS})
+            if (${src} MATCHES ".*\\.cu$")
+                list(APPEND cu_srcs ${src})
+            elseif(${src} MATCHES ".*\\.cc$")
+                list(APPEND cc_srcs ${src})
+            else()
+                message(FATAL_ERROR "${TARGET} Source file ${src} should only be .cc or .cu")
+            endif()
+        endforeach()
+    endif()
 
     list(LENGTH cc_srcs cc_srcs_len)
     if (${cc_srcs_len} EQUAL 0)
         message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file")
     endif()
 
-    list(LENGTH cu_srcs cu_srcs_len)
-    list(LENGTH op_library_DEPS dep_len)
-    if (${cu_srcs_len} EQUAL 0 AND ${dep_len} EQUAL 0)
-        message(WARNING "The op library ${TARGET} not support GPU!")
-    endif()
-
     if (WITH_GPU)
         nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS}
                 ${op_common_deps})
@@ -46,22 +50,22 @@ endfunction()
 
 add_subdirectory(math)
 
-list(REMOVE_ITEM GENERAL_OPS
-     net_op
-     minus_op
-     mul_op
-     recurrent_op
-     scale_op)
-
-op_library(net_op SRCS net_op.cc)
-op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)
-op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
+set(DEPS_OPS
+    identity_op
+    minus_op
+    mul_op
+    recurrent_op
+    scale_op)
+op_library(identity_op DEPS scale_op)
+op_library(minus_op DEPS scale_op)
+op_library(mul_op DEPS math_function)
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
   DEPS framework_proto tensor operator net_op)
-op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
+op_library(scale_op DEPS net_op)
 
+list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
-    op_library(${src} SRCS ${src}.cc ${src}.cu)
+    op_library(${src})
 endforeach()
 
 set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")

paddle/operators/add_op.cc

@@ -57,7 +57,6 @@ class AddOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker, add_two_grad, ops::AddOpGrad);
+REGISTER_OP(add, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad);
 
-REGISTER_OP_CPU_KERNEL(add_two,
-                       ops::AddKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(add, ops::AddKernel<paddle::platform::CPUPlace, float>);

paddle/operators/add_op.cu

@@ -12,10 +12,7 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#define EIGEN_USE_GPU
-#include "paddle/framework/op_registry.h"
 #include "paddle/operators/add_op.h"
 
 namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(add_two,
-                       ops::AddKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(add, ops::AddKernel<paddle::platform::GPUPlace, float>);

paddle/operators/cos_sim_op.h

@@ -23,6 +23,9 @@ using Tensor = framework::Tensor;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 
 template <typename Place, typename T>
 class CosSimKernel : public framework::OpKernel {
@@ -43,14 +46,14 @@ class CosSimKernel : public framework::OpKernel {
     auto new_dims = framework::make_ddim({dims[0], size / dims[0]});
     auto x = EigenMatrix<T>::From(*input_x, new_dims);
     auto y = EigenMatrix<T>::From(*input_y, new_dims);
-    auto z = EigenMatrix<T>::From(*output_z);
-    auto x_norm = EigenMatrix<T>::From(*output_x_norm);
-    auto y_norm = EigenMatrix<T>::From(*output_y_norm);
+    auto z = EigenVector<T>::Flatten(*output_z);
+    auto x_norm = EigenVector<T>::Flatten(*output_x_norm);
+    auto y_norm = EigenVector<T>::Flatten(*output_y_norm);
 
     auto place = context.GetEigenDevice<Place>();
-    auto xy = (x * y).sum(Eigen::array<int, 1>({1}));
-    x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({1})).sqrt();
-    y_norm.device(place) = y.square().sum(Eigen::array<int, 1>({1})).sqrt();
+    auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
+    x_norm.device(place) = x.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
+    y_norm.device(place) = y.square().sum(Eigen::array<int, 1>({{1}})).sqrt();
     z.device(place) = xy / x_norm / y_norm;
   }
 };

paddle/operators/gaussian_random_op.cc

@@ -19,12 +19,12 @@ template <typename T>
 class CPUGaussianRandomKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    float mean = context.GetAttr<float>("mean");
-    float std = context.GetAttr<float>("std");
+    float mean = context.Attr<float>("mean");
+    float std = context.Attr<float>("std");
     auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
 
-    unsigned int seed = static_cast<unsigned int>(context.GetAttr<int>("seed"));
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
     std::minstd_rand engine;
     if (seed == 0) {
       seed = std::random_device()();
@@ -45,7 +45,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext& context) const override {
     auto* tensor = context.Output<framework::Tensor>("Out");
-    auto dims = GetAttr<std::vector<int>>("dims");
+    auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());
     for (auto dim : dims) {

paddle/operators/gaussian_random_op.cu

@@ -42,13 +42,13 @@ class GPUGaussianRandomKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
-    unsigned int seed = static_cast<unsigned int>(context.GetAttr<int>("seed"));
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
     if (seed == 0) {
       std::random_device rd;
       seed = rd();
     }
-    T mean = static_cast<T>(context.GetAttr<float>("mean"));
-    T std = static_cast<T>(context.GetAttr<float>("std"));
+    T mean = static_cast<T>(context.Attr<float>("mean"));
+    T std = static_cast<T>(context.Attr<float>("std"));
     thrust::counting_iterator<unsigned int> index_sequence_begin(0);
     ssize_t N = framework::product(tensor->dims());
     thrust::transform(index_sequence_begin, index_sequence_begin + N,

paddle/operators/identity_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/scale_op.h"
+
+namespace paddle {
+namespace operators {
+
+// The identity operator is an alias of the scale operator. This is also an
+// example for creating an alias for an existing operator.
+template <typename AttrType>
+class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  IdentityOpMaker(framework::OpProto *proto,
+                  framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input tensor of identity operator.");
+    AddOutput("Out", "The output tensor of identity operator.");
+    AddComment(R"DOC(
+The identity operator is an alias of the scale operator
+with the attribute scale fixed to 1.0.
+)DOC");
+  }
+};
+
+template <typename AttrType>
+class IdentityOp : public NetOp {
+ public:
+  IdentityOp(const std::string &type, const framework::VariableNameMap &inputs,
+             const framework::VariableNameMap &outputs,
+             const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    AppendOp(framework::OpRegistry::CreateOp(
+        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
+        {{"scale", static_cast<AttrType>(1)}}));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp<float>,
+                             ops::IdentityOpMaker<float>);

paddle/operators/math/CMakeLists.txt

 
 if(WITH_GPU)
-    nv_library(math_function SRCS math_function.cc math_function.cu DEPS cblas device_context)
+    nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc 
+    im2col.cu DEPS cblas device_context)
 else()
-    cc_library(math_function SRCS math_function.cc DEPS cblas device_context)
+    cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context)
 endif()
 
 nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor)
+cc_test(im2col_test SRCS im2col_test.cc DEPS math_function tensor)

paddle/operators/math/im2col.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/math/im2col.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [input_channels, filter_height, filter_width, output_height, output_width]
+ */
+template <class T>
+class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                    platform::CPUPlace, T> {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[1];
+    int filter_width = col.dims()[2];
+    int output_height = col.dims()[3];
+    int output_width = col.dims()[4];
+    int channels_col = input_channels * filter_height * filter_width;
+
+    const T* im_data = im.data<T>();
+    T* col_data = col.data<T>();
+
+    for (int c = 0; c < channels_col; ++c) {
+      int w_offset = c % filter_width;
+      int h_offset = (c / filter_width) % filter_height;
+      int c_im = c / filter_width / filter_height;
+      for (int h = 0; h < output_height; ++h) {
+        for (int w = 0; w < output_width; ++w) {
+          int im_row_idx = h * stride_height + h_offset;
+          int im_col_idx = w * stride_width + w_offset;
+          if ((im_row_idx - padding_height) < 0 ||
+              (im_row_idx - padding_height) >= input_height ||
+              (im_col_idx - padding_width) < 0 ||
+              (im_col_idx - padding_width) >= input_width) {
+            col_data[(c * output_height + h) * output_width + w] = T(0);
+          } else {
+            im_row_idx += c_im * input_height - padding_height;
+            im_col_idx -= padding_width;
+            col_data[(c * output_height + h) * output_width + w] =
+                im_data[im_row_idx * input_width + im_col_idx];
+          }
+        }
+      }
+    }
+  }
+};
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [input_channels, filter_height, filter_width, output_height, output_width]
+ */
+template <class T>
+class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                    platform::CPUPlace, T> {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[1];
+    int filter_width = col.dims()[2];
+    int output_height = col.dims()[3];
+    int output_width = col.dims()[4];
+    int channels_col = input_channels * filter_height * filter_width;
+
+    T* im_data = im.data<T>();
+    const T* col_data = col.data<T>();
+
+    for (int c = 0; c < channels_col; ++c) {
+      int w_offset = c % filter_width;
+      int h_offset = (c / filter_width) % filter_height;
+      int c_im = c / filter_width / filter_height;
+      for (int h = 0; h < output_height; ++h) {
+        for (int w = 0; w < output_width; ++w) {
+          int im_row_idx = h * stride_height + h_offset;
+          int im_col_idx = w * stride_width + w_offset;
+          if ((im_row_idx - padding_height) >= 0 &&
+              (im_row_idx - padding_height) < input_height &&
+              (im_col_idx - padding_width) >= 0 &&
+              (im_col_idx - padding_width) < input_width) {
+            im_row_idx += c_im * input_height - padding_height;
+            im_col_idx -= padding_width;
+            im_data[im_row_idx * input_width + im_col_idx] +=
+                col_data[(c * output_height + h) * output_width + w];
+          }
+        }
+      }
+    }
+  }
+};
+
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::CPUPlace, float>;
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::CPUPlace, double>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::CPUPlace, float>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::CPUPlace, double>;
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [output_height, output_width, input_channels, filter_height, filter_width]
+ */
+template <class T>
+class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                    platform::CPUPlace, T> {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[3];
+    int filter_width = col.dims()[4];
+    int output_height = col.dims()[0];
+    int output_width = col.dims()[1];
+
+    const T* im_data = im.data<T>();
+    T* col_data = col.data<T>();
+
+    for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) {
+      for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) {
+        for (int channel = 0; channel < input_channels; ++channel) {
+          for (int filter_row_idx = 0; filter_row_idx < filter_height;
+               ++filter_row_idx) {
+            for (int filter_col_idx = 0; filter_col_idx < filter_width;
+                 ++filter_col_idx) {
+              int im_row_offset =
+                  col_row_idx * stride_height + filter_row_idx - padding_height;
+              int im_col_offset =
+                  col_col_idx * stride_width + filter_col_idx - padding_width;
+              int col_offset = (((col_row_idx * output_width + col_col_idx) *
+                                     input_channels +
+                                 channel) *
+                                    filter_height +
+                                filter_row_idx) *
+                                   filter_width +
+                               filter_col_idx;
+              if (im_row_offset < 0 || im_row_offset >= input_height ||
+                  im_col_offset < 0 || im_col_offset >= input_width) {
+                col_data[col_offset] = T(0);
+              } else {
+                int im_offset =
+                    (channel * input_height + im_row_offset) * input_width +
+                    im_col_offset;
+                col_data[col_offset] = im_data[im_offset];
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+};
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [output_height, output_width, input_channels, filter_height, filter_width]
+ */
+template <class T>
+class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                    platform::CPUPlace, T> {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[3];
+    int filter_width = col.dims()[4];
+    int output_height = col.dims()[0];
+    int output_width = col.dims()[1];
+
+    T* im_data = im.data<T>();
+    const T* col_data = col.data<T>();
+
+    for (int col_row_idx = 0; col_row_idx < output_height; ++col_row_idx) {
+      for (int col_col_idx = 0; col_col_idx < output_width; ++col_col_idx) {
+        for (int channel = 0; channel < input_channels; ++channel) {
+          for (int filter_row_idx = 0; filter_row_idx < filter_height;
+               ++filter_row_idx) {
+            for (int filter_col_idx = 0; filter_col_idx < filter_width;
+                 ++filter_col_idx) {
+              int im_row_offset =
+                  col_row_idx * stride_height + filter_row_idx - padding_height;
+              int im_col_offset =
+                  col_col_idx * stride_width + filter_col_idx - padding_width;
+              int col_offset = (((col_row_idx * output_width + col_col_idx) *
+                                     input_channels +
+                                 channel) *
+                                    filter_height +
+                                filter_row_idx) *
+                                   filter_width +
+                               filter_col_idx;
+              if (im_row_offset >= 0 && im_row_offset < input_height &&
+                  im_col_offset >= 0 && im_col_offset < input_width) {
+                int im_offset =
+                    (channel * input_height + im_row_offset) * input_width +
+                    im_col_offset;
+                im_data[im_offset] += col_data[col_offset];
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+};
+
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::CPUPlace, float>;
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::CPUPlace, double>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::CPUPlace, float>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::CPUPlace, double>;
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/im2col.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/math/im2col.h"
+#include "paddle/platform/cuda_helper.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <class T>
+__global__ void im2col(const T* data_im, int num_outs, int height, int width,
+                       int filter_height, int filter_width, int stride_height,
+                       int stride_width, int padding_height, int padding_width,
+                       int output_height, int output_width, T* data_col) {
+  int index = (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
+  if (index < num_outs) {
+    int w_out = index % output_width;
+    index /= output_width;
+    int h_out = index % output_height;
+    int channel_in = index / output_height;
+    int channel_out = channel_in * filter_height * filter_width;
+    int h_in = h_out * stride_height;
+    int w_in = w_out * stride_width;
+
+    data_col += (channel_out * output_height + h_out) * output_width + w_out;
+    for (int i = 0; i < filter_height; ++i) {
+      for (int j = 0; j < filter_width; ++j) {
+        int rIdx = int(h_in + i);
+        int cIdx = int(w_in + j);
+        if ((rIdx - (int)padding_height) >= (int)height ||
+            (rIdx - (int)padding_height) < 0 ||
+            (cIdx - (int)padding_width) >= (int)width ||
+            (cIdx - (int)padding_width) < 0) {
+          *data_col = 0;
+        } else {
+          rIdx = rIdx + channel_in * height - padding_height;
+          cIdx = cIdx - padding_width;
+          *data_col = data_im[rIdx * width + cIdx];
+        }
+        data_col += output_height * output_width;
+      }
+    }
+  }
+}
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [input_channels, filter_height, filter_width, output_height, output_width]
+ */
+template <class T>
+class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                    platform::GPUPlace, T> {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[1];
+    int filter_width = col.dims()[2];
+    int output_height = col.dims()[3];
+    int output_width = col.dims()[4];
+
+    int num_outputs = input_channels * output_height * output_width;
+    int blocks = (num_outputs + 1024 - 1) / 1024;
+    int block_x = 512;
+    int block_y = (blocks + 512 - 1) / 512;
+    dim3 threads(1024, 1);
+    dim3 grid(block_x, block_y);
+    im2col<T><<<
+        grid, threads, 0,
+        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+        im.data<T>(), num_outputs, input_height, input_width, filter_height,
+        filter_width, stride_height, stride_width, padding_height,
+        padding_width, output_height, output_width, col.data<T>());
+  }
+};
+
+template <class T>
+__global__ void col2im(size_t n, const T* data_col, size_t height, size_t width,
+                       size_t channels, size_t filter_height,
+                       size_t filter_width, size_t stride_height,
+                       size_t stride_width, size_t padding_height,
+                       size_t padding_width, size_t output_height,
+                       size_t output_width, T* data_im) {
+  size_t index =
+      (blockIdx.x * gridDim.y + blockIdx.y) * blockDim.x + threadIdx.x;
+  if (index < n) {
+    T val = 0;
+    int w = int(index % width);
+    int h = int((index / width) % height);
+    int c = int(index / (width * height));
+    if ((w - (int)padding_width) >= 0 &&
+        (w - (int)padding_width) < (width - 2 * padding_width) &&
+        (h - (int)padding_height) >= 0 &&
+        (h - padding_height) < (height - 2 * padding_height)) {
+      // compute the start and end of the output
+      int w_col_start = (w < (int)filter_width)
+                            ? 0
+                            : (w - int(filter_width)) / (int)stride_width + 1;
+      int w_col_end =
+          min((int)(w / (int)stride_width + 1), (int)(output_width));
+      int h_col_start = (h < (int)filter_height)
+                            ? 0
+                            : (h - (int)filter_height) / (int)stride_height + 1;
+      int h_col_end = min(int(h / stride_height + 1), int(output_height));
+      for (int h_col = h_col_start; h_col < h_col_end; ++h_col) {
+        for (int w_col = w_col_start; w_col < w_col_end; ++w_col) {
+          // the col location: [c * width * height + h_out, w_out]
+          int c_col = int(c * filter_height * filter_width) +
+                      (h - h_col * (int)stride_height) * (int)filter_width +
+                      (w - w_col * (int)stride_width);
+          val +=
+              data_col[(c_col * output_height + h_col) * output_width + w_col];
+        }
+      }
+      h -= padding_height;
+      w -= padding_width;
+      data_im[c * ((width - 2 * padding_width) *
+                   (height - 2 * padding_height)) +
+              h * (width - 2 * padding_width) + w] += val;
+    }
+  }
+}
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [input_channels, filter_height, filter_width, output_height, output_width]
+ */
+template <class T>
+class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                    platform::GPUPlace, T> {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[1];
+    int filter_width = col.dims()[2];
+    int output_height = col.dims()[3];
+    int output_width = col.dims()[4];
+
+    size_t num_kernels = input_channels * (input_height + 2 * padding_height) *
+                         (input_width + 2 * padding_width);
+
+    size_t blocks = (num_kernels + 1024 - 1) / 1024;
+    size_t block_x = 512;
+    size_t block_y = (blocks + 512 - 1) / 512;
+    dim3 threads(1024, 1);
+    dim3 grid(block_x, block_y);
+
+    // To avoid involving atomic operations, we will launch one kernel per
+    // bottom dimension, and then in the kernel add up the top dimensions.
+    col2im<T><<<
+        grid, threads, 0,
+        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+        num_kernels, col.data<T>(), input_height + 2 * padding_height,
+        input_width + 2 * padding_width, input_channels, filter_height,
+        filter_width, stride_height, stride_width, padding_height,
+        padding_width, output_height, output_width, im.data<T>());
+  }
+};
+
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::GPUPlace, float>;
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::GPUPlace, double>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::GPUPlace, float>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kCFO,
+                             platform::GPUPlace, double>;
+
+template <class T>
+__global__ void im2colOCF(const T* im_data, T* col_data, int input_channels,
+                          int input_height, int input_width, int filter_height,
+                          int filter_width, int stride_height, int stride_width,
+                          int padding_height, int padding_width,
+                          int output_height, int output_width) {
+  int swid = blockIdx.x;
+  int shid = blockIdx.y;
+  for (int channelid = threadIdx.z; channelid < input_channels;
+       channelid += blockDim.z) {
+    for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) {
+      for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) {
+        int width_offset = idx + swid * stride_width - padding_width;
+        int height_offset = idy + shid * stride_height - padding_height;
+        int im_offset = width_offset + height_offset * input_width +
+                        channelid * input_height * input_width;
+
+        int col_offset = idx + idy * filter_width +
+                         channelid * filter_height * filter_width +
+                         (shid * output_width + swid) *
+                             (input_channels * filter_height * filter_width);
+
+        if (height_offset >= input_height || height_offset < 0 ||
+            width_offset >= input_width || width_offset < 0) {
+          col_data[col_offset] = T(0);
+        } else {
+          col_data[col_offset] = im_data[im_offset];
+        }
+      }
+    }
+  }
+}
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [output_height, output_width, input_channels, filter_height, filter_width]
+ */
+template <class T>
+class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                    platform::GPUPlace, T> {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[3];
+    int filter_width = col.dims()[4];
+    int output_height = col.dims()[0];
+    int output_width = col.dims()[1];
+
+    int block_dim_x = 0;
+    int block_dim_y = 0;
+    if (filter_height <= 4 && filter_width <= 4) {
+      block_dim_x = 4;
+      block_dim_y = 4;
+    } else if (filter_height <= 8 && filter_width <= 8) {
+      block_dim_x = 8;
+      block_dim_y = 8;
+    } else if (filter_height <= 16 && filter_width <= 16) {
+      block_dim_x = 16;
+      block_dim_y = 16;
+    } else {
+      block_dim_x = 32;
+      block_dim_y = 32;
+    }
+
+    int block_dim_z = 1024 / block_dim_x / block_dim_y;
+    dim3 threads(block_dim_x, block_dim_y,
+                 std::min(block_dim_z, input_channels));
+    dim3 grid(output_width, output_height);
+    im2colOCF<T><<<
+        grid, threads, 0,
+        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+        im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
+        filter_height, filter_width, stride_height, stride_width,
+        padding_height, padding_width, output_height, output_width);
+  }
+};
+
+template <class T>
+__global__ void col2imOCF(T* im_data, const T* col_data, int input_channels,
+                          int input_height, int input_width, int filter_height,
+                          int filter_width, int stride_height, int stride_width,
+                          int padding_height, int padding_width,
+                          int output_height, int output_width) {
+  int swid = blockIdx.x;
+  int shid = blockIdx.y;
+  for (int channelid = threadIdx.z; channelid < input_channels;
+       channelid += blockDim.z) {
+    for (int idy = threadIdx.y; idy < filter_height; idy += blockDim.y) {
+      for (int idx = threadIdx.x; idx < filter_width; idx += blockDim.x) {
+        int width_offset = idx + swid * stride_width - padding_width;
+        int height_offset = idy + shid * stride_height - padding_height;
+        int im_offset = width_offset + height_offset * input_width +
+                        channelid * input_height * input_width;
+
+        int col_offset = idx + idy * filter_width +
+                         channelid * filter_height * filter_width +
+                         (shid * output_width + swid) *
+                             (input_channels * filter_height * filter_width);
+
+        if (height_offset >= 0 && height_offset < input_height &&
+            width_offset >= 0 && width_offset < input_width) {
+          paddle::platform::CudaAtomicAdd(im_data + im_offset,
+                                          col_data[col_offset]);
+        }
+      }
+    }
+  }
+}
+
+/*
+ * im = [input_channels, input_height, input_width]
+ * col =
+ *   [output_height, output_width, input_channels, filter_height, filter_width]
+ */
+template <class T>
+class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                    platform::GPUPlace, T> {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context) {
+    PADDLE_ENFORCE(im.dims().size() == 3);
+    PADDLE_ENFORCE(col.dims().size() == 5);
+    int input_channels = im.dims()[0];
+    int input_height = im.dims()[1];
+    int input_width = im.dims()[2];
+    int filter_height = col.dims()[3];
+    int filter_width = col.dims()[4];
+    int output_height = col.dims()[0];
+    int output_width = col.dims()[1];
+
+    int block_dim_x = 0;
+    int block_dim_y = 0;
+    if (filter_height <= 4 && filter_width <= 4) {
+      block_dim_x = 4;
+      block_dim_y = 4;
+    } else if (filter_height <= 8 && filter_width <= 8) {
+      block_dim_x = 8;
+      block_dim_y = 8;
+    } else if (filter_height <= 16 && filter_width <= 16) {
+      block_dim_x = 16;
+      block_dim_y = 16;
+    } else {
+      block_dim_x = 32;
+      block_dim_y = 32;
+    }
+
+    int block_dim_z = 1024 / block_dim_x / block_dim_y;
+    dim3 threads(block_dim_x, block_dim_y,
+                 std::min(block_dim_z, input_channels));
+    dim3 grid(output_width, output_height);
+    col2imOCF<T><<<
+        grid, threads, 0,
+        reinterpret_cast<platform::CUDADeviceContext*>(context)->stream()>>>(
+        im.data<T>(), col.data<T>(), input_channels, input_height, input_width,
+        filter_height, filter_width, stride_height, stride_width,
+        padding_height, padding_width, output_height, output_width);
+  }
+};
+
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::GPUPlace, float>;
+template class Im2ColFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::GPUPlace, double>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::GPUPlace, float>;
+template class Col2ImFunctor<paddle::operators::math::ColFormat::kOCF,
+                             platform::GPUPlace, double>;
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/im2col.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#pragma once
+
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+/* The storage format of the coldata in the Im2ColFunctor and Col2ImFunctor. */
+enum class ColFormat { kCFO = 0, kOCF = 1 };
+
+/*
+ * \brief Converts the image data of three dimensions(CHW) into a colData of
+ *        five dimensions in the Im2ColFunctor calculation,
+ *        And in the Col2ImFunctor calculation, it is reversed.
+ *
+ * \param imData   Image data.
+ * \param imShape  The shape of imData,
+ *                 [input_channels, input_height, input_width].
+ * \param colData  Column data.
+ * \param colShape The shape of colData.
+ *
+ * If the template argument Format is kCFO, the shape of colData is:
+ * [input_channels, filter_height, filter_width, output_height, output_width]
+ * So, it is easy to reshape into a convolution matrix for convolution
+ * calculation based on matrix multiplication.
+ * The shape of convolution matrix is [height, width], where the height is equal
+ * input_channels * filter_height * filter_width, and the width is equal
+ * output_height * output_width.
+ *
+ * Reshape:
+ *     shape of colData           shape of convolution matrix
+ *     [input_channels,
+ *      filter_height,
+ *      filter_width,      ======>      [height, width]
+ *      output_height,
+ *      output_width]
+ *
+ * If the template argument Format is kOCF, the shape of colData is:
+ * [output_height, output_width, input_channels, filter_height, filter_width]
+ * So, it is easy to reshape into a sequence matrix for rnn calculation.
+ * The shape of sequence matrix is [seq_length, step_size], where the seq_length
+ * is equal output_height * output_width, and the step_size is equal
+ * input_channels * filter_height * filter_width.
+ *
+ * Reshape:
+ *     shape of colData             shape of sequence matrix
+ *     [output_height,
+ *      output_width,
+ *      input_channels,    ======>    [seqLength, stepSize]
+ *      filter_height,
+ *      filter_width]
+ *
+ * \note The caller needs to ensure that imShape.inputChannels is equal to
+ *       colShape.inputChannels.
+ */
+template <ColFormat Format, typename Place, typename T>
+class Im2ColFunctor {
+ public:
+  void operator()(const framework::Tensor& im, framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context);
+};
+
+template <ColFormat Format, typename Place, typename T>
+class Col2ImFunctor {
+ public:
+  void operator()(framework::Tensor& im, const framework::Tensor& col,
+                  int stride_height, int stride_width, int padding_height,
+                  int padding_width, platform::DeviceContext* context);
+};
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/mul_op.cc

@@ -25,18 +25,27 @@ class MulOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto dim0 = ctx.Input<Tensor>("X")->dims();
-    auto dim1 = ctx.Input<Tensor>("Y")->dims();
-    PADDLE_ENFORCE_EQ(dim0.size(), 2,
-                      "input X(%s) should be a tensor with 2 dims, a matrix",
-                      ctx.op().Input("X"));
-    PADDLE_ENFORCE_EQ(dim1.size(), 2,
-                      "input Y(%s) should be a tensor with 2 dims, a matrix",
-                      ctx.op().Input("Y"));
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    int x_num_col_dims = Attr<int>("x_num_col_dims");
+    int y_num_col_dims = Attr<int>("y_num_col_dims");
+
+    PADDLE_ENFORCE(x_dims.size() > x_num_col_dims,
+                   "The rank of input tensor X(%s) should be larger than "
+                   "`mul_op`'s `x_num_col_dims`.",
+                   ctx.op().Input("X"));
+    PADDLE_ENFORCE(y_dims.size() > y_num_col_dims,
+                   "The rank of input tensor Y(%s) should be larger than "
+                   "`mul_op`'s `y_num_col_dims`.",
+                   ctx.op().Input("Y"));
+
+    auto x_mat_dims = framework::flatten_to_2d(x_dims, x_num_col_dims);
+    auto y_mat_dims = framework::flatten_to_2d(y_dims, y_num_col_dims);
+
     PADDLE_ENFORCE_EQ(
-        dim0[1], dim1[0],
+        x_mat_dims[1], y_mat_dims[0],
         "First matrix's width must be equal with second matrix's height.");
-    ctx.Output<Tensor>("Out")->Resize({dim0[0], dim1[1]});
+    ctx.Output<Tensor>("Out")->Resize({x_mat_dims[0], y_mat_dims[1]});
   }
 };
 
@@ -47,6 +56,23 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "The first input of mul op");
     AddInput("Y", "The second input of mul op");
     AddOutput("Out", "The output of mul op");
+    AddAttr<int>(
+        "x_num_col_dims",
+        R"DOC(mul_op can take tensors with more than two dimensions as input `X`, 
+            in that case, tensors will be reshaped to a matrix. The matrix's first 
+            dimension(column length) will be the product of tensor's last 
+            `num_col_dims` dimensions, and the matrix's second dimension(row length)
+            will be the product of tensor's first `rank - num_col_dims` dimensions.
+        )DOC")
+        .SetDefault(1)
+        .EqualGreaterThan(1);
+    AddAttr<int>(
+        "y_num_col_dims",
+        R"DOC(mul_op can take tensors with more than two dimensions as input `Y`,
+             in that case, tensors will be reshaped to a matrix. Just like input `X`.
+        )DOC")
+        .SetDefault(1)
+        .EqualGreaterThan(1);
     AddComment(R"DOC(
 Two Element Mul Operator.
 
@@ -70,10 +96,20 @@ class MulOpGrad : public framework::OperatorWithKernel {
     auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
     auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
     auto *y_grad = ctx.Output<Tensor>(framework::GradVarName("Y"));
-    PADDLE_ENFORCE(x_dims[0] == out_dims[0],
-                   "Out@GRAD M X N must equal to X dims 0, M ");
-    PADDLE_ENFORCE(y_dims[1] == out_dims[1],
-                   "Out@GRAD M X N must equal to Y dims 1, N ");
+
+    auto x_mat_dims =
+        framework::flatten_to_2d(x_dims, Attr<int>("x_num_col_dims"));
+    auto y_mat_dims =
+        framework::flatten_to_2d(y_dims, Attr<int>("y_num_col_dims"));
+
+    PADDLE_ENFORCE_EQ(
+        x_mat_dims[0], out_dims[0],
+        "The first dimension of Out@GRAD must equal to the first dimension of "
+        "the first operand.");
+    PADDLE_ENFORCE_EQ(
+        y_mat_dims[1], out_dims[1],
+        "The second dimension of Out@GRAD must equal to the second "
+        "dimension of the second operand.");
 
     if (x_grad) x_grad->Resize(x_dims);
     if (y_grad) y_grad->Resize(y_dims);

paddle/operators/rnn/recurrent_op_utils.cc

@@ -109,7 +109,7 @@ void InitArgument(const ArgumentName& name, Argument* arg,
   arg->step_scopes = op.Output(name.step_scopes);
 
   auto inlinks = op.Inputs(name.inlinks);
-  auto inlink_alias = op.GetAttr<std::vector<std::string>>(name.inlink_alias);
+  auto inlink_alias = op.Attr<std::vector<std::string>>(name.inlink_alias);
   PADDLE_ENFORCE(inlinks.size() == inlink_alias.size(),
                  "the size of inlinks and inlink_alias don't match:%d,%d",
                  inlinks.size(), inlink_alias.size());
@@ -121,7 +121,7 @@ void InitArgument(const ArgumentName& name, Argument* arg,
   }
 
   auto outlinks = op.Outputs(name.outlinks);
-  auto outlink_alias = op.GetAttr<std::vector<std::string>>(name.outlink_alias);
+  auto outlink_alias = op.Attr<std::vector<std::string>>(name.outlink_alias);
   PADDLE_ENFORCE(outlinks.size() == outlink_alias.size(),
                  "the size of outlinks and outlink_alias don't match:%d,%d",
                  outlinks.size(), outlink_alias.size());
@@ -135,8 +135,8 @@ void InitArgument(const ArgumentName& name, Argument* arg,
   auto boot_memories = op.Inputs(name.boot_memories);
 
   // attributes
-  auto memories = op.GetAttr<std::vector<std::string>>(name.memories);
-  auto pre_memories = op.GetAttr<std::vector<std::string>>(name.pre_memories);
+  auto memories = op.Attr<std::vector<std::string>>(name.memories);
+  auto pre_memories = op.Attr<std::vector<std::string>>(name.pre_memories);
 
   PADDLE_ENFORCE(memories.size() == boot_memories.size(),
                  "the size of memories, boot_memories don't match:%d,%d",

paddle/operators/scale_op.h

@@ -27,7 +27,7 @@ class ScaleKernel : public framework::OpKernel {
     auto* in = context.Input<framework::Tensor>("X");
     tensor->mutable_data<T>(in->place());
 
-    auto scale = static_cast<T>(context.GetAttr<AttrType>("scale"));
+    auto scale = static_cast<T>(context.Attr<AttrType>("scale"));
 
     auto eigen_out = framework::EigenVector<T>::Flatten(*tensor);
     auto eigen_in = framework::EigenVector<T>::Flatten(*in);