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

.gitignore

@@ -24,4 +24,5 @@ cmake-build-*
 python/paddle/v2/framework/core.so
 CMakeFiles
 cmake_install.cmake
-
+paddle/.timestamp
+python/paddlepaddle.egg-info/

.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

@@ -14,8 +14,8 @@
 
 cmake_minimum_required(VERSION 3.0)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
-set(PROJ_ROOT ${CMAKE_CURRENT_SOURCE_DIR})
-set(PROJ_BINARY_ROOT ${CMAKE_CURRENT_BINARY_DIR})
+set(PADDLE_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
+set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
 
 include(system)
 
@@ -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."    ${AVX_FOUND})
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
+option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    OFF)
+option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      OFF)
 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)
@@ -121,8 +121,8 @@ include(version)            # set PADDLE_VERSION
 include(coveralls)          # set code coverage
 
 
-include_directories("${PROJ_ROOT}")
-include_directories("${PROJ_ROOT}/paddle/cuda/include")
+include_directories("${PADDLE_SOURCE_DIR}")
+include_directories("${PADDLE_SOURCE_DIR}/paddle/cuda/include")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/client/c")
 include_directories(${Boost_INCLUDE_DIRS})
@@ -144,7 +144,7 @@ if(WITH_GPU)
 endif(WITH_GPU)
 
 if(WITH_MKLDNN)
-    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIBRARY} ${MKLDNN_IOMP_LIB})
+    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
 endif()
 
 if(USE_NNPACK)
@@ -164,10 +164,12 @@ if(WITH_GOLANG)
     add_subdirectory(go)
 endif(WITH_GOLANG)
 
+set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
 add_subdirectory(paddle)
 if(WITH_PYTHON)
   add_subdirectory(python)
 endif()
+
 if(WITH_DOC)
     add_subdirectory(doc)
 endif()

Dockerfile

@@ -64,13 +64,28 @@ RUN pip install --upgrade pip && \
     pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
     pip install pre-commit 'requests==2.9.2' 'ipython==5.3.0' && \
     pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
-    pip install rarfile
+    pip install opencv-python rarfile 'scipy>=0.19.0' 'nltk>=3.2.2'
 
 # 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 && \
     (cd /woboq \

cmake/configure.cmake

@@ -129,7 +129,7 @@ if(WITH_GOLANG)
     add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide
       COMMAND env GOPATH=${GOPATH} ${GLIDE} install
       COMMAND touch ${CMAKE_BINARY_DIR}/glide
-      DEPENDS ${PROJ_ROOT}/go/glide.lock
+      DEPENDS ${PADDLE_SOURCE_DIR}/go/glide.lock
       WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go"
       )
 

cmake/cpplint.cmake

@@ -52,7 +52,7 @@ macro(add_style_check_target TARGET_NAME)
 
         if(SOURCES_LIST)
             add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
-                COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
+                COMMAND "${PYTHON_EXECUTABLE}" "${PADDLE_SOURCE_DIR}/paddle/scripts/cpplint.py"
                         "--filter=${STYLE_FILTER}"
                         ${SOURCES_LIST}
                 COMMENT "cpplint: Checking source code style"

cmake/flags.cmake

@@ -9,10 +9,12 @@ function(CheckCompilerCXX11Flag)
         if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
             message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
         endif()
-        # 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)
+        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"

cmake/generic.cmake

@@ -411,7 +411,7 @@ function(py_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS)
     cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})  
     add_test(NAME ${TARGET_NAME}
-             COMMAND env PYTHONPATH=${PADDLE_PYTHON_PACKAGE_DIR}
+             COMMAND env PYTHONPATH=${PADDLE_PYTHON_BUILD_DIR}/lib-python
              python2 ${py_test_SRCS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
   endif()

cmake/package.cmake

@@ -12,7 +12,7 @@ set(CPACK_PACKAGE_DESCRIPTION "")
 set(CPACK_DEBIAN_PACKAGE_DEPENDS "libpython2.7-dev, libstdc++6, python-pip, curl, libgfortran3, python-pip-whl")
 set(CPACK_DEBIAN_PACKAGE_SECTION Devel)
 set(CPACK_DEBIAN_PACKAGE_VERSION ${PADDLE_VERSION})
-set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PROJ_ROOT}/paddle/scripts/deb/postinst")
+set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PADDLE_SOURCE_DIR}/paddle/scripts/deb/postinst")
 #set(CPACK_GENERATOR "DEB")
 # Start cpack
 include (CMakePackageConfigHelpers)

cmake/util.cmake

@@ -141,8 +141,8 @@ endmacro()
 function(create_resources res_file output_file)
   add_custom_command(
     OUTPUT ${output_file}
-    COMMAND python ARGS ${PROJ_ROOT}/cmake/make_resource.py ${res_file} ${output_file}
-    DEPENDS ${res_file} ${PROJ_ROOT}/cmake/make_resource.py)
+    COMMAND python ARGS ${PADDLE_SOURCE_DIR}/cmake/make_resource.py ${res_file} ${output_file}
+    DEPENDS ${res_file} ${PADDLE_SOURCE_DIR}/cmake/make_resource.py)
 endfunction()
 
 

cmake/version.cmake

@@ -4,7 +4,7 @@ set(tmp_version "HEAD")
 while ("${PADDLE_VERSION}" STREQUAL "")
   execute_process(
     COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 ${tmp_version}
-    WORKING_DIRECTORY ${PROJ_ROOT}
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
     OUTPUT_VARIABLE GIT_TAG_NAME
     RESULT_VARIABLE GIT_RESULT
     ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)

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/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：
 

doc/templates/conf.py.cn.in

@@ -13,7 +13,7 @@
 # serve to show the default.
 import sys
 import os, subprocess
-sys.path.insert(0, os.path.abspath('@PROJ_ROOT@/python'))
+sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python'))
 import shlex
 from recommonmark import parser, transform
 import paddle
@@ -24,7 +24,7 @@ AutoStructify = transform.AutoStructify
 # If extensions (or modules to document with autodoc) are in another directory,
 # add these directories to sys.path here. If the directory is relative to the
 # documentation root, use os.path.abspath to make it absolute, like shown here.
-templates_path = ["@PROJ_ROOT@/doc_theme/templates"]
+templates_path = ["@PADDLE_SOURCE_DIR@/doc_theme/templates"]
 
 # -- General configuration ------------------------------------------------
 
@@ -120,7 +120,7 @@ html_theme = 'sphinx_rtd_theme'
 # Add any paths that contain custom static files (such as style sheets) here,
 # relative to this directory. They are copied after the builtin static files,
 # so a file named "default.css" will overwrite the builtin "default.css".
-html_static_path = ['@PROJ_ROOT@/doc_theme/static']
+html_static_path = ['@PADDLE_SOURCE_DIR@/doc_theme/static']
 
 # Output file base name for HTML help builder.
 htmlhelp_basename = project + 'doc'

doc/templates/conf.py.en.in

@@ -13,7 +13,7 @@
 # serve to show the default.
 import sys
 import os, subprocess
-sys.path.insert(0, os.path.abspath('@PROJ_ROOT@/python'))
+sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python'))
 import shlex
 from recommonmark import parser, transform
 import paddle
@@ -25,7 +25,7 @@ AutoStructify = transform.AutoStructify
 # If extensions (or modules to document with autodoc) are in another directory,
 # add these directories to sys.path here. If the directory is relative to the
 # documentation root, use os.path.abspath to make it absolute, like shown here.
-templates_path = ["@PROJ_ROOT@/doc_theme/templates"]
+templates_path = ["@PADDLE_SOURCE_DIR@/doc_theme/templates"]
 
 # -- General configuration ------------------------------------------------
 
@@ -120,7 +120,7 @@ html_theme = 'sphinx_rtd_theme'
 # Add any paths that contain custom static files (such as style sheets) here,
 # relative to this directory. They are copied after the builtin static files,
 # so a file named "default.css" will overwrite the builtin "default.css".
-html_static_path = ['@PROJ_ROOT@/doc_theme/static']
+html_static_path = ['@PADDLE_SOURCE_DIR@/doc_theme/static']
 
 # Output file base name for HTML help builder.
 htmlhelp_basename = project + 'doc'

go/pserver/client/c/test/test_train.py

@@ -17,12 +17,10 @@ def main():
     # network config
     x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
     y_predict = paddle.layer.fc(input=x,
-                                param_attr=paddle.attr.Param(
-                                    name='w', learning_rate=1e-3),
+                                param_attr=paddle.attr.Param(name='w'),
                                 size=1,
                                 act=paddle.activation.Linear(),
-                                bias_attr=paddle.attr.Param(
-                                    name='b', learning_rate=1e-3))
+                                bias_attr=paddle.attr.Param(name='b'))
     y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1))
     cost = paddle.layer.mse_cost(input=y_predict, label=y)
 

paddle/api/CMakeLists.txt

@@ -19,9 +19,9 @@ add_library(paddle_api STATIC ${API_SOURCES})
 add_dependencies(paddle_api paddle_proto paddle_trainer_lib)
 
 INCLUDE(${SWIG_USE_FILE})
-INCLUDE_DIRECTORIES(${PROJ_ROOT}/paddle)
+INCLUDE_DIRECTORIES(${PADDLE_SOURCE_DIR}/paddle)
 
-FILE(GLOB PY_PADDLE_PYTHON_FILES ${PROJ_ROOT}/paddle/py_paddle/*.py)
+FILE(GLOB PY_PADDLE_PYTHON_FILES ${PADDLE_SOURCE_DIR}/paddle/py_paddle/*.py)
 
 SET_SOURCE_FILES_PROPERTIES(Paddle.i PROPERTIES CPLUSPLUS ON)
 
@@ -79,16 +79,16 @@ SWIG_LINK_LIBRARIES(swig_paddle
     ${START_END}
 )
 
-add_custom_command(OUTPUT ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so
-    COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/swig_paddle.py ${PROJ_ROOT}/paddle/py_paddle
-    COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/_swig_paddle.so ${PROJ_ROOT}/paddle/py_paddle
+add_custom_command(OUTPUT ${PADDLE_SOURCE_DIR}/paddle/py_paddle/_swig_paddle.so
+    COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/swig_paddle.py ${PADDLE_SOURCE_DIR}/paddle/py_paddle
+    COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/_swig_paddle.so ${PADDLE_SOURCE_DIR}/paddle/py_paddle
     COMMAND ${CMAKE_COMMAND} -E touch .timestamp
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle
     DEPENDS _swig_paddle
 )
 
 # TODO(yuyang18) : make wheel name calculated by cmake
-add_custom_target(python_api_wheel ALL DEPENDS ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so)
+add_custom_target(python_api_wheel ALL DEPENDS ${PADDLE_SOURCE_DIR}/paddle/py_paddle/_swig_paddle.so)
 
 if(WITH_TESTING)
     IF(NOT PY_PIP_FOUND)

paddle/api/ParameterUpdater.cpp

@@ -41,7 +41,7 @@ ParameterUpdater *ParameterUpdater::createNewRemoteUpdater(
       config->m->getConfig(), pserverSpec, useEtcd));
   return updater;
 #else
-  throw UnsupportError();
+  throw UnsupportError("not compiled with WITH_GOLANG");
 #endif
 }
 

paddle/capi/Arguments.cpp

@@ -90,6 +90,18 @@ paddle_error paddle_arguments_set_ids(paddle_arguments args,
   return kPD_NO_ERROR;
 }
 
+paddle_error paddle_arguments_set_frame_shape(paddle_arguments args,
+                                              uint64_t ID,
+                                              uint64_t frameHeight,
+                                              uint64_t frameWidth) {
+  if (args == nullptr) return kPD_NULLPTR;
+  auto a = castArg(args);
+  if (ID >= a->args.size()) return kPD_OUT_OF_RANGE;
+  a->args[ID].setFrameHeight(frameHeight);
+  a->args[ID].setFrameWidth(frameWidth);
+  return kPD_NO_ERROR;
+}
+
 paddle_error paddle_arguments_set_sequence_start_pos(paddle_arguments args,
                                                      uint64_t ID,
                                                      uint32_t nestedLevel,

paddle/capi/arguments.h

@@ -111,6 +111,20 @@ PD_API paddle_error paddle_arguments_set_ids(paddle_arguments args,
                                              uint64_t ID,
                                              paddle_ivector ids);
 
+/**
+ * @brief paddle_arguments_set_frame_shape Set the fram size of one argument
+ *        in array, which index is `ID`.
+ * @param [in] args arguments array
+ * @param [in] ID array index
+ * @param [in] frameHeight maximum height of input images
+ * @param [in] frameWidth maximum width of input images
+ * @return paddle_error
+ */
+PD_API paddle_error paddle_arguments_set_frame_shape(paddle_arguments args,
+                                                     uint64_t ID,
+                                                     uint64_t frameHeight,
+                                                     uint64_t frameWidth);
+
 /**
  * @brief PDArgsSetSequenceStartPos Set sequence start position vector of one
  *        argument in array, which index is `ID`.

paddle/capi/examples/model_inference/common/common.h

@@ -3,18 +3,21 @@
 #include <stdio.h>
 #include <stdlib.h>
 
-#define CHECK(stmt)                                                \
-  do {                                                             \
-    paddle_error __err__ = stmt;                                   \
-    if (__err__ != kPD_NO_ERROR) {                                 \
-      fprintf(stderr, "Invoke paddle error %d \n" #stmt, __err__); \
-      exit(__err__);                                               \
-    }                                                              \
+#define CHECK(stmt)                                                      \
+  do {                                                                   \
+    paddle_error __err__ = stmt;                                         \
+    if (__err__ != kPD_NO_ERROR) {                                       \
+      fprintf(stderr, "Invoke paddle error %d in " #stmt "\n", __err__); \
+      exit(__err__);                                                     \
+    }                                                                    \
   } while (0)
 
 void* read_config(const char* filename, long* size) {
   FILE* file = fopen(filename, "r");
-  if (file == NULL) return NULL;
+  if (file == NULL) {
+    fprintf(stderr, "Open %s error\n", filename);
+    return NULL;
+  }
   fseek(file, 0L, SEEK_END);
   *size = ftell(file);
   fseek(file, 0L, SEEK_SET);

paddle/capi/gradient_machine.cpp

@@ -54,6 +54,31 @@ paddle_error paddle_gradient_machine_create_for_inference(
   return kPD_NO_ERROR;
 }
 
+paddle_error paddle_gradient_machine_create_for_inference_with_parameters(
+    paddle_gradient_machine* machine, void* mergedModel, uint64_t size) {
+  if (mergedModel == nullptr) return kPD_NULLPTR;
+  std::istringstream is(std::string(static_cast<char*>(mergedModel), size));
+  int64_t modelConfigSize = 0;
+  is.read((char*)(&modelConfigSize), sizeof(modelConfigSize));
+  std::string modelConfigProtobuf;
+  modelConfigProtobuf.resize(modelConfigSize);
+  is.read(&modelConfigProtobuf[0], modelConfigSize);
+  paddle::TrainerConfig config;
+  if (!config.ParseFromString(modelConfigProtobuf) || !config.IsInitialized()) {
+    return kPD_PROTOBUF_ERROR;
+  }
+  auto ptr = new paddle::capi::CGradientMachine();
+  ptr->machine.reset(paddle::GradientMachine::create(
+      config.model_config(), CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE}));
+  std::vector<paddle::ParameterPtr>& parameters = ptr->machine->getParameters();
+  for (auto& para : parameters) {
+    para->load(is);
+  }
+
+  *machine = ptr;
+  return kPD_NO_ERROR;
+}
+
 paddle_error paddle_gradient_machine_destroy(paddle_gradient_machine machine) {
   delete cast(machine);
   return kPD_NO_ERROR;

paddle/capi/gradient_machine.h

@@ -36,6 +36,18 @@ typedef void* paddle_gradient_machine;
 PD_API paddle_error paddle_gradient_machine_create_for_inference(
     paddle_gradient_machine* machine, void* modelConfigProtobuf, int size);
 
+/**
+ * @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.
+ * @param [in] mergedModel
+ * @param [in] size
+ * @return paddle_error
+ */
+PD_API paddle_error
+paddle_gradient_machine_create_for_inference_with_parameters(
+    paddle_gradient_machine* machine, void* mergedModel, uint64_t size);
+
 /**
  * @brief Load parameter from disk.
  * @param machine Gradient Machine.

paddle/capi/tests/CMakeLists.txt

@@ -10,5 +10,5 @@ target_include_directories(capi_test_gradientMachine PUBLIC
   ${PADDLE_CAPI_INC_PATH})
 target_link_libraries(capi_test_gradientMachine paddle_capi)
 add_test(NAME capi_test_gradientMachine
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
-  WORKING_DIRECTORY ${PROJ_ROOT}/paddle/capi/tests)
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
+  WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests)

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)
@@ -15,26 +15,27 @@ 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 ${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)
@@ -43,12 +44,16 @@ 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
     uniform_random_op
+    gaussian_random_op
     fill_zeros_like_op)
 endif(WITH_PYTHON)

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

@@ -14,16 +14,15 @@ limitations under the License. */
 
 #pragma once
 
-#include <boost/variant.hpp>
 #include <functional>
 #include <string>
 #include <unordered_map>
 #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"
 
 namespace paddle {
 namespace framework {
@@ -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

@@ -21,15 +21,24 @@
 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(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 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::shared_ptr<OperatorBase> NOP() {
@@ -68,10 +77,11 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
   //  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);
-    }
+    ForEachVarName(forwardOp.inputs_,
+                   [&no_grad_names](const std::string& name) -> bool {
+                     no_grad_names.insert(GradVarName(name));
+                     return false;
+                   });
     return NOP();
   }
 
@@ -93,9 +103,11 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
       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;
+                     });
     }
     // Get unique ID for this method.
     auto uid = uniq_id++;
@@ -117,7 +129,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
       insert_position.push_back(
           {dup_op.back(),
            OpRegistry::CreateOp(
-               "add", {dup_outputs}, {name},
+               "add", {{"X", {dup_outputs}}}, {{"Out", {name}}},
                {{"input_format",
                  std::vector<int>{0, static_cast<int>(dup_outputs.size())}}})});
     }
@@ -131,24 +143,30 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
 
   } else {
     std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
-    for (std::string& grad_input : grad_op->inputs_) {
+
+    ForEachVarName(grad_op->inputs_, [&no_grad_names,
+                                      &net](std::string& grad_input) {
       if (no_grad_names.count(grad_input)) {
-        std::string prefix =
-            grad_input.substr(0, grad_input.size() - kGradVarSuffix.size());
+        // +1 for \0
+        std::string prefix = grad_input.substr(
+            0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
         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}, {}));
+        net->AddOp(OpRegistry::CreateOp("fill_zeros_like", {{"Src", {prefix}}},
+                                        {{"Dst", {grad_input}}}, {}));
       }
-    }
+      return false;
+    });
 
-    for (std::string& grad_output : grad_op->outputs_) {
-      if (no_grad_names.count(grad_output)) {
-        grad_output = kEmptyVarName;
-      }
-    }
+    ForEachVarName(grad_op->outputs_,
+                   [&no_grad_names](std::string& grad_output) {
+                     if (no_grad_names.count(grad_output)) {
+                       grad_output = kEmptyVarName;
+                     }
+                     return false;
+                   });
 
     if (net->ops_.empty()) {  // Current no aux op is added to network
       return grad_op;
@@ -167,7 +185,7 @@ std::shared_ptr<OperatorBase> Backward(
   std::unordered_set<std::string> no_grad_names;
   no_grad_names.reserve(no_grad_vars.size());
 
-  no_grad_names.insert(kEmptyVarName + kGradVarSuffix);
+  no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
 
   for (auto& name : no_grad_vars) {
     no_grad_names.insert(name + kGradVarSuffix);

paddle/framework/backward_test.cc

@@ -30,6 +30,7 @@ using DeviceContext = platform::DeviceContext;
 
 class EmptyOp : public OperatorBase {
  public:
+  using OperatorBase::OperatorBase;
   void InferShape(const Scope &scope) const override {}
   void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
 };
@@ -38,9 +39,9 @@ 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").AsNoGradient();
+    AddInput("b", "Bias of Add").AsNoGradient();
+    AddOutput("Out", "Out of Add").AsNoGradient();
     AddComment("Add Op");
   }
 };
@@ -49,8 +50,8 @@ 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 +62,7 @@ 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 +72,25 @@ 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) {
+    AddOp(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) {
+      AddOp(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 +99,8 @@ class FcOp : public operators::NetOp {
       }
     }
 
-    AddOp(OpRegistry::CreateOp("sigmoid", {Output(before_act)}, {Output("Out")},
-                               {}));
+    AddOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
+                               {{"Out", {Output("Out")}}}, {}));
     CompleteAddOp(false);
   }
 };
@@ -107,8 +112,8 @@ 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 +144,7 @@ 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("");
   }
@@ -165,27 +170,24 @@ REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker);
 REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp);
 
 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(1UL, gop->inputs_.size());
   ASSERT_EQ("rowwise_add_grad", gop->type_);
-  ASSERT_EQ("X" + f::kGradVarSuffix, gop->outputs_[0]);
-  ASSERT_EQ("b" + f::kGradVarSuffix, gop->outputs_[1]);
-
-  ASSERT_EQ("X" + f::kGradVarSuffix, gop->Output("X" + f::kGradVarSuffix));
+  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(),
-                      "X" + f::kGradVarSuffix),
-            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,
@@ -193,8 +195,12 @@ TEST(Backward, simple_op_not_need_grad) {
 }
 
 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());
@@ -216,8 +222,11 @@ TEST(Backward, net_fc_backward_normal) {
 
 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());
@@ -236,38 +245,49 @@ TEST(Backward, net_fc_backward_not_have_b) {
 
 TEST(Backward, net_input_of_network_not_need_grad) {
   ops::NetOp net;
-  net.AddOp(f::OpRegistry::CreateOp("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.AddOp(f::OpRegistry::CreateOp(
+      "fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}},
+      {{"mul_result", {"mul_tmp_0"}},
+       {"add_result", {"add_tmp_0"}},
+       {"Out", {"hidden0"}}},
+      {}));
+  net.AddOp(f::OpRegistry::CreateOp(
+      "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(out + f::kGradVarSuffix), all_output.end());
+    ASSERT_NE(all_outputs.find(f::GradVarName(out)), all_outputs.end());
   }
 
   // Not Generated X
-  ASSERT_EQ(all_output.find("X" + f::kGradVarSuffix), 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("A" + f::kGradVarSuffix));
+            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.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
+                                    {{"Out", {"out"}}}, {}));
+  net.AddOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
+                                    {{"Out", {"FinalOut"}}}, {}));
   net.CompleteAddOp();
 
   auto bwd = f::Backward(net, {});
@@ -278,31 +298,37 @@ TEST(Backward, net_shared_weight) {
 }
 
 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());
@@ -310,60 +336,77 @@ TEST(Backward, op_part_of_output_are_not_need) {
 
   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("Z" + f::kZeroVarSuffix, fill_zero.outputs_[0]);
+  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("Z" + f::kZeroVarSuffix, d_many_out.Input("z" + f::kGradVarSuffix));
-  ASSERT_EQ("Y" + f::kGradVarSuffix, d_many_out.Input("y" + f::kGradVarSuffix));
-  ASSERT_EQ("X" + f::kGradVarSuffix,
-            d_many_out.Output("x" + f::kGradVarSuffix));
+  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")));
+  ASSERT_EQ(f::GradVarName("X"), d_many_out.Output(f::GradVarName("x")));
 }
 
 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("A" + f::kGradVarSuffix), f::kEmptyVarName);
-  ASSERT_EQ(grad_mul.Output("B" + f::kGradVarSuffix), "b" + f::kGradVarSuffix);
-  ASSERT_EQ(grad_mul.Input("Out" + f::kGradVarSuffix),
-            "out" + f::kGradVarSuffix);
-  ASSERT_EQ(grad_mul.Input("A"), "a");
-  ASSERT_EQ(grad_mul.Input("B"), "b");
+  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("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.AddOp(f::OpRegistry::CreateOp(
+      "fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}},
+      {{"mul_result", {"mul_out1"}},
+       {"add_result", {"add_out1"}},
+       {"Out", {"out1"}}},
+      {}));
+  net.AddOp(f::OpRegistry::CreateOp(
+      "fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}},
+      {{"mul_result", {"mul_out2"}},
+       {"add_result", {"tmp_out2"}},
+       {"Out", {"out2"}}},
+      {}));
+  net.AddOp(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

@@ -14,13 +14,12 @@ limitations under the License. */
 
 #pragma once
 
-#include <boost/variant.hpp>
 #include <initializer_list>
 #include <stdexcept>
 #include <vector>
 #include "paddle/framework/dim.h"
 #include "paddle/platform/enforce.h"
-#include "unsupported/Eigen/CXX11/Tensor"
+#include "paddle/platform/variant.h"
 
 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/attribute.proto → paddle/framework/framework.proto

@@ -15,9 +15,6 @@ 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;
@@ -25,4 +22,61 @@ enum AttrType {
   INTS = 3;
   FLOATS = 4;
   STRINGS = 5;
-}
+}
\ No newline at end of file
+
+// 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 no_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 +13,52 @@ 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/framework.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 =
+static void TransOpArg(const OperatorBase* src_op,
+                       OperatorBase::VarNameMap* vars,
+                       const OpArgType& src_type, bool is_grad) {
+  const auto& src_inout =
       src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_;
-  const std::vector<int>* src_format = GetOpFormat(src_op, src_type);
+  auto& dst_inout = *vars;
 
-  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_);
+  const OpProto& proto = OpProtos().at(src_op->type_);
   const auto& src_arg_list =
       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.no_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 gop_type_it = OpRegistry::grad_ops().find(op->type_);
+  PADDLE_ENFORCE(gop_type_it != OpRegistry::grad_ops().end(),
+                 "Operator %s do not register gradient type", op->type_);
+  auto& grad_op_type = gop_type_it->second;
+  OperatorBase::VarNameMap inputs;
+  OperatorBase::VarNameMap outputs;
+  TransOpArg(op, &inputs, OpArgType::IN, false);   // I
+  TransOpArg(op, &inputs, OpArgType::OUT, false);  // O
+  TransOpArg(op, &inputs, OpArgType::OUT, true);   // OG
+  TransOpArg(op, &outputs, OpArgType::IN, true);   // IG
+  auto gop_it = OpRegistry::op_creators().find(grad_op_type);
+  PADDLE_ENFORCE(gop_it != OpRegistry::op_creators().end(),
+                 "Operator %s 's Gradient %s's creator cannot be found",
+                 op->type_, grad_op_type);
+
+  return gop_it->second(grad_op_type, inputs, outputs, op->attrs_);
 }
 
 }  // namespace framework

paddle/framework/grad_op_builder_test.cc

@@ -10,6 +10,7 @@ namespace framework {
 
 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 {}
@@ -20,10 +21,10 @@ class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
   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 +34,10 @@ 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().AsNoGradient();
+    AddInput("In3_mult", "another multiple input").AsDuplicable();
+    AddOutput("Out1_mult", "a multiple output").AsDuplicable();
+    AddOutput("Out2", "a single output").AsNoGradient();
     AddComment("op with inputs and outputs ignored in gradient calculating");
   }
 };
@@ -47,18 +48,18 @@ 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);
@@ -67,15 +68,15 @@ REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker);
 REGISTER_GRADIENT_OP(io_ignored, 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"}));
@@ -83,55 +84,49 @@ TEST(GradOpBuilder, MutiInOut) {
   EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
   EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
             std::vector<std::string>({"out2_1", "out2_2"}));
-  EXPECT_EQ(grad_test_op->Input("Out1" + f::kGradVarSuffix),
-            "out1" + f::kGradVarSuffix);
-  EXPECT_EQ(grad_test_op->Inputs("Out2_mult" + f::kGradVarSuffix),
+  EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")),
+            f::GradVarName("out1"));
+  EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")),
             std::vector<std::string>(
-                {"out2_1" + f::kGradVarSuffix, "out2_2" + f::kGradVarSuffix}));
+                {f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
 
-  ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
-  EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
-            "in1" + f::kGradVarSuffix);
-  EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
-            std::vector<std::string>({"in2_1" + f::kGradVarSuffix,
-                                      "in2_2" + f::kGradVarSuffix,
-                                      "in2_3" + f::kGradVarSuffix}));
-  EXPECT_EQ(grad_test_op->Output("In3" + f::kGradVarSuffix),
-            "in3" + f::kGradVarSuffix);
+  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"),
+                                      f::GradVarName("in2_2"),
+                                      f::GradVarName("in2_3")}));
+  EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3"));
 }
 
 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("Out1_mult" + f::kGradVarSuffix),
+  EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")),
             std::vector<std::string>(
-                {"out1_1" + f::kGradVarSuffix, "out1_2" + f::kGradVarSuffix}));
-  EXPECT_EQ(grad_test_op->Input("Out2" + f::kGradVarSuffix),
-            "out2" + f::kGradVarSuffix);
+                {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);
-  EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
-            "in1" + f::kGradVarSuffix);
-  EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
+  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>(
-                {"in2_1" + f::kGradVarSuffix, "in2_2" + f::kGradVarSuffix}));
-  EXPECT_EQ(grad_test_op->Outputs("In3_mult" + f::kGradVarSuffix),
+                {f::GradVarName("in2_1"), f::GradVarName("in2_2")}));
+  EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")),
             std::vector<std::string>(
-                {"in3_1" + f::kGradVarSuffix, "in3_2" + f::kGradVarSuffix}));
+                {f::GradVarName("in3_1"), f::GradVarName("in3_2")}));
 }

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_desc.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;
-
-import "attribute.proto";
-
-// AttrDesc is used to describe Attributes of an Operator. It contain's
-// name, type, and value of Attribute.
-//
-// e.g, for scale=3.0: name=scala, type=AttrType.FLOAT, value=3.0
-message AttrDesc {
-  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;
-};
-
-// Protocol Message to describe an Operator.
-//
-// In PaddlePaddle, Operator is used to do a certain computation such
-// as "add", "sub", "cosine", etc.
-//  (1) Operator needs to know the input and output variable names.
-//  (2) Some ops may have special attributes such as "scale" in "CosineOp".
-//
-// 3rd-party language can build this proto message and call
-// AddOp(const OpDesc& op_desc) of Paddle core to create an Operator.
-message OpDesc {
-  // input names of this Operator.
-  repeated string inputs = 1;
-
-  // output names of this Operator.
-  repeated string outputs = 2;
-
-  // type of this Operator, such as "add", "sub", "fc".
-  required string type = 3;
-
-  // Attributes of this Operator. e.g., scale=3.0 in cosine op.
-  repeated AttrDesc attrs = 4;
-};
\ No newline at end of file

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.h

@@ -20,8 +20,9 @@ limitations under the License. */
 #include <unordered_map>
 #include <unordered_set>
 #include "paddle/framework/attribute.h"
+#include "paddle/framework/framework.pb.h"
 #include "paddle/framework/grad_op_builder.h"
-#include "paddle/framework/op_desc.pb.h"
+#include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
 
 namespace paddle {
@@ -44,111 +45,58 @@ class OpProtoAndCheckerMaker {
 
  protected:
   struct VariableBuilder {
-    VarProto* var_;
-    std::function<void()> on_multiple_;
-    std::function<void()> on_temporary_;
+    OpProto::Var* var_;
 
-    VariableBuilder& SetMultiple() {
-      var_->set_multiple(true);
-      on_multiple_();
+    VariableBuilder& AsDuplicable() {
+      var_->set_duplicable(true);
       return *this;
     }
 
-    VariableBuilder& SetTemporary() {
-      PADDLE_ENFORCE(bool(on_temporary_), "Cannot set temporary");
-      var_->set_temporary(true);
-      on_temporary_();
+    VariableBuilder& AsIntermediate() {
+      var_->set_intermediate(true);
       return *this;
     }
 
-    VariableBuilder& IgnoreGradient() {
-      var_->set_ignore_gradient(true);
+    // TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it
+    // means that input/output is not needed when calculate gradient. It does
+    // not mean no gradient when backward. It should be changed soon.
+    VariableBuilder& AsNoGradient() {
+      var_->set_no_gradient(true);
       return *this;
     }
   };
 
   VariableBuilder AddInput(const std::string& name,
                            const std::string& comment) {
-    auto input = proto_->mutable_inputs()->Add();
-    *input->mutable_name() = name;
-    *input->mutable_comment() = comment;
-    return VariableBuilder{input, [=] { this->SetHasMultipleInput(); },
-                           nullptr};
+    auto* input = proto_->add_inputs();
+    input->set_name(name);
+    input->set_comment(comment);
+    return VariableBuilder{input};
   }
 
   VariableBuilder AddOutput(const std::string& name,
                             const std::string& comment) {
-    auto output = proto_->mutable_outputs()->Add();
-    *output->mutable_name() = name;
-    *output->mutable_comment() = comment;
-    return VariableBuilder{output, [=] { this->SetHasMultipleOutput(); },
-                           [=] { this->SetHasTemporaryOutput(); }};
+    auto* output = proto_->add_outputs();
+    output->set_name(name);
+    output->set_comment(comment);
+    return VariableBuilder{output};
   }
 
   template <typename T>
   TypedAttrChecker<T>& AddAttr(const std::string& name,
                                const std::string& comment,
                                bool generated = false) {
-    auto attr = proto_->mutable_attrs()->Add();
-    *attr->mutable_name() = name;
-    *attr->mutable_comment() = comment;
+    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_->mutable_comment()) = comment;
-  }
+  void AddComment(const std::string& comment) { proto_->set_comment(comment); }
 
  private:
-  void SetHasMultiple(const std::string& in_out, bool* flag) {
-    if (!*flag) {
-      AddAttr<std::vector<int>>(in_out + "_format",
-                                "The multiple index of " + in_out +
-                                    "\n"
-                                    R"DOC(
-This attribute is used by Paddle core framework. Paddle's Op support each input
-or output could be a list of variable. This attribute is used to show how that
-list organized.
-
-e.g.
-  input = ["a", "b", "c", "d", "e", "f"]
-  input_format = [0, 4, 5, 6]
-
-means
-  The number of all input variables this op is six, and they are segmented into
-  three inputs.
-
-  The first input is input[0:4], second is input[4:5], third is input[5:6].
-)DOC",
-                                /*generated*/ true);
-      *flag = true;
-    }
-  }
-
-  void SetHasMultipleInput() { SetHasMultiple("input", &has_multiple_input_); }
-  void SetHasMultipleOutput() {
-    SetHasMultiple("output", &has_multiple_output_);
-  }
-
-  void SetHasTemporaryOutput() {
-    if (!has_temporary_output_) {
-      AddAttr<std::vector<int>>("temporary_index",
-                                R"DOC(The temporary index of output.
-
-Not all output of Paddle Op is used by user. For faster computation, each op
-could output some its internal state to other op, other op could take that
-output to make compute faster.
-
-Add a mark to which output is temporary is helpful for future optimization.
-)DOC",
-                                /*generated*/ true)
-          .SetDefault(std::vector<int>());
-      has_temporary_output_ = true;
-    }
-  }
-
   void CheckNoDuplicatedInOutAttrs() {
     std::unordered_set<std::string> names;
     auto checker = [&](const std::string& name) {
@@ -169,89 +117,74 @@ Add a mark to which output is temporary is helpful for future optimization.
   OpProto* proto_;
   OpAttrChecker* op_checker_;
   bool validated_{false};
-  bool has_multiple_input_{false};
-  bool has_multiple_output_{false};
-  bool has_temporary_output_{false};
 };
 
 class OpRegistry {
-  using OpCreator = std::function<OperatorBase*()>;
-  using VarIndexMap = std::unordered_map<std::string, int>;
-  using VarNameList = std::vector<std::string>;
+  using VarNameMap = OperatorBase::VarNameMap;
+  using OpCreator = std::function<OperatorBase*(
+      const std::string& /*type*/, const VarNameMap& /*inputs*/,
+      const VarNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
 
  public:
   template <typename OpType, typename ProtoMakerType>
   static void RegisterOp(const std::string& op_type) {
-    op_creators()[op_type] = [] { return new OpType; };
+    op_creators()[op_type] = [](
+        const std::string& type, const VarNameMap& inputs,
+        const VarNameMap& outputs, const AttributeMap& attrs) {
+      return new OpType(type, inputs, outputs, attrs);
+    };
     OpAttrChecker& op_checker = op_checkers()[op_type];
-    OpProto& op_proto = protos()[op_type];
+    OpProto& op_proto = OpProtos()[op_type];
     auto maker = ProtoMakerType(&op_proto, &op_checker);
     maker.Validate();
-    *op_proto.mutable_type() = op_type;
+    op_proto.set_type(op_type);
     PADDLE_ENFORCE(
         op_proto.IsInitialized(),
         "Fail to initialize %s's OpProto, because %s is not initialized",
         op_type, op_proto.InitializationErrorString());
-
-    VarIndexMaps()[op_type].reset(new VarIndexMap());
-    auto& varmap = *VarIndexMaps()[op_type];
-    int idx = 0;
-    for (auto& var : op_proto.inputs()) {
-      varmap[var.name()] = idx++;
-    }
-    idx = 0;
-    for (auto& var : op_proto.outputs()) {
-      varmap[var.name()] = idx++;
-    }
   }
 
   template <typename GradOpType>
   static void RegisterGradOp(const std::string& op_type,
                              const std::string& grad_op_type) {
-    op_creators()[grad_op_type] = [] { return new GradOpType; };
+    op_creators()[grad_op_type] = [](
+        const std::string& type, const VarNameMap& inputs,
+        const VarNameMap& outputs, const AttributeMap& attrs) {
+      return new GradOpType(type, inputs, outputs, attrs);
+    };
     grad_ops()[op_type] = grad_op_type;
   }
 
   static std::shared_ptr<OperatorBase> CreateOp(const std::string& type,
-                                                const VarNameList& inputs,
-                                                const VarNameList& outputs,
-                                                const AttributeMap& attrs) {
+                                                const VarNameMap& inputs,
+                                                const VarNameMap& outputs,
+                                                AttributeMap attrs) {
     auto op_create_it = op_creators().find(type);
     PADDLE_ENFORCE(op_create_it != op_creators().end(),
                    "Operator %s cannot be found.", type);
+    op_checkers().at(type).Check(attrs);
 
-    auto op = op_create_it->second();
-    op->type_ = type;
-    op->inputs_ = inputs;
-    op->outputs_ = outputs;
-
-    op->attrs_ = attrs;
-    op_checkers().at(type).Check(op->attrs_);
+    auto op = op_create_it->second(type, inputs, outputs, attrs);
 
-    GenerateTempVariableName(op);
+    return std::shared_ptr<OperatorBase>(op);
+  }
 
-    {
-      auto var_index_it = VarIndexMaps().find(type);
-      if (var_index_it != VarIndexMaps().end()) {
-        op->in_out_idxs_ = var_index_it->second;
-      }
+  static VarNameMap 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));
     }
-
-    op->Init();
-    return std::shared_ptr<OperatorBase>(op);
+    return ret_val;
   }
 
   static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc) {
-    std::vector<std::string> inputs;
-    inputs.reserve((size_t)op_desc.inputs_size());
-    std::copy(op_desc.inputs().begin(), op_desc.inputs().end(),
-              std::back_inserter(inputs));
-
-    std::vector<std::string> outputs;
-    outputs.reserve((size_t)op_desc.outputs_size());
-    std::copy(op_desc.outputs().begin(), op_desc.outputs().end(),
-              std::back_inserter(outputs));
-
+    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);
@@ -264,26 +197,14 @@ class OpRegistry {
     PADDLE_ENFORCE(!op.IsNetOp(),
                    "Use framework::Backward to get backward ops");
     std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
-    grad_op->Init();
     return grad_op;
   }
 
-  static std::unordered_map<std::string, OpProto>& protos() {
-    static std::unordered_map<std::string, OpProto> protos_;
-    return protos_;
-  }
-
   static std::unordered_map<std::string, std::string>& grad_ops() {
     static std::unordered_map<std::string, std::string> grad_ops_;
     return grad_ops_;
   }
 
-  static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>>&
-  VarIndexMaps() {
-    static std::unordered_map<std::string, std::shared_ptr<VarIndexMap>> maps_;
-    return maps_;
-  }
-
   static std::unordered_map<std::string, OpCreator>& op_creators() {
     static std::unordered_map<std::string, OpCreator> op_creators_;
     return op_creators_;
@@ -294,35 +215,47 @@ class OpRegistry {
     static std::unordered_map<std::string, OpAttrChecker> op_checkers_;
     return op_checkers_;
   }
+};
 
-  static void GenerateTempVariableName(OperatorBase* op) {
-    static std::atomic<size_t> gUniqId(0UL);
-    for (auto& outname : op->outputs_) {
-      if (outname == kTempVarName) {
-        outname += op->type_;
-        outname += "@";
-        outname += std::to_string(gUniqId.fetch_add(1));
-      }
-    }
-  }
+class Registrar {
+ public:
+  // In our design, various kinds of classes, e.g., operators and kernels, have
+  // their corresponding registry and registrar. The action of registration is
+  // in the constructor of a global registrar variable, which, however, are not
+  // used in the code that calls package framework, and would be removed from
+  // the generated binary file by the linker. To avoid such removal, we add
+  // Touch to all registrar classes and make USE_OP macros to call this
+  // method. So, as long as the callee code calls USE_OP, the global
+  // registrar variable won't be removed by the linker.
+  void Touch() {}
 };
 
 template <typename OpType, typename ProtoMakerType>
-class OpRegisterHelper {
+class OpRegistrar : public Registrar {
  public:
-  explicit OpRegisterHelper(const char* op_type) {
+  explicit OpRegistrar(const char* op_type) {
     OpRegistry::RegisterOp<OpType, ProtoMakerType>(op_type);
   }
 };
 
 template <typename GradOpType>
-class GradOpRegisterHelper {
+class GradOpRegistrar : public Registrar {
  public:
-  GradOpRegisterHelper(const char* op_type, const char* grad_op_type) {
+  GradOpRegistrar(const char* op_type, const char* grad_op_type) {
     OpRegistry::RegisterGradOp<GradOpType>(op_type, grad_op_type);
   }
 };
 
+template <typename PlaceType, typename KernelType>
+class OpKernelRegistrar : public Registrar {
+ public:
+  explicit OpKernelRegistrar(const char* op_type) {
+    OperatorWithKernel::OpKernelKey key;
+    key.place_ = PlaceType();
+    OperatorWithKernel::AllOpKernels()[op_type][key].reset(new KernelType);
+  }
+};
+
 /**
  * check if MACRO is used in GLOBAL NAMESPACE.
  */
@@ -333,97 +266,121 @@ class GradOpRegisterHelper {
                 msg)
 
 /**
- * Macro to Register Operator.
+ * Macro to register Operator.
  */
-#define REGISTER_OP(__op_type, __op_class, __op_maker_class)                 \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(__reg_op__##__op_type,                      \
-                                 "REGISTER_OP must be in global namespace"); \
-  static ::paddle::framework::OpRegisterHelper<__op_class, __op_maker_class> \
-      __op_register_##__op_type##__(#__op_type);                             \
-  int __op_register_##__op_type##_handle__() { return 0; }
+#define REGISTER_OP(op_type, op_class, op_maker_class)                        \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                             \
+      __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
+  static ::paddle::framework::OpRegistrar<op_class, op_maker_class>           \
+      __op_registrar_##op_type##__(#op_type);                                 \
+  int TouchOpRegistrar_##op_type() {                                          \
+    __op_registrar_##op_type##__.Touch();                                     \
+    return 0;                                                                 \
+  }
 
 /**
- * Macro to Register Gradient Operator.
+ * Macro to register Gradient Operator.
  */
-#define REGISTER_GRADIENT_OP(__op_type, __grad_op_type, __grad_op_class)       \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                              \
-      __reg_gradient_op__##__op_type##__grad_op_type,                          \
-      "REGISTER_GRADIENT_OP must be in global namespace");                     \
-  static ::paddle::framework::GradOpRegisterHelper<__grad_op_class>            \
-      __op_gradient_register_##__op_type##__grad_op_type##__(#__op_type,       \
-                                                             #__grad_op_type); \
-  int __op_gradient_register_##__op_type##__grad_op_type##_handle__() {        \
-    return 0;                                                                  \
+#define REGISTER_GRADIENT_OP(op_type, grad_op_type, grad_op_class)           \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                            \
+      __reg_gradient_op__##op_type##_##grad_op_type,                         \
+      "REGISTER_GRADIENT_OP must be called in global namespace");            \
+  static ::paddle::framework::GradOpRegistrar<grad_op_class>                 \
+      __op_gradient_registrar_##op_type##_##grad_op_type##__(#op_type,       \
+                                                             #grad_op_type); \
+  int TouchOpGradientRegistrar_##op_type() {                                 \
+    __op_gradient_registrar_##op_type##_##grad_op_type##__.Touch();          \
+    return 0;                                                                \
   }
 
 /**
- * Macro to Forbid user register Gradient Operator.
+ * Macro to register OperatorKernel.
  */
-#define NO_GRADIENT(__op_type)                          \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                       \
-      __reg_gradient_op__##__op_type##__op_type##_grad, \
-      "NO_GRADIENT must be in global namespace")
+#define REGISTER_OP_KERNEL(op_type, DEVICE_TYPE, place_class, ...)        \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
+      __reg_op_kernel_##op_type##_##DEVICE_TYPE##__,                      \
+      "REGISTER_OP_KERNEL must be called in global namespace");           \
+  static ::paddle::framework::OpKernelRegistrar<place_class, __VA_ARGS__> \
+      __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type);      \
+  int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() {                \
+    __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__.Touch();          \
+    return 0;                                                             \
+  }
 
 /**
- * Macro to Register OperatorKernel.
+ * Macro to Forbid user register Gradient Operator.
  */
-#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, ...)             \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
-      __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
-      "REGISTER_OP_KERNEL must be in global namespace");                  \
-  struct __op_kernel_register__##type##__##DEVICE_TYPE##__ {              \
-    __op_kernel_register__##type##__##DEVICE_TYPE##__() {                 \
-      ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
-      key.place_ = PlaceType();                                           \
-      ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
-          .reset(new __VA_ARGS__());                                      \
-    }                                                                     \
-  };                                                                      \
-  static __op_kernel_register__##type##__##DEVICE_TYPE##__                \
-      __reg_kernel_##type##__##DEVICE_TYPE##__;                           \
-  int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
-
-// (type, KernelType)
-#define REGISTER_OP_GPU_KERNEL(type, ...) \
-  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
-
-// (type, KernelType)
-#define REGISTER_OP_CPU_KERNEL(type, ...) \
-  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
+#define NO_GRADIENT(op_type)                           \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                      \
+      __reg_gradient_op__##op_type##_##op_type##_grad, \
+      "NO_GRADIENT must be called in global namespace")
+
+#define REGISTER_OP_GPU_KERNEL(op_type, ...) \
+  REGISTER_OP_KERNEL(op_type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
+
+#define REGISTER_OP_CPU_KERNEL(op_type, ...) \
+  REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
 /**
  * Macro to mark what Operator and Kernel we will use and tell the compiler to
  * link them into target.
  */
-#define USE_OP_WITHOUT_KERNEL(op_type)                      \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                           \
-      __use_op_without_kernel_##op_type,                    \
-      "USE_OP_WITHOUT_KERNEL must be in global namespace"); \
-  extern int __op_register_##op_type##_handle__();          \
-  static int __use_op_ptr_##op_type##_without_kernel__      \
-      __attribute__((unused)) = __op_register_##op_type##_handle__()
-
-#define USE_OP_KERNEL(op_type, DEVICE_TYPE)                               \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
-      __use_op_kernel_##op_type##_##DEVICE_TYPE##__,                      \
-      "USE_OP_KERNEL must be in global namespace");                       \
-  extern int __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__(); \
-  static int __use_op_ptr_##op_type##_##DEVICE_TYPE##_kernel__            \
-      __attribute__((unused)) =                                           \
-          __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__()
-
-// use Operator with only cpu kernel.
-#define USE_OP_CPU(op_type)       \
-  USE_OP_WITHOUT_KERNEL(op_type); \
-  USE_OP_KERNEL(op_type, CPU)
+#define USE_OP_ITSELF(op_type)                                    \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                 \
+      __use_op_itself_##op_type,                                  \
+      "USE_OP_ITSELF must be called in global namespace");        \
+  extern int TouchOpRegistrar_##op_type();                        \
+  static int use_op_itself_##op_type##_ __attribute__((unused)) = \
+      TouchOpRegistrar_##op_type()
+
+// TODO(fengjiayi): Most ops' gradient op have not been compeleted. So we use
+// `NO_GRAD` to disable micro USE_OP_GRADIENT(op_type). Otherwise the code can't
+// be compiled. `NO_GRAD` should be removed after all gradient ops are
+// compeleted.
+#define NO_GRAD
+#ifndef NO_GRAD
+#define USE_OP_GRADIENT(op_type)                                    \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                   \
+      __use_op_gradient_##op_type,                                  \
+      "USE_OP_GRADIENT must be called in global namespace");        \
+  extern int TouchOpGradientRegistrar_##op_type();                  \
+  static int use_op_gradient_##op_type##_ __attribute__((unused)) = \
+      TouchOpGradientRegistrar_##op_type()
+#else
+#define USE_OP_GRADIENT(op_type)
+#endif
+
+#define USE_OP_DEVICE_KERNEL(op_type, DEVICE_TYPE)               \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                \
+      __use_op_kernel_##op_type##_##DEVICE_TYPE##__,             \
+      "USE_OP_DEVICE_KERNEL must be in global namespace");       \
+  extern int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE(); \
+  static int use_op_kernel_##op_type##_##DEVICE_TYPE##_          \
+      __attribute__((unused)) =                                  \
+          TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE()
+
+// TODO(fengjiayi): The following macros seems ugly, do we have better method?
 
 #ifdef PADDLE_ONLY_CPU
-#define USE_OP(op_type) USE_OP_CPU(op_type)
+#define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)
 #else
-#define USE_OP(op_type) \
-  USE_OP_CPU(op_type);  \
-  USE_OP_KERNEL(op_type, GPU)
+#define USE_OP_KERNEL(op_type)        \
+  USE_OP_DEVICE_KERNEL(op_type, CPU); \
+  USE_OP_DEVICE_KERNEL(op_type, GPU)
 #endif
 
+#define USE_NO_GRAD_OP(op_type) \
+  USE_OP_ITSELF(op_type);       \
+  USE_OP_KERNEL(op_type)
+
+#define USE_CPU_OP(op_type)           \
+  USE_OP_ITSELF(op_type);             \
+  USE_OP_DEVICE_KERNEL(op_type, CPU); \
+  USE_OP_GRADIENT(op_type)
+
+#define USE_OP(op_type)    \
+  USE_NO_GRAD_OP(op_type); \
+  USE_OP_GRADIENT(op_type)
+
 }  // 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!");
     };
@@ -49,6 +51,15 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 }  // namespace framework
 }  // namespace paddle
 
+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(cos_sim, paddle::framework::CosineOp,
             paddle::framework::CosineOpProtoAndCheckerMaker);
 REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
@@ -57,8 +68,8 @@ REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
 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();
@@ -78,8 +89,8 @@ 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");
@@ -103,8 +114,8 @@ 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());
 
@@ -116,20 +127,11 @@ TEST(OpRegistry, DefaultValue) {
   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 +171,6 @@ 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 {
@@ -33,84 +33,139 @@ 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);
+static std::unordered_map<std::string, OpProto>* g_op_protos = nullptr;
+std::unordered_map<std::string, OpProto>& OpProtos() {
+  if (g_op_protos == nullptr) {
+    g_op_protos = new std::unordered_map<std::string, OpProto>();
   }
+  return *g_op_protos;
 }
 
-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::string& OperatorBase::Input(const std::string& name) const {
+  auto& ins = Inputs(name);
+  PADDLE_ENFORCE_EQ(ins.size(), 1UL,
+                    "Op %s input %s should contain only one variable", type_,
+                    name);
+  return ins[0];
 }
 
-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_",
+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);
-  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);
-  }
+  return it->second;
+}
+
+const std::string& OperatorBase::Output(const std::string& name) const {
+  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 = OpProtos().find(type_);
+  PADDLE_ENFORCE(
+      it != OpProtos().end(),
+      "Operator %s not registered, cannot figure out intermediate outputs",
+      type_);
+
+  // get all OpProto::Var for outputs
+  for (auto& o : it->second.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;
 }
 
 }  // namespace framework

paddle/framework/operator.h

@@ -15,42 +15,43 @@ limitations under the License. */
 #pragma once
 
 #include <algorithm>
-#include <boost/variant.hpp>
 #include <string>
 #include <unordered_map>
 #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"
 #include "paddle/platform/place.h"
+#include "paddle/platform/variant.h"
 #include "paddle/utils/Error.h"
 
 namespace paddle {
 namespace framework {
 
 /// If a variable is a empty variable, that name will be used.
-const std::string kEmptyVarName = "@EMPTY@";
+constexpr char kEmptyVarName[] = "@EMPTY@";
 
 /// If a variable is a temporary variable, that name will be set in Python,
 /// but it will be convert to a unique name in scope after OpCreator.
-const std::string kTempVarName = "@TEMP@";
+constexpr char kTempVarName[] = "@TEMP@";
 
 /// If a variable's name has a certain suffix, it means that the
 /// variable is the gradient of another varibale.
 /// e.g. Variable "x@GRAD" is the gradient of varibale "x".
-const std::string kGradVarSuffix = "@GRAD";
+constexpr char kGradVarSuffix[] = "@GRAD";
 
 /// Variables with this suffix are supposed to be filled up with zeros.
-const std::string kZeroVarSuffix = "@ZERO";
+constexpr char kZeroVarSuffix[] = "@ZERO";
 
 inline std::string GradVarName(const std::string& var_name) {
   return var_name + kGradVarSuffix;
 }
 
+extern std::unordered_map<std::string, OpProto>& OpProtos();
+
 class OperatorBase;
 class InferShapeContext;
 class ExecutionContext;
@@ -63,6 +64,15 @@ class ExecutionContext;
  */
 class OperatorBase {
  public:
+  using VarNameMap = std::map<std::string, std::vector<std::string>>;
+
+  OperatorBase(const std::string& type, const VarNameMap& inputs,
+               const VarNameMap& outputs, const AttributeMap& attrs);
+
+  OperatorBase(const OperatorBase& o) = delete;
+  OperatorBase& operator=(const OperatorBase& o) = delete;
+  OperatorBase(OperatorBase&& o) = delete;
+
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -74,10 +84,6 @@ 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;
@@ -95,15 +101,19 @@ class OperatorBase {
 
   //! 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;
+
+  virtual std::vector<std::string> OutputVars(bool has_intermediate) const;
+
+  std::string Type() const { return type_; }
+  const AttributeMap& Attrs() const { return attrs_; }
 
  public:
   std::string type_;
@@ -111,30 +121,25 @@ class OperatorBase {
   // I (Inputs)
   // 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_;
 };
 
-class OperatorContext {
+class InferShapeContext {
  public:
-  OperatorContext(const OperatorBase* op, const Scope& scope)
-      : op_(*op), scope_(scope) {}
+  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 +171,9 @@ class OperatorContext {
     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>();
   }
@@ -234,12 +221,6 @@ class OperatorContext {
   const Scope& scope_;
 };
 
-class InferShapeContext : public OperatorContext {
- public:
-  InferShapeContext(const OperatorBase* op, const Scope& scope)
-      : OperatorContext(op, scope) {}
-};
-
 template <typename T>
 struct EigenDeviceConverter;
 
@@ -255,11 +236,11 @@ struct EigenDeviceConverter<platform::GPUPlace> {
 };
 #endif
 
-class ExecutionContext : public OperatorContext {
+class ExecutionContext : public InferShapeContext {
  public:
-  ExecutionContext(const OperatorBase* op, const Scope& scope,
+  ExecutionContext(const OperatorBase& op, const Scope& scope,
                    const platform::DeviceContext* device_context)
-      : OperatorContext(op, scope), device_context_(device_context) {}
+      : InferShapeContext(op, scope), device_context_(device_context) {}
 
   template <typename PlaceType,
             typename DeviceType =
@@ -268,6 +249,10 @@ class ExecutionContext : public OperatorContext {
 
   platform::Place GetPlace() const { return device_context_->GetPlace(); }
 
+  const platform::DeviceContext* device_context() const {
+    return device_context_;
+  }
+
   const platform::DeviceContext* device_context_;
 };
 
@@ -310,14 +295,18 @@ 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));
+    InferShape(InferShapeContext(*this, scope));
   }
 
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const final {
     auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
-    opKernel->Compute(ExecutionContext(this, scope, &dev_ctx));
+    opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
   }
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&

paddle/framework/operator_test.cc

@@ -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 +56,24 @@ class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 }  // namespace framework
 }  // namespace paddle
 
+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(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 +109,9 @@ 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 +128,15 @@ 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);
@@ -196,8 +193,9 @@ 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);
@@ -223,32 +221,15 @@ 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>();

paddle/framework/pybind.cc

@@ -22,6 +22,7 @@ limitations under the License. */
 #include "paddle/operators/net_op.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
+#include "paddle/string/to_string.h"
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
 #include "pybind11/stl.h"
@@ -29,17 +30,18 @@ limitations under the License. */
 namespace py = pybind11;
 
 USE_OP(add_two);
-USE_OP_CPU(onehot_cross_entropy);
-USE_OP_WITHOUT_KERNEL(fc);
-USE_OP(sgd);
+USE_CPU_OP(onehot_cross_entropy);
+USE_NO_GRAD_OP(sgd);
 USE_OP(mul);
 USE_OP(mean);
 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);
+
 namespace paddle {
 namespace framework {
 
@@ -54,30 +56,18 @@ void ExposeOperator(ClassType &m) {
              return op.type_;
            })
       .def("outputs",
-           [](const typename ClassType::type &op) -> std::vector<std::string> {
-             return op.outputs_;
-           })
+           [](const typename ClassType::type &op)
+               -> std::map<std::string, std::vector<std::string>> {
+                 return op.outputs_;
+               })
       .def("inputs",
-           [](const typename ClassType::type &op) -> std::vector<std::string> {
-             return op.inputs_;
+           [](const typename ClassType::type &op) { return op.inputs_; })
+      .def("__str__", &ClassType::type::DebugString)
+      .def("no_intermediate_outputs",
+           [](const typename ClassType::type &op) {
+             return op.OutputVars(false);
            })
-      .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);
+      .def("support_gpu", &ClassType::type::SupportGPU);
 }
 
 static size_t UniqueIntegerGenerator() {
@@ -170,7 +160,7 @@ 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 &protos = OpProtos();
     std::vector<py::bytes> ret_values;
     for (auto it = protos.begin(); it != protos.end(); ++it) {
       PADDLE_ENFORCE(it->second.IsInitialized(),
@@ -205,9 +195,13 @@ All parameter, weight, gradient are variables in Paddle.
                   });
   // clang-format on
 
-  py::class_<paddle::platform::GPUPlace>(m, "GPUPlace").def(py::init<int>());
+  py::class_<platform::GPUPlace>(m, "GPUPlace")
+      .def(py::init<int>())
+      .def("__str__", string::to_string<const platform::GPUPlace &>);
 
-  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace").def(py::init<>());
+  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
+      .def(py::init<>())
+      .def("__str__", string::to_string<const platform::CPUPlace &>);
 
   py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
       m, "Operator");

paddle/framework/tensor.h

@@ -79,11 +79,11 @@ class Tensor {
   inline const DDim& dims() const;
 
   /*! Resize the dimensions of the memory block. */
-  inline void Resize(const DDim& dims);
+  inline Tensor& Resize(const DDim& dims);
 
   /*! The internal of two tensors share the same memory block. */
   template <typename T>
-  inline void ShareDataWith(const Tensor& src);
+  inline Tensor& ShareDataWith(const Tensor& src);
 
   /**
    * @brief   Copy the content of external tensor to a new place.
@@ -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/framework/tensor_impl.h

@@ -23,9 +23,11 @@ template <typename T>
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
       holder_, "Tenosr holds no memory. Call Tensor::mutable_data first.");
-  PADDLE_ENFORCE_GE(holder_->size(), product(dims_) * sizeof(T) + offset_,
-                    "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
-                    "first to re-allocate memory.");
+  PADDLE_ENFORCE_GE(
+      holder_->size(), product(dims_) * sizeof(T) + offset_,
+      "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
+      "first to re-allocate memory.\n"
+      "or maybe the required data-type mismatches the data already stored.");
 }
 
 template <typename T>
@@ -78,9 +80,10 @@ inline T* Tensor::mutable_data(platform::Place place) {
 }
 
 template <typename T>
-inline void Tensor::ShareDataWith(const Tensor& src) {
+inline Tensor& Tensor::ShareDataWith(const Tensor& src) {
   src.check_memory_size<T>();
   *this = src;
+  return *this;
 }
 
 template <typename T>
@@ -136,7 +139,10 @@ inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
   return dst;
 }
 
-inline void Tensor::Resize(const DDim& dims) { dims_ = dims; }
+inline Tensor& Tensor::Resize(const DDim& dims) {
+  dims_ = dims;
+  return *this;
+}
 
 inline const DDim& Tensor::dims() const { return dims_; }
 

paddle/function/CMakeLists.txt

@@ -38,10 +38,11 @@ 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/operators/mean_op_test.cc → paddle/function/DepthwiseConvOpTest.cpp

@@ -13,13 +13,25 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <gtest/gtest.h>
+#include "ConvOpTest.h"
 
-#include <paddle/framework/op_registry.h>
+namespace paddle {
 
-USE_OP(mean);
+#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(MeanOp, GetOpProto) {
-  auto& protos = paddle::framework::OpRegistry::protos();
-  auto it = protos.find("mean");
-  ASSERT_NE(it, protos.end());
+TEST(DepthwiseConv, BackwardFilter) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradFilter-CPU", "DepthwiseConvGradFilter-GPU", backward_filter);
 }
+#endif
+
+}  // namespace paddle

paddle/operators/add_op_test.cc → paddle/function/GemmConvOpTest.cpp

@@ -13,16 +13,38 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <gtest/gtest.h>
-#define private public
-#include "paddle/framework/op_registry.h"
-
-USE_OP(add_two);
-
-TEST(AddOp, GetOpProto) {
-  auto& protos = paddle::framework::OpRegistry::protos();
-  auto it = protos.find("add_two");
-  ASSERT_NE(it, protos.end());
-  auto& op_creators = paddle::framework::OpRegistry::op_creators();
-  auto it1 = op_creators.find("add_two_grad");
-  ASSERT_NE(it1, op_creators.end());
+#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/function/nnpack/NNPACKConvOpTest.cpp

@@ -13,87 +13,18 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <gtest/gtest.h>
-#include "paddle/function/Function.h"
-#include "paddle/function/FunctionTest.h"
-
-DEFINE_string(algo,
-              "auto",
-              "The algorithm (auto, ft8x8, ft16x16, wt8x8, "
-              "implicit-gemm, or direct) for computing convolution of NNPACK.");
+#include "paddle/function/ConvOpTest.h"
 
 namespace paddle {
 
-#define IS_NNPACK_SUPPORT(algo, filterSize, stride)        \
-  if (algo == "direct" && filterSize != 1) continue;       \
-  if (algo == "direct" && batchSize != 1) continue;        \
-  if (algo == "wt8x8" && filterSize != 3) continue;        \
-  if (algo == "implicit-gemm" && batchSize != 1) continue; \
-  if (algo != "auto" && algo != "implicit-gemm" && stride > 1) continue;
-
-class ConvolutionTest {
-public:
-  ConvolutionTest(const std::string& conv1,
-                  const std::string& conv2,
-                  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, 128}) {
-              if (inputChannels < outputChannels) break;
-              for (size_t stride : {1, 2}) {
-                // if batchSize > 1 NNPACKConv only supports stride = 1
-                if (batchSize > 1 && stride > 1) break;
-                for (size_t padding : {0, 1}) {
-                  if (padding >= filterSize) break;
-                  size_t outputSize =
-                      (inputSize - filterSize + 2 * padding + stride) / stride;
-                  IS_NNPACK_SUPPORT(algo, filterSize, stride);
-                  LOG(INFO) << " 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<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU> test(
-                      conv1,
-                      conv2,
-                      FuncConfig()
-                          .set("paddings", paddings)
-                          .set("strides", strides)
-                          .set("groups", (size_t)1)
-                          .set("algo", algo));
-
-                  TensorShape shape0{
-                      batchSize, inputChannels, inputSize, inputSize};
-                  TensorShape shape1{
-                      outputChannels, inputChannels, filterSize, filterSize};
-                  TensorShape shape2{
-                      batchSize, outputChannels, outputSize, outputSize};
-                  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape0));
-                  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape1));
-                  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape2));
-                  test.run();
-                }
-              }
-            }
-          }
-        }
-      }
-    }
-  }
-};
+TEST(NNPACK, Forward) {
+  Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
+      "GemmConv-CPU", "NNPACKConv-CPU", forward);
+}
 
-TEST(Convolution, NNPACK) {
-  // NNPACK only supports stride = 1
-  ConvolutionTest test("GemmConv-CPU", "NNPACKConv-CPU", FLAGS_algo);
+TEST(NNPACK, Depthwise) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
+      "GemmConv-CPU", "NNPACKConv-CPU", forward);
 }
 
 }  // namespace paddle

paddle/gserver/CMakeLists.txt

@@ -23,6 +23,17 @@ endmacro()
 
 filter_test(GSERVER_HEADER)
 filter_test(GSERVER_SOURCES)
+
+if(NOT WITH_MKLDNN)
+    file(GLOB_RECURSE DNN_HEADER RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "MKLDNN*.h")
+    file(GLOB_RECURSE DNN_SOURCES RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "MKLDNN*.cpp")
+    list(REMOVE_ITEM GSERVER_HEADER ${DNN_HEADER})
+    list(REMOVE_ITEM GSERVER_SOURCES ${DNN_SOURCES})
+    message(STATUS "Skip compiling with MKLDNNLayers and MKLDNNActivations")
+else()
+    message(STATUS "Compile with MKLDNNLayers and MKLDNNActivations")
+endif()
+
 if(NOT WITH_GPU)
     list(REMOVE_ITEM GSERVER_HEADER
         layers/CudnnConvBaseLayer.h

paddle/gserver/activations/ActivationFunction.cpp

@@ -112,7 +112,6 @@ BEGIN_DEFINE_ACTIVATION(softmax)
 private:
 MatrixPtr sftMaxSum_;
 MatrixPtr sftMaxDot_;
-MatrixPtr one_;
 
 public:
 Error __must_check forward(Argument& act) {
@@ -138,14 +137,6 @@ Error __must_check backward(Argument& act) {
                            1,
                            /* trans */ false,
                            useGpu(act.deviceId));
-    if (!one_ || one_->getWidth() != outputG->getWidth()) {
-      Matrix::resizeOrCreate(one_,
-                             1,
-                             outputG->getWidth(),
-                             /* trans */ false,
-                             useGpu(act.deviceId));
-      one_->one();
-    }
 
     sftMaxDot_->dotMul(*outputG, *outputV);
     sftMaxSum_->colMerge(*sftMaxDot_);

paddle/gserver/layers/MKLDNNBase.h

+/* Copyright (c) 2017 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 "mkldnn.hpp"
+
+namespace paddle {
+
+typedef enum {
+  MKLDNN_BASE = 1,   // basical info of MKLDNN
+  MKLDNN_TESTS = 1,  // gtest info of MKLDNN
+  MKLDNN_SIZES = 2,  // size info of MKLDNN
+  MKLDNN_FMTS = 3,   // format info of MKLDNN
+  MKLDNN_ALL = 4,    // show all info of MKLDNN
+} MKLDNN_LOG_LEVEL;
+
+/**
+ * @brief MKLDNN CPU engine.
+ *
+ */
+class CPUEngine {
+public:
+  static CPUEngine& Instance() {
+    // Thread-safe in C++11.
+    static CPUEngine myInstance;
+    return myInstance;
+  }
+
+  // Disallow copy or move
+  CPUEngine(const CPUEngine&) = delete;             // Copy constructor
+  CPUEngine(CPUEngine&&) = delete;                  // Move constructor
+  CPUEngine& operator=(const CPUEngine&) = delete;  // Copy assignment
+  CPUEngine& operator=(CPUEngine&&) = delete;       // Move assignment
+
+  mkldnn::engine& getEngine() { return cpuEngine_; }
+
+protected:
+  CPUEngine() : cpuEngine_(mkldnn::engine::cpu, 0) {}
+  //    CPUEngine() : cpuEngine_(mkldnn::engine::cpu_lazy, 0) {}
+  ~CPUEngine() {}
+
+private:
+  mkldnn::engine cpuEngine_;
+};
+
+/**
+ * @brief MKLDNN Stream.
+ *
+ */
+class MKLDNNStream {
+public:
+  MKLDNNStream() : ready_(false) { resetState(); }
+
+  virtual ~MKLDNNStream() {}
+
+  /**
+   * @brief Submit stream
+   * @param prims The primitives vector
+   * @param block Waiting for the stream to complete
+   */
+  void submit(std::vector<mkldnn::primitive>& prims, bool block = true) {
+    resetState();
+    stream_->submit(prims).wait(block);
+    ready_ = false;
+  }
+
+  /**
+   * @brief Reset the mkldnn stream
+   */
+  void resetState() {
+    if (ready_) {
+      return;
+    }
+    // TODO(TJ): change me when mkldnn have method to reset this state
+    // stream_.reset(new mkldnn::stream(mkldnn::stream::kind::lazy));
+    stream_.reset(new mkldnn::stream(mkldnn::stream::kind::eager));
+    ready_ = true;
+  }
+
+private:
+  bool ready_;
+  std::shared_ptr<mkldnn::stream> stream_;
+};
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNFcLayer.cpp

+/* Copyright (c) 2017 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 "MKLDNNFcLayer.h"
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/Stat.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+typedef inner_product_forward fc_fwd;
+typedef inner_product_backward_weights fc_bwdWgt;
+typedef inner_product_backward_data fc_bwdData;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_fc, MKLDNNFcLayer);
+
+bool MKLDNNFcLayer::init(const LayerMap& layerMap,
+                         const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+
+  CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet";
+  CHECK_EQ(inputLayers_.size(), parameters_.size());
+  CHECK(!parameters_[0]->isSparse()) << "Do not support sparse yet";
+
+  // output size, cat not be changed
+  oc_ = getSize();
+  oh_ = 1;
+  ow_ = 1;
+
+  // input size can not change in FC
+  iLayerSize_ = inputLayers_[0]->getSize();
+  CHECK_EQ(parameters_[0]->getSize(), iLayerSize_ * oc_);
+
+  // create weight
+  weight_ =
+      std::unique_ptr<Weight>(new Weight(oc_, iLayerSize_, parameters_[0], 0));
+
+  // create biases
+  if (biasParameter_.get() != NULL) {
+    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_));
+  }
+  return true;
+}
+
+void MKLDNNFcLayer::convertWeightsFromPaddle() {
+  if (FLAGS_use_mkldnn_wgt) {
+    return;
+  }
+
+  if (hasInitedWgt_) {
+    return;
+  }
+
+  // The weight_ is transposed from initial paddle weight
+  MatrixPtr paddleWgt = Matrix::create(
+      weight_->getW()->getData(), iLayerSize_, oc_, false, false);
+
+  // TODO(TJ): remove this print when do not need differ weights
+  std::ostringstream ostr;
+  paddleWgt->print(ostr);
+  VLOG(MKLDNN_ALL) << "Initial Weight from paddle: " << std::endl << ostr.str();
+
+  // The mkldnn weight is transposed from initial paddle matrix
+  MatrixPtr paddleWgtT;
+  paddleWgt->transpose(paddleWgtT, true);
+  weight_->getW()->copyFrom(*paddleWgtT);
+  hasInitedWgt_ = true;
+}
+
+void MKLDNNFcLayer::convertWeightsToPaddle() {
+  MatrixPtr dnnWgt = weight_->getW();
+  MatrixPtr paddleWgt;
+  dnnWgt->transpose(paddleWgt, true);
+
+  // copy paddle weight and override on weight_
+  MatrixPtr dnnWgtT = Matrix::create(
+      dnnWgt->getData(), dnnWgt->getWidth(), dnnWgt->getHeight(), false, false);
+  dnnWgtT->copyFrom(*paddleWgt);
+}
+
+void MKLDNNFcLayer::reshape() {
+  const Argument& input = getInput(0);
+  int batchSize = input.getBatchSize();
+  if (bs_ == batchSize) {
+    return;
+  }
+  bs_ = batchSize;
+  ih_ = input.getFrameHeight();
+  iw_ = input.getFrameWidth();
+  if (ih_ == 0) {
+    ih_ = 1;
+  }
+  if (iw_ == 0) {
+    iw_ = 1;
+  }
+  hasSpatial_ = true;
+  if (ih_ == 1 && iw_ == 1) {
+    hasSpatial_ = false;
+  }
+  CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
+  ic_ = iLayerSize_ / (ih_ * iw_);
+  CHECK_EQ(size_t(ic_ * ih_ * iw_), iLayerSize_) << "not divisible";
+  CHECK_EQ(size_t(oc_), getSize());
+  printSizeInfo();
+
+  // reset output
+  output_.setFrameHeight(oh_);
+  output_.setFrameWidth(ow_);
+  resetOutput(bs_, oc_);
+
+  // reset mkldnn forward
+  resetFwd();
+  needResetBwd_ = true;
+
+  convertWeightsFromPaddle();
+}
+
+void MKLDNNFcLayer::resetFwd() {
+  bool hasBias = biases_ && biases_->getW();
+  real* iData = getInputValue(0)->getData();
+  real* oData = getOutputValue()->getData();
+  real* wData = weight_->getW()->getData();
+  real* bData = hasBias ? biases_->getW()->getData() : NULL;
+
+  // TODO(TJ): below create should be covered in MkldnnMatrix
+  // create memory desc
+  memory::desc iMD = hasSpatial_ ? createMD({bs_, ic_, ih_, iw_}, format::nchw)
+                                 : createMD({bs_, ic_}, format::nc);
+  memory::desc wMD = hasSpatial_ ? createMD({oc_, ic_, ih_, iw_}, format::oihw)
+                                 : createMD({oc_, ic_}, format::oi);
+  memory::desc bMD = bData != NULL ? createMD({oc_}, format::x)
+                                   : createMD({}, format::format_undef);
+  memory::desc oMD = createMD({bs_, oc_}, format::nc);
+
+  // create memory primitive desc and memory self
+  inVal_.reset(new memory(memory::primitive_desc(iMD, engine_), iData));
+  wgtVal_.reset(new memory(memory::primitive_desc(wMD, engine_), wData));
+  outVal_.reset(new memory(memory::primitive_desc(oMD, engine_), oData));
+
+  prop_kind pk = prop_kind::forward;
+  fc_fwd::desc fwdDesc = bData != NULL ? fc_fwd::desc(pk, iMD, wMD, bMD, oMD)
+                                       : fc_fwd::desc(pk, iMD, wMD, oMD);
+  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
+
+  if (bData != NULL) {
+    biasVal_.reset(new memory(memory::primitive_desc(bMD, engine_), bData));
+    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *biasVal_, *outVal_));
+  } else {
+    fwd_.reset(new fc_fwd(fwdPD, *inVal_, *wgtVal_, *outVal_));
+  }
+  pipelineFwd_.clear();
+  pipelineFwd_.push_back(*fwd_);
+}
+
+void MKLDNNFcLayer::resetBwd() {
+  if (!needResetBwd_) {
+    return;
+  }
+  needResetBwd_ = false;
+
+  bool hasBias = biases_ && biases_->getWGrad();
+  real* iData = getInputValue(0)->getData();
+  real* iDiff = getInputGrad(0) != nullptr ? getInputGrad(0)->getData() : NULL;
+  real* oDiff = getOutputGrad()->getData();
+  real* wDiff = weight_->getWGrad()->getData();
+  real* bDiff = hasBias ? biases_->getWGrad()->getData() : NULL;
+
+  /// backward weight
+  // create memory desc for backward memory
+  memory::desc iMD = hasSpatial_ ? createMD({bs_, ic_, ih_, iw_}, format::nchw)
+                                 : createMD({bs_, ic_}, format::nc);
+  memory::desc wMD = hasSpatial_ ? createMD({oc_, ic_, ih_, iw_}, format::oihw)
+                                 : createMD({oc_, ic_}, format::oi);
+  memory::desc oMD = createMD({bs_, oc_}, format::nc);
+  memory::desc bMD = bDiff != NULL ? createMD({oc_}, format::x)
+                                   : createMD({}, format::format_undef);
+
+  if (inVal_) {
+    // update data
+    inVal_->set_data_handle(iData);
+  } else {
+    inVal_.reset(new memory(memory::primitive_desc(iMD, engine_), iData));
+  }
+
+  // create memory primitive desc and memory self
+  wgtGrad_.reset(new memory(memory::primitive_desc(wMD, engine_), wDiff));
+  outGrad_.reset(new memory(memory::primitive_desc(oMD, engine_), oDiff));
+
+  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward, iMD, wMD, oMD);
+  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
+  fc_bwdWgt::desc bwdWgtDesc = bDiff != NULL
+                                   ? fc_bwdWgt::desc(iMD, wMD, bMD, oMD)
+                                   : fc_bwdWgt::desc(iMD, wMD, oMD);
+  fc_bwdWgt::primitive_desc bwdWgtPD =
+      fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
+
+  if (bDiff != NULL) {
+    biasGrad_.reset(new memory(memory::primitive_desc(bMD, engine_), bDiff));
+    bwdWgt_.reset(
+        new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_, *biasGrad_));
+  } else {
+    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *outGrad_, *wgtGrad_));
+  }
+  pipelineBwd_.clear();
+  pipelineBwd_.push_back(*bwdWgt_);
+
+  /// backward data
+  if (iDiff == NULL) {
+    return;
+  }
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(iMD, wMD, oMD);
+  fc_bwdData::primitive_desc bwdDataPD =
+      fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
+  inGrad_.reset(new memory(memory::primitive_desc(iMD, engine_), iDiff));
+  CHECK(wgtVal_) << "Should have weight memory";
+  bwdData_.reset(new fc_bwdData(bwdDataPD, *outGrad_, *wgtVal_, *inGrad_));
+  pipelineBwd_.push_back(*bwdData_);
+}
+
+void MKLDNNFcLayer::forward(PassType passType) {
+  Layer::forward(passType);
+  reshape();
+
+  {
+    REGISTER_TIMER_INFO("mkldnn_FwdTimer", getName().c_str());
+
+    // update input data
+    // since it might be changed if this is after data layer
+    real* iData = getInputValue(0)->getData();
+    inVal_->set_data_handle(iData);
+
+    // just submit forward pipeline
+    stream_->submit(pipelineFwd_);
+  }
+
+  /* activation */ {
+    REGISTER_TIMER_INFO("FwActTimer", getName().c_str());
+    forwardActivation();
+  }
+}
+
+void MKLDNNFcLayer::backward(const UpdateCallback& callback) {
+  /* Do derivation */ {
+    REGISTER_TIMER_INFO("BpActTimer", getName().c_str());
+    backwardActivation();
+  }
+
+  {
+    REGISTER_TIMER_INFO("mkldnn_bwdTimer", getName().c_str());
+    resetBwd();
+
+    // update diff
+    real* oDiff = getOutputGrad()->getData();
+    outGrad_->set_data_handle(oDiff);
+
+    // just sumbmit backward pipeline
+    stream_->submit(pipelineBwd_);
+  }
+
+  {
+    REGISTER_TIMER_INFO("WeightUpdate", getName().c_str());
+    weight_->getParameterPtr()->incUpdate(callback);
+    if (biases_ && biases_->getWGrad()) {
+      biases_->getParameterPtr()->incUpdate(callback);
+    }
+  }
+}
+}  // namespace paddle

paddle/gserver/layers/MKLDNNFcLayer.h

+/* Copyright (c) 2017 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 "MKLDNNLayer.h"
+#include "mkldnn.hpp"
+
+namespace paddle {
+
+/**
+ * @brief A subclass of MKLDNNLayer fc layer.
+ *
+ * The config file api is mkldnn_fc
+ */
+class MKLDNNFcLayer : public MKLDNNLayer {
+protected:
+  // input layer size, can not be change after init
+  size_t iLayerSize_;  // == ic * ih * iw
+
+  // if has already init the weight
+  bool hasInitedWgt_;
+
+  // if input layer has image size info (ih>1 && iw>1)
+  bool hasSpatial_;
+
+  // fc weight and bias
+  std::unique_ptr<Weight> weight_;
+  std::unique_ptr<Weight> biases_;
+
+public:
+  explicit MKLDNNFcLayer(const LayerConfig& config)
+      : MKLDNNLayer(config), hasInitedWgt_(false), hasSpatial_(true) {}
+
+  ~MKLDNNFcLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void convertWeightsFromPaddle() override;
+
+  void convertWeightsToPaddle() override;
+
+  void forward(PassType passType) override;
+
+  void backward(const UpdateCallback& callback) override;
+
+protected:
+  /**
+   * reshape the input image sizes
+   * and reset output buffer size
+   * and reset mkldnn forward
+   */
+  void reshape();
+
+  /**
+   * reset the forward primitve and memory
+   * only would be called when input size changes
+   */
+  void resetFwd();
+
+  /**
+   * reset the backward primitve and memory for mkldnn fc
+   * only would be called when needed
+   */
+  void resetBwd();
+};
+
+}  // namespace paddle

paddle/gserver/layers/MKLDNNLayer.h

+/* Copyright (c) 2017 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 <vector>
+#include "Layer.h"
+#include "MKLDNNBase.h"
+#include "mkldnn.hpp"
+
+DECLARE_bool(use_mkldnn);
+DECLARE_bool(use_mkldnn_wgt);
+
+namespace paddle {
+
+class MKLDNNLayer;
+typedef std::shared_ptr<MKLDNNLayer> MKLDNNLayerPtr;
+
+/**
+ * @brief Base class of MKLDNNlayer.
+ *
+ */
+class MKLDNNLayer : public Layer {
+protected:
+  // batch size
+  int bs_;
+  // input image channel, height and width
+  int ic_, ih_, iw_;
+  // output image channel, height and width
+  int oc_, oh_, ow_;
+
+  // backward also need reset after reset forward handle
+  bool needResetBwd_;
+
+  // mkldnn engine, stream and primivtives
+  mkldnn::engine engine_;
+  std::shared_ptr<MKLDNNStream> stream_;
+  std::shared_ptr<mkldnn::primitive> fwd_;
+  std::shared_ptr<mkldnn::primitive> bwdWgt_;
+  std::shared_ptr<mkldnn::primitive> bwdData_;
+  std::vector<mkldnn::primitive> pipelineFwd_;
+  std::vector<mkldnn::primitive> pipelineBwd_;
+
+  // TODO(TJ): change below memory as MKLDNNMatrixPtr type
+  std::shared_ptr<mkldnn::memory> inVal_;
+  std::shared_ptr<mkldnn::memory> inGrad_;
+  std::shared_ptr<mkldnn::memory> outVal_;
+  std::shared_ptr<mkldnn::memory> outGrad_;
+  std::shared_ptr<mkldnn::memory> wgtVal_;
+  std::shared_ptr<mkldnn::memory> wgtGrad_;
+  std::shared_ptr<mkldnn::memory> biasVal_;
+  std::shared_ptr<mkldnn::memory> biasGrad_;
+
+public:
+  explicit MKLDNNLayer(const LayerConfig& config)
+      : Layer(config),
+        bs_(0),
+        ic_(0),
+        ih_(0),
+        iw_(0),
+        oc_(0),
+        oh_(0),
+        ow_(0),
+        needResetBwd_(true),
+        engine_(mkldnn::engine::cpu, 0),
+        stream_(nullptr),
+        fwd_(nullptr),
+        bwdWgt_(nullptr),
+        bwdData_(nullptr) {}
+
+  ~MKLDNNLayer() {}
+
+  virtual bool init(const LayerMap& layerMap,
+                    const ParameterMap& parameterMap) {
+    if (!Layer::init(layerMap, parameterMap)) {
+      return false;
+    }
+
+    CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
+                            << "Please set WITH_MKLDNN=ON "
+                            << "and set use_mkldnn=True";
+    stream_.reset(new MKLDNNStream());
+    engine_ = CPUEngine::Instance().getEngine();
+
+    // TODO(TJ): deivecId
+    return true;
+  }
+
+  /**
+   * convert weight from paddle format to mkldnn format
+   * weight_ will be override
+   */
+  virtual void convertWeightsFromPaddle() {}
+
+  /**
+   * convert mkldnn weight to paddle format
+   * weight_ will be override
+   */
+  virtual void convertWeightsToPaddle() {}
+
+  /**
+   * print info about sizes
+   */
+  virtual void printSizeInfo() {
+    VLOG(MKLDNN_SIZES) << getName() << ": bs: " << bs_ << ", ic: " << ic_
+                       << ", ih: " << ih_ << ", iw: " << iw_ << ", oc: " << oc_
+                       << ", oh: " << oh_ << ", ow: " << ow_;
+  }
+
+  // TODO(TJ): move to MkldnnMatrix
+  // create memory desc
+  inline mkldnn::memory::desc createMD(
+      mkldnn::memory::dims dims,
+      mkldnn::memory::format fmt,
+      mkldnn::memory::data_type type = mkldnn::memory::data_type::f32) {
+    // TODO(TJ): isFmtSuppoted(fmt)
+    return mkldnn::memory::desc(dims, type, fmt);
+  }
+};
+
+}  // namespace paddle

paddle/gserver/tests/CMakeLists.txt

@@ -9,7 +9,7 @@ add_unittest_without_exec(test_ProtoDataProvider
 # mkdir will get error.
 add_test(NAME test_ProtoDataProvider
     COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoDataProvider
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
 
 ################# test_LayerGrad #######################
 add_unittest_without_exec(test_LayerGrad
@@ -18,6 +18,15 @@ add_unittest_without_exec(test_LayerGrad
 add_test(NAME test_LayerGrad
     COMMAND test_LayerGrad)
 
+########## test_Mkldnn layers and activations ##########
+if(WITH_MKLDNN)
+    add_unittest_without_exec(test_MKLDNN
+        test_MKLDNN.cpp
+        MKLDNNTester.cpp
+        LayerGradUtil.cpp)
+    add_test(NAME test_MKLDNN COMMAND test_MKLDNN)
+endif()
+
 ################ test_CRFLayerGrad ####################
 add_unittest_without_exec(test_CRFLayerGrad
     test_CRFLayerGrad.cpp
@@ -92,8 +101,8 @@ if(WITH_PYTHON)
         test_PyDataProvider.cpp)
 
     add_test(NAME test_PyDataProvider
-        COMMAND .set_python_path.sh -d ./gserver/tests:${PROJ_ROOT}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider
-        WORKING_DIRECTORY ${PROJ_ROOT}/paddle)
+        COMMAND .set_python_path.sh -d ./gserver/tests:${PADDLE_SOURCE_DIR}/python/ ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider
+        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
 endif()
 
 ############### test_RecurrentLayer #######################
@@ -106,7 +115,7 @@ if(NOT WITH_DOUBLE)
 
     add_test(NAME test_WarpCTCLayer
         COMMAND ${CMAKE_CURRENT_BINARY_DIR}/test_WarpCTCLayer --warpctc_dir=${WARPCTC_LIB_DIR}
-        WORKING_DIRECTORY ${PROJ_ROOT}/paddle)
+        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
 endif()
 
 ############### test_RecurrentGradientMachine ###############
@@ -116,20 +125,20 @@ add_unittest_without_exec(test_RecurrentGradientMachine
     test_RecurrentGradientMachine.cpp)
 add_test(NAME test_RecurrentGradientMachine
     COMMAND .set_python_path.sh -d
-            ${PROJ_ROOT}/python:${PROJ_ROOT}/paddle/gserver/tests
+            ${PADDLE_SOURCE_DIR}/python:${PADDLE_SOURCE_DIR}/paddle/gserver/tests
             ${CMAKE_CURRENT_BINARY_DIR}/test_RecurrentGradientMachine
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
 
 add_unittest_without_exec(test_NetworkCompare
     test_NetworkCompare.cpp)
 if(WITH_GPU)
     add_test(NAME test_NetworkCompare
-        COMMAND .set_python_path.sh -d ${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true
-        WORKING_DIRECTORY ${PROJ_ROOT}/paddle)
+        COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=true
+        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
 else()
     add_test(NAME test_NetworkCompare
-        COMMAND .set_python_path.sh -d ${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false
-        WORKING_DIRECTORY ${PROJ_ROOT}/paddle)
+        COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_NetworkCompare --use_gpu=false
+        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle)
 endif()
 
 
@@ -137,6 +146,6 @@ add_unittest_without_exec(test_PyDataProvider2
         test_PyDataProvider2.cpp)
 
 add_test(NAME test_PyDataProvider2
-   COMMAND .set_python_path.sh -d ${PROJ_ROOT}/paddle/gserver/tests:${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider2
-        WORKING_DIRECTORY ${PROJ_ROOT}/paddle
+   COMMAND .set_python_path.sh -d ${PADDLE_SOURCE_DIR}/paddle/gserver/tests:${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProvider2
+        WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle
 )

paddle/gserver/tests/LayerGradUtil.cpp

@@ -388,14 +388,23 @@ void initDataLayer(TestConfig testConf,
         data.grad->zeroMem();
         break;
       case INPUT_SELF_DEFINE_DATA: {
-        size_t height = testConf.inputDefs[i].selfDefinedData->getHeight();
-        size_t width = testConf.inputDefs[i].selfDefinedData->getWidth();
-        CHECK_GT(static_cast<int>(height), 0);
-        CHECK_GT(static_cast<int>(width), 0);
-        data.value = Matrix::create(height, width, false, useGpu);
-        data.grad = Matrix::create(height, width, false, useGpu);
-        data.value->copyFrom(*testConf.inputDefs[i].selfDefinedData);
-        data.grad->zeroMem();
+        if (testConf.inputDefs[i].ids.size()) {
+          data.ids = IVector::create(testConf.inputDefs[i].ids.size(), useGpu);
+          data.ids->copyFrom(testConf.inputDefs[i].ids.data(),
+                             testConf.inputDefs[i].ids.size());
+        } else if (testConf.inputDefs[i].selfDefinedData) {
+          size_t height = testConf.inputDefs[i].selfDefinedData->getHeight();
+          size_t width = testConf.inputDefs[i].selfDefinedData->getWidth();
+          CHECK_GT(static_cast<int>(height), 0);
+          CHECK_GT(static_cast<int>(width), 0);
+          data.value = Matrix::create(height, width, false, useGpu);
+          data.grad = Matrix::create(height, width, false, useGpu);
+          data.value->copyFrom(*testConf.inputDefs[i].selfDefinedData);
+          data.grad->zeroMem();
+        } else {
+          LOG(FATAL) << "No self-defined data are given.";
+          return;
+        }
 
         const std::vector<int>& labelSeqStartPositions =
             testConf.inputDefs[i].labelSeqStartPositions;

paddle/gserver/tests/LayerGradUtil.h

@@ -68,6 +68,7 @@ struct InputDef {
   std::vector<int> labelInitValue;
   std::vector<int> labelSeqStartPositions;
   std::vector<int> labelSubSeqStartPositions;
+  std::vector<int> ids;
   MatrixPtr selfDefinedData;
 
   InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) {
@@ -95,6 +96,23 @@ struct InputDef {
     isStatic = false;
   }
 
+  InputDef(InputType type,
+           string nameIn,
+           const std::vector<int>& ids,
+           const std::vector<int>& selfDefinedSeqStartPos = {},
+           const std::vector<int>& selfDefinedSubSeqStartPos = {})
+      : labelSeqStartPositions(selfDefinedSeqStartPos),
+        labelSubSeqStartPositions(selfDefinedSubSeqStartPos),
+        ids(ids) {
+    selfDefinedData = nullptr;
+    inputType = type;
+    name = nameIn;
+    dim = 0;
+    sparse = {""};
+    paraSize = 0;
+    isStatic = false;
+  }
+
   InputDef(InputType type,
            string nameIn,
            size_t dimIn,

paddle/gserver/tests/MKLDNNTester.cpp

+/* Copyright (c) 2017 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 "MKLDNNTester.h"
+#include "paddle/gserver/layers/MKLDNNBase.h"
+#include "paddle/gserver/layers/MKLDNNLayer.h"
+
+namespace paddle {
+
+// init data layer and test layer of both dnn and reference
+void MKLDNNTester::reset(const TestConfig& dnn,
+                         const TestConfig& ref,
+                         size_t batchSize) {
+  const bool trans = false;
+  const bool useGpu = false;
+
+  // clear
+  configs_.clear();
+  layerNames_.clear();
+  dataLayers_.clear();
+  datas_.clear();
+  layerMaps_.clear();
+  parameters_.clear();
+  testLayers_.clear();
+
+  // resize
+  configs_.resize(NUM);
+  layerNames_.resize(NUM);
+  dataLayers_.resize(NUM);
+  datas_.resize(NUM);
+  layerMaps_.resize(NUM);
+  parameters_.resize(NUM);
+  testLayers_.resize(NUM);
+
+  // reset configs and layer names
+  configs_[DNN] = dnn;
+  configs_[REF] = ref;
+  layerNames_[DNN] = "mkldnn";     // the first is mkldnn layer
+  layerNames_[REF] = "reference";  // second is reference layer
+
+  // reset others
+  for (size_t i = 0; i < NUM; ++i) {
+    configs_[i].layerConfig.set_name(layerNames_[i]);
+    initDataLayer(configs_[i],
+                  &(dataLayers_[i]),
+                  &(datas_[i]),
+                  &(layerMaps_[i]),
+                  layerNames_[i],
+                  batchSize,
+                  trans,
+                  useGpu);
+    initTestLayer(
+        configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i]));
+  }
+  dnnLayer_ = testLayers_[DNN];
+  refLayer_ = testLayers_[REF];
+  EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size());
+  EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
+
+  setInputImgSize();
+}
+
+void MKLDNNTester::setInputImgSize() {
+  for (size_t n = 0; n < dataLayers_.size(); ++n) {
+    for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
+      // TODO(TJ): fix me when concat and elewise ready
+      dataLayers_[n][i]->getOutput().setFrameHeight(ih_);
+      dataLayers_[n][i]->getOutput().setFrameWidth(iw_);
+    }
+  }
+}
+
+// init randome parameters of ref, and copy to mkldnn
+void MKLDNNTester::randomWgtDatas() {
+  EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
+  for (size_t i = 0; i < parameters_[REF].size(); ++i) {
+    const VectorPtr& dnnValue = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
+    const VectorPtr& refValue = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
+    parameters_[REF][i]->randomize();
+    dnnValue->copyFrom(*refValue);
+
+    VLOG(lvl_) << "Random weight data " << parameters_[DNN][i]->getName();
+    printVector(dnnValue);
+  }
+}
+
+// random botdata of ref layer and copy same to mkldnn
+void MKLDNNTester::randomBotDatas() {
+  CHECK_EQ(dataLayers_.size(), NUM);
+  for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) {
+    dataLayers_[REF][i]->getOutputValue()->randomizeUniform();
+    dataLayers_[DNN][i]->getOutputValue()->copyFrom(
+        *(dataLayers_[REF][i]->getOutputValue()));
+    VLOG(lvl_) << "Input " << i << " data:";
+    printMatrix(dataLayers_[REF][i]->getOutputValue());
+  }
+}
+
+void MKLDNNTester::randomTopDiffs() {
+  refLayer_->getOutputGrad()->randomizeUniform();
+  dnnLayer_->getOutputGrad()->copyFrom(*(refLayer_->getOutputGrad()));
+  VLOG(lvl_) << "Random dom Backward Input, TopDiff: ";
+  printMatrix(refLayer_->getOutputGrad());
+}
+
+void MKLDNNTester::checkForward() {
+  printTopDatas();
+  double delta = compareMatrix(testLayers_[DNN]->getOutputValue(),
+                               testLayers_[REF]->getOutputValue());
+  VLOG(MKLDNN_ALL) << "Check Forward";
+  EXPECT_LE(fabs(delta), eps_);
+}
+
+void MKLDNNTester::checkBackwardData() {
+  // TODO(TJ): uncomment me when batch norm ready
+  // const bool isBN = dnnLayer_->getType() == "mkldnn_batch_norm";
+  for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) {
+    const MatrixPtr& dnnDiff = dataLayers_[DNN][i]->getOutputGrad();
+    const MatrixPtr& refDiff = dataLayers_[REF][i]->getOutputGrad();
+    VLOG(lvl_) << "Mkldnn Backward Output BotDiff " << i;
+    printMatrix(dnnDiff);
+    VLOG(lvl_) << "Reference Backward Output BotDiff " << i;
+    printMatrix(refDiff);
+
+    double delta = compareMatrix(dnnDiff, refDiff);
+    EXPECT_LE(fabs(delta), eps_);
+    // TODO(TJ): uncomment me when batch norm ready
+    // if (isBN) {
+    //  // the other two inputs in batch norm are for moving mean and var
+    //  break;
+    // }
+  }
+}
+
+void MKLDNNTester::checkBackwardWgts() {
+  CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
+  vector<VectorPtr> dnnWgts;  // used to temply save mkldnn weights
+  saveWgt(parameters_[DNN], dnnWgts);
+
+  const MKLDNNLayerPtr dnnlayer =
+      std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
+  CHECK(dnnlayer);
+  dnnlayer->convertWeightsToPaddle();
+  for (size_t i = 0; i < parameters_[DNN].size(); ++i) {
+    const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
+    const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
+    VLOG(lvl_) << "Mkldnn Output weight " << parameters_[DNN][i]->getName();
+    printVector(dnn);
+    VLOG(lvl_) << "Reference Output weight " << parameters_[REF][i]->getName();
+    printVector(ref);
+
+    double delta = compareVector(dnn, ref);
+    EXPECT_LE(fabs(delta), eps_);
+  }
+
+  VLOG(MKLDNN_ALL) << "Restore dnn weights before comapre";
+  restoreWgt(dnnWgts, parameters_[DNN]);
+}
+
+void MKLDNNTester::saveWgt(const vector<ParameterPtr>& from,
+                           vector<VectorPtr>& to) {
+  const bool useGpu = false;
+  to.resize(from.size());
+  for (size_t i = 0; i < to.size(); ++i) {
+    const VectorPtr& wgt = from[i]->getBuf(PARAMETER_VALUE);
+    to[i] = Vector::create(wgt->getSize(), useGpu);
+    to[i]->copyFrom(*wgt);
+  }
+}
+
+void MKLDNNTester::restoreWgt(const vector<VectorPtr>& from,
+                              vector<ParameterPtr>& to) {
+  CHECK_EQ(from.size(), to.size());
+  for (size_t i = 0; i < from.size(); ++i) {
+    const VectorPtr& wgt = to[i]->getBuf(PARAMETER_VALUE);
+    wgt->copyFrom(*from[i]);
+  }
+}
+
+// clear parameters grad
+void MKLDNNTester::clearWgtDiffs() {
+  for (size_t n = 0; n < parameters_.size(); ++n) {
+    for (size_t i = 0; i < parameters_[n].size(); ++i) {
+      const VectorPtr& grad = parameters_[n][i]->getBuf(PARAMETER_GRADIENT);
+      if (grad) {
+        grad->zeroMem();
+      }
+    }
+  }
+}
+
+void MKLDNNTester::clearBotDiffs() {
+  // dnn and ref
+  for (size_t n = 0; n < dataLayers_.size(); ++n) {
+    // all inputs layers
+    for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
+      dataLayers_[n][i]->getOutputGrad()->zeroMem();
+    }
+  }
+}
+
+void MKLDNNTester::clearBotDiffs(int n) {
+  CHECK_LT(n, NUM);
+  // all inputs layers
+  for (size_t i = 0; i < dataLayers_[n].size(); ++i) {
+    dataLayers_[n][i]->getOutputGrad()->zeroMem();
+  }
+}
+
+void MKLDNNTester::clearTopDatas() {
+  for (size_t i = 0; i < testLayers_.size(); ++i) {
+    testLayers_[i]->getOutputValue()->zeroMem();
+  }
+}
+
+void MKLDNNTester::printTopDatas() {
+  if (!log_) {
+    return;
+  }
+
+  for (int n = 0; n < NUM; ++n) {
+    VLOG(lvl_) << testLayers_[n]->getType() << " forward output TopData: ";
+    printMatrix(testLayers_[n]->getOutputValue());
+  }
+}
+
+void MKLDNNTester::printMatrix(const MatrixPtr& m) {
+  if (!log_) {
+    return;
+  }
+
+  std::ostringstream ostr;
+  m->print(ostr);
+  VLOG(lvl_) << std::endl << ostr.str();
+}
+
+void MKLDNNTester::printVector(const VectorPtr& v) {
+  if (!log_) {
+    return;
+  }
+
+  std::ostringstream ostr;
+  v->print(ostr, v->getSize());
+  VLOG(lvl_) << std::endl << ostr.str();
+}
+
+double MKLDNNTester::getDelta(const real* d1,
+                              const real* d2,
+                              size_t len,
+                              const float failRate,
+                              const float thres) {
+  double delta = 0, sum = 0;
+  int failCnt = 0;
+  const double eps = 1e-5;
+  double maxOut = 0;
+  for (size_t i = 0; i < len; ++i) {
+    double ref = fabs(d2[i]);
+    double diff = fabs(d1[i] - d2[i]);
+    delta += diff;
+    sum += ref;
+    if (ref > eps && fabs(d1[i]) > eps && diff / ref > thres) {
+      maxOut = std::max(maxOut, diff / ref);
+      failCnt++;
+    }
+  }
+  EXPECT_TRUE(std::isnormal(sum));
+  EXPECT_FALSE(std::isinf(sum));
+  EXPECT_FALSE(std::isnan(delta));
+  VLOG(MKLDNN_ALL) << "reference avg data: " << sum / len
+                   << ", delta: " << delta / sum << ", failCnt:" << failCnt;
+  return (failCnt / (float)len) > failRate ? maxOut : delta / sum;
+}
+
+double MKLDNNTester::compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2) {
+  CHECK_EQ(m1->getElementCnt(), m2->getElementCnt());
+  return getDelta(m1->getData(), m2->getData(), m1->getElementCnt());
+}
+
+double MKLDNNTester::compareVector(const VectorPtr& v1, const VectorPtr& v2) {
+  CHECK_EQ(v1->getSize(), v2->getSize());
+  return getDelta(v1->getData(), v2->getData(), v1->getSize());
+}
+
+void MKLDNNTester::runOnce() {
+  // test forward
+  randomBotDatas();
+  dnnLayer_->forward(PASS_TRAIN);
+  refLayer_->forward(PASS_TRAIN);
+  checkForward();
+
+  // test backward
+  randomTopDiffs();
+  dnnLayer_->backward(nullptr);
+  refLayer_->backward(nullptr);
+  checkBackwardData();
+  checkBackwardWgts();
+
+  // clear buffers
+  // ref code will addto the diff, dnn code will writeto it
+  // and clearTopDatas() and clearWgtDiffs() should be coverd by test layers
+  clearBotDiffs(REF);
+}
+
+void MKLDNNTester::run(const TestConfig& dnn,
+                       const TestConfig& ref,
+                       size_t batchSize,
+                       size_t inputImgH,
+                       size_t inputImgW,
+                       size_t iter,
+                       float epsilon,
+                       bool log,
+                       int level) {
+  VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " << dnn.layerConfig.type()
+                     << " vs " << ref.layerConfig.type();
+  ih_ = inputImgH;
+  iw_ = inputImgW;
+  iter_ = iter;
+  eps_ = epsilon;
+  log_ = log;
+  lvl_ = level;
+
+  // Firstly test FLAGS_use_mkldnn_wgt = false
+  FLAGS_use_mkldnn_wgt = false;
+  // reset and run once
+  reset(dnn, ref, batchSize);
+  randomWgtDatas();
+  clearWgtDiffs();
+  clearBotDiffs();
+  for (size_t i = 0; i < iter_; ++i) {
+    VLOG(MKLDNN_TESTS) << "Check Iteration " << i;
+    runOnce();
+  }
+
+  // Then test FLAGS_use_mkldnn_wgt = true
+  FLAGS_use_mkldnn_wgt = true;
+  // after run once the mkldnn weight has been stored in dnnlayer
+  // then save the weights and restart again
+  vector<VectorPtr> dnnWgts, refWgts;
+  CHECK_EQ(parameters_[DNN].size(), parameters_[REF].size());
+  saveWgt(parameters_[DNN], dnnWgts);
+  saveWgt(parameters_[REF], refWgts);
+
+  // restart again with flag true
+  reset(dnn, ref, batchSize);
+
+  // restore wgt
+  restoreWgt(dnnWgts, parameters_[DNN]);
+  restoreWgt(refWgts, parameters_[REF]);
+  clearWgtDiffs();
+  clearBotDiffs();
+
+  for (size_t i = 0; i < iter_; ++i) {
+    VLOG(MKLDNN_TESTS) << "Check Iteration " << i;
+    runOnce();
+  }
+}
+
+}  //  namespace paddle

paddle/gserver/tests/MKLDNNTester.h

+/* Copyright (c) 2017 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 <string>
+#include <vector>
+#include "LayerGradUtil.h"
+#include "paddle/gserver/layers/MKLDNNBase.h"
+
+namespace paddle {
+
+/**
+ * @brief test the functionality of Mkldnnlayers
+ * refer to paddle original function
+ */
+class MKLDNNTester {
+  enum {
+    DNN = 0,  // MKLDNN layer
+    REF = 1,  // Reference layer
+    NUM = 2,  // Number of total
+  };
+
+protected:
+  std::vector<TestConfig> configs_;
+  vector<string> layerNames_;
+  vector<vector<DataLayerPtr>> dataLayers_;
+  vector<vector<Argument>> datas_;
+  vector<LayerMap> layerMaps_;
+  vector<vector<ParameterPtr>> parameters_;
+  vector<LayerPtr> testLayers_;
+  LayerPtr dnnLayer_, refLayer_;
+
+  /// run some iterations, all the result should pass
+  size_t iter_;
+  /// whether to print out the details
+  bool log_;
+  /// vlog level to print the matrix details datas
+  int lvl_;
+  /// epsilon
+  float eps_;
+  /// input image size, default 1
+  size_t ih_, iw_;
+
+public:
+  explicit MKLDNNTester(size_t iter = 3, float epsilon = 1e-4) {
+    iter_ = iter;
+    eps_ = epsilon;
+    log_ = false;
+    lvl_ = MKLDNN_ALL;
+  }
+
+  ~MKLDNNTester() {}
+
+public:
+  void run(const TestConfig& dnn,
+           const TestConfig& ref,
+           size_t batchSize,
+           size_t inputImgH = 1,
+           size_t inputImgW = 1,
+           size_t iter = 3,
+           float epsilon = 1e-4,
+           bool log = false,
+           int level = MKLDNN_ALL);
+  void setLogLevel(int lvl) { lvl_ = lvl; }
+
+private:
+  void reset(const TestConfig& dnn, const TestConfig& ref, size_t batchSize);
+  void setInputImgSize();
+  void runOnce();
+
+  void randomWgtDatas();
+  void randomBotDatas();
+  void randomTopDiffs();
+
+  void checkForward();
+  void checkBackwardData();
+  void checkBackwardWgts();
+
+  void clearWgtDiffs();
+  void clearBotDiffs();
+  void clearBotDiffs(int n);  // clear specific layer
+  void clearTopDatas();
+
+  void printTopDatas();
+  void printMatrix(const MatrixPtr& m);
+  void printVector(const VectorPtr& v);
+
+  void saveWgt(const vector<ParameterPtr>& from, vector<VectorPtr>& to);
+  void restoreWgt(const vector<VectorPtr>& from, vector<ParameterPtr>& to);
+
+  double compareMatrix(const MatrixPtr& m1, const MatrixPtr& m2);
+  double compareVector(const VectorPtr& v1, const VectorPtr& v2);
+
+  /**
+   * Get delta percent
+   * if many(>failRate) wrong(abs(dnn-ref)/abs(ref)>thres) points return the
+   * max(diff/ref)
+   * else return sum(abs(a-b)) / sum(abs(b))
+   * The return value should smaller than eps when passing.
+   */
+  double getDelta(const real* d1,
+                  const real* d2,
+                  size_t len,
+                  const float failRate = 1e-3,
+                  const float thres = 0.1);
+};
+
+}  //  namespace paddle

paddle/gserver/tests/test_KmaxSeqScore.cpp

@@ -96,6 +96,11 @@ TEST(Layer, kmaxSeqScoreLayer) {
   MatrixPtr inValue =
       Matrix::create(subSeqStartPosition.back(), 1, false, false);
 
+  std::vector<bool> mode = {false};
+#ifndef PADDLE_ONLY_CPU
+  mode.push_back(true);
+#endif
+
   for (auto hasSubseq : {false, true}) {
     vector<vector<int>> groundTruth;
     inValue->randomizeUniform();
@@ -104,7 +109,7 @@ TEST(Layer, kmaxSeqScoreLayer) {
                          hasSubseq ? subSeqStartPosition : seqStartPosition,
                          beamSize);
 
-    for (auto useGpu : {false, true}) {
+    for (auto useGpu : mode) {
       TestConfig config;
       config.layerConfig.set_type("kmax_seq_score");
       config.layerConfig.set_beam_size(beamSize);

paddle/gserver/tests/test_MKLDNN.cpp

+/* Copyright (c) 2017 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 <string>
+#include <vector>
+#include "MKLDNNTester.h"
+#include "ModelConfig.pb.h"
+
+using namespace paddle;  // NOLINT
+
+DECLARE_bool(thread_local_rand_use_global_seed);
+DECLARE_bool(use_gpu);
+DECLARE_bool(use_mkldnn);
+
+struct testFCDesc {
+  int bs;
+  int ic;
+  int oc;
+  int ih, iw;  // oh == ow == 1
+};
+
+void testFcLayer(const testFCDesc& pm) {
+  const std::string compareTypes[] = {"mkldnn_fc", "fc"};
+  TestConfig cfg;
+  cfg.layerConfig.set_type(compareTypes[0]);
+  cfg.layerConfig.set_size(pm.oc);
+  cfg.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       /* size of input layer= */ size_t(pm.ic * pm.ih * pm.iw),
+       /* size of weight= */ size_t(pm.oc * pm.ic * pm.ih * pm.iw)});
+  cfg.layerConfig.add_inputs();
+
+  MKLDNNTester tester;
+  for (auto biasSize : {pm.oc, 0}) {
+    cfg.biasSize = biasSize;
+    TestConfig ref = cfg;
+    ref.layerConfig.set_type(compareTypes[1]);
+    for (auto bs : {pm.bs, 1}) {
+      tester.run(cfg, ref, bs, pm.ih, pm.iw);
+    }
+  }
+}
+
+TEST(MKLDNNLayer, FcLayer) {
+  testFcLayer({/*bs*/ 2, /*ic*/ 2, /*oc*/ 3, /*ih*/ 1, /*iw*/ 1});
+  testFcLayer({/*bs*/ 3, /*ic*/ 7, /*oc*/ 19, /*ih*/ 1, /*iw*/ 1});
+  testFcLayer({/*bs*/ 8, /*ic*/ 16, /*oc*/ 32, /*ih*/ 13, /*iw*/ 13});
+  testFcLayer({/*bs*/ 4, /*ic*/ 12, /*oc*/ 18, /*ih*/ 13, /*iw*/ 11});
+  testFcLayer({/*bs*/ 2, /*ic*/ 64, /*oc*/ 32, /*ih*/ 16, /*iw*/ 16});
+  testFcLayer({/*bs*/ 15, /*ic*/ 3, /*oc*/ 6, /*ih*/ 16, /*iw*/ 16});
+}
+
+// TODO(TJ): add branch test
+
+int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
+  FLAGS_use_gpu = false;
+  FLAGS_use_mkldnn = true;
+  initMain(argc, argv);
+  FLAGS_thread_local_rand_use_global_seed = true;
+  srand(1);
+  return RUN_ALL_TESTS();
+}

paddle/math/CMakeLists.txt

@@ -15,13 +15,13 @@
 file(GLOB MATH_HEADERS . *.h)
 file(GLOB MATH_SOURCES . *.cpp)
 set(MATH_SOURCES
-    "${PROJ_ROOT}/paddle/math/BaseMatrix.cu"
-    "${PROJ_ROOT}/paddle/math/TrainingAlgorithmOp.cu"
+    "${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu"
+    "${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu"
     ${MATH_SOURCES})
 if(NOT WITH_GPU)
     # then compile BaseMatrix.cu as c++ file
-    compile_cu_as_cpp("${PROJ_ROOT}/paddle/math/BaseMatrix.cu")
-    compile_cu_as_cpp("${PROJ_ROOT}/paddle/math/TrainingAlgorithmOp.cu")
+    compile_cu_as_cpp("${PADDLE_SOURCE_DIR}/paddle/math/BaseMatrix.cu")
+    compile_cu_as_cpp("${PADDLE_SOURCE_DIR}/paddle/math/TrainingAlgorithmOp.cu")
     add_library(paddle_math STATIC
         ${MATH_SOURCES})
 else()

paddle/math/CpuSparseMatrix.h

@@ -302,6 +302,10 @@ public:
   bool isSparse() const { return true; }
 
 private:
+  using Matrix::mul;
   using Matrix::copyFrom;
+  using Matrix::rowMax;
+  using Matrix::print;
+  using Matrix::subMatrix;
 };
 }  // namespace paddle

paddle/math/SparseMatrix.h

@@ -231,6 +231,9 @@ public:
 private:
   using Matrix::mul;
   using Matrix::copyFrom;
+  using Matrix::rowMax;
+  using Matrix::print;
+  using Matrix::subMatrix;
 };
 
 }  // namespace paddle

paddle/operators/CMakeLists.txt

@@ -41,31 +41,29 @@ function(op_library TARGET)
     endif()
 endfunction()
 
+add_subdirectory(math)
+cc_test(gather_test SRCS gather_test.cc DEPS tensor)
+
 cc_library(net_op SRCS net_op.cc DEPS op_registry)
 cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
 
 op_library(add_op SRCS add_op.cc add_op.cu)
-cc_test(add_op_test SRCS add_op_test.cc DEPS add_op)
 
 op_library(mean_op SRCS mean_op.cc mean_op.cu)
-cc_test(mean_op_test SRCS mean_op_test.cc DEPS mean_op)
 
-op_library(mul_op SRCS mul_op.cc mul_op.cu)
+op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
 op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
 
 op_library(sigmoid_op SRCS sigmoid_op.cc sigmoid_op.cu)
 op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
+op_library(gaussian_random_op SRCS gaussian_random_op.cc gaussian_random_op.cu)
 op_library(cross_entropy_op SRCS cross_entropy_op.cc cross_entropy_op.cu)
 op_library(fill_zeros_like_op SRCS fill_zeros_like_op.cc fill_zeros_like_op.cu)
 
 op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)
-cc_test(sgd_op_test SRCS sgd_op_test.cc DEPS sgd_op)
 
-op_library(fc_op
-    SRCS fc_op.cc
-    DEPS mul_op rowwise_add_op sigmoid_op softmax_op net_op)
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
-    DEPS op_desc tensor op_registry operator net_op)
+    DEPS framework_proto tensor op_registry operator net_op)
 cc_test(recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op)
 op_library(uniform_random_op
         SRCS uniform_random_op.cc uniform_random_op.cu)

paddle/operators/add_op.cc

@@ -18,16 +18,15 @@ namespace paddle {
 namespace operators {
 
 class AddOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.InputSize(), 2);
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1);
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "Inputs of AddOp must all be set");
-    PADDLE_ENFORCE(ctx.OutputVar(0) != nullptr,
-                   "Outputs of AddOp must all be set");
-    PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims() == ctx.Input<Tensor>(1)->dims(),
-                   "Two input of Add Op's dimension must be same.");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
+                      ctx.Input<Tensor>("Y")->dims(),
+                      "Two input of Add Op's dimension must be same.");
+    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -47,6 +46,9 @@ The equation is: Out = X + Y
 };
 
 class AddOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {}
 };

paddle/operators/add_op.h

@@ -28,9 +28,9 @@ template <typename Place, typename T>
 class AddKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto input0 = context.Input<Tensor>(0);
-    auto input1 = context.Input<Tensor>(1);
-    auto output = context.Output<Tensor>(0);
+    auto* input0 = context.Input<Tensor>("X");
+    auto* input1 = context.Input<Tensor>("Y");
+    auto* output = context.Output<Tensor>("Out");
 
     output->mutable_data<T>(context.GetPlace());
 

paddle/operators/cross_entropy_op.cc

@@ -18,26 +18,25 @@ namespace paddle {
 namespace operators {
 
 class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.InputSize(), 2,
-                      "Input size of OnehotCrossEntropyOp must be two");
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1,
-                      "Output size of OnehotCrossEntropyOp must be one");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0),
-                            "0-th input of OnehotCrossEntropyOp should be set");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(1),
-                            "1-th input of OnehotCrossEntropyOp should be set");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0),
-                            "Outputs of OnehotCrossEntropyOp must all be set");
-    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>(0)->dims().size(), 2);
-    PADDLE_ENFORCE_EQ(ctx.Output<Tensor>(0)->dims().size(), 1,
-                      "label's dimension must be 1.");
-    ctx.Output<Tensor>(0)->Resize({ctx.Input<Tensor>(0)->dims()[0]});
+    auto *X = ctx.Input<Tensor>("X");
+    auto *label = ctx.Input<Tensor>("label");
+
+    PADDLE_ENFORCE_EQ(X->dims().size(), 2, "X's dimension must be 2.");
+    PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1.");
+    PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]);
+    ctx.Output<Tensor>("Y")->Resize({X->dims()[0]});
   }
 };
 
 class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     auto X_grad = ctx.Output<Tensor>(framework::GradVarName("X"));

paddle/operators/cross_entropy_op.h

@@ -45,7 +45,7 @@ class OnehotCrossEntropyOpKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto X = ctx.Input<Tensor>("X");
     const T* Xdata = X->data<T>();
-    const int* label_data = ctx.Input<Tensor>(1)->data<int>();
+    const int* label_data = ctx.Input<Tensor>("label")->data<int>();
     auto Y = ctx.Output<Tensor>("Y");
 
     Y->mutable_data<T>(ctx.GetPlace());

paddle/operators/fill_zeros_like_op.cc

@@ -18,18 +18,13 @@ namespace paddle {
 namespace operators {
 
 class FillZerosLikeOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL,
-                      "Input size of FillZerosLikeOp must be one.");
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL,
-                      "Output size of AddOp must be one.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0),
-                            "Input of FillZerosLikeOp must be set.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0),
-                            "Output of FillZerosLikeOp must be set.");
-    ctx.Output<framework::Tensor>(0)->Resize(
-        ctx.Input<framework::Tensor>(0)->dims());
+    ctx.Output<framework::Tensor>("Dst")->Resize(
+        ctx.Input<framework::Tensor>("Src")->dims());
   }
 };
 

paddle/operators/fill_zeros_like_op.h

@@ -23,7 +23,7 @@ template <typename Place, typename T>
 class FillZerosLikeKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* output = context.Output<framework::Tensor>(0);
+    auto* output = context.Output<framework::Tensor>("Dst");
     output->mutable_data<T>(context.GetPlace());
     auto t = framework::EigenVector<T>::Flatten(*output);
     t.device(context.GetEigenDevice<Place>()) = t.constant(T(0));

paddle/operators/gather.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.h>
+#include <cstring>
+
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/place.h"
+
+namespace paddle {
+namespace operators {
+
+// Implementation of CPU copy
+template <typename T>
+void CPUGather(const T* params, const int* indices, const int slice_size,
+               const int index_size, T* output) {
+  const size_t slice_bytes = slice_size * sizeof(T);
+
+  for (size_t i = 0; i < index_size; ++i) {
+    int index_ = indices[i];
+    memcpy(output + i * slice_size, params + index_ * slice_size, slice_bytes);
+  }
+}
+
+// Implementation of GPU copy:
+template <typename T>
+void GPUGather(const T* src, const int* index, const int slice_size,
+               const int index_size, T* output);
+
+/**
+ * Return a new tensor from source tensor, gathered according to index
+ * input[src]: type-T source Tensor
+ * input[index]: type-int index Tensor (1-D)
+ * return: output tensor
+ */
+template <typename T>
+void Gather(const platform::Place& place, const paddle::framework::Tensor* src,
+            const paddle::framework::Tensor* index,
+            paddle::framework::Tensor* output) {
+  // check index of shape 1-D
+  PADDLE_ENFORCE(index->dims().size() == 1);
+  int index_size = index->dims()[0];
+
+  auto src_dims = src->dims();
+  paddle::framework::DDim output_dims(src_dims);
+  output_dims[0] = index_size;
+
+  // slice size
+  int slice_size = 1;
+  for (size_t i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
+
+  // Gathering
+  if (platform::is_cpu_place(place)) {
+    CPUGather<T>(src->data<T>(), index->data<int>(), slice_size, index_size,
+                 output->data<T>());
+  }
+}
+
+}  // namespace operators
+}  // namespace paddle

paddle/framework/op_desc_test.cc → paddle/operators/gather_test.cc

@@ -12,24 +12,37 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include "paddle/operators/gather.h"
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/place.h"
+
 #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());
-}
+#include <iostream>
+#include <string>
+
+TEST(Gather, GatherData) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  using namespace paddle::operators;
+
+  Tensor* src = new Tensor();
+  Tensor* index = new Tensor();
+  Tensor* output = new Tensor();
+
+  int* p_src = nullptr;
+  int* p_index = nullptr;
+  p_src = src->mutable_data<int>(make_ddim({3, 4}), CPUPlace());
+  p_index = index->mutable_data<int>(make_ddim({2}), CPUPlace());
+
+  for (size_t i = 0; i < 12; ++i) p_src[i] = i;
+  p_index[0] = 1;
+  p_index[1] = 0;
\ No newline at end of file
+
+  int* p_output = output->mutable_data<int>(make_ddim({2, 4}), CPUPlace());
+
+  Gather<int>(CPUPlace(), src, index, output);
+
+  for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
+  for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
+}

paddle/operators/gaussian_random_op.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <random>
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class GaussianRandomKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    float mean = context.op_.GetAttr<float>("mean");
+    float std = context.op_.GetAttr<float>("std");
+    auto* tensor = context.Output<framework::Tensor>(0);
+    T* data = tensor->mutable_data<T>(context.GetPlace());
+
+    // TODO(dzh): attribute does not support unsigned int.
+    // And we need a global random seed configuration.
+    int seed = context.op_.GetAttr<int>("seed");
+    if (seed == 0) {
+      seed = std::random_device()();
+    }
+    std::mt19937 g(seed);
+    std::normal_distribution<T> distribution(mean, std);
+    ssize_t size = framework::product(tensor->dims());
+    for (int i = 0; i < size; ++i) {
+      data[i] = distribution(g);
+    }
+  }
+};
+
+class GaussianRandomOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext& context) const override {
+    auto* tensor = context.Output<framework::Tensor>(0);
+    auto dims = GetAttr<std::vector<int>>("dims");
+    PADDLE_ENFORCE(dims.size() > 0UL,
+                   "dims can be one int or array. dims must be set.");
+    tensor->Resize(framework::make_ddim(dims));
+  }
+};
+
+class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  GaussianRandomOpMaker(framework::OpProto* proto,
+                        framework::OpAttrChecker* op_checker)
+      : framework::OpProtoAndCheckerMaker(proto, op_checker) {
+    AddOutput("Out", "output matrix of random op");
+    AddComment(R"DOC(
+GaussianRandom operator.
+Use to initialize tensor with gaussian random generator.
+)DOC");
+
+    AddAttr<std::vector<int>>("dims", "The dimension of random tensor.");
+    AddAttr<float>("mean", "mean value of random.").SetDefault(.0f);
+    AddAttr<float>("std", "minimum value of random value.").SetDefault(1.0f);
+    AddAttr<int>("seed",
+                 "Random seed of generator."
+                 "0 means use system wide seed")
+        .SetDefault(0);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(gaussian_random, ops::GaussianRandomOp, ops::GaussianRandomOpMaker);
+REGISTER_OP_CPU_KERNEL(gaussian_random, ops::GaussianRandomKernel<float>);

paddle/operators/fc_op.cc → paddle/operators/gaussian_random_op.cu

@@ -12,65 +12,42 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/operators/net_op.h"
+#include <memory>
+#include <random>
+#include "paddle/platform/dynload/curand.h"
+#include "paddle/platform/gpu_info.h"
 
-#include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace operators {
 
-using OpRegistry = framework::OpRegistry;
-
-class FullyConnectedOp : public NetOp {
+template <typename T>
+class GaussianRandomKernel : public framework::OpKernel {
  public:
-  void Init() override {
-    AddOp(OpRegistry::CreateOp("mul",
-                               {
-                                   Input("X"), Input("W"),
-                               },
-                               {Output("before_act")}, {}));
-    auto b = Input("b");
-    if (b != framework::kEmptyVarName) {
-      AddOp(OpRegistry::CreateOp("rowwise_add",
-                                 {Output("before_act"), Input("b")},
-                                 {Output("before_act")}, {}));
+  void Compute(const framework::ExecutionContext& context) const override {
+    float mean = context.op_.GetAttr<float>("mean");
+    float std = context.op_.GetAttr<float>("std");
+    auto* tensor = context.Output<framework::Tensor>(0);
+    T* data = tensor->mutable_data<T>(context.GetPlace());
+
+    int seed = context.op_.GetAttr<int>("seed");
+    if (seed == 0) {
+      std::random_device rd;
+      seed = rd();
     }
-
-    auto activation = GetAttr<std::string>("activation");
-    AddOp(OpRegistry::CreateOp(activation, {Output("before_act")},
-                               {Output("Y")}, {}));
-    CompleteAddOp(false);
+    curandGenerator_t g;
+    PADDLE_ENFORCE(platform::dynload::curandCreateGenerator(
+        &g, CURAND_RNG_PSEUDO_DEFAULT));
+    PADDLE_ENFORCE(
+        platform::dynload::curandSetPseudoRandomGeneratorSeed(g, seed));
+    platform::dynload::curandGenerateNormal(
+        g, data, framework::product(tensor->dims()), mean, std);
   }
 };
 
-class FullyConnectedOpMaker : public framework::OpProtoAndCheckerMaker {
- public:
-  FullyConnectedOpMaker(framework::OpProto *proto,
-                        framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "the input of fc operator");
-    AddInput("W", "the weight of fc operator");
-    AddInput("b", "the bias of fc operator");
-
-    AddOutput("Y", "the output of fc operator");
-    AddOutput("before_act", "the before activation output of fc operator")
-        .SetTemporary();
-    AddAttr<std::string>("activation", "The activation key for fc layer")
-        .SetDefault("sigmoid")
-        .InEnum({"sigmoid", "softmax"});
-
-    //! TODO(yuyang18): Complete comment;
-    AddComment("FullyConnected Operator");
-  }
-};
 }  // namespace operators
 }  // namespace paddle
 
-USE_OP(mul);
-USE_OP(rowwise_add);
-USE_OP(sigmoid);
-USE_OP(softmax);
-
 namespace ops = paddle::operators;
-REGISTER_OP(fc, ops::FullyConnectedOp, ops::FullyConnectedOpMaker);
+REGISTER_OP_GPU_KERNEL(gaussian_random, ops::GaussianRandomKernel<float>);

paddle/operators/math/CMakeLists.txt

+if(WITH_MKLML)
+    set(BLAS_LIB mklml)
+else()
+    set(BLAS_LIB cblas)
+endif()
+
+if(WITH_GPU)
+    nv_library(math_function SRCS math_function.cc math_function.cu DEPS ${BLAS_LIB} device_context)
+else()
+    cc_library(math_function SRCS math_function.cc DEPS ${BLAS_LIB} device_context)
+endif()
+
+nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor)

paddle/operators/math/math_function.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <>
+void gemm<platform::CPUPlace, float>(const CBLAS_TRANSPOSE transA,
+                                     const CBLAS_TRANSPOSE transB, const int M,
+                                     const int N, const int K,
+                                     const float alpha, const float* A,
+                                     const float* B, const float beta, float* C,
+                                     platform::DeviceContext* context) {
+  int lda = K;
+  int ldb = N;
+  int ldc = N;
+  cblas_sgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
+              beta, C, ldc);
+}
+
+template <>
+void gemm<platform::CPUPlace, double>(const CBLAS_TRANSPOSE transA,
+                                      const CBLAS_TRANSPOSE transB, const int M,
+                                      const int N, const int K,
+                                      const double alpha, const double* A,
+                                      const double* B, const double beta,
+                                      double* C,
+                                      platform::DeviceContext* context) {
+  int lda = K;
+  int ldb = N;
+  int ldc = N;
+  cblas_dgemm(CblasRowMajor, transA, transB, M, N, K, alpha, A, lda, B, ldb,
+              beta, C, ldc);
+}
+
+template <>
+void matmul<platform::CPUPlace, float>(const framework::Tensor& matrix_a,
+                                       bool trans_a,
+                                       const framework::Tensor& matrix_b,
+                                       bool trans_b, float alpha,
+                                       framework::Tensor* matrix_out,
+                                       float beta,
+                                       platform::DeviceContext* context) {
+  auto dim_a = matrix_a.dims();
+  auto dim_b = matrix_b.dims();
+  auto dim_out = matrix_out->dims();
+  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+                 "The input and output of matmul be matrix");
+
+  PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
+                     platform::is_cpu_place(matrix_b.place()) &&
+                     platform::is_cpu_place(matrix_out->place()),
+                 "Matrix must all be in CPUPlace");
+
+  int M = dim_out[0];
+  int N = dim_out[1];
+  int K = (trans_a == false) ? dim_a[1] : dim_a[0];
+
+  CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
+  CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
+
+  gemm<platform::CPUPlace, float>(
+      transA, transB, M, N, K, alpha, matrix_a.data<float>(),
+      matrix_b.data<float>(), beta, matrix_out->data<float>(), context);
+}
+
+template <>
+void matmul<platform::CPUPlace, double>(const framework::Tensor& matrix_a,
+                                        bool trans_a,
+                                        const framework::Tensor& matrix_b,
+                                        bool trans_b, double alpha,
+                                        framework::Tensor* matrix_out,
+                                        double beta,
+                                        platform::DeviceContext* context) {
+  auto dim_a = matrix_a.dims();
+  auto dim_b = matrix_b.dims();
+  auto dim_out = matrix_out->dims();
+  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+                 "The input and output of matmul be matrix");
+
+  PADDLE_ENFORCE(platform::is_cpu_place(matrix_a.place()) &&
+                     platform::is_cpu_place(matrix_b.place()) &&
+                     platform::is_cpu_place(matrix_out->place()),
+                 "Matrix must all be in CPUPlace");
+
+  int M = dim_out[0];
+  int N = dim_out[1];
+  int K = (trans_a == false) ? dim_a[1] : dim_a[0];
+
+  CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
+  CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
+
+  gemm<platform::CPUPlace, double>(
+      transA, transB, M, N, K, alpha, matrix_a.data<double>(),
+      matrix_b.data<double>(), beta, matrix_out->data<double>(), context);
+}
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/math_function.cu

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <>
+void gemm<platform::GPUPlace, float>(const CBLAS_TRANSPOSE transA,
+                                     const CBLAS_TRANSPOSE transB, const int M,
+                                     const int N, const int K,
+                                     const float alpha, const float* A,
+                                     const float* B, const float beta, float* C,
+                                     platform::DeviceContext* context) {
+  // Note that cublas follows fortran order, so the order is different from
+  // the cblas convention.
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
+  cublasOperation_t cuTransA =
+      (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
+  cublasOperation_t cuTransB =
+      (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
+
+  PADDLE_ENFORCE(platform::dynload::cublasSgemm(
+      reinterpret_cast<platform::CUDADeviceContext*>(context)->cublas_handle(),
+      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
+}
+
+template <>
+void gemm<platform::GPUPlace, double>(const CBLAS_TRANSPOSE transA,
+                                      const CBLAS_TRANSPOSE transB, const int M,
+                                      const int N, const int K,
+                                      const double alpha, const double* A,
+                                      const double* B, const double beta,
+                                      double* C,
+                                      platform::DeviceContext* context) {
+  // Note that cublas follows fortran order, so the order is different from
+  // the cblas convention.
+  int lda = (transA == CblasNoTrans) ? K : M;
+  int ldb = (transB == CblasNoTrans) ? N : K;
+  cublasOperation_t cuTransA =
+      (transA == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
+  cublasOperation_t cuTransB =
+      (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
+  PADDLE_ENFORCE(platform::dynload::cublasDgemm(
+      reinterpret_cast<platform::CUDADeviceContext*>(context)->cublas_handle(),
+      cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
+}
+
+template <>
+void matmul<platform::GPUPlace, float>(const framework::Tensor& matrix_a,
+                                       bool trans_a,
+                                       const framework::Tensor& matrix_b,
+                                       bool trans_b, float alpha,
+                                       framework::Tensor* matrix_out,
+                                       float beta,
+                                       platform::DeviceContext* context) {
+  auto dim_a = matrix_a.dims();
+  auto dim_b = matrix_b.dims();
+  auto dim_out = matrix_out->dims();
+  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+                 "The input and output of matmul be matrix");
+
+  PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) &&
+                     platform::is_gpu_place(matrix_b.place()) &&
+                     platform::is_gpu_place(matrix_out->place()),
+                 "Matrix must all be in GPUPlace");
+
+  int M = dim_out[0];
+  int N = dim_out[1];
+  int K = (trans_a == false) ? dim_a[1] : dim_a[0];
+
+  CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
+  CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
+
+  gemm<platform::GPUPlace, float>(
+      transA, transB, M, N, K, alpha, matrix_a.data<float>(),
+      matrix_b.data<float>(), beta, matrix_out->data<float>(), context);
+}
+
+template <>
+void matmul<platform::GPUPlace, double>(const framework::Tensor& matrix_a,
+                                        bool trans_a,
+                                        const framework::Tensor& matrix_b,
+                                        bool trans_b, double alpha,
+                                        framework::Tensor* matrix_out,
+                                        double beta,
+                                        platform::DeviceContext* context) {
+  auto dim_a = matrix_a.dims();
+  auto dim_b = matrix_b.dims();
+  auto dim_out = matrix_out->dims();
+  PADDLE_ENFORCE(dim_a.size() == 2 && dim_b.size() == 2 && dim_out.size() == 2,
+                 "The input and output of matmul be matrix");
+
+  PADDLE_ENFORCE(platform::is_gpu_place(matrix_a.place()) &&
+                     platform::is_gpu_place(matrix_b.place()) &&
+                     platform::is_gpu_place(matrix_out->place()),
+                 "Matrix must all be in GPUPlace");
+
+  int M = dim_out[0];
+  int N = dim_out[1];
+  int K = (trans_a == false) ? dim_a[1] : dim_a[0];
+
+  CBLAS_TRANSPOSE transA = (trans_a == false) ? CblasNoTrans : CblasTrans;
+  CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
+
+  gemm<platform::GPUPlace, double>(
+      transA, transB, M, N, K, alpha, matrix_a.data<double>(),
+      matrix_b.data<double>(), beta, matrix_out->data<double>(), context);
+}
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/math_function.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
+#ifdef PADDLE_USE_MKLML
+#include <mkl_cblas.h>
+#include <mkl_lapacke.h>
+#include <mkl_vml_functions.h>
+#endif
+
+#ifdef PADDLE_USE_MKL
+#include <mkl.h>
+#include <mkl_lapacke.h>
+#endif
+
+#ifdef PADDLE_USE_ATLAS
+extern "C" {
+#include <cblas.h>
+#include <clapack.h>
+}
+#endif
+
+#ifdef PADDLE_USE_OPENBLAS
+#include <cblas.h>
+#include <lapacke.h>
+#endif
+
+#ifndef LAPACK_FOUND
+extern "C" {
+#include <cblas.h>
+int LAPACKE_sgetrf(int matrix_layout, int m, int n, float* a, int lda,
+                   int* ipiv);
+int LAPACKE_dgetrf(int matrix_layout, int m, int n, double* a, int lda,
+                   int* ipiv);
+int LAPACKE_sgetri(int matrix_layout, int n, float* a, int lda,
+                   const int* ipiv);
+int LAPACKE_dgetri(int matrix_layout, int n, double* a, int lda,
+                   const int* ipiv);
+}
+#endif
+
+#include <cmath>
+
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/enforce.h"
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+// Support continuous memory now
+// If transA = N, and transB = N
+// Then matrixA: M * K, matrixB: K * N matrixC : M * N
+// For more detailed info, please refer to
+// http://www.netlib.org/lapack/explore-html/d4/de2/sgemm_8f.html
+template <typename Place, typename T>
+void gemm(const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB,
+          const int M, const int N, const int K, const T alpha, const T* A,
+          const T* B, const T beta, T* C, platform::DeviceContext* context);
+
+// matrix multiply with continuous memory
+template <typename Place, typename T>
+void matmul(const framework::Tensor& matrix_a, bool trans_a,
+            const framework::Tensor& matrix_b, bool trans_b, T alpha,
+            framework::Tensor* matrix_out, T beta,
+            platform::DeviceContext* context);
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

paddle/operators/math/math_function_test.cc

+#include "paddle/operators/math/math_function.h"
+#include "gtest/gtest.h"
+
+#ifndef PADDLE_ONLY_CPU
+TEST(math_function, notrans_mul_trans) {
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor out_gpu;
+  paddle::framework::Tensor out;
+
+  auto* cpu_place = new paddle::platform::CPUPlace();
+  float* input1_ptr = input1.mutable_data<float>({2, 3}, *cpu_place);
+  float arr[6] = {0, 1, 2, 3, 4, 5};
+  memcpy(input1_ptr, arr, 6 * sizeof(float));
+
+  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  paddle::platform::DeviceContext* context =
+      new paddle::platform::CUDADeviceContext(*gpu_place);
+
+  input1_gpu.CopyFrom<float>(input1, *gpu_place);
+  input2_gpu.CopyFrom<float>(input1, *gpu_place);
+
+  out_gpu.mutable_data<float>({2, 2}, *gpu_place);
+
+  paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
+      input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0, context);
+
+  out.CopyFrom<float>(out_gpu, *cpu_place);
+
+  float* out_ptr = out.data<float>();
+  EXPECT_EQ(out_ptr[0], 5);
+  EXPECT_EQ(out_ptr[1], 14);
+  EXPECT_EQ(out_ptr[2], 14);
+  EXPECT_EQ(out_ptr[3], 50);
+}
+
+TEST(math_function, trans_mul_notrans) {
+  paddle::framework::Tensor input1;
+  paddle::framework::Tensor input1_gpu;
+  paddle::framework::Tensor input2_gpu;
+  paddle::framework::Tensor out_gpu;
+  paddle::framework::Tensor out;
+
+  auto* cpu_place = new paddle::platform::CPUPlace();
+  float* input1_ptr = input1.mutable_data<float>({2, 3}, *cpu_place);
+  float arr[6] = {0, 1, 2, 3, 4, 5};
+  memcpy(input1_ptr, arr, 6 * sizeof(float));
+
+  auto* gpu_place = new paddle::platform::GPUPlace(0);
+  paddle::platform::DeviceContext* context =
+      new paddle::platform::CUDADeviceContext(*gpu_place);
+
+  input1_gpu.CopyFrom<float>(input1, *gpu_place);
+  input2_gpu.CopyFrom<float>(input1, *gpu_place);
+
+  out_gpu.mutable_data<float>({3, 3}, *gpu_place);
+
+  paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
+      input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0, context);
+
+  out.CopyFrom<float>(out_gpu, *cpu_place);
+
+  float* out_ptr = out.data<float>();
+  EXPECT_EQ(out_ptr[0], 9);
+  EXPECT_EQ(out_ptr[1], 12);
+  EXPECT_EQ(out_ptr[2], 15);
+  EXPECT_EQ(out_ptr[3], 12);
+  EXPECT_EQ(out_ptr[4], 17);
+  EXPECT_EQ(out_ptr[5], 22);
+  EXPECT_EQ(out_ptr[6], 15);
+  EXPECT_EQ(out_ptr[7], 22);
+  EXPECT_EQ(out_ptr[8], 29);
+}
+#endif

paddle/operators/mean_op.cc

@@ -18,13 +18,14 @@ namespace paddle {
 namespace operators {
 
 class MeanOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.InputSize(), 1, "Input size of AddOp must be one");
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of AddOp must be one");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "input should be set");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), "output should be set");
-    ctx.Output<Tensor>(0)->Resize(framework::make_ddim({1}));
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input of MeanOp must be initialized.");
+    ctx.Output<Tensor>("Out")->Resize({1});
   }
 };
 
@@ -33,15 +34,18 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
   MeanOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input of mean op");
-    AddOutput("Out", "The output of mean op").IgnoreGradient();
+    AddOutput("Out", "The output of mean op").AsNoGradient();
     AddComment("Mean Operator");
   }
 };
 
 class MeanGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<Tensor>("X" + framework::kGradVarSuffix)
+    ctx.Output<Tensor>(framework::GradVarName("X"))
         ->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };

paddle/operators/mean_op.h

@@ -31,14 +31,14 @@ template <typename Place, typename T>
 class MeanKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto input = context.Input<Tensor>(0);
-    auto output = context.Output<Tensor>(0);
+    auto* input = context.Input<Tensor>("X");
+    auto* output = context.Output<Tensor>("Out");
 
     output->mutable_data<T>(context.GetPlace());
 
     auto X = EigenVector<T>::Flatten(*input);
     auto y = EigenScalar<T>::From(*output);
-    auto place = context.GetEigenDevice<Place>();
+    auto& place = context.GetEigenDevice<Place>();
 
     y.device(place) = X.mean();
   }
@@ -48,10 +48,10 @@ template <typename Place, typename T>
 class MeanGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto OG = context.Input<Tensor>("Out" + framework::kGradVarSuffix);
+    auto OG = context.Input<Tensor>(framework::GradVarName("Out"));
     PADDLE_ENFORCE(framework::product(OG->dims()) == 1,
                    "Mean Gradient should be scalar");
-    auto IG = context.Output<Tensor>("X" + framework::kGradVarSuffix);
+    auto IG = context.Output<Tensor>(framework::GradVarName("X"));
     IG->mutable_data<T>(context.GetPlace());
 
     T ig_size = (T)framework::product(IG->dims());

paddle/operators/mul_op.cc

@@ -13,16 +13,19 @@
    limitations under the License. */
 
 #include "paddle/operators/mul_op.h"
+#include "paddle/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {
 
 class MulOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 2, "The mul op must take two inputs");
-    auto dim0 = ctx.Input<Tensor>(0)->dims();
-    auto dim1 = ctx.Input<Tensor>(1)->dims();
+    auto dim0 = ctx.Input<Tensor>("X")->dims();
+    auto dim1 = ctx.Input<Tensor>("Y")->dims();
     PADDLE_ENFORCE_EQ(dim0.size(), 2,
                       "input X(%s) should be a tensor with 2 dims, a matrix",
                       ctx.op_.Input("X"));
@@ -32,8 +35,7 @@ class MulOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(
         dim0[1], dim1[0],
         "First matrix's width must be equal with second matrix's height.");
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "The mul op takes only one output");
-    ctx.Output<Tensor>(0)->Resize({dim0[0], dim1[1]});
+    ctx.Output<Tensor>("Out")->Resize({dim0[0], dim1[1]});
   }
 };
 
@@ -53,6 +55,9 @@ The equation is: Out = X * Y
 };
 
 class MulOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {}
   std::string DebugString() const override {

paddle/operators/mul_op.cu

@@ -16,5 +16,4 @@
 #include "paddle/operators/mul_op.h"
 
 namespace ops = paddle::operators;
-
 REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);

paddle/operators/mul_op.h

@@ -13,6 +13,9 @@
    limitations under the License. */
 
 #pragma once
+
+#include "paddle/operators/math/math_function.h"
+
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
 
@@ -30,17 +33,14 @@ class MulKernel : public framework::OpKernel {
   void Compute(const framework::ExecutionContext& context) const override {
     Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair = {
         {Eigen::IndexPair<Eigen::DenseIndex>(1, 0)}};
-
-    auto input0 = context.Input<Tensor>("X");
-    auto input1 = context.Input<Tensor>("Y");
-    auto output = context.Output<Tensor>(0);
-
+    auto* input0 = context.Input<Tensor>("X");
+    auto* input1 = context.Input<Tensor>("Y");
+    auto* output = context.Output<Tensor>("Out");
     output->mutable_data<T>(context.GetPlace());
-
     auto X = EigenMatrix<T>::From(*input0);
     auto Y = EigenMatrix<T>::From(*input1);
     auto Z = EigenMatrix<T>::From(*output);
-    auto place = context.GetEigenDevice<Place>();
+    auto& place = context.GetEigenDevice<Place>();
 
     Z.device(place) = X.contract(Y, dim_pair);
   }

paddle/operators/net_op.cc

@@ -15,48 +15,42 @@
 */
 
 #include "paddle/operators/net_op.h"
+#include <set>
+#include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace operators {
 
+const char NetOp::kAll[] = "all";
+
 void NetOp::CompleteAddOp(bool calc) {
   add_op_done_ = true;
   if (!calc) return;
-  std::unordered_set<std::string> input_set;
-  std::unordered_set<std::string> output_set;
-  std::unordered_set<std::string> temp_output;
+  std::set<std::string> input_set;
+  std::set<std::string> output_set;
   for (auto& op : ops_) {
     for (auto& ipt : op->inputs_) {
-      if (!Contains(output_set, ipt)) {  // Not other op's output
-        input_set.insert(ipt);
-      } else {
-        temp_output.insert(ipt);
+      for (auto& var_name : ipt.second) {
+        if (!Contains(output_set, var_name)) {  // Not other op's output
+          input_set.insert(var_name);
+        } else {
+          intermediate_outputs_.insert(var_name);
+        }
       }
     }
 
     for (auto& opt : op->outputs_) {
-      output_set.insert(opt);
-    }
-  }
-
-  inputs_.reserve(input_set.size());
-  std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs_));
-  std::sort(inputs_.begin(), inputs_.end());
-
-  outputs_.reserve(output_set.size());
-  std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs_));
-  std::sort(outputs_.begin(), outputs_.end());
-
-  std::vector<int> tmp_index;
-  tmp_index.reserve(temp_output.size());
-  int output_len = static_cast<int>(outputs_.size());
-  for (int i = 0; i < output_len; ++i) {
-    if (Contains(temp_output, outputs_[i])) {
-      tmp_index.push_back(i);
+      for (auto& var_name : opt.second) {
+        output_set.insert(var_name);
+      }
     }
   }
-
-  attrs_["temporary_index"] = tmp_index;
+  auto& inputs = inputs_[kAll];
+  inputs.reserve(input_set.size());
+  std::copy(input_set.begin(), input_set.end(), std::back_inserter(inputs));
+  auto& outputs = outputs_[kAll];
+  outputs.reserve(output_set.size());
+  std::copy(output_set.begin(), output_set.end(), std::back_inserter(outputs));
 }
 
 std::string NetOp::DebugString() const {
@@ -73,5 +67,25 @@ std::string NetOp::DebugString() const {
 
 bool NetOp::IsNetOp() const { return true; }
 
+std::vector<std::string> NetOp::OutputVars(bool has_intermediate) const {
+  if (has_intermediate) {
+    return this->outputs_.at(kAll);
+  }
+  auto& all = this->outputs_.at(kAll);
+  std::vector<std::string> ret_val;
+  for (auto& each : all) {
+    if (!Contains(intermediate_outputs_, each)) {
+      ret_val.push_back(each);
+    }
+  }
+  return ret_val;
+}
+
+NetOp::NetOp(const std::string& type,
+             const framework::OperatorBase::VarNameMap& inputs,
+             const framework::OperatorBase::VarNameMap& outputs,
+             const framework::AttributeMap& attrs)
+    : OperatorBase(type, inputs, outputs, attrs) {}
+
 }  // namespace operators
 }  // namespace paddle

paddle/operators/net_op.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/framework/framework.pb.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
@@ -35,6 +36,11 @@ namespace operators {
  */
 class NetOp : public framework::OperatorBase {
  public:
+  static const char kAll[];
+  NetOp() : framework::OperatorBase("plain_net", {}, {}, {}) {}
+  NetOp(const std::string& type, const VarNameMap& inputs,
+        const VarNameMap& outputs, const framework::AttributeMap& attrs);
+
   /**
    * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch
@@ -90,11 +96,13 @@ class NetOp : public framework::OperatorBase {
   std::string DebugString() const override;
 
   bool IsNetOp() const override;
+  std::vector<std::string> OutputVars(bool has_intermediate) const override;
 
   std::vector<std::shared_ptr<OperatorBase>> ops_;
 
  private:
   bool add_op_done_{false};
+  std::set<std::string> intermediate_outputs_;
 
   template <typename T, typename KeyType>
   static bool Contains(T container, KeyType key) {

paddle/operators/net_op_test.cc

@@ -12,6 +12,7 @@ static int run_cnt = 0;
 
 class TestOp : public framework::OperatorBase {
  public:
+  using framework::OperatorBase::OperatorBase;
   void InferShape(const Scope& scope) const override { ++infer_shape_cnt; }
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {
@@ -21,6 +22,7 @@ class TestOp : public framework::OperatorBase {
 
 class EmptyOp : public framework::OperatorBase {
  public:
+  using framework::OperatorBase::OperatorBase;
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope, const DeviceContext& dev_ctx) const override {}
 };
@@ -42,40 +44,32 @@ TEST(OpKernel, all) {
   auto net = std::make_shared<NetOp>();
   ASSERT_NE(net, nullptr);
 
-  auto op1 = std::make_shared<TestOp>();
-  op1->inputs_ = {"x", "w1", "b1"};
-  op1->outputs_ = {"y"};
+  auto op1 = std::shared_ptr<TestOp>(
+      new TestOp("test", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
+                 {{"Out", {"y"}}}, {}));
   net->AddOp(op1);
 
-  auto op2 = std::make_shared<TestOp>();
-  op2->inputs_ = {"y", "w2", "b2"};
-  op2->outputs_ = {"z"};
+  auto op2 = std::shared_ptr<TestOp>(
+      new TestOp("test", {{"X", {"y"}}, {"W", {"w2"}}, {"b", {"b2"}}},
+                 {{"Out", {"z"}}}, {}));
   net->AddOp(op2);
 
   net->CompleteAddOp();
-  AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"}, net->inputs_);
-  AssertSameVectorWithoutOrder({"y", "z"}, net->outputs_);
-  auto tmp_idx_iter = net->attrs_.find("temporary_index");
-  ASSERT_NE(net->attrs_.end(), tmp_idx_iter);
-  auto& tmp_idx = boost::get<std::vector<int>>(tmp_idx_iter->second);
-  ASSERT_EQ(1UL, tmp_idx.size());
-  ASSERT_EQ("y", net->outputs_[tmp_idx[0]]);
+  AssertSameVectorWithoutOrder({"x", "w1", "b1", "w2", "b2"},
+                               net->inputs_.at(NetOp::kAll));
+  AssertSameVectorWithoutOrder({"y", "z"}, net->outputs_.at(NetOp::kAll));
 
-  Scope scope;
-  platform::CPUDeviceContext dev_ctx;
+  auto final_outs = net->OutputVars(false);
 
-  net->InferShape(scope);
-  net->Run(scope, dev_ctx);
-  ASSERT_EQ(2, infer_shape_cnt);
-  ASSERT_EQ(2, run_cnt);
-  ASSERT_THROW(net->AddOp(op2), platform::EnforceNotMet);
+  ASSERT_EQ(final_outs.size(), 1UL);
+  ASSERT_EQ(final_outs[0], "z");
 }
 
 TEST(NetOp, insert_op) {
   NetOp net;
-  auto op1 = std::make_shared<EmptyOp>();
-  op1->inputs_ = {"x", "w1", "b1"};
-  op1->outputs_ = {"y"};
+  auto op1 = std::shared_ptr<EmptyOp>(
+      new EmptyOp("empty", {{"X", {"x"}}, {"W", {"w1"}}, {"b", {"b1"}}},
+                  {{"Out", {"y"}}}, {}));
   net.AddOp(op1);
   net.InsertOp(0, op1);
   ASSERT_EQ(2UL, net.ops_.size());

paddle/operators/recurrent_op.cc

@@ -91,12 +91,17 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
       // create step net's temp inputs
       for (auto& input : net_op->inputs_) {
         // the weight are located in parent scope
-        if (!step_scope.FindVar(input))
-          step_scope.NewVar(input)->GetMutable<Tensor>();
+        for (auto& var_name : input.second) {
+          if (!step_scope.FindVar(var_name)) {
+            step_scope.NewVar(var_name)->GetMutable<Tensor>();
+          }
+        }
       }
       // create stepnet's outputs
       for (const auto& output : net_op->outputs_) {
-        step_scope.NewVar(output);
+        for (auto& var_name : output.second) {
+          step_scope.NewVar(var_name);
+        }
       }
       step_scopes->emplace_back(&step_scope);
     }
@@ -130,8 +135,11 @@ const rnn::ArgumentName RecurrentGradientOp::kArgName{
     "inlink@grad", "inlink_alias", "outlink_alias",
     "memories",    "pre_memories", "boot_memories@grad"};
 
-void RecurrentOp::Init() {
-  OperatorBase::Init();
+RecurrentOp::RecurrentOp(const std::string& type,
+                         const framework::OperatorBase::VarNameMap& inputs,
+                         const framework::OperatorBase::VarNameMap& outputs,
+                         const framework::AttributeMap& attrs)
+    : OperatorBase(type, inputs, outputs, attrs) {
   std::unique_ptr<rnn::Argument> arg(new rnn::Argument());
   rnn::InitArgument(kArgName, arg.get(), *this);
   alg_.Init(std::move(arg));
@@ -147,13 +155,13 @@ class RecurrentAlgorithmProtoAndCheckerMaker
     // inputs and outputs stored in proto
     AddInput(name.inlinks,
              "the inputs that need to be segmented for each step.")
-        .SetMultiple();
+        .AsDuplicable();
     AddInput(name.boot_memories, "variables to initialize memories.")
-        .SetMultiple();
+        .AsDuplicable();
     AddInput(name.step_net, "network shared by all steps.");
 
     AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
-        .SetMultiple();
+        .AsDuplicable();
     AddOutput(name.step_scopes, "step scopes");
 
     // Attributes stored in AttributeMap
@@ -225,8 +233,11 @@ void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const {
   LinkBootMemoryGradients(step_scopes[0], true /*infer_shape_mode*/);
 }
 
-void RecurrentGradientOp::Init() {
-  OperatorBase::Init();
+RecurrentGradientOp::RecurrentGradientOp(
+    const std::string& type, const framework::OperatorBase::VarNameMap& inputs,
+    const framework::OperatorBase::VarNameMap& outputs,
+    const framework::AttributeMap& attrs)
+    : OperatorBase(type, inputs, outputs, attrs) {
   std::unique_ptr<rnn::Argument> arg(new rnn::Argument());
   rnn::InitArgument(kArgName, arg.get(), *this);
   alg_.Init(std::move(arg));

paddle/operators/recurrent_op.h

@@ -101,11 +101,11 @@ class RecurrentGradientAlgorithm {
 
 class RecurrentOp final : public framework::OperatorBase {
  public:
-  void Init() override;
-
+  RecurrentOp(const std::string& type, const VarNameMap& inputs,
+              const VarNameMap& outputs, const framework::AttributeMap& attrs);
   /**
-   * InferShape must be called before Run.
-   */
+     * InferShape must be called before Run.
+     */
   void InferShape(const framework::Scope& scope) const override {
     alg_.InferShape(scope);
   }
@@ -123,7 +123,9 @@ class RecurrentOp final : public framework::OperatorBase {
 
 class RecurrentGradientOp final : public framework::OperatorBase {
  public:
-  void Init() override;
+  RecurrentGradientOp(const std::string& type, const VarNameMap& inputs,
+                      const VarNameMap& outputs,
+                      const framework::AttributeMap& attrs);
 
   /**
    * InferShape must be called before Run.

paddle/operators/recurrent_op_test.cc

@@ -25,157 +25,7 @@
 namespace paddle {
 namespace operators {
 
-using framework::make_ddim;
-using framework::DDim;
-using framework::Tensor;
-using framework::Variable;
-using framework::Scope;
-using framework::OpRegistry;
-
-class RecurrentOpTest : public ::testing::Test {
- protected:
-  virtual void SetUp() override {
-    CreateGlobalVariables();
-    CreateStepNet();
-    CreateRNNOp();
-  }
-
-  virtual void TearDown() override {}
-
-  void CreateGlobalVariables() {
-    // create input, and init content
-    LOG(INFO) << "create global variable x";
-    for (auto inlink : std::vector<std::string>{"x", "x0", "x1", "h"}) {
-      Variable* x = scope_.NewVar(inlink);
-      DDim dims = make_ddim(std::vector<int>{
-          10 /*sent size*/, 20 /*batch size*/, 30 /*input dim*/});
-      x->GetMutable<Tensor>()->mutable_data<float>(dims, platform::CPUPlace());
-    }
-    // create output alias just for test
-    for (auto inlink : std::vector<std::string>{"h@alias"}) {
-      Variable* x = scope_.NewVar(inlink);
-      DDim dims =
-          make_ddim(std::vector<int>{20 /*batch size*/, 30 /*input dim*/});
-      x->GetMutable<Tensor>()->mutable_data<float>(dims, platform::CPUPlace());
-    }
-
-    LOG(INFO) << "create global variable w";
-    Variable* w = scope_.NewVar("rnn/w");
-    w->GetMutable<Tensor>()->mutable_data<float>(
-        make_ddim(std::vector<int>{30, 30}), platform::CPUPlace());
-
-    for (auto boot : std::vector<std::string>{"h_boot"}) {
-      LOG(INFO) << "create global variable " << boot;
-      Variable* h_boot = scope_.NewVar(boot);
-      h_boot->GetMutable<Tensor>()->mutable_data<float>(
-          make_ddim(std::vector<int>{20 /*batch size*/, 30 /*input dim*/}),
-          platform::CPUPlace());
-    }
-
-    LOG(INFO) << "create variable step_scopes";
-    scope_.NewVar("step_scopes");
-
-    LOG(INFO) << "create variable h";
-    scope_.NewVar("h");
-  }
-
-  void CreateRNNOp() {
-    framework::OpDesc op_desc;
-
-    op_desc.set_type("recurrent_op");
-    // inlinks 0
-    op_desc.add_inputs("x");
-    op_desc.add_inputs("x0");
-    op_desc.add_inputs("x1");
-    // boot_memories 3
-    op_desc.add_inputs("h_boot");
-    // step net 5
-    op_desc.add_inputs("step_net");
-    // outlinks 6
-    op_desc.add_outputs("h");
-    // step scopes 7
-    op_desc.add_outputs("step_scopes");
-
-    auto _input_format = std::vector<int>{
-        0,  // in_link
-        3,  // memories
-        4   // step_net
-    };
-    auto input_format = op_desc.add_attrs();
-    input_format->set_name("input_format");
-    input_format->set_type(paddle::framework::AttrType::INTS);
-    for (auto i : _input_format) {
-      input_format->add_ints(i);
-    }
-
-    auto output_format = op_desc.add_attrs();
-    output_format->set_name("output_format");
-    output_format->set_type(paddle::framework::AttrType::INTS);
-    for (auto i : std::vector<int>{0, 1, 2}) {
-      output_format->add_ints(i);
-    }
-
-    auto inlink_alias = op_desc.add_attrs();
-    inlink_alias->set_name("inlink_alias");
-    inlink_alias->set_type(paddle::framework::AttrType::STRINGS);
-
-    auto outlink_alias = op_desc.add_attrs();
-    outlink_alias->set_name("outlink_alias");
-    outlink_alias->set_type(paddle::framework::AttrType::STRINGS);
-
-    auto pre_memories = op_desc.add_attrs();
-    pre_memories->set_name("pre_memories");
-    pre_memories->set_type(paddle::framework::AttrType::STRINGS);
-
-    auto memories = op_desc.add_attrs();
-    memories->set_name("memories");
-    memories->set_type(paddle::framework::AttrType::STRINGS);
-
-    // create inlink_alias
-    for (const auto& item :
-         std::vector<std::string>{"x@alias", "x0@alias", "x1@alias"}) {
-      inlink_alias->add_strings(item);
-    }
-    // pre memories
-    for (const auto& item : std::vector<std::string>{"rnn/h@pre"}) {
-      pre_memories->add_strings(item);
-    }
-    // memories
-    for (const auto& item : std::vector<std::string>{"rnn/h"}) {
-      memories->add_strings(item);
-    }
-    // output alias
-    for (const auto& item : std::vector<std::string>{"h@alias"}) {
-      outlink_alias->add_strings(item);
-    }
-
-    rnn_op_ = OpRegistry::CreateOp(op_desc);
-
-    LOG(INFO) << "rnn_op finish init";
-  }
-
-  void CreateStepNet() {
-    LOG(INFO) << "create variable step_net";
-    Variable* var = scope_.NewVar("step_net");
-    auto net = var->GetMutable<NetOp>();
-    net->AddOp(
-        OpRegistry::CreateOp("mul", {"rnn/h@pre", "rnn/w"}, {"rnn/s"}, {}));
-
-    net->AddOp(
-        OpRegistry::CreateOp("add_two", {"x@alias", "rnn/s"}, {"rnn/h"}, {}));
-    net->CompleteAddOp();
-  }
-
-  // father scope
-  Scope scope_;
-  std::shared_ptr<framework::OperatorBase> rnn_op_;
-};
-
-TEST_F(RecurrentOpTest, Run) {
-  platform::CPUDeviceContext ctx;
-  rnn_op_->InferShape(scope_);
-  rnn_op_->Run(scope_, ctx);
-}
+using namespace paddle::framework;
 
 class RecurrentGradientAlgorithmTest : public ::testing::Test {
  protected:
@@ -281,11 +131,13 @@ class RecurrentGradientAlgorithmTest : public ::testing::Test {
     LOG(INFO) << "create variable step_net";
     Variable* var = scope_.NewVar("step_net");
     auto net = var->GetMutable<NetOp>();
-    net->AddOp(OpRegistry::CreateOp("mul", {"rnn/h_pre", "rnn/w", "rnn/s_grad"},
-                                    {"rnn/h_pre_grad", "rnn/w_grad"}, {}));
+    // TODO(qingqing) modify backward op create for RNNOp unit test
+    // and the unit test will be removed to Python.
+    // net->AddOp(OpRegistry::CreateOp("mul", {"X", {"rnn/h_pre", "rnn/w",
+    //     "rnn/s_grad"}}, {"Y", {"rnn/h_pre_grad", "rnn/w_grad"}}, {}));
 
-    net->AddOp(OpRegistry::CreateOp("add_two", {"rnn/h_grad"},
-                                    {"rnn/x_grad", "rnn/s_grad"}, {}));
+    // net->AddOp(OpRegistry::CreateOp("add_two", {"X", {"rnn/h_grad"}},
+    //			    {"Y", {"rnn/x_grad"}}, {"Out", "rnn/s_grad"}}, {}));
     net->CompleteAddOp();
   }
 
@@ -359,7 +211,8 @@ TEST(RecurrentOp, LinkMemories) {
   memories.push_back(mem_attr);
 
   for (size_t i = 1; i < len; ++i) {
-    rnn::LinkMemories(step_scopes, memories, i, -1, false /*infer_shape_mode*/);
+    rnn::LinkMemories(step_scopes, memories, i, -1, false
+                      /*infer_shape_mode*/);
   }
   // check
   for (size_t i = 0; i < len - 1; ++i) {
@@ -375,7 +228,8 @@ TEST(RecurrentOp, LinkMemories) {
   }
 
   for (int i = len - 2; i >= 0; --i) {
-    rnn::LinkMemories(step_scopes, memories, i, 1, false /*infer_shape_mode*/);
+    rnn::LinkMemories(step_scopes, memories, i, 1, false
+                      /*infer_shape_mode*/);
   }
   // check
   for (int i = len - 2; i >= 0; --i) {
@@ -395,4 +249,4 @@ TEST(RecurrentOp, LinkMemories) {
 
 USE_OP(add_two);
 USE_OP(mul);
-USE_OP_WITHOUT_KERNEL(recurrent_op);
+USE_OP_ITSELF(recurrent_op);

paddle/operators/rowwise_add_op.cc

@@ -18,18 +18,19 @@ namespace paddle {
 namespace operators {
 
 class RowwiseAddOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 2UL,
-                   "Two inputs is needed by rowwise add");
-    auto dim0 = ctx.Input<Tensor>(0)->dims();
-    auto dim1 = ctx.Input<Tensor>(1)->dims();
+    auto dim0 = ctx.Input<Tensor>("X")->dims();
+    auto dim1 = ctx.Input<Tensor>("b")->dims();
 
     PADDLE_ENFORCE(dim0.size() == 2, "Input 0 must be matrix");
     PADDLE_ENFORCE(dim1.size() == 1, "The second input must be vector");
     PADDLE_ENFORCE(dim0[1] == dim1[0], "The width of two input must be same");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1, "The output size must be 1");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    PADDLE_ENFORCE(ctx.OutputSize("Out") == 1, "The output size must be 1");
+    ctx.Output<Tensor>("Out")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 

paddle/operators/rowwise_add_op.h

@@ -31,11 +31,11 @@ template <typename Place, typename T>
 class RowwiseAddKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto out = context.Output<Tensor>(0);
+    auto out = context.Output<Tensor>("Out");
     out->mutable_data<T>(context.GetPlace());
 
-    auto input = EigenMatrix<T>::From(*context.Input<Tensor>(0));
-    auto bias = EigenVector<T>::From(*context.Input<Tensor>(1));
+    auto input = EigenMatrix<T>::From(*context.Input<Tensor>("X"));
+    auto bias = EigenVector<T>::From(*context.Input<Tensor>("b"));
     auto output = EigenMatrix<T>::From(*out);
 
     const int bias_size = bias.dimension(0);

paddle/operators/sgd_op.cc

@@ -18,16 +18,15 @@ namespace paddle {
 namespace operators {
 
 class SGDOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.InputSize(), 2, "Input size of SGDOp must be two");
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of SGDOp must be one");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "inputs[0] mast be set");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(1), "inputs[1] mast be set");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(0), "outputs[0] mast be set");
-    PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims() == ctx.Input<Tensor>(1)->dims(),
-                   "Two input of SGD Op's dimension must be same.");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    PADDLE_ENFORCE(
+        ctx.Input<Tensor>("param")->dims() == ctx.Input<Tensor>("grad")->dims(),
+        "Two input of SGD Op's dimension must be same.");
+    ctx.Output<Tensor>("param_out")->Resize(ctx.Input<Tensor>("param")->dims());
   }
 };
 

paddle/operators/sgd_op_test.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include <gtest/gtest.h>
-#include <paddle/framework/op_registry.h>
-USE_OP(sgd);
-TEST(SGDOp, GetOpProto) {
-  auto& protos = paddle::framework::OpRegistry::protos();
-  auto it = protos.find("sgd");
-  ASSERT_NE(it, protos.end());
-}

paddle/operators/sigmoid_op.cc

@@ -18,11 +18,12 @@ namespace paddle {
 namespace operators {
 
 class SigmoidOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1, "Sigmoid Op only have one input");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1, "Sigmoid Op only have one output");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -38,6 +39,9 @@ class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
 };
 
 class SigmoidOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());

paddle/operators/sigmoid_op.h

@@ -28,8 +28,8 @@ template <typename Place, typename T>
 class SigmoidKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto input = context.Input<Tensor>(0);
-    auto output = context.Output<Tensor>(0);
+    auto input = context.Input<Tensor>("X");
+    auto output = context.Output<Tensor>("Y");
     output->mutable_data<T>(context.GetPlace());
 
     // The clipping is used in Paddle's raw implenmention

paddle/operators/softmax_op.cc

@@ -18,14 +18,13 @@ namespace paddle {
 namespace operators {
 
 class SoftmaxOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL,
-                      "Only one input is need for softmax");
-    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims().size(), 2UL,
-                      "The input of softmax op must be matrix");
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL,
-                      "Only one output is need for softmax");
+    PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
+                   "The input of softmax op must be matrix");
     ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
@@ -42,13 +41,12 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
 };
 
 class SoftmaxOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_EQ(ctx.InputSize(), 3UL,
-                      "Input of SoftmaxOpGrad should be 3, X, Y, YG");
-    PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL,
-                      "Output of SoftmaxOpGrad should be 1");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) should not be null");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")),
                             "Input(Y@GRAD) should not be null");
     PADDLE_ENFORCE(ctx.Input<Tensor>("Y")->dims() ==

paddle/operators/uniform_random_op.cc

@@ -27,7 +27,7 @@ template <typename T>
 class CPUUniformRandomKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* tensor = context.Output<framework::Tensor>(0);
+    auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
     unsigned int seed =
         static_cast<unsigned int>(context.op_.GetAttr<int>("seed"));
@@ -46,11 +46,14 @@ class CPUUniformRandomKernel : public framework::OpKernel {
 };
 
 class UniformRandomOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
     PADDLE_ENFORCE(GetAttr<float>("min") < GetAttr<float>("max"),
                    "uniform_random's min must less then max");
-    auto* tensor = ctx.Output<framework::Tensor>(0);
+    auto* tensor = ctx.Output<framework::Tensor>("Out");
     auto dims = GetAttr<std::vector<int>>("dims");
     tensor->Resize(framework::make_ddim(dims));
   }

paddle/operators/uniform_random_op.cu

@@ -46,12 +46,13 @@ template <typename T>
 class GPUUniformRandomKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* tensor = context.Output<framework::Tensor>(0);
+    auto* tensor = context.Output<framework::Tensor>("Out");
     T* data = tensor->mutable_data<T>(context.GetPlace());
     unsigned int seed =
         static_cast<unsigned int>(context.op_.GetAttr<int>("seed"));
     if (seed == 0) {
-      seed = std::random_device()();
+      std::random_device rd;
+      seed = rd();
     }
     T min = static_cast<T>(context.op_.GetAttr<float>("min"));
     T max = static_cast<T>(context.op_.GetAttr<float>("max"));

paddle/platform/dynload/cublas.h

@@ -48,13 +48,13 @@ extern void *cublas_dso_handle;
   };                                                                \
   extern DynLoad__##__name __name
 #else
-#define DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name) \
-  struct DynLoad__##__name {                     \
-    inline template <typename... Args>           \
-    cublasStatus_t operator()(Args... args) {    \
-      return __name(args...);                    \
-    }                                            \
-  };                                             \
+#define DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name)     \
+  struct DynLoad__##__name {                         \
+    template <typename... Args>                      \
+    inline cublasStatus_t operator()(Args... args) { \
+      return __name(args...);                        \
+    }                                                \
+  };                                                 \
   extern DynLoad__##__name __name
 #endif
 
@@ -62,12 +62,12 @@ extern void *cublas_dso_handle;
   DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name)
 
 #define CUBLAS_BLAS_ROUTINE_EACH(__macro) \
-  __macro(cublasSgemv);                   \
-  __macro(cublasDgemv);                   \
-  __macro(cublasSgemm);                   \
-  __macro(cublasDgemm);                   \
-  __macro(cublasSgeam);                   \
-  __macro(cublasDgeam);                   \
+  __macro(cublasSgemv_v2);                \
+  __macro(cublasDgemv_v2);                \
+  __macro(cublasSgemm_v2);                \
+  __macro(cublasDgemm_v2);                \
+  __macro(cublasSgeam_v2);                \
+  __macro(cublasDgeam_v2);                \
   __macro(cublasCreate_v2);               \
   __macro(cublasDestroy_v2);              \
   __macro(cublasSetStream_v2);            \

paddle/platform/dynload/curand.h

@@ -55,6 +55,7 @@ extern void *curand_dso_handle;
   __macro(curandSetPseudoRandomGeneratorSeed); \
   __macro(curandGenerateUniform);              \
   __macro(curandGenerateUniformDouble);        \
+  __macro(curandGenerateNormal);               \
   __macro(curandDestroyGenerator);
 
 CURAND_RAND_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CURAND_WRAP);

paddle/platform/enforce.h

@@ -15,11 +15,12 @@ limitations under the License. */
 #pragma once
 
 #include <execinfo.h>
-#include <paddle/string/printf.h>
 #include <iomanip>
 #include <sstream>
 #include <stdexcept>
 #include <string>
+#include "paddle/string/printf.h"
+#include "paddle/string/to_string.h"
 
 #ifndef PADDLE_ONLY_CPU
 
@@ -194,8 +195,8 @@ inline void throw_on_error(T e) {
 #define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...)        \
   PADDLE_ENFORCE(__VAL0 __CMP __VAL1,                                         \
                  "enforce %s " #__CMP " %s failed, %s " #__INV_CMP " %s\n%s", \
-                 #__VAL0, #__VAL1, std::to_string(__VAL0),                    \
-                 std::to_string(__VAL1),                                      \
+                 #__VAL0, #__VAL1, paddle::string::to_string(__VAL0),         \
+                 paddle::string::to_string(__VAL1),                           \
                  paddle::string::Sprintf("" __VA_ARGS__));
 
 }  // namespace platform

paddle/platform/enforce_test.cc

@@ -9,6 +9,8 @@ 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 <array>
+#include <iostream>
 #include <memory>
 
 #include "gtest/gtest.h"
@@ -83,7 +85,7 @@ TEST(ENFORCE_NE, FAIL) {
   } catch (paddle::platform::EnforceNotMet error) {
     caught_exception = true;
     EXPECT_TRUE(HasPrefix(StringPiece(error.what()),
-                          "enforce 1.0 != 1UL failed, 1.000000 == 1"))
+                          "enforce 1.0 != 1UL failed, 1 == 1"))
         << error.what() << " does not have expected prefix";
   }
   EXPECT_TRUE(caught_exception);
@@ -176,3 +178,39 @@ TEST(ENFORCE_NOT_NULL, FAIL) {
   }
   EXPECT_TRUE(caught_exception);
 }
+
+struct Dims {
+  size_t dims_[4];
+
+  bool operator==(const Dims& o) const {
+    for (size_t i = 0; i < 4; ++i) {
+      if (dims_[i] != o.dims_[i]) return false;
+    }
+    return true;
+  }
+};
+
+std::ostream& operator<<(std::ostream& os, const Dims& d) {
+  for (size_t i = 0; i < 4; ++i) {
+    if (i == 0) {
+      os << "[";
+    }
+    os << d.dims_[i];
+    if (i == 4 - 1) {
+      os << "]";
+    } else {
+      os << ", ";
+    }
+  }
+  return os;
+}
+
+TEST(ENFORCE_USER_DEFINED_CLASS, EQ) {
+  Dims a{{1, 2, 3, 4}}, b{{1, 2, 3, 4}};
+  PADDLE_ENFORCE_EQ(a, b);
+}
+
+TEST(ENFORCE_USER_DEFINED_CLASS, NE) {
+  Dims a{{1, 2, 3, 4}}, b{{5, 6, 7, 8}};
+  ASSERT_THROW(PADDLE_ENFORCE_EQ(a, b), paddle::platform::EnforceNotMet);
+}
\ No newline at end of file

paddle/platform/place.h

@@ -14,8 +14,8 @@ limitations under the License. */
 
 #pragma once
 
-#include <boost/variant.hpp>
 #include <iostream>
+#include "paddle/platform/variant.h"
 
 namespace paddle {
 namespace platform {

paddle/platform/variant.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 <boost/config.hpp>
+
+#ifndef PADDLE_ONLY_CPU
+
+// Because boost's variadic templates has bug on nvcc, boost will disable
+// variadic template support when GPU enabled on nvcc.
+// Define BOOST_NO_CXX11_VARIADIC_TEMPLATES on gcc/clang to generate same
+// function symbols.
+//
+// https://github.com/PaddlePaddle/Paddle/issues/3386
+#ifndef BOOST_NO_CXX11_VARIADIC_TEMPLATES
+#define BOOST_NO_CXX11_VARIADIC_TEMPLATES
+#endif
+#endif
+
+#include <boost/variant.hpp>

paddle/pserver/test/CMakeLists.txt

@@ -3,7 +3,7 @@ add_unittest_without_exec(socket_test
     SocketTest.cpp)
 
 add_test(NAME socket_test
-    COMMAND ${PROJ_ROOT}/paddle/.set_port.sh -p port
+    COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port
         ${CMAKE_CURRENT_BINARY_DIR}/socket_test --loop_time=10)
 
 ####################### test_ProtoServer ####################
@@ -12,7 +12,7 @@ add_unittest_without_exec(test_ProtoServer
 
 IF(NOT ON_TRAVIS)
     add_test(NAME test_ProtoServer
-        COMMAND ${PROJ_ROOT}/paddle/.set_port.sh -p port
+        COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port
             ${CMAKE_CURRENT_BINARY_DIR}/test_ProtoServer)
 ENDIF(NOT ON_TRAVIS)
 
@@ -24,5 +24,5 @@ ENDIF(NOT ON_TRAVIS)
 add_unittest_without_exec(test_ParameterServer2
     test_ParameterServer2.cpp)
 add_test(NAME test_ParameterServer2
-    COMMAND ${PROJ_ROOT}/paddle/.set_port.sh -p port -n 4
+    COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port -n 4
         ${CMAKE_CURRENT_BINARY_DIR}/test_ParameterServer2)

paddle/scripts/docker/build.sh

@@ -31,7 +31,7 @@ Configuring cmake in /paddle/build ...
       -DWITH_DOC=OFF
       -DWITH_GPU=${WITH_GPU:-OFF}
       -DWITH_AVX=${WITH_AVX:-OFF}
-      -DWITH_GOLANG=${WITH_GOLANG:-OFF}
+      -DWITH_GOLANG=${WITH_GOLANG:-ON}
       -DWITH_SWIG_PY=ON
       -DWITH_C_API=${WITH_C_API:-OFF}
       -DWITH_PYTHON=${WITH_PYTHON:-ON}
@@ -51,7 +51,7 @@ cmake .. \
       -DWITH_DOC=OFF \
       -DWITH_GPU=${WITH_GPU:-OFF} \
       -DWITH_AVX=${WITH_AVX:-OFF} \
-      -DWITH_GOLANG=${WITH_GOLANG:-OFF} \
+      -DWITH_GOLANG=${WITH_GOLANG:-ON} \
       -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \
       -DWITH_C_API=${WITH_C_API:-OFF} \
       -DWITH_PYTHON=${WITH_PYTHON:-ON} \
@@ -74,11 +74,11 @@ cat <<EOF
 Running unit tests ...
 ========================================
 EOF
+    ctest --output-on-failure
     # make install should also be test when unittest
     make install -j `nproc`
     pip install /usr/local/opt/paddle/share/wheels/*.whl
     paddle version
-    ctest --output-on-failure
 fi
 
 
@@ -130,7 +130,7 @@ fi
 
 # generate deb package for current build
 # FIXME(typhoonzero): should we remove paddle/scripts/deb ?
-if [[ ${WITH_DEB:-OFF} == "ON" ]]; then
+if [[ ${WITH_DEB:-ON} == "ON" ]]; then
     cat <<EOF
 ========================================
 Generating .deb package ...

paddle/string/CMakeLists.txt

@@ -2,3 +2,4 @@ cc_library(stringpiece SRCS piece.cc)
 cc_test(stringpiece_test SRCS piece_test.cc DEPS stringpiece glog gflags)
 
 cc_test(stringprintf_test SRCS printf_test.cc DEPS glog gflags)
+cc_test(to_string_test SRCS to_string_test.cc)

paddle/framework/details/lod_tensor.h → paddle/string/to_string.h

@@ -13,34 +13,28 @@
    limitations under the License. */
 
 #pragma once
-
-#include <memory>
+#include <sstream>
+#include <string>
 
 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 string {
+template <typename T>
+inline std::string to_string(T v) {
+  std::ostringstream sout;
+  sout << v;
+  return sout.str();
+}
+
+// Faster std::string/const char* type
+template <>
+inline std::string to_string(std::string v) {
+  return v;
+}
+
+template <>
+inline std::string to_string(const char* v) {
+  return std::string(v);
+}
+
+}  // namespace string
 }  // namespace paddle

paddle/string/to_string_test.cc

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/string/to_string.h"
+#include <gtest/gtest.h>
+
+constexpr char kOutputString[] = "User Defined Output";
+class UserDefinedClass {
+public:
+};
+
+std::ostream& operator<<(std::ostream& s, const UserDefinedClass& ins) {
+  s << kOutputString;
+  return s;
+}
+
+TEST(to_string, normal) {
+  using namespace paddle::string;
+  ASSERT_EQ("10", to_string(10));
+  ASSERT_EQ("abc", to_string("abc"));
+  ASSERT_EQ("1.2", to_string(1.2));
+}
+
+TEST(to_string, user_defined) {
+  using namespace paddle::string;
+  UserDefinedClass instance;
+  ASSERT_EQ(kOutputString, to_string(instance));
+}
\ No newline at end of file

paddle/trainer/NewRemoteParameterUpdater.cpp

@@ -66,28 +66,92 @@ void NewRemoteParameterUpdater::init(
   // from parameter server
   if (paddle_begin_init_params(parameterClient_)) {
     LOG(INFO) << "paddle_begin_init_params start";
+    // NOTE: convert V1 OptimizatioinConfig proto to V2 OptimizerConfig.
+    // This makes golang pserver compatible with handy V1 demos.
+    // TODO(wuyi): Refine or remove these ugly converting lines
+    OptimizerConfig optimizerConfigV2;
+    if (trainerConfig_.learning_method() == "momentum") {
+      optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::SGD);
+    } else if (trainerConfig_.learning_method() == "adagrad") {
+      optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adagrad);
+      optimizerConfigV2.mutable_adagrad()->set_epsilon(
+          trainerConfig_.ada_epsilon());
+    } else if (trainerConfig_.learning_method() == "adadelta") {
+      optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adagrad);
+      optimizerConfigV2.mutable_adadelta()->set_epsilon(
+          trainerConfig_.ada_epsilon());
+      optimizerConfigV2.mutable_adadelta()->set_rho(trainerConfig_.ada_rou());
+    } else if (trainerConfig_.learning_method() == "adam") {
+      optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::Adam);
+      optimizerConfigV2.mutable_adam()->set_beta_1(trainerConfig_.adam_beta1());
+      optimizerConfigV2.mutable_adam()->set_beta_2(trainerConfig_.adam_beta2());
+      optimizerConfigV2.mutable_adam()->set_epsilon(
+          trainerConfig_.adam_epsilon());
+    } else {
+      LOG(ERROR) << "got unsupported v1 optimizer config: "
+                 << trainerConfig_.learning_method();
+      optimizerConfigV2.set_optimizer(paddle::OptimizerConfig::SGD);
+    }
+
+    if (trainerConfig_.learning_rate_schedule() == "constant") {
+      optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Const);
+      optimizerConfigV2.mutable_const_lr()->set_learning_rate(
+          trainerConfig_.learning_rate());
+    } else if (trainerConfig_.learning_rate_schedule() == "linear") {
+      optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Linear);
+      optimizerConfigV2.mutable_linear_lr()->set_learning_rate(
+          trainerConfig_.learning_rate());
+      optimizerConfigV2.mutable_linear_lr()->set_lr_decay_a(
+          trainerConfig_.learning_rate_decay_a());
+      optimizerConfigV2.mutable_linear_lr()->set_lr_decay_b(
+          trainerConfig_.learning_rate_decay_b());
+    } else {
+      LOG(ERROR) << "got unsupported v1 learning_rate_schedule config: "
+                 << trainerConfig_.learning_rate_schedule() << ", set to const";
+      optimizerConfigV2.set_lr_policy(paddle::OptimizerConfig::Const);
+    }
+
+    // overwrite optimizerConfigV2 for per-parameter(layer) configs
     for (int i = 0; i < parameterSize(); ++i) {
       auto paramConfig = parameters_[i]->getConfig();
-      LOG(INFO) << "old param config: " << paramConfig.DebugString();
-      // FIXME(typhoonzero): convert old paramConfig to optimizerConfig
-      OptimizerConfig optimizeConfigV2;
-      auto sgdConfigV2 = optimizeConfigV2.mutable_sgd();
-      sgdConfigV2->set_momentum(paramConfig.momentum());
-      sgdConfigV2->set_decay(paramConfig.decay_rate());
-      optimizeConfigV2.set_lr_policy(paddle::OptimizerConfig::Const);
-      auto constlr = optimizeConfigV2.mutable_const_lr();
+      if (paramConfig.has_momentum() &&
+          trainerConfig_.learning_method() == "momentum") {
+        optimizerConfigV2.mutable_sgd()->set_momentum(paramConfig.momentum());
+      }
       if (paramConfig.has_learning_rate()) {
-        constlr->set_learning_rate(paramConfig.learning_rate());
-      } else {
-        constlr->set_learning_rate(trainerConfig_.learning_rate());
+        switch (optimizerConfigV2.lr_policy()) {
+          case 0:
+            optimizerConfigV2.mutable_const_lr()->set_learning_rate(
+                paramConfig.learning_rate());
+            break;
+          case 1:
+            optimizerConfigV2.mutable_linear_lr()->set_learning_rate(
+                paramConfig.learning_rate());
+            break;
+        }
       }
-      if (trainerConfig_.algorithm() == "sgd") {
-        optimizeConfigV2.set_optimizer(paddle::OptimizerConfig::SGD);
-        // FIXME: config all algorithms
-      } else {
-        optimizeConfigV2.set_optimizer(paddle::OptimizerConfig::SGD);
+      if (paramConfig.has_decay_rate()) {
+        switch (optimizerConfigV2.optimizer()) {
+          case 1:  // SGD
+            optimizerConfigV2.mutable_sgd()->set_decay(
+                paramConfig.decay_rate());
+            break;
+          case 2:  // Adadelta
+            optimizerConfigV2.mutable_adadelta()->set_decay(
+                paramConfig.decay_rate());
+            break;
+          case 3:  // Adagrad
+            optimizerConfigV2.mutable_adagrad()->set_decay(
+                paramConfig.decay_rate());
+            break;
+          case 4:  // Adam
+            optimizerConfigV2.mutable_adam()->set_decay(
+                paramConfig.decay_rate());
+            break;
+        }
       }
-      std::string bytes = optimizeConfigV2.SerializeAsString();
+      // send param and config to pserver
+      std::string bytes = optimizerConfigV2.SerializeAsString();
       const char *array = bytes.data();
       int size = (int)bytes.size();
       paddle_init_param(

paddle/trainer/TrainerConfigHelper.cpp

@@ -28,6 +28,8 @@ DECLARE_bool(with_cost);
 DECLARE_bool(with_gpu);
 DECLARE_bool(parallel_nn);
 DECLARE_string(config_args);
+DECLARE_bool(use_mkldnn);
+DECLARE_bool(use_mkldnn_wgt);
 
 const char *kConfigParserModuleName = "paddle.trainer.config_parser";
 const char *kConfigParserFuncName = "parse_config_and_serialize";
@@ -44,6 +46,8 @@ TrainerConfigHelper::TrainerConfigHelper(const std::string &configFilePath)
   configArgs << "trainer_id=" << FLAGS_trainer_id << ",local=" << FLAGS_local
              << ",with_cost=" << FLAGS_with_cost << ",use_gpu=" << FLAGS_use_gpu
              << ",parallel_nn=" << FLAGS_parallel_nn
+             << ",use_mkldnn=" << FLAGS_use_mkldnn
+             << ",use_mkldnn_wgt=" << FLAGS_use_mkldnn_wgt
              << ",cudnn_version=" << hl_get_cudnn_lib_version();
   if (!FLAGS_config_args.empty()) {
     configArgs << "," << FLAGS_config_args;

paddle/trainer/tests/CMakeLists.txt

@@ -2,19 +2,19 @@
 add_unittest_without_exec(test_Compare
     test_Compare.cpp)
 add_test(NAME test_Compare
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python
         ${CMAKE_CURRENT_BINARY_DIR}/test_Compare
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 
 ################# test_Trainer ###########################
 add_unittest_without_exec(test_Trainer
     test_Trainer.cpp)
 add_test(NAME test_Trainer
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
-        ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/paddle/trainer/tests/gen_proto_data.py &&
-        ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
+        ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/paddle/trainer/tests/gen_proto_data.py &&
+        ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
         ${CMAKE_CURRENT_BINARY_DIR}/test_Trainer
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 
 ############### test_TrainerOnePass ##########################
 if(WITH_PYTHON)
@@ -23,60 +23,60 @@ if(WITH_PYTHON)
   add_unittest_without_exec(test_TrainerOnePass
       test_TrainerOnePass.cpp)
   add_test(NAME test_TrainerOnePass
-    COMMAND  ${PROJ_ROOT}/paddle/.set_python_path.sh -d
-          ${PROJ_ROOT}/python/:${PROJ_ROOT}/paddle/trainer/tests
-          ${PROJ_ROOT}/paddle/.set_port.sh -p port ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass
-      WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+    COMMAND  ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
+          ${PADDLE_SOURCE_DIR}/python/:${PADDLE_SOURCE_DIR}/paddle/trainer/tests
+          ${PADDLE_SOURCE_DIR}/paddle/.set_port.sh -p port ${CMAKE_CURRENT_BINARY_DIR}/test_TrainerOnePass
+      WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 endif()
 ################ test_CompareTwoNets ######################
 add_unittest_without_exec(test_CompareTwoNets
     test_CompareTwoNets.cpp)
 add_test(NAME test_CompareTwoNets
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
         ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoNets
             --config_file_a=trainer/tests/sample_trainer_config_qb_rnn.conf --config_file_b=trainer/tests/sample_trainer_config_rnn.conf
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 
 ############### test_CompareTwoOpts ###################
 add_unittest_without_exec(test_CompareTwoOpts
     test_CompareTwoOpts.cpp)
 add_test(NAME test_CompareTwoOpts
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
         ${CMAKE_CURRENT_BINARY_DIR}/test_CompareTwoOpts
             --config_file_a=trainer/tests/sample_trainer_config_opt_a.conf --config_file_b=trainer/tests/sample_trainer_config_opt_b.conf
             --num_passes=1 --need_high_accuracy=0
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 
 ################# test_CompareSparse ##################
 add_unittest_without_exec(test_CompareSparse
     test_CompareSparse.cpp)
 if(NOT ON_TRAVIS)
   add_test(NAME test_CompareSparse
-    COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
+    COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
               ./.set_port.sh -p port -n 6
                   ${CMAKE_CURRENT_BINARY_DIR}/test_CompareSparse
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 endif()
 ################# test_recurrent_machine_generation ###############
 add_unittest_without_exec(test_recurrent_machine_generation
     test_recurrent_machine_generation.cpp)
 add_test(NAME test_recurrent_machine_generation
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
         ${CMAKE_CURRENT_BINARY_DIR}/test_recurrent_machine_generation
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 
 #################### test_PyDataProviderWrapper #########################
 add_unittest_without_exec(test_PyDataProviderWrapper
     test_PyDataProviderWrapper.cpp)
 
 add_test(NAME test_PyDataProviderWrapper
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d
-        ${PROJ_ROOT}/python/:${PROJ_ROOT}/paddle/trainer/tests
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d
+        ${PADDLE_SOURCE_DIR}/python/:${PADDLE_SOURCE_DIR}/paddle/trainer/tests
         ${CMAKE_CURRENT_BINARY_DIR}/test_PyDataProviderWrapper
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)
 
 #################### test_config_parser #########################
 add_test(NAME test_config_parser
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
-        ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/paddle/trainer/tests/config_parser_test.py
-    WORKING_DIRECTORY ${PROJ_ROOT}/paddle/)
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
+        ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/paddle/trainer/tests/config_parser_test.py
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/)

paddle/utils/Flags.cpp

@@ -20,6 +20,14 @@ DEFINE_bool(use_gpu, false, "Only support CPU training");
 DEFINE_bool(use_gpu, true, "Whether to use GPU for training");
 #endif
 
+#ifdef PADDLE_USE_MKLDNN
+// TODO(TJ): change to true when MKLDNN layers support multi-inputs
+DEFINE_bool(use_mkldnn, false, "Default still keep use CPU training");
+#else
+DEFINE_bool(use_mkldnn, false, "Only support CPU training");
+#endif
+
+DEFINE_bool(use_mkldnn_wgt, false, "Init weight from CPU weight");
 DEFINE_bool(parallel_nn,
             false,
             "Whether to use multi-threads to calculate one neural network."

paddle/utils/Flags.h

@@ -40,3 +40,5 @@ DECLARE_bool(show_layer_stat);
 DECLARE_string(predict_file);
 DECLARE_bool(prev_batch_state);
 DECLARE_string(init_model_path);
+DECLARE_bool(use_mkldnn);
+DECLARE_bool(use_mkldnn_wgt);

paddle/utils/tests/CMakeLists.txt

@@ -13,6 +13,6 @@ add_executable(
 link_paddle_exe(test_CustomStackTracePrint)
 if(NOT APPLE)
     add_test(NAME test_CustomStackTracePrint
-        COMMAND ${PROJ_ROOT}/paddle/utils/tests/test_CustomStackTracePrint.sh
+        COMMAND ${PADDLE_SOURCE_DIR}/paddle/utils/tests/test_CustomStackTracePrint.sh
         WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 endif()

proto/CMakeLists.txt

@@ -9,13 +9,13 @@ foreach(filename ${proto_filenames})
     get_filename_component(ABS_FIL ${filename} ABSOLUTE)
     get_filename_component(FIL_WE ${filename} NAME_WE)
     set(CUR_PROTO_GEN_PY
-            ${PROJ_ROOT}/paddle/python/paddle/proto/${FIL_WE}_pb2.py)
+            ${PADDLE_SOURCE_DIR}/paddle/python/paddle/proto/${FIL_WE}_pb2.py)
     set(PROTO_GEN_PY
             ${CUR_PROTO_GEN_PY}
             ${PROTO_GEN_PY})
     add_custom_command(OUTPUT ${CUR_PROTO_GEN_PY}
             COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
-            ARGS "--python_out=${PROJ_ROOT}/python/paddle/proto"
+            ARGS "--python_out=${PADDLE_SOURCE_DIR}/python/paddle/proto"
             "-I" ${CMAKE_CURRENT_SOURCE_DIR} ${ABS_FIL}
             DEPENDS ${ABS_FIL} protoc)
 endforeach()

python/CMakeLists.txt

-set(OUTPUT_DIR
-    "${CMAKE_CURRENT_BINARY_DIR}/build")
 
 file(GLOB TRAINER_PY_FILES . ./paddle/trainer/*.py)
 file(GLOB HELPERS_PY_FILES . ./paddle/trainer_config_helpers/*.py)
@@ -18,7 +16,7 @@ SET(COPY_PADDLE_MASTER "")
 if(WITH_GOLANG)
   SET(COPY_PADDLE_MASTER "copy_paddle_master")
   add_custom_command(TARGET ${COPY_PADDLE_MASTER}
-    COMMAND cp ${paddle_master_LIB_PATH} ${PROJ_ROOT}/python/paddle/v2/master/
+    COMMAND cp ${paddle_master_LIB_PATH} ${PADDLE_SOURCE_DIR}/python/paddle/v2/master/
     )
   add_dependencies(copy_paddle_master paddle_master)
 endif(WITH_GOLANG)
@@ -27,19 +25,21 @@ configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in
     ${CMAKE_CURRENT_BINARY_DIR}/setup.py)
 
 
-add_custom_command(OUTPUT ${PROJ_ROOT}/python/paddle/v2/framework/core.so
-        COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${PROJ_ROOT}/python/paddle/v2/framework/core.so
+add_custom_command(OUTPUT ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/core.so
+        COMMAND cmake -E copy $<TARGET_FILE:paddle_pybind> ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/core.so
         DEPENDS paddle_pybind)
-add_custom_target(copy_paddle_pybind ALL DEPENDS ${PROJ_ROOT}/python/paddle/v2/framework/core.so)
+add_custom_target(copy_paddle_pybind ALL DEPENDS ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/core.so)
 
 
-add_custom_command(OUTPUT ${OUTPUT_DIR}/.timestamp
+add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
     COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
-    COMMAND ${CMAKE_COMMAND} -E touch ${OUTPUT_DIR}/.timestamp
+    COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
+    COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
+    COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
     DEPENDS gen_proto_py copy_paddle_pybind framework_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
 
 add_custom_target(paddle_python ALL DEPENDS
-    ${OUTPUT_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model python_api_wheel)
+    ${PADDLE_PYTHON_BUILD_DIR}/.timestamp paddle_pserver_main paddle_trainer paddle_merge_model python_api_wheel)
 
 set(PADDLE_PYTHON_PACKAGE_DIR ${CMAKE_CURRENT_BINARY_DIR}/dist/)
 

python/paddle/trainer/config_parser.py

@@ -1604,6 +1604,8 @@ class MultiClassCrossEntropySelfNormCostLayer(LayerBase):
 
 @config_layer('fc')
 class FCLayer(LayerBase):
+    layer_type = 'fc'
+
     def __init__(self,
                  name,
                  size,
@@ -1611,14 +1613,27 @@ class FCLayer(LayerBase):
                  bias=True,
                  error_clipping_threshold=None,
                  **xargs):
-        super(FCLayer, self).__init__(name, 'fc', size, inputs=inputs, **xargs)
+        use_mkldnn = bool(int(g_command_config_args.get("use_mkldnn", 0)))
+        use_mkldnn_wgt = bool(
+            int(g_command_config_args.get("use_mkldnn_wgt", 0)))
+        if use_mkldnn:
+            self.layer_type = 'mkldnn_fc'
+            config_assert(
+                len(inputs) == 1,
+                "MkldnnFCLayer support one and only one input!")
+        super(FCLayer, self).__init__(
+            name, self.layer_type, size, inputs=inputs, **xargs)
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
             psize = self.config.size * input_layer.size
             dims = [input_layer.size, self.config.size]
             format = self.inputs[input_index].format
             sparse = format == "csr" or format == "csc"
-
+            if use_mkldnn:
+                config_assert(not sparse,
+                              "MkldnnFCLayer do not support sparse format yet")
+                if use_mkldnn_wgt:
+                    dims = [self.config.size, input_layer.size]
             if sparse:
                 psize = self.inputs[input_index].nnz
             else:
@@ -1631,6 +1646,11 @@ class FCLayer(LayerBase):
             self.config.error_clipping_threshold = error_clipping_threshold
 
 
+@config_layer('mkldnn_fc')
+class MkldnnFcLayer(FCLayer):
+    layer_type = 'mkldnn_fc'
+
+
 @config_layer('selective_fc')
 class SelectiveFCLayer(LayerBase):
     def __init__(self,

python/paddle/trainer_config_helpers/tests/CMakeLists.txt

 #################### test_config_parser #########################
 add_test(NAME layers_test
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
-        ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/layers_test.py
-    WORKING_DIRECTORY ${PROJ_ROOT}/python/paddle)
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
+        ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/python/paddle/trainer_config_helpers/tests/layers_test.py
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/python/paddle)
 
 add_test(NAME test_reset_hook
-  COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python/
-        ${PYTHON_EXECUTABLE} ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/test_reset_hook.py
-    WORKING_DIRECTORY ${PROJ_ROOT}/python/paddle)
+  COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python/
+        ${PYTHON_EXECUTABLE} ${PADDLE_SOURCE_DIR}/python/paddle/trainer_config_helpers/tests/test_reset_hook.py
+    WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/python/paddle)
 
 add_paddle_exe(protobuf_equal ProtobufEqualMain.cpp)
 add_test(NAME test_layerHelpers
   COMMAND
-  ${PROJ_ROOT}/python/paddle/trainer_config_helpers/tests/configs/run_tests.sh ${PYTHON_EXECUTABLE}
+  ${PADDLE_SOURCE_DIR}/python/paddle/trainer_config_helpers/tests/configs/run_tests.sh ${PYTHON_EXECUTABLE}
   ${CMAKE_CURRENT_BINARY_DIR}/protobuf_equal
 )

python/paddle/v2/framework/.gitignore

+proto

python/paddle/v2/framework/op.py

 import paddle.v2.framework.core as core
-import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2
-import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2
-import paddle.v2.framework.proto.attribute_pb2 as attribute_pb2
+import paddle.v2.framework.proto.framework_pb2 as framework_pb2
 
 
 def get_all_op_protos():
@@ -12,11 +10,15 @@ def get_all_op_protos():
     protostrs = core.get_all_op_protos()
     ret_values = []
     for pbstr in protostrs:
-        op_proto = op_proto_pb2.OpProto.FromString(str(pbstr))
+        op_proto = framework_pb2.OpProto.FromString(str(pbstr))
         ret_values.append(op_proto)
     return ret_values
 
 
+def is_str(s):
+    return isinstance(s, str) or isinstance(s, unicode)
+
+
 class OpDescCreationMethod(object):
     """
     A Functor object to convert user input(use key word args) to OpDesc based on
@@ -27,7 +29,7 @@ class OpDescCreationMethod(object):
     """
 
     def __init__(self, op_proto):
-        if not isinstance(op_proto, op_proto_pb2.OpProto):
+        if not isinstance(op_proto, framework_pb2.OpProto):
             raise TypeError("Argument should be OpProto")
         self.__op_proto__ = op_proto
 
@@ -39,26 +41,34 @@ class OpDescCreationMethod(object):
         """
         if len(args) != 0:
             raise ValueError("Only keyword arguments is supported by Paddle")
-        op_desc = op_desc_pb2.OpDesc()
-
-        # Inputs
-        ipts, ipt_format, _ = OpDescCreationMethod.extract_input_or_output(
-            "input", kwargs, self.__op_proto__.inputs)
-        op_desc.inputs.extend(ipts)
-        if ipt_format is not None:
-            op_desc.attrs.extend([ipt_format])
-
-        # Outputs
-        outs, out_format, tmp_index = OpDescCreationMethod.extract_input_or_output(
-            "output", kwargs, self.__op_proto__.outputs)
-        op_desc.outputs.extend(outs)
-        if out_format is not None:
-            op_desc.attrs.extend([out_format])
-        if len(tmp_index) != 0:
-            tmp_index_attr = op_desc.attrs.add()
-            tmp_index_attr.type = attribute_pb2.INTS
-            tmp_index_attr.name = "temporary_index"
-            tmp_index_attr.ints.extend(tmp_index)
+        op_desc = framework_pb2.OpDesc()
+
+        for input_parameter in self.__op_proto__.inputs:
+            input_arguments = kwargs.get(input_parameter.name, [])
+            if is_str(input_arguments):
+                input_arguments = [input_arguments]
+
+            if not input_parameter.duplicable and len(input_arguments) > 1:
+                raise ValueError("Input %s only accepts one input, but give %d"
+                                 % (input_parameter.name, len(input_arguments)))
+
+            ipt = op_desc.inputs.add()
+            ipt.parameter = input_parameter.name
+            ipt.arguments.extend(input_arguments)
+
+        for output_parameter in self.__op_proto__.outputs:
+            output_arguments = kwargs.get(output_parameter.name, [])
+            if is_str(output_arguments):
+                output_arguments = [output_arguments]
+
+            if not output_parameter.duplicable and len(output_arguments) > 1:
+                raise ValueError(
+                    "Output %s only accepts one output, but give %d" %
+                    (output_parameter.name, len(output_arguments)))
+
+            out = op_desc.outputs.add()
+            out.parameter = output_parameter.name
+            out.arguments.extend(output_arguments)
 
         # Types
         op_desc.type = self.__op_proto__.type
@@ -72,17 +82,17 @@ class OpDescCreationMethod(object):
                 new_attr = op_desc.attrs.add()
                 new_attr.name = attr.name
                 new_attr.type = attr.type
-                if attr.type == attribute_pb2.INT:
+                if attr.type == framework_pb2.INT:
                     new_attr.i = user_defined_attr
-                elif attr.type == attribute_pb2.FLOAT:
+                elif attr.type == framework_pb2.FLOAT:
                     new_attr.f = user_defined_attr
-                elif attr.type == attribute_pb2.STRING:
+                elif attr.type == framework_pb2.STRING:
                     new_attr.s = user_defined_attr
-                elif attr.type == attribute_pb2.INTS:
+                elif attr.type == framework_pb2.INTS:
                     new_attr.ints.extend(user_defined_attr)
-                elif attr.type == attribute_pb2.FLOATS:
+                elif attr.type == framework_pb2.FLOATS:
                     new_attr.floats.extend(user_defined_attr)
-                elif attr.type == attribute_pb2.STRINGS:
+                elif attr.type == framework_pb2.STRINGS:
                     new_attr.strings.extend(user_defined_attr)
                 else:
                     raise NotImplementedError("Not support attribute type " +
@@ -90,50 +100,6 @@ class OpDescCreationMethod(object):
 
         return op_desc
 
-    @staticmethod
-    def extract_input_or_output(in_out, kwargs, meta):
-        """
-        Extract input variable names or output variable names from key-word 
-        arguments, which base on VarProtos.
-        
-        :param in_out: "input" or "output"
-        :param kwargs: key-word arguments that user inputted.
-        :param meta: a list of VarProto
-        :return: The three object will be return. The variable names. The 
-        input_format or output_format attribute(None if the input or output is 
-        not multiple). The temporary variable index list.
-        """
-        multiple = OpDescCreationMethod.any_is_true((m.multiple for m in meta))
-        tmp_index = []
-        retv = []
-        if multiple:
-            var_format = op_desc_pb2.AttrDesc()
-            var_format.type = attribute_pb2.INTS
-            var_format.name = "%s_format" % in_out
-            var_format.ints.append(0)
-
-            for var in meta:
-                var_name = var.name
-
-                if var.temporary:
-                    var_name = [core.var_names.temp()]
-                    tmp_index.append(len(retv))
-                else:
-                    var_name = kwargs.get(var_name, [])
-                if not isinstance(var_name, list):
-                    var_name = [var_name]
-                retv.extend(var_name)
-                var_format.ints.append(len(var_name) + var_format.ints[-1])
-            return retv, var_format, tmp_index
-        else:
-            for var in meta:
-                if var.temporary:
-                    retv.append(kwargs.get(var.name, core.var_names.temp()))
-                    tmp_index.append(len(retv))
-                else:
-                    retv.append(kwargs.get(var.name, core.var_names.empty()))
-            return retv, None, tmp_index
-
     @staticmethod
     def any_is_true(generator):
         """
@@ -146,13 +112,12 @@ class OpDescCreationMethod(object):
 
 
 class OpInfo(object):
-    def __init__(self, name, method, inputs, outputs, attrs, no_temp_outputs):
+    def __init__(self, name, method, inputs, outputs, attrs):
         self.name = name
         self.method = method
         self.inputs = inputs
         self.outputs = outputs
         self.attrs = attrs
-        self.no_temp_outputs = no_temp_outputs
 
 
 def create_op_creation_method(op_proto):
@@ -170,10 +135,7 @@ def create_op_creation_method(op_proto):
         name=op_proto.type,
         inputs=[var.name for var in op_proto.inputs],
         outputs=[var.name for var in op_proto.outputs],
-        attrs=[attr.name for attr in op_proto.attrs],
-        no_temp_outputs=[
-            var.name for var in op_proto.outputs if not var.temporary
-        ])
+        attrs=[attr.name for attr in op_proto.attrs])
 
 
 class OperatorFactory(object):
@@ -214,8 +176,5 @@ class OperatorFactory(object):
     def get_op_attr_names(self, type):
         return self.get_op_info(type).attrs
 
-    def get_op_no_temp_output_names(self, type):
-        return self.get_op_info(type).no_temp_outputs
-
 
 Operator = OperatorFactory()  # Default global factory

python/paddle/v2/framework/tests/CMakeLists.txt

 py_test(test_net SRCS test_net.py)
 
-py_test(test_fc_op SRCS test_fc_op.py)
 py_test(test_scope SRCS test_scope.py)
 
 py_test(test_tensor SRCS test_tensor.py)
@@ -21,5 +20,8 @@ py_test(gradient_checker SRCS gradient_checker.py)
 py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py)
 
 py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py)
+
 py_test(test_operator SRCS test_operator.py)
+# py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
 py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
+py_test(test_recurrent_op SRCS test_recurrent_op.py)

python/paddle/v2/framework/tests/gradient_checker.py

@@ -53,15 +53,18 @@ def get_numeric_gradient(op,
         tensor.set(input_values[var_name], core.CPUPlace())
 
     # Create all output variable in local_scope
-    for output in op.outputs():
-        if local_scope.find_var(output) is None:
-            local_scope.new_var(output).get_tensor()
-
+    opts = op.outputs()
+    for key in opts:
+        for output in opts[key]:
+            if local_scope.find_var(output) is None:
+                local_scope.new_var(output).get_tensor()
     op.infer_shape(local_scope)
 
     # allocate output memory
-    for output in op.outputs():
-        local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace())
+    for key in opts:
+        for output in opts[key]:
+            local_scope.find_var(output).get_tensor().alloc_float(core.CPUPlace(
+            ))
 
     # TODO(yuyang18): Only CPU is support now.
     cpu_ctx = core.DeviceContext.create(core.CPUPlace())
@@ -73,34 +76,60 @@ def get_numeric_gradient(op,
     def product(dim):
         return reduce(lambda a, b: a * b, dim, 1)
 
+    # get the input tensor that we want to get it's numeric gradient.
     tensor_to_check = local_scope.find_var(input_to_check).get_tensor()
     tensor_size = product(tensor_to_check.get_dims())
+    # prepare a numpy array to store the gradient.
     gradient_flat = numpy.zeros(shape=(tensor_size, ), dtype='float32')
+
+    # 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()
 
-        tensor_to_check.set_float_element(i, origin)  # restore old value
+        # 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())
 
 
 class GradientChecker(unittest.TestCase):
-    def __is_close(self, numeric_grads, scope, max_relative_error):
+    def assert_is_close(self, numeric_grads, scope, max_relative_error,
+                        msg_prefix):
         for name in numeric_grads:
-            op_grad = numpy.array(
-                scope.find_var(grad_var_name(name)).get_tensor())
-            is_close = numpy.allclose(
-                numeric_grads[name], op_grad, rtol=max_relative_error, atol=100)
-            if not is_close:
-                return False
-        return True
+            b = numpy.array(scope.find_var(grad_var_name(name)).get_tensor())
+            a = numeric_grads[name]
+
+            abs_a = numpy.abs(a)
+            # if abs_a is nearly zero, then use abs error for a, not relative
+            # error.
+            abs_a[abs_a < 1e-3] = 1
+
+            diff_mat = numpy.abs(a - b) / abs_a
+            max_diff = numpy.max(diff_mat)
+
+            def err_msg():
+                offset = numpy.argmax(diff_mat > max_relative_error)
+                return "%s Variable %s max gradient diff %f over limit %f, the first " \
+                       "error element is %d" % (
+                       msg_prefix, name, max_diff, max_relative_error, offset)
+
+            self.assertLessEqual(max_diff, max_relative_error, err_msg())
 
     def check_grad(self,
                    forward_op,
@@ -124,19 +153,24 @@ class GradientChecker(unittest.TestCase):
         if no_grad_set is None:
             no_grad_set = set()
 
-        tmp_outs = forward_op.temp_outputs()
-        no_tmp_out = filter(lambda name: name not in tmp_outs,
-                            forward_op.outputs())
+        no_tmp_out = forward_op.no_intermediate_outputs()
         if len(no_tmp_out) != 1:
             raise ValueError("non temp out_names should be 1")
 
-        in_names = forward_op.inputs()
+        inputs = forward_op.inputs()
+        in_names = [item for k in inputs for item in inputs[k]]
+        outputs = forward_op.outputs()
+        out_names = [item for k in outputs for item in outputs[k]]
+
         for no_grad in no_grad_set:
             if no_grad not in in_names:
                 raise ValueError("no_grad should be in in_names")
 
         backward_op = core.Operator.backward(forward_op, no_grad_set)
 
+        bwd_outputs = backward_op.outputs()
+        bwd_out_names = [item for k in bwd_outputs for item in bwd_outputs[k]]
+
         places = [core.CPUPlace()]
         if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu():
             places.append(core.GPUPlace(0))
@@ -145,7 +179,8 @@ class GradientChecker(unittest.TestCase):
         # get numeric gradient
         for check_name in inputs_to_check:
             numeric_grad[check_name] = \
-                get_numeric_gradient(forward_op, input_vars, output_name, check_name)
+                get_numeric_gradient(forward_op, input_vars, output_name,
+                                     check_name)
 
         # get operator gradient according to different device
         for place in places:
@@ -161,7 +196,7 @@ class GradientChecker(unittest.TestCase):
                 var.set(value, place)
 
             # create output var
-            for out_name in forward_op.outputs():
+            for out_name in out_names:
                 scope.new_var(out_name).get_tensor()
 
             # infer the shape of output var and compute/set value of output var
@@ -171,7 +206,7 @@ class GradientChecker(unittest.TestCase):
             # create output grad var
             # set shape as the output var
             # set value of this grad to ones
-            for name in forward_op.outputs():
+            for name in out_names:
                 out_tensor = scope.find_var(name).get_tensor()
                 grad_tensor = scope.new_var(grad_var_name(name)).get_tensor()
                 grad_tensor.set_dims(out_tensor.shape())
@@ -179,7 +214,7 @@ class GradientChecker(unittest.TestCase):
                 grad_tensor.set(data, place)
 
             # create input grad var
-            for name in backward_op.outputs():
+            for name in bwd_out_names:
                 scope.new_var(name).get_tensor()
 
             # infer the shape of input gradient var and compute/set it's value
@@ -187,15 +222,8 @@ class GradientChecker(unittest.TestCase):
             backward_op.infer_shape(scope)
             backward_op.run(scope, ctx)
 
-            if isinstance(place, core.CPUPlace):
-                msg = "CPU kernel gradient is not close to numeric gradient"
-            else:
-                if isinstance(place, core.GPUPlace):
-                    msg = "GPU kernel gradient is not close to numeric gradient"
-                else:
-                    raise ValueError("unknown place " + type(place))
-            self.assertTrue(
-                self.__is_close(numeric_grad, scope, max_relative_error), msg)
+            self.assert_is_close(numeric_grad, scope, max_relative_error,
+                                 "Gradient Check On %s" % str(place))
 
 
 if __name__ == '__main__':

python/paddle/v2/framework/tests/test_add_two_op.py

@@ -19,14 +19,5 @@ class TestAddOp(unittest.TestCase):
         self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
 
 
-class TestAddGradOp(unittest.TestCase):
-    def test_add_grad(self):
-        op = Operator('add_two', X="X", Y="Y", Out="Out")
-        backward_op = core.Operator.backward(op, set())
-        self.assertEqual(backward_op.type(), "add_two_grad")
-        expected = '''Op(add_two_grad), inputs:(X, Y, Out, Out@GRAD), outputs:(X@GRAD, Y@GRAD).'''
-        self.assertEqual(expected, str(backward_op))
-
-
 if __name__ == '__main__':
     unittest.main()

python/paddle/v2/framework/tests/test_fc_op.py

-import paddle.v2.framework.core as core
-import unittest
-import numpy
-from paddle.v2.framework.op import Operator
-
-
-class TestFc(unittest.TestCase):
-    def test_fc(self):
-        scope = core.Scope()
-        place = core.CPUPlace()
-        x = scope.new_var("X")
-
-        x_tensor = x.get_tensor()
-        x_tensor.set_dims([1000, 784])
-        x_tensor.alloc_float(place)
-
-        w = scope.new_var("W")
-        w_tensor = w.get_tensor()
-        w_tensor.set_dims([784, 100])
-        w_tensor.alloc_float(place)
-
-        w_tensor.set(numpy.random.random((784, 100)).astype("float32"), place)
-
-        # Set a real numpy array here.
-        # x_tensor.set(numpy.array([]))
-
-        op = Operator("fc", X="X", Y="Y", W="W")
-
-        for out in op.outputs():
-            if scope.find_var(out) is None:
-                scope.new_var(out).get_tensor()
-
-        tensor = scope.find_var("Y").get_tensor()
-        op.infer_shape(scope)
-        self.assertEqual([1000, 100], tensor.shape())
-
-        ctx = core.DeviceContext.create(place)
-
-        op.run(scope, ctx)
-
-        # After complete all ops, check Y is expect or not.
-
-
-if __name__ == '__main__':
-    unittest.main()

python/paddle/v2/framework/tests/test_gaussian_random_op.py

+import unittest
+import paddle.v2.framework.core as core
+from paddle.v2.framework.op import Operator
+import numpy
+
+
+class GaussianRandomTest(unittest.TestCase):
+    def test_cpu(self):
+        self.gaussian_random_test(place=core.CPUPlace())
+
+    def test_gpu(self):
+        if core.is_compile_gpu():
+            self.gaussian_random_test(place=core.GPUPlace(0))
+
+    def gaussian_random_test(self, place):
+        scope = core.Scope()
+        scope.new_var("Out").get_tensor()
+
+        op = Operator(
+            "gaussian_random",
+            Out="Out",
+            dims=[1000, 784],
+            mean=.0,
+            std=1.,
+            seed=10)
+
+        op.infer_shape(scope)
+        context = core.DeviceContext.create(place)
+        op.run(scope, context)
+        tensor = numpy.array(scope.find_var("Out").get_tensor())
+        self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1)
+        self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_net.py

@@ -3,6 +3,15 @@ from paddle.v2.framework.op import Operator
 import unittest
 
 
+def fc(X, W, Y):
+    ret_v = core.Net.create()
+
+    ret_v.add_op(Operator("mul", X="X", Y="W", Out="pre_activation"))
+    ret_v.add_op(Operator("sigmoid", X="pre_activation", Y=Y))
+    ret_v.complete_add_op(True)
+    return ret_v
+
+
 class TestNet(unittest.TestCase):
     def test_net_all(self):
         net = core.Net.create()
@@ -10,18 +19,18 @@ class TestNet(unittest.TestCase):
         net.add_op(op1)
 
         net2 = core.Net.create()
-        net2.add_op(Operator("fc", X="X", W="w", Y="fc.out"))
+        net2.add_op(fc(X="X", W="w", Y="fc.out"))
         net2.complete_add_op(True)
         net.add_op(net2)
         net.complete_add_op(True)
 
         expected = '''
-Op(plain_net), inputs:(@EMPTY@, X, Y, w), outputs:(@TEMP@fc@0, Out, fc.out).
-    Op(add_two), inputs:(X, Y), outputs:(Out).
-    Op(plain_net), inputs:(@EMPTY@, X, w), outputs:(@TEMP@fc@0, fc.out).
-        Op(fc), inputs:(X, w, @EMPTY@), outputs:(fc.out, @TEMP@fc@0).
-            Op(mul), inputs:(X, w), outputs:(@TEMP@fc@0).
-            Op(sigmoid), inputs:(@TEMP@fc@0), outputs:(fc.out).
+Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}.
+    Op(add_two), inputs:{X[X], Y[Y]}, outputs:{Out[Out]}.
+    Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
+        Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
+            Op(mul), inputs:{X[X], Y[W]}, outputs:{Out[pre_activation]}.
+            Op(sigmoid), inputs:{X[pre_activation]}, outputs:{Y[fc.out]}.
 '''
         self.assertEqual(expected, "\n" + str(net))
 

python/paddle/v2/framework/tests/test_operator.py

 import unittest
 import paddle.v2.framework.op as op
 import paddle.v2.framework.core as core
-import paddle.v2.framework.proto.op_proto_pb2 as op_proto_pb2
-import paddle.v2.framework.proto.op_desc_pb2 as op_desc_pb2
-import paddle.v2.framework.proto.attribute_pb2 as attribute_pb2
+import paddle.v2.framework.proto.framework_pb2 as framework_pb2
 
 
 class TestGetAllProtos(unittest.TestCase):
@@ -17,7 +15,7 @@ class TestGetAllProtos(unittest.TestCase):
 
 class TestOpDescCreationMethod(unittest.TestCase):
     def test_plain_input_output(self):
-        op_proto = op_proto_pb2.OpProto()
+        op_proto = framework_pb2.OpProto()
         op_proto.type = "test"
         ipt = op_proto.inputs.add()
         ipt.name = "X"
@@ -37,25 +35,32 @@ class TestOpDescCreationMethod(unittest.TestCase):
 
         method = op.OpDescCreationMethod(op_proto)
         output = method(X="a", Y="b", Z="c")
-
-        expected = op_desc_pb2.OpDesc()
+        expected = framework_pb2.OpDesc()
         expected.type = "test"
-        expected.inputs.extend(["a", "b"])
-        expected.outputs.append("c")
+        ipt_0 = expected.inputs.add()
+        ipt_0.parameter = "X"
+        ipt_0.arguments.extend(["a"])
+        ipt_1 = expected.inputs.add()
+        ipt_1.parameter = 'Y'
+        ipt_1.arguments.extend(['b'])
+        opt = expected.outputs.add()
+        opt.parameter = "Z"
+        opt.arguments.extend(["c"])
+
         self.assertEqual(expected, output)
 
     def test_multiple_input_plain_output(self):
-        op_proto = op_proto_pb2.OpProto()
+        op_proto = framework_pb2.OpProto()
         op_proto.type = "fc"
         ipt = op_proto.inputs.add()
         ipt.name = "X"
         ipt.comment = ""
-        ipt.multiple = True
+        ipt.duplicable = True
 
         ipt = op_proto.inputs.add()
         ipt.name = "W"
         ipt.comment = ""
-        ipt.multiple = True
+        ipt.duplicable = True
 
         ipt = op_proto.inputs.add()
         ipt.name = "b"
@@ -70,30 +75,50 @@ class TestOpDescCreationMethod(unittest.TestCase):
         method = op.OpDescCreationMethod(op_proto)
 
         generated1 = method(X="x", W="w", b="b", Y="y")
-        expected1 = op_desc_pb2.OpDesc()
-        expected1.inputs.extend(['x', 'w', 'b'])
-        expected1.outputs.extend(['y'])
+        expected1 = framework_pb2.OpDesc()
+        tmp = expected1.inputs.add()
+        tmp.parameter = "X"
+        tmp.arguments.extend(['x'])
+
+        tmp = expected1.inputs.add()
+        tmp.parameter = 'W'
+        tmp.arguments.extend(['w'])
+
+        tmp = expected1.inputs.add()
+        tmp.parameter = 'b'
+        tmp.arguments.extend(['b'])
+
+        tmp = expected1.outputs.add()
+        tmp.parameter = 'Y'
+        tmp.arguments.extend(['y'])
         expected1.type = 'fc'
-        attr = expected1.attrs.add()
-        attr.name = 'input_format'
-        attr.type = attribute_pb2.INTS
-        attr.ints.extend([0, 1, 2, 3])
         self.assertEqual(expected1, generated1)
 
         generated2 = method(
             X=['x1', 'x2', 'x3'], b='b', W=['w1', 'w2', 'w3'], Y='y')
-        expected2 = op_desc_pb2.OpDesc()
-        expected2.inputs.extend(['x1', 'x2', 'x3', 'w1', 'w2', 'w3', 'b'])
-        expected2.outputs.extend(['y'])
+        expected2 = framework_pb2.OpDesc()
+
+        tmp = expected2.inputs.add()
+        tmp.parameter = "X"
+        tmp.arguments.extend(['x1', 'x2', 'x3'])
+
+        tmp = expected2.inputs.add()
+        tmp.parameter = 'W'
+        tmp.arguments.extend(['w1', 'w2', 'w3'])
+
+        tmp = expected2.inputs.add()
+        tmp.parameter = 'b'
+        tmp.arguments.extend(['b'])
+
+        tmp = expected2.outputs.add()
+        tmp.parameter = 'Y'
+        tmp.arguments.extend(['y'])
+
         expected2.type = 'fc'
-        attr = expected2.attrs.add()
-        attr.name = 'input_format'
-        attr.type = attribute_pb2.INTS
-        attr.ints.extend([0, 3, 6, 7])
         self.assertEqual(expected2, generated2)
 
     def test_attrs(self):
-        op_proto = op_proto_pb2.OpProto()
+        op_proto = framework_pb2.OpProto()
         op_proto.type = "test"
         ipt = op_proto.inputs.add()
         ipt.name = 'X'
@@ -105,12 +130,12 @@ class TestOpDescCreationMethod(unittest.TestCase):
             attr.comment = ""
             attr.type = type
 
-        __add_attr__("int_attr", attribute_pb2.INT)
-        __add_attr__("float_attr", attribute_pb2.FLOAT)
-        __add_attr__("string_attr", attribute_pb2.STRING)
-        __add_attr__("ints_attr", attribute_pb2.INTS)
-        __add_attr__("floats_attr", attribute_pb2.FLOATS)
-        __add_attr__("strings_attr", attribute_pb2.STRINGS)
+        __add_attr__("int_attr", framework_pb2.INT)
+        __add_attr__("float_attr", framework_pb2.FLOAT)
+        __add_attr__("string_attr", framework_pb2.STRING)
+        __add_attr__("ints_attr", framework_pb2.INTS)
+        __add_attr__("floats_attr", framework_pb2.FLOATS)
+        __add_attr__("strings_attr", framework_pb2.STRINGS)
 
         op_proto.comment = ""
         self.assertTrue(op_proto.IsInitialized())
@@ -126,76 +151,52 @@ class TestOpDescCreationMethod(unittest.TestCase):
             floats_attr=[0.2, 3.2, 4.5],
             strings_attr=["a", "b", "c"])
 
-        expected = op_desc_pb2.OpDesc()
+        expected = framework_pb2.OpDesc()
         expected.type = "test"
-        expected.inputs.extend(['a'])
+
+        ipt = expected.inputs.add()
+        ipt.parameter = "X"
+        ipt.arguments.extend(['a'])
+
         attr = expected.attrs.add()
         attr.name = "int_attr"
-        attr.type = attribute_pb2.INT
+        attr.type = framework_pb2.INT
         attr.i = 10
 
         attr = expected.attrs.add()
         attr.name = "float_attr"
-        attr.type = attribute_pb2.FLOAT
+        attr.type = framework_pb2.FLOAT
         attr.f = 3.2
 
         attr = expected.attrs.add()
         attr.name = "string_attr"
-        attr.type = attribute_pb2.STRING
+        attr.type = framework_pb2.STRING
         attr.s = "test_str"
 
         attr = expected.attrs.add()
         attr.name = "ints_attr"
-        attr.type = attribute_pb2.INTS
+        attr.type = framework_pb2.INTS
         attr.ints.extend([0, 1, 2, 3, 4])
 
         attr = expected.attrs.add()
         attr.name = 'floats_attr'
-        attr.type = attribute_pb2.FLOATS
+        attr.type = framework_pb2.FLOATS
         attr.floats.extend([0.2, 3.2, 4.5])
 
         attr = expected.attrs.add()
         attr.name = 'strings_attr'
-        attr.type = attribute_pb2.STRINGS
+        attr.type = framework_pb2.STRINGS
         attr.strings.extend(['a', 'b', 'c'])
 
         self.assertEqual(expected, generated)
 
-    def test_input_temporary_output(self):
-        op_proto = op_proto_pb2.OpProto()
-        op_proto.type = "test"
-        out = op_proto.outputs.add()
-        out.name = "OUT"
-        out.comment = ""
-
-        out = op_proto.outputs.add()
-        out.name = "TMP"
-        out.comment = ""
-        out.temporary = True
-
-        out = op_proto.outputs.add()
-        out.name = "OUT2"
-        out.comment = ""
-        op_proto.comment = ""
-
-        method = op.OpDescCreationMethod(op_proto)
-        generated = method(OUT="a", OUT2="b")
-        desc = op_desc_pb2.OpDesc()
-        desc.outputs.extend(["a", core.var_names.temp(), "b"])
-        desc.type = "test"
-        attr = desc.attrs.add()
-        attr.name = "temporary_index"
-        attr.type = attribute_pb2.INTS
-        attr.ints.append(2)
-        self.assertEqual(generated, desc)
-
 
 class TestOpCreations(unittest.TestCase):
     def test_all(self):
         add_op = op.Operator("add_two", X="a", Y="b", Out="z")
         self.assertIsNotNone(add_op)
         # Invoke C++ DebugString()
-        self.assertEqual('Op(add_two), inputs:(a, b), outputs:(z).',
+        self.assertEqual('Op(add_two), inputs:{X[a], Y[b]}, outputs:{Out[z]}.',
                          str(add_op))
 
 

python/paddle/v2/framework/tests/test_protobuf.py

-import paddle.v2.framework.proto.op_proto_pb2 as op_proto_lib
-import paddle.v2.framework.proto.attribute_pb2 as attr_type_lib
+import paddle.v2.framework.proto.framework_pb2 as framework_pb2
 import unittest
 
 
 class TestFrameworkProto(unittest.TestCase):
     def test_all(self):
-        op_proto = op_proto_lib.OpProto()
+        op_proto = framework_pb2.OpProto()
         ipt0 = op_proto.inputs.add()
         ipt0.name = "a"
         ipt0.comment = "the input of cosine op"
@@ -19,7 +18,7 @@ class TestFrameworkProto(unittest.TestCase):
         attr = op_proto.attrs.add()
         attr.name = "scale"
         attr.comment = "scale of cosine op"
-        attr.type = attr_type_lib.FLOAT
+        attr.type = framework_pb2.FLOAT
         op_proto.type = "cos"
         self.assertTrue(op_proto.IsInitialized())
 

python/paddle/v2/framework/tests/test_recurrent_op.py

@@ -2,19 +2,74 @@ import logging
 import paddle.v2.framework.core as core
 import unittest
 import numpy as np
-import paddle.v2.framework.create_op_creation_methods as creation
+from paddle.v2.framework.op import Operator
 
-ops = creation.op_creations
 
+def py_sigmoid(x):
+    return 1. / (1. + np.exp(-x))
 
-def create_tensor(scope, name, shape):
+
+class PySimpleRNN(object):
+    '''
+    A simple implementation of RNN based on numpy, to futhur test RecurrentOp's alogorithm
+    '''
+
+    def __init__(self, input_dim=30, batch_size=50, weight_dim=15, sent_len=11):
+        self.x = np.random.normal(size=(sent_len, batch_size, input_dim))
+        self.W = np.random.normal(size=(input_dim, input_dim))
+        self.U = np.random.normal(size=(input_dim, input_dim))
+        self.h_boot = np.random.normal(size=(batch_size, input_dim))
+
+        # memories
+        self.mems = [
+            np.zeros(shape=(batch_size, input_dim)) for i in range(sent_len)
+        ]
+
+    def forward(self):
+        xs = self.segment_inputs()
+        for step_id in range(self.x.shape[0]):
+            self.step(step_id, xs[step_id])
+        return self.concat_outputs()
+
+    def segment_inputs(self):
+        return [self.x[i] for i in range(self.x.shape[0])]
+
+    def concat_outputs(self):
+        return np.array(self.mems)
+
+    def step(self, step_id, x):
+        '''
+        run a step
+        '''
+        mem = self.mems[step_id]
+        if step_id > 0:
+            pre_mem = self.mems[step_id - 1]
+        else:
+            pre_mem = self.h_boot
+        xW = np.matmul(x, self.W)
+        hU = np.matmul(mem, self.U)
+
+        sum = xW + hU
+        self.mems[step_id] = py_sigmoid(sum)
+
+
+class PySimpleRNNTest(unittest.TestCase):
+    def setUp(self):
+        self.rnn = PySimpleRNN()
+
+    def test_forward(self):
+        output = self.rnn.forward()
+        print 'output', output
+
+
+def create_tensor(scope, name, shape, np_data):
     tensor = scope.new_var(name).get_tensor()
     tensor.set_dims(shape)
-    tensor.set(np.random.random(shape), core.CPUPlace())
+    tensor.set(np_data, core.CPUPlace())
     return tensor
 
 
-class TestRNN(unittest.TestCase):
+class TestRecurrentOp(unittest.TestCase):
     '''
     Test RNNOp
 
@@ -28,7 +83,7 @@ class TestRNN(unittest.TestCase):
     memories:
         - h
     outputs:
-        - h
+       - h
     '''
 
     input_dim = 30
@@ -36,33 +91,45 @@ class TestRNN(unittest.TestCase):
     weight_dim = 15
     sent_len = 11
 
-    def init(self):
+    def setUp(self):
+        self.py_rnn = PySimpleRNN(self.input_dim, self.batch_size,
+                                  self.weight_dim, self.sent_len)
 
+    def forward(self):
         self.scope = core.Scope()
-
         self.create_global_variables()
         self.create_step_net()
         rnn_op = self.create_rnn_op()
         ctx = core.DeviceContext.create(core.CPUPlace())
-        print 'infer_shape'
         rnn_op.infer_shape(self.scope)
-
         rnn_op.run(self.scope, ctx)
+        return np.array(self.scope.find_var("h").get_tensor())
 
     def create_global_variables(self):
         # create inlink
+        x_np_data = self.py_rnn.x
         create_tensor(self.scope, "x",
-                      [self.sent_len, self.batch_size, self.input_dim])
-        create_tensor(self.scope, "W", [self.input_dim, self.input_dim])
-        create_tensor(self.scope, "U", [self.input_dim, self.input_dim])
-        create_tensor(self.scope, "h_boot", [self.batch_size, self.input_dim])
+                      [self.sent_len, self.batch_size, self.input_dim],
+                      x_np_data)
+        W_np_data = self.py_rnn.W
+        create_tensor(self.scope, "W", [self.input_dim, self.input_dim],
+                      W_np_data)
+
+        U_np_data = self.py_rnn.U
+        create_tensor(self.scope, "U", [self.input_dim, self.input_dim],
+                      U_np_data)
+
+        h_boot_np_data = self.py_rnn.h_boot
+        create_tensor(self.scope, "h_boot", [self.batch_size, self.input_dim],
+                      h_boot_np_data)
         self.scope.new_var("step_scopes")
         self.scope.new_var("h@alias")
         self.scope.new_var("h")
 
     def create_rnn_op(self):
         # create RNNOp
-        rnnop = ops.recurrent_op(
+        rnnop = Operator(
+            "recurrent_op",
             # inputs
             inlinks=["x"],
             boot_memories=["h_boot"],
@@ -81,17 +148,25 @@ class TestRNN(unittest.TestCase):
         var = self.scope.new_var("stepnet")
         stepnet = var.get_net()
 
-        x_fc_op = ops.fc(X="x@alias", W="W", Y="Wx")
-        h_fc_op = ops.fc(X="h@pre", W="U", Y="Uh")
-        sum_op = ops.add_two(X="Wx", Y="Uh", Out="sum")
-        sig_op = ops.sigmoid(X="sum", Y="h@alias")
+        # x_fc_op = Operator("fc", X="x@alias", W="W", Y="Wx")
+        # h_fc_op = Operator("fc", X="h@pre", W="U", Y="Uh")
+        x_fc_op = Operator("mul", X="x@alias", Y="W", Out="Wx")
+        h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
+        sum_op = Operator("add_two", X="Wx", Y="Uh", Out="sum")
+        sig_op = Operator("sigmoid", X="sum", Y="h@alias")
 
         for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
             stepnet.add_op(op)
         stepnet.complete_add_op(True)
 
-    def test_recurrent(self):
-        self.init()
+    def test_forward(self):
+        print 'test recurrent op forward'
+        pd_output = self.forward()
+        py_output = self.py_rnn.forward()
+        print 'pd_output', pd_output
+        print
+        print 'py_output', py_output
+        self.assertEqual(pd_output.shape, py_output.shape)
 
 
 if __name__ == '__main__':

python/paddle/v2/optimizer.py

-import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils
-import paddle.trainer_config_helpers.optimizers as v1_optimizers
+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 """
 Optimizers(update equation) for SGD method.
 
-TODO(zhihong) : create new optimizer with proto config, add new optimizer here
-
 TODO(yuyang18): Complete comments.
 """
 
+import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils
+import paddle.trainer_config_helpers.optimizers as v1_optimizers
+from paddle.proto.OptimizerConfig_pb2 import OptimizerConfig
+
 __all__ = [
     'Momentum', 'Adam', 'Adamax', 'AdaGrad', 'DecayedAdaGrad', 'AdaDelta',
     'RMSProp', 'ModelAverage', 'L2Regularization'
@@ -70,7 +83,8 @@ class Optimizer(object):
                         gradient_machine.prefetch(in_args)
                         parameter_updater.getParametersRemote()
 
-        :param pserver_spec: pserver location, eg: localhost:3000
+        :param pserver_spec: pserver location, eg: localhost:3000, if use etcd,
+        pserver_spec should be the etcd endpoints, eg: http://localhost:2379
         :return: parameter_updater
         """
         if is_local:

python/paddle/v2/parameters.py

+# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import numpy as np
 from paddle.proto.ParameterConfig_pb2 import ParameterConfig
 import paddle.trainer.config_parser as cp
@@ -113,16 +127,7 @@ class Parameters(object):
         """
         return iter(self.__param_conf__)
 
-    def __getitem__(self, key):
-        """
-        Get parameter by parameter name. It uses Python dict syntax.
-
-        :note: It will always copy the parameter from C++ side.
-        :param key: Parameter name
-        :type key: basestring
-        :return: parameter value
-        :rtype: np.ndarray
-        """
+    def __getter_inner(self, key, param_type):
         import py_paddle.swig_paddle as api
         shape = self.get_shape(key)
 
@@ -138,7 +143,7 @@ class Parameters(object):
                     each_gradient_machine, key)
                 # for simplify implementation now, we always copy from C++
                 assert isinstance(param, api.Parameter)
-                val = param.getBuf(api.PARAMETER_VALUE)
+                val = param.getBuf(param_type)
                 assert isinstance(val, api.Vector)
                 val = val.copyToNumpyArray()
                 return val
@@ -146,6 +151,19 @@ class Parameters(object):
 
             raise RuntimeError("Unexpected branch")
 
+    def __getitem__(self, key):
+        """
+        Get parameter by parameter name. It uses Python dict syntax.
+
+        :note: It will always copy the parameter from C++ side.
+        :param key: Parameter name
+        :type key: basestring
+        :return: parameter value
+        :rtype: np.ndarray
+        """
+        import py_paddle.swig_paddle as api
+        return self.__getter_inner(key, api.PARAMETER_VALUE)
+
     def get_shape(self, key):
         """
         get shape of the parameter.
@@ -202,6 +220,19 @@ class Parameters(object):
         """
         return self.__getitem__(key=parameter_name)
 
+    def get_grad(self, key):
+        """
+        Get grandient by parameter name.
+
+        :note: It will always copy the parameter from C++ side.
+        :param key: parameter name
+        :type key: basestring
+        :return: The grandient matrix.
+        :rtype: np.ndarray
+        """
+        import py_paddle.swig_paddle as api
+        return self.__getter_inner(key, api.PARAMETER_GRADIENT)
+
     def set(self, parameter_name, value):
         """
         Set parameter by parameter name & matrix.
@@ -250,7 +281,13 @@ class Parameters(object):
         size = reduce(lambda a, b: a * b, param.shape)
         f.write(struct.pack("IIQ", 0, 4, size))
         param = param.astype(np.float32)
-        f.write(param.tostring())
+        s = param.tostring()
+        wrote_size = 0
+        buf = buffer(s, wrote_size, 65535)
+        while buf:  # f.write crashes with big data blog.
+            f.write(buf)
+            wrote_size += 65535
+            buf = buffer(s, wrote_size, 65535)
 
     def deserialize(self, name, f):
         """

python/paddle/v2/trainer.py

@@ -161,14 +161,14 @@ class SGD(object):
                     self.__parameter_updater__.update(each_param)
                 cost_sum = out_args.sum()
                 cost = cost_sum / len(data_batch)
-                self.__parameter_updater__.finishBatch(cost)
-                batch_evaluator.finish()
                 event_handler(
                     v2_event.EndIteration(
                         pass_id=pass_id,
                         batch_id=batch_id,
                         cost=cost,
                         evaluator=batch_evaluator))
+                self.__parameter_updater__.finishBatch(cost)
+                batch_evaluator.finish()
 
             self.__parameter_updater__.finishPass()
             pass_evaluator.finish()

python/requirements.txt

+requests==2.9.2
+numpy>=1.12
+protobuf==3.1
+recordio
+matplotlib
+rarfile
+scipy>=0.19.0
+Pillow
+nltk>=3.2.2

python/setup.py.in

 from setuptools import setup, Distribution
-
 class BinaryDistribution(Distribution):
     def has_ext_modules(foo):
         return True
@@ -18,15 +17,8 @@ packages=['paddle',
           'paddle.v2.framework.proto',
           'py_paddle']
 
-setup_requires=["requests",
-                "numpy>=1.12",
-                "protobuf==3.1",
-                "recordio",
-                "matplotlib",
-                "rarfile",
-                "scipy>=0.19.0",
-                "Pillow",
-                "nltk>=3.2.2"]
+with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f:
+    setup_requires = f.read().splitlines()
 
 if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
     setup_requires+=["opencv-python"]
@@ -45,14 +37,14 @@ setup(name='paddlepaddle',
           '': '${CMAKE_CURRENT_SOURCE_DIR}',
           # The paddle.v2.framework.proto will be generated while compiling.
           # So that package points to other directory.
-          'paddle.v2.framework.proto': '${PROJ_BINARY_ROOT}/paddle/framework',
-          'py_paddle': '${PROJ_ROOT}/paddle/py_paddle'
+          'paddle.v2.framework.proto': '${PADDLE_BINARY_DIR}/paddle/framework',
+          'py_paddle': '${PADDLE_SOURCE_DIR}/paddle/py_paddle'
       },
-      scripts=['${PROJ_BINARY_ROOT}/paddle/scripts/paddle'],
+      scripts=['${PADDLE_BINARY_DIR}/paddle/scripts/paddle'],
       distclass=BinaryDistribution,
       data_files=[('/usr/local/opt/paddle/bin',
-                       ['${PROJ_BINARY_ROOT}/paddle/scripts/paddle_usage',
-                        '${PROJ_BINARY_ROOT}/paddle/trainer/paddle_trainer',
-                        '${PROJ_BINARY_ROOT}/paddle/trainer/paddle_merge_model',
-                        '${PROJ_BINARY_ROOT}/paddle/pserver/paddle_pserver_main'])]
+                       ['${PADDLE_BINARY_DIR}/paddle/scripts/paddle_usage',
+                        '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer',
+                        '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model',
+                        '${PADDLE_BINARY_DIR}/paddle/pserver/paddle_pserver_main'])]
 )