Merge branch 'develop' into add_sequence_slice_layer

.clang_format.hook

+#!/bin/bash
+set -e
+
+readonly VERSION="3.8"
+
+version=$(clang-format -version)
+
+if ! [[ $version == *"$VERSION"* ]]; then
+    echo "clang-format version check failed."
+    echo "a version contains '$VERSION' is needed, but get '$version'"
+    echo "you can install the right version, and make an soft-link to '\$PATH' env"
+    exit -1
+fi
+
+clang-format $@

.pre-commit-config.yaml

@@ -19,10 +19,10 @@
     -   id: end-of-file-fixer
 -   repo: local
     hooks:
-    -   id: clang-format
+    -   id: clang-format-with-version-check
         name: clang-format
         description: Format files with ClangFormat.
-        entry: clang-format -i
+        entry: ./.clang_format.hook -i
         language: system
         files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
 -   repo: https://github.com/PaddlePaddle/pre-commit-golang

.travis.yml

@@ -37,8 +37,8 @@ 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 numpy wheel 'protobuf==3.1' sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit requests==2.9.2 LinkChecker
-  - pip install rarfile nltk==3.2.2 scipy==0.19.0 recordio matplotlib Pillow
+  - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
+  - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
   - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
   - go get -u github.com/alecthomas/gometalinter

CMakeLists.txt

@@ -36,8 +36,8 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    OFF)
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      OFF)
+option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
+option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
@@ -55,6 +55,7 @@ option(WITH_C_API       "Compile PaddlePaddle with C-API(Prediction)"   OFF)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
+option(USE_EIGEN_FOR_BLAS   "Use matrix multiplication in Eigen"        OFF)
 
 # CMAKE_BUILD_TYPE
 if(NOT CMAKE_BUILD_TYPE)
@@ -137,9 +138,9 @@ set(EXTERNAL_LIBS
 )
 
 if(WITH_GPU)
-    list(APPEND EXTERNAL_LIB ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
+    list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
     if(NOT WITH_DSO)
-        list(APPEND EXTERNAL_LIB ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
+        list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
     endif(NOT WITH_DSO)
 endif(WITH_GPU)
 

Dockerfile

@@ -34,9 +34,6 @@ RUN apt-get update && \
     net-tools && \
     apt-get clean -y
 
-# paddle is using numpy.flip, which is introduced since 1.12.0
-RUN pip --no-cache-dir install 'numpy>=1.12.0'
-
 # Install Go and glide
 RUN wget -qO- https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
     tar -xz -C /usr/local && \
@@ -58,33 +55,22 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
 # FIXME: due to temporary ipykernel dependency issue, specify ipykernel jupyter
 # version util jupyter fixes this issue.
 RUN pip install --upgrade pip && \
-    pip install -U 'protobuf==3.1.0' && \
-    pip install -U wheel pillow BeautifulSoup && \
+    pip install -U wheel && \
     pip install -U docopt PyYAML sphinx && \
-    pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
-    pip install pre-commit 'requests==2.9.2' 'ipython==5.3.0' && \
+    pip install -U sphinx-rtd-theme==0.1.9 recommonmark
+
+RUN pip install pre-commit 'ipython==5.3.0' && \
     pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
-    pip install opencv-python rarfile 'scipy>=0.19.0' 'nltk>=3.2.2'
+    pip install opencv-python
+
+COPY ./python/requirements.txt /root/
+RUN pip install -r /root/requirements.txt
 
 # To fix https://github.com/PaddlePaddle/Paddle/issues/1954, we use
 # the solution in https://urllib3.readthedocs.io/en/latest/user-guide.html#ssl-py2
 RUN apt-get install -y libssl-dev libffi-dev
 RUN pip install certifi urllib3[secure]
 
-# TODO(qijun) The template library Eigen doesn't work well with GCC 5 
-# coming with the default Docker image, so we switch to use GCC 4.8 
-# by default. And I will check Eigen library later.
-
-RUN ln -sf gcc-4.8 /usr/bin/gcc && \
-    ln -sf gcc-ar-4.8 /usr/bin/gcc-ar && \
-    ln -sf gcc-nm-4.8 /usr/bin/gcc-nm && \
-    ln -sf gcc-ranlib-4.8 /usr/bin/gcc-ranlib && \
-    ln -sf gcc-4.8 /usr/bin/x86_64-linux-gnu-gcc && \
-    ln -sf gcc-ar-4.8 /usr/bin/x86_64-linux-gnu-gcc-ar && \
-    ln -sf gcc-nm-4.8 /usr/bin/x86_64-linux-gnu-gcc-nm && \
-    ln -sf gcc-ranlib-4.8 /usr/bin/x86_64-linux-gnu-gcc-ranlib && \
-    ln -sf g++-4.8 /usr/bin/g++ && \
-    ln -sf g++-4.8 /usr/bin/x86_64-linux-gnu-g++ 
 
 # Install woboq_codebrowser to /woboq
 RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \

cmake/configure.cmake

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

cmake/cudnn.cmake

@@ -2,7 +2,7 @@ if(NOT WITH_GPU)
     return()
 endif()
 
-set(CUDNN_ROOT "" CACHE PATH "CUDNN ROOT")
+set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT")
 find_path(CUDNN_INCLUDE_DIR cudnn.h
     PATHS ${CUDNN_ROOT} ${CUDNN_ROOT}/include
     $ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/include ${CUDA_TOOLKIT_INCLUDE}

cmake/external/openblas.cmake

@@ -73,10 +73,18 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
 # linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
 SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
 FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
-ADD_LIBRARY(cblas STATIC ${dummyfile})
+IF(${CBLAS_PROVIDER} MATCHES MKL)
+    ADD_LIBRARY(cblas SHARED ${dummyfile})
+ELSE()
+    ADD_LIBRARY(cblas STATIC ${dummyfile})
+ENDIF()
 TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
 
 IF(NOT ${CBLAS_FOUND})
     ADD_DEPENDENCIES(cblas extern_openblas)
     LIST(APPEND external_project_dependencies cblas)
+ELSE()
+    IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
+        ADD_DEPENDENCIES(cblas mklml)
+    ENDIF()
 ENDIF(NOT ${CBLAS_FOUND})

cmake/flags.cmake

@@ -9,13 +9,6 @@ function(CheckCompilerCXX11Flag)
         if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
             message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
         endif()
-        if(NOT ANDROID)
-            # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
-            # Use Debug mode instead for now.
-            if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9)
-                set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
-            endif()
-        endif()
     elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
         # cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang"
         # Apple Clang is a different compiler than upstream Clang which havs different version numbers.
@@ -160,7 +153,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
 
 # Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
 # So, don't set these flags here.
-LIST(APPEND CUDA_NVCC_FLAGS -std=c++11 --default-stream per-thread)
+LIST(APPEND CUDA_NVCC_FLAGS -std=c++11)
 LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math)
 
 if(CMAKE_BUILD_TYPE  STREQUAL "Debug")

doc/api/v2/config/layer.rst

@@ -367,6 +367,11 @@ trans
 ..  autoclass:: paddle.v2.layer.trans
     :noindex:
 
+scale_shift
+-----------
+..  autoclass:: paddle.v2.layer.scale_shift
+    :noindex:
+
 Sampling Layers
 ===============
 

doc/design/auto_gradient_check.md

+## Auto Gradient Checker Design
+
+## Backgraound：
+- Operator forward computing is easy to check if the result is right because it has a clear definition. **But** backpropagation is a notoriously difficult algorithm to debug and get right:
+  - 1. you should get the right backpropagation formula according to the forward computation.
+  - 2. you should implement it right in CPP.
+  - 3. it's difficult to prepare test data.
+
+- Auto gradient check gets a numeric gradient by forward Operator and use it as a reference of the backward Operator's result. It has several advantages:
+  - 1. numeric gradient checker only need forward operator.
+  - 2. user only need to prepare the input data for forward Operator.
+
+## Mathematical Theory
+The following two document from stanford has a detailed explanation of how to get numeric gradient and why it's useful.
+
+- [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization)
+- [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96)
+
+
+## Numeric Gradient Implementation
+### Python Interface
+```python
+def get_numeric_gradient(op,
+                         input_values,
+                         output_name,
+                         input_to_check,
+                         delta=0.005,
+                         local_scope=None):
+    """
+    Get Numeric Gradient for an operator's input.
+
+    :param op: C++ operator instance, could be an network
+    :param input_values: The input variables. Should be an dictionary, key is
+    variable name. Value is numpy array.
+    :param output_name: The final output variable name.
+    :param input_to_check: The input variable need to get gradient.
+    :param delta: The perturbation value for numeric gradient method. The
+    smaller delta is, the more accurate result will get. But if that delta is
+     too small, it could occur numerical stability problem.
+    :param local_scope: The local scope used for get_numeric_gradient.
+    :return: The gradient array in numpy format.
+    """
+```
+
+### Explaination:
+
+- Why need `output_name`
+  - One Operator may have multiple Output, you can get independent gradient from each Output. So user should set one output to calculate.
+
+- Why need `input_to_check`
+  - One operator may have multiple inputs. Gradient Op can calculate the gradient of these Inputs at the same time. But Numeric Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times.
+
+
+### Core Algorithm Implementation
+
+
+```python
+    # we only compute gradient of one element each time.
+    # we use a for loop to compute the gradient of every element.
+    for i in xrange(tensor_size):
+        # get one input element throw it's index i.
+        origin = tensor_to_check.get_float_element(i)
+
+        # add delta to it, run op and then get the sum of the result tensor.
+        x_pos = origin + delta
+        tensor_to_check.set_float_element(i, x_pos)
+        y_pos = get_output()
+
+        # plus delta to this element, run op and get the sum of the result tensor.
+        x_neg = origin - delta
+        tensor_to_check.set_float_element(i, x_neg)
+        y_neg = get_output()
+
+        # restore old value
+        tensor_to_check.set_float_element(i, origin)
+
+        # compute the gradient of this element and store it into a numpy array.
+        gradient_flat[i] = (y_pos - y_neg) / delta / 2
+
+    # reshape the gradient result to the shape of the source tensor.
+    return gradient_flat.reshape(tensor_to_check.get_dims())
+```
+
+## Auto Graident Checker Framework
+
+Each Operator Kernel has three kinds of Gradient:
+
+- 1. Numeric Gradient
+- 2. CPU Operator Gradient
+- 3. GPU Operator Gradient(if supported)
+
+Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as the reference value.
+
+- 1. calculate the numeric gradient.
+- 2. calculate CPU kernel Gradient with the backward Operator and compare it with the numeric gradient.
+- 3. calculate GPU kernel Gradient with the backward Operator and compare it with the numeric gradient.(if support GPU)
+
+#### Python Interface
+
+```python
+    def check_grad(self,
+                   forward_op,
+                   input_vars,
+                   inputs_to_check,
+                   output_name,
+                   no_grad_set=None,
+                   only_cpu=False,
+                   max_relative_error=0.005):
+        """
+        :param forward_op: used to create backward_op
+        :param input_vars: numpy value of input variable. The following
+            computation will use these variables.
+        :param inputs_to_check: inputs var names that should check gradient.
+        :param output_name: output name that used to
+        :param max_relative_error: The relative tolerance parameter.
+        :param no_grad_set: used when create backward ops
+        :param only_cpu: only compute and check gradient on cpu kernel.
+        :return:
+        """
+```
+
+### How to check if two numpy array is close enough?
+if `abs_numeric_grad` is nearly zero, then use abs error for numeric_grad, not relative
+
+```python
+numeric_grad = ...
+operator_grad = numpy.array(scope.find_var(grad_var_name(name)).get_tensor())
+
+abs_numeric_grad = numpy.abs(numeric_grad)
+# if abs_numeric_grad is nearly zero, then use abs error for numeric_grad, not relative
+# error.
+abs_numeric_grad[abs_numeric_grad < 1e-3] = 1
+
+diff_mat = numpy.abs(abs_numeric_grad - operator_grad) / abs_numeric_grad
+max_diff = numpy.max(diff_mat)
+```
+
+
+#### Notes：
+1，The Input data for auto gradient checker should be reasonable to avoid numeric problem.
+
+
+#### Refs:
+
+- [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization)
+- [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96)

doc/design/cluster_train/large_model_dist_train.md

+# Alalysis of large model distributed training in Paddle
+
+***NOTE: This is only some note for how we implemeted this scheme in V1, not a new design.***
+
+## What is it
+
+We often encounter cases that the embedding layer parameters(sparse) are so large that we can not store it in the trainer's memory when training. So we need to put them to several servers, and fetch them row by row instead of fetch all of the parameters.
+
+## How to use
+
+Specify command-line argument like  `--loadsave_parameters_in_pserver=true --ports_num_for_sparse=1  --use_old_updater=1` when starting the paddle trainer. And also add something like `--ports_num_for_sparse=1 --pserver_num_threads=5` when starting pserver processes.
+
+Accrodingly, configure your embedding layers like:
+
+```python
+SPARSE_REMOTE=True
+
+w1 = data_layer(name="w1", size=dict_size)
+emb1 = embedding_layer(input=w1, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE))
+w2 = data_layer(name="w2", size=dict_size)
+emb2 = embedding_layer(input=w2, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE))
+...
+```
+
+## Implementation details
+
+```c++
+enum MatType {
+  MAT_NORMAL,
+  MAT_NORMAL_SHARED,
+  MAT_VALUE_SHARED,
+  MAT_SPARSE_ROW_IDS,
+  MAT_SPARSE_ROW_AUTO_GROW,
+  MAT_CACHE_ROW,
+  MAT_SPARSE_ROW,
+  MAT_SPARSE_ROW_PREFETCH,
+  MAT_SPARSE_ROW_PREFETCH_FULL_SIZE,
+};
+```
+
+`MAT_SPARSE_ROW_PREFETCH` is what we use when configured to fetch only row of matrix when training.
+
+In `trainer_internal.cpp:L93 trainOneBatch`:
+
+```c++
+  if (config_->getOptConfig().use_sparse_remote_updater()) {
+    REGISTER_TIMER("prefetch");
+    gradientMachine_->prefetch(inArgs);
+    parameterUpdater_->getParametersRemote();
+  }
+```
+
+When doing actual network forward and backward, at the beginning of each batch, the trainer will try to download one row of data from pserver.
+
+In `trainer/RemoteParameterUpdater.cpp`: `parameterUpdater_->getParametersRemote();`:
+
+```c++
+if (fullSize) {
+    ...
+} else {
+getParams = [&] {
+    parameterClient_->getParameterSparse(
+        /* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType);
+};
+applyL1 = [](Parameter& para, real decayRate) {
+    para.getMat(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate);
+};
+}
+```
+
+Calling `parameterClient_->getParameterSparse` will do remote call to pserver's `getParameterSparse`:
+
+```c++
+void ParameterServer2::getParameterSparse(const SendParameterRequest& request,
+                                          std::vector<Buffer>& inputBuffers,
+                                          SendParameterResponse* response,
+                                          std::vector<Buffer>* outputBuffers) {
+  (void)inputBuffers;
+  auto& buffer = *readWriteBuffer_;
+  size_t numReals = 0;
+  for (const auto& block : request.blocks()) {
+    numReals += getParameterConfig(block).dims(1);
+  }
+  buffer.resize(numReals);
+
+  VLOG(3) << "pserver: getParameterSparse, numReals=" << numReals;
+
+  ReadLockGuard guard(parameterMutex_);
+  size_t offset = 0;
+  for (const auto& block : request.blocks()) {
+    size_t width = getParameterConfig(block).dims(1);
+    Buffer buf = {buffer.data() + offset, width};
+    int type = request.send_back_parameter_type();
+    sendBackParameterSparse(block, type, response, &buf, width, outputBuffers);
+    offset += width;
+  }
+}
+```
+
+`getParameterConfig(block).dims(1)` returns the width of the current "parameter block"(a shard of parameter object),
+then `getParameterSparse` remote call returns only one row of data to the client.

doc/design/mkldnn/README.MD

@@ -101,6 +101,7 @@ if use_mkldnn
 5. 在**Argument**里添加两个`MkldnnMatrixPtr`，取名为`mkldnnValue`和`mkldnnGrad`，用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名)，用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。
 6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑，用于判断`deviceId`，并针对device在MKL-DNN和CPU之间不统一的情况，做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。
 7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。
+8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况，所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面，一直保存的是0，所以可以充分利用这个信息，定义一个枚举处理所有MKLDNN的参数格式，从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。
 
 ## References
 

doc/getstarted/build_and_install/build_from_source_en.md

@@ -68,7 +68,7 @@ As a simple example, consider the following:
 
 1. **BLAS Dependencies(optional)**
   
-    CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically.
+    CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically.
     To utilize preinstalled BLAS， you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
 
     ```bash
@@ -131,9 +131,9 @@ As a simple example, consider the following:
     To build GPU version, you will need the following installed:
 
         1. a CUDA-capable GPU
-        2. A supported version of Linux with a gcc compiler and toolchain
+        2. A supported version of Linux with a GCC compiler and toolchain
         3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
-        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+        4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
 
     The CUDA development environment relies on tight integration with the host development environment,
     including the host compiler and C runtime libraries, and is therefore only supported on
@@ -172,6 +172,7 @@ export PATH=<path to install>/bin:$PATH
 # install PaddlePaddle Python modules.
 sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
 ```
+
 ## <span id="centos">Build on Centos 7</span>
 
 ### Install Dependencies
@@ -192,9 +193,9 @@ sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
     To build GPU version, you will need the following installed:
 
         1. a CUDA-capable GPU
-        2. A supported version of Linux with a gcc compiler and toolchain
+        2. A supported version of Linux with a GCC compiler and toolchain
         3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
-        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+        4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
 
     The CUDA development environment relies on tight integration with the host development environment,
     including the host compiler and C runtime libraries, and is therefore only supported on
@@ -222,7 +223,7 @@ mkdir build && cd build
 ``` 
 
 Finally, you can build and install PaddlePaddle:
-
+  
 ```bash
 # you can add build option here, such as:    
 cmake3 .. -DCMAKE_INSTALL_PREFIX=<path to install>

doc/getstarted/build_and_install/docker_install_cn.rst

@@ -74,13 +74,13 @@ PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以
 
    .. code-block:: bash
 
-      docker run -it --rm paddlepaddle/paddle:0.10.0-dev /bin/bash
+      docker run -it --rm -v $(pwd):/paddle  paddlepaddle/paddle:0.10.0-dev /bin/bash
 
    或者，可以以后台进程方式运行容器：
 
    .. code-block:: bash
 
-      docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:0.10.0-dev
+      docker run -d -p 2202:22 -p 8888:8888 -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /usr/sbin/sshd -D
 
    然后用密码 :code:`root` SSH进入容器：
 

paddle/capi/gradient_machine.cpp

@@ -146,3 +146,19 @@ paddle_error paddle_gradient_machine_randomize_param(
   m->machine->randParameters();
   return kPD_NO_ERROR;
 }
+
+paddle_error paddle_gradient_machine_get_layer_output(
+    paddle_gradient_machine machine,
+    const char* layerName,
+    paddle_arguments args) {
+  auto m = cast(machine);
+  auto out = paddle::capi::cast<paddle::capi::CArguments>(args);
+  if (m == nullptr || layerName == nullptr || out == nullptr ||
+      m->machine == nullptr) {
+    return kPD_NULLPTR;
+  }
+
+  auto layerOutput = m->machine->getLayerOutput(layerName);
+  out->args.push_back(layerOutput);
+  return kPD_NO_ERROR;
+}

paddle/capi/gradient_machine.h

@@ -39,7 +39,11 @@ PD_API paddle_error paddle_gradient_machine_create_for_inference(
 /**
  * @brief Create a gradient machine used for model inference, using config with
  *        parameters which is generated by `paddle merge_model`.
- * @param [out] machine that used for model inference.
+ *        Example:
+ *          paddle merge_model \
+ *                 --model_dir="pass-00000" \
+ *                 --model_file="merged_model.paddle"
+ * @param [out] machine that used for model inference
  * @param [in] mergedModel
  * @param [in] size
  * @return paddle_error
@@ -97,6 +101,18 @@ paddle_gradient_machine_randomize_param(paddle_gradient_machine machine);
 PD_API paddle_error
 paddle_gradient_machine_destroy(paddle_gradient_machine machine);
 
+/**
+ * @brief Get the output of the layer named `layerName`.
+ * @param [in] gradient machine that have run a inference
+ * @param [in] layerName name of specified layer
+ * @param [out] args output of the specified layer
+ * @return paddle_error
+ */
+PD_API paddle_error
+paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine,
+                                         const char* layerName,
+                                         paddle_arguments args);
+
 #ifdef __cplusplus
 }
 #endif

paddle/framework/CMakeLists.txt

@@ -7,7 +7,7 @@ cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context)
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
 cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
-cc_library(lod_tensor SRCS lod_tensor.cc details/lod_tensor.cc DEPS ddim place 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)
 
 cc_test(variable_test SRCS variable_test.cc)
@@ -15,42 +15,41 @@ cc_test(variable_test SRCS variable_test.cc)
 cc_library(scope SRCS scope.cc)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
-proto_library(attribute_proto SRCS attribute.proto)
-proto_library(op_proto SRCS op_proto.proto DEPS attribute_proto)
-proto_library(op_desc SRCS op_desc.proto DEPS attribute_proto)
-cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
-cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
+proto_library(framework_proto SRCS framework.proto)
 
-cc_library(attribute SRCS attribute.cc DEPS op_desc op_proto)
+cc_library(attribute SRCS attribute.cc DEPS framework_proto)
 
-cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope attribute)
+cc_library(operator SRCS operator.cc DEPS framework_proto device_context tensor scope attribute)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
-cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS op_proto operator)
-cc_library(op_registry SRCS op_registry.cc DEPS op_desc grad_op_builder)
+cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator)
+cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op)
 
-py_proto_compile(framework_py_proto SRCS attribute.proto op_proto.proto op_desc.proto)
+py_proto_compile(framework_py_proto SRCS framework.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
 add_custom_command(TARGET framework_py_proto POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E make_directory ${PROJ_ROOT}/python/paddle/v2/framework/proto
-    COMMAND cp *.py ${PROJ_ROOT}/python/paddle/v2/framework/proto/
+    COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto
+    COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto/
     COMMENT "Copy generated python proto into directory paddle/v2/framework/proto."
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 
 cc_library(backward SRCS backward.cc DEPS net_op)
-cc_test(backward_test SRCS backward_test.cc DEPS backward)
+cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context)
 
 if(WITH_PYTHON)
 cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
-    fc_op
     sgd_op
     add_op
+    mul_op
+    rowwise_add_op
+    sigmoid_op
+    softmax_op
     mean_op
     cross_entropy_op
     recurrent_op

paddle/framework/attribute.cc

@@ -44,7 +44,7 @@ AttrType AttrTypeID<std::vector<std::string>>() {
   return STRINGS;
 }
 
-Attribute GetAttrValue(const AttrDesc& attr_desc) {
+Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
   switch (attr_desc.type()) {
     case paddle::framework::AttrType::INT: {
       return attr_desc.i();

paddle/framework/attribute.h

@@ -20,8 +20,7 @@ limitations under the License. */
 #include <unordered_set>
 #include <vector>
 
-#include "paddle/framework/attribute.pb.h"
-#include "paddle/framework/op_desc.pb.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/variant.h"
 
@@ -37,7 +36,7 @@ typedef std::unordered_map<std::string, Attribute> AttributeMap;
 template <typename T>
 AttrType AttrTypeID();
 
-Attribute GetAttrValue(const AttrDesc& attr_desc);
+Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
 
 // check whether a value(attribute) fit a certain limit
 template <typename T>

paddle/framework/backward.cc

@@ -15,31 +15,44 @@
 #include "paddle/framework/backward.h"
 
 #include <list>
+#include <memory>
+
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
+#include "paddle/operators/recurrent_op.h"
 
 namespace paddle {
 namespace framework {
 
-static bool AllInSet(const std::vector<std::string>& names,
-                     const std::string& suffix,
-                     const std::unordered_set<std::string>& set) {
+template <typename Map, typename T>
+static void ForEachVarName(const Map& names, T callback) {
   for (auto& name : names) {
-    if (set.find(name + suffix) == set.end()) {
-      return false;
+    for (auto& n : name.second) {
+      if (callback(n)) return;
     }
   }
-  return true;
 }
 
-static std::shared_ptr<OperatorBase> NOP() {
-  auto net_op = std::make_shared<operators::NetOp>();
-  net_op->type_ = "@NOP@";
+// return whether all the names + suffixes in the set
+static bool AllInSet(
+    const std::map<std::string, std::vector<std::string>>& names,
+    const std::string& suffix, const std::unordered_set<std::string>& set) {
+  bool all_in_set = true;
+  ForEachVarName(names, [&all_in_set, &set, &suffix](const std::string& n) {
+    all_in_set = set.find(n + suffix) != set.end();
+    return !all_in_set;
+  });
+  return all_in_set;
+}
+
+static std::unique_ptr<OperatorBase> NOP() {
+  auto net_op = new operators::NetOp();
+  net_op->SetType("@NOP@");
   net_op->CompleteAddOp();
-  return net_op;
+  return std::unique_ptr<OperatorBase>(net_op);
 }
 
-//  Get backward operator from a forward operator, recursively implementation.
+//  Get backward operator from a forward operator, a recursive implementation.
 //
 //  no_grad_names the gradient variable names without gradient calculating.
 //
@@ -47,122 +60,152 @@ static std::shared_ptr<OperatorBase> NOP() {
 //  BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and
 //  pass `uniq_id` through recursive calling.
 //
-//  returns The backward operator. For simple situation, it is a simple
-//  operator. For complex situation, it is a NetOp.
+//  returns The backward operator. In a simple situation, it may be a simple
+//  operator, in a complex situation, it maybe a NetOp.
 //
 //  See Backward.h for details
-static std::shared_ptr<OperatorBase> BackwardRecursive(
-    const OperatorBase& forwardOp,
-    std::unordered_set<std::string>& no_grad_names, size_t& uniq_id);
-std::shared_ptr<OperatorBase> BackwardRecursive(
+static std::unique_ptr<OperatorBase> BackwardRecursive(
     const OperatorBase& forwardOp,
     std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
   //  If all input gradients of forwarding operator do not need to calculate,
   //  just return an NOP. Not return null ptr because NOP does not take
   //  too much time for calculation, but it is useful for simplifying logic.
-  if (AllInSet(forwardOp.inputs_, kGradVarSuffix, no_grad_names)) {
+  if (AllInSet(forwardOp.Inputs() /*names*/, kGradVarSuffix /*suffix*/,
+               no_grad_names /*set*/)) {
     return NOP();
   }
 
   //  All output gradients of forwarding operator do not need to calculate.
   //  Then all input gradients cannot be computed at all, and we put them into
   //  `no_grad_names` set. Return an NOP.
-  if (AllInSet(forwardOp.outputs_, kGradVarSuffix, no_grad_names)) {
-    for (auto& name : forwardOp.inputs_) {
-      // Mark all input is not need
-      no_grad_names.insert(name + kGradVarSuffix);
-    }
+  if (AllInSet(forwardOp.Outputs() /*names*/, kGradVarSuffix /*suffix*/,
+               no_grad_names /*set*/)) {
+    ForEachVarName(forwardOp.Inputs(),
+                   [&no_grad_names](const std::string& name) -> bool {
+                     no_grad_names.insert(GradVarName(name));
+                     return false;
+                   });
     return NOP();
   }
 
   // Returned gradient network
-  auto net = std::make_shared<operators::NetOp>();
+  auto net = std::unique_ptr<operators::NetOp>(new operators::NetOp());
 
   if (forwardOp.IsNetOp()) {
     // Because forwardOp is a net op, it can static_cast.
     auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
 
     // Map from output gradient variable name to operator's indices in
-    // backward net. That operator generates that variable.
+    // backward net's ops_. That operator generates that variable.
     std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;
 
     size_t local_op_id = 0;
-    // reversely travel forwardNet
+    // reversely travel forwardNet and collect all duplicate outputs.
     for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
          ++it, ++local_op_id) {
-      auto fwd = *it;
+      auto& fwd = *it;
       auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
-      net->AddOp(bwd);
-      for (auto& out : bwd->outputs_) {
-        dup_output_ops[out].emplace_back(local_op_id);
-      }
+      ForEachVarName(bwd->Outputs(),
+                     [&dup_output_ops, local_op_id](const std::string& out) {
+                       dup_output_ops[out].emplace_back(local_op_id);
+                       return false;
+                     });
+      net->AppendOp(std::move(bwd));
     }
     // Get unique ID for this method.
     auto uid = uniq_id++;
     // TODO(dzh): more comment
-    using Pos = std::pair<size_t, std::shared_ptr<OperatorBase>>;
+    // multiple operators which have the same output (y for example) may
+    // overwrite the same y variable when backward, special operations are token
+    // to handle this case. For each duplicate output, rename it to an alias
+    // (original name with a offset), append an `add` op for its operator,
+    // and finally sum all the alias variable to the final output variable y.
+    using Pos = std::pair<size_t, std::unique_ptr<OperatorBase>>;
     std::list<Pos> insert_position;
     for (auto& dup_output_op : dup_output_ops) {
       const std::string& name = dup_output_op.first;
       auto& dup_op = dup_output_op.second;
+      // no duplicate output
       if (dup_op.size() == 1) continue;
-      std::vector<std::string> dup_outputs;
 
+      // process the duplicate outputs
+      std::vector<std::string> dup_outputs;
       for (size_t i = 0; i < dup_op.size(); ++i) {
+        // rename each duplicate output to an alias
         auto op_offset = dup_op[i];
         dup_outputs.push_back(name + "@RENAME@" + std::to_string(uid) + "@" +
                               std::to_string(i));
         net->ops_[op_offset]->Rename(name, dup_outputs.back());
       }
+      // collect all the offset to append `add` op for each alias
       insert_position.push_back(
-          {dup_op.back(),
-           OpRegistry::CreateOp(
-               "add", {dup_outputs}, {name},
-               {{"input_format",
-                 std::vector<int>{0, static_cast<int>(dup_outputs.size())}}})});
+          {dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}},
+                                               {{"Out", {name}}}, {})});
     }
 
+    // make sure the inserted `add` ops follow the BFS order.
     insert_position.sort(
         [](const Pos& l, const Pos& r) { return l.first > r.first; });
 
     for (auto& pos : insert_position) {
-      net->InsertOp(pos.first + 1, pos.second);
+      net->InsertOp(pos.first + 1, std::move(pos.second));
     }
-
   } else {
-    std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
-    for (std::string& grad_input : grad_op->inputs_) {
+    std::unique_ptr<OperatorBase> grad_op(OpRegistry::CreateGradOp(forwardOp));
+
+    ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
+                                          const std::string& grad_input) {
       if (no_grad_names.count(grad_input)) {
         // +1 for \0
         std::string prefix = grad_input.substr(
             0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
-        grad_input = prefix + kZeroVarSuffix;
+        grad_op->Rename(grad_input, prefix + kZeroVarSuffix);
 
         // If part of input gradient of that operator is not calculated, fill
         // zero variables to that input gradient.
-        net->AddOp(OpRegistry::CreateOp("fill_zeros_like", {prefix},
-                                        {grad_input}, {}));
-      }
-    }
-
-    for (std::string& grad_output : grad_op->outputs_) {
-      if (no_grad_names.count(grad_output)) {
-        grad_output = kEmptyVarName;
+        net->AppendOp(OpRegistry::CreateOp("fill_zeros_like",
+                                           {{"Src", {prefix}}},
+                                           {{"Dst", {grad_input}}}, {}));
       }
+      return false;
+    });
+
+    ForEachVarName(grad_op->Outputs(),
+                   [&no_grad_names, &grad_op](const std::string& grad_output) {
+                     if (no_grad_names.count(grad_output)) {
+                       grad_op->Rename(grad_output, kEmptyVarName);
+                     }
+                     return false;
+                   });
+
+    // process recurrent gradient op as a special operator.
+    if (forwardOp.Type() == "recurrent_op") {
+      // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or
+      // this will result in infinite loop.
+      const auto& rnnop =
+          *static_cast<const operators::RecurrentOp*>(&forwardOp);
+      auto rnn_grad_op =
+          static_cast<operators::RecurrentGradientOp*>(grad_op.get());
+      const auto& stepnet_op =
+          *static_cast<const OperatorBase*>(&rnnop.stepnet());
+      // create stepnet's gradient op
+      rnn_grad_op->set_stepnet(
+          BackwardRecursive(stepnet_op, no_grad_names, uniq_id));
     }
 
     if (net->ops_.empty()) {  // Current no aux op is added to network
       return grad_op;
     }
-    net->AddOp(grad_op);
+    net->AppendOp(std::move(grad_op));
   }
-  net->type_ = "@GENERATED_BACKWARD@";
+  net->SetType("@GENERATED_BACKWARD@");
   net->CompleteAddOp();
-  return net;
+  return std::unique_ptr<OperatorBase>(
+      static_cast<OperatorBase*>(net.release()));
 }
 
 // See header for comments
-std::shared_ptr<OperatorBase> Backward(
+std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars) {
   std::unordered_set<std::string> no_grad_names;

paddle/framework/backward.h

@@ -20,7 +20,7 @@ namespace framework {
 
 // Create the backward operator from a forward operator.
 // TODO(yuyang18): Add more API reference comment.
-extern std::shared_ptr<OperatorBase> Backward(
+extern std::unique_ptr<OperatorBase> Backward(
     const OperatorBase& forwardOp,
     const std::unordered_set<std::string>& no_grad_vars);
 }  // namespace framework

paddle/framework/backward_test.cc

@@ -28,19 +28,13 @@ using OpAttrChecker = framework::OpAttrChecker;
 using Scope = framework::Scope;
 using DeviceContext = platform::DeviceContext;
 
-class EmptyOp : public OperatorBase {
- public:
-  void InferShape(const Scope &scope) const override {}
-  void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
-};
-
 class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
  public:
   RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "Input X of Add").IgnoreGradient();
-    AddInput("b", "Bias of Add").IgnoreGradient();
-    AddOutput("Out", "Out of Add").IgnoreGradient();
+    AddInput("X", "Input X of Add").NotInGradient();
+    AddInput("b", "Bias of Add").NotInGradient();
+    AddOutput("Out", "Out of Add").NotInGradient();
     AddComment("Add Op");
   }
 };
@@ -49,8 +43,8 @@ class MulOpMaker : public OpProtoAndCheckerMaker {
  public:
   MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("A", "A");
-    AddInput("B", "B");
+    AddInput("X", "A");
+    AddInput("Y", "B");
     AddOutput("Out", "Out");
     AddComment("Mul");
   }
@@ -61,7 +55,7 @@ class SigmoidOpMaker : public OpProtoAndCheckerMaker {
   SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "X");
-    AddOutput("Y", "Y");
+    AddOutput("Out", "Y");
     AddComment("Sigmoid");
   }
 };
@@ -71,21 +65,25 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
   NoGradOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "X input");
-    AddOutput("Y", "Y output");
+    AddOutput("Out", "Y output");
     AddComment("NoGradOp, same input output. no Grad");
   }
 };
 
 class FcOp : public operators::NetOp {
  public:
-  void Init() override {
-    AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")},
-                               {Output("mul_result")}, {}));
-    auto b_name = Input("b");
+  FcOp(const std::string &type, const VarNameMap &inputs,
+       const VarNameMap &outputs, const AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    AppendOp(OpRegistry::CreateOp("mul",
+                                  {{"X", {Input("X")}}, {"Y", {Input("W")}}},
+                                  {{"Out", {Output("mul_result")}}}, {}));
+    auto input_b = Inputs("b");
     std::string before_act = "mul_result";
-    if (b_name != kEmptyVarName) {
-      AddOp(OpRegistry::CreateOp("rowwise_add", {Output("mul_result"), b_name},
-                                 {Output("add_result")}, {}));
+    if (input_b.size() != 0) {
+      AppendOp(OpRegistry::CreateOp(
+          "rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}},
+          {{"Out", {Output("add_result")}}}, {}));
       before_act = "add_result";
     } else {
       auto out_varname = Output("add_result");
@@ -94,8 +92,8 @@ class FcOp : public operators::NetOp {
       }
     }
 
-    AddOp(OpRegistry::CreateOp("sigmoid", {Output(before_act)}, {Output("Out")},
-                               {}));
+    AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
+                                  {{"Out", {Output("Out")}}}, {}));
     CompleteAddOp(false);
   }
 };
@@ -107,8 +105,8 @@ class FcOpMaker : public OpProtoAndCheckerMaker {
     AddInput("X", "x");
     AddInput("W", "w");
     AddInput("b", "b");
-    AddOutput("mul_result", "").SetTemporary();
-    AddOutput("add_result", "").SetTemporary();
+    AddOutput("mul_result", "").AsIntermediate();
+    AddOutput("add_result", "").AsIntermediate();
     AddOutput("Out", "");
     AddComment("");
   }
@@ -139,7 +137,7 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
  public:
   AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "x").SetMultiple();
+    AddInput("X", "x").AsDuplicable();
     AddOutput("Y", "y");
     AddComment("");
   }
@@ -150,51 +148,48 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
 namespace f = paddle::framework;
 namespace ops = paddle::operators;
 using EnforceNotMet = paddle::platform::EnforceNotMet;
-REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker);
-REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp);
-REGISTER_OP(mul, f::EmptyOp, f::MulOpMaker);
-REGISTER_GRADIENT_OP(mul, mul_grad, f::EmptyOp);
-REGISTER_OP(sigmoid, f::EmptyOp, f::SigmoidOpMaker);
-REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, f::EmptyOp);
-REGISTER_OP(nograd, f::EmptyOp, f::NoGradOpMaker);
-REGISTER_OP(fill_zeros_like, f::EmptyOp, f::FillZeroOpMaker);
-REGISTER_OP(add, f::EmptyOp, f::AddOpMaker);
-REGISTER_GRADIENT_OP(add, add_grad, f::EmptyOp);
-REGISTER_OP(fc, f::FcOp, f::FcOpMaker);
-REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker);
-REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp);
+REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
+            f::NOP);
+REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
+REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
+REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
+REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
+REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
+REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
+REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
+            f::NOP);
 
 TEST(Backward, simple_op_grad) {
-  auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
+  auto fwd = f::OpRegistry::CreateOp(
+      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
   ASSERT_NE(fwd, nullptr);
   auto gop = f::OpRegistry::CreateGradOp(*fwd);
-  ASSERT_EQ(4UL, gop->inputs_.size());
-  ASSERT_EQ(f::kEmptyVarName, gop->inputs_[0]);
-  ASSERT_EQ("rowwise_add_grad", gop->type_);
-  ASSERT_EQ(f::GradVarName("X"), gop->outputs_[0]);
-  ASSERT_EQ(f::GradVarName("b"), gop->outputs_[1]);
-
-  ASSERT_EQ(f::GradVarName("X"), gop->Output(f::GradVarName("X")));
+  ASSERT_EQ(1UL, gop->Inputs().size());
+  ASSERT_EQ("rowwise_add_grad", gop->Type());
+  ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X")));
+  ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b")));
 }
 
 TEST(Backward, simple_op_not_need_grad) {
-  auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
+  auto fwd = f::OpRegistry::CreateOp(
+      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
   ASSERT_NE(fwd, nullptr);
-  auto gop = f::Backward(*fwd, {"X"});
-  ASSERT_EQ(std::find(gop->outputs_.begin(), gop->outputs_.end(),
-                      f::GradVarName("X")),
-            gop->outputs_.end());
+  auto gop = f::Backward(*fwd, {"x"});
+  ASSERT_EQ(gop->Output(f::GradVarName("X")), f::kEmptyVarName);
 
-  auto no_input_gop = f::Backward(*fwd, {"X", "b"});
+  auto no_input_gop = f::Backward(*fwd, {"x", "b"});
   ASSERT_NE(no_input_gop, nullptr);
   ASSERT_TRUE(no_input_gop->IsNetOp());
-  ASSERT_EQ(0UL,
-            std::static_pointer_cast<ops::NetOp>(no_input_gop)->ops_.size());
+  ASSERT_EQ(0UL, static_cast<ops::NetOp *>(no_input_gop.get())->ops_.size());
 }
 
 TEST(Backward, net_fc_backward_normal) {
-  std::shared_ptr<f::OperatorBase> fwd = f::OpRegistry::CreateOp(
-      "fc", {"X", "w", "b"}, {"mul_result", "add_result", "out"}, {});
+  std::shared_ptr<f::OperatorBase> fwd =
+      f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
+                              {{"mul_result", {"mul_res"}},
+                               {"add_result", {"add_re"}},
+                               {"Out", {"out"}}},
+                              {});
   ASSERT_NE(fwd, nullptr);
   std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
   ASSERT_TRUE(gop->IsNetOp());
@@ -205,19 +200,22 @@ TEST(Backward, net_fc_backward_normal) {
   ASSERT_EQ(3UL, net->ops_.size());
 
   f::OperatorBase &d_sigmoid = *net->ops_[0];
-  ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
+  ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
 
   f::OperatorBase &d_add = *net->ops_[1];
-  ASSERT_EQ("rowwise_add_grad", d_add.type_);
+  ASSERT_EQ("rowwise_add_grad", d_add.Type());
 
   f::OperatorBase &d_mul = *net->ops_[2];
-  ASSERT_EQ("mul_grad", d_mul.type_);
+  ASSERT_EQ("mul_grad", d_mul.Type());
 }
 
 TEST(Backward, net_fc_backward_not_have_b) {
   std::shared_ptr<f::OperatorBase> fwd =
-      f::OpRegistry::CreateOp("fc", {"X", "w", f::kEmptyVarName},
-                              {"mul_result", "add_result", "tmp"}, {});
+      f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {}}},
+                              {{"mul_result", {"mul_res"}},
+                               {"add_result", {"add_res"}},
+                               {"Out", {"tmp"}}},
+                              {});
   ASSERT_NE(fwd, nullptr);
   std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
   ASSERT_TRUE(gop->IsNetOp());
@@ -228,96 +226,113 @@ TEST(Backward, net_fc_backward_not_have_b) {
   ASSERT_EQ(2UL, net->ops_.size());
 
   f::OperatorBase &d_sigmoid = *net->ops_[0];
-  ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
+  ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
 
   f::OperatorBase &d_mul = *net->ops_[1];
-  ASSERT_EQ("mul_grad", d_mul.type_);
+  ASSERT_EQ("mul_grad", d_mul.Type());
 }
 
 TEST(Backward, net_input_of_network_not_need_grad) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp("fc", {"X", "W1", "b1"},
-                                    {"mul_tmp_0", "add_tmp_0", "hidden0"}, {}));
-  net.AddOp(f::OpRegistry::CreateOp("fc", {"hidden0", "W2", "b2"},
-                                    {"mul_tmp_1", "add_tmp_1", "hidden1"}, {}));
+  net.AppendOp(f::OpRegistry::CreateOp(
+      "fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}},
+      {{"mul_result", {"mul_tmp_0"}},
+       {"add_result", {"add_tmp_0"}},
+       {"Out", {"hidden0"}}},
+      {}));
+  net.AppendOp(f::OpRegistry::CreateOp(
+      "fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}},
+      {{"mul_result", {"mul_tmp_1"}},
+       {"add_result", {"add_tmp_1"}},
+       {"Out", {"hidden1"}}},
+      {}));
   net.CompleteAddOp();
-  auto bwd = Backward(net, {"X"});  // X@GRAD is not need.
+  auto bwd = Backward(net, {"x"});  // x@GRAD is not need.
   ASSERT_TRUE(bwd->IsNetOp());
   auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
 
-  std::unordered_set<std::string> all_output = std::unordered_set<std::string>(
-      bwd_net->outputs_.begin(), bwd_net->outputs_.end());
-  all_output.erase(f::kEmptyVarName);
+  auto output_vars = bwd_net->OutputVars(true);
+  std::unordered_set<std::string> all_outputs =
+      std::unordered_set<std::string>(output_vars.begin(), output_vars.end());
+  all_outputs.erase(f::kEmptyVarName);
 
   for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
-    ASSERT_NE(all_output.find(f::GradVarName(out)), all_output.end());
+    ASSERT_NE(all_outputs.find(f::GradVarName(out)), all_outputs.end());
   }
 
   // Not Generated X
-  ASSERT_EQ(all_output.find(f::GradVarName("X")), all_output.end());
+  ASSERT_EQ(all_outputs.find(f::GradVarName("X")), all_outputs.end());
 
   ASSERT_EQ(2UL, bwd_net->ops_.size());
   ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
   auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
   ASSERT_EQ(3UL, first_fc_grad->ops_.size());
   ASSERT_EQ(f::kEmptyVarName,
-            first_fc_grad->ops_[2]->Output(f::GradVarName("A")));
+            first_fc_grad->ops_[2]->Output(f::GradVarName("X")));
 }
 
 TEST(Backward, net_shared_weight) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp("mul", {"X", "W"}, {"Out"}, {}));
-  net.AddOp(f::OpRegistry::CreateOp("mul", {"Out", "W"}, {"FinalOut"}, {}));
+  net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
+                                       {{"Out", {"out"}}}, {}));
+  net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
+                                       {{"Out", {"FinalOut"}}}, {}));
   net.CompleteAddOp();
 
   auto bwd = f::Backward(net, {});
   ASSERT_TRUE(bwd->IsNetOp());
   auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
   ASSERT_EQ(3UL, bwd_net->ops_.size());
-  ASSERT_EQ("add", bwd_net->ops_[2]->type_);
+  ASSERT_EQ("add", bwd_net->ops_[2]->Type());
 }
 
 TEST(Backward, op_register_grad_not_for_network) {
-  auto fwd = f::OpRegistry::CreateOp(
-      "fc", {"X", "W", "b"}, {"mul_out", "add_out", "out1"},
-      {{"temporary_index", std::vector<int>{0, 1}}});
+  auto fwd =
+      f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
+                              {{"mul_result", {"mul_out"}},
+                               {"add_result", {"add_out"}},
+                               {"Out", {"out1"}}},
+                              {{"temporary_index", std::vector<int>{0, 1}}});
 
   ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet);
 }
 
 TEST(Backward, op_all_input_are_not_need) {
-  auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
-  auto backward = f::Backward(*fwd, {"X", "b"});
+  auto fwd = f::OpRegistry::CreateOp(
+      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
+  auto backward = f::Backward(*fwd, {"x", "b"});
   ASSERT_TRUE(backward->IsNetOp());
   auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_TRUE(net->ops_.empty());
 }
 
 TEST(Backward, op_all_output_are_not_need) {
-  auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
-  auto backward = f::Backward(*fwd, {"Out"});
+  auto fwd = f::OpRegistry::CreateOp(
+      "rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
+  auto backward = f::Backward(*fwd, {"out"});
   ASSERT_TRUE(backward->IsNetOp());
   auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_TRUE(net->ops_.empty());
 }
 
 TEST(Backward, op_part_of_output_are_not_need) {
-  auto fwd = f::OpRegistry::CreateOp("many_output_op", {"X"}, {"Y", "Z"}, {});
+  auto fwd = f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}},
+                                     {{"y", {"Y"}}, {"z", {"Z"}}}, {});
   auto backward = f::Backward(*fwd, {"Z"});
   ASSERT_TRUE(backward->IsNetOp());
   auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_EQ(net->ops_.size(), 2UL);
 
   auto &fill_zero = *net->ops_[0];
-  ASSERT_EQ("fill_zeros_like", fill_zero.type_);
-  ASSERT_EQ(1UL, fill_zero.inputs_.size());
-  ASSERT_EQ("Z", fill_zero.inputs_[0]);
-  ASSERT_EQ(1UL, fill_zero.outputs_.size());
-  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.outputs_[0]);
+  ASSERT_EQ("fill_zeros_like", fill_zero.Type());
+  ASSERT_EQ(1UL, fill_zero.Inputs("Src").size());
+  ASSERT_EQ("Z", fill_zero.Input("Src"));
+  ASSERT_EQ(1UL, fill_zero.Outputs("Dst").size());
+  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Dst"));
 
   auto &d_many_out = *net->ops_[1];
-  ASSERT_EQ("many_output_op_grad", d_many_out.type_);
-  ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size());  // I/O/OG
+  ASSERT_EQ("many_output_op_grad", d_many_out.Type());
+  ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.Inputs().size());  // I/O/OG
   ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix,
             d_many_out.Input(f::GradVarName("z")));
   ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y")));
@@ -325,44 +340,62 @@ TEST(Backward, op_part_of_output_are_not_need) {
 }
 
 TEST(Backward, op_part_of_input_are_not_need) {
-  auto fwd = f::OpRegistry::CreateOp("mul", {"a", "b"}, {"out"}, {});
+  auto fwd = f::OpRegistry::CreateOp("mul", {{"X", {"a"}}, {"Y", {"b"}}},
+                                     {{"Out", {"out"}}}, {});
   auto backward = f::Backward(*fwd, {"a"});
   auto &grad_mul = *backward;
-  ASSERT_EQ(grad_mul.type_, "mul_grad");
-  ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
-  ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
-  ASSERT_EQ(grad_mul.Output(f::GradVarName("A")), f::kEmptyVarName);
-  ASSERT_EQ(grad_mul.Output(f::GradVarName("B")), f::GradVarName("b"));
+  ASSERT_EQ(grad_mul.Type(), "mul_grad");
+  ASSERT_EQ(grad_mul.Inputs().size(), 2UL + 1UL + 1UL);
+  ASSERT_EQ(grad_mul.Outputs().size(), 2UL);
+  ASSERT_EQ(grad_mul.Output(f::GradVarName("X")), f::kEmptyVarName);
+  ASSERT_EQ(grad_mul.Output(f::GradVarName("Y")), f::GradVarName("b"));
   ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out"));
-  ASSERT_EQ(grad_mul.Input("A"), "a");
-  ASSERT_EQ(grad_mul.Input("B"), "b");
+  ASSERT_EQ(grad_mul.Input("X"), "a");
+  ASSERT_EQ(grad_mul.Input("Y"), "b");
   ASSERT_EQ(grad_mul.Input("Out"), "out");
 }
 
 TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp("fc", {"x1", "w1", "b1"},
-                                    {"mul_out1", "add_out1", "out1"}, {}));
-  net.AddOp(f::OpRegistry::CreateOp("fc", {"out1", "w2", "b2"},
-                                    {"mul_out2", "tmp_out2", "out2"}, {}));
-  net.AddOp(f::OpRegistry::CreateOp("fc", {"out2", "w3", "b3"},
-                                    {"mul_out3", "tmp_out3", "out3"}, {}));
+  net.AppendOp(f::OpRegistry::CreateOp(
+      "fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}},
+      {{"mul_result", {"mul_out1"}},
+       {"add_result", {"add_out1"}},
+       {"Out", {"out1"}}},
+      {}));
+  net.AppendOp(f::OpRegistry::CreateOp(
+      "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}},
+      {{"mul_result", {"mul_out2"}},
+       {"add_result", {"tmp_out2"}},
+       {"Out", {"out2"}}},
+      {}));
+  net.AppendOp(f::OpRegistry::CreateOp(
+      "fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}},
+      {{"mul_result", {"mul_out3"}},
+       {"add_result", {"tmp_out3"}},
+       {"Out", {"out3"}}},
+      {}));
   net.CompleteAddOp();
+
   auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"});
   ASSERT_TRUE(backward->IsNetOp());
   auto bwd_net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_EQ(bwd_net->ops_.size(), 3UL);
   auto &grad_fc = *bwd_net->ops_[0];
-  EXPECT_EQ(grad_fc.inputs_.size(),
-            3UL       /* external input number */
+
+  const char *all = paddle::operators::NetOp::kAll;
+  EXPECT_EQ(grad_fc.Inputs(all).size(),
+            2UL       /* external input number */
                 + 1UL /* external output number*/
                 + 1UL /* number of gradient of external output*/
                 + 2U /* internal variable number*/);
-  EXPECT_EQ(grad_fc.outputs_.size(), 2UL       /* input number of mul*/
-                                         + 2UL /* input number of rowwise_add */
-                                         + 1UL /* input number of sigmod */);
-  EXPECT_EQ(bwd_net->ops_[1]->inputs_.size(), 0UL);
-  EXPECT_EQ(bwd_net->ops_[1]->outputs_.size(), 0UL);
-  EXPECT_EQ(bwd_net->ops_[2]->inputs_.size(), 0UL);
-  EXPECT_EQ(bwd_net->ops_[2]->outputs_.size(), 0UL);
+  EXPECT_EQ(grad_fc.Outputs(all).size(),
+            2UL       /* input number of mul*/
+                + 2UL /* input number of rowwise_add
+                       */
+                + 1UL /* input number of sigmod */);
+  EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL);
 }

paddle/framework/ddim.cc

@@ -283,6 +283,5 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
 DDim::DDim(std::initializer_list<int> init_list) {
   *this = make_ddim(init_list);
 }
-
 }  // namespace framework
 }  // namespace paddle

paddle/framework/ddim.h

@@ -20,7 +20,6 @@ limitations under the License. */
 #include "paddle/framework/dim.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/variant.h"
-#include "unsupported/Eigen/CXX11/Tensor"
 
 namespace paddle {
 namespace framework {

paddle/framework/details/lod_tensor.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-   http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License. */
-
-#include "paddle/framework/lod_tensor.h"
-
-#include <memory>
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-using LOD = LODTensor::LOD;
-
-std::shared_ptr<LOD> SliceLOD(const LOD &lod, size_t level_begin,
-                              size_t level_end) {
-  auto new_lod = std::make_shared<LOD>();
-  new_lod->reserve(level_end - level_begin);
-  for (size_t i = level_begin; i < level_end; i++) {
-    new_lod->emplace_back(lod[i]);
-  }
-  return new_lod;
-}
-
-std::shared_ptr<LOD> SliceLOD(const LOD &lod, size_t level, size_t elem_begin,
-                              size_t elem_end, bool tensor_shared) {
-  // slice the lod.
-  auto new_lod = std::make_shared<LOD>();
-  new_lod->reserve(lod.size() - level);
-  auto start = lod.at(level)[elem_begin];
-  auto end = lod.at(level)[elem_end];
-
-  for (auto it = lod.begin() + level; it != lod.end(); it++) {
-    auto it_begin = std::find(it->begin(), it->end(), start);
-    auto it_end = std::find(it_begin, it->end(), end);
-    PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
-    PADDLE_ENFORCE(it_end != it->end(), "error in parsing lod info");
-    new_lod->emplace_back(it_begin, it_end + 1);
-    if (!tensor_shared) {
-      // reset offset if tensor is copyed and sliced.
-      std::transform(new_lod->back().begin(), new_lod->back().end(),
-                     new_lod->back().begin(),
-                     [start](int v) { return v - start; });
-      PADDLE_ENFORCE(new_lod->back().front() == 0, "error in slice LOD");
-    }
-  }
-  return new_lod;
-}
-
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle

paddle/framework/details/lod_tensor.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 <memory>
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-/*
- * Slice levels from LOD.
- *
- * @lod: LOD to slice.
- * @level_begin: level to begin slice.
- * @level_end: level to end slice.
- */
-std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
-                                         size_t level_begin, size_t level_end);
-
-/*
- * Slice elements from a level of LOD.
- *
- * @lod: LOD to slice.
- * @level: which level to slice.
- * @elem_begin: element's index to begin slice.
- * @elem_end: element's index to end slice.
- */
-std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
-                                         size_t level, size_t elem_begin,
-                                         size_t elem_end, bool tensor_shared);
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle

paddle/framework/framework.proto

+/* 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. */
+
+syntax = "proto2";
+package paddle.framework;
+
+enum AttrType {
+  INT = 0;
+  FLOAT = 1;
+  STRING = 2;
+  INTS = 3;
+  FLOATS = 4;
+  STRINGS = 5;
+}
+
+// OpDesc describes an instance of a C++ framework::OperatorBase
+// derived class type.
+message OpDesc {
+
+  message Attr {
+    required string name = 1;
+    required AttrType type = 2;
+    optional int32 i = 3;
+    optional float f = 4;
+    optional string s = 5;
+    repeated int32 ints = 6;
+    repeated float floats = 7;
+    repeated string strings = 8;
+  };
+
+  message Var {
+    required string parameter = 1;
+    repeated string arguments = 2;
+  };
+
+  required string type = 3;
+  repeated Var inputs = 1;
+  repeated Var outputs = 2;
+  repeated Attr attrs = 4;
+};
+
+// OpProto describes a C++ framework::OperatorBase derived class.
+message OpProto {
+
+  // VarProto describes the C++ type framework::Variable.
+  message Var {
+    required string name = 1;
+    required string comment = 2;
+
+    optional bool duplicable = 3 [ default = false ];
+    optional bool intermediate = 4 [ default = false ];
+    optional bool not_in_gradient = 5 [ default = false ];
+  }
+
+  // AttrProto describes the C++ type Attribute.
+  message Attr {
+    required string name = 1;
+    required AttrType type = 2;
+    required string comment = 3;
+    // If that attribute is generated, it means the Paddle third
+    // language binding has responsibility to fill that
+    // attribute. End-User should not set that attribute.
+    optional bool generated = 4 [ default = false ];
+  }
+
+  required string type = 1;
+  repeated Var inputs = 2;
+  repeated Var outputs = 3;
+  repeated Attr attrs = 4;
+  required string comment = 5;
+}

paddle/framework/grad_op_builder.cc

@@ -13,90 +13,53 @@ express or implied. See the License for the specific language governing
 permissions and limitations under the License. */
 
 #include "paddle/framework/grad_op_builder.h"
-#include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace framework {
-
-class OpRegistry;
-
-using VarIndexMap = std::unordered_map<std::string, int>;
-
 enum class OpArgType { IN, OUT };
 
-static std::vector<int>* GetOpFormat(OperatorBase* op, const OpArgType& type) {
-  std::string key = type == OpArgType::IN ? "input_format" : "output_format";
-  return op->attrs_.count(key)
-             ? &boost::get<std::vector<int>>(op->attrs_.at(key))
-             : nullptr;
-}
-
-static const std::vector<int>* GetOpFormat(const OperatorBase* op,
-                                           const OpArgType& type) {
-  std::string key = type == OpArgType::IN ? "input_format" : "output_format";
-  return op->attrs_.count(key)
-             ? &boost::get<std::vector<int>>(op->attrs_.at(key))
-             : nullptr;
-}
-
-static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op,
-                       const OpArgType& src_type, const OpArgType& dst_type,
-                       int& idx, bool is_grad) {
-  const std::vector<std::string>& src_inout =
-      src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_;
-  const std::vector<int>* src_format = GetOpFormat(src_op, src_type);
-
-  std::vector<std::string>& dst_inout =
-      dst_type == OpArgType::IN ? dst_op->inputs_ : dst_op->outputs_;
-  std::vector<int>* dst_format = GetOpFormat(dst_op, dst_type);
-  const OpProto& proto = OpRegistry::protos().at(src_op->type_);
+static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
+                       bool is_grad, OperatorBase::VarNameMap* vars) {
+  const auto& src_inout =
+      src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
+  auto& dst_inout = *vars;
+  const OpProto* proto = OpRegistry::op_info_map().at(src_op->Type()).proto_;
   const auto& src_arg_list =
-      src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
-
+      src_type == OpArgType::IN ? proto->inputs() : proto->outputs();
   for (const auto& arg : src_arg_list) {
-    std::string src_name = arg.name();
-    std::string dst_name = is_grad ? src_name + kGradVarSuffix : src_name;
-    (*dst_op->in_out_idxs_)[dst_name] = idx++;
-    int src_arg_idx = src_op->in_out_idxs_->at(src_name);
-    int src_begin =
-        src_format == nullptr ? src_arg_idx : src_format->at(src_arg_idx);
-    int src_end = src_format == nullptr ? src_arg_idx + 1
-                                        : src_format->at(src_arg_idx + 1);
-    for (int i = src_begin; i < src_end; ++i) {
-      std::string s =
-          is_grad ? src_inout[i] + kGradVarSuffix
-                  : (arg.ignore_gradient() ? kEmptyVarName : src_inout[i]);
-      dst_inout.emplace_back(s);
-    }
-    if (dst_format != nullptr) {
-      dst_format->push_back(dst_inout.size());
+    if (arg.not_in_gradient() && !is_grad) continue;
+    const std::string src_name = arg.name();
+    std::string dst_name = is_grad ? GradVarName(src_name) : src_name;
+    dst_inout[dst_name].reserve(src_inout.at(src_name).size());
+    for (auto& var_name : src_inout.at(src_name)) {
+      std::string s = is_grad ? GradVarName(var_name) : var_name;
+      dst_inout[dst_name].emplace_back(s);
     }
   }
 }
 
 OperatorBase* BuildGradOp(const OperatorBase* op) {
-  std::string grad_op_type = OpRegistry::grad_ops().at(op->type_);
-  OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)();
-  grad_op->type_ = grad_op_type;
-  grad_op->attrs_ = op->attrs_;
-  grad_op->attrs_.erase("input_format");
-  grad_op->attrs_.erase("output_format");
-  if (GetOpFormat(op, OpArgType::IN) != nullptr) {
-    grad_op->attrs_["output_format"] = std::vector<int>({0});
-  }
-  if (GetOpFormat(op, OpArgType::IN) != nullptr ||
-      GetOpFormat(op, OpArgType::OUT) != nullptr) {
-    grad_op->attrs_["input_format"] = std::vector<int>({0});
-  }
-  grad_op->in_out_idxs_.reset(new VarIndexMap());
-  int in_idx = 0;
-  int out_idx = 0;
-  TransOpArg(op, grad_op, OpArgType::IN, OpArgType::IN, in_idx, false);   // I
-  TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, false);  // G
-  TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, true);   // OG
-  TransOpArg(op, grad_op, OpArgType::IN, OpArgType::OUT, out_idx, true);  // IG
-  return grad_op;
+  auto it = OpRegistry::op_info_map().find(op->Type());
+  PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
+                 "'%s' has not been registered.", op->Type());
+  PADDLE_ENFORCE(it->second.proto_ != nullptr, "'%s' has no OpProto.",
+                 op->Type());
+  std::string grad_op_type = it->second.grad_op_type_;
+  PADDLE_ENFORCE(!grad_op_type.empty(), "'%s' has no gradient operator.",
+                 op->Type());
+
+  OperatorBase::VarNameMap inputs;
+  OperatorBase::VarNameMap outputs;
+  TransOpArg(op, OpArgType::IN, false, &inputs);   // I
+  TransOpArg(op, OpArgType::OUT, false, &inputs);  // O
+  TransOpArg(op, OpArgType::OUT, true, &inputs);   // OG
+  TransOpArg(op, OpArgType::IN, true, &outputs);   // IG
+
+  it = OpRegistry::op_info_map().find(grad_op_type);
+  PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
+                 "'%s' has not been registered.", grad_op_type);
+  return it->second.creator_(grad_op_type, inputs, outputs, op->Attrs());
 }
 
 }  // namespace framework

paddle/framework/grad_op_builder_test.cc

@@ -8,22 +8,15 @@ USE_OP(add_two);
 namespace paddle {
 namespace framework {
 
-class NOP : public OperatorBase {
- public:
-  void InferShape(const Scope &scope) const override {}
-  void Run(const Scope &scope,
-           const platform::DeviceContext &dev_ctx) const override {}
-};
-
 class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
  public:
   MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("In1", "a single input");
-    AddInput("In2_mult", "a multiple input").SetMultiple();
+    AddInput("In2_mult", "a multiple input").AsDuplicable();
     AddInput("In3", "another single input");
     AddOutput("Out1", "a single output");
-    AddOutput("Out2_mult", "a multiple output").SetMultiple();
+    AddOutput("Out2_mult", "a multiple output").AsDuplicable();
     AddComment("test op with multiple inputs and outputs");
   }
 };
@@ -33,10 +26,10 @@ class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
   IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("In1", "a single input");
-    AddInput("In2_mult", "a multiple input").SetMultiple().IgnoreGradient();
-    AddInput("In3_mult", "another multiple input").SetMultiple();
-    AddOutput("Out1_mult", "a multiple output").SetMultiple();
-    AddOutput("Out2", "a single output").IgnoreGradient();
+    AddInput("In2_mult", "a multiple input").AsDuplicable().NotInGradient();
+    AddInput("In3_mult", "another multiple input").AsDuplicable();
+    AddOutput("Out1_mult", "a multiple output").AsDuplicable();
+    AddOutput("Out2", "a single output").NotInGradient();
     AddComment("op with inputs and outputs ignored in gradient calculating");
   }
 };
@@ -47,35 +40,33 @@ class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
 namespace f = paddle::framework;
 
 TEST(GradOpBuilder, AddTwo) {
-  std::shared_ptr<f::OperatorBase> add_op(
-      f::OpRegistry::CreateOp("add_two", {"x", "y"}, {"out"}, {}));
+  std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
+      "add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
   std::shared_ptr<f::OperatorBase> grad_add_op =
       f::OpRegistry::CreateGradOp(*add_op);
-  EXPECT_EQ(static_cast<int>(grad_add_op->inputs_.size()), 4);
-  EXPECT_EQ(static_cast<int>(grad_add_op->outputs_.size()), 2);
+  EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);
+  EXPECT_EQ(grad_add_op->Outputs().size(), 2UL);
   EXPECT_EQ(grad_add_op->Input("X"), "x");
   EXPECT_EQ(grad_add_op->Input("Y"), "y");
   EXPECT_EQ(grad_add_op->Input("Out"), "out");
-  EXPECT_EQ(grad_add_op->Input("Out@GRAD"), "out@GRAD");
-  EXPECT_EQ(grad_add_op->Output("X@GRAD"), "x@GRAD");
-  EXPECT_EQ(grad_add_op->Output("Y@GRAD"), "y@GRAD");
+  EXPECT_EQ(grad_add_op->Input(f::GradVarName("Out")), f::GradVarName("out"));
+  EXPECT_EQ(grad_add_op->Output(f::GradVarName("X")), f::GradVarName("x"));
+  EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y"));
 }
 
-REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker);
-REGISTER_GRADIENT_OP(mult_io, mult_io_grad, f::NOP);
-REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker);
-REGISTER_GRADIENT_OP(io_ignored, io_ignored_grad, f::NOP);
+REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);
+REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP);
 
 TEST(GradOpBuilder, MutiInOut) {
-  f::AttributeMap attrs{{"input_format", std::vector<int>{0, 1, 4, 5}},
-                        {"output_format", std::vector<int>{0, 1, 3}}};
   std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
-      "mult_io", {"in1", "in2_1", "in2_2", "in2_3", "in3"},
-      {"out1", "out2_1", "out2_2"}, attrs));
+      "mult_io", {{"In1", {"in1"}},
+                  {"In2_mult", {"in2_1", "in2_2", "in2_3"}},
+                  {"In3", {"in3"}}},
+      {{"Out1", {"out1"}}, {"Out2_mult", {"out2_1", "out2_2"}}}, {}));
   std::shared_ptr<f::OperatorBase> grad_test_op =
       f::OpRegistry::CreateGradOp(*test_op);
 
-  ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
+  ASSERT_EQ(grad_test_op->Inputs().size(), 3UL + 2UL + 2UL);
   EXPECT_EQ(grad_test_op->Input("In1"), "in1");
   EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
             std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
@@ -89,7 +80,7 @@ TEST(GradOpBuilder, MutiInOut) {
             std::vector<std::string>(
                 {f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
 
-  ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
+  ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
   EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
   EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
             std::vector<std::string>({f::GradVarName("in2_1"),
@@ -99,31 +90,28 @@ TEST(GradOpBuilder, MutiInOut) {
 }
 
 TEST(GradOpBuilder, IOIgnoredInGradient) {
-  f::AttributeMap attrs{{"input_format", std::vector<int>{0, 1, 3, 5}},
-                        {"output_format", std::vector<int>{0, 2, 3}}};
   std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
-      "io_ignored", {"in1", "in2_1", "in2_2", "in3_1", "in3_2"},
-      {"out1_1", "out1_2", "out2"}, attrs));
+      "io_ignored", {{"In1", {"in1"}},
+                     {"In2_mult", {"in2_1", "in2_2"}},
+                     {"In3_mult", {"in3_1", "in3_2"}}},
+      {{"Out1_mult", {"out1_1", "out1_2"}}, {"Out2", {"out2"}}}, {}));
   std::shared_ptr<f::OperatorBase> grad_test_op =
       f::OpRegistry::CreateGradOp(*test_op);
 
   // 'In2' and 'Out2' are ignored in gradient calculating
-  ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
+  ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL);
   EXPECT_EQ(grad_test_op->Input("In1"), "in1");
-  EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
-            std::vector<std::string>({f::kEmptyVarName, f::kEmptyVarName}));
   EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
             std::vector<std::string>({"in3_1", "in3_2"}));
   EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
             std::vector<std::string>({"out1_1", "out1_2"}));
-  EXPECT_EQ(grad_test_op->Input("Out2"), f::kEmptyVarName);
   EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")),
             std::vector<std::string>(
                 {f::GradVarName("out1_1"), f::GradVarName("out1_2")}));
   EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")),
             f::GradVarName("out2"));
 
-  ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
+  ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
   EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
   EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
             std::vector<std::string>(

paddle/framework/lod_tensor.cc

@@ -19,32 +19,59 @@
 namespace paddle {
 namespace framework {
 
-LODTensor LODTensor::SliceShared(size_t level_begin, size_t level_end) const {
-  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
-  auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end);
-  // slice levels just need to update LOD info, each level will contains the
-  // whole tensor_, so no need to modify tensor_.
-  return LODTensor(tensor_, new_lod);
+LODTensor::LOD LODTensor::LOD::SliceLevels(size_t level_begin,
+                                           size_t level_end) const {
+  LOD new_lod;
+  new_lod.reserve(level_end - level_begin);
+  for (size_t i = level_begin; i < level_end; i++) {
+    new_lod.emplace_back(at(i));
+  }
+  return new_lod;
 }
 
-LODTensor LODTensor::SliceShared(size_t level, size_t elem_begin,
-                                 size_t elem_end) const {
-  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
-  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
-                 NumLevels());
-  PADDLE_ENFORCE(elem_begin < NumElements(level),
-                 "element begin [%d] out of range [%d]", elem_begin,
-                 NumElements(level));
-  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
-                 "element end [%d] out of range [%d]", elem_end,
-                 NumElements(level));
-
-  auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
-                                   true /*tensor_shared*/);
-
-  // slice elements just need to update LOD info, because offsets are not
-  // changed, so the original tensor_ can be reused.
-  return LODTensor(tensor_, new_lod);
+LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
+                                            size_t elem_end) const {
+  // slice the lod.
+  LOD new_lod;
+  new_lod.reserve(size() - level);
+  auto start = this->at(level)[elem_begin];
+  auto end = this->at(level)[elem_end];
+
+  for (auto it = this->begin() + level; it != this->end(); it++) {
+    auto it_begin = std::find(it->begin(), it->end(), start);
+    auto it_end = std::find(it_begin, it->end(), end);
+    PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
+    PADDLE_ENFORCE(it_end != it->end(), "error in parsing lod info");
+    new_lod.emplace_back(it_begin, it_end + 1);
+    // reset offset if tensor is copyed and sliced.
+    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_LE(new_lod.size(), this->size());
+  return new_lod;
+}
+
+bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
+  if (a.size() != b.size()) {
+    return false;
+  }
+
+  for (size_t i = 0; i < a.size(); i++) {
+    const auto& a_level = a[i];
+    const auto& b_level = b[i];
+    if (a_level.size() != b_level.size()) {
+      return false;
+    }
+    for (size_t j = 0; j < a_level.size(); j++) {
+      if (a_level[j] != b_level[j]) {
+        return false;
+      }
+    }
+  }
+
+  return true;
 }
 
 }  // namespace framework

paddle/framework/lod_tensor.h

@@ -15,7 +15,7 @@
 #pragma once
 
 #include <memory>
-#if (!PADDLE_ONLY_CPU)
+#if !defined(PADDLE_ONLY_CPU)
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
 #endif
@@ -31,30 +31,29 @@ namespace framework {
  * LODTensor (Level of details Tensor)
  * see https://en.wikipedia.org/wiki/Level_of_details for reference.
  */
-class LODTensor {
+class LODTensor : public Tensor {
  public:
 // Level save offsets of each unit.
 #ifdef PADDLE_ONLY_CPU
-  using Level = std::vector<size_t>;
+  template <typename T>
+  using Vector = std::vector<T>;
 #else
-  using Level = thrust::device_vector<size_t>;
+  template <typename T>
+  using Vector = thrust::host_vector<T>;
 #endif
-  // LOD stores offsets of each level of units, the largest units level first,
+  // LoD stores offsets of each level of units, the largest units level first,
   // then the smaller units level. Each Level stores the offsets of units in
   // Tesor.
-  typedef std::vector<Level> LOD;
+  class LOD : public std::vector<Vector<size_t>> {
+   public:
+    LOD SliceLevels(size_t level_begin, size_t level_end) const;
+    LOD SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) const;
+  };
 
   LODTensor() {}
-  LODTensor(const std::shared_ptr<Tensor> &tensor,
-            const std::shared_ptr<LOD> &lod) {
-    Reset(tensor, lod);
-  }
+  explicit LODTensor(const LOD &lod) : lod_(lod) {}
 
-  void Reset(const std::shared_ptr<Tensor> &tensor,
-             const std::shared_ptr<LOD> &lod) {
-    tensor_ = tensor;
-    lod_start_pos_ = lod;
-  }
+  virtual Tensor *Clone() const { return new LODTensor(lod_); }
 
   /*
    * Get a element from LOD.
@@ -65,16 +64,14 @@ class LODTensor {
     PADDLE_ENFORCE(elem < NumElements(level),
                    "element begin [%d] out of range [%d]", elem,
                    NumElements(level));
-    return (*lod_start_pos_)[level][elem];
+    return (lod_)[level][elem];
   }
 
   /*
    * 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_start_pos_ ? lod_start_pos_->size() : 0UL;
-  }
+  size_t NumLevels() const { return lod_.size(); }
   /*
    * Number of elements in a level.
    */
@@ -82,64 +79,71 @@ class LODTensor {
     PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
                    NumLevels());
     // the last offset is the end of last element
-    return lod_start_pos_->at(level).size() - 1;
+    return lod_[level].size() - 1;
   }
 
-  /*
-   * Slice of levels[level_begin:level_end], with tensor copied.
-   */
-  template <typename T>
-  LODTensor SliceCopied(size_t level_begin, size_t level_end,
-                        const platform::Place &dst_place) const;
-
   /*
    * Slice of levels[level_begin:level_end], with tensor shared.
    */
-  LODTensor SliceShared(size_t level_begin, size_t level_end) const;
-
-  /*
-   * Slice of elements of a level, [elem_begin: elem_end], with tensor copied.
-   * @note: low performance in slice lod_start_pos_.
-   */
   template <typename T>
-  LODTensor SliceCopied(size_t level, size_t elem_begin, size_t elem_end,
-                        const platform::Place &dst_place) const;
+  LODTensor SliceLevels(size_t level_begin, size_t level_end) const;
 
   /*
    * Slice of elements of a level, [elem_begin: elem_end], with tensor shared.
-   * @note: low performance in slice lod_start_pos_.
-   */
-  LODTensor SliceShared(size_t level, size_t elem_begin, size_t elem_end) const;
-
-  /*
-   * Copy other's lod_start_pos_, to share LOD info.
-   * @note: the LOD info should not be changed.
+   * @note: low performance in slice lod_.
    */
-  void ShareLOD(const LODTensor &other) {
-    lod_start_pos_ = other.lod_start_pos_;
-  }
+  template <typename T>
+  LODTensor SliceInLevel(size_t level, size_t elem_begin,
+                         size_t elem_end) const;
 
   /*
-   * Copy other's lod_start_pos_'s content, free to mutate.
+   * Copy other's lod_'s content, free to mutate.
    */
-  void CopyLOD(const LODTensor &other) {
-    lod_start_pos_ = std::make_shared<LOD>(*other.lod_start_pos_);
-  }
+  void CopyLOD(const LODTensor &other) { lod_ = other.lod_; }
   /*
    * Determine whether LODTensor has a valid LOD info.
    */
-  bool HasLOD() const { return bool(lod_start_pos_); }
-  LOD *lod() const { return lod_start_pos_.get(); }
+  const LOD &lod() const { return lod_; }
+  LOD *mutable_lod() { return &lod_; }
 
-  std::shared_ptr<Tensor> &tensor() { return tensor_; }
-  Tensor *raw_tensor() { return tensor_.get(); }
+  virtual ~LODTensor() {}
 
  private:
-  std::shared_ptr<LOD> lod_start_pos_;
-  std::shared_ptr<Tensor> tensor_;
+  LOD lod_;
 };
 
+bool operator==(const LODTensor::LOD &a, const LODTensor::LOD &b);
+
+template <typename T>
+LODTensor LODTensor::SliceLevels(size_t level_begin, size_t level_end) const {
+  auto new_lod = lod_.SliceLevels(level_begin, level_end);
+  // slice levels just need to update LOD info, each level will contains the
+  // whole tensor_, so no need to modify tensor_.
+  LODTensor new_tensor(new_lod);
+  new_tensor.ShareDataWith<T>(*this);
+  return new_tensor;
+}
+
+template <typename T>
+LODTensor LODTensor::SliceInLevel(size_t level, size_t elem_begin,
+                                  size_t elem_end) const {
+  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
+                 NumLevels());
+  PADDLE_ENFORCE(elem_begin < NumElements(level),
+                 "element begin [%d] out of range [%d]", elem_begin,
+                 NumElements(level));
+  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
+                 "element end [%d] out of range [%d]", elem_end,
+                 NumElements(level));
+
+  auto new_lod = lod_.SliceInLevel(level, elem_begin, elem_end);
+
+  // slice elements just need to update LOD info, because offsets are not
+  // changed, so the original tensor_ can be reused.
+  LODTensor new_tensor(new_lod);
+  new_tensor.ShareDataWith<T>(*this);
+  return new_tensor;
+}
+
 }  // namespace framework
 }  // namespace paddle
-
-#include "paddle/framework/lod_tensor_impl.h"

paddle/framework/lod_tensor_impl.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-   http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License. */
-
-#pragma once
-
-#include "paddle/framework/details/lod_tensor.h"
-
-namespace paddle {
-namespace framework {
-
-template <typename T>
-LODTensor LODTensor::SliceCopied(size_t level_begin, size_t level_end,
-                                 const platform::Place &dst_place) const {
-  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
-  auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end);
-  auto new_tensor = std::make_shared<Tensor>();
-  new_tensor->CopyFrom<T>(*tensor_, dst_place);
-
-  return LODTensor(new_tensor, new_lod);
-}
-
-template <typename T>
-LODTensor LODTensor::SliceCopied(size_t level, size_t elem_begin,
-                                 size_t elem_end,
-                                 const platform::Place &dst_place) const {
-  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
-  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
-                 NumLevels());
-  PADDLE_ENFORCE(elem_begin < NumElements(level),
-                 "element begin [%d] out of range [%d]", elem_begin,
-                 NumElements(level));
-  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
-                 "element end [%d] out of range [%d]", elem_end,
-                 NumElements(level));
-
-  auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
-                                   false /*tensor_shared*/);
-
-  auto start_idx = new_lod->front().front();
-  auto end_idx = new_lod->front().back() - 1 /*the next element's start*/;
-  auto sliced_tensor = tensor_->Slice<T>(start_idx, end_idx);
-  auto new_tensor = std::make_shared<Tensor>();
-  new_tensor->CopyFrom<T>(sliced_tensor, dst_place);
-
-  return LODTensor(new_tensor, new_lod);
-}
-
-}  // namespace framework
-}  // namespace paddle

paddle/framework/lod_tensor_test.cc

@@ -15,6 +15,7 @@
 
 #include <glog/logging.h>
 #include <gtest/gtest.h>
+#include <algorithm>
 #include <memory>
 
 namespace paddle {
@@ -29,22 +30,28 @@ class LODTensorTester : public ::testing::Test {
     // 0 10 20
     // 0 5 10 15 20
     // 0 2 5 7 10 12 15 20
-    auto lod = std::make_shared<LODTensor::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});
+    LODTensor::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});
 
-    auto tensor = std::make_shared<Tensor>();
-    tensor->Resize({20 /*batch size*/, 128 /*dim*/});
+    ASSERT_EQ(lod.size(), 3UL);
+
+    tensor.Resize({20 /*batch size*/, 128 /*dim*/});
     // malloc memory
-    tensor->mutable_data<float>(place);
+    tensor.mutable_data<float>(place);
+
+    lod_tensor.reset(new LODTensor(lod));
+    lod_tensor->Resize({20 /*batch size*/, 128 /*dim*/});
 
-    lod_tensor->Reset(tensor, lod);
+    lod_tensor->ShareDataWith<float>(tensor);
+    // lod_tensor->ShareDataWith<Tensor>(tensor);
   }
 
  protected:
   std::unique_ptr<LODTensor> lod_tensor;
   platform::CPUPlace place;
+  Tensor tensor;
 };
 
 TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor->NumLevels(), 3UL); }
@@ -55,110 +62,54 @@ TEST_F(LODTensorTester, NumElements) {
   ASSERT_EQ(lod_tensor->NumElements(2), 8UL);
 }
 
-TEST_F(LODTensorTester, SliceShared_Level) {
-  // slice 1 level
-  for (size_t level = 0; level < 3UL; ++level) {
-    auto new_lod_tensor = lod_tensor->SliceShared(level, level + 1);
-    ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
-    ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
-  }
-  // slice 2 level
-  for (size_t level = 0; level < 2UL; ++level) {
-    auto new_lod_tensor = lod_tensor->SliceShared(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(), lod_tensor->tensor());
-  }
-}
-
-TEST_F(LODTensorTester, SliceCopied_Level) {
+TEST_F(LODTensorTester, SliceLevels) {
   // slice 1 level
   for (size_t level = 0; level < 3UL; ++level) {
-    auto new_lod_tensor =
-        lod_tensor->SliceCopied<float>(level, level + 1, place);
+    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
     ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
-    // ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
-    // TODO(superjom) add tensor comparation here.
+    // ASSERT_EQ(new_lod_tensor, *lod_tensor);
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    auto new_lod_tensor =
-        lod_tensor->SliceCopied<float>(level, level + 2, place);
+    auto new_lod_tensor = lod_tensor->SliceLevels<float>(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(), lod_tensor->tensor());
-    // TODO(superjom) add tensor comparation here.
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
   }
 }
 
-TEST_F(LODTensorTester, SliceShared_Element) {
-  size_t level = 0;
-  auto new_lod_tensor = lod_tensor->SliceShared(level, 0, 2);
-  ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_EQ(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
-
-  level = 1;
-  new_lod_tensor = lod_tensor->SliceShared(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.raw_tensor(), lod_tensor->raw_tensor());
-}
-
-TEST_F(LODTensorTester, SliceCopied_Element) {
+TEST_F(LODTensorTester, SliceInLevel) {
   size_t level = 0;
-  auto new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, place);
-  ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
+  auto new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
+  EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, place);
+  new_lod_tensor = lod_tensor->SliceInLevel<float>(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_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
-
-  level = 1;
-  // LOD is
-  //    0 5 10
-  //    0 2 5 7 10
-  new_lod_tensor = lod_tensor->SliceCopied<float>(level, 1, 3, place);
-  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.lod_element(0, 0), 0UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(0, 1), 5UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(1, 0), 0UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(1, 1), 2UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(1, 2), 5UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(1, 3), 7UL);
-
-  // TODO(superjom) compare the content of these tensors
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
 }
 
 TEST_F(LODTensorTester, ShareLOD) {
   LODTensor new_lod_tensor;
-  new_lod_tensor.ShareLOD(*lod_tensor);
+  new_lod_tensor.CopyLOD(*lod_tensor);
   ASSERT_EQ(new_lod_tensor.lod(), lod_tensor->lod());
 }
 
 TEST_F(LODTensorTester, CopyLOD) {
   LODTensor new_lod_tensor;
   new_lod_tensor.CopyLOD(*lod_tensor);
-  ASSERT_NE(new_lod_tensor.lod(), lod_tensor->lod());
+  bool equals = std::equal(lod_tensor->lod().begin(), lod_tensor->lod().end(),
+                           new_lod_tensor.lod().begin());
+  ASSERT_TRUE(equals);
 }
 
 }  // namespace framework

paddle/framework/op_proto.proto

-/* 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. */
-
-// Protocol Message for 3rd-party language binding.
-//
-// Paddle Python package will use `OpProto` to generate op creation methods.
-// The op creation methods take user's input and generate `OpDesc` proto
-// message,
-// then pass `OpDesc` to C++ side and create Op pointer.
-//
-syntax = "proto2";
-package paddle.framework;
-
-import "attribute.proto";
-
-// Attribute protocol message for 3rd-party language binding.
-// It will store the Op support what attribute and what type.
-message AttrProto {
-  // Supported attribute name. e.g. `scale` for cosine op.
-  required string name = 1;
-
-  // Supported attribute type.
-  required AttrType type = 2;
-
-  // Supported attribute comments. It helps 3rd-party language generate
-  // doc-string.
-  required string comment = 3;
-
-  // If that attribute is generated, it means the Paddle third language
-  // binding has responsibility to fill that attribute. End-User should
-  // not set that attribute.
-  optional bool generated = 4 [ default = false ];
-}
-
-// Input or output message for 3rd-party language binding.
-// It contains parameter name and its comments.
-message VarProto {
-  // Input or output name in that op creation function.
-  // e.g. `cos(a, b, output, ...)`, "a", "b", "output" are names.
-  required string name = 1;
-
-  // The comment for that input. It helps 3rd-party language generate
-  // doc-string.
-  required string comment = 2;
-
-  // Is that input/output could be a list or not.
-  // If so, that Op should write a attributed named `input_format` or
-  // `output_format`.
-  //
-  // e.g.
-  //   If the op is a fc op, the inputs are `X`, `W`, `b`. The `X` and `W`
-  //   could be multiple, so the multiple of `X` and `W` is True, and OpDesc
-  //   will hold a attribute of them.
-  //
-  //   The Op desc of same fc could be
-  //   {
-  //      "type": "fc",
-  //      "input": ["X1", "X2", "W1", "W2", "b"],
-  //      "output": "fc.out",
-  //      "attrs" : {
-  //        "input_format": [0, 2, 4, 5]
-  //      }
-  //   }
-  //
-  optional bool multiple = 3 [ default = false ];
-
-  // It marks that output is a temporary output. That output is not used by
-  // user, but used by other op internally as input. If other op is not use
-  // that output, it could be optimized early.
-  //
-  // Attribute temporary_index will be set in OpDesc if there is some
-  // outputs are temporary.
-  //
-  // output = [ "xxx.out1", "xxx.tmp", "xxx.out2"],
-  // attrs = {
-  //   "temporary_index": [1]
-  // }
-  optional bool temporary = 4 [ default = false ];
-
-  // The gradient of operator can be ignored immediately
-  // e.g. operator AddOp, y = x1 + x2, the gradient of dy/dx1, dy/dx2
-  // can be ignored for the future optimized on graph.
-  optional bool ignore_gradient = 6;
-}
-
-// Op protocol message for 3rd-party language binding.
-// It contains all information for generating op creation method.
-message OpProto {
-  // The input information to generate op creation method.
-  repeated VarProto inputs = 1;
-
-  // The output information to generate op creation method.
-  repeated VarProto outputs = 2;
-
-  // The attribute information to generate op creation method.
-  repeated AttrProto attrs = 3;
-
-  // The comments for that Op. It helps 3rd-party language generate
-  // doc-string. The whole documentation of that Op is generated by comment,
-  // inputs, outputs, attrs together.
-  required string comment = 4;
-
-  // The type of that Op.
-  required string type = 5;
-}

paddle/framework/op_proto_test.cc

-#include <gtest/gtest.h>
-#include <paddle/framework/op_proto.pb.h>
-
-TEST(TestOpProto, ALL) {
-  paddle::framework::OpProto proto;
-  {
-    auto ipt = proto.mutable_inputs()->Add();
-    *ipt->mutable_name() = "a";
-    *ipt->mutable_comment() = "the one input of cosine op";
-  }
-  {
-    auto ipt = proto.mutable_inputs()->Add();
-    *ipt->mutable_name() = "b";
-    *ipt->mutable_comment() = "the other input of cosine op";
-  }
-  {
-    auto opt = proto.mutable_outputs()->Add();
-    *opt->mutable_name() = "output";
-    *opt->mutable_comment() = "the output of cosine op";
-  }
-  {
-    auto attr = proto.mutable_attrs()->Add();
-    *attr->mutable_name() = "scale";
-    attr->set_type(paddle::framework::AttrType::FLOAT);
-    *attr->mutable_comment() = "the scale attribute of cosine op";
-  }
-  proto.set_type("cos");
-  *proto.mutable_comment() = "cosine op, output = scale * cos(a, b)";
-
-  ASSERT_TRUE(proto.IsInitialized());
-}
\ No newline at end of file

paddle/framework/op_registry.cc

@@ -17,5 +17,48 @@ limitations under the License. */
 #include <vector>
 
 namespace paddle {
-namespace framework {}  // namespace framework
+namespace framework {
+
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
+                                                   const VarNameMap& inputs,
+                                                   const VarNameMap& outputs,
+                                                   AttributeMap attrs) {
+  auto it = op_info_map().find(type);
+  PADDLE_ENFORCE(it != op_info_map().end(),
+                 "Operator '%s' has not been registered.", type);
+  it->second.checker_->Check(attrs);
+  auto op = it->second.creator_(type, inputs, outputs, attrs);
+  return std::unique_ptr<OperatorBase>(op);
+}
+
+std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
+  VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
+  VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
+  AttributeMap attrs;
+  for (auto& attr : op_desc.attrs()) {
+    attrs[attr.name()] = GetAttrValue(attr);
+  }
+
+  return CreateOp(op_desc.type(), inputs, outputs, attrs);
+}
+
+OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
+    const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
+  VarNameMap ret_val;
+  for (auto& var : op_desc_vars) {
+    auto& var_names = ret_val[var.parameter()];
+    auto& var_names_in_proto = var.arguments();
+    var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
+    std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
+              std::back_inserter(var_names));
+  }
+  return ret_val;
+}
+
+std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
+  PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
+  return std::unique_ptr<OperatorBase>(BuildGradOp(&op));
+}
+
+}  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry_test.cc

@@ -7,6 +7,7 @@ namespace paddle {
 namespace framework {
 class CosineOp : public OperatorBase {
  public:
+  using OperatorBase::OperatorBase;
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}
   void InferShape(const Scope& scope) const override {}
@@ -27,6 +28,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 
 class MyTestOp : public OperatorBase {
  public:
+  using OperatorBase::OperatorBase;
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}
@@ -36,8 +38,8 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
   MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("input", "input of cosine op").SetMultiple();
-    AddOutput("output", "output of cosine op").SetTemporary();
+    AddInput("input", "input of cosine op").AsDuplicable();
+    AddOutput("output", "output of cosine op").AsIntermediate();
     auto my_checker = [](int i) {
       PADDLE_ENFORCE(i % 2 == 0, "'test_attr' must be even!");
     };
@@ -49,16 +51,24 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_OP(cos_sim, paddle::framework::CosineOp,
-            paddle::framework::CosineOpProtoAndCheckerMaker);
-REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
-            paddle::framework::MyTestOpProtoAndCheckerMaker);
+static void BuildVar(const std::string& param_name,
+                     std::initializer_list<const char*> arguments,
+                     paddle::framework::OpDesc::Var* var) {
+  var->set_parameter(param_name);
+  for (auto& arg_name : arguments) {
+    var->add_arguments(arg_name);
+  }
+}
+REGISTER_OP_WITHOUT_GRADIENT(cos_sim, paddle::framework::CosineOp,
+                             paddle::framework::CosineOpProtoAndCheckerMaker);
+REGISTER_OP_WITHOUT_GRADIENT(my_test_op, paddle::framework::MyTestOp,
+                             paddle::framework::MyTestOpProtoAndCheckerMaker);
 
 TEST(OpRegistry, CreateOp) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("cos_sim");
-  op_desc.add_inputs("aa");
-  op_desc.add_outputs("bb");
+  BuildVar("input", {"aa"}, op_desc.add_inputs());
+  BuildVar("output", {"bb"}, op_desc.add_outputs());
 
   float scale = 3.3;
   auto attr = op_desc.mutable_attrs()->Add();
@@ -66,8 +76,7 @@ TEST(OpRegistry, CreateOp) {
   attr->set_type(paddle::framework::AttrType::FLOAT);
   attr->set_f(scale);
 
-  std::shared_ptr<paddle::framework::OperatorBase> op =
-      paddle::framework::OpRegistry::CreateOp(op_desc);
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
@@ -78,8 +87,8 @@ TEST(OpRegistry, CreateOp) {
 TEST(OpRegistry, IllegalAttr) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("cos_sim");
-  op_desc.add_inputs("aa");
-  op_desc.add_outputs("bb");
+  BuildVar("input", {"aa"}, op_desc.add_inputs());
+  BuildVar("output", {"bb"}, op_desc.add_outputs());
 
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
@@ -103,33 +112,23 @@ TEST(OpRegistry, IllegalAttr) {
 TEST(OpRegistry, DefaultValue) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("cos_sim");
-  op_desc.add_inputs("aa");
-  op_desc.add_outputs("bb");
+  BuildVar("input", {"aa"}, op_desc.add_inputs());
+  BuildVar("output", {"bb"}, op_desc.add_outputs());
 
   ASSERT_TRUE(op_desc.IsInitialized());
 
-  std::shared_ptr<paddle::framework::OperatorBase> op =
-      paddle::framework::OpRegistry::CreateOp(op_desc);
+  auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::framework::Scope scope;
   paddle::platform::CPUDeviceContext dev_ctx;
   op->Run(scope, dev_ctx);
   ASSERT_EQ(op->GetAttr<float>("scale"), 1.0);
 }
 
-static void SetInputFormat(paddle::framework::OpDesc* desc) {
-  auto attr = desc->add_attrs();
-  attr->set_name("input_format");
-  attr->set_type(paddle::framework::INTS);
-  attr->mutable_ints()->Add(0);
-  attr->mutable_ints()->Add(1);
-}
-
 TEST(OpRegistry, CustomChecker) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("my_test_op");
-  op_desc.add_inputs("ii");
-  op_desc.add_outputs("oo");
-  SetInputFormat(&op_desc);
+  BuildVar("input", {"ii"}, op_desc.add_inputs());
+  BuildVar("output", {"oo"}, op_desc.add_outputs());
 
   // attr 'test_attr' is not set
   bool caught = false;
@@ -169,7 +168,6 @@ TEST(OpRegistry, CustomChecker) {
   attr->set_name("test_attr");
   attr->set_type(paddle::framework::AttrType::INT);
   attr->set_i(4);
-  SetInputFormat(&op_desc);
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   paddle::platform::CPUDeviceContext dev_ctx;
   paddle::framework::Scope scope;

paddle/framework/operator.cc

@@ -12,9 +12,9 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <algorithm>
-
 #include "paddle/framework/operator.h"
+#include <algorithm>
+#include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace framework {
@@ -34,83 +34,172 @@ ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
 #endif
 
 const std::string& OperatorBase::Input(const std::string& name) const {
-  PADDLE_ENFORCE_NOT_NULL(in_out_idxs_,
-                          "Input Output Indices could not be nullptr");
-  auto it = in_out_idxs_->find(name);
-  PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_",
-                 name);
-  if (attrs_.count("input_format") == 0) {
-    return inputs_.at((size_t)it->second);
-  } else {
-    const auto& input_format = GetAttr<std::vector<int>>("input_format");
-    int idx = input_format[it->second];
-    return inputs_.at((size_t)idx);
-  }
+  auto& ins = Inputs(name);
+  PADDLE_ENFORCE_EQ(ins.size(), 1UL,
+                    "Op %s input %s should contain only one variable", type_,
+                    name);
+  return ins[0];
 }
 
-std::vector<std::string> OperatorBase::Inputs(const std::string& name) const {
-  PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "IO Idx could not be nullptr");
-  auto input_format = GetAttr<std::vector<int>>("input_format");
-  auto offset = in_out_idxs_->at(name);
-  PADDLE_ENFORCE(input_format.at(static_cast<size_t>(offset) + 1) <=
-                     static_cast<int>(inputs_.size()),
-                 "Input Out Of Range");
-
-  return std::vector<std::string>{
-      inputs_.begin() + input_format.at(offset),
-      inputs_.begin() + input_format.at(offset + 1)};
+const std::vector<std::string>& OperatorBase::Inputs(
+    const std::string& name) const {
+  auto it = inputs_.find(name);
+  PADDLE_ENFORCE(it != inputs_.end(), "Op %s do not have input %s", type_,
+                 name);
+  return it->second;
 }
 
 const std::string& OperatorBase::Output(const std::string& name) const {
-  PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "InOut Indice could not be nullptr");
-  auto it = in_out_idxs_->find(name);
-  PADDLE_ENFORCE(it != in_out_idxs_->end(), "no key [%s] in in_out_idxs_",
-                 name);
-  if (attrs_.count("output_format") == 0) {
-    return outputs_.at((size_t)it->second);
-  } else {
-    const auto& output_format = GetAttr<std::vector<int>>("output_format");
-    int idx = output_format[it->second];
-    return outputs_.at((size_t)idx);
-  }
+  auto& outs = Outputs(name);
+  PADDLE_ENFORCE_EQ(outs.size(), 1UL,
+                    "Op %s output %s should contain only one variable", type_,
+                    name);
+  return outs[0];
 }
 
-std::vector<std::string> OperatorBase::Outputs(const std::string& name) const {
-  PADDLE_ENFORCE_NOT_NULL(in_out_idxs_, "InOut Indice could not be nullptr");
-  auto output_format = GetAttr<std::vector<int>>("output_format");
-  auto offset = in_out_idxs_->at(name);
-  PADDLE_ENFORCE(output_format.at(static_cast<size_t>(offset) + 1) <=
-                     static_cast<int>(outputs_.size()),
-                 "Output Out of Range");
-  return std::vector<std::string>{
-      outputs_.begin() + output_format.at(offset),
-      outputs_.begin() + output_format.at(offset + 1)};
+const std::vector<std::string>& OperatorBase::Outputs(
+    const std::string& name) const {
+  auto it = outputs_.find(name);
+  PADDLE_ENFORCE(it != outputs_.end(), "Op %s does not have output %s", type_,
+                 name);
+  return it->second;
 }
 
 std::string OperatorBase::DebugString() const {
   std::stringstream ss;
-  ss << "Op(" << type_ << "), inputs:(";
-  for (size_t i = 0; i < inputs_.size(); ++i) {
-    ss << inputs_[i];
-    if (i != inputs_.size() - 1) {
+  ss << "Op(" << type_ << "), inputs:{";
+  for (auto it = inputs_.begin(); it != inputs_.end();) {
+    auto& input = *it;
+    ss << input.first << "[";
+    for (size_t i = 0; i < input.second.size(); ++i) {
+      ss << input.second[i];
+      if (i != input.second.size() - 1) {
+        ss << ", ";
+      }
+    }
+    ss << "]";
+    ++it;
+    if (it != inputs_.end()) {
       ss << ", ";
     }
   }
-  ss << "), outputs:(";
-  for (size_t i = 0; i < outputs_.size(); ++i) {
-    ss << outputs_[i];
-    if (i != outputs_.size() - 1) {
+  ss << "}, outputs:{";
+  for (auto it = outputs_.begin(); it != outputs_.end();) {
+    auto& output = *it;
+    ss << output.first << "[";
+    for (size_t i = 0; i < output.second.size(); ++i) {
+      ss << output.second[i];
+      if (i != output.second.size() - 1) {
+        ss << ", ";
+      }
+    }
+    ss << "]";
+    ++it;
+    if (it != outputs_.end()) {
       ss << ", ";
     }
   }
-  ss << ").";
+  ss << "}.";
   return ss.str();
 }
 
 void OperatorBase::Rename(const std::string& old_name,
                           const std::string& new_name) {
-  std::replace(inputs_.begin(), inputs_.end(), old_name, new_name);
-  std::replace(outputs_.begin(), outputs_.end(), old_name, new_name);
+  for (auto& input : inputs_) {
+    std::replace(input.second.begin(), input.second.end(), old_name, new_name);
+  }
+  for (auto& output : outputs_) {
+    std::replace(output.second.begin(), output.second.end(), old_name,
+                 new_name);
+  }
+}
+
+OperatorBase::OperatorBase(const std::string& type,
+                           const OperatorBase::VarNameMap& inputs,
+                           const OperatorBase::VarNameMap& outputs,
+                           const AttributeMap& attrs)
+    : type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
+  static std::atomic<size_t> gUniqId(0UL);
+  for (auto& output : outputs_) {
+    for (auto& output_name : output.second) {
+      if (output_name == kTempVarName) {
+        output_name += type_;
+        output_name += "@";
+        output_name += std::to_string(gUniqId.fetch_add(1));
+      }
+    }
+  }
+}
+
+std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
+  std::vector<std::string> ret_val;
+  if (has_intermediate) {
+    // push all outputs into 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;
+  }
+  auto it = OpRegistry::op_info_map().find(type_);
+  PADDLE_ENFORCE(
+      it != OpRegistry::op_info_map().end(),
+      "Operator %s not registered, cannot figure out intermediate outputs",
+      type_);
+  PADDLE_ENFORCE(
+      it->second.proto_ != nullptr,
+      "Operator %s has no OpProto, cannot figure out intermediate outputs",
+      type_);
+
+  // get all OpProto::Var for outputs
+  for (auto& o : it->second.proto_->outputs()) {
+    // ignore all intermediate output
+    if (o.intermediate()) continue;
+    auto out = outputs_.find(o.name());
+    if (out != outputs_.end()) {
+      ret_val.reserve(ret_val.size() + out->second.size());
+      ret_val.insert(ret_val.end(), out->second.begin(), out->second.end());
+    }
+  }
+  return ret_val;
+}
+
+void OpProtoAndCheckerMaker::Validate() {
+  validated_ = true;
+  CheckNoDuplicatedInOutAttrs();
+}
+
+OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput(
+    const std::string& name, const std::string& comment) {
+  auto* input = proto_->add_inputs();
+  input->set_name(name);
+  input->set_comment(comment);
+  return OpProtoAndCheckerMaker::VariableBuilder{input};
+}
+
+OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput(
+    const std::string& name, const std::string& comment) {
+  auto* output = proto_->add_outputs();
+  output->set_name(name);
+  output->set_comment(comment);
+  return OpProtoAndCheckerMaker::VariableBuilder{output};
+}
+
+void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() {
+  std::unordered_set<std::string> names;
+  auto checker = [&](const std::string& name) {
+    PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
+    names.insert(name);
+  };
+  for (auto& attr : proto_->attrs()) {
+    checker(attr.name());
+  }
+  for (auto& input : proto_->inputs()) {
+    checker(input.name());
+  }
+  for (auto& output : proto_->outputs()) {
+    checker(output.name());
+  }
 }
 
 }  // namespace framework

paddle/framework/operator.h

@@ -20,8 +20,7 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/attribute.h"
-#include "paddle/framework/op_desc.pb.h"
-#include "paddle/framework/op_proto.pb.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/framework/scope.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
@@ -63,6 +62,11 @@ class ExecutionContext;
  */
 class OperatorBase {
  public:
+  using VarNameMap = std::map<std::string, std::vector<std::string>>;
+
+  OperatorBase(const std::string& type, const VarNameMap& inputs,
+               const VarNameMap& outputs, const AttributeMap& attrs);
+
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -74,10 +78,6 @@ class OperatorBase {
 
   virtual std::string DebugString() const;
 
-  /// Init will be called after CreateOperator, you can put some initialization
-  /// logic here.
-  virtual void Init() {}
-
   /// InferShape infer the size of Variables used by this Operator with
   /// information inside scope
   virtual void InferShape(const Scope& scope) const = 0;
@@ -93,31 +93,134 @@ class OperatorBase {
   /// rename inputs outputs name
   void Rename(const std::string& old_name, const std::string& new_name);
 
+  const VarNameMap& Inputs() const { return inputs_; }
+  const VarNameMap& Outputs() const { return outputs_; }
   //! Get a input with argument's name described in `op_proto`
   const std::string& Input(const std::string& name) const;
-
   //! Get a input which has multiple variables.
-  //! TODO add a vector_view to prevent memory copy.
-  std::vector<std::string> Inputs(const std::string& name) const;
+  const std::vector<std::string>& Inputs(const std::string& name) const;
+
   //! Get a output with argument's name described in `op_proto`
   const std::string& Output(const std::string& name) const;
   //! Get an output which has multiple variables.
   //! TODO add a vector_view to prevent memory copy.
-  std::vector<std::string> Outputs(const std::string& name) const;
+  const std::vector<std::string>& Outputs(const std::string& name) const;
 
- public:
+  virtual std::vector<std::string> OutputVars(bool has_intermediate) const;
+
+  const std::string& Type() const { return type_; }
+  void SetType(const std::string& type) { type_ = type; }
+  const AttributeMap& Attrs() const { return attrs_; }
+
+  // Return a new operator instance, which is as same as this.
+  // Use unique_ptr to prevent caller forget to delete this pointer.
+  virtual std::unique_ptr<OperatorBase> Clone() const = 0;
+
+ protected:
   std::string type_;
   // NOTE: in case of OpGrad, inputs_ contains:
-  // I (Inputs)
+  // I (Inputs)opear
   // O (Outputs)
   // OG (Output Gradients)
-  std::vector<std::string> inputs_;
+  VarNameMap inputs_;
+
   // NOTE: in case of OpGrad, outputs_ contains
   // IG (Inputs Gradients)
-  std::vector<std::string> outputs_;
+  VarNameMap outputs_;
   AttributeMap attrs_;
-  // store the arguments' offset described in op_desc.
-  std::shared_ptr<std::unordered_map<std::string, int>> in_out_idxs_;
+};
+
+// Macro for define a clone method.
+// If you are writing an kernel operator, `Clone` will be defined when you
+// register it. i.e. `Clone` method is not needed to define by yourself.
+#define DEFINE_OP_CLONE_METHOD(CLS)                       \
+  std::unique_ptr<OperatorBase> Clone() const final {     \
+    return std::unique_ptr<OperatorBase>(new CLS(*this)); \
+  }
+
+// Macro for define a default constructor for Operator.
+// You can also use
+//   using PARENT_CLASS::PARENT_CLASS;
+// to use parent's constructor.
+#define DEFINE_OP_CONSTRUCTOR(CLS, PARENT_CLS)                                 \
+  CLS(const std::string& type, const VarNameMap& inputs,                       \
+      const VarNameMap& outputs, const paddle::framework::AttributeMap& attrs) \
+      : PARENT_CLS(type, inputs, outputs, attrs) {}
+
+class NOP : public OperatorBase {
+ public:
+  using OperatorBase::OperatorBase;
+  void InferShape(const Scope& scope) const override {}
+  void Run(const Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {}
+  std::unique_ptr<OperatorBase> Clone() const override {
+    return std::unique_ptr<OperatorBase>(new NOP(*this));
+  }
+};
+
+// this class not only make proto but also init attribute checkers.
+class OpProtoAndCheckerMaker {
+ public:
+  OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : proto_(proto), op_checker_(op_checker) {}
+
+  ~OpProtoAndCheckerMaker() {
+    PADDLE_ENFORCE(validated_, "should call Validate after build");
+  }
+
+  void Validate();
+
+ protected:
+  struct VariableBuilder {
+    OpProto::Var* var_;
+
+    VariableBuilder& AsDuplicable() {
+      var_->set_duplicable(true);
+      return *this;
+    }
+
+    VariableBuilder& AsIntermediate() {
+      var_->set_intermediate(true);
+      return *this;
+    }
+
+    VariableBuilder& NotInGradient() {
+      var_->set_not_in_gradient(true);
+      return *this;
+    }
+  };
+
+  VariableBuilder AddInput(const std::string& name, const std::string& comment);
+
+  VariableBuilder AddOutput(const std::string& name,
+                            const std::string& comment);
+
+  template <typename T>
+  TypedAttrChecker<T>& AddAttr(const std::string& name,
+                               const std::string& comment,
+                               bool generated = false) {
+    auto* attr = proto_->add_attrs();
+    attr->set_name(name);
+    attr->set_comment(comment);
+    attr->set_generated(generated);
+    attr->set_type(AttrTypeID<T>());
+    return op_checker_->AddAttrChecker<T>(name);
+  }
+
+  void AddComment(const std::string& comment) { proto_->set_comment(comment); }
+
+ private:
+  void CheckNoDuplicatedInOutAttrs();
+
+  OpProto* proto_;
+  OpAttrChecker* op_checker_;
+  bool validated_{false};
+};
+
+class NOPMaker : public OpProtoAndCheckerMaker {
+ public:
+  NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {}
 };
 
 class InferShapeContext {
@@ -125,16 +228,12 @@ class InferShapeContext {
   InferShapeContext(const OperatorBase& op, const Scope& scope)
       : op_(op), scope_(scope) {}
 
-  size_t InputSize() const { return op_.inputs_.size(); }
-
-  size_t OutputSize() const { return op_.outputs_.size(); }
-
-  const Variable* InputVar(const size_t index) const {
-    return scope_.FindVar(op_.inputs_.at(index));
+  size_t InputSize(const std::string& name) const {
+    return op_.Inputs(name).size();
   }
 
-  Variable* OutputVar(const size_t index) const {
-    return scope_.FindVar(op_.outputs_.at(index));
+  size_t OutputSize(const std::string& name) const {
+    return op_.Outputs(name).size();
   }
 
   const Variable* InputVar(const std::string& name) const {
@@ -166,27 +265,9 @@ class InferShapeContext {
     return res;
   }
 
-  template <typename T>
-  const T* Input(const size_t index) const {
-    auto var = InputVar(index);
-    PADDLE_ENFORCE_NOT_NULL(var, "Input(%d) should not be nullptr", index);
-    return &var->Get<T>();
-  }
-
-  template <typename T>
-  T* Output(const size_t index) const {
-    auto var = OutputVar(index);
-    PADDLE_ENFORCE_NOT_NULL(
-        var,
-        "Output(%d) not be nullptr, which means variable [%s] does not "
-        "exist in scope",
-        index, op_.outputs_[index]);
-    return var->GetMutable<T>();
-  }
-
   template <typename T>
   const T* Input(const std::string& name) const {
-    auto var = InputVar(name);
+    auto* var = InputVar(name);
     PADDLE_ENFORCE_NOT_NULL(var, "Input(%s) should not be nullptr", name);
     return &var->Get<T>();
   }
@@ -223,7 +304,7 @@ class InferShapeContext {
                    [&](const std::string& sub_name) {
                      auto var = scope_.FindVar(sub_name);
                      PADDLE_ENFORCE_NOT_NULL(
-                         var, "MultiOutput(%s:%s) should not be nullptr", name,
+                         var, "MultiOutput(%s:%s) should not be nullptr.", name,
                          sub_name);
                      return var->GetMutable<T>();
                    });
@@ -262,6 +343,10 @@ class ExecutionContext : public InferShapeContext {
 
   platform::Place GetPlace() const { return device_context_->GetPlace(); }
 
+  const platform::DeviceContext* device_context() const {
+    return device_context_;
+  }
+
   const platform::DeviceContext* device_context_;
 };
 
@@ -304,6 +389,10 @@ class OperatorWithKernel : public OperatorBase {
   using OpKernelMap =
       std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;
 
+  OperatorWithKernel(const std::string& type, const VarNameMap& inputs,
+                     const VarNameMap& outputs, const AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+
   void InferShape(const Scope& scope) const override {
     InferShape(InferShapeContext(*this, scope));
   }

paddle/framework/operator_test.cc

@@ -23,20 +23,22 @@ static int op_run_num = 0;
 
 class OpWithoutKernelTest : public OperatorBase {
  public:
-  void Init() override { x = 1; }
+  OpWithoutKernelTest(const std::string& type, const VarNameMap& inputs,
+                      const VarNameMap& outputs, const AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs), x(1) {}
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {
-    op_run_num++;
-    ASSERT_EQ((int)inputs_.size(), 1);
-    ASSERT_EQ((int)outputs_.size(), 1);
-    ASSERT_EQ(scope.FindVar(inputs_[0]), nullptr);
+    ++op_run_num;
+    ASSERT_EQ(static_cast<int>(inputs_.size()), 1);
+    ASSERT_EQ(static_cast<int>(outputs_.size()), 1);
+    ASSERT_EQ(scope.FindVar(inputs_.at("input")[0]), nullptr);
     ASSERT_EQ(x, 1);
-    ASSERT_NE(scope.FindVar(outputs_[0]), nullptr);
+    ASSERT_NE(scope.FindVar(outputs_.at("output")[0]), nullptr);
   }
 
  public:
-  float x = 0;
+  int x{0};
 };
 
 class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
@@ -54,14 +56,25 @@ class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest,
-            paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
+static void BuildVar(const std::string& param_name,
+                     std::initializer_list<const char*> arguments,
+                     paddle::framework::OpDesc::Var* var) {
+  var->set_parameter(param_name);
+  for (auto& arg_name : arguments) {
+    *var->mutable_arguments()->Add() = arg_name;
+  }
+}
+
+REGISTER_OP_WITHOUT_GRADIENT(
+    test_operator, paddle::framework::OpWithoutKernelTest,
+    paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
 
 TEST(OperatorBase, all) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("test_operator");
-  *op_desc.mutable_inputs()->Add() = "IN1";
-  *op_desc.mutable_outputs()->Add() = "OUT1";
+  BuildVar("input", {"IN1"}, op_desc.add_inputs());
+  BuildVar("output", {"OUT1"}, op_desc.add_outputs());
+
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
   attr->set_type(paddle::framework::AttrType::FLOAT);
@@ -97,6 +110,9 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 static int cpu_kernel_run_num = 0;
 
 class OpWithKernelTest : public OperatorWithKernel {
+ public:
+  using OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {}
 };
@@ -113,33 +129,15 @@ class CPUKernelTest : public OpKernel {
   }
 };
 
-// multiple inputs test
-class OperatorMultiInputsTest : public OperatorBase {
- public:
-  void Init() override { x = 1; }
-  void InferShape(const Scope& scope) const override {}
-  void Run(const Scope& scope,
-           const platform::DeviceContext& dev_ctx) const override {
-    ASSERT_EQ(scope.FindVar(inputs_[0]), nullptr);
-    ASSERT_EQ(x, 1);
-    ASSERT_NE(scope.FindVar(outputs_[0]), nullptr);
-    ASSERT_EQ(Input("x"), "IN1");
-    ASSERT_EQ(Input("y"), "OUT1");
-  }
-
- public:
-  float x = 0;
-};
-
 class OpKernelTestMultiInputsProtoAndCheckerMaker
     : public OpProtoAndCheckerMaker {
  public:
   OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto,
                                               OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("xs", "inputs of test op").SetMultiple();
+    AddInput("xs", "inputs of test op").AsDuplicable();
     AddInput("k", "input of test op");
-    AddOutput("ys", "outputs of test op").SetMultiple();
+    AddOutput("ys", "outputs of test op").AsDuplicable();
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .LargerThan(0.0);
@@ -187,8 +185,9 @@ class CPUKernalMultiInputsTest : public OpKernel {
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_OP(op_with_kernel, paddle::framework::OpWithKernelTest,
-            paddle::framework::OpKernelTestProtoAndCheckerMaker);
+REGISTER_OP_WITHOUT_GRADIENT(
+    op_with_kernel, paddle::framework::OpWithKernelTest,
+    paddle::framework::OpKernelTestProtoAndCheckerMaker);
 REGISTER_OP_CPU_KERNEL(op_with_kernel,
                        paddle::framework::CPUKernelTest<float, float>);
 
@@ -196,8 +195,9 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel,
 TEST(OpKernel, all) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("op_with_kernel");
-  *op_desc.mutable_inputs()->Add() = "IN1";
-  *op_desc.mutable_outputs()->Add() = "OUT1";
+  BuildVar("x", {"IN1"}, op_desc.add_inputs());
+  BuildVar("y", {"OUT1"}, op_desc.add_outputs());
+
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
   attr->set_type(paddle::framework::AttrType::FLOAT);
@@ -212,8 +212,9 @@ TEST(OpKernel, all) {
   ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
 }
 
-REGISTER_OP(op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
-            paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
+REGISTER_OP_WITHOUT_GRADIENT(
+    op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
+    paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
 REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
                        paddle::framework::CPUKernalMultiInputsTest);
 
@@ -223,32 +224,15 @@ TEST(OpKernel, multi_inputs) {
 
   OpDesc op_desc;
   op_desc.set_type("op_multi_inputs_with_kernel");
-  *op_desc.mutable_inputs()->Add() = "x0";
-  *op_desc.mutable_inputs()->Add() = "x1";
-  *op_desc.mutable_inputs()->Add() = "x2";
-  *op_desc.mutable_inputs()->Add() = "k0";
-  *op_desc.mutable_outputs()->Add() = "y0";
-  *op_desc.mutable_outputs()->Add() = "y1";
+  BuildVar("xs", {"x0", "x1", "x2"}, op_desc.add_inputs());
+  BuildVar("k", {"k0"}, op_desc.add_inputs());
+  BuildVar("ys", {"y0", "y1"}, op_desc.add_outputs());
+
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
   attr->set_type(paddle::framework::AttrType::FLOAT);
   attr->set_f(3.14);
 
-  auto attr0 = op_desc.mutable_attrs()->Add();
-  attr0->set_name("input_format");
-  attr0->set_type(paddle::framework::AttrType::INTS);
-  auto input_format = attr0->mutable_ints();
-  input_format->Add(0);  // x0
-  input_format->Add(3);  // k
-  input_format->Add(4);  // end
-
-  auto attr1 = op_desc.mutable_attrs()->Add();
-  attr1->set_name("output_format");
-  attr1->set_type(paddle::framework::AttrType::INTS);
-  auto output_format = attr1->mutable_ints();
-  output_format->Add(0);  // y0
-  output_format->Add(2);  // y1
-
   paddle::platform::CPUDeviceContext cpu_device_context;
   paddle::framework::Scope scope;
   scope.NewVar("x0")->GetMutable<Tensor>();
@@ -261,3 +245,21 @@ TEST(OpKernel, multi_inputs) {
   auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
   op->Run(scope, cpu_device_context);
 }
+
+class OperatorClone : public paddle::framework::OperatorBase {
+ public:
+  DEFINE_OP_CLONE_METHOD(OperatorClone);
+  OperatorClone(const std::string& type, const VarNameMap& inputs,
+                const VarNameMap& outputs,
+                const paddle::framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void InferShape(const paddle::framework::Scope& scope) const override {}
+  void Run(const paddle::framework::Scope& scope,
+           const paddle::platform::DeviceContext& dev_ctx) const override {}
+};
+
+TEST(Operator, Clone) {
+  OperatorClone a("ABC", {}, {}, {});
+  auto b = a.Clone();
+  ASSERT_EQ(a.Type(), b->Type());
+}
\ No newline at end of file

paddle/framework/pybind.cc

@@ -20,6 +20,7 @@ limitations under the License. */
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/tensor_py.h"
 #include "paddle/operators/net_op.h"
+#include "paddle/operators/recurrent_op.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
 #include "paddle/string/to_string.h"
@@ -30,8 +31,7 @@ limitations under the License. */
 namespace py = pybind11;
 
 USE_OP(add_two);
-USE_OP_CPU(onehot_cross_entropy);
-USE_OP_WITHOUT_KERNEL(fc);
+USE_OP(onehot_cross_entropy);
 USE_OP(sgd);
 USE_OP(mul);
 USE_OP(mean);
@@ -39,7 +39,7 @@ USE_OP(sigmoid);
 USE_OP(softmax);
 USE_OP(rowwise_add);
 USE_OP(fill_zeros_like);
-USE_OP_WITHOUT_KERNEL(recurrent_op);
+USE_OP_ITSELF(recurrent_op);
 USE_OP(gaussian_random);
 USE_OP(uniform_random);
 
@@ -48,41 +48,6 @@ namespace framework {
 
 using Tensor = framework::Tensor;
 
-template <typename ClassType>
-void ExposeOperator(ClassType &m) {
-  m.def("infer_shape", &ClassType::type::InferShape)
-      .def("run", &ClassType::type::Run)
-      .def("type",
-           [](const typename ClassType::type &op) -> std::string {
-             return op.type_;
-           })
-      .def("outputs",
-           [](const typename ClassType::type &op) -> std::vector<std::string> {
-             return op.outputs_;
-           })
-      .def("inputs",
-           [](const typename ClassType::type &op) -> std::vector<std::string> {
-             return op.inputs_;
-           })
-      .def("support_gpu", &ClassType::type::SupportGPU)
-      .def("temp_outputs",
-           [](const typename ClassType::type &op) -> std::vector<std::string> {
-             auto iter = op.attrs_.find("temporary_index");
-             std::vector<std::string> ret;
-             if (iter == op.attrs_.end()) {
-               return ret;
-             } else {
-               auto tmp_idx = boost::get<std::vector<int>>(iter->second);
-               ret.reserve(tmp_idx.size());
-               for (auto &index : tmp_idx) {
-                 ret.push_back(op.outputs_.at(index));
-               }
-               return ret;
-             }
-           })
-      .def("__str__", &ClassType::type::DebugString);
-}
-
 static size_t UniqueIntegerGenerator() {
   static std::atomic<size_t> generator;
   return generator.fetch_add(1);
@@ -173,13 +138,16 @@ All parameter, weight, gradient are variables in Paddle.
   //! @note: Be careful! PyBind will return std::string as an unicode, not
   //! Python str. If you want a str object, you should cast them in Python.
   m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
-    auto &protos = OpRegistry::protos();
+    auto &op_info_map = OpRegistry::op_info_map();
     std::vector<py::bytes> ret_values;
-    for (auto it = protos.begin(); it != protos.end(); ++it) {
-      PADDLE_ENFORCE(it->second.IsInitialized(),
-                     "OpProto must all be initialized");
+    for (auto it = op_info_map.begin(); it != op_info_map.end(); ++it) {
+      const OpProto *proto = it->second.proto_;
+      if (proto == nullptr) {
+        continue;
+      }
+      PADDLE_ENFORCE(proto->IsInitialized(), "OpProto must all be initialized");
       std::string str;
-      PADDLE_ENFORCE(it->second.SerializeToString(&str),
+      PADDLE_ENFORCE(proto->SerializeToString(&str),
                      "Serialize OpProto Error. This could be a bug of Paddle.");
       ret_values.push_back(py::bytes(str));
     }
@@ -216,47 +184,69 @@ All parameter, weight, gradient are variables in Paddle.
       .def(py::init<>())
       .def("__str__", string::to_string<const platform::CPUPlace &>);
 
-  py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
-      m, "Operator");
-
-  operator_base.def_static("create", [](py::bytes protobin) {
-    OpDesc desc;
-    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
-                   "Cannot parse user input to OpDesc");
-    PADDLE_ENFORCE(desc.IsInitialized(),
-                   "User OpDesc is not initialized, reason %s",
-                   desc.InitializationErrorString());
-    return OpRegistry::CreateOp(desc);
-  });
-
-  operator_base.def("backward",
-                    [](const OperatorBase &forwardOp,
-                       const std::unordered_set<std::string> &no_grad_vars) {
-                      return Backward(forwardOp, no_grad_vars);
-                    });
-
-  ExposeOperator(operator_base);
-
-  py::class_<operators::NetOp, std::shared_ptr<operators::NetOp>> net(m, "Net");
-
-  net.def_static("create",
-                 []() -> std::shared_ptr<operators::NetOp> {
-                   auto retv = std::make_shared<operators::NetOp>();
-                   retv->type_ = "plain_net";
-                   return retv;
-                 })
-      .def("add_op", &operators::NetOp::AddOp)
-      .def("add_op",
-           [](operators::NetOp &self,
-              const std::shared_ptr<operators::NetOp> &net) -> void {
-             self.AddOp(std::static_pointer_cast<OperatorBase>(net));
+  py::class_<OperatorBase>(m, "Operator")
+      .def_static("create",
+                  [](py::bytes protobin) {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    return OpRegistry::CreateOp(desc);
+                  })
+      .def("backward",
+           [](const OperatorBase &forwardOp,
+              const std::unordered_set<std::string> &no_grad_vars) {
+             return Backward(forwardOp, no_grad_vars).release();
            })
+      .def("infer_shape", &OperatorBase::InferShape)
+      .def("run", &OperatorBase::Run)
+      .def("type",
+           [](const OperatorBase &op) -> std::string { return op.Type(); })
+      .def("outputs",
+           [](const OperatorBase &op)
+               -> std::map<std::string, std::vector<std::string>> {
+                 return op.Outputs();
+               })
+      .def("inputs", [](const OperatorBase &op) { return op.Inputs(); })
+      .def("__str__", &OperatorBase::DebugString)
+      .def("no_intermediate_outputs",
+           [](const OperatorBase &op) { return op.OutputVars(false); })
+      .def("support_gpu", &OperatorBase::SupportGPU);
+
+  py::class_<operators::NetOp, OperatorBase>(m, "Net")
+      .def_static("create",
+                  []() -> operators::NetOp * {
+                    auto *retv = new operators::NetOp;
+                    retv->SetType("plain_net");
+                    return retv;
+                  })
+      .def("append_op", [](operators::NetOp &self,
+                           const OperatorBase &op) { self.AppendOp(op); })
       .def("complete_add_op", &operators::NetOp::CompleteAddOp)
       .def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
         self->CompleteAddOp();
       });
 
-  ExposeOperator(net);
+  // recurrent_op
+  py::class_<operators::RecurrentOp, OperatorBase>(m, "RecurrentOp")
+      .def_static(
+          "create",
+          [](py::bytes protobin) -> operators::RecurrentOp * {
+            OpDesc desc;
+            PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                           "Cannot parse user input to OpDesc");
+            PADDLE_ENFORCE(desc.IsInitialized(),
+                           "User OpDesc is not initialized, reason %s",
+                           desc.InitializationErrorString());
+            auto rnn_op = OpRegistry::CreateOp(desc);
+            return static_cast<operators::RecurrentOp *>(rnn_op.release());
+          })
+      .def("set_stepnet", [](operators::RecurrentOp &self,
+                             const operators::NetOp &net) -> void {
+        self.set_stepnet(net.Clone());
+      });
 
   m.def("unique_integer", UniqueIntegerGenerator);
 

paddle/framework/tensor.h

@@ -105,6 +105,8 @@ class Tensor {
   template <typename T>
   inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
 
+  platform::Place place() const { return holder_->place(); }
+
  private:
   template <typename T>
   inline void check_memory_size() const;

paddle/function/CMakeLists.txt

@@ -4,6 +4,10 @@ file(GLOB cpp_files . *Op.cpp)
 list(APPEND h_files Function.h)
 list(APPEND cpp_files Function.cpp)
 list(APPEND cpp_files BufferArg.cpp)
+list(APPEND cpp_files GemmFunctor.cpp)
+if(USE_EIGEN_FOR_BLAS)
+  list(APPEND cpp_files EigenGemm.cpp)
+endif(USE_EIGEN_FOR_BLAS)
 
 if(WITH_GPU)
     file(GLOB cu_files . *OpGpu.cu)
@@ -38,10 +42,11 @@ if(WITH_GPU)
     add_simple_unittest(RowConvOpTest)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
+    add_simple_unittest(DepthwiseConvOpTest)
 endif()
 
-add_simple_unittest(ConvOpTest)
 add_simple_unittest(Im2ColTest)
+add_simple_unittest(GemmConvOpTest)
 endif()
 
 add_style_check_target(paddle_function ${h_files})

paddle/function/ConvOpTest.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 <memory>
-#include "Function.h"
-#include "FunctionTest.h"
-
-namespace paddle {
-
-enum TestType {
-  kForwardTest = 0,
-  kBackwardInputTest = 1,
-  kBackwardFilterTest = 2,
-};
-
-template <DeviceType DType1, DeviceType DType2>
-class ConvolutionTest {
-public:
-  ConvolutionTest(const std::string& conv1,
-                  const std::string& conv2,
-                  TestType type,
-                  bool useGroups = true,
-                  std::string algo = "auto") {
-    for (size_t batchSize : {1, 32}) {
-      for (size_t inputSize : {7, 14, 54}) {
-        for (size_t filterSize : {1, 3, 5}) {
-          for (size_t inputChannels : {3, 64}) {
-            for (size_t outputChannels : {3, 64}) {
-              if (inputChannels > outputChannels) break;
-              size_t groups;
-              if (!useGroups) {
-                groups = 1;
-              } else {
-                if (outputChannels % inputChannels != 0) continue;
-                groups = inputChannels;
-              }
-
-              for (size_t stride : {1, 2}) {
-                for (size_t padding : {0, 1}) {
-                  if (padding >= filterSize) break;
-                  size_t outputSize =
-                      (inputSize - filterSize + 2 * padding + stride) / stride;
-                  VLOG(3) << " batchSize=" << batchSize
-                          << " inputChannels=" << inputChannels
-                          << " inputHeight=" << inputSize
-                          << " inputWidth=" << inputSize
-                          << " outputChannels=" << outputChannels
-                          << " filterHeight=" << filterSize
-                          << " filterWidth=" << filterSize
-                          << " outputHeight=" << outputSize
-                          << " outputWidth=" << outputSize
-                          << " stride=" << stride << " padding=" << padding;
-
-                  std::vector<size_t> paddings = {padding, padding};
-                  std::vector<size_t> strides = {stride, stride};
-                  Compare2Function<DType1, DType2> test(
-                      conv1,
-                      conv2,
-                      FuncConfig()
-                          .set("paddings", paddings)
-                          .set("strides", strides)
-                          .set("groups", groups)
-                          .set("algo", algo));
-
-                  TensorShape input{
-                      batchSize, inputChannels, inputSize, inputSize};
-
-                  TensorShape filter;
-                  if (groups > 1)
-                    filter = TensorShape({groups,
-                                          outputChannels / groups,
-                                          inputChannels / groups,
-                                          filterSize,
-                                          filterSize});
-                  else
-                    filter = TensorShape({outputChannels,
-                                          inputChannels,
-                                          filterSize,
-                                          filterSize});
-                  TensorShape output{
-                      batchSize, outputChannels, outputSize, outputSize};
-
-                  if (type == kForwardTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.run();
-                  } else if (type == kBackwardInputTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO);
-                    test.run();
-                  } else if (type == kBackwardFilterTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter),
-                                    ADD_TO);
-                    test.run();
-                  }
-                }
-              }
-            }
-          }
-        }
-      }
-    }
-  }
-};
-
-// Mainly used to test cases where the height and width (input, filter)
-// are not equal.
-template <DeviceType DType1, DeviceType DType2>
-class ConvolutionTest2 {
-public:
-  ConvolutionTest2(const std::string& conv1,
-                   const std::string& conv2,
-                   TestType type,
-                   bool useGroups = true,
-                   std::string algo = "auto") {
-    for (size_t batchSize : {16}) {
-      for (size_t inputHeight : {7, 31}) {
-        for (size_t inputWidth : {10, 54}) {
-          for (size_t filterHeight : {1, 5}) {
-            for (size_t filterWidth : {3, 7}) {
-              for (size_t inputChannels : {7}) {
-                for (size_t outputChannels : {7}) {
-                  size_t groups;
-                  if (!useGroups) {
-                    groups = 1;
-                  } else {
-                    if (outputChannels % inputChannels != 0) continue;
-                    groups = inputChannels;
-                  }
-
-                  size_t stride = 1;
-                  size_t padding = 0;
-                  size_t outputHeight =
-                      (inputHeight - filterHeight + 2 * padding + stride) /
-                      stride;
-                  size_t outputWidth =
-                      (inputWidth - filterWidth + 2 * padding + stride) /
-                      stride;
-                  VLOG(3) << " batchSize=" << batchSize
-                          << " inputChannels=" << inputChannels
-                          << " inputHeight=" << inputHeight
-                          << " inputWidth=" << inputWidth
-                          << " outputChannels=" << outputChannels
-                          << " filterHeight=" << filterHeight
-                          << " filterWidth=" << filterWidth
-                          << " outputHeight=" << outputHeight
-                          << " outputWidth=" << outputWidth
-                          << " stride=" << stride << " padding=" << padding;
-
-                  std::vector<size_t> paddings = {padding, padding};
-                  std::vector<size_t> strides = {stride, stride};
-                  Compare2Function<DType1, DType2> test(
-                      conv1,
-                      conv2,
-                      FuncConfig()
-                          .set("paddings", paddings)
-                          .set("strides", strides)
-                          .set("groups", groups)
-                          .set("algo", algo));
-
-                  TensorShape input{
-                      batchSize, inputChannels, inputHeight, inputWidth};
-
-                  TensorShape filter;
-                  if (groups > 1)
-                    filter = TensorShape({groups,
-                                          outputChannels / groups,
-                                          inputChannels / groups,
-                                          filterHeight,
-                                          filterWidth});
-                  else
-                    filter = TensorShape({outputChannels,
-                                          inputChannels,
-                                          filterHeight,
-                                          filterWidth});
-                  TensorShape output{
-                      batchSize, outputChannels, outputHeight, outputWidth};
-
-                  if (type == kForwardTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.run();
-                  } else if (type == kBackwardInputTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO);
-                    test.run();
-                  } else if (type == kBackwardFilterTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter),
-                                    ADD_TO);
-                    test.run();
-                  }
-                }
-              }
-            }
-          }
-        }
-      }
-    }
-  }
-};
-
-// ======Start Convolution TEST======
-
-TEST(Forward, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU> test(
-      "NaiveConv-CPU", "GemmConv-CPU", kForwardTest, false);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU> test2(
-      "NaiveConv-CPU", "GemmConv-CPU", kForwardTest, false);
-}
-
-#ifndef PADDLE_ONLY_CPU
-TEST(Forward, GEMM2) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConv-CPU", "GemmConv-GPU", kForwardTest, false);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConv-CPU", "GemmConv-GPU", kForwardTest, false);
-}
-
-TEST(BackwardInput, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConvGradInput-CPU",
-      "GemmConvGradInput-GPU",
-      kBackwardInputTest,
-      false);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConvGradInput-CPU",
-      "GemmConvGradInput-GPU",
-      kBackwardInputTest,
-      false);
-}
-
-TEST(BackwardFilter, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConvGradFilter-CPU",
-      "GemmConvGradFilter-GPU",
-      kBackwardFilterTest,
-      false);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConvGradFilter-CPU",
-      "GemmConvGradFilter-GPU",
-      kBackwardFilterTest,
-      false);
-}
-#endif
-// ======End Convolution TEST======
-
-// ======Start DepthwiseConvolution TEST======
-
-// TODO(zhaolong) The depthwise convolution cpu test will be added when the cpu
-// version of depthwiseConv is implemented.
-
-#ifndef PADDLE_ONLY_CPU
-
-TEST(DepthwiseConvForward, GEMM2) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConv-CPU", "DepthwiseConv-GPU", kForwardTest);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConv-CPU", "DepthwiseConv-GPU", kForwardTest);
-}
-
-TEST(DepthwiseConvBackwardInput, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConvGradInput-CPU",
-      "DepthwiseConvGradInput-GPU",
-      kBackwardInputTest);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConvGradInput-CPU",
-      "DepthwiseConvGradInput-GPU",
-      kBackwardInputTest);
-}
-
-TEST(DepthwiseConvBackwardFilter, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConvGradFilter-CPU",
-      "DepthwiseConvGradFilter-GPU",
-      kBackwardFilterTest);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConvGradFilter-CPU",
-      "DepthwiseConvGradFilter-GPU",
-      kBackwardFilterTest);
-}
-
-#endif
-// ======End DepthwiseConvolution TEST======
-
-}  // namespace paddle

paddle/function/ConvOpTest.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. */
+
+#include "FunctionTest.h"
+
+namespace paddle {
+
+template <DeviceType DType1, DeviceType DType2>
+void forward(Compare2Function<DType1, DType2>& test,
+             const TensorShape& input,
+             const TensorShape& filter,
+             const TensorShape& output) {
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output));
+  test.run();
+}
+
+template <DeviceType DType1, DeviceType DType2>
+void backward_input(Compare2Function<DType1, DType2>& test,
+                    const TensorShape& input,
+                    const TensorShape& filter,
+                    const TensorShape& output) {
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO);
+  test.run();
+}
+
+template <DeviceType DType1, DeviceType DType2>
+void backward_filter(Compare2Function<DType1, DType2>& test,
+                     const TensorShape& input,
+                     const TensorShape& filter,
+                     const TensorShape& output) {
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter), ADD_TO);
+  test.run();
+}
+
+template <DeviceType DType1, DeviceType DType2>
+using Function = void (*)(Compare2Function<DType1, DType2>& test,
+                          const TensorShape& input,
+                          const TensorShape& filter,
+                          const TensorShape& output);
+
+/**
+ * \brief A basic convolution function test interface.
+ *
+ * \param conv1         type name of convolution function 1.
+ * \param conv2         type name of convolution function 2.
+ * \param function      test function, can be one of the forward, backward_input
+ *                      backward_filter function.
+ * Example:
+ * 1. Compare GemmConv's CPU and GPU implementation:
+ *   Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+ *      "GemmConv-CPU", "GemmConv-GPU", forward);
+ */
+template <DeviceType DType1, DeviceType DType2>
+void Convolution(const std::string& conv1,
+                 const std::string& conv2,
+                 Function<DType1, DType2> function) {
+  for (size_t batchSize : {1, 5}) {
+    for (size_t inputSize : {7, 14, 31}) {
+      for (size_t filterSize : {1, 3, 5}) {
+        for (size_t inputChannels : {3, 16}) {
+          for (size_t outputChannels : {3, 16}) {
+            if (outputChannels < inputChannels) continue;
+            for (size_t stride : {1, 2}) {
+              for (size_t padding : {0, 1}) {
+                if (padding >= filterSize) break;
+
+                // NNPACK only supports stride = 1 if batchSize > 1
+                if ((conv1 == "NNPACKConv-CPU" || conv2 == "NNPACKConv-CPU") &&
+                    batchSize > 1 && stride > 1)
+                  break;
+
+                size_t outputSize =
+                    (inputSize - filterSize + 2 * padding + stride) / stride;
+                VLOG(3) << " batchSize=" << batchSize
+                        << " inputChannels=" << inputChannels
+                        << " inputHeight=" << inputSize
+                        << " inputWidth=" << inputSize
+                        << " outputChannels=" << outputChannels
+                        << " filterHeight=" << filterSize
+                        << " filterWidth=" << filterSize
+                        << " outputHeight=" << outputSize
+                        << " outputWidth=" << outputSize << " stride=" << stride
+                        << " padding=" << padding;
+
+                std::vector<size_t> paddings = {padding, padding};
+                std::vector<size_t> strides = {stride, stride};
+                Compare2Function<DType1, DType2> test(
+                    conv1,
+                    conv2,
+                    FuncConfig()
+                        .set("paddings", paddings)
+                        .set("strides", strides)
+                        .set("groups", (size_t)1)
+                        .set("algo", (std::string) "auto"));
+
+                TensorShape input{
+                    batchSize, inputChannels, inputSize, inputSize};
+                TensorShape filter{
+                    outputChannels, inputChannels, filterSize, filterSize};
+                TensorShape output{
+                    batchSize, outputChannels, outputSize, outputSize};
+
+                function(test, input, filter, output);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief A convolution function test interface for
+ *        image height is not equal image width.
+ */
+template <DeviceType DType1, DeviceType DType2>
+void Convolution2(const std::string& conv1,
+                  const std::string& conv2,
+                  Function<DType1, DType2> function) {
+  for (size_t batchSize : {4}) {
+    for (size_t inputHeight : {7, 31}) {
+      for (size_t inputWidth : {10, 54}) {
+        for (size_t filterHeight : {1, 5}) {
+          for (size_t filterWidth : {3, 7}) {
+            for (size_t inputChannels : {7}) {
+              for (size_t outputChannels : {7}) {
+                size_t stride = 1;
+                size_t padding = 0;
+                size_t outputHeight =
+                    (inputHeight - filterHeight + 2 * padding + stride) /
+                    stride;
+                size_t outputWidth =
+                    (inputWidth - filterWidth + 2 * padding + stride) / stride;
+                VLOG(3) << " batchSize=" << batchSize
+                        << " inputChannels=" << inputChannels
+                        << " inputHeight=" << inputHeight
+                        << " inputWidth=" << inputWidth
+                        << " outputChannels=" << outputChannels
+                        << " filterHeight=" << filterHeight
+                        << " filterWidth=" << filterWidth
+                        << " outputHeight=" << outputHeight
+                        << " outputWidth=" << outputWidth
+                        << " stride=" << stride << " padding=" << padding;
+
+                std::vector<size_t> paddings = {padding, padding};
+                std::vector<size_t> strides = {stride, stride};
+                Compare2Function<DType1, DType2> test(
+                    conv1,
+                    conv2,
+                    FuncConfig()
+                        .set("paddings", paddings)
+                        .set("strides", strides)
+                        .set("groups", (size_t)1)
+                        .set("algo", (std::string) "auto"));
+
+                TensorShape input{
+                    batchSize, inputChannels, inputHeight, inputWidth};
+                TensorShape filter{
+                    outputChannels, inputChannels, filterHeight, filterWidth};
+                TensorShape output{
+                    batchSize, outputChannels, outputHeight, outputWidth};
+
+                function(test, input, filter, output);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief A convolution function test interface for depthwise convolution.
+ */
+template <DeviceType DType1, DeviceType DType2>
+void DepthwiseConvolution(const std::string& conv1,
+                          const std::string& conv2,
+                          Function<DType1, DType2> function) {
+  for (size_t batchSize : {1, 32}) {
+    for (size_t inputSize : {7, 14, 54}) {
+      for (size_t filterSize : {3, 4}) {
+        for (size_t inputChannels : {32}) {
+          for (size_t outputChannels : {32, 64}) {
+            for (size_t stride : {1, 2}) {
+              for (size_t padding : {0, 1}) {
+                // NNPACK only supports stride = 1 if batchSize > 1,
+                // and there has some bug when batchSize > 1 and groups != 1
+                if ((conv1 == "NNPACKConv-CPU" || conv2 == "NNPACKConv-CPU") &&
+                    batchSize > 1)
+                  break;
+
+                size_t outputSize =
+                    (inputSize - filterSize + 2 * padding + stride) / stride;
+                VLOG(3) << " batchSize=" << batchSize
+                        << " inputChannels=" << inputChannels
+                        << " inputHeight=" << inputSize
+                        << " inputWidth=" << inputSize
+                        << " outputChannels=" << outputChannels
+                        << " filterHeight=" << filterSize
+                        << " filterWidth=" << filterSize
+                        << " outputHeight=" << outputSize
+                        << " outputWidth=" << outputSize << " stride=" << stride
+                        << " padding=" << padding;
+
+                std::vector<size_t> paddings = {padding, padding};
+                std::vector<size_t> strides = {stride, stride};
+                size_t groups = inputChannels;
+                Compare2Function<DType1, DType2> test(
+                    conv1,
+                    conv2,
+                    FuncConfig()
+                        .set("paddings", paddings)
+                        .set("strides", strides)
+                        .set("groups", groups)
+                        .set("algo", (std::string) "auto"));
+
+                TensorShape input{
+                    batchSize, inputChannels, inputSize, inputSize};
+                TensorShape filter{groups,
+                                   outputChannels / groups,
+                                   inputChannels / groups,
+                                   filterSize,
+                                   filterSize};
+                TensorShape output{
+                    batchSize, outputChannels, outputSize, outputSize};
+
+                function(test, input, filter, output);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/function/DepthwiseConvOp.cpp

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

paddle/function/DepthwiseConvOpGpu.cu

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

paddle/framework/op_desc_test.cc → paddle/function/DepthwiseConvOpTest.cpp

@@ -13,23 +13,25 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <gtest/gtest.h>
-#include <paddle/framework/op_desc.pb.h>
-
-TEST(OpDesc, Create) {
-  paddle::framework::OpDesc op_desc;
-  op_desc.set_type("add");
-  op_desc.add_inputs("X");
-  op_desc.add_inputs("Y");
-  op_desc.add_outputs("Z");
-
-  auto attr = op_desc.mutable_attrs()->Add();
-  attr->set_type(paddle::framework::AttrType::FLOAT);
-  attr->set_f(3.14);
-
-  // required field name is not set, so IsInitialized should be false.
-  ASSERT_FALSE(op_desc.IsInitialized());
-
-  attr->set_name("add");
-  // after all required fields are set, IsInitialized should be true now.
-  ASSERT_TRUE(op_desc.IsInitialized());
-}
\ No newline at end of file
+#include "ConvOpTest.h"
+
+namespace paddle {
+
+#ifndef PADDLE_ONLY_CPU
+TEST(DepthwiseConv, Forward) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConv-CPU", "DepthwiseConv-GPU", forward);
+}
+
+TEST(DepthwiseConv, BackwardInput) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradInput-CPU", "DepthwiseConvGradInput-GPU", backward_input);
+}
+
+TEST(DepthwiseConv, BackwardFilter) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradFilter-CPU", "DepthwiseConvGradFilter-GPU", backward_filter);
+}
+#endif
+
+}  // namespace paddle

paddle/function/EigenGemm.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <glog/logging.h>
+#include "unsupported/Eigen/CXX11/Tensor"
+
+namespace paddle {
+
+template <class T>
+struct EigenBlasGemm {
+  typedef Eigen::TensorMap<Eigen::Tensor<T, 2, Eigen::RowMajor, int>,
+                           Eigen::Aligned>
+      Matrix;
+
+  static void compute(const bool transA,
+                      const bool transB,
+                      const int M,
+                      const int N,
+                      const int K,
+                      const T alpha,
+                      const T* A,
+                      const int lda,
+                      const T* B,
+                      const int ldb,
+                      const T beta,
+                      T* C,
+                      const int ldc) {
+    Eigen::array<int, 2> sizeA;
+    if (transA) {
+      sizeA[0] = K;
+      sizeA[1] = M;
+      CHECK_EQ(M, lda);
+    } else {
+      sizeA[0] = M;
+      sizeA[1] = K;
+      CHECK_EQ(K, lda);
+    }
+    Eigen::array<int, 2> sizeB;
+    if (transB) {
+      sizeB[0] = N;
+      sizeB[1] = K;
+      CHECK_EQ(K, ldb);
+    } else {
+      sizeB[0] = K;
+      sizeB[1] = N;
+      CHECK_EQ(N, ldb);
+    }
+    Eigen::array<int, 2> sizeC;
+    sizeC[0] = M;
+    sizeC[1] = N;
+    CHECK_EQ(N, ldc);
+
+    const Matrix a(const_cast<T*>(A), sizeA);
+    const Matrix b(const_cast<T*>(B), sizeB);
+    Matrix c(C, sizeC);
+
+    typedef typename Eigen::Tensor<T, 2>::DimensionPair DimPair;
+    Eigen::array<DimPair, 1> dims;
+    dims[0] = DimPair(1, 0);
+    dims[0].first = transA ? 0 : 1;
+    dims[0].second = transB ? 1 : 0;
+
+    Eigen::DefaultDevice device;
+    if (alpha == T(1) && beta == T(0)) {
+      c.device(device) = a.contract(b, dims);
+    } else if (alpha == T(1) && beta == T(1)) {
+      c.device(device) += a.contract(b, dims);
+    } else {
+      c.device(device) = alpha * a.contract(b, dims) + beta * c;
+    }
+  }
+};
+
+#ifdef PADDLE_TYPE_DOUBLE
+template class EigenBlasGemm<double>;
+#else
+template class EigenBlasGemm<float>;
+#endif
+
+}  // namespace paddle

paddle/function/GemmConvOp.cpp

@@ -85,7 +85,6 @@ public:
     }
 
     Im2ColFunctor<kCFO, Device, real> im2col;
-    GemmFunctor<Device, real> gemm;
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -108,19 +107,19 @@ public:
         int M = outputChannels / groups_;
         int N = outputHeight * outputWidth;
         int K = inputChannels / groups_ * filterHeight * filterWidth;
-        gemm(CblasNoTrans,
-             CblasNoTrans,
-             M,
-             N,
-             K,
-             1.0f,
-             filterData + g * filterOffset,
-             K,
-             colData,
-             N,
-             beta,
-             outputData + g * outputOffset,
-             N);
+        BlasGemm<Device, real>::compute(false,
+                                        false,
+                                        M,
+                                        N,
+                                        K,
+                                        1.0f,
+                                        filterData + g * filterOffset,
+                                        K,
+                                        colData,
+                                        N,
+                                        beta,
+                                        outputData + g * outputOffset,
+                                        N);
       }
       inputData += inputChannels * inputHeight * inputWidth;
       outputData += outputChannels * outputHeight * outputWidth;
@@ -188,8 +187,6 @@ public:
     }
 
     Col2ImFunctor<kCFO, Device, real> col2im;
-    GemmFunctor<Device, real> gemm;
-
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -205,19 +202,19 @@ public:
           colData = inputGrad + g * inputOffset;
           scale = 1.0f;
         }
-        gemm(CblasTrans,
-             CblasNoTrans,
-             M,
-             N,
-             K,
-             1.0f,
-             filterData + g * filterOffset,
-             M,
-             outputGrad + g * outputOffset,
-             N,
-             scale,
-             colData,
-             N);
+        BlasGemm<Device, real>::compute(true,
+                                        false,
+                                        M,
+                                        N,
+                                        K,
+                                        1.0f,
+                                        filterData + g * filterOffset,
+                                        M,
+                                        outputGrad + g * outputOffset,
+                                        N,
+                                        scale,
+                                        colData,
+                                        N);
         if (needIm2col) {
           col2im(inputGrad + g * inputOffset,
                  imShape,
@@ -299,7 +296,6 @@ public:
     }
 
     Im2ColFunctor<kCFO, Device, real> im2col;
-    GemmFunctor<Device, real> gemm;
     size_t inputOffset = imShape.getElements();
     size_t outputOffset =
         (outputChannels / groups_) * outputHeight * outputWidth;
@@ -321,19 +317,19 @@ public:
         int M = outputChannels / groups_;
         int K = outputHeight * outputWidth;
         int N = inputChannels / groups_ * filterHeight * filterWidth;
-        gemm(CblasNoTrans,
-             CblasTrans,
-             M,
-             N,
-             K,
-             1.0f,
-             outputGrad + g * outputOffset,
-             K,
-             colData,
-             K,
-             i == 0 ? beta : 1.0f,
-             filterGrad + g * filterOffset,
-             N);
+        BlasGemm<Device, real>::compute(false,
+                                        true,
+                                        M,
+                                        N,
+                                        K,
+                                        1.0f,
+                                        outputGrad + g * outputOffset,
+                                        K,
+                                        colData,
+                                        K,
+                                        i == 0 ? beta : 1.0f,
+                                        filterGrad + g * filterOffset,
+                                        N);
       }
       inputData += inputChannels * inputHeight * inputWidth;
       outputGrad += outputChannels * outputHeight * outputWidth;

paddle/framework/attribute.proto → paddle/function/GemmConvOpTest.cpp

@@ -12,17 +12,39 @@ 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. */
 
-syntax = "proto2";
-package paddle.framework;
-
-// Attribute Type for paddle's Op.
-// Op contains many attributes. Each type of attributes could be different.
-// The AttrType will be shared between AttrDesc and AttrProto.
-enum AttrType {
-  INT = 0;
-  FLOAT = 1;
-  STRING = 2;
-  INTS = 3;
-  FLOATS = 4;
-  STRINGS = 5;
-}
\ No newline at end of file
+#include <gtest/gtest.h>
+#include "ConvOpTest.h"
+
+namespace paddle {
+
+TEST(GemmConv, NaiveConv) {
+  Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
+      "NaiveConv-CPU", "GemmConv-CPU", forward);
+  Convolution2<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
+      "NaiveConv-CPU", "GemmConv-CPU", forward);
+}
+
+#ifndef PADDLE_ONLY_CPU
+TEST(GemmConv, Forward) {
+  Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConv-CPU", "GemmConv-GPU", forward);
+  Convolution2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConv-CPU", "GemmConv-GPU", forward);
+}
+
+TEST(GemmConv, BackwardInput) {
+  Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradInput-CPU", "GemmConvGradInput-GPU", backward_input);
+  Convolution2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradInput-CPU", "GemmConvGradInput-GPU", backward_input);
+}
+
+TEST(GemmConv, BackwardFilter) {
+  Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradFilter-CPU", "GemmConvGradFilter-GPU", backward_filter);
+  Convolution2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradFilter-CPU", "GemmConvGradFilter-GPU", backward_filter);
+}
+#endif
+
+}  // namespace paddle

paddle/function/nnpack/NNPACKConvOp.cpp

@@ -196,30 +196,30 @@ public:
         CHECK_EQ(status, nnp_status_success);
       }
     } else {
-      for (size_t g = 0; g < groups_; g++) {
-        // only supports stride = 1
-        CHECK_EQ(strideH(), 1);
-        CHECK_EQ(strideW(), 1);
-        nnp_status status =
-            nnp_convolution_output(algorithm_,
-                                   batchSize,
-                                   inputChannels / groups_,
-                                   outputChannels / groups_,
-                                   inputSize,
-                                   padding,
-                                   kernelSize,
-                                   inputData + inputOffset * g,
-                                   filterData + filterOffset * g,
-                                   nullptr, /* bias */
-                                   outputData + outputOffset * g,
-                                   bufferPtr,
-                                   sizePtr,
-                                   nnp_activation_identity,
-                                   nullptr,
-                                   threadpool_, /* threadpool */
-                                   nullptr);
-        CHECK_EQ(status, nnp_status_success);
-      }
+      // only supports stride = 1
+      CHECK_EQ(strideH(), 1);
+      CHECK_EQ(strideW(), 1);
+
+      // TODO(hedaoyuan): There has some bug when batchSize > 1 and groups_ > 1.
+      CHECK_EQ(groups_, static_cast<size_t>(1));
+      nnp_status status = nnp_convolution_output(algorithm_,
+                                                 batchSize,
+                                                 inputChannels,
+                                                 outputChannels,
+                                                 inputSize,
+                                                 padding,
+                                                 kernelSize,
+                                                 inputData,
+                                                 filterData,
+                                                 nullptr, /* bias */
+                                                 outputData,
+                                                 bufferPtr,
+                                                 sizePtr,
+                                                 nnp_activation_identity,
+                                                 nullptr,
+                                                 threadpool_, /* threadpool */
+                                                 nullptr);
+      CHECK_EQ(status, nnp_status_success);
     }
   }
 

paddle/gserver/gradientmachines/NeuralNetwork.cpp

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