Merge remote-tracking branch 'baidu/develop' into port_blas

.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)
 
@@ -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)

Dockerfile

@@ -71,6 +71,21 @@ RUN pip install --upgrade pip && \
 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

@@ -9,11 +9,13 @@ function(CheckCompilerCXX11Flag)
         if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
             message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
         endif()
+        if(NOT ANDROID)
             # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
             # Use Debug mode instead for now.
             if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9)
                 set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
             endif()
+        endif()
     elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
         # cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang"
         # Apple Clang is a different compiler than upstream Clang which havs different version numbers.

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

@@ -7,14 +7,17 @@
   do {                                                                   \
     paddle_error __err__ = stmt;                                         \
     if (__err__ != kPD_NO_ERROR) {                                       \
-      fprintf(stderr, "Invoke paddle error %d \n" #stmt, __err__); \
+      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)
@@ -36,8 +36,8 @@ py_proto_compile(framework_py_proto SRCS attribute.proto op_proto.proto op_desc.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
 add_custom_command(TARGET framework_py_proto POST_BUILD
-    COMMAND ${CMAKE_COMMAND} -E make_directory ${PROJ_ROOT}/python/paddle/v2/framework/proto
-    COMMAND cp *.py ${PROJ_ROOT}/python/paddle/v2/framework/proto/
+    COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto
+    COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto/
     COMMENT "Copy generated python proto into directory paddle/v2/framework/proto."
     WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 
@@ -48,9 +48,12 @@ if(WITH_PYTHON)
 cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
-    fc_op
     sgd_op
     add_op
+    mul_op
+    rowwise_add_op
+    sigmoid_op
+    softmax_op
     mean_op
     cross_entropy_op
     recurrent_op

paddle/framework/backward_test.cc

@@ -30,6 +30,8 @@ using DeviceContext = platform::DeviceContext;
 
 class EmptyOp : public OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(EmptyOp, OperatorBase)
+
   void InferShape(const Scope &scope) const override {}
   void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
 };

paddle/framework/ddim.h

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

paddle/framework/details/lod_tensor.cc

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

paddle/framework/details/lod_tensor.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-   http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License. */
-
-#pragma once
-
-#include <memory>
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-/*
- * Slice levels from LOD.
- *
- * @lod: LOD to slice.
- * @level_begin: level to begin slice.
- * @level_end: level to end slice.
- */
-std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
-                                         size_t level_begin, size_t level_end);
-
-/*
- * Slice elements from a level of LOD.
- *
- * @lod: LOD to slice.
- * @level: which level to slice.
- * @elem_begin: element's index to begin slice.
- * @elem_end: element's index to end slice.
- */
-std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
-                                         size_t level, size_t elem_begin,
-                                         size_t elem_end, bool tensor_shared);
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle

paddle/framework/grad_op_builder.cc

@@ -19,45 +19,44 @@ permissions and limitations under the License. */
 namespace paddle {
 namespace framework {
 
-class OpRegistry;
-
-using VarIndexMap = std::unordered_map<std::string, int>;
+typedef std::vector<int> Ints;
 
 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;
+const Ints* AttrFormat(const AttributeMap& attrs, const std::string& key) {
+  return (attrs.count(key) > 0) ? &boost::get<Ints>(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;
+Ints* AttrFormat(AttributeMap& attrs, const std::string& key) {
+  return (attrs.count(key) > 0) ? &boost::get<Ints>(attrs.at(key)) : nullptr;
 }
 
-static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op,
-                       const OpArgType& src_type, const OpArgType& dst_type,
+static void TransOpArg(const OperatorBase* src_op,
+                       std::vector<std::string>& grad_inputs,
+                       std::vector<std::string>& grad_outputs,
+                       AttributeMap& grad_attrs,
+                       std::unordered_map<std::string, int>& grad_idxs,
+                       const std::string& src_type, const std::string& dst_type,
                        int& idx, bool is_grad) {
   const std::vector<std::string>& src_inout =
-      src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_;
-  const std::vector<int>* src_format = GetOpFormat(src_op, src_type);
+      (src_type == "input_format") ? src_op->inputs_ : src_op->outputs_;
+
+  const std::vector<int>* src_format = AttrFormat(src_op->Attrs(), src_type);
 
   std::vector<std::string>& dst_inout =
-      dst_type == OpArgType::IN ? dst_op->inputs_ : dst_op->outputs_;
-  std::vector<int>* dst_format = GetOpFormat(dst_op, dst_type);
+      (dst_type == "input_format") ? grad_inputs : grad_outputs;
+
+  std::vector<int>* dst_format = AttrFormat(grad_attrs, dst_type);
+
   const OpProto& proto = OpRegistry::protos().at(src_op->type_);
+
   const auto& src_arg_list =
-      src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
+      (src_type == "input_format") ? 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++;
+    grad_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);
@@ -76,26 +75,42 @@ static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op,
 }
 
 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});
+  const std::string& grad_op_type = OpRegistry::grad_ops().at(op->Type());
+
+  AttributeMap grad_attrs(op->Attrs());
+  grad_attrs.erase("input_format");
+  grad_attrs.erase("output_format");
+  if (op->Attrs().count("input_format") > 0) {
+    grad_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});
+  if (op->Attrs().count("input_format") > 0 ||
+      op->Attrs().count("output_format") > 0) {
+    grad_attrs["input_format"] = std::vector<int>({0});
   }
-  grad_op->in_out_idxs_.reset(new VarIndexMap());
+
+  std::vector<std::string> grad_inputs, grad_outputs;
+
+  using VarIndexMap = std::unordered_map<std::string, int>;
+  VarIndexMap* grad_idxs = 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
+  TransOpArg(op, grad_inputs, grad_outputs, grad_attrs, *grad_idxs,
+             "input_format", "input_format", in_idx, false);  // I
+  TransOpArg(op, grad_inputs, grad_outputs, grad_attrs, *grad_idxs,
+             "output_format", "input_format", in_idx, false);  // G
+  TransOpArg(op, grad_inputs, grad_outputs, grad_attrs, *grad_idxs,
+             "output_format", "input_format", in_idx, true);  // OG
+  TransOpArg(op, grad_inputs, grad_outputs, grad_attrs, *grad_idxs,
+             "input_format", "output_format", out_idx, true);  // IG
+
+  OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)();
+
+  grad_op->type_ = grad_op_type;
+  grad_op->inputs_ = grad_inputs;
+  grad_op->outputs_ = grad_outputs;
+  grad_op->attrs_ = grad_attrs;
+  grad_op->in_out_idxs_.reset(grad_idxs);
+
   return grad_op;
 }
 

paddle/framework/grad_op_builder_test.cc

@@ -10,6 +10,8 @@ namespace framework {
 
 class NOP : public OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(NOP, OperatorBase)
+
   void InferShape(const Scope &scope) const override {}
   void Run(const Scope &scope,
            const platform::DeviceContext &dev_ctx) const override {}

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,
+LODTensor::LOD LODTensor::LOD::SliceInLevel(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);
+  // 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

@@ -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_.
+   * @note: low performance in slice lod_.
    */
-  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.
-   */
-  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_registry.h

@@ -69,18 +69,18 @@ class OpProtoAndCheckerMaker {
 
   VariableBuilder AddInput(const std::string& name,
                            const std::string& comment) {
-    auto input = proto_->mutable_inputs()->Add();
-    *input->mutable_name() = name;
-    *input->mutable_comment() = comment;
+    VarProto* input = proto_->add_inputs();
+    input->set_name(name);
+    input->set_comment(comment);
     return VariableBuilder{input, [=] { this->SetHasMultipleInput(); },
                            nullptr};
   }
 
   VariableBuilder AddOutput(const std::string& name,
                             const std::string& comment) {
-    auto output = proto_->mutable_outputs()->Add();
-    *output->mutable_name() = name;
-    *output->mutable_comment() = comment;
+    VarProto* output = proto_->add_outputs();
+    output->set_name(name);
+    output->set_comment(comment);
     return VariableBuilder{output, [=] { this->SetHasMultipleOutput(); },
                            [=] { this->SetHasTemporaryOutput(); }};
   }
@@ -89,17 +89,15 @@ class OpProtoAndCheckerMaker {
   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;
+    AttrProto* 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) {
@@ -187,7 +185,7 @@ class OpRegistry {
     OpProto& op_proto = protos()[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",
@@ -307,22 +305,45 @@ class OpRegistry {
   }
 };
 
+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 +354,121 @@ 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)       \
+#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__() {        \
+      __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)                          \
+#define REGISTER_OP_KERNEL(op_type, DEVICE_TYPE, place_class, ...)        \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
-      __reg_gradient_op__##__op_type##__op_type##_grad, \
-      "NO_GRADIENT must be in 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, ...)             \
+#define NO_GRADIENT(op_type)                           \
   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__)
+      __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)                      \
+#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_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__()
+      __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_KERNEL(op_type, DEVICE_TYPE)                               \
+#define USE_OP_DEVICE_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__            \
+      "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)) =                                  \
-          __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__()
+          TouchOpKernelRegistrar_##op_type##_##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)
+// 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 +7,8 @@ namespace paddle {
 namespace framework {
 class CosineOp : public OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(CosineOp, OperatorBase)
+
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}
   void InferShape(const Scope& scope) const override {}
@@ -27,6 +29,8 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 
 class MyTestOp : public OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(MyTestOp, OperatorBase)
+
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}

paddle/framework/operator.h

@@ -63,6 +63,17 @@ class ExecutionContext;
  */
 class OperatorBase {
  public:
+  OperatorBase() {}  // TODO(yi): This constructor is to be removed.
+  OperatorBase(const std::string& type, const std::vector<std::string>& inputs,
+               const std::vector<std::string>& outputs,
+               const AttributeMap& attrs,
+               std::unordered_map<std::string, int>* in_out_idxs)
+      : type_(type),
+        inputs_(inputs),
+        outputs_(outputs),
+        attrs_(attrs),
+        in_out_idxs_(in_out_idxs) {}
+
   virtual ~OperatorBase() {}
 
   template <typename T>
@@ -95,16 +106,24 @@ 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;
+
   //! 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::string Type() const { return type_; }
+  const std::vector<std::string> Inputs() const { return inputs_; }
+  const std::vector<std::string> Outputs() const { return outputs_; }
+  const AttributeMap& Attrs() const { return attrs_; }
+  const std::unordered_map<std::string, int>* InOutIdx() const {
+    return in_out_idxs_.get();
+  }
+
  public:
   std::string type_;
   // NOTE: in case of OpGrad, inputs_ contains:
@@ -285,6 +304,14 @@ class OpKernel {
 
 class OperatorWithKernel : public OperatorBase {
  public:
+  OperatorWithKernel() {}  // TODO(yi): This constructor is to be removed.
+  OperatorWithKernel(const std::string& type,
+                     const std::vector<std::string>& inputs,
+                     const std::vector<std::string>& outputs,
+                     const AttributeMap& attrs,
+                     std::unordered_map<std::string, int>* in_out_idxs)
+      : OperatorBase(type, inputs, outputs, attrs, in_out_idxs) {}
+
   struct OpKernelKey {
     platform::Place place_;
 
@@ -334,5 +361,15 @@ class OperatorWithKernel : public OperatorBase {
   virtual void InferShape(const InferShapeContext& ctx) const = 0;
 };
 
+#define DEFINE_OPERATOR_CTOR(Class, ParentClass)                         \
+ public:                                                                 \
+  Class() { /* TODO(yi): This constructor is to be removed. */           \
+  }                                                                      \
+  Class(const std::string& type, const std::vector<std::string>& inputs, \
+        const std::vector<std::string>& outputs,                         \
+        const ::paddle::framework::AttributeMap& attrs,                  \
+        std::unordered_map<std::string, int>* in_out_idxs)               \
+      : ParentClass(type, inputs, outputs, attrs, in_out_idxs) {}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/operator_test.cc

@@ -23,6 +23,8 @@ static int op_run_num = 0;
 
 class OpWithoutKernelTest : public OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(OpWithoutKernelTest, OperatorBase)
+
   void Init() override { x = 1; }
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
@@ -97,6 +99,8 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 static int cpu_kernel_run_num = 0;
 
 class OpWithKernelTest : public OperatorWithKernel {
+ public:
+  DEFINE_OPERATOR_CTOR(OpWithKernelTest, OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {}
 };
@@ -116,6 +120,8 @@ class CPUKernelTest : public OpKernel {
 // multiple inputs test
 class OperatorMultiInputsTest : public OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(OperatorMultiInputsTest, OperatorBase)
+
   void Init() override { x = 1; }
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,

paddle/framework/pybind.cc

@@ -30,16 +30,15 @@ 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);
 

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/function/DepthwiseConvOpTest.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 "ConvOpTest.h"
+
+namespace paddle {
+
+#ifndef PADDLE_ONLY_CPU
+TEST(DepthwiseConv, Forward) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConv-CPU", "DepthwiseConv-GPU", forward);
+}
+
+TEST(DepthwiseConv, BackwardInput) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradInput-CPU", "DepthwiseConvGradInput-GPU", backward_input);
+}
+
+TEST(DepthwiseConv, BackwardFilter) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+      "GemmConvGradFilter-CPU", "DepthwiseConvGradFilter-GPU", backward_filter);
+}
+#endif
+
+}  // namespace paddle

paddle/function/GemmConvOpTest.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 "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

@@ -196,22 +196,23 @@ 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_,
+
+      // 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 / groups_,
-                                   outputChannels / groups_,
+                                                 inputChannels,
+                                                 outputChannels,
                                                  inputSize,
                                                  padding,
                                                  kernelSize,
-                                   inputData + inputOffset * g,
-                                   filterData + filterOffset * g,
+                                                 inputData,
+                                                 filterData,
                                                  nullptr, /* bias */
-                                   outputData + outputOffset * g,
+                                                 outputData,
                                                  bufferPtr,
                                                  sizePtr,
                                                  nnp_activation_identity,
@@ -221,7 +222,6 @@ public:
       CHECK_EQ(status, nnp_status_success);
     }
   }
-  }
 
   static void create_nnpack_threadpool() {
     if (FLAGS_nnpack_num_threads && threadpool_ == nullptr) {

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/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_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

@@ -62,9 +62,6 @@ 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)
 
-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)
 cc_test(recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op)

paddle/operators/add_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 class AddOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(AddOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_EQ(ctx.InputSize(), 2);
@@ -47,6 +48,7 @@ The equation is: Out = X + Y
 };
 
 class AddOpGrad : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(AddOpGrad, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {}
 };

paddle/operators/cross_entropy_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(OnehotCrossEntropyOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_EQ(ctx.InputSize(), 2,
@@ -38,6 +39,8 @@ class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
 };
 
 class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(OnehotCrossEntropyGradientOp,
+                       framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     auto X_grad = ctx.Output<Tensor>(framework::GradVarName("X"));

paddle/operators/fc_op.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-   Licensed under the Apache License, Version 2.0 (the "License");
-   you may not use this file except in compliance with the License.
-   You may obtain a copy of the License at
-
-   http://www.apache.org/licenses/LICENSE-2.0
-
-   Unless required by applicable law or agreed to in writing, software
-   distributed under the License is distributed on an "AS IS" BASIS,
-   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-   See the License for the specific language governing permissions and
-   limitations under the License. */
-
-#include "paddle/operators/net_op.h"
-
-#include "paddle/framework/eigen.h"
-#include "paddle/framework/op_registry.h"
-
-namespace paddle {
-namespace operators {
-
-using OpRegistry = framework::OpRegistry;
-
-class FullyConnectedOp : public NetOp {
- 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")}, {}));
-    }
-
-    auto activation = GetAttr<std::string>("activation");
-    AddOp(OpRegistry::CreateOp(activation, {Output("before_act")},
-                               {Output("Y")}, {}));
-    CompleteAddOp(false);
-  }
-};
-
-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);

paddle/operators/fill_zeros_like_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 class FillZerosLikeOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(FillZerosLikeOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL,

paddle/operators/gaussian_random_op.cc

@@ -43,6 +43,7 @@ class GaussianRandomKernel : public framework::OpKernel {
 };
 
 class GaussianRandomOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(GaussianRandomOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext& context) const override {
     auto* tensor = context.Output<framework::Tensor>(0);

paddle/operators/gaussian_random_op.cu

@@ -33,7 +33,8 @@ class GaussianRandomKernel : public framework::OpKernel {
 
     int seed = context.op_.GetAttr<int>("seed");
     if (seed == 0) {
-      seed = std::random_device()();
+      std::random_device rd;
+      seed = rd();
     }
     curandGenerator_t g;
     PADDLE_ENFORCE(platform::dynload::curandCreateGenerator(

paddle/operators/mean_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 class MeanOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(MeanOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_EQ(ctx.InputSize(), 1, "Input size of AddOp must be one");
@@ -39,6 +40,7 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
 };
 
 class MeanGradOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(MeanGradOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     ctx.Output<Tensor>(framework::GradVarName("X"))

paddle/operators/mul_op.cc

@@ -19,6 +19,7 @@ namespace paddle {
 namespace operators {
 
 class MulOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(MulOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE(ctx.InputSize() == 2, "The mul op must take two inputs");
@@ -54,6 +55,7 @@ The equation is: Out = X * Y
 };
 
 class MulOpGrad : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(MulOpGrad, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {}
   std::string DebugString() const override {

paddle/operators/net_op.h

@@ -35,6 +35,8 @@ namespace operators {
  */
 class NetOp : public framework::OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(NetOp, framework::OperatorBase)
+
   /**
    * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch

paddle/operators/net_op_test.cc

@@ -12,6 +12,8 @@ static int run_cnt = 0;
 
 class TestOp : public framework::OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(TestOp, framework::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 +23,8 @@ class TestOp : public framework::OperatorBase {
 
 class EmptyOp : public framework::OperatorBase {
  public:
+  DEFINE_OPERATOR_CTOR(EmptyOp, framework::OperatorBase)
+
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope, const DeviceContext& dev_ctx) const override {}
 };

paddle/operators/recurrent_op.h

@@ -100,6 +100,7 @@ class RecurrentGradientAlgorithm {
 };
 
 class RecurrentOp final : public framework::OperatorBase {
+  DEFINE_OPERATOR_CTOR(RecurrentOp, framework::OperatorBase)
  public:
   void Init() override;
 

paddle/operators/recurrent_op_test.cc

@@ -395,4 +395,4 @@ 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

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 class RowWiseAddOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(RowWiseAddOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE(ctx.InputSize() == 2UL,

paddle/operators/sgd_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 class SGDOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(SGDOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_EQ(ctx.InputSize(), 2, "Input size of SGDOp must be two");

paddle/operators/sigmoid_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 class SigmoidOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(SigmoidOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE(ctx.InputSize() == 1, "Sigmoid Op only have one input");
@@ -38,6 +39,7 @@ class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
 };
 
 class SigmoidOpGrad : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(SigmoidOpGrad, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());

paddle/operators/softmax_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 class SoftmaxOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(SoftmaxOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL,
@@ -42,6 +43,7 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
 };
 
 class SoftmaxOpGrad : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(SoftmaxOpGrad, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_EQ(ctx.InputSize(), 3UL,

paddle/operators/uniform_random_op.cc

@@ -46,6 +46,7 @@ class CPUUniformRandomKernel : public framework::OpKernel {
 };
 
 class UniformRandomOp : public framework::OperatorWithKernel {
+  DEFINE_OPERATOR_CTOR(UniformRandomOp, framework::OperatorWithKernel)
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
     PADDLE_ENFORCE(GetAttr<float>("min") < GetAttr<float>("max"),

paddle/operators/uniform_random_op.cu

@@ -51,7 +51,8 @@ class GPUUniformRandomKernel : public framework::OpKernel {
     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/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} \
@@ -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/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 (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;
         }
-      if (trainerConfig_.algorithm() == "sgd") {
-        optimizeConfigV2.set_optimizer(paddle::OptimizerConfig::SGD);
-        // FIXME: config all algorithms
-      } else {
-        optimizeConfigV2.set_optimizer(paddle::OptimizerConfig::SGD);
       }
-      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

@@ -16,7 +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} ${PADDLE_SOURCE_DIR}/python/paddle/v2/master/
     )
   add_dependencies(copy_paddle_master paddle_master)
 endif(WITH_GOLANG)
@@ -25,10 +25,10 @@ configure_file(${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 ${PADDLE_PYTHON_BUILD_DIR}/.timestamp

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/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)
@@ -25,3 +24,4 @@ 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

@@ -73,21 +73,35 @@ 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())
 
 

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_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(plain_net), inputs:(W, X, Y), outputs:(Out, fc.out, pre_activation).
     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:(W, X), outputs:(fc.out, pre_activation).
+        Op(plain_net), inputs:(W, X), outputs:(fc.out, pre_activation).
+            Op(mul), inputs:(X, W), outputs:(pre_activation).
+            Op(sigmoid), inputs:(pre_activation), outputs:(fc.out).
 '''
         self.assertEqual(expected, "\n" + str(net))
 

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
 
@@ -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'])]
 )