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

.gitignore

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

.travis.yml

@@ -37,8 +37,8 @@ before_install:
   - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
   # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
   # protobuf version.
-  - pip install numpy wheel 'protobuf==3.1' sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit requests==2.9.2 LinkChecker
-  - pip install rarfile
+  - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
+  - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
   - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
   - go get -u github.com/alecthomas/gometalinter

CMakeLists.txt

@@ -14,8 +14,8 @@
 
 cmake_minimum_required(VERSION 3.0)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
-set(PROJ_ROOT ${CMAKE_CURRENT_SOURCE_DIR})
-set(PROJ_BINARY_ROOT ${CMAKE_CURRENT_BINARY_DIR})
+set(PADDLE_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
+set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
 
 include(system)
 
@@ -36,8 +36,8 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
+option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    OFF)
+option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
@@ -121,8 +121,8 @@ include(version)            # set PADDLE_VERSION
 include(coveralls)          # set code coverage
 
 
-include_directories("${PROJ_ROOT}")
-include_directories("${PROJ_ROOT}/paddle/cuda/include")
+include_directories("${PADDLE_SOURCE_DIR}")
+include_directories("${PADDLE_SOURCE_DIR}/paddle/cuda/include")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/client/c")
 include_directories(${Boost_INCLUDE_DIRS})
@@ -144,7 +144,7 @@ if(WITH_GPU)
 endif(WITH_GPU)
 
 if(WITH_MKLDNN)
-    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIBRARY} ${MKLDNN_IOMP_LIB})
+    list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
 endif()
 
 if(USE_NNPACK)
@@ -164,10 +164,12 @@ if(WITH_GOLANG)
     add_subdirectory(go)
 endif(WITH_GOLANG)
 
+set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
 add_subdirectory(paddle)
 if(WITH_PYTHON)
   add_subdirectory(python)
 endif()
+
 if(WITH_DOC)
     add_subdirectory(doc)
 endif()

Dockerfile

@@ -28,7 +28,7 @@ RUN apt-get update && \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
     python-matplotlib gcc-4.8 g++-4.8 \
-    automake locales clang-format-3.8 swig doxygen cmake  \
+    automake locales clang-format swig doxygen cmake  \
     liblapack-dev liblapacke-dev libboost-dev \
     clang-3.8 llvm-3.8 libclang-3.8-dev \
     net-tools && \
@@ -64,13 +64,28 @@ RUN pip install --upgrade pip && \
     pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
     pip install pre-commit 'requests==2.9.2' 'ipython==5.3.0' && \
     pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
-    pip install rarfile
+    pip install opencv-python rarfile 'scipy>=0.19.0' 'nltk>=3.2.2'
 
 # To fix https://github.com/PaddlePaddle/Paddle/issues/1954, we use
 # the solution in https://urllib3.readthedocs.io/en/latest/user-guide.html#ssl-py2
 RUN apt-get install -y libssl-dev libffi-dev
 RUN pip install certifi urllib3[secure]
 
+# TODO(qijun) The template library Eigen doesn't work well with GCC 5 
+# coming with the default Docker image, so we switch to use GCC 4.8 
+# by default. And I will check Eigen library later.
+
+RUN ln -sf gcc-4.8 /usr/bin/gcc && \
+    ln -sf gcc-ar-4.8 /usr/bin/gcc-ar && \
+    ln -sf gcc-nm-4.8 /usr/bin/gcc-nm && \
+    ln -sf gcc-ranlib-4.8 /usr/bin/gcc-ranlib && \
+    ln -sf gcc-4.8 /usr/bin/x86_64-linux-gnu-gcc && \
+    ln -sf gcc-ar-4.8 /usr/bin/x86_64-linux-gnu-gcc-ar && \
+    ln -sf gcc-nm-4.8 /usr/bin/x86_64-linux-gnu-gcc-nm && \
+    ln -sf gcc-ranlib-4.8 /usr/bin/x86_64-linux-gnu-gcc-ranlib && \
+    ln -sf g++-4.8 /usr/bin/g++ && \
+    ln -sf g++-4.8 /usr/bin/x86_64-linux-gnu-g++ 
+
 # Install woboq_codebrowser to /woboq
 RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \
     (cd /woboq \

cmake/configure.cmake

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

cmake/cpplint.cmake

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

cmake/flags.cmake

@@ -9,10 +9,12 @@ function(CheckCompilerCXX11Flag)
         if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
             message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
         endif()
-        # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
-        # Use Debug mode instead for now.
-        if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9) 
-            set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
+        if(NOT ANDROID)
+            # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
+            # Use Debug mode instead for now.
+            if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9)
+                set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
+            endif()
         endif()
     elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
         # cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang"

cmake/generic.cmake

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

cmake/package.cmake

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

cmake/util.cmake

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

cmake/version.cmake

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

doc/api/v2/config/layer.rst

@@ -257,6 +257,11 @@ seq_concat
 ..  autoclass:: paddle.v2.layer.seq_concat
     :noindex:
 
+kmax_sequence_score
+-------------------
+..  autoclass:: paddle.v2.layer.kmax_sequence_score
+    :noindex:
+
 sub_nested_seq
 --------------
 ..  autoclass:: paddle.v2.layer.sub_nested_seq

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/templates/conf.py.cn.in

@@ -13,22 +13,18 @@
 # serve to show the default.
 import sys
 import os, subprocess
+sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python'))
 import shlex
 from recommonmark import parser, transform
-try:
-   import py_paddle
-   import paddle
-   import paddle.v2
-except ImportError:
-   print("Must install paddle python package before generating documentation")
-   sys.exit(1)
+import paddle
+import paddle.v2
 
 MarkdownParser = parser.CommonMarkParser
 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 ------------------------------------------------
 
@@ -124,7 +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

@@ -13,15 +13,11 @@
 # serve to show the default.
 import sys
 import os, subprocess
+sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python'))
 import shlex
 from recommonmark import parser, transform
-try:
-   import py_paddle
-   import paddle
-   import paddle.v2
-except ImportError:
-   print("Must install paddle python package before generating documentation")
-   sys.exit(1)
+import paddle
+import paddle.v2
 
 
 MarkdownParser = parser.CommonMarkParser
@@ -29,7 +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 ------------------------------------------------
 
@@ -124,7 +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/glide.lock

 hash: 1b9b07408ca7fac27a374dc2ccd2433e4bff090484008a037df967284949a582
-updated: 2017-08-03T21:46:51.744995189Z
+updated: 2017-08-07T23:37:48.867469328Z
 imports:
 - name: github.com/beorn7/perks
   version: 4c0e84591b9aa9e6dcfdf3e020114cd81f89d5f9
@@ -10,7 +10,7 @@ imports:
 - name: github.com/cockroachdb/cmux
   version: 112f0506e7743d64a6eb8fedbcff13d9979bbf92
 - name: github.com/coreos/etcd
-  version: c31bec0f29facff13f7c3e3d948e55dd6689ed42
+  version: d0d1a87aa96ae14914751d42264262cb69eda170
   subpackages:
   - alarm
   - auth
@@ -24,6 +24,7 @@ imports:
   - error
   - etcdserver
   - etcdserver/api
+  - etcdserver/api/etcdhttp
   - etcdserver/api/v2http
   - etcdserver/api/v2http/httptypes
   - etcdserver/api/v3client
@@ -210,11 +211,6 @@ testImports:
   version: 04cdfd42973bb9c8589fd6a731800cf222fde1a9
   subpackages:
   - spew
-- name: github.com/docker/docker
-  version: b6d164e6c46d8115b146e4c3ac93784e9ef8b49e
-  subpackages:
-  - pkg/ioutils
-  - pkg/longpath
 - name: github.com/pmezard/go-difflib
   version: d8ed2627bdf02c080bf22230dbb337003b7aba2d
   subpackages:

go/master/service_test.go

 package master_test
 
 import (
+	"io/ioutil"
+	"net/url"
 	"os"
+	"strings"
 	"testing"
 	"time"
 
 	"github.com/PaddlePaddle/Paddle/go/master"
 	"github.com/coreos/etcd/clientv3"
 	"github.com/coreos/etcd/embed"
-	"github.com/docker/docker/pkg/ioutils"
 	"github.com/stretchr/testify/assert"
 )
 
 func TestNewServiceWithEtcd(t *testing.T) {
 	// setup an embed etcd server
-	etcdDir, err := ioutils.TempDir("", "")
+	etcdDir, err := ioutil.TempDir("", "")
 	if err != nil {
 		t.Fatal(err)
 	}
 	cfg := embed.NewConfig()
+	lpurl, _ := url.Parse("http://localhost:0")
+	lcurl, _ := url.Parse("http://localhost:0")
+	cfg.LPUrls = []url.URL{*lpurl}
+	cfg.LCUrls = []url.URL{*lcurl}
 	cfg.Dir = etcdDir
 	e, err := embed.StartEtcd(cfg)
 	if err != nil {
@@ -30,15 +36,13 @@ func TestNewServiceWithEtcd(t *testing.T) {
 			t.Fatal(err)
 		}
 	}()
-	select {
-	case <-e.Server.ReadyNotify():
-		t.Log("Server is ready!")
-	case <-time.After(60 * time.Second):
-		e.Server.Stop() // trigger a shutdown
-		t.Fatal("Server took too long to start!")
-	}
 
-	ep := []string{"127.0.0.1:2379"}
+	<-e.Server.ReadyNotify()
+
+	port := strings.Split(e.Clients[0].Addr().String(), ":")[1]
+	endpoint := "127.0.0.1:" + port
+
+	ep := []string{endpoint}
 	masterAddr := "127.0.0.1:3306"
 	store, err := master.NewEtcdClient(ep, masterAddr, master.DefaultLockPath, master.DefaultAddrPath, master.DefaultStatePath, 30)
 	if err != nil {

go/pserver/client/c/cclient.go

@@ -90,8 +90,12 @@ func cArrayToSlice(p unsafe.Pointer, len int) []byte {
 
 type selector bool
 
-func (s selector) Select() bool {
-	return bool(s)
+func (s selector) Select() (bool, error) {
+	return bool(s), nil
+}
+
+func (s selector) Done() error {
+	return nil
 }
 
 type lister []client.Server
@@ -114,11 +118,10 @@ func paddle_new_pserver_client(addrs *C.char, selected int) C.paddle_pserver_cli
 }
 
 //export paddle_new_etcd_pserver_client
-func paddle_new_etcd_pserver_client(etcdEndpoints *C.char, selected int) C.paddle_pserver_client {
-	// TODO(Longfei: use etcd lock to decide which trainer to initialize the parameters)
+func paddle_new_etcd_pserver_client(etcdEndpoints *C.char) C.paddle_pserver_client {
 	addr := C.GoString(etcdEndpoints)
 	etcdClient := client.NewEtcd(addr)
-	c := client.NewClient(etcdClient, etcdClient.Desired(), selector(selected != 0))
+	c := client.NewClient(etcdClient, etcdClient.Desired(), etcdClient)
 	return add(c)
 }
 
@@ -136,7 +139,12 @@ func paddle_pserver_client_release(client C.paddle_pserver_client) {
 //export paddle_begin_init_params
 func paddle_begin_init_params(client C.paddle_pserver_client) C.int {
 	c := get(client)
-	if selected := c.BeginInitParams(); selected {
+	selected, err := c.BeginInitParams()
+	if err != nil {
+		panic(err)
+	}
+
+	if selected {
 		return 1
 	}
 	return 0

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)
 

go/pserver/client/client.go

@@ -27,9 +27,13 @@ import (
 
 // TODO(helin): add RPC call retry logic
 
-// Selector selects if the client should initialize parameter servers.
+// Selector selects if the client should initialize parameters and
+// reports the initialization process done.
 type Selector interface {
-	Select() bool
+	// Select selects if the client should initialize parameter servers.
+	Select() (bool, error)
+	// Done indicates the initialization process is done.
+	Done() error
 }
 
 // Server is the identification of a parameter Server.
@@ -115,7 +119,7 @@ func (c *Client) monitorPservers(l Lister, pserverNum int) {
 // servers. Other trainers will be blocked until the initialization is
 // done, and they need to get the initialized parameters from
 // parameter servers using GetParams.
-func (c *Client) BeginInitParams() bool {
+func (c *Client) BeginInitParams() (bool, error) {
 	return c.sel.Select()
 }
 

go/pserver/client/client_test.go

@@ -124,8 +124,12 @@ func initEtcdClient() {
 
 type selector bool
 
-func (s selector) Select() bool {
-	return bool(s)
+func (s selector) Select() (bool, error) {
+	return bool(s), nil
+}
+
+func (s selector) Done() error {
+	return nil
 }
 
 type lister []client.Server
@@ -135,7 +139,11 @@ func (l lister) List() []client.Server {
 }
 
 func testClient(t *testing.T, c *client.Client) {
-	selected := c.BeginInitParams()
+	selected, err := c.BeginInitParams()
+	if err != nil {
+		t.Fatal(err)
+	}
+
 	if !selected {
 		t.Fatal("should be selected.")
 	}

go/pserver/client/etcd_client.go

@@ -16,53 +16,60 @@ package client
 
 import (
 	"context"
+	"errors"
+	"fmt"
 	"strconv"
 	"strings"
 	"time"
 
 	"github.com/PaddlePaddle/Paddle/go/pserver"
 	"github.com/coreos/etcd/clientv3"
+	"github.com/coreos/etcd/clientv3/concurrency"
 	log "github.com/sirupsen/logrus"
 )
 
 const (
 	defaultEtcdTimeout time.Duration = 5 * time.Second
+
+	initLockPath = "/init_ps/lock"
+	initDonePath = "/init_ps/done"
+	initDoneVal  = "1"
 )
 
-// EtcdClient is used by pserver client that is a part of trainer process.
+// Etcd is used by pserver client that is a part of trainer process.
 // TODO:
-// 1. add watcher to watch the change state of pservers)
-// 1. add etcd lock)
-type EtcdClient struct {
+// 1. add watcher to watch the change state of pservers.
+type Etcd struct {
 	client    *clientv3.Client
 	timeout   time.Duration
 	endpoints []string
+	lock      *concurrency.Mutex
 }
 
 // Desired read ps desired number from etcd.
-func (p *EtcdClient) Desired() int {
+func (e *Etcd) Desired() int {
 	var psDesired int
 	for {
-		ctx, cancel := context.WithTimeout(context.Background(), p.timeout)
-		resp, err := p.client.Get(ctx, pserver.PsDesired)
+		ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
+		resp, err := e.client.Get(ctx, pserver.PsDesired)
 		cancel()
 		if err != nil {
 			log.Errorf("Get ps dresire number failed! recnnectiong..., %v", err)
-			time.Sleep(p.timeout)
+			time.Sleep(e.timeout)
 			continue
 		}
 
 		kvs := resp.Kvs
 		if len(kvs) == 0 {
 			log.Infoln("Waiting for ps desired registered ...")
-			time.Sleep(p.timeout)
+			time.Sleep(e.timeout)
 			continue
 		}
 
 		psDesired, err = strconv.Atoi(string(resp.Kvs[0].Value))
 		if err != nil {
 			log.Errorf("psDesired %d invalid %v", psDesired, err)
-			time.Sleep(p.timeout)
+			time.Sleep(e.timeout)
 			continue
 		}
 
@@ -73,26 +80,26 @@ func (p *EtcdClient) Desired() int {
 }
 
 // List return the pserver list read from etcd.
-func (p *EtcdClient) List() []Server {
-	psDesired := p.Desired()
+func (e *Etcd) List() []Server {
+	psDesired := e.Desired()
 
 	servers := make([]Server, psDesired)
 	for {
 		for i := 0; i < psDesired; i++ {
-			ctx, cancel := context.WithTimeout(context.Background(), p.timeout)
+			ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
 			psKey := pserver.PsPath + strconv.Itoa(i)
 			log.Debugf("checking %s", psKey)
-			resp, err := p.client.Get(ctx, psKey)
+			resp, err := e.client.Get(ctx, psKey)
 			cancel()
 			if err != nil {
 				log.Infof("Get psKey= %s error, %v", psKey, err)
-				time.Sleep(p.timeout)
+				time.Sleep(e.timeout)
 				continue
 			}
 			kvs := resp.Kvs
 			if len(kvs) == 0 {
 				log.Infof("Waiting for ps addr registered ...")
-				time.Sleep(p.timeout)
+				time.Sleep(e.timeout)
 				continue
 			}
 
@@ -100,7 +107,7 @@ func (p *EtcdClient) List() []Server {
 			// TODO(Longfei) check the ps address
 			if psAddr == "" {
 				log.Infof("Get psKey = %s, psAddr is empty", psKey)
-				time.Sleep(p.timeout)
+				time.Sleep(e.timeout)
 				continue
 			}
 			log.Debugf("got value (%s) for key: %s", psAddr, psKey)
@@ -113,7 +120,7 @@ func (p *EtcdClient) List() []Server {
 }
 
 // NewEtcd create a etcd client to return the state of pserver on etcd.
-func NewEtcd(endpoints string) *EtcdClient {
+func NewEtcd(endpoints string) *Etcd {
 	ep := strings.Split(endpoints, ",")
 	var cli *clientv3.Client
 	var err error
@@ -130,10 +137,118 @@ func NewEtcd(endpoints string) *EtcdClient {
 		break
 	}
 	log.Infof("Connected to etcd: %s\n", endpoints)
-	client := &EtcdClient{
+	client := &Etcd{
 		client:    cli,
 		timeout:   defaultEtcdTimeout,
 		endpoints: ep,
 	}
 	return client
 }
+
+// Select indicates if the current trainer is selected to initialize
+// the pserver parameters.
+func (e *Etcd) Select() (bool, error) {
+	sess, err := concurrency.NewSession(e.client, concurrency.WithTTL(5))
+	if err != nil {
+		return false, err
+	}
+
+	lock := concurrency.NewMutex(sess, initLockPath)
+	log.Infof("Trying to acquire lock at %s.", initLockPath)
+	// Do not use timeout context here, since we don't know how
+	// long does it take for other trainers to initialize the
+	// parameters.
+	err = lock.Lock(context.Background())
+	if err != nil {
+		return false, err
+	}
+	log.Infof("Successfully acquired lock at %s.", initLockPath)
+
+	get := clientv3.OpGet(initDonePath)
+	ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
+	tresp, err := e.client.Txn(ctx).If(lock.IsOwner()).Then(get).Commit()
+	cancel()
+	if err != nil {
+		return false, err
+	}
+
+	if !tresp.Succeeded {
+		return false, errors.New("no longer the owner of the lock")
+	}
+
+	resp := tresp.Responses[0].GetResponseRange()
+
+	if len(resp.Kvs) == 0 {
+		// Key value not set, select current trainer.
+		e.lock = lock
+		log.Infoln("Trainer selected.")
+		return true, nil
+	}
+
+	if string(resp.Kvs[0].Value) == initDoneVal {
+		log.Infoln("Initialization is already done.")
+		ctx, cancel = context.WithTimeout(context.Background(), e.timeout)
+		err = lock.Unlock(ctx)
+		cancel()
+		if err != nil {
+			log.Errorln(err)
+		}
+		return false, nil
+	}
+
+	return false, fmt.Errorf("key %s have unexpected value: %v", initDonePath, resp.Kvs[0].Value)
+}
+
+// Done indicates the parameter initialization process is done.
+func (e *Etcd) Done() error {
+	if e.lock == nil {
+		return errors.New("lock is nil, Done called unexpectedly")
+	}
+
+	put := clientv3.OpPut(initDonePath, initDoneVal)
+	ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
+	tresp, err := e.client.Txn(ctx).If(e.lock.IsOwner()).Then(put).Commit()
+	cancel()
+	if err != nil {
+		return err
+	}
+
+	if !tresp.Succeeded {
+		return errors.New("no longer the owner of the lock")
+	}
+
+	ctx, cancel = context.WithTimeout(context.Background(), e.timeout)
+	err = e.lock.Unlock(ctx)
+	cancel()
+	if err != nil {
+		log.Errorln(err)
+	} else {
+		e.lock = nil
+	}
+
+	return nil
+}
+
+// Close closes the etcd client.
+func (e *Etcd) Close() error {
+	var err error
+	if e.lock != nil {
+		ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
+		err = e.lock.Unlock(ctx)
+		cancel()
+		if err == nil {
+			e.lock = nil
+		}
+	}
+
+	cErr := e.client.Close()
+	if cErr != nil {
+		if err != nil {
+			log.Errorln(cErr)
+			return err
+		}
+		return cErr
+	}
+
+	return err
+}

go/pserver/client/etcd_client_test.go

+package client_test
+
+import (
+	"io/ioutil"
+	"net/url"
+	"os"
+	"strings"
+	"sync"
+	"testing"
+
+	"github.com/PaddlePaddle/Paddle/go/pserver/client"
+	"github.com/coreos/etcd/embed"
+)
+
+func TestSelector(t *testing.T) {
+	etcdDir, err := ioutil.TempDir("", "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	cfg := embed.NewConfig()
+	lpurl, _ := url.Parse("http://localhost:0")
+	lcurl, _ := url.Parse("http://localhost:0")
+	cfg.LPUrls = []url.URL{*lpurl}
+	cfg.LCUrls = []url.URL{*lcurl}
+	cfg.Dir = etcdDir
+	e, err := embed.StartEtcd(cfg)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	defer func() {
+		e.Close()
+		if err := os.RemoveAll(etcdDir); err != nil {
+			t.Fatal(err)
+		}
+	}()
+
+	<-e.Server.ReadyNotify()
+
+	port := strings.Split(e.Clients[0].Addr().String(), ":")[1]
+	endpoint := "127.0.0.1:" + port
+
+	var mu sync.Mutex
+	selectedCount := 0
+	var wg sync.WaitGroup
+	selectAndDone := func(c *client.Etcd) {
+		defer wg.Done()
+
+		selected, err := c.Select()
+		if err != nil {
+			panic(err)
+		}
+
+		if selected {
+			mu.Lock()
+			selectedCount++
+			mu.Unlock()
+			err = c.Done()
+			if err != nil {
+				t.Fatal(err)
+			}
+		}
+	}
+
+	c0 := client.NewEtcd(endpoint)
+	c1 := client.NewEtcd(endpoint)
+	c2 := client.NewEtcd(endpoint)
+	c3 := client.NewEtcd(endpoint)
+	wg.Add(3)
+	go selectAndDone(c0)
+	go selectAndDone(c1)
+	go selectAndDone(c2)
+	wg.Wait()
+
+	// simulate trainer crashed and restarted after the
+	// initialization process.
+	wg.Add(1)
+	go selectAndDone(c3)
+	wg.Wait()
+
+	mu.Lock()
+	if selectedCount != 1 {
+		t.Fatal("selected count wrong:", selectedCount)
+	}
+	mu.Unlock()
+
+	err = c0.Close()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	err = c1.Close()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	err = c2.Close()
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	err = c3.Close()
+	if err != nil {
+		t.Fatal(err)
+	}
+}

paddle/api/CMakeLists.txt

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

paddle/api/ParameterUpdater.cpp

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

paddle/capi/Arguments.cpp

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

paddle/capi/arguments.h

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

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

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

paddle/capi/gradient_machine.cpp

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

paddle/capi/gradient_machine.h

@@ -36,6 +36,18 @@ typedef void* paddle_gradient_machine;
 PD_API paddle_error paddle_gradient_machine_create_for_inference(
     paddle_gradient_machine* machine, void* modelConfigProtobuf, int size);
 
+/**
+ * @brief Create a gradient machine used for model inference, using config with
+ *        parameters which is generated by `paddle merge_model`.
+ * @param [out] machine that used for model inference.
+ * @param [in] mergedModel
+ * @param [in] size
+ * @return paddle_error
+ */
+PD_API paddle_error
+paddle_gradient_machine_create_for_inference_with_parameters(
+    paddle_gradient_machine* machine, void* mergedModel, uint64_t size);
+
 /**
  * @brief Load parameter from disk.
  * @param machine Gradient Machine.

paddle/capi/tests/CMakeLists.txt

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

paddle/framework/CMakeLists.txt

@@ -7,6 +7,9 @@ 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_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor)
+
 cc_test(variable_test SRCS variable_test.cc)
 
 cc_library(scope SRCS scope.cc)
@@ -32,6 +35,11 @@ py_proto_compile(framework_py_proto SRCS attribute.proto op_proto.proto op_desc.
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
+add_custom_command(TARGET framework_py_proto POST_BUILD
+    COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto
+    COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto/
+    COMMENT "Copy generated python proto into directory paddle/v2/framework/proto."
+    WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 
 cc_library(backward SRCS backward.cc DEPS net_op)
 cc_test(backward_test SRCS backward_test.cc DEPS backward)
@@ -40,11 +48,16 @@ if(WITH_PYTHON)
 cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
-	fc_op
-	sgd_op
-	add_op
-	mean_op
-	cross_entropy_op
-	fill_zeros_like_op
-	recurrent_op)
+    sgd_op
+    add_op
+    mul_op
+    rowwise_add_op
+    sigmoid_op
+    softmax_op
+    mean_op
+    cross_entropy_op
+    recurrent_op
+    uniform_random_op
+    gaussian_random_op
+    fill_zeros_like_op)
 endif(WITH_PYTHON)

paddle/framework/attribute.h

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #pragma once
 
-#include <boost/variant.hpp>
 #include <functional>
 #include <string>
 #include <unordered_map>
@@ -24,6 +23,7 @@ limitations under the License. */
 #include "paddle/framework/attribute.pb.h"
 #include "paddle/framework/op_desc.pb.h"
 #include "paddle/platform/enforce.h"
+#include "paddle/platform/variant.h"
 
 namespace paddle {
 namespace framework {

paddle/framework/backward.cc

@@ -13,6 +13,7 @@
    limitations under the License. */
 
 #include "paddle/framework/backward.h"
+
 #include <list>
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
@@ -132,8 +133,9 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
     for (std::string& grad_input : grad_op->inputs_) {
       if (no_grad_names.count(grad_input)) {
-        std::string prefix =
-            grad_input.substr(0, grad_input.size() - kGradVarSuffix.size());
+        // +1 for \0
+        std::string prefix = grad_input.substr(
+            0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
         grad_input = prefix + kZeroVarSuffix;
 
         // If part of input gradient of that operator is not calculated, fill
@@ -166,7 +168,7 @@ std::shared_ptr<OperatorBase> Backward(
   std::unordered_set<std::string> no_grad_names;
   no_grad_names.reserve(no_grad_vars.size());
 
-  no_grad_names.insert(kEmptyVarName + kGradVarSuffix);
+  no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
 
   for (auto& name : no_grad_vars) {
     no_grad_names.insert(name + kGradVarSuffix);

paddle/framework/backward_test.cc

@@ -17,16 +17,23 @@
 #include <gtest/gtest.h>
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/net_op.h"
-#include "paddle/operators/type_alias.h"
 
 namespace paddle {
 namespace framework {
 
+using OperatorBase = framework::OperatorBase;
+using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
+using OpProto = framework::OpProto;
+using OpAttrChecker = framework::OpAttrChecker;
+using Scope = framework::Scope;
+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 platform::DeviceContext &dev_ctx) const override {}
+  void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
 };
 
 class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
@@ -71,7 +78,7 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
   }
 };
 
-class FcOp : public ops::NetOp {
+class FcOp : public operators::NetOp {
  public:
   void Init() override {
     AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")},
@@ -143,6 +150,7 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
 }  // namespace paddle
 
 namespace f = paddle::framework;
+namespace ops = paddle::operators;
 using EnforceNotMet = paddle::platform::EnforceNotMet;
 REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker);
 REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp);
@@ -165,10 +173,10 @@ TEST(Backward, simple_op_grad) {
   ASSERT_EQ(4UL, gop->inputs_.size());
   ASSERT_EQ(f::kEmptyVarName, gop->inputs_[0]);
   ASSERT_EQ("rowwise_add_grad", gop->type_);
-  ASSERT_EQ("X" + f::kGradVarSuffix, gop->outputs_[0]);
-  ASSERT_EQ("b" + f::kGradVarSuffix, gop->outputs_[1]);
+  ASSERT_EQ(f::GradVarName("X"), gop->outputs_[0]);
+  ASSERT_EQ(f::GradVarName("b"), gop->outputs_[1]);
 
-  ASSERT_EQ("X" + f::kGradVarSuffix, gop->Output("X" + f::kGradVarSuffix));
+  ASSERT_EQ(f::GradVarName("X"), gop->Output(f::GradVarName("X")));
 }
 
 TEST(Backward, simple_op_not_need_grad) {
@@ -176,7 +184,7 @@ TEST(Backward, simple_op_not_need_grad) {
   ASSERT_NE(fwd, nullptr);
   auto gop = f::Backward(*fwd, {"X"});
   ASSERT_EQ(std::find(gop->outputs_.begin(), gop->outputs_.end(),
-                      "X" + f::kGradVarSuffix),
+                      f::GradVarName("X")),
             gop->outputs_.end());
 
   auto no_input_gop = f::Backward(*fwd, {"X", "b"});
@@ -244,18 +252,18 @@ TEST(Backward, net_input_of_network_not_need_grad) {
   all_output.erase(f::kEmptyVarName);
 
   for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
-    ASSERT_NE(all_output.find(out + f::kGradVarSuffix), all_output.end());
+    ASSERT_NE(all_output.find(f::GradVarName(out)), all_output.end());
   }
 
   // Not Generated X
-  ASSERT_EQ(all_output.find("X" + f::kGradVarSuffix), all_output.end());
+  ASSERT_EQ(all_output.find(f::GradVarName("X")), all_output.end());
 
   ASSERT_EQ(2UL, bwd_net->ops_.size());
   ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
   auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
   ASSERT_EQ(3UL, first_fc_grad->ops_.size());
   ASSERT_EQ(f::kEmptyVarName,
-            first_fc_grad->ops_[2]->Output("A" + f::kGradVarSuffix));
+            first_fc_grad->ops_[2]->Output(f::GradVarName("A")));
 }
 
 TEST(Backward, net_shared_weight) {
@@ -307,15 +315,15 @@ TEST(Backward, op_part_of_output_are_not_need) {
   ASSERT_EQ(1UL, fill_zero.inputs_.size());
   ASSERT_EQ("Z", fill_zero.inputs_[0]);
   ASSERT_EQ(1UL, fill_zero.outputs_.size());
-  ASSERT_EQ("Z" + f::kZeroVarSuffix, fill_zero.outputs_[0]);
+  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.outputs_[0]);
 
   auto &d_many_out = *net->ops_[1];
   ASSERT_EQ("many_output_op_grad", d_many_out.type_);
   ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size());  // I/O/OG
-  ASSERT_EQ("Z" + f::kZeroVarSuffix, d_many_out.Input("z" + f::kGradVarSuffix));
-  ASSERT_EQ("Y" + f::kGradVarSuffix, d_many_out.Input("y" + f::kGradVarSuffix));
-  ASSERT_EQ("X" + f::kGradVarSuffix,
-            d_many_out.Output("x" + f::kGradVarSuffix));
+  ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix,
+            d_many_out.Input(f::GradVarName("z")));
+  ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y")));
+  ASSERT_EQ(f::GradVarName("X"), d_many_out.Output(f::GradVarName("x")));
 }
 
 TEST(Backward, op_part_of_input_are_not_need) {
@@ -325,10 +333,9 @@ TEST(Backward, op_part_of_input_are_not_need) {
   ASSERT_EQ(grad_mul.type_, "mul_grad");
   ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
   ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
-  ASSERT_EQ(grad_mul.Output("A" + f::kGradVarSuffix), f::kEmptyVarName);
-  ASSERT_EQ(grad_mul.Output("B" + f::kGradVarSuffix), "b" + f::kGradVarSuffix);
-  ASSERT_EQ(grad_mul.Input("Out" + f::kGradVarSuffix),
-            "out" + f::kGradVarSuffix);
+  ASSERT_EQ(grad_mul.Output(f::GradVarName("A")), f::kEmptyVarName);
+  ASSERT_EQ(grad_mul.Output(f::GradVarName("B")), f::GradVarName("b"));
+  ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out"));
   ASSERT_EQ(grad_mul.Input("A"), "a");
   ASSERT_EQ(grad_mul.Input("B"), "b");
   ASSERT_EQ(grad_mul.Input("Out"), "out");

paddle/framework/ddim.h

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

paddle/framework/details/lod_tensor.cc

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

paddle/framework/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 {}
@@ -83,21 +85,19 @@ TEST(GradOpBuilder, MutiInOut) {
   EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
   EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
             std::vector<std::string>({"out2_1", "out2_2"}));
-  EXPECT_EQ(grad_test_op->Input("Out1" + f::kGradVarSuffix),
-            "out1" + f::kGradVarSuffix);
-  EXPECT_EQ(grad_test_op->Inputs("Out2_mult" + f::kGradVarSuffix),
+  EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")),
+            f::GradVarName("out1"));
+  EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")),
             std::vector<std::string>(
-                {"out2_1" + f::kGradVarSuffix, "out2_2" + f::kGradVarSuffix}));
+                {f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
 
   ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
-  EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
-            "in1" + f::kGradVarSuffix);
-  EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
-            std::vector<std::string>({"in2_1" + f::kGradVarSuffix,
-                                      "in2_2" + f::kGradVarSuffix,
-                                      "in2_3" + f::kGradVarSuffix}));
-  EXPECT_EQ(grad_test_op->Output("In3" + f::kGradVarSuffix),
-            "in3" + f::kGradVarSuffix);
+  EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
+  EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
+            std::vector<std::string>({f::GradVarName("in2_1"),
+                                      f::GradVarName("in2_2"),
+                                      f::GradVarName("in2_3")}));
+  EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3"));
 }
 
 TEST(GradOpBuilder, IOIgnoredInGradient) {
@@ -119,19 +119,18 @@ TEST(GradOpBuilder, IOIgnoredInGradient) {
   EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
             std::vector<std::string>({"out1_1", "out1_2"}));
   EXPECT_EQ(grad_test_op->Input("Out2"), f::kEmptyVarName);
-  EXPECT_EQ(grad_test_op->Inputs("Out1_mult" + f::kGradVarSuffix),
+  EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")),
             std::vector<std::string>(
-                {"out1_1" + f::kGradVarSuffix, "out1_2" + f::kGradVarSuffix}));
-  EXPECT_EQ(grad_test_op->Input("Out2" + f::kGradVarSuffix),
-            "out2" + f::kGradVarSuffix);
+                {f::GradVarName("out1_1"), f::GradVarName("out1_2")}));
+  EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")),
+            f::GradVarName("out2"));
 
   ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
-  EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
-            "in1" + f::kGradVarSuffix);
-  EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
+  EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
+  EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
             std::vector<std::string>(
-                {"in2_1" + f::kGradVarSuffix, "in2_2" + f::kGradVarSuffix}));
-  EXPECT_EQ(grad_test_op->Outputs("In3_mult" + f::kGradVarSuffix),
+                {f::GradVarName("in2_1"), f::GradVarName("in2_2")}));
+  EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")),
             std::vector<std::string>(
-                {"in3_1" + f::kGradVarSuffix, "in3_2" + f::kGradVarSuffix}));
+                {f::GradVarName("in3_1"), f::GradVarName("in3_2")}));
 }

paddle/framework/lod_tensor.cc

+/* 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 <glog/logging.h>
+
+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 LODTensor::SliceShared(size_t level, size_t elem_begin,
+                                 size_t elem_end) const {
+  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
+  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
+                 NumLevels());
+  PADDLE_ENFORCE(elem_begin < NumElements(level),
+                 "element begin [%d] out of range [%d]", elem_begin,
+                 NumElements(level));
+  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
+                 "element end [%d] out of range [%d]", elem_end,
+                 NumElements(level));
+
+  auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
+                                   true /*tensor_shared*/);
+
+  // slice elements just need to update LOD info, because offsets are not
+  // changed, so the original tensor_ can be reused.
+  return LODTensor(tensor_, new_lod);
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/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>
+#if (!PADDLE_ONLY_CPU)
+#include <thrust/device_vector.h>
+#include <thrust/host_vector.h>
+#endif
+
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/enforce.h"
+
+namespace paddle {
+namespace framework {
+
+/*
+ * LODTensor (Level of details Tensor)
+ * see https://en.wikipedia.org/wiki/Level_of_details for reference.
+ */
+class LODTensor {
+ public:
+// Level save offsets of each unit.
+#ifdef PADDLE_ONLY_CPU
+  using Level = std::vector<size_t>;
+#else
+  using Level = thrust::device_vector<size_t>;
+#endif
+  // 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;
+
+  LODTensor() {}
+  LODTensor(const std::shared_ptr<Tensor> &tensor,
+            const std::shared_ptr<LOD> &lod) {
+    Reset(tensor, lod);
+  }
+
+  void Reset(const std::shared_ptr<Tensor> &tensor,
+             const std::shared_ptr<LOD> &lod) {
+    tensor_ = tensor;
+    lod_start_pos_ = lod;
+  }
+
+  /*
+   * Get a element from LOD.
+   */
+  size_t lod_element(size_t level, size_t elem) const {
+    PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
+                   NumLevels());
+    PADDLE_ENFORCE(elem < NumElements(level),
+                   "element begin [%d] out of range [%d]", elem,
+                   NumElements(level));
+    return (*lod_start_pos_)[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;
+  }
+  /*
+   * Number of elements in a level.
+   */
+  size_t NumElements(size_t level = 0) const {
+    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;
+  }
+
+  /*
+   * 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;
+
+  /*
+   * Slice of elements of a level, [elem_begin: elem_end], with tensor shared.
+   * @note: low performance in slice lod_start_pos_.
+   */
+  LODTensor SliceShared(size_t level, size_t elem_begin, size_t elem_end) const;
+
+  /*
+   * Copy other's lod_start_pos_, to share LOD info.
+   * @note: the LOD info should not be changed.
+   */
+  void ShareLOD(const LODTensor &other) {
+    lod_start_pos_ = other.lod_start_pos_;
+  }
+
+  /*
+   * Copy other's lod_start_pos_'s content, free to mutate.
+   */
+  void CopyLOD(const LODTensor &other) {
+    lod_start_pos_ = std::make_shared<LOD>(*other.lod_start_pos_);
+  }
+  /*
+   * Determine whether LODTensor has a valid LOD info.
+   */
+  bool HasLOD() const { return bool(lod_start_pos_); }
+  LOD *lod() const { return lod_start_pos_.get(); }
+
+  std::shared_ptr<Tensor> &tensor() { return tensor_; }
+  Tensor *raw_tensor() { return tensor_.get(); }
+
+ private:
+  std::shared_ptr<LOD> lod_start_pos_;
+  std::shared_ptr<Tensor> 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

+/*
+  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 <glog/logging.h>
+#include <gtest/gtest.h>
+#include <memory>
+
+namespace paddle {
+namespace framework {
+
+class LODTensorTester : public ::testing::Test {
+ public:
+  virtual void SetUp() override {
+    lod_tensor.reset(new LODTensor);
+    // tensor's batch_size: 30
+    // 3 levels
+    // 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});
+
+    auto tensor = std::make_shared<Tensor>();
+    tensor->Resize({20 /*batch size*/, 128 /*dim*/});
+    // malloc memory
+    tensor->mutable_data<float>(place);
+
+    lod_tensor->Reset(tensor, lod);
+  }
+
+ protected:
+  std::unique_ptr<LODTensor> lod_tensor;
+  platform::CPUPlace place;
+};
+
+TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor->NumLevels(), 3UL); }
+
+TEST_F(LODTensorTester, NumElements) {
+  ASSERT_EQ(lod_tensor->NumElements(0), 2UL);
+  ASSERT_EQ(lod_tensor->NumElements(1), 4UL);
+  ASSERT_EQ(lod_tensor->NumElements(2), 8UL);
+}
+
+TEST_F(LODTensorTester, 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) {
+  // slice 1 level
+  for (size_t level = 0; level < 3UL; ++level) {
+    auto new_lod_tensor =
+        lod_tensor->SliceCopied<float>(level, level + 1, place);
+    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.
+  }
+  // slice 2 level
+  for (size_t level = 0; level < 2UL; ++level) {
+    auto new_lod_tensor =
+        lod_tensor->SliceCopied<float>(level, level + 2, place);
+    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.
+  }
+}
+
+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) {
+  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());
+
+  level = 1;
+  new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, 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_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
+}
+
+TEST_F(LODTensorTester, ShareLOD) {
+  LODTensor new_lod_tensor;
+  new_lod_tensor.ShareLOD(*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());
+}
+
+}  // namespace framework
+}  // namespace paddle

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",
@@ -260,12 +258,6 @@ class OpRegistry {
     return CreateOp(op_desc.type(), inputs, outputs, attrs);
   }
 
-  static bool SupportGPU(const std::string& op_type) {
-    OperatorWithKernel::OpKernelKey key;
-    key.place_ = platform::GPUPlace();
-    return OperatorWithKernel::AllOpKernels().at(op_type).count(key) != 0;
-  }
-
   static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
     PADDLE_ENFORCE(!op.IsNetOp(),
                    "Use framework::Backward to get backward ops");
@@ -313,22 +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.
  */
@@ -339,97 +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)       \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                              \
-      __reg_gradient_op__##__op_type##__grad_op_type,                          \
-      "REGISTER_GRADIENT_OP must be in global namespace");                     \
-  static ::paddle::framework::GradOpRegisterHelper<__grad_op_class>            \
-      __op_gradient_register_##__op_type##__grad_op_type##__(#__op_type,       \
-                                                             #__grad_op_type); \
-  int __op_gradient_register_##__op_type##__grad_op_type##_handle__() {        \
-    return 0;                                                                  \
+#define REGISTER_GRADIENT_OP(op_type, grad_op_type, grad_op_class)           \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                            \
+      __reg_gradient_op__##op_type##_##grad_op_type,                         \
+      "REGISTER_GRADIENT_OP must be called in global namespace");            \
+  static ::paddle::framework::GradOpRegistrar<grad_op_class>                 \
+      __op_gradient_registrar_##op_type##_##grad_op_type##__(#op_type,       \
+                                                             #grad_op_type); \
+  int TouchOpGradientRegistrar_##op_type() {                                 \
+    __op_gradient_registrar_##op_type##_##grad_op_type##__.Touch();          \
+    return 0;                                                                \
   }
 
 /**
- * Macro to Forbid user register Gradient Operator.
+ * Macro to register OperatorKernel.
  */
-#define NO_GRADIENT(__op_type)                          \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                       \
-      __reg_gradient_op__##__op_type##__op_type##_grad, \
-      "NO_GRADIENT must be in global namespace")
+#define REGISTER_OP_KERNEL(op_type, DEVICE_TYPE, place_class, ...)        \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
+      __reg_op_kernel_##op_type##_##DEVICE_TYPE##__,                      \
+      "REGISTER_OP_KERNEL must be called in global namespace");           \
+  static ::paddle::framework::OpKernelRegistrar<place_class, __VA_ARGS__> \
+      __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__(#op_type);      \
+  int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() {                \
+    __op_kernel_registrar_##op_type##_##DEVICE_TYPE##__.Touch();          \
+    return 0;                                                             \
+  }
 
 /**
- * Macro to Register OperatorKernel.
+ * Macro to Forbid user register Gradient Operator.
  */
-#define REGISTER_OP_KERNEL(type, DEVICE_TYPE, PlaceType, ...)             \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
-      __reg_op_kernel_##type##_##DEVICE_TYPE##__,                         \
-      "REGISTER_OP_KERNEL must be in global namespace");                  \
-  struct __op_kernel_register__##type##__##DEVICE_TYPE##__ {              \
-    __op_kernel_register__##type##__##DEVICE_TYPE##__() {                 \
-      ::paddle::framework::OperatorWithKernel::OpKernelKey key;           \
-      key.place_ = PlaceType();                                           \
-      ::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
-          .reset(new __VA_ARGS__());                                      \
-    }                                                                     \
-  };                                                                      \
-  static __op_kernel_register__##type##__##DEVICE_TYPE##__                \
-      __reg_kernel_##type##__##DEVICE_TYPE##__;                           \
-  int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
-
-// (type, KernelType)
-#define REGISTER_OP_GPU_KERNEL(type, ...) \
-  REGISTER_OP_KERNEL(type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
-
-// (type, KernelType)
-#define REGISTER_OP_CPU_KERNEL(type, ...) \
-  REGISTER_OP_KERNEL(type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
+#define NO_GRADIENT(op_type)                           \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                      \
+      __reg_gradient_op__##op_type##_##op_type##_grad, \
+      "NO_GRADIENT must be called in global namespace")
+
+#define REGISTER_OP_GPU_KERNEL(op_type, ...) \
+  REGISTER_OP_KERNEL(op_type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
+
+#define REGISTER_OP_CPU_KERNEL(op_type, ...) \
+  REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
 
 /**
  * Macro to mark what Operator and Kernel we will use and tell the compiler to
  * link them into target.
  */
-#define USE_OP_WITHOUT_KERNEL(op_type)                      \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                           \
-      __use_op_without_kernel_##op_type,                    \
-      "USE_OP_WITHOUT_KERNEL must be in global namespace"); \
-  extern int __op_register_##op_type##_handle__();          \
-  static int __use_op_ptr_##op_type##_without_kernel__      \
-      __attribute__((unused)) = __op_register_##op_type##_handle__()
-
-#define USE_OP_KERNEL(op_type, DEVICE_TYPE)                               \
-  STATIC_ASSERT_GLOBAL_NAMESPACE(                                         \
-      __use_op_kernel_##op_type##_##DEVICE_TYPE##__,                      \
-      "USE_OP_KERNEL must be in global namespace");                       \
-  extern int __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__(); \
-  static int __use_op_ptr_##op_type##_##DEVICE_TYPE##_kernel__            \
-      __attribute__((unused)) =                                           \
-          __op_kernel_register_##op_type##_handle_##DEVICE_TYPE##__()
-
-// use Operator with only cpu kernel.
-#define USE_OP_CPU(op_type)       \
-  USE_OP_WITHOUT_KERNEL(op_type); \
-  USE_OP_KERNEL(op_type, CPU)
+#define USE_OP_ITSELF(op_type)                                    \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                 \
+      __use_op_itself_##op_type,                                  \
+      "USE_OP_ITSELF must be called in global namespace");        \
+  extern int TouchOpRegistrar_##op_type();                        \
+  static int use_op_itself_##op_type##_ __attribute__((unused)) = \
+      TouchOpRegistrar_##op_type()
+
+// TODO(fengjiayi): Most ops' gradient op have not been compeleted. So we use
+// `NO_GRAD` to disable micro USE_OP_GRADIENT(op_type). Otherwise the code can't
+// be compiled. `NO_GRAD` should be removed after all gradient ops are
+// compeleted.
+#define NO_GRAD
+#ifndef NO_GRAD
+#define USE_OP_GRADIENT(op_type)                                    \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                   \
+      __use_op_gradient_##op_type,                                  \
+      "USE_OP_GRADIENT must be called in global namespace");        \
+  extern int TouchOpGradientRegistrar_##op_type();                  \
+  static int use_op_gradient_##op_type##_ __attribute__((unused)) = \
+      TouchOpGradientRegistrar_##op_type()
+#else
+#define USE_OP_GRADIENT(op_type)
+#endif
+
+#define USE_OP_DEVICE_KERNEL(op_type, DEVICE_TYPE)               \
+  STATIC_ASSERT_GLOBAL_NAMESPACE(                                \
+      __use_op_kernel_##op_type##_##DEVICE_TYPE##__,             \
+      "USE_OP_DEVICE_KERNEL must be in global namespace");       \
+  extern int TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE(); \
+  static int use_op_kernel_##op_type##_##DEVICE_TYPE##_          \
+      __attribute__((unused)) =                                  \
+          TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE()
+
+// TODO(fengjiayi): The following macros seems ugly, do we have better method?
 
 #ifdef PADDLE_ONLY_CPU
-#define USE_OP(op_type) USE_OP_CPU(op_type)
+#define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU)
 #else
-#define USE_OP(op_type) \
-  USE_OP_CPU(op_type);  \
-  USE_OP_KERNEL(op_type, GPU)
+#define USE_OP_KERNEL(op_type)        \
+  USE_OP_DEVICE_KERNEL(op_type, CPU); \
+  USE_OP_DEVICE_KERNEL(op_type, GPU)
 #endif
 
+#define USE_NO_GRAD_OP(op_type) \
+  USE_OP_ITSELF(op_type);       \
+  USE_OP_KERNEL(op_type)
+
+#define USE_CPU_OP(op_type)           \
+  USE_OP_ITSELF(op_type);             \
+  USE_OP_DEVICE_KERNEL(op_type, CPU); \
+  USE_OP_GRADIENT(op_type)
+
+#define USE_OP(op_type)    \
+  USE_NO_GRAD_OP(op_type); \
+  USE_OP_GRADIENT(op_type)
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/op_registry_test.cc

@@ -7,6 +7,8 @@ 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

@@ -15,7 +15,6 @@ limitations under the License. */
 #pragma once
 
 #include <algorithm>
-#include <boost/variant.hpp>
 #include <string>
 #include <unordered_map>
 #include <vector>
@@ -27,25 +26,26 @@ limitations under the License. */
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
 #include "paddle/platform/place.h"
+#include "paddle/platform/variant.h"
 #include "paddle/utils/Error.h"
 
 namespace paddle {
 namespace framework {
 
 /// If a variable is a empty variable, that name will be used.
-const std::string kEmptyVarName = "@EMPTY@";
+constexpr char kEmptyVarName[] = "@EMPTY@";
 
 /// If a variable is a temporary variable, that name will be set in Python,
 /// but it will be convert to a unique name in scope after OpCreator.
-const std::string kTempVarName = "@TEMP@";
+constexpr char kTempVarName[] = "@TEMP@";
 
 /// If a variable's name has a certain suffix, it means that the
 /// variable is the gradient of another varibale.
 /// e.g. Variable "x@GRAD" is the gradient of varibale "x".
-const std::string kGradVarSuffix = "@GRAD";
+constexpr char kGradVarSuffix[] = "@GRAD";
 
 /// Variables with this suffix are supposed to be filled up with zeros.
-const std::string kZeroVarSuffix = "@ZERO";
+constexpr char kZeroVarSuffix[] = "@ZERO";
 
 inline std::string GradVarName(const std::string& var_name) {
   return var_name + kGradVarSuffix;
@@ -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>
@@ -88,21 +99,31 @@ class OperatorBase {
 
   virtual bool IsNetOp() const { return false; }
 
+  virtual bool SupportGPU() const { return false; }
+
   /// rename inputs outputs name
   void Rename(const std::string& old_name, const std::string& new_name);
 
   //! 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:
@@ -118,10 +139,10 @@ class OperatorBase {
   std::shared_ptr<std::unordered_map<std::string, int>> in_out_idxs_;
 };
 
-class OperatorContext {
+class InferShapeContext {
  public:
-  OperatorContext(const OperatorBase* op, const Scope& scope)
-      : op_(*op), scope_(scope) {}
+  InferShapeContext(const OperatorBase& op, const Scope& scope)
+      : op_(op), scope_(scope) {}
 
   size_t InputSize() const { return op_.inputs_.size(); }
 
@@ -232,12 +253,6 @@ class OperatorContext {
   const Scope& scope_;
 };
 
-class InferShapeContext : public OperatorContext {
- public:
-  InferShapeContext(const OperatorBase* op, const Scope& scope)
-      : OperatorContext(op, scope) {}
-};
-
 template <typename T>
 struct EigenDeviceConverter;
 
@@ -253,11 +268,11 @@ struct EigenDeviceConverter<platform::GPUPlace> {
 };
 #endif
 
-class ExecutionContext : public OperatorContext {
+class ExecutionContext : public InferShapeContext {
  public:
-  ExecutionContext(const OperatorBase* op, const Scope& scope,
+  ExecutionContext(const OperatorBase& op, const Scope& scope,
                    const platform::DeviceContext* device_context)
-      : OperatorContext(op, scope), device_context_(device_context) {}
+      : InferShapeContext(op, scope), device_context_(device_context) {}
 
   template <typename PlaceType,
             typename DeviceType =
@@ -285,6 +300,14 @@ 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_;
 
@@ -308,14 +331,14 @@ class OperatorWithKernel : public OperatorBase {
   using OpKernelMap =
       std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;
 
-  void InferShape(const Scope& scope) const {
-    InferShape(InferShapeContext(this, scope));
+  void InferShape(const Scope& scope) const override {
+    InferShape(InferShapeContext(*this, scope));
   }
 
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const final {
     auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
-    opKernel->Compute(ExecutionContext(this, scope, &dev_ctx));
+    opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
   }
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&
@@ -324,9 +347,25 @@ class OperatorWithKernel : public OperatorBase {
     return g_all_op_kernels;
   }
 
+  bool SupportGPU() const override {
+    OperatorWithKernel::OpKernelKey key;
+    key.place_ = platform::GPUPlace();
+    return OperatorWithKernel::AllOpKernels().at(type_).count(key) != 0;
+  }
+
  protected:
   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

@@ -18,13 +18,11 @@ limitations under the License. */
 
 #include "paddle/framework/backward.h"
 #include "paddle/framework/op_registry.h"
-#include "paddle/framework/operator.h"
-#include "paddle/framework/scope.h"
 #include "paddle/framework/tensor_py.h"
 #include "paddle/operators/net_op.h"
-#include "paddle/operators/type_alias.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
+#include "paddle/string/to_string.h"
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
 #include "pybind11/stl.h"
@@ -32,18 +30,23 @@ limitations under the License. */
 namespace py = pybind11;
 
 USE_OP(add_two);
-USE_OP_CPU(onehot_cross_entropy);
-USE_OP_WITHOUT_KERNEL(fc);
-USE_OP(sgd);
+USE_CPU_OP(onehot_cross_entropy);
+USE_NO_GRAD_OP(sgd);
 USE_OP(mul);
 USE_OP(mean);
 USE_OP(sigmoid);
 USE_OP(softmax);
 USE_OP(rowwise_add);
 USE_OP(fill_zeros_like);
-USE_OP_WITHOUT_KERNEL(recurrent_op);
+USE_OP_ITSELF(recurrent_op);
+USE_OP(gaussian_random);
+USE_OP(uniform_random);
+
 namespace paddle {
 namespace framework {
+
+using Tensor = framework::Tensor;
+
 template <typename ClassType>
 void ExposeOperator(ClassType &m) {
   m.def("infer_shape", &ClassType::type::InferShape)
@@ -56,6 +59,26 @@ void ExposeOperator(ClassType &m) {
            [](const typename ClassType::type &op) -> std::vector<std::string> {
              return op.outputs_;
            })
+      .def("inputs",
+           [](const typename ClassType::type &op) -> std::vector<std::string> {
+             return op.inputs_;
+           })
+      .def("support_gpu", &ClassType::type::SupportGPU)
+      .def("temp_outputs",
+           [](const typename ClassType::type &op) -> std::vector<std::string> {
+             auto iter = op.attrs_.find("temporary_index");
+             std::vector<std::string> ret;
+             if (iter == op.attrs_.end()) {
+               return ret;
+             } else {
+               auto tmp_idx = boost::get<std::vector<int>>(iter->second);
+               ret.reserve(tmp_idx.size());
+               for (auto &index : tmp_idx) {
+                 ret.push_back(op.outputs_.at(index));
+               }
+               return ret;
+             }
+           })
       .def("__str__", &ClassType::type::DebugString);
 }
 
@@ -129,8 +152,8 @@ All parameter, weight, gradient are variables in Paddle.
            [](Variable &self) -> Tensor * { return self.GetMutable<Tensor>(); },
            py::return_value_policy::reference)
       .def("get_net",
-           [](Variable &self) -> ops::NetOp * {
-             return self.GetMutable<ops::NetOp>();
+           [](Variable &self) -> operators::NetOp * {
+             return self.GetMutable<operators::NetOp>();
            },
            py::return_value_policy::reference);
 
@@ -184,9 +207,13 @@ All parameter, weight, gradient are variables in Paddle.
                   });
   // clang-format on
 
-  py::class_<paddle::platform::GPUPlace>(m, "GPUPlace").def(py::init<int>());
+  py::class_<platform::GPUPlace>(m, "GPUPlace")
+      .def(py::init<int>())
+      .def("__str__", string::to_string<const platform::GPUPlace &>);
 
-  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace").def(py::init<>());
+  py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
+      .def(py::init<>())
+      .def("__str__", string::to_string<const platform::CPUPlace &>);
 
   py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
       m, "Operator");
@@ -201,8 +228,6 @@ All parameter, weight, gradient are variables in Paddle.
     return OpRegistry::CreateOp(desc);
   });
 
-  operator_base.def_static("support_gpu", &OpRegistry::SupportGPU);
-
   operator_base.def("backward",
                     [](const OperatorBase &forwardOp,
                        const std::unordered_set<std::string> &no_grad_vars) {
@@ -211,23 +236,24 @@ All parameter, weight, gradient are variables in Paddle.
 
   ExposeOperator(operator_base);
 
-  py::class_<ops::NetOp, std::shared_ptr<ops::NetOp>> net(m, "Net");
+  py::class_<operators::NetOp, std::shared_ptr<operators::NetOp>> net(m, "Net");
 
   net.def_static("create",
-                 []() -> std::shared_ptr<ops::NetOp> {
-                   auto retv = std::make_shared<ops::NetOp>();
+                 []() -> std::shared_ptr<operators::NetOp> {
+                   auto retv = std::make_shared<operators::NetOp>();
                    retv->type_ = "plain_net";
                    return retv;
                  })
-      .def("add_op", &ops::NetOp::AddOp)
-      .def(
-          "add_op",
-          [](ops::NetOp &self, const std::shared_ptr<ops::NetOp> &net) -> void {
-            self.AddOp(std::static_pointer_cast<OperatorBase>(net));
-          })
-      .def("complete_add_op", &ops::NetOp::CompleteAddOp)
-      .def("complete_add_op",
-           [](std::shared_ptr<ops::NetOp> &self) { self->CompleteAddOp(); });
+      .def("add_op", &operators::NetOp::AddOp)
+      .def("add_op",
+           [](operators::NetOp &self,
+              const std::shared_ptr<operators::NetOp> &net) -> void {
+             self.AddOp(std::static_pointer_cast<OperatorBase>(net));
+           })
+      .def("complete_add_op", &operators::NetOp::CompleteAddOp)
+      .def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
+        self->CompleteAddOp();
+      });
 
   ExposeOperator(net);
 

paddle/framework/tensor.h

@@ -18,6 +18,8 @@ limitations under the License. */
 #include <cstring>
 #include <memory>
 #include <typeindex>
+#include <vector>
+
 #include "paddle/framework/ddim.h"
 #include "paddle/memory/memory.h"
 #include "paddle/platform/device_context.h"
@@ -77,11 +79,11 @@ class Tensor {
   inline const DDim& dims() const;
 
   /*! Resize the dimensions of the memory block. */
-  inline void Resize(const DDim& dims);
+  inline Tensor& Resize(const DDim& dims);
 
   /*! The internal of two tensors share the same memory block. */
   template <typename T>
-  inline void ShareDataWith(const Tensor& src);
+  inline Tensor& ShareDataWith(const Tensor& src);
 
   /**
    * @brief   Copy the content of external tensor to a new place.

paddle/framework/tensor_impl.h

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

paddle/framework/tensor_test.cc

@@ -19,7 +19,7 @@ TEST(Tensor, Dims) {
   using namespace paddle::framework;
   using namespace paddle::platform;
   Tensor tt;
-  tt.Resize(make_ddim({2, 3, 4}));
+  tt.Resize({2, 3, 4});
   DDim dims = tt.dims();
   ASSERT_EQ(arity(dims), 3);
   for (int i = 0; i < 3; ++i) {

paddle/function/CMakeLists.txt

@@ -38,10 +38,11 @@ if(WITH_GPU)
     add_simple_unittest(RowConvOpTest)
     add_simple_unittest(BlockExpandOpTest)
     add_simple_unittest(CropOpTest)
+    add_simple_unittest(DepthwiseConvOpTest)
 endif()
 
-add_simple_unittest(ConvOpTest)
 add_simple_unittest(Im2ColTest)
+add_simple_unittest(GemmConvOpTest)
 endif()
 
 add_style_check_target(paddle_function ${h_files})

paddle/function/ConvOpTest.cpp

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include <gtest/gtest.h>
-#include <memory>
-#include "Function.h"
-#include "FunctionTest.h"
-
-namespace paddle {
-
-enum TestType {
-  kForwardTest = 0,
-  kBackwardInputTest = 1,
-  kBackwardFilterTest = 2,
-};
-
-template <DeviceType DType1, DeviceType DType2>
-class ConvolutionTest {
-public:
-  ConvolutionTest(const std::string& conv1,
-                  const std::string& conv2,
-                  TestType type,
-                  bool useGroups = true,
-                  std::string algo = "auto") {
-    for (size_t batchSize : {1, 32}) {
-      for (size_t inputSize : {7, 14, 54}) {
-        for (size_t filterSize : {1, 3, 5}) {
-          for (size_t inputChannels : {3, 64}) {
-            for (size_t outputChannels : {3, 64}) {
-              if (inputChannels > outputChannels) break;
-              size_t groups;
-              if (!useGroups) {
-                groups = 1;
-              } else {
-                if (outputChannels % inputChannels != 0) continue;
-                groups = inputChannels;
-              }
-
-              for (size_t stride : {1, 2}) {
-                for (size_t padding : {0, 1}) {
-                  if (padding >= filterSize) break;
-                  size_t outputSize =
-                      (inputSize - filterSize + 2 * padding + stride) / stride;
-                  VLOG(3) << " batchSize=" << batchSize
-                          << " inputChannels=" << inputChannels
-                          << " inputHeight=" << inputSize
-                          << " inputWidth=" << inputSize
-                          << " outputChannels=" << outputChannels
-                          << " filterHeight=" << filterSize
-                          << " filterWidth=" << filterSize
-                          << " outputHeight=" << outputSize
-                          << " outputWidth=" << outputSize
-                          << " stride=" << stride << " padding=" << padding;
-
-                  std::vector<size_t> paddings = {padding, padding};
-                  std::vector<size_t> strides = {stride, stride};
-                  Compare2Function<DType1, DType2> test(
-                      conv1,
-                      conv2,
-                      FuncConfig()
-                          .set("paddings", paddings)
-                          .set("strides", strides)
-                          .set("groups", groups)
-                          .set("algo", algo));
-
-                  TensorShape input{
-                      batchSize, inputChannels, inputSize, inputSize};
-
-                  TensorShape filter;
-                  if (groups > 1)
-                    filter = TensorShape({groups,
-                                          outputChannels / groups,
-                                          inputChannels / groups,
-                                          filterSize,
-                                          filterSize});
-                  else
-                    filter = TensorShape({outputChannels,
-                                          inputChannels,
-                                          filterSize,
-                                          filterSize});
-                  TensorShape output{
-                      batchSize, outputChannels, outputSize, outputSize};
-
-                  if (type == kForwardTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.run();
-                  } else if (type == kBackwardInputTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO);
-                    test.run();
-                  } else if (type == kBackwardFilterTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter),
-                                    ADD_TO);
-                    test.run();
-                  }
-                }
-              }
-            }
-          }
-        }
-      }
-    }
-  }
-};
-
-// Mainly used to test cases where the height and width (input, filter)
-// are not equal.
-template <DeviceType DType1, DeviceType DType2>
-class ConvolutionTest2 {
-public:
-  ConvolutionTest2(const std::string& conv1,
-                   const std::string& conv2,
-                   TestType type,
-                   bool useGroups = true,
-                   std::string algo = "auto") {
-    for (size_t batchSize : {16}) {
-      for (size_t inputHeight : {7, 31}) {
-        for (size_t inputWidth : {10, 54}) {
-          for (size_t filterHeight : {1, 5}) {
-            for (size_t filterWidth : {3, 7}) {
-              for (size_t inputChannels : {7}) {
-                for (size_t outputChannels : {7}) {
-                  size_t groups;
-                  if (!useGroups) {
-                    groups = 1;
-                  } else {
-                    if (outputChannels % inputChannels != 0) continue;
-                    groups = inputChannels;
-                  }
-
-                  size_t stride = 1;
-                  size_t padding = 0;
-                  size_t outputHeight =
-                      (inputHeight - filterHeight + 2 * padding + stride) /
-                      stride;
-                  size_t outputWidth =
-                      (inputWidth - filterWidth + 2 * padding + stride) /
-                      stride;
-                  VLOG(3) << " batchSize=" << batchSize
-                          << " inputChannels=" << inputChannels
-                          << " inputHeight=" << inputHeight
-                          << " inputWidth=" << inputWidth
-                          << " outputChannels=" << outputChannels
-                          << " filterHeight=" << filterHeight
-                          << " filterWidth=" << filterWidth
-                          << " outputHeight=" << outputHeight
-                          << " outputWidth=" << outputWidth
-                          << " stride=" << stride << " padding=" << padding;
-
-                  std::vector<size_t> paddings = {padding, padding};
-                  std::vector<size_t> strides = {stride, stride};
-                  Compare2Function<DType1, DType2> test(
-                      conv1,
-                      conv2,
-                      FuncConfig()
-                          .set("paddings", paddings)
-                          .set("strides", strides)
-                          .set("groups", groups)
-                          .set("algo", algo));
-
-                  TensorShape input{
-                      batchSize, inputChannels, inputHeight, inputWidth};
-
-                  TensorShape filter;
-                  if (groups > 1)
-                    filter = TensorShape({groups,
-                                          outputChannels / groups,
-                                          inputChannels / groups,
-                                          filterHeight,
-                                          filterWidth});
-                  else
-                    filter = TensorShape({outputChannels,
-                                          inputChannels,
-                                          filterHeight,
-                                          filterWidth});
-                  TensorShape output{
-                      batchSize, outputChannels, outputHeight, outputWidth};
-
-                  if (type == kForwardTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.run();
-                  } else if (type == kBackwardInputTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO);
-                    test.run();
-                  } else if (type == kBackwardFilterTest) {
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
-                    test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
-                    test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter),
-                                    ADD_TO);
-                    test.run();
-                  }
-                }
-              }
-            }
-          }
-        }
-      }
-    }
-  }
-};
-
-// ======Start Convolution TEST======
-
-TEST(Forward, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU> test(
-      "NaiveConv-CPU", "GemmConv-CPU", kForwardTest, false);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU> test2(
-      "NaiveConv-CPU", "GemmConv-CPU", kForwardTest, false);
-}
-
-#ifndef PADDLE_ONLY_CPU
-TEST(Forward, GEMM2) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConv-CPU", "GemmConv-GPU", kForwardTest, false);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConv-CPU", "GemmConv-GPU", kForwardTest, false);
-}
-
-TEST(BackwardInput, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConvGradInput-CPU",
-      "GemmConvGradInput-GPU",
-      kBackwardInputTest,
-      false);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConvGradInput-CPU",
-      "GemmConvGradInput-GPU",
-      kBackwardInputTest,
-      false);
-}
-
-TEST(BackwardFilter, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConvGradFilter-CPU",
-      "GemmConvGradFilter-GPU",
-      kBackwardFilterTest,
-      false);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConvGradFilter-CPU",
-      "GemmConvGradFilter-GPU",
-      kBackwardFilterTest,
-      false);
-}
-#endif
-// ======End Convolution TEST======
-
-// ======Start DepthwiseConvolution TEST======
-
-// TODO(zhaolong) The depthwise convolution cpu test will be added when the cpu
-// version of depthwiseConv is implemented.
-
-#ifndef PADDLE_ONLY_CPU
-
-TEST(DepthwiseConvForward, GEMM2) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConv-CPU", "DepthwiseConv-GPU", kForwardTest);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConv-CPU", "DepthwiseConv-GPU", kForwardTest);
-}
-
-TEST(DepthwiseConvBackwardInput, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConvGradInput-CPU",
-      "DepthwiseConvGradInput-GPU",
-      kBackwardInputTest);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConvGradInput-CPU",
-      "DepthwiseConvGradInput-GPU",
-      kBackwardInputTest);
-}
-
-TEST(DepthwiseConvBackwardFilter, GEMM) {
-  ConvolutionTest<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test(
-      "GemmConvGradFilter-CPU",
-      "DepthwiseConvGradFilter-GPU",
-      kBackwardFilterTest);
-  ConvolutionTest2<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU> test2(
-      "GemmConvGradFilter-CPU",
-      "DepthwiseConvGradFilter-GPU",
-      kBackwardFilterTest);
-}
-
-#endif
-// ======End DepthwiseConvolution TEST======
-
-}  // namespace paddle

paddle/function/ConvOpTest.h

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "FunctionTest.h"
+
+namespace paddle {
+
+template <DeviceType DType1, DeviceType DType2>
+void forward(Compare2Function<DType1, DType2>& test,
+             const TensorShape& input,
+             const TensorShape& filter,
+             const TensorShape& output) {
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, output));
+  test.run();
+}
+
+template <DeviceType DType1, DeviceType DType2>
+void backward_input(Compare2Function<DType1, DType2>& test,
+                    const TensorShape& input,
+                    const TensorShape& filter,
+                    const TensorShape& output) {
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, filter));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, input), ADD_TO);
+  test.run();
+}
+
+template <DeviceType DType1, DeviceType DType2>
+void backward_filter(Compare2Function<DType1, DType2>& test,
+                     const TensorShape& input,
+                     const TensorShape& filter,
+                     const TensorShape& output) {
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, output));
+  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, input));
+  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, filter), ADD_TO);
+  test.run();
+}
+
+template <DeviceType DType1, DeviceType DType2>
+using Function = void (*)(Compare2Function<DType1, DType2>& test,
+                          const TensorShape& input,
+                          const TensorShape& filter,
+                          const TensorShape& output);
+
+/**
+ * \brief A basic convolution function test interface.
+ *
+ * \param conv1         type name of convolution function 1.
+ * \param conv2         type name of convolution function 2.
+ * \param function      test function, can be one of the forward, backward_input
+ *                      backward_filter function.
+ * Example:
+ * 1. Compare GemmConv's CPU and GPU implementation:
+ *   Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_GPU>(
+ *      "GemmConv-CPU", "GemmConv-GPU", forward);
+ */
+template <DeviceType DType1, DeviceType DType2>
+void Convolution(const std::string& conv1,
+                 const std::string& conv2,
+                 Function<DType1, DType2> function) {
+  for (size_t batchSize : {1, 5}) {
+    for (size_t inputSize : {7, 14, 31}) {
+      for (size_t filterSize : {1, 3, 5}) {
+        for (size_t inputChannels : {3, 16}) {
+          for (size_t outputChannels : {3, 16}) {
+            if (outputChannels < inputChannels) continue;
+            for (size_t stride : {1, 2}) {
+              for (size_t padding : {0, 1}) {
+                if (padding >= filterSize) break;
+
+                // NNPACK only supports stride = 1 if batchSize > 1
+                if ((conv1 == "NNPACKConv-CPU" || conv2 == "NNPACKConv-CPU") &&
+                    batchSize > 1 && stride > 1)
+                  break;
+
+                size_t outputSize =
+                    (inputSize - filterSize + 2 * padding + stride) / stride;
+                VLOG(3) << " batchSize=" << batchSize
+                        << " inputChannels=" << inputChannels
+                        << " inputHeight=" << inputSize
+                        << " inputWidth=" << inputSize
+                        << " outputChannels=" << outputChannels
+                        << " filterHeight=" << filterSize
+                        << " filterWidth=" << filterSize
+                        << " outputHeight=" << outputSize
+                        << " outputWidth=" << outputSize << " stride=" << stride
+                        << " padding=" << padding;
+
+                std::vector<size_t> paddings = {padding, padding};
+                std::vector<size_t> strides = {stride, stride};
+                Compare2Function<DType1, DType2> test(
+                    conv1,
+                    conv2,
+                    FuncConfig()
+                        .set("paddings", paddings)
+                        .set("strides", strides)
+                        .set("groups", (size_t)1)
+                        .set("algo", (std::string) "auto"));
+
+                TensorShape input{
+                    batchSize, inputChannels, inputSize, inputSize};
+                TensorShape filter{
+                    outputChannels, inputChannels, filterSize, filterSize};
+                TensorShape output{
+                    batchSize, outputChannels, outputSize, outputSize};
+
+                function(test, input, filter, output);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief A convolution function test interface for
+ *        image height is not equal image width.
+ */
+template <DeviceType DType1, DeviceType DType2>
+void Convolution2(const std::string& conv1,
+                  const std::string& conv2,
+                  Function<DType1, DType2> function) {
+  for (size_t batchSize : {4}) {
+    for (size_t inputHeight : {7, 31}) {
+      for (size_t inputWidth : {10, 54}) {
+        for (size_t filterHeight : {1, 5}) {
+          for (size_t filterWidth : {3, 7}) {
+            for (size_t inputChannels : {7}) {
+              for (size_t outputChannels : {7}) {
+                size_t stride = 1;
+                size_t padding = 0;
+                size_t outputHeight =
+                    (inputHeight - filterHeight + 2 * padding + stride) /
+                    stride;
+                size_t outputWidth =
+                    (inputWidth - filterWidth + 2 * padding + stride) / stride;
+                VLOG(3) << " batchSize=" << batchSize
+                        << " inputChannels=" << inputChannels
+                        << " inputHeight=" << inputHeight
+                        << " inputWidth=" << inputWidth
+                        << " outputChannels=" << outputChannels
+                        << " filterHeight=" << filterHeight
+                        << " filterWidth=" << filterWidth
+                        << " outputHeight=" << outputHeight
+                        << " outputWidth=" << outputWidth
+                        << " stride=" << stride << " padding=" << padding;
+
+                std::vector<size_t> paddings = {padding, padding};
+                std::vector<size_t> strides = {stride, stride};
+                Compare2Function<DType1, DType2> test(
+                    conv1,
+                    conv2,
+                    FuncConfig()
+                        .set("paddings", paddings)
+                        .set("strides", strides)
+                        .set("groups", (size_t)1)
+                        .set("algo", (std::string) "auto"));
+
+                TensorShape input{
+                    batchSize, inputChannels, inputHeight, inputWidth};
+                TensorShape filter{
+                    outputChannels, inputChannels, filterHeight, filterWidth};
+                TensorShape output{
+                    batchSize, outputChannels, outputHeight, outputWidth};
+
+                function(test, input, filter, output);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+/**
+ * \brief A convolution function test interface for depthwise convolution.
+ */
+template <DeviceType DType1, DeviceType DType2>
+void DepthwiseConvolution(const std::string& conv1,
+                          const std::string& conv2,
+                          Function<DType1, DType2> function) {
+  for (size_t batchSize : {1, 32}) {
+    for (size_t inputSize : {7, 14, 54}) {
+      for (size_t filterSize : {3, 4}) {
+        for (size_t inputChannels : {32}) {
+          for (size_t outputChannels : {32, 64}) {
+            for (size_t stride : {1, 2}) {
+              for (size_t padding : {0, 1}) {
+                // NNPACK only supports stride = 1 if batchSize > 1,
+                // and there has some bug when batchSize > 1 and groups != 1
+                if ((conv1 == "NNPACKConv-CPU" || conv2 == "NNPACKConv-CPU") &&
+                    batchSize > 1)
+                  break;
+
+                size_t outputSize =
+                    (inputSize - filterSize + 2 * padding + stride) / stride;
+                VLOG(3) << " batchSize=" << batchSize
+                        << " inputChannels=" << inputChannels
+                        << " inputHeight=" << inputSize
+                        << " inputWidth=" << inputSize
+                        << " outputChannels=" << outputChannels
+                        << " filterHeight=" << filterSize
+                        << " filterWidth=" << filterSize
+                        << " outputHeight=" << outputSize
+                        << " outputWidth=" << outputSize << " stride=" << stride
+                        << " padding=" << padding;
+
+                std::vector<size_t> paddings = {padding, padding};
+                std::vector<size_t> strides = {stride, stride};
+                size_t groups = inputChannels;
+                Compare2Function<DType1, DType2> test(
+                    conv1,
+                    conv2,
+                    FuncConfig()
+                        .set("paddings", paddings)
+                        .set("strides", strides)
+                        .set("groups", groups)
+                        .set("algo", (std::string) "auto"));
+
+                TensorShape input{
+                    batchSize, inputChannels, inputSize, inputSize};
+                TensorShape filter{groups,
+                                   outputChannels / groups,
+                                   inputChannels / groups,
+                                   filterSize,
+                                   filterSize};
+                TensorShape output{
+                    batchSize, outputChannels, outputSize, outputSize};
+
+                function(test, input, filter, output);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace paddle

paddle/operators/mean_op_test.cc → paddle/function/DepthwiseConvOpTest.cpp

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

paddle/function/FunctionTest.cpp

@@ -93,8 +93,8 @@ TEST(Arguments, Matrix) {
   MatrixPtr matrix = Matrix::create(100, 200);
   CheckBufferArg check = [=](const BufferArg& arg) {
     EXPECT_EQ(arg.shape().ndims(), 2U);
-    EXPECT_EQ(arg.shape()[0], 100);
-    EXPECT_EQ(arg.shape()[1], 200);
+    EXPECT_EQ(arg.shape()[0], 100U);
+    EXPECT_EQ(arg.shape()[1], 200U);
     EXPECT_EQ(arg.data(), matrix->getData());
 
     EXPECT_EQ(arg.matrix<DEVICE_TYPE_CPU>().getHeight(), matrix->getHeight());
@@ -112,8 +112,8 @@ TEST(Arguments, Matrix) {
 TEST(Arguments, Vector) {
   VectorPtr vector = Vector::create(100, false);
   CheckBufferArg check = [=](const BufferArg& arg) {
-    EXPECT_EQ(arg.shape().ndims(), 1);
-    EXPECT_EQ(arg.shape()[0], 100);
+    EXPECT_EQ(arg.shape().ndims(), 1U);
+    EXPECT_EQ(arg.shape()[0], 100U);
     EXPECT_EQ(arg.data(), vector->getData());
 
     CpuVector inVector = arg.vector<real, DEVICE_TYPE_CPU>();
@@ -131,9 +131,9 @@ TEST(Arguments, Vector) {
 TEST(Arguments, CpuSparseMatrix) {
   CpuSparseMatrix sparse(200, 300, 50);
   CheckBufferArg check = [=](const BufferArg& arg) {
-    EXPECT_EQ(arg.shape().ndims(), 2);
-    EXPECT_EQ(arg.shape()[0], 200);
-    EXPECT_EQ(arg.shape()[1], 300);
+    EXPECT_EQ(arg.shape().ndims(), 2U);
+    EXPECT_EQ(arg.shape()[0], 200U);
+    EXPECT_EQ(arg.shape()[1], 300U);
     EXPECT_EQ(arg.data(), sparse.getData());
     // CHECK_EQ(arg.sparse().nnz(), 50);
     // CHECK_EQ(arg.sparse().dataFormat(), SPARSE_CSR_FORMAT);
@@ -152,10 +152,10 @@ TEST(Arguments, CpuSparseMatrix) {
 TEST(Arguments, BufferArg) {
   BufferArg arg(nullptr, VALUE_TYPE_FLOAT, {1, 2, 3});
   CheckBufferArg check = [=](const BufferArg& arg) {
-    EXPECT_EQ(arg.shape().ndims(), 3);
-    EXPECT_EQ(arg.shape()[0], 1);
-    EXPECT_EQ(arg.shape()[1], 2);
-    EXPECT_EQ(arg.shape()[2], 3);
+    EXPECT_EQ(arg.shape().ndims(), 3U);
+    EXPECT_EQ(arg.shape()[0], 1U);
+    EXPECT_EQ(arg.shape()[1], 2U);
+    EXPECT_EQ(arg.shape()[2], 3U);
   };
 
   BufferArgs argments;

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

@@ -13,10 +13,38 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <gtest/gtest.h>
-#include <paddle/framework/op_registry.h>
-USE_OP(sgd);
-TEST(SGDOp, GetOpProto) {
-  auto& protos = paddle::framework::OpRegistry::protos();
-  auto it = protos.find("sgd");
-  ASSERT_NE(it, protos.end());
+#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/TensorShapeTest.cpp

@@ -44,7 +44,7 @@ TEST(TensorShape, GetAndSet) {
   EXPECT_EQ(t.ndims(), 3U);
   EXPECT_EQ(t.getElements(), 6U);
 
-  EXPECT_EQ(t[1], 2);
+  EXPECT_EQ(t[1], 2U);
   t.setDim(1, 100);
   EXPECT_EQ(t.getElements(), 300U);
   EXPECT_EQ(t[1], 100U);

paddle/function/nnpack/NNPACKConvOp.cpp

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

paddle/function/nnpack/NNPACKConvOpTest.cpp

@@ -13,87 +13,18 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include <gtest/gtest.h>
-#include "paddle/function/Function.h"
-#include "paddle/function/FunctionTest.h"
-
-DEFINE_string(algo,
-              "auto",
-              "The algorithm (auto, ft8x8, ft16x16, wt8x8, "
-              "implicit-gemm, or direct) for computing convolution of NNPACK.");
+#include "paddle/function/ConvOpTest.h"
 
 namespace paddle {
 
-#define IS_NNPACK_SUPPORT(algo, filterSize, stride)        \
-  if (algo == "direct" && filterSize != 1) continue;       \
-  if (algo == "direct" && batchSize != 1) continue;        \
-  if (algo == "wt8x8" && filterSize != 3) continue;        \
-  if (algo == "implicit-gemm" && batchSize != 1) continue; \
-  if (algo != "auto" && algo != "implicit-gemm" && stride > 1) continue;
-
-class ConvolutionTest {
-public:
-  ConvolutionTest(const std::string& conv1,
-                  const std::string& conv2,
-                  std::string algo = "auto") {
-    for (size_t batchSize : {1, 32}) {
-      for (size_t inputSize : {7, 14, 54}) {
-        for (size_t filterSize : {1, 3, 5}) {
-          for (size_t inputChannels : {3, 64}) {
-            for (size_t outputChannels : {3, 64, 128}) {
-              if (inputChannels < outputChannels) break;
-              for (size_t stride : {1, 2}) {
-                // if batchSize > 1 NNPACKConv only supports stride = 1
-                if (batchSize > 1 && stride > 1) break;
-                for (size_t padding : {0, 1}) {
-                  if (padding >= filterSize) break;
-                  size_t outputSize =
-                      (inputSize - filterSize + 2 * padding + stride) / stride;
-                  IS_NNPACK_SUPPORT(algo, filterSize, stride);
-                  LOG(INFO) << " batchSize=" << batchSize
-                            << " inputChannels=" << inputChannels
-                            << " inputHeight=" << inputSize
-                            << " inputWidth=" << inputSize
-                            << " outputChannels=" << outputChannels
-                            << " filterHeight=" << filterSize
-                            << " filterWidth=" << filterSize
-                            << " outputHeight=" << outputSize
-                            << " outputWidth=" << outputSize
-                            << " stride=" << stride << " padding=" << padding;
-
-                  std::vector<size_t> paddings = {padding, padding};
-                  std::vector<size_t> strides = {stride, stride};
-                  Compare2Function<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU> test(
-                      conv1,
-                      conv2,
-                      FuncConfig()
-                          .set("paddings", paddings)
-                          .set("strides", strides)
-                          .set("groups", (size_t)1)
-                          .set("algo", algo));
-
-                  TensorShape shape0{
-                      batchSize, inputChannels, inputSize, inputSize};
-                  TensorShape shape1{
-                      outputChannels, inputChannels, filterSize, filterSize};
-                  TensorShape shape2{
-                      batchSize, outputChannels, outputSize, outputSize};
-                  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape0));
-                  test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape1));
-                  test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape2));
-                  test.run();
-                }
-              }
-            }
-          }
-        }
-      }
-    }
-  }
-};
+TEST(NNPACK, Forward) {
+  Convolution<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
+      "GemmConv-CPU", "NNPACKConv-CPU", forward);
+}
 
-TEST(Convolution, NNPACK) {
-  // NNPACK only supports stride = 1
-  ConvolutionTest test("GemmConv-CPU", "NNPACKConv-CPU", FLAGS_algo);
+TEST(NNPACK, Depthwise) {
+  DepthwiseConvolution<DEVICE_TYPE_CPU, DEVICE_TYPE_CPU>(
+      "GemmConv-CPU", "NNPACKConv-CPU", forward);
 }
 
 }  // namespace paddle

paddle/gserver/CMakeLists.txt

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

paddle/gserver/activations/ActivationFunction.cpp

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

paddle/gserver/layers/KmaxSeqScoreLayer.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 "Layer.h"
+
+namespace paddle {
+
+class KmaxSeqScoreLayer : public Layer {
+private:
+  MatrixPtr scores_;
+  size_t beamSize_;
+  void kmaxScorePerSeq(const real* score,
+                       real* sortedRes,
+                       const ICpuGpuVectorPtr seqStartPos);
+
+public:
+  explicit KmaxSeqScoreLayer(const LayerConfig& config) : Layer(config) {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void forward(PassType passType) override;
+  void backward(const UpdateCallback& callback = nullptr) override;
+};
+
+REGISTER_LAYER(kmax_seq_score, KmaxSeqScoreLayer);
+
+bool KmaxSeqScoreLayer::init(const LayerMap& layerMap,
+                             const ParameterMap& parameterMap) {
+  bool ret = Layer::init(layerMap, parameterMap);
+  CHECK_EQ(1U, inputLayers_.size());
+
+  beamSize_ = config_.beam_size();
+  CHECK_GE(beamSize_, 1U);
+
+  setNeedSequenceInfo(false);
+  setNeedGradient(false);
+  return ret;
+}
+
+void KmaxSeqScoreLayer::kmaxScorePerSeq(const real* scores,
+                                        real* sortedIds,
+                                        const ICpuGpuVectorPtr seqStartPos) {
+  int* starts = seqStartPos->getMutableData(false);
+  std::vector<real> indices;
+  for (size_t i = 0; i < seqStartPos->getSize() - 1; ++i) {
+    int seqLen = starts[i + 1] - starts[i];
+    int k = std::min(static_cast<int>(beamSize_), seqLen);
+
+    indices.resize(seqLen, 0);
+    std::iota(begin(indices), end(indices), 0.);
+    std::vector<real> tmpScore(scores + starts[i], scores + starts[i + 1]);
+    std::partial_sort(
+        begin(indices),
+        begin(indices) + k,
+        end(indices),
+        [&](size_t a, size_t b) { return tmpScore[a] > tmpScore[b]; });
+    memcpy(sortedIds + (i * beamSize_), indices.data(), k * sizeof(real));
+  }
+}
+
+void KmaxSeqScoreLayer::forward(PassType passType) {
+  Layer::forward(passType);
+
+  const Argument& input = getInput(0);
+  const MatrixPtr inputScore = getInputValue(0);
+
+  CHECK(input.hasSeq() || input.hasSubseq())
+      << "input of " << getName()
+      << " must be a sequence or a nested sequence.";
+  CHECK_EQ(input.value->getWidth(), 1UL)
+      << "input of " << getName()
+      << " is score over a sequence or a nested sequence, so its width "
+      << " must be 1.";
+
+  if (useGpu_) {
+    // this Layer runs only in CPU, if the model is runing on GPU,
+    // then copy the input to this layer from GPU to CPU.
+    Matrix::resizeOrCreate(scores_,
+                           inputScore->getHeight(),
+                           1,
+                           false /* trans */,
+                           false /* useGpu */);
+    scores_->copyFrom(*inputScore);
+  } else {
+    scores_ = inputScore;
+  }
+
+  Matrix::resizeOrCreate(
+      output_.value,
+      input.hasSubseq() ? input.getNumSubSequences() : input.getNumSequences(),
+      beamSize_,
+      false,
+      false);
+  output_.value->one();
+  output_.value->mulScalar(-1.);
+
+  kmaxScorePerSeq(scores_->getData(),
+                  output_.value->getData(),
+                  input.hasSubseq() ? input.subSequenceStartPositions
+                                    : input.sequenceStartPositions);
+}
+
+void KmaxSeqScoreLayer::backward(const UpdateCallback& callback) {}
+
+}  // namespace paddle

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/layers/SubNestedSequenceLayer.cpp

@@ -96,7 +96,7 @@ void SubNestedSequenceLayer::calSelectedCols(
   for (size_t i = 0; i < seqNum; ++i) {
     for (size_t j = 0; j < beamSize; ++j) {
       if (selectedIndices->getElement(i, j) == -1.) break;
-      int selSubSeqIdx = selectedIndices->getElement(i, j);
+      size_t selSubSeqIdx = selectedIndices->getElement(i, j);
       CHECK_GT(inputSeqInfoVec_[i].size() - 1, selSubSeqIdx);
 
       size_t subSeqLen = inputSeqInfoVec_[i][selSubSeqIdx + 1] -
@@ -135,7 +135,7 @@ void SubNestedSequenceLayer::forward(PassType passType) {
   CHECK(inputSeq.hasSubseq()) << "The first input of SubNestSequence layer "
                               << "must be a nested sequence.";
   const MatrixPtr selectedIndices = getInputValue(1);
-  CHECK_EQ(inputSeq.getNumSequences(), selectedIndices->getHeight());
+  CHECK_EQ(size_t(inputSeq.getNumSequences()), selectedIndices->getHeight());
 
   if (dynamic_cast<GpuMatrix*>(selectedIndices.get())) {
     /*

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
@@ -66,6 +75,16 @@ add_unittest_without_exec(test_BatchNorm
 
 add_test(NAME test_BatchNorm
     COMMAND test_BatchNorm)
+
+
+################# test_KmaxSeqScore #######################
+add_unittest_without_exec(test_KmaxSeqScore
+    test_KmaxSeqScore.cpp
+    LayerGradUtil.cpp)
+
+add_test(NAME test_KmaxSeqScore
+    COMMAND test_KmaxSeqScore)
+
 ################## test_Evaluator #######################
 add_unittest(test_Evaluator
     test_Evaluator.cpp)
@@ -82,8 +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 #######################
@@ -96,7 +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 ###############
@@ -106,20 +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()
 
 
@@ -127,6 +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/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/add_op.cu

@@ -16,4 +16,6 @@
 #include "paddle/framework/op_registry.h"
 #include "paddle/operators/add_op.h"
 
-REGISTER_OP_GPU_KERNEL(add_two, ops::AddKernel<ops::GPUPlace, float>);
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(add_two,
+                       ops::AddKernel<paddle::platform::GPUPlace, float>);

paddle/operators/cross_entropy_op.cu

@@ -14,3 +14,8 @@
 
 #define EIGEN_USE_GPU
 #include "paddle/operators/cross_entropy_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    onehot_cross_entropy,
+    ops::OnehotCrossEntropyOpKernel<paddle::platform::GPUPlace, float>);