Merge remote-tracking branch 'origin/develop' into impl

.gitignore

@@ -22,7 +22,9 @@ cmake-build-*
 
 # generated while compiling
 python/paddle/v2/framework/core.so
+paddle/pybind/pybind.h
 CMakeFiles
 cmake_install.cmake
 paddle/.timestamp
 python/paddlepaddle.egg-info/
+paddle/pybind/pybind.h

.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 +65,11 @@ 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")
+    elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
+        # TODO: support glog for Android api 16 ~ 19 in the future
+        message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
     endif()
 
     set(WITH_GPU OFF CACHE STRING

Dockerfile.android

@@ -4,9 +4,16 @@ 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
+ARG ANDROID_API
+
+ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
+ENV ANDROID_API=${ANDROID_API:-21}
+
 ENV HOME=/root \
     ANDROID_NDK_HOME=/opt/android-ndk-linux \
-    ANDROID_STANDALONE_TOOLCHAIN=/opt/android-toolchain-gcc
+    ANDROID_TOOLCHAINS_DIR=/opt/toolchains
 
 RUN apt-get update && \
     apt-get install -y \
@@ -15,12 +22,11 @@ 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
@@ -37,13 +43,12 @@ RUN pip install --upgrade pip && \
     pip install pre-commit
 
 # Android NDK
-RUN mkdir /opt/android-ndk-tmp && \
+RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
+    mkdir -p /opt/android-ndk-tmp && \
     cd /opt/android-ndk-tmp && \
     wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
     unzip -q android-ndk-r14b-linux-x86_64.zip && \
     mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
-    ${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-21 --install-dir=${ANDROID_STANDALONE_TOOLCHAIN} && \
-    rm -rf /opt/android-ndk-tmp && \
-    rm -rf ${ANDROID_NDK_HOME}
+    rm -rf /opt/android-ndk-tmp
 
 CMD ["bash", "/paddle/paddle/scripts/docker/build_android.sh"]

cmake/cpplint.cmake

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

cmake/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 ${ANDROID_STANDALONE_TOOLCHAIN})
+    ELSE(ANDROID_NDK)
+        # TODO: use android ndk
+    ENDIF()
+
     IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
         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()
-        ENDIF()
-        IF(ANDROID_ABI STREQUAL "arm64-v8a")
+    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/gflags.cmake

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

cmake/external/glog.cmake

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

cmake/external/openblas.cmake

@@ -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})
@@ -69,6 +73,26 @@ IF(NOT ${CBLAS_FOUND})
         UPDATE_COMMAND      ""
         CONFIGURE_COMMAND   ""
     )
+
+    IF(WITH_C_API)
+        INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
+        # Because libopenblas.a is a symbolic link of another library, thus need to
+        # install the whole directory.
+        IF(ANDROID)
+            SET(TMP_INSTALL_DIR third_party/openblas/lib/${ANDROID_ABI})
+        ELSE()
+            SET(TMP_INSTALL_DIR third_party/openblas/lib)
+        ENDIF()
+        INSTALL(CODE "execute_process(
+            COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
+                    destination ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
+            )"
+        )
+        INSTALL(CODE "MESSAGE(STATUS \"Installing: \"
+                \"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
+            )"
+        )
+    ENDIF()
 ENDIF(NOT ${CBLAS_FOUND})
 
 MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")

cmake/external/protobuf.cmake

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

cmake/external/zlib.cmake

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

doc/design/block.md

+# Design Doc: Block and Scope
+
+## The Representation of Computation
+
+Both deep learning systems and programming languages help users describe computation procedures.  These systems use various representations of computation:
+
+- Caffe, Torch, and Paddle: sequences of layers.
+- TensorFlow, Caffe2, Mxnet: graphs of operators.
+- PaddlePaddle: nested blocks, like C++ and Java programs.
+
+## Block in Programming Languages and Deep Learning
+
+In programming languages, a block is a pair of curly braces that includes local variables definitions and a sequence of instructions, or operators.
+
+Blocks work with control flow structures like `if`, `else`, and `for`, which have equivalents in deep learning:
+
+| programming languages | PaddlePaddle          |
+|-----------------------|-----------------------|
+| for, while loop       | RNN, WhileOp          |
+| if, if-else, switch   | IfElseOp, SwitchOp    |
+| sequential execution  | a sequence of layers  |
+
+A key difference is that a C++ program describes a one pass computation, whereas a deep learning program describes both the forward and backward passes.
+
+## Stack Frames and the Scope Hierarchy
+
+The existence of the backward makes the execution of a block of traditional programs and PaddlePaddle different to each other:
+
+| programming languages | PaddlePaddle                  |
+|-----------------------|-------------------------------|
+| stack                 | scope hierarchy               |
+| stack frame           | scope                         |
+| push at entering block| push at entering block        |
+| pop at leaving block  | destroy at minibatch completes|
+
+1. In traditional programs:
+
+   - When the execution enters the left curly brace of a block, the runtime pushes a frame into the stack, where it realizes local variables.
+   - After the execution leaves the right curly brace, the runtime pops the frame.
+   - The maximum number of frames in the stack is the maximum depth of nested blocks.
+
+1. In PaddlePaddle
+
+   - When the execution enters a block, PaddlePaddle adds a new scope, where it realizes variables.
+   - PaddlePaddle doesn't pop a scope after the execution of the block because variables therein are to be used by the backward pass.  So it has a stack forest known as a *scope hierarchy*.
+   - The height of the highest tree is the maximum depth of nested blocks.
+   - After the process of a minibatch, PaddlePaddle destroys the scope hierarchy.
+
+## Use Blocks in C++ and PaddlePaddle Programs
+
+Let us consolidate the discussion by presenting some examples.
+
+### Blocks with `if-else` and `IfElseOp`
+
+The following C++ programs shows how blocks are used with the `if-else` structure:
+
+```c++
+int x = 10;
+int y = 20;
+int out;
+bool cond = false;
+if (cond) {
+  int z = x + y;
+  out = softmax(z);
+} else {
+  int z = fc(x);
+  out = z;
+}
+```
+
+An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator](./if_else_op.md) is as follows:
+
+```python
+import paddle as pd
+
+x = var(10)
+y = var(20)
+cond = var(false)
+ie = pd.create_ifelseop(inputs=[x], output_num=1)
+with ie.true_block():
+    x = ie.inputs(true, 0)
+    z = operator.add(x, y)
+    ie.set_output(true, 0, operator.softmax(z))
+with ie.false_block():
+    x = ie.inputs(false, 0)
+    z = layer.fc(x)
+    ie.set_output(true, 0, operator.softmax(z))
+out = b(cond)
+```
+
+In both examples, the left branch computes `softmax(x+y)` and the right branch computes `fc(x)`.
+
+A difference is that variables in the C++ program contain scalar values, whereas those in the PaddlePaddle programs are mini-batches of instances.  The `ie.input(true, 0)` invocation returns instances in the 0-th input, `x`, that corresponds to true values in `cond` as the local variable `x`, where `ie.input(false, 0)` returns instances corresponding to false values.
+
+### Blocks with `for` and `RNNOp`
+
+The following RNN model from the [RNN design doc](./rnn.md)
+
+```python
+x = sequence([10, 20, 30])
+m = var(0)
+W = tensor()
+U = tensor()
+
+rnn = create_rnn(inputs=[input])
+with rnn.stepnet() as net:
+  x = net.set_inputs(0)
+  h = net.add_memory(init=m)
+  fc_out = pd.matmul(W, x)
+  hidden_out = pd.matmul(U, h.pre(n=1))
+  sum = pd.add_two(fc_out, hidden_out)
+  act = pd.sigmoid(sum)
+  h.update(act)                       # update memory with act
+  net.set_outputs(0, act, hidden_out) # two outputs
+
+o1, o2 = rnn()
+print o1, o2
+```
+
+has its equivalent C++ program as follows
+
+```c++
+int* x = {10, 20, 30};
+int m = 0;
+int W = some_value();
+int U = some_other_value();
+
+int mem[sizeof(x) / sizeof(x[0]) + 1];
+int o1[sizeof(x) / sizeof(x[0]) + 1];
+int o2[sizeof(x) / sizeof(x[0]) + 1];
+for (int i = 1; i <= sizeof(x)/sizeof(x[0]); ++i) {
+  int x = x[i-1];
+  if (i == 1) mem[0] = m;
+  int fc_out = W * x;
+  int hidden_out = Y * mem[i-1];
+  int sum = fc_out + hidden_out;
+  int act = sigmoid(sum);
+  mem[i] = act;
+  o1[i] = act;
+  o2[i] = hidden_out;
+}
+
+print_array(o1);
+print_array(o2);
+```
+
+
+## Compilation and Execution
+
+Like TensorFlow programs, a PaddlePaddle program is written in Python.  The first part describes a neural network as a protobuf message, and the rest part executes the message for training or inference.
+
+The generation of this protobuf message is like what a compiler generates a binary executable file.  The execution of the message that the OS executes the binary file.
+
+## The "Binary Executable File Format"
+
+The definition of the protobuf message is as follows:
+
+```protobuf
+message BlockDesc {
+  repeated VarDesc vars = 1;
+  repeated OpDesc ops = 2;
+}
+```
+
+The step net in above RNN example would look like
+
+```
+BlockDesc {
+  vars = {
+    VarDesc {...} // x
+    VarDesc {...} // h
+    VarDesc {...} // fc_out
+    VarDesc {...} // hidden_out
+    VarDesc {...} // sum
+    VarDesc {...} // act
+  }
+  ops = {
+    OpDesc {...} // matmul
+    OpDesc {...} // add_two
+    OpDesc {...} // sigmoid
+  }
+};
+```
+
+Also, the RNN operator in above example is serialized into a protobuf message of type `OpDesc` and would look like:
+
+```
+OpDesc {
+  inputs = {0} // the index of x
+  outputs = {5, 3} // indices of act and hidden_out
+  attrs {
+    "memories" : {1} // the index of h
+    "step_net" : <above step net>
+  }
+};
+```
+
+This `OpDesc` value is in the `ops` field of the `BlockDesc` value representing the global block.
+
+
+## The Compilation of Blocks
+
+During the generation of the Protobuf message, the Block should store VarDesc (the Protobuf message which describes Variable) and OpDesc (the Protobuf message which describes Operator).
+
+VarDesc in a block should have its name scope to avoid local variables affect parent block's name scope.
+Child block's name scopes should inherit the parent's so that OpDesc in child block can reference a VarDesc that stored in parent block. For example
+
+```python
+a = pd.Varaible(shape=[20, 20])
+b = pd.fc(a, params=["fc.w", "fc.b"])
+
+rnn = pd.create_rnn()
+with rnn.stepnet() as net:
+    x = net.set_inputs(a)
+    # reuse fc's parameter
+    fc_without_b = pd.get_variable("fc.w")
+    net.set_outputs(fc_without_b)
+
+out = rnn()
+```
+the method `pd.get_variable` can help retrieve a Variable by a name, a Variable may store in a parent block, but might be retrieved in a child block, so block should have a variable scope that supports inheritance.
+
+In compiler design, the symbol table is a data structure created and maintained by compilers to store information about the occurrence of various entities such as variable names, function names, classes, etc.
+
+To store the definition of variables and operators, we define a C++ class `SymbolTable`, like the one used in compilers.
+
+`SymbolTable` can do the following stuff:
+
+- store the definitions (some names and attributes) of variables and operators,
+- to verify if a variable was declared,
+- to make it possible to implement type checking (offer Protobuf message pointers to `InferShape` handlers).
+
+
+```c++
+// Information in SymbolTable is enough to trace the dependency graph. So maybe
+// the Eval() interface takes a SymbolTable is enough.
+class SymbolTable {
+ public:
+  SymbolTable(SymbolTable* parent) : parent_(parent) {}
+
+  OpDesc* NewOp(const string& name="");
+
+  // TODO determine whether name is generated by python or C++
+  // currently assume that a unique name will be generated by C++ if the
+  // argument name left default.
+  VarDesc* NewVar(const string& name="");
+
+  // find a VarDesc by name, if recursive true, find parent's SymbolTable
+  // recursively.
+  // this interface is introduced to support InferShape, find protobuf messages
+  // of variables and operators, pass pointers into InferShape.
+  // operator
+  //
+  // NOTE maybe some C++ classes such as VarDescBuilder and OpDescBuilder should
+  // be proposed and embedded into pybind to enable python operate on C++ pointers.
+  VarDesc* FindVar(const string& name, bool recursive=true);
+
+  OpDesc* FindOp(const string& name);
+
+  BlockDesc Compile() const;
+
+ private:
+  SymbolTable* parent_;
+
+  map<string, OpDesc> ops_;
+  map<string, VarDesc> vars_;
+};
+```
+
+After all the description of variables and operators is added into SymbolTable,
+the block has enough information to run.
+
+The `Block` class takes a `BlockDesc` as input, and provide `Run` and `InferShape` functions.
+
+
+```c++
+namespace {
+
+class Block : OperatorBase {
+public:
+  Block(const BlockDesc& desc) desc_(desc) {}
+
+  void InferShape(const framework::Scope& scope) const override {
+    if (!symbols_ready_) {
+      CreateVariables(scope);
+      CreateOperators();
+    }
+    // should run InferShape first.
+    for (auto& op : runtime_table_.ops()) {
+      op->InferShape(scope);
+    }
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
+    PADDLE_ENFORCE(symbols_ready_, "operators and variables should be created first.");
+    for (auto& op : runtime_table_.ops()) {
+      op->Run(scope, dev_ctx);
+    }
+  }
+
+  void CreateVariables(const framework::Scope& scope);
+  void CreateOperators();
+
+  // some other necessary interfaces of NetOp are list below
+  // ...
+
+private:
+  BlockDesc desc_;
+  bool symbols_ready_{false};
+};
+```
+
+## The Execution of Blocks
+
+Block inherits from OperatorBase, which has a Run method.
+Block's Run method will run its operators sequentially.
+
+There is another important interface called `Eval`, which take some arguments called targets, and generate a minimal graph which takes targets as the end points and creates a new Block,
+after `Run`, `Eval` will get the latest value and return the targets.
+
+The definition of Eval is as follows:
+
+```c++
+// clean a block description by targets using the corresponding dependency graph.
+// return a new BlockDesc with minimal number of operators.
+// NOTE not return a Block but the block's description so that this can be distributed
+// to a cluster.
+BlockDesc Prune(const BlockDesc& desc, vector<string> targets);
+
+void Block::Eval(const vector<string>& targets,
+                 const framework::Scope& scope,
+                 const platform::DeviceContext& dev_ctx) {
+  BlockDesc min_desc = Prune(desc_, targets);
+  Block min_block(min_desc);
+  min_block.Run(scope, dev_ctx);
+}
+```

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/if_else_op.md

-IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has M (M<=N) instances, each corresponds to a true element in `cond`.
-
-```python
-import paddle as pd
-
-x = var()
-y = var()
-cond = var()
-
-b = pd.create_ifop(inputs=[x], output_num=1)
-with b.true_block():
-    x = b.inputs(0)
-    z = operator.add(x, y)
-    b.set_output(0, operator.softmax(z))
-
-out = b(cond)
-```
-
-If we want the output still has N instances, we can use IfElseOp with a default value, whose minibatch size must be N:
+IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack().
 
 ```python
 import paddle as pd
@@ -39,7 +21,7 @@ with b.false_block():
 out = b(cond)
 ```
 
-If only true_block is set in an IfElseOp, we can have a default value for false as:
+If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as:
 ```python
 import paddle as pd
 

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/ops/dist_train.md

+# Design Doc: Operation Graph Based Parameter Server
+
+## Abstract
+
+We propose an approach to implement the parameter server. In this
+approach, there is no fundamental difference between the trainer and
+the parameter server: they both run subgraphs, but subgraphs of
+different purposes.
+
+## Background
+
+The previous implementations of the parameter server does not run a
+subgraph. parameter initialization, optimizer computation, network
+communication and checkpointing are implemented twice on both the
+trainer and the parameter server.
+
+It would be great if we can write code once and use them on both the
+trainer and the parameter server: reduces code duplication and
+improves extensibility. Given that after the current refactor, we are
+representing everything as a computing graph on the
+trainer. Representing everything as a computing graph on the parameter
+server becomes a natural extension.
+
+## Design
+
+### Graph Converter
+
+The *graph converter* converts the user-defined operation (OP) graph
+into subgraphs to be scheduled on different nodes with the following
+steps:
+
+1. OP placement: the OPs will be placed on different nodes according
+   to heuristic that minimizes estimated total computation
+   time. Currently we will use a simple heuristic that puts parameter
+   varable on parameter server workers and everything else on trainer
+   workers.
+
+1. Add communication OPs to enable the communication between nodes.
+
+We will need these OPs: *Send*, *Recv*, *Enqueue*, *Dequeue*.
+
+Below is an example of converting the user defined graph to the
+subgraphs for the trainer and the parameter server:
+
+<img src="src/local-graph.png" width="300"/>
+
+After converting:
+
+<img src="src/dist-graph.png" width="700"/>
+
+1. The parameter variable W and it's optimizer subgraph are placed on the parameter server.
+1. Operators are added to the subgraphs.
+   - *Send* sends data to the connected *Recv* operator.  The
+	 scheduler on the receive node will only schedule *Recv* operator
+	 to run when the *Send* operator has ran (the *Send* OP will mark
+	 the *Recv* OP runnable automatically).
+   - *Enueue* enqueues the input variable, it can block until space
+     become available in the queue.
+   - *Dequeue* outputs configurable numbers of tensors from the
+     queue. It will block until the queue have the required number of
+     tensors.
+
+
+### Benefits
+
+- Model parallelism become easier to implement: it's an extension to
+  the trainer - parameter server approach. we already have the
+  communication OPs, but need to extend the graph converter's
+  placement functionality.
+
+- User-defined optimizer is easier to add - user can now express it as
+  a subgraph.
+
+- No more duplication logic inside the trainer and the parameter
+  server mentioned in the background section.
+
+### Challenges
+
+- It might be hard for the graph converter to cut a general graph
+  (without any hint for which subgraph is the optimizer). We may need
+  to label which subgraph inside the OP graph is the optimizer.
+
+- It's important to balance the parameter shards of on multiple
+  parameter server. If a single parameter is very big (some
+  word-embedding, fully connected, softmax layer), we need to
+  automatically partition the single parameter onto different
+  parameter servers when possible (only element-wise optimizer depends
+  on the parameter variable).
+
+### Discussion
+
+- In the "Aync SGD" figure, the "W" variable on the parameter server
+  could be read and wrote concurrently, what is our locking strategy?
+  E.g., each variable have a lock cpp method to be invoked by every
+  OP, or, have a lock OP.
+
+- Can the Enqueue OP be implemented under our current tensor design
+  (puts the input tensor into the queue tensor)?
+
+- *Dequeue* OP will have variable numbers of output (depends on the
+  `min_count` attribute), does our current design support it? (similar
+  question for the *Add* OP)
+
+
+### References:
+[1] [TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems](https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45166.pdf)

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

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

doc/design/ops/images/rnn.dot

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

doc/design/ops/images/rnn_2level_data.dot

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

doc/design/ops/rnn.md

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

doc/design/refactorization.md

+# Design Doc: Refactorization Overview
+
+The goal of refactorizaiton include:
+
+1. Make it easy for external contributors to write new elementory computaiton operations.
+1. Make the codebase clean and readable.
+1. Introduce a new design of computation representation -- a computation graph of operators and variables.
+1. The graph representation helps implementing auto-scalable and auto fault recoverable distributed computing.
+
+## Computation Graphs
+
+1. PaddlePaddle represent the computation, training and inference of DL models, by computation graphs.
+
+  1. Please dig into [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a solid example.
+
+1. Users write Python programs to describe the graphs and run it (locally or remotely).
+
+1. A graph is composed of *variabels* and *operators*.
+
+1. The description of graphs must be able to be serialized/deserialized, so it
+
+   1. could to be sent to the cloud for distributed execution, and
+   1. be sent to clients for mobile or enterprise deployment.
+
+1. The Python program do
+
+   1. *compilation*: runs a Python program to generate a protobuf message representation of the graph and send it to
+      1. the C++ library `libpaddle.so` for local execution,
+      1. the master process of a distributed training job for training, or
+      1. the server process of a Kubernetes serving job for distributed serving.
+   1. *execution*: according to the protobuf message, constructs instances of class `Variable` and `OperatorBase`, and run them.
+
+## Description and Realization
+
+At compile time, the Python program generates protobuf message representation of the graph, or the description of the graph.
+
+At runtime, the C++ program realizes the graph and run it.
+
+| | Representation (protobuf messages) | Realization (C++ class objects) |
+|---|---|---|
+|Data|[VarDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L107)|[Variable](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24)|
+|Operation|[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35)|[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64)|
+|Block|BlockDesc|Block|
+
+The word *graph* is exchangable with *block* in this document.  A graph represent computation steps and local variables as a C++/Java program block, or a pair of { and }.
+
+## Compilation and Execution
+
+1. Run an applicaton Python program to describe the graph.  In particular,
+
+   1. create VarDesc to represent local/intermediate variables,
+   1. create operators and set attributes,
+   1. validate attribute values,
+   1. inference the type and the shape of variables,
+   1. plan for memory-reuse for variables,
+   1. generate backward and optimization part of the Graph.
+   1. possiblly split the graph for distributed training.
+
+1. The invocation of `train` or `infer` in the application Python program:
+
+   1. create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block,
+      1. realize local variables defined in the BlockDesc message in the new scope,
+      1. a scope is similar to the stack frame in programming languages,
+
+   1. create an instance of class `Block`, in which,
+      1. realize operators in the BlockDesc message,
+
+   1. run the Block by calling
+      1. `Block::Eval(vector<Variable>* targets)` for forward and backward computations, or
+      1. `Block::Eval(vector<Operator>* targets)` for optimization.
+
+
+## Intermediate Representation (IR)
+
+```text
+Compile Time -> IR -> Runtime
+```
+
+### Benefit
+
+- Optimization
+  ```text
+  Compile Time -> IR -> Optimized IR -> Runtime
+  ```
+- Send automatically partitioned IR to different nodes.
+  - Automatic data parallel
+    ```text
+    Compile Time
+    |-> Single GPU IR
+        |-> [trainer-IR-0, trainer-IR-1, pserver-IR]
+            |-> Node-0 (runs trainer-IR-0)
+            |-> Node-1 (runs trainer-IR-1)
+            |-> Node-2 (runs pserver-IR)
+    ```
+  - Automatic model parallel (planned for future)
+
+---
+
+# Operator/OpWithKernel/OpKernel
+
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot)
+
+---
+
+# Operator
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot)
+
+* `Operator` is the fundamental building block as the user interface.
+    * Operator stores input/output variable name, and attributes.
+    * The `InferShape` interface is used to infer output variable shapes by its input shapes.
+    * Use `Run` to compute `input variables` to `output variables`.
+
+---
+
+# OpWithKernel/Kernel
+
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot)
+
+* `OpWithKernel` inherits `Operator`.
+* `OpWithKernel` contains a Kernel map.
+    * `OpWithKernel::Run` get device's kernel, and invoke `OpKernel::Compute`.
+    * `OpKernelKey` is the map key. Only device place now, but may be data type later.
+
+---
+
+# Why separate Kernel and Operator
+
+* Separate GPU and CPU code.
+    * Make Paddle can run without GPU.
+* Make one operator (which is user interface) can contain many implementations.
+    * Same mul op, different FP16, FP32 Kernel. different MKL, eigen kernel.
+---
+
+# Libraries for Kernel development
+
+* `Eigen::Tensor` contains basic math and element-wise functions.
+    * Note that `Eigen::Tensor` has broadcast implementation.
+    * Limit number of `tensor.device(dev) = ` in your code.
+* `thrust::tranform` and `std::transform`.
+    * `thrust` has the same API as C++ standard library. Using `transform` can quickly implement a customized elementwise kernel.
+    * `thrust` has more complex API, like `scan`, `reduce`, `reduce_by_key`.
+* Hand-writing `GPUKernel` and `CPU` code
+    * Do not write `.h`. CPU Kernel should be in `.cc`. CPU kernel should be in `.cu`. (`GCC` cannot compile GPU code.)
+---
+# Operator Register
+
+## Why register is necessary?
+We need a method to build mappings between Op type names and Op classes.
+
+## How to do the register?
+
+Maintain a map, whose key is the type name and value is corresponding Op constructor.
+
+---
+# The Registry Map
+
+### `OpInfoMap`
+
+`op_type(string)` -> `OpInfo`
+
+`OpInfo`:
+
+- **`creator`**: The Op constructor.
+- **`grad_op_type`**: The type of the gradient Op.
+- **`proto`**: The Op's Protobuf, including inputs, outputs and required attributes.
+- **`checker`**: Used to check attributes.
+
+---
+# Related Concepts
+
+### Op_Maker
+It's constructor takes `proto` and `checker`. They are compeleted during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37))
+
+### Register Macros
+```cpp
+REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class)
+REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
+```
+
+### `USE` Macros
+make sure the registration process is executed and linked.
+
+---
+# Register Process
+1. Write Op class, as well as its gradient Op class if there is.
+2. Write Op maker class. In the constructor, describe its inputs, outputs, and attributes.
+3. Invoke macro `REGISTER_OP`. The macro will
+	1. call maker class to complete `proto` and `checker`
+	2. with the completed `proto` and `checker`, build a new key-value pair in the `OpInfoMap`
+
+4. Invoke `USE` macro in where the Op is used to make sure it is linked.
+
+---
+# Backward Module (1/2)
+### Create Backward Operator
+- Mapping from forwarding Op to backward Op
+![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png)
+
+---
+# Backward Module (2/2)
+### Build Backward Network
+- **Input** graph of forwarding operators
+- **Output** graph of backward operators
+- **corner case in construction**
+	- shared variable => insert `Add` operator
+	- no gradient => insert `fill_zero_grad` operator
+	- recursive netOp => call `Backward` recursively
+	- RNN Op => recursively call `Backward` on stepnet
+
+
+---
+# Scope, Variable, Tensor
+
+* `Tensor` is an n-dimension array with type.
+	* Only dims and data pointers are stored in `Tensor`.
+	* All operators on `Tensor` is written in `Operator` or global functions.
+	* variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md)
+* `Variable` is the inputs and outputs of an operator. Not just `Tensor`.
+	* step_scopes in RNN is a variable and not a tensor.
+* `Scope` is where variables store at.
+	* map<string/*var name */, Variable>
+	* `Scope` has a hierarchical structure. The local scope can get variable from its parent scope.
+
+---
+# Block (in design)
+## the difference with original RNNOp
+- as an operator is more intuitive than `RNNOp`,
+- offers new interface `Eval(targets)` to deduce the minimal block to `Run`,
+- fits the compile-time/ runtime separation design.
+  - during the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc`
+  - when graph executes, a Block with `BlockDesc` passed in creates `Op` and `Var` then `Run`
+
+---
+# Milestone
+- take Paddle/books as the main line, the requirement of the models motivates framework refactoring,
+- model migration
+  - framework development gives **priority support** to model migration, for example,
+    - the MNIST demo needs a Python interface,
+    - the RNN models require the framework to support `LoDTensor`.
+  - determine some timelines,
+  - heavily-relied Ops need to be migrated first,
+  - different models can be migrated parallelly.
+- improve the framework at the same time
+- accept imperfection, concentrated on solving the specific problem at the right price.
+
+---
+# Control the migration quality
+- compare the performance of migrated models with old ones.
+- follow google C style
+- build the automatic workflow of generating Python/C++ documentations
+  - the documentation of layers and ops should be written inside the code
+  - take the documentation quality into account when doing PR
+  - preview the documentations, read and improve them from users' perspective

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

@@ -23,15 +23,18 @@
 - `framework::OperatorWithKernel`：继承自OperatorBase，Op有计算函数，称作有Kernel。
 - `class OpProtoAndCheckerMaker`：描述该Op的输入、输出、属性、注释,主要用于Python API接口生成
 
-依据是否包含kernel，将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下：
+依据是否包含kernel，可以将Op分为两种：包含Kernel的Op和不包含kernel的Op，前者Op的定义继承自`OperatorBase`，后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写，简单总结Op需要包含的内容如下：
 
 
  内容            | 定义位置
 --------------  | :----------------------
 OpProtoMake定义  | `.cc`文件，Backward Op不需要定义OpProtoMake
 Op定义           | `.cc`文件
-Kernel实现       | CPU、GPU共享Kernel在`.h`文件，否则，CPU可以在`.cc`文件，GPU可在`.cu`文件。 
-注册Op           | Op注册在`.cc`文件；Kernel注册CPU在`.cc`文件，GPU在`.cu`文件
+Kernel实现       | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，GPU 实现在`.cu`文件中。
+注册Op           | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中
+
+
+实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
 
 
 下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。
@@ -42,9 +45,11 @@ Kernel实现       | CPU、GPU共享Kernel在`.h`文件，否则，CPU可以在`
 
 ### 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 {
  public:
   MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
@@ -60,19 +65,19 @@ 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：
+[`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，该命名尽可能的规范。
+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>
 class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
@@ -88,16 +93,18 @@ The equation is: Out = scale*X
 };
 ```
 
- 在这个例子里，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。
-  - `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。
+- `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。
+
+- `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：
 
-```c++
+```cpp
 class MulOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
@@ -122,13 +129,13 @@ class MulOp : public framework::OperatorWithKernel {
 
 [`MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L22)继承自`OperatorWithKernel`。`public`成员：
 
-```c++
+```cpp
 using framework::OperatorWithKernel::OperatorWithKernel;
 ```
 
 这句表示使用基类`OperatorWithKernel`的构造函数，也可写成：
 
-```c++
+```cpp
 MulOp(const std::string &type, const framework::VariableNameMap &inputs,
       const framework::VariableNameMap &outputs,
       const framework::AttributeMap &attrs)
@@ -140,13 +147,26 @@ 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
 
-```C++
-template <typename Place, typename T>
-class MulKernel : public framework::OpKernel {
+`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");
@@ -157,55 +177,54 @@ 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:
+  };
+  ```
 
-  - `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)。
+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 T` : 表示数据类型，如`float`, `double`等。
+`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)。
 
-`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。
+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)。
 
-注意，不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel`，`OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。 
 
-到此前向Op实现完成，需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似，这里不再重复。但注意，反向Op没有`ProtoMaker`。
+到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。
+反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。
 
 ### 4. 注册Operator
 
-在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
+- 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
 
-```c++
-namespace ops = paddle::operators;
-REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
-REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
-REGISTER_OP_CPU_KERNEL(mul_grad,
+    ```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`，
+   在上面的代码中：
+
+    - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`。
     - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
     - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
 
-在 `.cu`文件中注册GPU Kernel。
 
-```c++
-namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
-REGISTER_OP_GPU_KERNEL(mul_grad,
-                       ops::MulGradKernel<paddle::platform::GPUPlace, float>);
-```
+- 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：
 
-### 5. 编译
+    ```cpp
+    // if use Eigen unsupported module before include head files
+    #define EIGEN_USE_GPU
 
-在[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)文件中添加编译。
+    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>);
+    ```
 
-```
-op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
-```
+### 5. 编译
 
-下面命令可以编译：
+运行下面命令可以进行编译：
 
 ```
 make mul_op
@@ -213,58 +232,28 @@ make mul_op
 
 ## 绑定Python
 
-- 绑定Python 
- 
-    在 [`paddle/pybind/pybind.cc 
-`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc)文件中添加该类：
-
-    ```
-    USE_OP(mul);
-    ```
-    如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`:
-    
-    ```
-    USE_CPU_ONLY_OP(gather);
-    ```
-    
-    如果OP不带Kernel，则使用`USE_NO_KENREL_OP`:
+系统会对新增的op自动绑定Python，并链接到生成的lib库中。
 
-    ```
-    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
+## 实现单元测试
 
-   在 [`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件添加类到`DEPS`中，使得该Op可以链接到生成的lib库中。
+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)。
 
-   ```
-   if(WITH_PYTHON)
-     cc_library(paddle_pybind SHARED
-     SRCS pybind.cc
-     DEPS pybind python backward
-     mul_op
-     minus_op)
-   endif(WITH_PYTHON)
-   ```
+### 前向Operator单元测试
 
-## 实现单元测试
+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：
 
-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)。
+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。
 
-### 前向Operator单测
 
-前向Op单测继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`，具体单测流程在`OpTestMeta`里完成。需在`setUp`函数定义输入输出和属性参数，以及Python对比的输出值。
+  ```python
+  import unittest
+  import numpy as np
+  from gradient_checker import GradientChecker, create_op
+  from op_test_util import OpTestMeta
 
-```
-import unittest
-import numpy as np
-from gradient_checker import GradientChecker, create_op
-from op_test_util import OpTestMeta
-
-class TestMulOp(unittest.TestCase):
+  class TestMulOp(unittest.TestCase):
       __metaclass__ = OpTestMeta
 
       def setUp(self):
@@ -274,57 +263,84 @@ class TestMulOp(unittest.TestCase):
               'Y': np.random.random((84, 100)).astype("float32")
           }
           self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
-```
-   首先需要`import`必要的包,下面详细解释其他值：
+  ```
 
-   - `self.type = "mul" ` : 定义类型，和注册的类型一致。
-   - `self.inputs` : 定义输入，类型为Numpy.array，并初始化。
-   - `self.outputs` : 定义输出，并得到Python结算结果。
+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单测
 
-反向Op单测继承自`GradientChecker`，而`GradientChecker`集成自`unittest.TestCase`，所以反向单测函数需要`test_`开头。
+### 反向Operator单元测试
 
- ```
- class MulGradOpTest(GradientChecker):
-    def test_mul(self):
-        op = create_op("mul")
-        inputs = {
+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**。
+
+```python
+class TestMulGradOp(GradientChecker):
+    def setUp(self):
+        self.op = create_op("mul")
+        self.inputs = {
             'X': np.random.random((32, 84)).astype("float32"),
             'Y': np.random.random((84, 100)).astype("float32")
         }
-        self.compare_grad(op, inputs)      
+
+    def test_cpu_gpu_compare(self):
+        self.compare_grad(self.op, self.inputs)
+
+    def test_normal(self):
         # mul op will enlarge the relative error
         self.check_grad(
-            op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5)
- ```
+            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)
 
-   - 调用`create_op("mul")`创建反向Op对应的前向Op。
-   - 定义输入`inputs`。
-   - 调用`compare_grad`函数对比CPU、GPU计算结果。
-   - 调用`check_grad`检查梯度稳定性，这里采用数值法检测梯度正确性。
-      - 第一个参数`op` : 前向op。
-      - 第二个参数`inputs` : 输入词典，词典的Key和`ProtoMaker`定义保持一致。
-      - 第三个参数`set(["X", "Y"])` : 指定对输入变量`X`、`Y`做梯度检测。
-      - 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
+    def test_ignore_x(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["Y"],
+            "Out",
+            max_relative_error=0.5,
+            no_grad_set={"X"})
 
+    def test_ignore_y(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["X"],
+            "Out",
+            max_relative_error=0.5,
+            no_grad_set={"Y"})
+```
 
-### 编译和执行 
+下面解释代码中一些关键的地方:
 
-单测完成之后，在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)里添加编译：
+- 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。
 
-```
-py_test(test_mul_op SRCS test_mul_op.py)
-```
 
-编译时需要打开`WITH_TESTING`, 即 `cmake paddle_dir -DWITH_TESTING=ON`，编译成功之后执行单测命令为：
+### 编译和执行单元测试
 
-```
+`python/paddle/v2/framework/tests` 目录下新增的 `test_*.py` 单元测试会被自动加入工程进行编译。
+
+请注意，**不同于Op的编译测试，运行单元测试测时需要编译整个工程**，并且编译时需要打开`WITH_TESTING`, 即`cmake paddle_dir -DWITH_TESTING=ON`。编译成功后，执行下面的命令来运行单元测试：
+
+```bash
 make test ARGS="-R test_mul_op -V"
 ```
+
 或者:
 
-```
+```bash
 ctest -R test_mul_op
 ```
+
+## 注意事项
+
+- 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，这将会导致编译出错。
+- 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。
+- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
+- 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。

doc/howto/dev/use_eigen_cn.md

+## 在Paddle中如何使用Eigen
+
+神经网络本质上是一个计算图，计算需要的数据存放在`Tensor`中，而计算过程是由`Operartor`来描述的。在执行时，`Operator`调用对应`OpKernel`中的`Compute`接口，实现对`Tensor`的操作。
+
+
+### Eigen Tensor模块
+
+Eigen Tensor模块对element-wise计算提供了强大的支持，并且书写一份代码，可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块，因此可能测试不够完备，文档较少。
+
+关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md)
+
+
+### paddle::framework::Tensor
+
+Paddle Tensor定义在framework目录下，其主要接口如下：
+
+```cpp
+class Tensor {
+ public:
+  /*! Return a pointer to mutable memory block. */
+  template <typename T>
+  inline T* data();
+  
+  /**
+   * @brief   Return a pointer to mutable memory block.
+   * @note    If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(platform::Place place);
+  
+  /**
+   * @brief     Return a pointer to mutable memory block.
+   *
+   * @param[in] dims    The dimensions of the memory block.
+   * @param[in] place   The place of the memory block.
+   *
+   * @note      If not exist, then allocation.
+   */
+  template <typename T>
+  inline T* mutable_data(DDim dims, platform::Place place);
+  
+  /*! Resize the dimensions of the memory block. */
+  inline Tensor& Resize(const DDim& dims);
+  
+  /*! Return the dimensions of the memory block. */
+  inline const DDim& dims() const;
+
+ private:  
+  /*! holds the memory block if allocated. */
+  std::shared_ptr<Placeholder> holder_;
+  
+  /*! points to dimensions of memory block. */
+  DDim dim_;
+};
+```
+
+`Placeholder`的作用是延迟分配内存，即我们可以先定义一个Tensor，然后使用Resize接口设置Tensor的大小，最后再调用mutable_data接口分配实际的内存。
+
+```cpp
+paddle::framework::Tensor t;
+paddle::platform::CPUPlace place;
+// set size first
+t.Resize({2, 3});
+// allocate memory on CPU later
+t.mutable_data(place);
+```
+
+### paddle::framework::Tensor使用样例
+下面以AddOp为例说明Tensor的使用过程：
+
+- InferShape
+
+在运行神经网络计算图时，我们先调用每个`Operator`的`InferShape`接口，根据输入Tensor的大小来设置输出Tensor的大小，`Resize`接口会被调用。
+
+```cpp
+void InferShape(const framework::InferShapeContext &ctx) const override {
+  PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
+                    ctx.Input<Tensor>("Y")->dims(),
+                    "Two input of Add Op's dimension must be same.");
+  ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
+}
+```
+
+
+- Run
+
+`Operator`的`Run`接口最终会调用对应`OpKernel`的`Compute`接口，在这时真正的分配内存，`mutable_data`接口会被调用。
+
+```cpp
+void Compute(const framework::ExecutionContext& context) const override {
+  auto* input0 = context.Input<Tensor>("X");
+  auto* input1 = context.Input<Tensor>("Y");
+  auto* output = context.Output<Tensor>("Out");
+
+  output->mutable_data<T>(context.GetPlace());
+
+  auto x = EigenVector<T>::Flatten(*input0);
+  auto y = EigenVector<T>::Flatten(*input1);
+  auto z = EigenVector<T>::Flatten(*output);
+
+  auto place = context.GetEigenDevice<Place>();
+
+  z.device(place) = x + y;
+}
+```
+
+
+### paddle::framework::Tensor到EigenTensor的转换
+
+如上一小节所示，在具体的计算中，我们需要先把输入Tensor和输出Tensor转换为Eigen支持的格式。我们在[eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen.h)中提供了一些全局函数用来实现paddle::framework::Tensor到EigenTensor/EigenMatrix/EigenVector/EigenScalar的转换。
+
+以EigenTensor为例，做一个介绍
+
+```cpp
+Tensor t;
+float* p = t.mutable_data<float>(make_ddim({1, 2, 3}), platform::CPUPlace());
+for (int i = 0; i < 1 * 2 * 3; i++) {
+  p[i] = static_cast<float>(i);
+}
+
+EigenTensor<float, 3>::Type et = EigenTensor<float, 3>::From(t);
+```
+
+From是EigenTensor模板提供的一个接口，可以实现从paddle::framework::Tensor到对EigenTensor的转换。由于Tensor的rank是模板参数，因此在转换时需要显示的指定。
+
+在Eigen中，不同rank的Tensor是不同类型，Vector是rank为1的Tensor。需要额外注意的是，EigenVector<T>::From方法是把paddle中的一维Tensor转为Eigen的一维Tensor，在这里用EigenVector来表示；而EigenVector<T>::Flatten方法是把paddle中的一个Tensor进行reshape操作，压扁成为Eigen的一维Tensor，类型仍然为EigenVector。
+
+更多的转换方法请参考eigen_test.cc中的[单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/eigen_test.cc)。
+
+
+
+### 实现计算
+
+当需要完成计算时，我们需要等式左边的EigenTensor调用device接口。在这里需要注意的是，这里的EigenTensor之间的运算只是改变了原有Tensor中的数据，而不会改变原有Tensor的shape信息。
+
+```cpp
+auto x = EigenVector<T>::Flatten(*input0);
+auto y = EigenVector<T>::Flatten(*input1);
+auto z = EigenVector<T>::Flatten(*output);
+auto place = context.GetEigenDevice<Place>();
+z.device(place) = x + y;
+```
+
+在这段代码中，input0/input1/output可以是任意维度的Tensor。我们调用了EigenVector的Flatten接口，把任意维度的Tensor转为了一维的EigenVector。而在计算结束之后，input0/input1/output的原有shape信息不变。如果想改变原有Tensor的shape信息，可以调用Resize接口进行改变。
+
+由于Eigen Tensor模块的文档较少，我们可以参考TensorFlow的[kernels](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/core/kernels)模块下的相关`OpKernel`的计算代码。

doc/howto/dev/write_docs_cn.rst

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

paddle/capi/CMakeLists.txt

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

paddle/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/cuda/include/hl_cuda_cudnn.h

@@ -22,10 +22,10 @@ limitations under the License. */
  */
 typedef enum {
   HL_POOLING_MAX = 0,
-  // average includes padded values
-  HL_POOLING_AVERAGE = 1,
   // average does not include padded values
-  HL_POOLING_AVERAGE_EXCLUDE_PADDING = 2,
+  HL_POOLING_AVERAGE = 1,
+  // average includes padded values
+  HL_POOLING_AVERAGE_INCLUDE_PADDING = 2,
   HL_POOLING_END
 } hl_pooling_mode_t;
 

paddle/cuda/include/hl_tensor_ops.h

@@ -461,7 +461,7 @@ class add<float32x4_t> {
 public:
   INLINE float32x4_t operator()(const float32x4_t a,
                                 const float32x4_t b) const {
-    return vmulq_f32(a, b);
+    return vaddq_f32(a, b);
   }
 };
 

paddle/cuda/src/hl_cuda_cnn.cu

@@ -211,13 +211,11 @@ __global__ void KeAvgPoolForward(const int nthreads,
 
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (hend - hstart) * (wend - wstart);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * height * width;
@@ -299,12 +297,14 @@ __global__ void KeAvgPoolBackward(const int nthreads,
     outGrad += (frameNum * outStride + offsetC * pooledH * pooledW);
 
     for (int ph = phstart; ph < phend; ++ph) {
+      int hstart = ph * strideH - padH;
+      int hend = min(hstart + sizeY, height);
+      hstart = max(hstart, 0);
       for (int pw = pwstart; pw < pwend; ++pw) {
         // figure out the pooling size
-        int hstart = ph * strideH - padH;
         int wstart = pw * strideW - padW;
-        int hend = min(hstart + sizeY, height + padH);
-        int wend = min(wstart + sizeX, width + padW);
+        int wend = min(wstart + sizeX, width);
+        wstart = max(wstart, 0);
         int poolsize = (hend - hstart) * (wend - wstart);
         gradient += outGrad[ph * pooledW + pw] / poolsize;
       }
@@ -600,16 +600,13 @@ __global__ void KeAvgPool3DForward(const int nthreads,
     int dstart = pd * strideD - padD;
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int dend = min(dstart + sizeZ, depth + padD);
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
+    int dend = min(dstart + sizeZ, depth);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     dstart = max(dstart, 0);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    dend = min(dend, depth);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * depth * height * width;
@@ -712,15 +709,18 @@ __global__ void KeAvgPool3DBackward(const int nthreads,
     outGrad += (frameNum * channels + offsetC) * pooledD * pooledH * pooledW;
 
     for (int pd = pdstart; pd < pdend; ++pd) {
+      int dstart = pd * strideD - padD;
+      int dend = min(dstart + sizeZ, depth);
+      dstart = max(dstart, 0);
       for (int ph = phstart; ph < phend; ++ph) {
+        int hstart = ph * strideH - padH;
+        int hend = min(hstart + sizeY, height);
+        hstart = max(hstart, 0);
         for (int pw = pwstart; pw < pwend; ++pw) {
           // figure out the pooling size
-          int dstart = pd * strideD - padD;
-          int hstart = ph * strideH - padH;
           int wstart = pw * strideW - padW;
-          int dend = min(dstart + sizeZ, depth + padD);
-          int hend = min(hstart + sizeY, height + padH);
-          int wend = min(wstart + sizeX, width + padW);
+          int wend = min(wstart + sizeX, width);
+          wstart = max(wstart, 0);
           int poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart);
           gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize;
         }

paddle/cuda/src/hl_cuda_cudnn.cc

@@ -432,11 +432,11 @@ void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc,
       cudnn_mode = CUDNN_POOLING_MAX;
       break;
     case HL_POOLING_AVERAGE:
-      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
-      break;
-    case HL_POOLING_AVERAGE_EXCLUDE_PADDING:
       cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
       break;
+    case HL_POOLING_AVERAGE_INCLUDE_PADDING:
+      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
+      break;
     default:
       LOG(FATAL) << "parameter mode error";
   }

paddle/framework/CMakeLists.txt

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

paddle/framework/attribute.cc

@@ -43,6 +43,10 @@ template <>
 AttrType AttrTypeID<std::vector<std::string>>() {
   return STRINGS;
 }
+template <>
+AttrType AttrTypeID<std::vector<std::pair<int, int>>>() {
+  return INT_PAIRS;
+}
 
 Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
   switch (attr_desc.type()) {
@@ -76,6 +80,14 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
       }
       return val;
     }
+    case paddle::framework::AttrType::INT_PAIRS: {
+      std::vector<std::pair<int, int>> val(attr_desc.int_pairs_size());
+      for (int i = 0; i < attr_desc.int_pairs_size(); ++i) {
+        val[i].first = attr_desc.int_pairs(i).first();
+        val[i].second = attr_desc.int_pairs(i).second();
+      }
+      return val;
+    }
   }
   PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
   return boost::blank();

paddle/framework/attribute.h

@@ -28,7 +28,8 @@ namespace paddle {
 namespace framework {
 
 typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
-                       std::vector<float>, std::vector<std::string>>
+                       std::vector<float>, std::vector<std::string>,
+                       std::vector<std::pair<int, int>>>
     Attribute;
 
 typedef std::unordered_map<std::string, Attribute> AttributeMap;
@@ -40,11 +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:
@@ -109,8 +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 +2,31 @@
 
 ## 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, many model is solved by the the backpropagation algorithm(known as BP) at present. Technically it caculates the gradient of the loss function, then distributed back through the networks. Follows the chain rule, so we need a module chains the gradient operators/expressions together with to construct the backward pass. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. 
 
-## Backward Operator Registry
+## Implementation
 
-A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs and output gradients and then calculate its input gradients.
+In this design doc, we exported only one API for generating the backward pass.
+
+```c++
+std::unique_ptr<OperatorBase> Backward(const OperatorBase& forwardOp,
+    const std::unordered_set<std::string>& no_grad_vars);
+```
+
+The implementation behind it can be divided into two parts, **Backward Operator Creating** and **Backward Operator Building**.
+
+### Backward Operator Registry
+
+A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs, and output gradients and then calculate its input gradients.
 
 |                        | forward operator | backward operator 
 | ---------------------- | ---------------- |------------------------- |		
 | **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);
@@ -25,9 +36,9 @@ REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
 
 `mul_grad` is the type of backward operator, and `MulOpGrad` is its class name.
 
-## Backward Opeartor Creating
+### Backward Opeartor Creating
 
-Given a certain forward operator, we can get its corresponding backward 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);
@@ -37,46 +48,53 @@ 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`.
 
 4. Building backward operator with `inputs`, `outputs` and forward operator's attributes.
 
-## Backward Network Building
-
-A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and put them together.
+### Backward Network Building
 
-In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network. 
-
-given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`.
+A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and append them together one by one. There is some corner case need to process specially.
 
 1. Op 
 
-   when the input forward network is a Op, return its gradient Operator Immediately.
+   When the input forward network is an Op, return its gradient Operator Immediately. If all of its outputs are in no gradient set, then return a special `NOP`.
 
 2. NetOp 
 
-   when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp.
+   In our design, the network itself is also a kind of operator(**NetOp**). So the operators contained by a big network may be some small network. When the input forward network is a NetOp, it needs to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to the forward NetOp.
+
+3. RnnOp
+
+   RnnOp is a nested stepnet operator.  Backward module need to recusively call `Backward` for every stepnet.
+
+4. Sharing Variables
+
+   **sharing variables**. As illustrated in the pictures, two operator's share the same variable name of W@GRAD, which will overwrite their sharing input variable. 
+
+<p align="center">
+<img src="./images/duplicate_op.png" width="50%" ><br/>
 
-   **shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable.  
+​	pic 1. Sharing variables in operators. 
 
-   <p align="center">
-   <img src="./images/duplicate_op.png" width="70%" ><br/>
+</p>
 
-   1. shared variable in two operators. 
+​	Sharing variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator to replace the overwrite links. 
 
-   </p>
+<p align="center">
+<img src="images/duplicate_op2.png" width="40%" ><br/>
 
-   Share variable between operators or same input variable used in multiple operators 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. 
+​	pic 2. Replace sharing variable's gradient with `Add` operator.
 
-   <p align="center">
-   <img src="images/duplicate_op2.png" width="90%" ><br/>
+</p>
 
-   2. replace shared variable gradient with `Add` Operator
+​	Because our framework finds variables accord to their names, we need to rename the output links. We add a suffix of number to represent its position in clockwise. 
 
-   </p>
+5. Part of Gradient is Zero.
 
+   In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator,  we need to fill a same shape gradient matrix in the position. In our implement, we insert a special `fillZeroLike` operator.
 
 
-​	Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.
+Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.

paddle/framework/ddim.cc

@@ -21,16 +21,16 @@ namespace framework {
 /// @cond HIDDEN
 
 template <int i>
-Dim<i> make_dim(const int* d) {
+Dim<i> make_dim(const int64_t* d) {
   return Dim<i>(*d, make_dim<i - 1>(d + 1));
 }
 
 template <>
-Dim<1> make_dim<1>(const int* d) {
+Dim<1> make_dim<1>(const int64_t* d) {
   return Dim<1>(*d);
 }
 
-void make_ddim(DDim& ddim, const int* dims, int n) {
+void make_ddim(DDim& ddim, const int64_t* dims, int n) {
   switch (n) {
     case 1:
       ddim = make_dim<1>(dims);
@@ -67,13 +67,13 @@ void make_ddim(DDim& ddim, const int* dims, int n) {
 
 /// @endcond
 
-DDim make_ddim(std::initializer_list<int> dims) {
+DDim make_ddim(std::initializer_list<int64_t> dims) {
   DDim result(make_dim(0));
   make_ddim(result, dims.begin(), dims.size());
   return result;
 }
 
-DDim make_ddim(const std::vector<int>& dims) {
+DDim make_ddim(const std::vector<int64_t>& dims) {
   DDim result(make_dim(0));
   make_ddim(result, &dims[0], dims.size());
   return result;
@@ -81,12 +81,12 @@ DDim make_ddim(const std::vector<int>& dims) {
 
 /// @cond HIDDEN
 // XXX For some reason, putting this in an anonymous namespace causes errors
-class DynamicMutableIndexer : public boost::static_visitor<int&> {
+class DynamicMutableIndexer : public boost::static_visitor<int64_t&> {
  public:
   explicit DynamicMutableIndexer(int idx) : idx_(idx) {}
 
   template <int D>
-  int& operator()(Dim<D>& dim) const {
+  int64_t& operator()(Dim<D>& dim) const {
     return dim[idx_];
   }
 
@@ -94,12 +94,12 @@ class DynamicMutableIndexer : public boost::static_visitor<int&> {
   int idx_;
 };
 
-class DynamicConstIndexer : public boost::static_visitor<int> {
+class DynamicConstIndexer : public boost::static_visitor<int64_t> {
  public:
   explicit DynamicConstIndexer(int idx) : idx_(idx) {}
 
   template <int D>
-  int operator()(const Dim<D>& dim) const {
+  int64_t operator()(const Dim<D>& dim) const {
     return dim[idx_];
   }
 
@@ -109,22 +109,22 @@ class DynamicConstIndexer : public boost::static_visitor<int> {
 
 /// @endcond
 
-int& DDim::operator[](int idx) {
+int64_t& DDim::operator[](int idx) {
   return boost::apply_visitor(DynamicMutableIndexer(idx), var);
 }
 
-int DDim::operator[](int idx) const {
+int64_t DDim::operator[](int idx) const {
   return boost::apply_visitor(DynamicConstIndexer(idx), var);
 }
 
-ssize_t DDim::size() const { return arity(*this); }
+int64_t DDim::size() const { return arity(*this); }
 
 bool DDim::operator==(DDim d) const {
   if (var.which() != d.getVar().which()) {
     return false;
   } else {
-    std::vector<int> v1 = vectorize(*this);
-    std::vector<int> v2 = vectorize(d);
+    std::vector<int64_t> v1 = vectorize(*this);
+    std::vector<int64_t> v2 = vectorize(d);
 
     for (unsigned int i = 0; i < v1.size(); i++) {
       if (v1[i] != v2[i]) {
@@ -139,10 +139,10 @@ bool DDim::operator==(DDim d) const {
 bool DDim::operator!=(DDim d) const { return !(*this == d); }
 
 DDim DDim::operator+(DDim d) const {
-  std::vector<int> v1 = vectorize(*this);
-  std::vector<int> v2 = vectorize(d);
+  std::vector<int64_t> v1 = vectorize(*this);
+  std::vector<int64_t> v2 = vectorize(d);
 
-  std::vector<int> v3;
+  std::vector<int64_t> v3;
 
   assert(v1.size() == v2.size());
 
@@ -154,10 +154,10 @@ DDim DDim::operator+(DDim d) const {
 }
 
 DDim DDim::operator*(DDim d) const {
-  std::vector<int> v1 = vectorize(*this);
-  std::vector<int> v2 = vectorize(d);
+  std::vector<int64_t> v1 = vectorize(*this);
+  std::vector<int64_t> v2 = vectorize(d);
 
-  std::vector<int> v3;
+  std::vector<int64_t> v3;
 
   assert(v1.size() == v2.size());
 
@@ -168,15 +168,15 @@ DDim DDim::operator*(DDim d) const {
   return make_ddim(v3);
 }
 
-int get(const DDim& ddim, int idx) { return ddim[idx]; }
+int64_t get(const DDim& ddim, int idx) { return ddim[idx]; }
 
 void set(DDim& ddim, int idx, int value) { ddim[idx] = value; }
 
 /// @cond HIDDEN
 struct VectorizeVisitor : public boost::static_visitor<> {
-  std::vector<int>& vector;
+  std::vector<int64_t>& vector;
 
-  explicit VectorizeVisitor(std::vector<int>& v) : vector(v) {}
+  explicit VectorizeVisitor(std::vector<int64_t>& v) : vector(v) {}
 
   template <typename T>
   void operator()(const T& t) {
@@ -188,31 +188,31 @@ struct VectorizeVisitor : public boost::static_visitor<> {
 };
 /// @endcond
 
-std::vector<int> vectorize(const DDim& ddim) {
-  std::vector<int> result;
+std::vector<int64_t> vectorize(const DDim& ddim) {
+  std::vector<int64_t> result;
   VectorizeVisitor visitor(result);
   boost::apply_visitor(visitor, ddim);
   return result;
 }
 
-struct ProductVisitor : public boost::static_visitor<ssize_t> {
+struct ProductVisitor : public boost::static_visitor<int64_t> {
   template <int D>
-  ssize_t operator()(const Dim<D>& dim) {
+  int64_t operator()(const Dim<D>& dim) {
     return product(dim);
   }
 };
 
-ssize_t product(const DDim& ddim) {
+int64_t product(const DDim& ddim) {
   ProductVisitor visitor;
   return boost::apply_visitor(visitor, ddim);
 }
 
 struct SliceVectorizeVisitor : public boost::static_visitor<> {
-  std::vector<int>& vector;
+  std::vector<int64_t>& vector;
   int begin;
   int end;
 
-  SliceVectorizeVisitor(std::vector<int>& v, int b, int e)
+  SliceVectorizeVisitor(std::vector<int64_t>& v, int b, int e)
       : vector(v), begin(b), end(e) {
     PADDLE_ENFORCE(begin < end,
                    "Begin index must be less than end index in ddim slice.");
@@ -240,7 +240,7 @@ struct SliceVectorizeVisitor : public boost::static_visitor<> {
 };
 
 DDim slice_ddim(const DDim& dim, int begin, int end) {
-  std::vector<int> vec;
+  std::vector<int64_t> vec;
   vec.reserve(end - begin);
   SliceVectorizeVisitor visitor(vec, begin, end);
   boost::apply_visitor(visitor, dim);
@@ -280,8 +280,17 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
   return os;
 }
 
-DDim::DDim(std::initializer_list<int> init_list) {
+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

@@ -40,7 +40,7 @@ struct DDim {
   template <int D>
   explicit DDim(const Dim<D>& in) : var(in) {}
 
-  /*implicit*/ DDim(std::initializer_list<int> init_list);
+  /*implicit*/ DDim(std::initializer_list<int64_t> init_list);
 
   template <int D>
   DDim& operator=(const Dim<D>& in) {
@@ -48,8 +48,8 @@ struct DDim {
     return *this;
   }
 
-  int& operator[](int idx);
-  int operator[](int idx) const;
+  int64_t& operator[](int idx);
+  int64_t operator[](int idx) const;
 
   template <typename Visitor>
   typename Visitor::result_type apply_visitor(Visitor& visitor) {
@@ -71,15 +71,15 @@ struct DDim {
 
   DDim operator*(DDim d) const;
 
-  ssize_t size() const;
+  int64_t size() const;
 };
 
 /**
- * \brief Make a DDim from std::vector<int>
+ * \brief Make a DDim from std::vector<int64_t>
  *
  * \param dims An vector of ints. Must be sized between [1, 9]
  */
-DDim make_ddim(const std::vector<int>& dims);
+DDim make_ddim(const std::vector<int64_t>& dims);
 
 /**
  * \brief Make a DDim from an initializer list
@@ -87,14 +87,14 @@ DDim make_ddim(const std::vector<int>& dims);
  * \param dims An initializer list of ints. Must be sized between [1, 9]
  *
  */
-DDim make_ddim(std::initializer_list<int> dims);
+DDim make_ddim(std::initializer_list<int64_t> dims);
 
-int get(const DDim& dim, int idx);
+int64_t get(const DDim& dim, int idx);
 void set(DDim& dim, int idx, int val);
 
-std::vector<int> vectorize(const DDim& ddim);
+std::vector<int64_t> vectorize(const DDim& ddim);
 
-ssize_t product(const DDim& ddim);
+int64_t product(const DDim& ddim);
 
 /**
  * \brief Slice a ddim
@@ -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/ddim_test.cc

@@ -12,7 +12,7 @@ TEST(DDim, Equality) {
   EXPECT_EQ(ddim[2], 5);
 
   // construct a DDim from a vector
-  std::vector<int> vec({9, 1, 5});
+  std::vector<int64_t> vec({9, 1, 5});
   paddle::framework::DDim vddim = paddle::framework::make_ddim(vec);
   EXPECT_EQ(ddim[0], 9);
   EXPECT_EQ(ddim[1], 1);
@@ -25,7 +25,7 @@ TEST(DDim, Equality) {
   EXPECT_EQ(paddle::framework::get(ddim, 0), 6);
 
   // vectorize a DDim
-  std::vector<int> res_vec = paddle::framework::vectorize(vddim);
+  std::vector<int64_t> res_vec = paddle::framework::vectorize(vddim);
   EXPECT_EQ(res_vec[0], 9);
   EXPECT_EQ(res_vec[1], 1);
   EXPECT_EQ(res_vec[2], 5);

paddle/framework/dim.h

@@ -17,13 +17,13 @@ struct Dim {
   static constexpr int dimensions = i;
 
   template <typename... Args>
-  HOSTDEVICE Dim(int _head, Args... _tail) : head(_head), tail(_tail...) {
+  HOSTDEVICE Dim(int64_t _head, Args... _tail) : head(_head), tail(_tail...) {
     static_assert(sizeof...(_tail) == i - 1,
                   "Dim initialized with the wrong number of parameters");
   }
 
   HOSTDEVICE
-  Dim(int _head, const Dim<i - 1>& _tail) : head(_head), tail(_tail) {}
+  Dim(int64_t _head, const Dim<i - 1>& _tail) : head(_head), tail(_tail) {}
 
   HOSTDEVICE
   Dim() : head(0), tail() {}
@@ -31,12 +31,12 @@ struct Dim {
   /** Construct a Dim from a linear index and size.  Uses Fortran order
    * indexing. */
   HOSTDEVICE
-  Dim(int idx, const Dim<i>& size)
+  Dim(int64_t idx, const Dim<i>& size)
       : head(idx % size.head), tail(idx / size.head, size.tail) {}
 
   /** Construct a Dim with each dimension set to the given index */
   HOSTDEVICE
-  Dim(int idx) : head(idx), tail(idx) {}
+  Dim(int64_t idx) : head(idx), tail(idx) {}
 
   HOSTDEVICE
   bool operator==(const Dim<i>& o) const {
@@ -47,13 +47,13 @@ struct Dim {
   bool operator!=(const Dim<i>& o) const { return !(*this == o); }
 
   HOSTDEVICE
-  int& operator[](int idx);
+  int64_t& operator[](int idx);
   HOSTDEVICE
-  int operator[](int idx) const;
+  int64_t operator[](int idx) const;
 
   HOST std::string to_string() const;
 
-  int head;
+  int64_t head;
   Dim<i - 1> tail;
 };
 
@@ -63,7 +63,7 @@ struct Dim<1> {
   static constexpr int dimensions = 1;
 
   HOSTDEVICE
-  Dim(int _head) : head(_head) {}
+  Dim(int64_t _head) : head(_head) {}
 
   HOSTDEVICE
   Dim() : head(0) {}
@@ -86,11 +86,11 @@ struct Dim<1> {
   bool operator!=(const Dim<1>& o) const { return !(*this == o); }
 
   HOSTDEVICE
-  int& operator[](int idx);
+  int64_t& operator[](int idx);
   HOSTDEVICE
-  int operator[](int idx) const;
+  int64_t operator[](int idx) const;
 
-  int head;
+  int64_t head;
 };
 
 namespace {
@@ -100,12 +100,12 @@ template <int i>
 struct DimGetter {
   // Return a copy if Dim is const
   template <typename D>
-  HOSTDEVICE static int impl(const D& d) {
+  HOSTDEVICE static int64_t impl(const D& d) {
     return DimGetter<i - 1>::impl(d.tail);
   }
   // Return a reference if Dim is mutable
   template <typename D>
-  HOSTDEVICE static int& impl(D& d) {
+  HOSTDEVICE static int64_t& impl(D& d) {
     return DimGetter<i - 1>::impl(d.tail);
   }
 };
@@ -115,18 +115,18 @@ template <>
 struct DimGetter<0> {
   // Return a copy if Dim is const
   template <typename D>
-  HOSTDEVICE static int impl(const D& d) {
+  HOSTDEVICE static int64_t impl(const D& d) {
     return d.head;
   }
   // Return a reference if Dim is mutable
   template <typename D>
-  HOSTDEVICE static int& impl(D& d) {
+  HOSTDEVICE static int64_t& impl(D& d) {
     return d.head;
   }
 };
 
 template <int D>
-HOSTDEVICE int& indexer(Dim<D>& dim, int idx) {
+HOSTDEVICE int64_t& indexer(Dim<D>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx < 0) {
     throw std::invalid_argument("Tried to access a negative dimension");
@@ -141,7 +141,7 @@ HOSTDEVICE int& indexer(Dim<D>& dim, int idx) {
 }
 
 template <>
-HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) {
+HOSTDEVICE int64_t& indexer<1>(Dim<1>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx != 0) {
     throw std::invalid_argument("Invalid index");
@@ -153,7 +153,7 @@ HOSTDEVICE int& indexer<1>(Dim<1>& dim, int idx) {
 }
 
 template <int D>
-HOSTDEVICE int indexer(const Dim<D>& dim, int idx) {
+HOSTDEVICE int64_t indexer(const Dim<D>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx < 0) {
     throw std::invalid_argument("Tried to access a negative dimension");
@@ -168,7 +168,7 @@ HOSTDEVICE int indexer(const Dim<D>& dim, int idx) {
 }
 
 template <>
-HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) {
+HOSTDEVICE int64_t indexer<1>(const Dim<1>& dim, int idx) {
 #ifndef __CUDA_ARCH__
   if (idx != 0) {
     throw std::invalid_argument("Invalid index");
@@ -182,73 +182,76 @@ HOSTDEVICE int indexer<1>(const Dim<1>& dim, int idx) {
 }  // namespace
 // Static access to constant Dim
 template <int i, int l>
-HOSTDEVICE int get(const Dim<l>& d) {
+HOSTDEVICE int64_t get(const Dim<l>& d) {
   return DimGetter<i>::impl(d);
 }
 
 // Static access to mutable Dim
 template <int i, int l>
-HOSTDEVICE int& get(Dim<l>& d) {
+HOSTDEVICE int64_t& get(Dim<l>& d) {
   return DimGetter<i>::impl(d);
 }
 
 // Dynamic access to constant Dim
 template <int l>
-HOSTDEVICE int Dim<l>::operator[](int i) const {
+HOSTDEVICE int64_t Dim<l>::operator[](int i) const {
   return indexer(*this, i);
 }
 
 // Dynamic access to mutable Dim
 template <int l>
-HOSTDEVICE int& Dim<l>::operator[](int i) {
+HOSTDEVICE int64_t& Dim<l>::operator[](int i) {
   return indexer(*this, i);
 }
 
 // Dynamic access to constant Dim
-inline HOSTDEVICE int Dim<1>::operator[](int i) const {
+inline HOSTDEVICE int64_t Dim<1>::operator[](int i) const {
   return indexer(*this, i);
 }
 
 // Dynamic access to mutable Dim
-inline HOSTDEVICE int& Dim<1>::operator[](int i) { return indexer(*this, i); }
+inline HOSTDEVICE int64_t& Dim<1>::operator[](int i) {
+  return indexer(*this, i);
+}
 
 // Dynamic access to constant Dim
 // without std::enable_if will try to instantiate this on get<0>(d)
 template <int l>
-HOSTDEVICE typename std::enable_if<(l > 0), int>::type get(const Dim<l>& d,
+HOSTDEVICE typename std::enable_if<(l > 0), int64_t>::type get(const Dim<l>& d,
                                                                int i) {
   return d[i];
 }
 
 // Dynamic access to mutable Dim
 template <int l>
-HOSTDEVICE typename std::enable_if<(l > 0), int&>::type get(Dim<l>& d, int i) {
+HOSTDEVICE typename std::enable_if<(l > 0), int64_t&>::type get(Dim<l>& d,
+                                                                int i) {
   return d[i];
 }
 
 // Dot product of two dims
 template <int i>
-HOSTDEVICE int linearize(const Dim<i>& a, const Dim<i>& b) {
+HOSTDEVICE int64_t linearize(const Dim<i>& a, const Dim<i>& b) {
   return a.head * b.head + linearize(a.tail, b.tail);
 }
 
 // Base case dot product of two Dims
 // Notice it is inline because it is no longer a template
 template <>
-HOSTDEVICE inline int linearize(const Dim<1>& a, const Dim<1>& b) {
+HOSTDEVICE inline int64_t linearize(const Dim<1>& a, const Dim<1>& b) {
   return a.head * b.head;
 }
 
 // Product of a Dim
 template <int i>
-HOSTDEVICE int product(const Dim<i>& a, int prod = 1) {
+HOSTDEVICE int64_t product(const Dim<i>& a, int prod = 1) {
   return prod * a.head * product(a.tail);
 }
 
 // Base case product of a Dim
 // Notice it is inline because it is no longer a template
 template <>
-HOSTDEVICE inline int product(const Dim<1>& a, int prod) {
+HOSTDEVICE inline int64_t product(const Dim<1>& a, int prod) {
   return prod * a.head;
 }
 

paddle/framework/dim_test.cu

@@ -8,7 +8,7 @@ __global__ void test(paddle::framework::Dim<2>* o) {
   o[0] = paddle::framework::make_dim(5, 6);
 }
 
-__global__ void dyn_idx_gpu(int* o) {
+__global__ void dyn_idx_gpu(int64_t* o) {
   auto d = paddle::framework::make_dim(5, 6);
   o[0] = d[1];
 }
@@ -47,9 +47,9 @@ TEST(Dim, Equality) {
   EXPECT_EQ(b[1], 11);
 
   // dynamic access on GPU
-  thrust::device_vector<int> r(1);
+  thrust::device_vector<int64_t> r(1);
   dyn_idx_gpu<<<1, 1>>>(thrust::raw_pointer_cast(r.data()));
-  int res = r[0];
+  int64_t res = r[0];
   EXPECT_EQ(res, 6);
 
   // ex_prefix_mul

paddle/framework/eigen.h

@@ -28,7 +28,7 @@ struct EigenDim {
   static Type From(const DDim& dims) {
     PADDLE_ENFORCE(arity(dims) == D, "D must match arity(DDim)");
     Type ret;
-    for (int d = 0; d < arity(dims); d++) {
+    for (int64_t d = 0; d < arity(dims); d++) {
       ret[d] = dims[d];
     }
     return ret;
@@ -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

@@ -22,8 +22,14 @@ enum AttrType {
   INTS = 3;
   FLOATS = 4;
   STRINGS = 5;
+  INT_PAIRS = 6;
 }
 
+message IntPair {
+  required int32 first = 1;
+  required int32 second = 2;
+};
+
 // OpDesc describes an instance of a C++ framework::OperatorBase
 // derived class type.
 message OpDesc {
@@ -37,6 +43,7 @@ message OpDesc {
     repeated int32 ints = 6;
     repeated float floats = 7;
     repeated string strings = 8;
+    repeated IntPair int_pairs = 9;
   };
 
   message Var {
@@ -80,3 +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

@@ -18,8 +18,10 @@
 #ifndef PADDLE_ONLY_CPU
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
+#include <thrust/system/cuda/experimental/pinned_allocator.h>
 #endif
 
+#include <glog/logging.h>
 #include "paddle/framework/ddim.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/enforce.h"
@@ -32,37 +34,35 @@ template <typename T>
 using Vector = std::vector<T>;
 #else
 template <typename T>
-using Vector = thrust::host_vector<T>;
+using Vector = thrust::host_vector<
+    T, thrust::system::cuda::experimental::pinned_allocator<T>>;
 #endif
 
-using LOD = std::vector<Vector<size_t>>;
+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 Tensor {
  public:
-  LODTensor() {}
-  LODTensor(const LOD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
+  LoDTensor() {}
 
-  void set_lod(const LOD& lod) { lod_ = lod; }
+  explicit LoDTensor(const LoD& lod) : lod_(lod) {}
 
-  void set_tensor(Tensor* tensor) { tensor_ = tensor; }
+  void set_lod(const LoD& lod) { lod_ = lod; }
 
-  Tensor& tensor() { return *tensor_; }
-
-  LOD lod() { return lod_; }
+  LoD lod() const { 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(); }
@@ -100,8 +100,7 @@ class LODTensor {
   void SliceInLevel(size_t level, size_t elem_begin, size_t elem_end);
 
  private:
-  LOD lod_;
-  Tensor* tensor_;  // not owned
+  LoD lod_;
 };
 }  // namespace framework
 }  // namespace paddle

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
@@ -29,76 +29,71 @@ 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});
 
     ASSERT_EQ(lod.size(), 3UL);
 
-    tensor.Resize({20 /*batch size*/, 128 /*dim*/});
+    lod_tensor_.Resize({20 /*batch size*/, 128 /*dim*/});
     // malloc memory
-    tensor.mutable_data<float>(place);
+    lod_tensor_.mutable_data<float>(place);
 
-    lod_tensor.set_lod(lod);
-    lod_tensor.set_tensor(&tensor);
+    lod_tensor_.set_lod(lod);
   }
 
  protected:
   platform::CPUPlace place;
-  Tensor tensor;
-  LODTensor lod_tensor;
+  LoDTensor lod_tensor_;
 };
 
-TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
+TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor_.NumLevels(), 3UL); }
 
-TEST_F(LODTensorTester, NumElements) {
-  ASSERT_EQ(lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(lod_tensor.NumElements(2), 8UL);
+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));
-    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-              lod_tensor.tensor().data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    LODTensor new_lod_tensor = lod_tensor;
+    LoDTensor new_lod_tensor = lod_tensor_;
     new_lod_tensor.SliceLevels(level, level + 2);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
-    ASSERT_EQ(new_lod_tensor.NumElements(1), lod_tensor.NumElements(level + 1));
-    ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-              lod_tensor.tensor().data<float>());
+    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
+    ASSERT_EQ(new_lod_tensor.NumElements(1),
+              lod_tensor_.NumElements(level + 1));
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
   }
 }
 
-TEST_F(LODTensorTester, SliceInLevel) {
+TEST_F(LoDTensorTester, SliceInLevel) {
   size_t level = 0;
-  LODTensor new_lod_tensor = lod_tensor;
+  LoDTensor new_lod_tensor = lod_tensor_;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
   EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
   EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
   EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-            lod_tensor.tensor().data<float>());
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor;
+  new_lod_tensor = lod_tensor_;
   new_lod_tensor.SliceInLevel(level, 0, 2);
   ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
-            lod_tensor.tensor().data<float>());
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
 }
 
 }  // namespace framework

paddle/framework/lod_tensor_test.cu

+/*
+  Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+  Licensed under the Apache License, Version 2.0 (the "License");
+  you may not use this file except in compliance with the License.
+  You may obtain a copy of the License at
+  http://www.apache.org/licenses/LICENSE-2.0
+  Unless required by applicable law or agreed to in writing, software
+  distributed under the License is distributed on an "AS IS" BASIS,
+  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+  See the License for the specific language governing permissions and
+  limitations under the License.
+*/
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/platform/assert.h"
+
+#include <gtest/gtest.h>
+
+__global__ void test(size_t* a, int size) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < size;
+       i += blockDim.x * gridDim.x) {
+    a[i] *= 2;
+  }
+}
+
+TEST(LoDTensor, LoDInGPU) {
+  paddle::framework::LoDTensor lod_tensor;
+  paddle::platform::GPUPlace place(0);
+
+  paddle::framework::LoD src_lod;
+  src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});
+
+  lod_tensor.Resize({14, 16});
+  lod_tensor.mutable_data<float>(place);
+
+  lod_tensor.set_lod(src_lod);
+  CHECK_EQ(lod_tensor.lod_element(0, 2), 4);
+  CHECK_EQ(lod_tensor.lod_element(0, 4), 8);
+
+  auto lod = lod_tensor.lod();
+
+  test<<<1, 8>>>(lod[0].data(), lod[0].size());
+  cudaDeviceSynchronize();
+
+  for (size_t i = 0; i < src_lod[0].size(); ++i) {
+    CHECK_EQ(lod[0].data()[i], src_lod[0].data()[i] * 2);
+  }
+}

paddle/framework/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) {
@@ -172,38 +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
-
-class TestAttrProtoMaker : public pd::OpProtoAndCheckerMaker {
- public:
-  TestAttrProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddAttr<float>("scale", "scale of test op");
-    AddAttr<float>("scale", "scale of test op");
-  }
-};
-
-TEST(ProtoMaker, DuplicatedAttr) {
-  pd::OpProto op_proto;
-  pd::OpAttrChecker op_checker;
-  auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker);
-  ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
-}
-
-class TestInOutProtoMaker : public pd::OpProtoAndCheckerMaker {
- public:
-  TestInOutProtoMaker(pd::OpProto* proto, pd::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("input", "input of test op");
-    AddInput("input", "input of test op");
-  }
-};
-
-TEST(ProtoMaker, DuplicatedInOut) {
-  pd::OpProto op_proto;
-  pd::OpAttrChecker op_checker;
-  auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
-  ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
-}

paddle/framework/operator.cc

@@ -22,14 +22,14 @@ namespace framework {
 template <>
 Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
     platform::CPUPlace, Eigen::DefaultDevice>() const {
-  return *device_context_->get_eigen_device<Eigen::DefaultDevice>();
+  return *device_context_.get_eigen_device<Eigen::DefaultDevice>();
 }
 
 #ifndef PADDLE_ONLY_CPU
 template <>
 Eigen::GpuDevice&
 ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
-  return *device_context_->get_eigen_device<Eigen::GpuDevice>();
+  return *device_context_.get_eigen_device<Eigen::GpuDevice>();
 }
 #endif
 
@@ -123,6 +123,15 @@ OperatorBase::OperatorBase(const std::string& type,
   CheckAllInputOutputSet();
 }
 
+std::vector<std::string> OperatorBase::InputVars() const {
+  std::vector<std::string> ret_val;
+  for (auto& o : outputs_) {
+    ret_val.reserve(ret_val.size() + o.second.size());
+    ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
+  }
+  return ret_val;
+}
+
 std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
   std::vector<std::string> ret_val;
   if (has_intermediate) {
@@ -177,6 +186,48 @@ void OperatorBase::GenerateTemporaryNames() {
   }
 }
 
+template <>
+const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const {
+  auto* var = InputVar(name);
+  return var == nullptr ? nullptr : GetTensorFromVar(var);
+}
+
+template <>
+const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+    const std::string& name) const {
+  auto names = op().Inputs(name);
+  std::vector<const Tensor*> res;
+  res.reserve(names.size());
+  std::transform(names.begin(), names.end(), std::back_inserter(res),
+                 [&](const std::string& sub_name) {
+                   auto var = scope_.FindVar(sub_name);
+                   return var == nullptr ? nullptr : GetTensorFromVar(var);
+                 });
+  return res;
+}
+
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
+  auto* var = OutputVar(name);
+  return var == nullptr ? nullptr : const_cast<Tensor*>(GetTensorFromVar(var));
+}
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const {
+  auto names = op().Outputs(name);
+  std::vector<Tensor*> res;
+  res.reserve(names.size());
+  std::transform(names.begin(), names.end(), std::back_inserter(res),
+                 [&](const std::string& sub_name) {
+                   auto var = scope().FindVar(sub_name);
+                   return var == nullptr
+                              ? nullptr
+                              : const_cast<Tensor*>(GetTensorFromVar(var));
+                 });
+  return res;
+}
+
 void OpProtoAndCheckerMaker::Validate() {
   validated_ = true;
   CheckNoDuplicatedInOutAttrs();

paddle/framework/operator.h

@@ -22,6 +22,7 @@ limitations under the License. */
 #include "op_info.h"
 #include "paddle/framework/attribute.h"
 #include "paddle/framework/framework.pb.h"
+#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
@@ -69,7 +70,7 @@ 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));
@@ -94,11 +95,14 @@ class OperatorBase {
 
   const VariableNameMap& Inputs() const { return inputs_; }
   const VariableNameMap& Outputs() const { return outputs_; }
+
   //! Get a input with argument's name described in `op_proto`
   std::string Input(const std::string& name) const;
   //! Get a input which has multiple variables.
   const std::vector<std::string>& Inputs(const std::string& name) const;
 
+  std::vector<std::string> InputVars() const;
+
   //! Get a output with argument's name described in `op_proto`
   std::string Output(const std::string& name) const;
   //! Get an output which has multiple variables.
@@ -238,8 +242,8 @@ 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 {
@@ -311,9 +315,9 @@ class InferShapeContext {
   }
 
   template <typename T>
-  std::vector<const T*> MultiOutput(const std::string& name) const {
+  std::vector<T*> MultiOutput(const std::string& name) const {
     auto names = op_.Outputs(name);
-    std::vector<const T*> res;
+    std::vector<T*> res;
     res.reserve(names.size());
     std::transform(names.begin(), names.end(), std::back_inserter(res),
                    [&](const std::string& sub_name) {
@@ -323,11 +327,27 @@ class InferShapeContext {
     return res;
   }
 
+  const Tensor* GetTensorFromVar(const Variable* var) const {
+    if (var->IsType<LoDTensor>()) {
+      return &var->Get<LoDTensor>();
+    }
+    PADDLE_ENFORCE(var->IsType<Tensor>(),
+                   "The Input(%s) must be LoDTensor or Tensor.");
+    return &var->Get<Tensor>();
+  }
+
  private:
   const OperatorBase& op_;
   const Scope& scope_;
 };
 
+template <>
+const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const;
+
+template <>
+const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
+    const std::string& name) const;
+
 template <typename T>
 struct EigenDeviceConverter;
 
@@ -346,7 +366,7 @@ struct EigenDeviceConverter<platform::GPUPlace> {
 class ExecutionContext : public InferShapeContext {
  public:
   ExecutionContext(const OperatorBase& op, const Scope& scope,
-                   const platform::DeviceContext* device_context)
+                   const platform::DeviceContext& device_context)
       : InferShapeContext(op, scope), device_context_(device_context) {}
 
   template <typename PlaceType,
@@ -354,15 +374,44 @@ class ExecutionContext : public InferShapeContext {
                 typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
   DeviceType& GetEigenDevice() const;
 
-  platform::Place GetPlace() const { return device_context_->GetPlace(); }
+  platform::Place GetPlace() const { return device_context_.GetPlace(); }
 
-  const platform::DeviceContext* device_context() const {
+  const platform::DeviceContext& device_context() const {
     return device_context_;
   }
 
-  const platform::DeviceContext* device_context_;
+  // redefine Output function,
+  // use Variable::Get instead of Variable::GetMutable
+  template <typename T>
+  T* Output(const std::string& name) const {
+    auto var = OutputVar(name);
+    return var == nullptr ? nullptr : const_cast<T*>(&var->Get<T>());
+  }
+
+  // redefine MultiOutput function.
+  // use Variable::Get instead of Variable::GetMutable
+  template <typename T>
+  std::vector<T*> MultiOutput(const std::string& name) const {
+    auto names = op().Outputs(name);
+    std::vector<T*> res;
+    res.reserve(names.size());
+    std::transform(
+        names.begin(), names.end(), std::back_inserter(res),
+        [&](const std::string& sub_name) { return Output<T>(sub_name); });
+    return res;
+  }
+
+ private:
+  const platform::DeviceContext& device_context_;
 };
 
+template <>
+Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
+
+template <>
+std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
+    const std::string& name) const;
+
 class OpKernel {
  public:
   /**
@@ -413,7 +462,7 @@ class OperatorWithKernel : public OperatorBase {
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const final {
     auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
-    opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
+    opKernel->Compute(ExecutionContext(*this, scope, dev_ctx));
   }
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&

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");
   }
 };
@@ -264,3 +264,37 @@ TEST(Operator, Clone) {
   auto b = a.Clone();
   ASSERT_EQ(a.Type(), b->Type());
 }
+
+class TestAttrProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
+ public:
+  TestAttrProtoMaker(paddle::framework::OpProto* proto,
+                     paddle::framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddAttr<float>("scale", "scale of test op");
+    AddAttr<float>("scale", "scale of test op");
+  }
+};
+
+TEST(ProtoMaker, DuplicatedAttr) {
+  paddle::framework::OpProto op_proto;
+  paddle::framework::OpAttrChecker op_checker;
+  auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker);
+  ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
+}
+
+class TestInOutProtoMaker : public paddle::framework::OpProtoAndCheckerMaker {
+ public:
+  TestInOutProtoMaker(paddle::framework::OpProto* proto,
+                      paddle::framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("input", "input of test op");
+    AddInput("input", "input of test op");
+  }
+};
+
+TEST(ProtoMaker, DuplicatedInOut) {
+  paddle::framework::OpProto op_proto;
+  paddle::framework::OpAttrChecker op_checker;
+  auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
+  ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
+}
\ No newline at end of file

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;
 
@@ -78,6 +81,9 @@ class Tensor {
   /*! Return the dimensions of the memory block. */
   inline const DDim& dims() const;
 
+  /*! Return the numel of the memory block. */
+  inline int64_t numel() const;
+
   /*! Resize the dimensions of the memory block. */
   inline Tensor& Resize(const DDim& dims);
 
@@ -159,6 +165,12 @@ class Tensor {
   /*! points to dimensions of memory block. */
   DDim dims_;
 
+  /**
+   * A cache of the number of elements in a tensor.
+   * Would be 0 for an uninitialized tensor.
+   */
+  int64_t numel_;
+
   /**
    * @brief   A PlaceHolder may be shared by more than one tensor.
    *

paddle/framework/tensor_impl.h

@@ -22,9 +22,9 @@ namespace framework {
 template <typename T>
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
-      holder_, "Tenosr holds no memory. Call Tensor::mutable_data first.");
+      holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
   PADDLE_ENFORCE_GE(
-      holder_->size(), product(dims_) * sizeof(T) + offset_,
+      holder_->size(), numel() * sizeof(T) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
       "first to re-allocate memory.\n"
       "or maybe the required data-type mismatches the data already stored.");
@@ -54,11 +54,11 @@ inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
 template <typename T>
 inline T* Tensor::mutable_data(platform::Place place) {
   static_assert(std::is_pod<T>::value, "T must be POD");
-  PADDLE_ENFORCE_GT(product(dims_), 0,
+  PADDLE_ENFORCE_GT(numel(), 0,
                     "Tensor's numel must be larger than zero to call "
                     "Tensor::mutable_data. Call Tensor::set_dim first.");
   /* some versions of boost::variant don't have operator!= */
-  size_t size = product(dims_) * sizeof(T);
+  int64_t size = numel() * sizeof(T);
   if (holder_ == nullptr || !(holder_->place() == place) ||
       holder_->size() < size + offset_) {
     if (platform::is_cpu_place(place)) {
@@ -97,7 +97,7 @@ inline void Tensor::CopyFrom(const Tensor& src,
 
   auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
 
-  auto size = product(src.dims_) * sizeof(T);
+  auto size = src.numel() * sizeof(T);
 
   if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) {
     memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
@@ -131,7 +131,7 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   PADDLE_ENFORCE_LT(begin_idx, end_idx,
                     "Begin index must be less than end index.");
   PADDLE_ENFORCE_NE(dims_[0], 1, "Can not slice a tensor with dims_[0] = 1.");
-  int base = product(dims_) / dims_[0];
+  size_t base = numel() / dims_[0];
   Tensor dst;
   dst.holder_ = holder_;
   DDim dst_dims = dims_;
@@ -143,10 +143,21 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
 
 inline Tensor& Tensor::Resize(const DDim& dims) {
   dims_ = dims;
+  numel_ = product(dims_);
   return *this;
 }
 
 inline const DDim& Tensor::dims() const { return dims_; }
 
+inline int64_t Tensor::numel() const { return numel_; }
+
+template <typename T>
+inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) {
+  Tensor res;
+  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

@@ -36,7 +36,7 @@ TEST(Tensor, DataAssert) {
   } catch (paddle::platform::EnforceNotMet err) {
     caught = true;
     std::string msg =
-        "holder_ should not be null\nTenosr holds no memory. Call "
+        "holder_ should not be null\nTensor holds no memory. Call "
         "Tensor::mutable_data first.";
     const char* what = err.what();
     for (size_t i = 0; i < msg.length(); ++i) {
@@ -112,7 +112,7 @@ TEST(Tensor, ShareDataWith) {
     } catch (paddle::platform::EnforceNotMet err) {
       caught = true;
       std::string msg =
-          "holder_ should not be null\nTenosr holds no memory. Call "
+          "holder_ should not be null\nTensor holds no memory. Call "
           "Tensor::mutable_data first.";
       const char* what = err.what();
       for (size_t i = 0; i < msg.length(); ++i) {
@@ -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);
+}

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) {
@@ -114,8 +106,9 @@ void MulOp<DEVICE_TYPE_CPU>(CpuMatrix& out,
                             real scaleT,
                             bool aTrans,
                             bool bTrans) {
-  GEMM(aTrans ? CblasTrans : CblasNoTrans,
-       bTrans ? CblasTrans : CblasNoTrans,
+  BlasGemm<DEVICE_TYPE_CPU, real>::compute(
+      aTrans,
+      bTrans,
       out.getHeight(),
       out.getWidth(),
       !aTrans ? a.getWidth() : a.getHeight(),

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/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/BatchNormBaseLayer.cpp

@@ -62,14 +62,18 @@ void BatchNormBaseLayer::calFeatureMapSize() {
   const ImageConfig& conf = config_.inputs(0).image_conf();
   imageH_ = inputLayers_[0]->getOutput().getFrameHeight();
   imageW_ = inputLayers_[0]->getOutput().getFrameWidth();
+  imageD_ = inputLayers_[0]->getOutput().getFrameDepth();
+
+  if (0 == imageD_) imageD_ = conf.img_size_z();
   if (imageH_ == 0 && imageW_ == 0) {
     imageH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
     imageW_ = conf.img_size();
   } else {
     getOutput().setFrameHeight(imageH_);
     getOutput().setFrameWidth(imageW_);
+    getOutput().setFrameDepth(imageD_);
   }
-  imgPixels_ = imageH_ * imageW_;
+  imgPixels_ = imageH_ * imageW_ * imageD_;
 }
 
 }  // namespace paddle

paddle/gserver/layers/BatchNormBaseLayer.h

@@ -80,6 +80,7 @@ protected:
 
   /// Height or width of input image feature.
   /// Both of them are 1 if the input is fully-connected layer.
+  int imageD_;
   int imageH_;
   int imageW_;
   /// Height * Width.

paddle/gserver/layers/Conv3DLayer.cpp

@@ -83,8 +83,8 @@ void Conv3DLayer::forward(PassType passType) {
   int outWidth = getSize();
   resetOutput(batchSize, outWidth);
 
-  for (size_t i = 0; i != inputLayers_.size(); ++i) {
   REGISTER_TIMER_INFO("FwdConv3D", getName().c_str());
+  for (size_t i = 0; i != inputLayers_.size(); ++i) {
     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);
   }
 
-  for (size_t i = 0; i != inputLayers_.size(); ++i) {
   REGISTER_TIMER_INFO("BwdConv3D", getName().c_str());
+  for (size_t i = 0; i != inputLayers_.size(); ++i) {
     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/CudnnBatchNormLayer.cpp

@@ -37,7 +37,7 @@ bool CudnnBatchNormLayer::init(const LayerMap& layerMap,
 }
 
 void CudnnBatchNormLayer::reshape(int batchSize) {
-  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_, imageW_);
+  hl_tensor_reshape(ioDesc_, batchSize, channels_, imageH_ * imageD_, imageW_);
 }
 
 void CudnnBatchNormLayer::forward(PassType passType) {
@@ -104,7 +104,7 @@ void CudnnBatchNormLayer::forward(PassType passType) {
                                    EPS,
                                    batchSize,
                                    channels_,
-                                   imageH_,
+                                   imageH_ * imageD_,
                                    imageW_);
     }
   }

paddle/gserver/layers/CudnnPoolLayer.cpp

@@ -29,9 +29,9 @@ bool CudnnPoolLayer::typeCheck(const std::string &poolType,
     if (mode) {
       *mode = HL_POOLING_AVERAGE;
     }
-  } else if (poolType == "cudnn-avg-excl-pad-pool") {
+  } else if (poolType == "cudnn-avg-incl-pad-pool") {
     if (mode) {
-      *mode = HL_POOLING_AVERAGE_EXCLUDE_PADDING;
+      *mode = HL_POOLING_AVERAGE_INCLUDE_PADDING;
     }
   } else {
     return false;

paddle/gserver/layers/DetectionOutputLayer.cpp

@@ -139,7 +139,13 @@ void DetectionOutputLayer::forward(PassType passType) {
                                        allDecodedBBoxes,
                                        &allIndices);
 
+  if (numKept > 0) {
     resetOutput(numKept, 7);
+  } else {
+    MatrixPtr outV = getOutputValue();
+    if (outV) outV->resize(0, 0);
+    return;
+  }
   MatrixPtr outV = getOutputValue();
   getDetectionOutput(confBuffer_->getData(),
                      numKept,

paddle/gserver/layers/DetectionUtil.cpp

@@ -469,7 +469,7 @@ size_t getDetectionIndices(
     const size_t numClasses,
     const size_t backgroundId,
     const size_t batchSize,
-    const size_t confThreshold,
+    const real confThreshold,
     const size_t nmsTopK,
     const real nmsThreshold,
     const size_t keepTopK,

paddle/gserver/layers/DetectionUtil.h

@@ -275,7 +275,7 @@ size_t getDetectionIndices(
     const size_t numClasses,
     const size_t backgroundId,
     const size_t batchSize,
-    const size_t confThreshold,
+    const real confThreshold,
     const size_t nmsTopK,
     const real nmsThreshold,
     const size_t keepTopK,