diff --git a/.gitignore b/.gitignore
index 4f21fefda9f64a0392881971a715b97c234030e3..351b8204100dfd71e94cb3efa2e946b44b9e4285 100644
--- a/.gitignore
+++ b/.gitignore
@@ -27,3 +27,4 @@ CMakeFiles
 cmake_install.cmake
 paddle/.timestamp
 python/paddlepaddle.egg-info/
+paddle/pybind/pybind.h
diff --git a/.travis.yml b/.travis.yml
index e217c8f5a740ef5ab7315656ed7839ffa219c805..d0e2696f100e55f320e410afd6a3038db647f76f 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -36,10 +36,6 @@ before_install:
   # protobuf version.
   - sudo pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
   - sudo pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
-  - curl https://glide.sh/get | bash
-  - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
-  - go get -u github.com/alecthomas/gometalinter
-  - gometalinter --install
   - |
     function timeout() { perl -e 'alarm shift; exec @ARGV' "$@"; }
 script:
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5739c2a26039426ab544f762e9401445f01e7de7..4b564b48265897d8b412603baf181030e2b00f82 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -27,7 +27,7 @@ if(NOT CMAKE_CROSSCOMPILING)
 endif(NOT CMAKE_CROSSCOMPILING)
 find_package(Git REQUIRED)
 find_package(Threads REQUIRED)
-if(NOT ANDROID)
+if(NOT ANDROID AND NOT IOS)
     find_package(Boost QUIET)
 endif()
 
@@ -64,27 +64,29 @@ if(NOT CMAKE_BUILD_TYPE)
       FORCE)
 endif()
 
-if(ANDROID)
-    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
-        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
-    elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
-        # TODO: support glog for Android api 16 ~ 19 in the future
-        message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
+if(ANDROID OR IOS)
+    if(ANDROID)
+        if(AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
+            message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
+        elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
+            # TODO: support glog for Android api 16 ~ 19 in the future
+            message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
+        endif()
     endif()
 
     set(WITH_GPU OFF CACHE STRING
-        "Disable GPU when cross-compiling for Android" FORCE)
+        "Disable GPU when cross-compiling for Android and iOS" FORCE)
     set(WITH_AVX OFF CACHE STRING
-        "Disable AVX when cross-compiling for Android" FORCE)
+        "Disable AVX when cross-compiling for Android and iOS" FORCE)
     set(WITH_PYTHON OFF CACHE STRING
-        "Disable PYTHON when cross-compiling for Android" FORCE)
+        "Disable PYTHON when cross-compiling for Android and iOS" FORCE)
     set(WITH_RDMA OFF CACHE STRING
-        "Disable RDMA when cross-compiling for Android" FORCE)
+        "Disable RDMA when cross-compiling for Android and iOS" FORCE)
     set(WITH_MKLDNN OFF CACHE STRING
-        "Disable MKLDNN when cross-compiling for Android" FORCE)
+        "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
     set(WITH_MKLML OFF CACHE STRING
-        "Disable MKLML package when cross-compiling for Android" FORCE)
-endif(ANDROID)
+        "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+endif()
 
 set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
   "A path setting third party libraries download & build directories.")
diff --git a/cmake/cblas.cmake b/cmake/cblas.cmake
index 854066fd1d205c337fbdbe08997d88251095c799..8fdc382f0c1c453a01dba884a3dad216e1c3092c 100644
--- a/cmake/cblas.cmake
+++ b/cmake/cblas.cmake
@@ -171,3 +171,10 @@ if (REFERENCE_CBLAS_INCLUDE_DIR AND REFERENCE_CBLAS_LIBRARY)
   add_definitions(-DPADDLE_USE_REFERENCE_CBLAS)
   message(STATUS "Found reference-cblas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
 endif()
+
+if(IOS_USE_VECLIB_FOR_BLAS AND VECLIB_FOUND)
+  set(CBLAS_FOUND ON)
+  set(CBLAS_PROVIDER vecLib)
+  set(CBLAS_INC_DIR ${VECLIB_INC_DIR})
+  add_definitions(-DPADDLE_USE_VECLIB)
+endif()
diff --git a/cmake/cross_compiling/ios.cmake b/cmake/cross_compiling/ios.cmake
new file mode 100644
index 0000000000000000000000000000000000000000..0b38943952f7fb9052368fe95eb31dd7592d8a47
--- /dev/null
+++ b/cmake/cross_compiling/ios.cmake
@@ -0,0 +1,350 @@
+# 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.
+
+# This is a toolchain file for cross-compiling for iOS, and the
+# configuration largely refers to public toolchain file:
+#    https://raw.githubusercontent.com/leetal/ios-cmake/master/ios.toolchain.cmake
+# and
+#    https://github.com/cristeab/ios-cmake
+#
+# Supports options:
+# IOS_PLATFORM = OS (default) or SIMULATOR
+#   This decides if SDKS will be selected from the iPhoneOS.platform or iPhoneSimulator.platform folders
+#   OS - the default, used to build for iPhone and iPad physical devices, which have an arm arch.
+#   SIMULATOR - used to build for the Simulator platforms, which have an x86 arch.
+# IOS_ARCH
+#   The archectures wanted to support, such "arm64", "armv7;arm64"
+# IOS_DEPLOYMENT_TARGET
+#   The minimum iOS deployment version, such as "7.0"
+# IOS_ENABLE_BITCODE = ON (default) or OFF
+# IOS_USE_VECLIB_FOR_BLAS = OFF (default) or ON
+# IOS_DEVELOPER_ROOT = automatic(default) or /path/to/platform/Developer folder
+#   By default this location is automatcially chosen based on the IOS_PLATFORM value above.
+#   If set manually, it will override the default location and force the user of a particular Developer Platform
+# IOS_SDK_ROOT = automatic(default) or /path/to/platform/Developer/SDKs/SDK folder
+#   By default this location is automatcially chosen based on the IOS_DEVELOPER_ROOT value.
+#   In this case it will always be the most up-to-date SDK found in the IOS_DEVELOPER_ROOT path.
+#   If set manually, this will force the use of a specific SDK version
+
+# Macros:
+# set_xcode_property (TARGET XCODE_PROPERTY XCODE_VALUE)
+#  A convenience macro for setting xcode specific properties on targets
+#  example: set_xcode_property (myioslib IPHONEOS_DEPLOYMENT_TARGET "3.1")
+# find_host_package (PROGRAM ARGS)
+#  A macro used to find executable programs on the host system, not within the iOS environment.
+#  Thanks to the android-cmake project for providing the command
+
+if(NOT IOS)
+  return()
+endif()
+
+set(CMAKE_SYSTEM_NAME Darwin)
+
+# Get the Xcode version being used.
+execute_process(COMMAND xcodebuild -version
+                OUTPUT_VARIABLE XCODE_VERSION
+                RESULT_VARIABLE XCODE_VERSION_RESULT
+                ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+if(NOT ${XCODE_VERSION_RESULT})
+  string(REGEX MATCH "Xcode [0-9\\.]+" XCODE_VERSION "${XCODE_VERSION}")
+  string(REGEX REPLACE "Xcode ([0-9\\.]+)" "\\1" XCODE_VERSION "${XCODE_VERSION}")
+  message(STATUS "Building with Xcode version: ${XCODE_VERSION}")
+else()
+  message(FATAL_ERROR "Cannot execute xcodebuild, please check whether xcode is installed.")
+endif()
+
+# Required as of cmake 2.8.10
+set(CMAKE_OSX_DEPLOYMENT_TARGET "" CACHE STRING "Force unset of the deployment target for iOS" FORCE)
+
+# Setup iOS platform unless specified manually with IOS_PLATFORM
+if(NOT DEFINED IOS_PLATFORM)
+  set(IOS_PLATFORM "OS")
+endif()
+set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING "Type of iOS Platform")
+
+# Set the architecture for iOS
+if(NOT DEFINED IOS_ARCH)
+  if(IOS_PLATFORM STREQUAL "OS")
+    # FIXME(liuyiqun): support "armv7;armv7s;arm64" future
+    set(IOS_ARCH "arm64")
+  elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
+    set(IOS_ARCH "i386;x86_64")
+  elseif(IOS_PLATFORM STREQUAL "WATCHOS")
+    set(IOS_ARCH armv7k)
+  endif()
+endif()
+set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string  "Build architecture for iOS")
+
+# Specify minimum iOS deployment version
+if(NOT DEFINED IOS_DEPLOYMENT_TARGET)
+  set(IOS_DEPLOYMENT_TARGET "7.0")
+endif()
+set(IOS_DEPLOYMENT_TARGET ${IOS_DEPLOYMENT_TARGET} CACHE STRING "Minimum iOS version")
+
+# Whether to enable bitcode
+if(NOT DEFINED IOS_ENABLE_BITCODE)
+  set(IOS_ENABLE_BITCODE ON)
+endif()
+set(IOS_ENABLE_BITCODE ${IOS_ENABLE_BITCODE} CACHE BOOL "Whether to enable bitcode")
+
+if(NOT DEFINED IOS_USE_VECLIB_FOR_BLAS)
+  set(IOS_USE_VECLIB_FOR_BLAS OFF)
+endif()
+set(IOS_USE_VECLIB_FOR_BLAS ${IOS_UES_VECLIB_FOR_BLAS} CACHE BOOL "Whether to use veclib")
+
+# Check the platform selection and setup for developer root
+if(${IOS_PLATFORM} STREQUAL "OS")
+  set(IOS_PLATFORM_LOCATION "iPhoneOS.platform")
+  set(XCODE_IOS_PLATFORM iphoneos)
+
+  # This causes the installers to properly locate the output libraries
+  set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-iphoneos")
+elseif(${IOS_PLATFORM} STREQUAL "SIMULATOR")
+  set(IOS_PLATFORM_LOCATION "iPhoneSimulator.platform")
+  set(XCODE_IOS_PLATFORM iphonesimulator)
+
+  # This causes the installers to properly locate the output libraries
+  set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-iphonesimulator")
+elseif(${IOS_PLATFORM} STREQUAL "WATCHOS")
+  set(IOS_PLATFORM_LOCATION "WatchOS.platform")
+  set(XCODE_IOS_PLATFORM watchos)
+
+  # This causes the installers to properly locate the output libraries
+  set(CMAKE_XCODE_EFFECTIVE_PLATFORMS "-watchos")
+else(${IOS_PLATFORM} STREQUAL "OS")
+  message(FATAL_ERROR "Unsupported IOS_PLATFORM value selected. Please set to\n"
+          "\t OS, SIMULATOR, or WATCHOS.")
+endif()
+
+# Check iOS developer toolchain
+if(NOT DEFINED IOS_DEVELOPER_ROOT)
+  # Setup iOS developer location
+  execute_process(COMMAND xcode-select -print-path
+                  OUTPUT_VARIABLE XCODE_DEVELOPER_DIR
+                  RESULT_VARIABLE XCODE_DEVELOPER_DIR_RESULT
+                  ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+  # Xcode 4.3 changed the installation location, choose the most recent one available
+  if(${XCODE_VERSION} VERSION_LESS "4.3.0")
+    set(IOS_DEVELOPER_ROOT "/Developer/Platforms/${IOS_PLATFORM_LOCATION}/Developer")
+  else()
+    set(IOS_DEVELOPER_ROOT "${XCODE_DEVELOPER_DIR}/Platforms/${IOS_PLATFORM_LOCATION}/Developer")
+  endif()
+endif()
+if(EXISTS ${IOS_DEVELOPER_ROOT})
+  set(IOS_DEVELOPER_ROOT ${IOS_DEVELOPER_ROOT} CACHE PATH "Location of iOS Platform")
+else()
+  message(FATAL_ERROR "Invalid IOS_DEVELOPER_ROOT: ${IOS_DEVELOPER_ROOT} does not exist.")
+endif()
+
+# Check iOS SDK
+if(NOT DEFINED IOS_SDK_ROOT)
+  # Find and use the most recent iOS sdk
+  file(GLOB IOS_SDK_LISTS "${IOS_DEVELOPER_ROOT}/SDKs/*")
+  if(IOS_SDK_LISTS)
+    list(SORT IOS_SDK_LISTS)
+    list(REVERSE IOS_SDK_LISTS)
+    list(GET IOS_SDK_LISTS 0 IOS_SDK_ROOT)
+  else(IOS_SDK_LISTS)
+    message(FATAL_ERROR "No iOS SDK's found in default search path ${IOS_DEVELOPER_ROOT}."
+            " Please manually set IOS_SDK_ROOT or install the iOS SDK.")
+  endif(IOS_SDK_LISTS)
+endif()
+if(EXISTS ${IOS_SDK_ROOT})
+  set(IOS_SDK_ROOT ${IOS_SDK_ROOT} CACHE PATH "Location of the selected iOS SDK")
+  message(STATUS "iOS toolchain: ${IOS_SDK_ROOT}")
+else()
+  message(FATAL_ERROR "Invalid IOS_SDK_ROOT: ${IOS_SDK_ROOT} does not exist.")
+endif()
+
+# Set the sysroot default to the most recent SDK
+set(CMAKE_OSX_SYSROOT ${IOS_SDK_ROOT} CACHE PATH "Sysroot used for iOS support")
+
+# Get version of iOS SDK
+execute_process(COMMAND xcodebuild -sdk ${CMAKE_OSX_SYSROOT} -version SDKVersion
+                OUTPUT_VARIABLE IOS_SDK_VERSION
+                RESULT_VARIABLE IOS_SDK_VERSION_RESULT
+                ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+if(${IOS_SDK_VERSION_RESULT})
+  string(REGEX MATCH "(([0-9]+)\\.)+([0-9]+)" IOS_SDK_VERSION "${IOS_SDK_ROOT}")
+endif()
+if(NOT IOS_SDK_VERSION)
+  message(WARNING "Cannot get SDK's version.")
+  set(IOS_SDK_VERSION 1)
+endif()
+set(CMAKE_SYSTEM_VERSION ${IOS_SDK_VERSION})
+
+# Find the C & C++ compilers for the specified SDK.
+if(NOT CMAKE_C_COMPILER)
+  # Default to use clang
+  execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find clang
+                  OUTPUT_VARIABLE IOS_C_COMPILER
+                  RESULT_VARIABLE IOS_C_COMPILER_RESULT
+                  ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+  if(${IOS_C_COMPILER_RESULT})
+    get_filename_component(IOS_C_COMPILER clang PROGRAM)
+  endif()
+else(NOT CMAKE_C_COMPILER)
+  # User can set it in cmake command
+  get_filename_component(IOS_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
+endif(NOT CMAKE_C_COMPILER)
+if(NOT EXISTS ${IOS_C_COMPILER})
+  message(FATAL_ERROR "Cannot find C compiler: ${IOS_C_COMPILER}")
+endif()
+
+if(NOT CMAKE_CXX_COMPILER)
+  # Default to use clang++
+  execute_process(COMMAND xcrun -sdk ${CMAKE_OSX_SYSROOT} -find clang++
+                  OUTPUT_VARIABLE IOS_CXX_COMPILER
+                  RESULT_VARIABLE IOS_CXX_COMPILER_RESULT
+                  ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+  if(${IOS_CXX_COMPILER_RESULT})
+    get_filename_component(IOS_CXX_COMPILER clang++ PROGRAM)
+  endif()
+else(NOT CMAKE_CXX_COMPILER)
+  # User can set it in cmake command
+  get_filename_component(IOS_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
+endif(NOT CMAKE_CXX_COMPILER)
+if(NOT EXISTS ${IOS_CXX_COMPILER})
+  message(FATAL_ERROR "Cannot find CXX compiler: ${IOS_CXX_COMPILER}")
+endif()
+
+set(CMAKE_C_COMPILER ${IOS_C_COMPILER} CACHE PATH "C compiler" FORCE)
+set(CMAKE_CXX_COMPILER ${IOS_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
+
+set(CMAKE_C_OSX_COMPATIBILITY_VERSION_FLAG "-compatibility_version ")
+set(CMAKE_C_OSX_CURRENT_VERSION_FLAG "-current_version ")
+set(CMAKE_CXX_OSX_COMPATIBILITY_VERSION_FLAG "${CMAKE_C_OSX_COMPATIBILITY_VERSION_FLAG}")
+set(CMAKE_CXX_OSX_CURRENT_VERSION_FLAG "${CMAKE_C_OSX_CURRENT_VERSION_FLAG}")
+
+# Set iOS specific C/C++ flags
+if(IOS_PLATFORM STREQUAL "OS")
+  if(XCODE_VERSION VERSION_LESS "7.0")
+    set(XCODE_IOS_PLATFORM_VERSION_FLAGS "-mios-version-min=${IOS_DEPLOYMENT_TARGET}")
+  else()
+    # Xcode 7.0+ uses flags we can build directly from XCODE_IOS_PLATFORM.
+    set(XCODE_IOS_PLATFORM_VERSION_FLAGS "-m${XCODE_IOS_PLATFORM}-version-min=${IOS_DEPLOYMENT_TARGET}")
+  endif()
+else()
+  set(XCODE_IOS_FLATFORM_VERSION_FLAGS "-mios-simulator-version-min=${IOS_DEPLOYMENT_TARGET}")
+endif()
+
+if(IOS_ENABLE_BITCODE)
+  set(XCODE_IOS_BITCODE_FLAGS "${IOS_COMPILER_FLAGS} -fembed-bitcode")
+else()
+  set(XCODE_IOS_BITCODE_FLAGS "")
+endif()
+
+set(IOS_COMPILER_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${XCODE_IOS_BITCODE_FLAGS}")
+
+# Hidden visibilty is required for cxx on iOS 
+set(CMAKE_C_FLAGS "${IOS_COMPILER_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags")
+set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
+
+set(IOS_LINK_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} -Wl,-search_paths_first")
+
+if(IOS_USE_VECLIB_FOR_BLAS)
+  # Find vecLib for iOS
+  set(VECLIB_SEARCH_DIRS
+      ${IOS_SDK_ROOT}/System/Library/Frameworks/Accelerate.framework/Versions/Current/Frameworks
+      ${IOS_SDK_ROOT}/System/Library/Frameworks/Accelerate.framework/Frameworks
+      )
+  find_path(VECLIB_INC_DIR vecLib.h PATHS ${VECLIB_SEARCH_DIRS}/vecLib.framework/Headers)
+
+  include(FindPackageHandleStandardArgs)
+  find_package_handle_standard_args(vecLib DEFAULT_MSG VECLIB_INC_DIR)
+
+  if(VECLIB_FOUND)
+    if(VECLIB_INC_DIR MATCHES "^/System/Library/Frameworks/vecLib.framework.*")
+      set(IOS_LINK_FLAGS ${IOS_LINK_FLAGS} -lcblas "-framework vecLib")
+      message(STATUS "Found standalone vecLib.framework")
+    else()
+      set(IOS_LINK_FLAGS ${IOS_LINK_FLAGS} -lcblas "-framework Accelerate")
+      message(STATUS "Found vecLib as part of Accelerate.framework")
+    endif()
+
+  endif()
+endif()
+
+set(CMAKE_C_LINK_FLAGS "${IOS_LINK_FLAGS} ${CMAKE_C_LINK_FLAGS}")
+set(CMAKE_CXX_LINK_FLAGS "${IOS_LINK_FLAGS} ${CMAKE_CXX_LINK_FLAGS}")
+
+set(CMAKE_PLATFORM_HAS_INSTALLNAME 1)
+if(NOT IOS_ENABLE_BITCODE)
+  set(CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS "-dynamiclib -headerpad_max_install_names")
+  set(CMAKE_SHARED_MODULE_CREATE_C_FLAGS "-bundle -headerpad_max_install_names")
+else()
+  set(CMAKE_SHARED_LIBRARY_CREATE_C_FLAGS "-dynamiclib")
+  set(CMAKE_SHARED_MODULE_CREATE_C_FLAGS "-bundle")
+endif()
+set(CMAKE_SHARED_MODULE_LOADER_C_FLAG "-Wl,-bundle_loader,")
+set(CMAKE_SHARED_MODULE_LOADER_CXX_FLAG "-Wl,-bundle_loader,")
+set(CMAKE_FIND_LIBRARY_SUFFIXES ".dylib" ".so" ".a")
+
+# hack: if a new cmake (which uses CMAKE_INSTALL_NAME_TOOL) runs on an old build tree
+# (where install_name_tool was hardcoded) and where CMAKE_INSTALL_NAME_TOOL isn't in the cache
+# and still cmake didn't fail in CMakeFindBinUtils.cmake (because it isn't rerun)
+# hardcode CMAKE_INSTALL_NAME_TOOL here to install_name_tool, so it behaves as it did before, Alex
+if(NOT DEFINED CMAKE_INSTALL_NAME_TOOL)
+  find_program(CMAKE_INSTALL_NAME_TOOL install_name_tool)
+endif()
+
+# Set the find root to the iOS developer roots and to user defined paths
+set(CMAKE_FIND_ROOT_PATH ${IOS_DEVELOPER_ROOT} ${IOS_SDK_ROOT} ${CMAKE_PREFIX_PATH}
+    CACHE string  "iOS find search path root")
+
+# default to searching for frameworks first
+set(CMAKE_FIND_FRAMEWORK FIRST)
+
+# set up the default search directories for frameworks
+set(CMAKE_SYSTEM_FRAMEWORK_PATH
+    ${IOS_SDK_ROOT}/System/Library/Frameworks
+    ${IOS_SDK_ROOT}/System/Library/PrivateFrameworks
+    ${IOS_SDK_ROOT}/Developer/Library/Frameworks
+    )
+
+# only search the iOS sdks, not the remainder of the host filesystem
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+
+message(STATUS "iOS: Targeting iOS '${CMAKE_SYSTEM_VERSION}', "
+        "building for '${IOS_PLATFORM}' platform, with architecture '${CMAKE_OSX_ARCHITECTURES}'")
+message(STATUS "System CMAKE_C_FLAGS: ${CMAKE_C_FLAGS}")
+message(STATUS "System CMAKE_CXX_FLAGS: ${CMAKE_CXX_FLAGS}")
+
+# Used in ExternalProject command
+string(REPLACE ";" "\\$<SEMICOLON>" EXTERNAL_IOS_ARCHITECTURES "${CMAKE_OSX_ARCHITECTURES}")
+set(EXTERNAL_OPTIONAL_ARGS
+    -DCMAKE_OSX_SYSROOT=${CMAKE_OSX_SYSROOT}
+    -DCMAKE_OSX_ARCHITECTURES=${EXTERNAL_IOS_ARCHITECTURES})
+
+# This little macro lets you set any XCode specific property
+macro(set_xcode_property TARGET XCODE_PROPERTY XCODE_VALUE)
+  set_property (TARGET ${TARGET} PROPERTY XCODE_ATTRIBUTE_${XCODE_PROPERTY} ${XCODE_VALUE})
+endmacro(set_xcode_property)
+
+# This macro lets you find executable programs on the host system
+macro(find_host_package)
+  set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+  set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY NEVER)
+  set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE NEVER)
+  set(IOS FALSE)
+
+  find_package(${ARGN})
+
+  set(IOS TRUE)
+  set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM ONLY)
+  set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+  set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+endmacro(find_host_package)
diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake
index 01a2f4d5fa357ca882162247cc52299a3d1d3030..957f8271e4841836956b0c3f2cf3d8c88a31192a 100644
--- a/cmake/external/gflags.cmake
+++ b/cmake/external/gflags.cmake
@@ -39,13 +39,14 @@ ExternalProject_Add(
     PREFIX          ${GFLAGS_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-    CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-    CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
-    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-    CMAKE_ARGS      -DBUILD_TESTING=OFF
-    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                    -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DBUILD_TESTING=OFF
+                    -DCMAKE_BUILD_TYPE=Release
+                    ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GFLAGS_INSTALL_DIR}
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=Release
diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake
index b450a3016667dcb4ab229fe7ec8aaae8609d8171..b3fef738ccc0b5886bb0a32501bb7b7adade0ff1 100644
--- a/cmake/external/glog.cmake
+++ b/cmake/external/glog.cmake
@@ -34,16 +34,17 @@ ExternalProject_Add(
     PREFIX          ${GLOG_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-    CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-    CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
-    CMAKE_ARGS      -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib
-    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-    CMAKE_ARGS      -DWITH_GFLAGS=ON
-    CMAKE_ARGS      -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags
-    CMAKE_ARGS      -DBUILD_TESTING=OFF
-    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                    -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
+                    -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DWITH_GFLAGS=ON
+                    -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags
+                    -DBUILD_TESTING=OFF
+                    -DCMAKE_BUILD_TYPE=Release
+                    ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GLOG_INSTALL_DIR}
                      -DCMAKE_INSTALL_LIBDIR:PATH=${GLOG_INSTALL_DIR}/lib
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
diff --git a/cmake/external/gtest.cmake b/cmake/external/gtest.cmake
index e3970073a1a0b946fa1db6642799719d7a9fcf4f..6a2a79b7631b32e8a099797de509af64533bbb95 100644
--- a/cmake/external/gtest.cmake
+++ b/cmake/external/gtest.cmake
@@ -48,15 +48,16 @@ IF(WITH_TESTING)
         PREFIX          ${GTEST_SOURCES_DIR}
         UPDATE_COMMAND  ""
         CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-        CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-        CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-        CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-        CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GTEST_INSTALL_DIR}
-        CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-        CMAKE_ARGS      -DBUILD_GMOCK=ON
-        CMAKE_ARGS      -Dgtest_disable_pthreads=ON
-        CMAKE_ARGS      -Dgtest_force_shared_crt=ON
-        CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                        -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+                        -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                        -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                        -DCMAKE_INSTALL_PREFIX=${GTEST_INSTALL_DIR}
+                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                        -DBUILD_GMOCK=ON
+                        -Dgtest_disable_pthreads=ON
+                        -Dgtest_force_shared_crt=ON
+                        -DCMAKE_BUILD_TYPE=Release
+                        ${EXTERNAL_OPTIONAL_ARGS}
         CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${GTEST_INSTALL_DIR}
                          -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                          -DCMAKE_BUILD_TYPE:STRING=Release
diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake
index 4fc8d43fc10891603b79c01a1c769cae21c52655..143b57a954e4e6b2bf273535ebdf0fa8e3dab768 100644
--- a/cmake/external/openblas.cmake
+++ b/cmake/external/openblas.cmake
@@ -29,30 +29,41 @@ IF(NOT ${CBLAS_FOUND})
         "${CBLAS_INSTALL_DIR}/lib/${CMAKE_STATIC_LIBRARY_PREFIX}openblas${CMAKE_STATIC_LIBRARY_SUFFIX}"
         CACHE FILEPATH "openblas library." FORCE)
 
-    IF(APPLE)
-        SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
-        SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
-    ELSE()
-        SET(COMMON_ARGS CC=${CMAKE_C_COMPILER} NO_SHARED=1 NO_LAPACK=1 libs)
-    ENDIF()
+    SET(OPENBLAS_CC "${CMAKE_C_COMPILER}")
 
     IF(CMAKE_CROSSCOMPILING)
+        SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER})
+        GET_FILENAME_COMPONENT(CROSS_SUFFIX ${CMAKE_C_COMPILER} DIRECTORY)
+        SET(CROSS_SUFFIX ${CROSS_SUFFIX}/)
         IF(ANDROID)
             # arm_soft_fp_abi branch of OpenBLAS to support softfp
             #   https://github.com/xianyi/OpenBLAS/tree/arm_soft_fp_abi
             SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
             IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
-                SET(TARGET "ARMV7")
+                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0)
             ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
-                SET(TARGET "ARMV8")
+                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0)
+            ENDIF()
+        ELSEIF(IOS)
+            # FIXME(liuyiqun): support multiple architectures
+            SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
+            SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}")
+            IF(CMAKE_OSX_ARCHITECTURES MATCHES "armv7")
+                SET(OPENBLAS_CC "${OPENBLAS_CC} -arch armv7")
+                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0)
+            ELSEIF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
+                SET(OPENBLAS_CC "${OPENBLAS_CC} -arch arm64")
+                SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0 CROSS_SUFFIX=${CROSS_SUFFIX})
             ENDIF()
-            SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=${TARGET} ARM_SOFTFP_ABI=1 USE_THREAD=0)
         ELSEIF(RPI)
             # use hardfp
             SET(OPENBLAS_COMMIT "v0.2.20")
-            SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 USE_THREAD=0)
+            SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 USE_THREAD=0)
         ENDIF()
     ELSE()
+        IF(APPLE)
+            SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
+        ENDIF()
         SET(OPENBLAS_COMMIT "v0.2.20")
         SET(OPTIONAL_ARGS "")
         IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
@@ -60,6 +71,8 @@ IF(NOT ${CBLAS_FOUND})
         ENDIF()
     ENDIF()
 
+    SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
+
     ExternalProject_Add(
         extern_openblas
         ${EXTERNAL_PROJECT_LOG_ARGS}
diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake
index a887be2e2ae5e21562fc15c775bb24cc1553480e..7cf7ba85cca4c248dcc74e078124c0b3815ee380 100644
--- a/cmake/external/protobuf.cmake
+++ b/cmake/external/protobuf.cmake
@@ -173,7 +173,8 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
             "-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}"
             "-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}"
             "-Dprotobuf_WITH_ZLIB=ON"
-            "-DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}")
+            "-DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}"
+            ${EXTERNAL_OPTIONAL_ARGS})
         SET(OPTIONAL_CACHE_ARGS "-DZLIB_ROOT:STRING=${ZLIB_ROOT}")
     ENDIF()
 
diff --git a/cmake/external/python.cmake b/cmake/external/python.cmake
index 490c87d67ed79a238dd506127cd4d9855fab6626..46c68cce324f565ec9985ef1a280d6d933f88f1f 100644
--- a/cmake/external/python.cmake
+++ b/cmake/external/python.cmake
@@ -12,16 +12,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-INCLUDE(ExternalProject)
+IF(NOT WITH_PYTHON)
+    return()
+ENDIF()
+
 INCLUDE(python_module)
 
 FIND_PACKAGE(PythonInterp 2.7)
-IF(WITH_PYTHON)
-    FIND_PACKAGE(PythonLibs 2.7)
-    # Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE.
-    ADD_LIBRARY(python SHARED IMPORTED GLOBAL)
-    SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES})
-ENDIF(WITH_PYTHON)
+FIND_PACKAGE(PythonLibs 2.7)
+# Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE.
+ADD_LIBRARY(python SHARED IMPORTED GLOBAL)
+SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES})
 
 SET(py_env "")
 IF(PYTHONINTERP_FOUND)
@@ -36,9 +37,5 @@ IF(PYTHONINTERP_FOUND)
     ENDIF()
 ENDIF(PYTHONINTERP_FOUND)
 
-IF(WITH_PYTHON)
-    INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR})
-    INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR})
-ELSE()
-    SET(PYTHON_LIBRARIES "")
-ENDIF()
+INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR})
+INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR})
diff --git a/cmake/external/swig.cmake b/cmake/external/swig.cmake
index 744c766ee7b067058b2cb4aa7f7b761cbb9778d4..ce088ae7eaa3355f2f9761e8c421da0d7ef89fa7 100644
--- a/cmake/external/swig.cmake
+++ b/cmake/external/swig.cmake
@@ -12,6 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+IF(NOT WITH_SWIG_PY)
+    return()
+ENDIF()
+
 FIND_PACKAGE(SWIG)
 
 IF(NOT SWIG_FOUND)
diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake
index 2d7daed9bcd5b8d854ffae6dc1ea191d154c16fe..bb258c7b5581fc22b44f4fe15c119f8081f4767e 100644
--- a/cmake/external/warpctc.cmake
+++ b/cmake/external/warpctc.cmake
@@ -16,25 +16,14 @@ INCLUDE(ExternalProject)
 
 SET(WARPCTC_SOURCES_DIR ${THIRD_PARTY_PATH}/warpctc)
 SET(WARPCTC_INSTALL_DIR ${THIRD_PARTY_PATH}/install/warpctc)
-SET(WARPCTC_INCLUDE_DIR "${WARPCTC_INSTALL_DIR}/include" CACHE PATH "Warp-ctc Directory" FORCE)
 
-INCLUDE_DIRECTORIES(${WARPCTC_INCLUDE_DIR})
-
-SET(WARPCTC_LIB_DIR "${WARPCTC_INSTALL_DIR}/lib" CACHE PATH "Warp-ctc Library Directory" FORCE)
-
-IF(WIN32)
-    SET(WARPCTC_LIBRARIES
-        "${WARPCTC_INSTALL_DIR}/lib/warpctc.dll" CACHE FILEPATH "Warp-ctc Library" FORCE)
-ELSE(WIN32)
-    IF(APPLE)
-        SET(_warpctc_SHARED_SUFFIX dylib)
-    ELSE(APPLE)
-        SET(_warpctc_SHARED_SUFFIX so)
-    ENDIF(APPLE)
-
-    SET(WARPCTC_LIBRARIES
-        "${WARPCTC_INSTALL_DIR}/lib/libwarpctc.${_warpctc_SHARED_SUFFIX}" CACHE FILEPATH "Warp-ctc Library" FORCE)
-ENDIF(WIN32)
+SET(WARPCTC_INCLUDE_DIR "${WARPCTC_INSTALL_DIR}/include"
+    CACHE PATH "Warp-ctc Directory" FORCE)
+# Used in unit test test_WarpCTCLayer
+SET(WARPCTC_LIB_DIR "${WARPCTC_INSTALL_DIR}/lib"
+    CACHE PATH "Warp-ctc Library Directory" FORCE)
+SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/libwarpctc${CMAKE_SHARED_LIBRARY_SUFFIX}"
+    CACHE FILEPATH "Warp-ctc Library" FORCE)
 
 IF(CMAKE_CXX_COMPILER_ID STREQUAL "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" )
     SET(USE_OMP OFF)
@@ -49,22 +38,26 @@ ExternalProject_Add(
     PREFIX          ${WARPCTC_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-    CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-    CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR}
-    CMAKE_ARGS      -DWITH_GPU=${WITH_GPU}
-    CMAKE_ARGS      -DWITH_OMP=${USE_OMP}
-    CMAKE_ARGS      -DWITH_TORCH=OFF
-    CMAKE_ARGS      -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON
-    CMAKE_ARGS      -DBUILD_SHARED=ON
-    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                    -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR}
+                    -DWITH_GPU=${WITH_GPU}
+                    -DWITH_OMP=${USE_OMP}
+                    -DWITH_TORCH=OFF
+                    -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON
+                    -DBUILD_SHARED=ON
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DCMAKE_BUILD_TYPE=Release
+                    ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS -DCMAKE_BUILD_TYPE:STRING=Release
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_INSTALL_PREFIX:PATH=${WARPCTC_INSTALL_DIR}
 )
 
+MESSAGE(STATUS "warp-ctc library: ${WARPCTC_LIBRARIES}")
+INCLUDE_DIRECTORIES(${WARPCTC_INCLUDE_DIR})
+
 ADD_LIBRARY(warpctc STATIC IMPORTED GLOBAL)
 SET_PROPERTY(TARGET warpctc PROPERTY IMPORTED_LOCATION ${WARPCTC_LIBRARIES})
 ADD_DEPENDENCIES(warpctc extern_warpctc)
diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake
index 5aecab90ca3cecdfdba0eac178a6ba07dfcb8745..c496a52b780364f3014f8fa3dfbc944a7aa7430e 100644
--- a/cmake/external/zlib.cmake
+++ b/cmake/external/zlib.cmake
@@ -34,15 +34,16 @@ ExternalProject_Add(
     GIT_TAG         "v1.2.8"
     PREFIX          ${ZLIB_SOURCES_DIR}
     UPDATE_COMMAND  ""
-    CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
     CMAKE_ARGS      -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-    CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR}
-    CMAKE_ARGS      -DBUILD_SHARED_LIBS=OFF
-    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-    CMAKE_ARGS      -DCMAKE_MACOSX_RPATH=ON
-    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                    -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR}
+                    -DBUILD_SHARED_LIBS=OFF
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DCMAKE_MACOSX_RPATH=ON
+                    -DCMAKE_BUILD_TYPE=Release
+                    ${EXTERNAL_OPTIONAL_ARGS}
     CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ZLIB_INSTALL_DIR}
                      -DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
                      -DCMAKE_BUILD_TYPE:STRING=Release
diff --git a/cmake/flags.cmake b/cmake/flags.cmake
index ff246b2eb4ed97dd14d45763569b661cefd203c8..4593ae6180b6d7deb61d897eb634b17ac0bb1683 100644
--- a/cmake/flags.cmake
+++ b/cmake/flags.cmake
@@ -128,8 +128,10 @@ set(GPU_COMMON_FLAGS
 )
 
 if (APPLE)
-    # On Mac OS X build fat binaries with x86_64 architectures by default.
-    set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
+    if(NOT CMAKE_CROSSCOMPILING)
+        # On Mac OS X build fat binaries with x86_64 architectures by default.
+        set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
+    endif()
 else()
     set(GPU_COMMON_FLAGS
         -Wall
diff --git a/cmake/system.cmake b/cmake/system.cmake
index adf5e2c539740076ad1808353522c7467d765e64..396bd1a0797edea0522bb1f02349373563b7726a 100644
--- a/cmake/system.cmake
+++ b/cmake/system.cmake
@@ -24,11 +24,10 @@ IF(WIN32)
     SET(HOST_SYSTEM "win32")
 ELSE(WIN32)
     IF(APPLE)
-        EXEC_PROGRAM (sw_vers ARGS -productVersion OUTPUT_VARIABLE MACOSX_VERSION)
-        STRING(REGEX MATCH "[0-9]+.[0-9]+" VERSION "${MACOSX_VERSION}")
-        SET(MACOS_VERSION ${VERSION})
         SET(HOST_SYSTEM "macosx")
-        IF(NOT DEFINED ENV{MACOSX_DEPLOYMENT_TARGET})
+        EXEC_PROGRAM(sw_vers ARGS -productVersion OUTPUT_VARIABLE HOST_SYSTEM_VERSION)
+        STRING(REGEX MATCH "[0-9]+.[0-9]+" MACOS_VERSION "${HOST_SYSTEM_VERSION}")
+        IF(NOT DEFINED $ENV{MACOSX_DEPLOYMENT_TARGET})
             # Set cache variable - end user may change this during ccmake or cmake-gui configure.
             SET(CMAKE_OSX_DEPLOYMENT_TARGET ${MACOS_VERSION} CACHE STRING
                 "Minimum OS X version to target for deployment (at runtime); newer APIs weak linked. Set to empty string for default value.")
@@ -49,6 +48,8 @@ ELSE(WIN32)
             ELSEIF(LINUX_ISSUE MATCHES "Fedora")
                 SET(HOST_SYSTEM "fedora")
             ENDIF()
+
+            STRING(REGEX MATCH "(([0-9]+)\\.)+([0-9]+)" HOST_SYSTEM_VERSION "${LINUX_ISSUE}")
         ENDIF(EXISTS "/etc/issue")
 
         IF(EXISTS "/etc/redhat-release")
@@ -70,7 +71,7 @@ CMAKE_HOST_SYSTEM_INFORMATION(RESULT CPU_CORES QUERY NUMBER_OF_LOGICAL_CORES)
 
 MARK_AS_ADVANCED(HOST_SYSTEM CPU_CORES)
 
-MESSAGE(STATUS "Found Paddle host system: ${HOST_SYSTEM}")
+MESSAGE(STATUS "Found Paddle host system: ${HOST_SYSTEM}, version: ${HOST_SYSTEM_VERSION}")
 MESSAGE(STATUS "Found Paddle host system's CPU: ${CPU_CORES} cores")
 
 # configuration for cross-compiling
@@ -82,6 +83,9 @@ IF(DEFINED CMAKE_SYSTEM_NAME)
     ELSEIF(${CMAKE_SYSTEM_NAME} STREQUAL "RPi")
         SET(RPI TRUE)
         INCLUDE(cross_compiling/raspberry_pi)
+    ELSEIF(${CMAKE_SYSTEM_NAME} STREQUAL "iOS")
+        SET(IOS TRUE)
+        INCLUDE(cross_compiling/ios)
     ENDIF()
 ENDIF()
 
diff --git a/cmake/util.cmake b/cmake/util.cmake
index 0da4969d310368ab27b0ed65237813c07d6e59f0..e814cad36f2a8ce95a2dc9fabc35cb39506d4cd7 100644
--- a/cmake/util.cmake
+++ b/cmake/util.cmake
@@ -25,7 +25,9 @@ function(target_circle_link_libraries TARGET_NAME)
             endif()
         endforeach()
         if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang" OR "${CMAKE_CXX_COMPILER_ID}" STREQUAL "AppleClang")
-            list(APPEND LIBS "-undefined dynamic_lookup")
+            if(IOS AND NOT IOS_ENABLE_BITCODE)
+                list(APPEND LIBS "-undefined dynamic_lookup")
+            endif()
         endif()
         list(REVERSE libsInArgn)
         target_link_libraries(${TARGET_NAME}
diff --git a/doc/design/block.md b/doc/design/block.md
new file mode 100644
index 0000000000000000000000000000000000000000..be8800122035984df281692fc40009c397565046
--- /dev/null
+++ b/doc/design/block.md
@@ -0,0 +1,338 @@
+# Design Doc: Block and Scope
+
+## The Representation of Computation
+
+Both deep learning systems and programming languages help users describe computation procedures.  These systems use various representations of computation:
+
+- Caffe, Torch, and Paddle: sequences of layers.
+- TensorFlow, Caffe2, Mxnet: graphs of operators.
+- PaddlePaddle: nested blocks, like C++ and Java programs.
+
+## Block in Programming Languages and Deep Learning
+
+In programming languages, a block is a pair of curly braces that includes local variables definitions and a sequence of instructions, or operators.
+
+Blocks work with control flow structures like `if`, `else`, and `for`, which have equivalents in deep learning:
+
+| programming languages | PaddlePaddle          |
+|-----------------------|-----------------------|
+| for, while loop       | RNN, WhileOp          |
+| if, if-else, switch   | IfElseOp, SwitchOp    |
+| sequential execution  | a sequence of layers  |
+
+A key difference is that a C++ program describes a one pass computation, whereas a deep learning program describes both the forward and backward passes.
+
+## Stack Frames and the Scope Hierarchy
+
+The existence of the backward makes the execution of a block of traditional programs and PaddlePaddle different to each other:
+
+| programming languages | PaddlePaddle                  |
+|-----------------------|-------------------------------|
+| stack                 | scope hierarchy               |
+| stack frame           | scope                         |
+| push at entering block| push at entering block        |
+| pop at leaving block  | destroy at minibatch completes|
+
+1. In traditional programs:
+
+   - When the execution enters the left curly brace of a block, the runtime pushes a frame into the stack, where it realizes local variables.
+   - After the execution leaves the right curly brace, the runtime pops the frame.
+   - The maximum number of frames in the stack is the maximum depth of nested blocks.
+
+1. In PaddlePaddle
+
+   - When the execution enters a block, PaddlePaddle adds a new scope, where it realizes variables.
+   - PaddlePaddle doesn't pop a scope after the execution of the block because variables therein are to be used by the backward pass.  So it has a stack forest known as a *scope hierarchy*.
+   - The height of the highest tree is the maximum depth of nested blocks.
+   - After the process of a minibatch, PaddlePaddle destroys the scope hierarchy.
+
+## Use Blocks in C++ and PaddlePaddle Programs
+
+Let us consolidate the discussion by presenting some examples.
+
+### Blocks with `if-else` and `IfElseOp`
+
+The following C++ programs shows how blocks are used with the `if-else` structure:
+
+```c++
+int x = 10;
+int y = 20;
+int out;
+bool cond = false;
+if (cond) {
+  int z = x + y;
+  out = softmax(z);
+} else {
+  int z = fc(x);
+  out = z;
+}
+```
+
+An equivalent PaddlePaddle program from the design doc of the [IfElseOp operator](./if_else_op.md) is as follows:
+
+```python
+import paddle as pd
+
+x = var(10)
+y = var(20)
+cond = var(false)
+ie = pd.create_ifelseop(inputs=[x], output_num=1)
+with ie.true_block():
+    x = ie.inputs(true, 0)
+    z = operator.add(x, y)
+    ie.set_output(true, 0, operator.softmax(z))
+with ie.false_block():
+    x = ie.inputs(false, 0)
+    z = layer.fc(x)
+    ie.set_output(true, 0, operator.softmax(z))
+out = b(cond)
+```
+
+In both examples, the left branch computes `softmax(x+y)` and the right branch computes `fc(x)`.
+
+A difference is that variables in the C++ program contain scalar values, whereas those in the PaddlePaddle programs are mini-batches of instances.  The `ie.input(true, 0)` invocation returns instances in the 0-th input, `x`, that corresponds to true values in `cond` as the local variable `x`, where `ie.input(false, 0)` returns instances corresponding to false values.
+
+### Blocks with `for` and `RNNOp`
+
+The following RNN model from the [RNN design doc](./rnn.md)
+
+```python
+x = sequence([10, 20, 30])
+m = var(0)
+W = tensor()
+U = tensor()
+
+rnn = create_rnn(inputs=[input])
+with rnn.stepnet() as net:
+  x = net.set_inputs(0)
+  h = net.add_memory(init=m)
+  fc_out = pd.matmul(W, x)
+  hidden_out = pd.matmul(U, h.pre(n=1))
+  sum = pd.add_two(fc_out, hidden_out)
+  act = pd.sigmoid(sum)
+  h.update(act)                       # update memory with act
+  net.set_outputs(0, act, hidden_out) # two outputs
+
+o1, o2 = rnn()
+print o1, o2
+```
+
+has its equivalent C++ program as follows
+
+```c++
+int* x = {10, 20, 30};
+int m = 0;
+int W = some_value();
+int U = some_other_value();
+
+int mem[sizeof(x) / sizeof(x[0]) + 1];
+int o1[sizeof(x) / sizeof(x[0]) + 1];
+int o2[sizeof(x) / sizeof(x[0]) + 1];
+for (int i = 1; i <= sizeof(x)/sizeof(x[0]); ++i) {
+  int x = x[i-1];
+  if (i == 1) mem[0] = m;
+  int fc_out = W * x;
+  int hidden_out = Y * mem[i-1];
+  int sum = fc_out + hidden_out;
+  int act = sigmoid(sum);
+  mem[i] = act;
+  o1[i] = act;
+  o2[i] = hidden_out;
+}
+
+print_array(o1);
+print_array(o2);
+```
+
+
+## Compilation and Execution
+
+Like TensorFlow programs, a PaddlePaddle program is written in Python.  The first part describes a neural network as a protobuf message, and the rest part executes the message for training or inference.
+
+The generation of this protobuf message is like what a compiler generates a binary executable file.  The execution of the message that the OS executes the binary file.
+
+## The "Binary Executable File Format"
+
+The definition of the protobuf message is as follows:
+
+```protobuf
+message BlockDesc {
+  repeated VarDesc vars = 1;
+  repeated OpDesc ops = 2;
+}
+```
+
+The step net in above RNN example would look like
+
+```
+BlockDesc {
+  vars = {
+    VarDesc {...} // x
+    VarDesc {...} // h
+    VarDesc {...} // fc_out
+    VarDesc {...} // hidden_out
+    VarDesc {...} // sum
+    VarDesc {...} // act
+  }
+  ops = {
+    OpDesc {...} // matmul
+    OpDesc {...} // add_two
+    OpDesc {...} // sigmoid
+  }
+};
+```
+
+Also, the RNN operator in above example is serialized into a protobuf message of type `OpDesc` and would look like:
+
+```
+OpDesc {
+  inputs = {0} // the index of x
+  outputs = {5, 3} // indices of act and hidden_out
+  attrs {
+    "memories" : {1} // the index of h
+    "step_net" : <above step net>
+  }
+};
+```
+
+This `OpDesc` value is in the `ops` field of the `BlockDesc` value representing the global block.
+
+
+## The Compilation of Blocks
+
+During the generation of the Protobuf message, the Block should store VarDesc (the Protobuf message which describes Variable) and OpDesc (the Protobuf message which describes Operator).
+
+VarDesc in a block should have its name scope to avoid local variables affect parent block's name scope.
+Child block's name scopes should inherit the parent's so that OpDesc in child block can reference a VarDesc that stored in parent block. For example
+
+```python
+a = pd.Varaible(shape=[20, 20])
+b = pd.fc(a, params=["fc.w", "fc.b"])
+
+rnn = pd.create_rnn()
+with rnn.stepnet() as net:
+    x = net.set_inputs(a)
+    # reuse fc's parameter
+    fc_without_b = pd.get_variable("fc.w")
+    net.set_outputs(fc_without_b)
+
+out = rnn()
+```
+the method `pd.get_variable` can help retrieve a Variable by a name, a Variable may store in a parent block, but might be retrieved in a child block, so block should have a variable scope that supports inheritance.
+
+In compiler design, the symbol table is a data structure created and maintained by compilers to store information about the occurrence of various entities such as variable names, function names, classes, etc.
+
+To store the definition of variables and operators, we define a C++ class `SymbolTable`, like the one used in compilers.
+
+`SymbolTable` can do the following stuff:
+
+- store the definitions (some names and attributes) of variables and operators,
+- to verify if a variable was declared,
+- to make it possible to implement type checking (offer Protobuf message pointers to `InferShape` handlers).
+
+
+```c++
+// Information in SymbolTable is enough to trace the dependency graph. So maybe
+// the Eval() interface takes a SymbolTable is enough.
+class SymbolTable {
+ public:
+  SymbolTable(SymbolTable* parent) : parent_(parent) {}
+
+  OpDesc* NewOp(const string& name="");
+
+  // TODO determine whether name is generated by python or C++
+  // currently assume that a unique name will be generated by C++ if the
+  // argument name left default.
+  VarDesc* NewVar(const string& name="");
+
+  // find a VarDesc by name, if recursive true, find parent's SymbolTable
+  // recursively.
+  // this interface is introduced to support InferShape, find protobuf messages
+  // of variables and operators, pass pointers into InferShape.
+  // operator
+  //
+  // NOTE maybe some C++ classes such as VarDescBuilder and OpDescBuilder should
+  // be proposed and embedded into pybind to enable python operate on C++ pointers.
+  VarDesc* FindVar(const string& name, bool recursive=true);
+
+  OpDesc* FindOp(const string& name);
+
+  BlockDesc Compile() const;
+
+ private:
+  SymbolTable* parent_;
+
+  map<string, OpDesc> ops_;
+  map<string, VarDesc> vars_;
+};
+```
+
+After all the description of variables and operators is added into SymbolTable,
+the block has enough information to run.
+
+The `Block` class takes a `BlockDesc` as input, and provide `Run` and `InferShape` functions.
+
+
+```c++
+namespace {
+
+class Block : OperatorBase {
+public:
+  Block(const BlockDesc& desc) desc_(desc) {}
+
+  void InferShape(const framework::Scope& scope) const override {
+    if (!symbols_ready_) {
+      CreateVariables(scope);
+      CreateOperators();
+    }
+    // should run InferShape first.
+    for (auto& op : runtime_table_.ops()) {
+      op->InferShape(scope);
+    }
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
+    PADDLE_ENFORCE(symbols_ready_, "operators and variables should be created first.");
+    for (auto& op : runtime_table_.ops()) {
+      op->Run(scope, dev_ctx);
+    }
+  }
+
+  void CreateVariables(const framework::Scope& scope);
+  void CreateOperators();
+
+  // some other necessary interfaces of NetOp are list below
+  // ...
+
+private:
+  BlockDesc desc_;
+  bool symbols_ready_{false};
+};
+```
+
+## The Execution of Blocks
+
+Block inherits from OperatorBase, which has a Run method.
+Block's Run method will run its operators sequentially.
+
+There is another important interface called `Eval`, which take some arguments called targets, and generate a minimal graph which takes targets as the end points and creates a new Block,
+after `Run`, `Eval` will get the latest value and return the targets.
+
+The definition of Eval is as follows:
+
+```c++
+// clean a block description by targets using the corresponding dependency graph.
+// return a new BlockDesc with minimal number of operators.
+// NOTE not return a Block but the block's description so that this can be distributed
+// to a cluster.
+BlockDesc Prune(const BlockDesc& desc, vector<string> targets);
+
+void Block::Eval(const vector<string>& targets,
+                 const framework::Scope& scope,
+                 const platform::DeviceContext& dev_ctx) {
+  BlockDesc min_desc = Prune(desc_, targets);
+  Block min_block(min_desc);
+  min_block.Run(scope, dev_ctx);
+}
+```
diff --git a/doc/design/if_else_op.md b/doc/design/if_else_op.md
index 7370c2a24fa644a64e738f202bac9b9209642e08..954a19c0733358c235eae3cffe134c23dac94c95 100644
--- a/doc/design/if_else_op.md
+++ b/doc/design/if_else_op.md
@@ -1,22 +1,4 @@
-IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has M (M<=N) instances, each corresponds to a true element in `cond`.
-
-```python
-import paddle as pd
-
-x = var()
-y = var()
-cond = var()
-
-b = pd.create_ifop(inputs=[x], output_num=1)
-with b.true_block():
-    x = b.inputs(0)
-    z = operator.add(x, y)
-    b.set_output(0, operator.softmax(z))
-
-out = b(cond)
-```
-
-If we want the output still has N instances, we can use IfElseOp with a default value, whose minibatch size must be N:
+IfOp should have only one branch. An IfOp operator takes a `cond` variable whose value must be a vector of N boolean elements. Its return value has N instances. If cond[i] == True, input instance input[i] will go through true_block() and generate output[i]; otherwise it will produce output from false_bloack().
 
 ```python
 import paddle as pd
@@ -39,7 +21,7 @@ with b.false_block():
 out = b(cond)
 ```
 
-If only true_block is set in an IfElseOp, we can have a default value for false as:
+If only true_block is set in an IfElseOp, a special case is that we can have a default value for false as:
 ```python
 import paddle as pd
 
diff --git a/doc/design/refactorization.md b/doc/design/refactorization.md
new file mode 100644
index 0000000000000000000000000000000000000000..e105861e926411a269b0b52dd4688744912c9ab3
--- /dev/null
+++ b/doc/design/refactorization.md
@@ -0,0 +1,253 @@
+# Design Doc: Refactorization Overview
+
+The goal of refactorizaiton include:
+
+1. Make it easy for external contributors to write new elementory computaiton operations.
+1. Make the codebase clean and readable.
+1. Introduce a new design of computation representation -- a computation graph of operators and variables.
+1. The graph representation helps implementing auto-scalable and auto fault recoverable distributed computing.
+
+## Computation Graphs
+
+1. PaddlePaddle represent the computation, training and inference of DL models, by computation graphs.
+
+  1. Please dig into [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a solid example.
+
+1. Users write Python programs to describe the graphs and run it (locally or remotely).
+
+1. A graph is composed of *variabels* and *operators*.
+
+1. The description of graphs must be able to be serialized/deserialized, so it
+
+   1. could to be sent to the cloud for distributed execution, and
+   1. be sent to clients for mobile or enterprise deployment.
+
+1. The Python program do
+
+   1. *compilation*: runs a Python program to generate a protobuf message representation of the graph and send it to
+      1. the C++ library `libpaddle.so` for local execution,
+      1. the master process of a distributed training job for training, or
+      1. the server process of a Kubernetes serving job for distributed serving.
+   1. *execution*: according to the protobuf message, constructs instances of class `Variable` and `OperatorBase`, and run them.
+
+## Description and Realization
+
+At compile time, the Python program generates protobuf message representation of the graph, or the description of the graph.
+
+At runtime, the C++ program realizes the graph and run it.
+
+| | Representation (protobuf messages) | Realization (C++ class objects) |
+|---|---|---|
+|Data|[VarDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L107)|[Variable](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24)|
+|Operation|[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35)|[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64)|
+|Block|BlockDesc|Block|
+
+The word *graph* is exchangable with *block* in this document.  A graph represent computation steps and local variables as a C++/Java program block, or a pair of { and }.
+
+## Compilation and Execution
+
+1. Run an applicaton Python program to describe the graph.  In particular,
+
+   1. create VarDesc to represent local/intermediate variables,
+   1. create operators and set attributes,
+   1. validate attribute values,
+   1. inference the type and the shape of variables,
+   1. plan for memory-reuse for variables,
+   1. generate backward and optimization part of the Graph.
+   1. possiblly split the graph for distributed training.
+
+1. The invocation of `train` or `infer` in the application Python program:
+
+   1. create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block,
+      1. realize local variables defined in the BlockDesc message in the new scope,
+      1. a scope is similar to the stack frame in programming languages,
+
+   1. create an instance of class `Block`, in which,
+      1. realize operators in the BlockDesc message,
+
+   1. run the Block by calling
+      1. `Block::Eval(vector<Variable>* targets)` for forward and backward computations, or
+      1. `Block::Eval(vector<Operator>* targets)` for optimization.
+
+
+## Intermediate Representation (IR)
+
+```text
+Compile Time -> IR -> Runtime
+```
+
+### Benefit
+
+- Optimization
+  ```text
+  Compile Time -> IR -> Optimized IR -> Runtime
+  ```
+- Send automatically partitioned IR to different nodes.
+  - Automatic data parallel
+    ```text
+    Compile Time
+    |-> Single GPU IR
+        |-> [trainer-IR-0, trainer-IR-1, pserver-IR]
+            |-> Node-0 (runs trainer-IR-0)
+            |-> Node-1 (runs trainer-IR-1)
+            |-> Node-2 (runs pserver-IR)
+    ```
+  - Automatic model parallel (planned for future)
+
+---
+
+# Operator/OpWithKernel/OpKernel
+
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot)
+
+---
+
+# Operator
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot)
+
+* `Operator` is the fundamental building block as the user interface.
+    * Operator stores input/output variable name, and attributes.
+    * The `InferShape` interface is used to infer output variable shapes by its input shapes.
+    * Use `Run` to compute `input variables` to `output variables`.
+
+---
+
+# OpWithKernel/Kernel
+
+![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot)
+
+* `OpWithKernel` inherits `Operator`.
+* `OpWithKernel` contains a Kernel map.
+    * `OpWithKernel::Run` get device's kernel, and invoke `OpKernel::Compute`.
+    * `OpKernelKey` is the map key. Only device place now, but may be data type later.
+
+---
+
+# Why separate Kernel and Operator
+
+* Separate GPU and CPU code.
+    * Make Paddle can run without GPU.
+* Make one operator (which is user interface) can contain many implementations.
+    * Same mul op, different FP16, FP32 Kernel. different MKL, eigen kernel.
+---
+
+# Libraries for Kernel development
+
+* `Eigen::Tensor` contains basic math and element-wise functions.
+    * Note that `Eigen::Tensor` has broadcast implementation.
+    * Limit number of `tensor.device(dev) = ` in your code.
+* `thrust::tranform` and `std::transform`.
+    * `thrust` has the same API as C++ standard library. Using `transform` can quickly implement a customized elementwise kernel.
+    * `thrust` has more complex API, like `scan`, `reduce`, `reduce_by_key`.
+* Hand-writing `GPUKernel` and `CPU` code
+    * Do not write `.h`. CPU Kernel should be in `.cc`. CPU kernel should be in `.cu`. (`GCC` cannot compile GPU code.)
+---
+# Operator Register
+
+## Why register is necessary?
+We need a method to build mappings between Op type names and Op classes.
+
+## How to do the register?
+
+Maintain a map, whose key is the type name and value is corresponding Op constructor.
+
+---
+# The Registry Map
+
+### `OpInfoMap`
+
+`op_type(string)` -> `OpInfo`
+
+`OpInfo`:
+
+- **`creator`**: The Op constructor.
+- **`grad_op_type`**: The type of the gradient Op.
+- **`proto`**: The Op's Protobuf, including inputs, outputs and required attributes.
+- **`checker`**: Used to check attributes.
+
+---
+# Related Concepts
+
+### Op_Maker
+It's constructor takes `proto` and `checker`. They are compeleted during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37))
+
+### Register Macros
+```cpp
+REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class)
+REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
+```
+
+### `USE` Macros
+make sure the registration process is executed and linked.
+
+---
+# Register Process
+1. Write Op class, as well as its gradient Op class if there is.
+2. Write Op maker class. In the constructor, describe its inputs, outputs, and attributes.
+3. Invoke macro `REGISTER_OP`. The macro will
+	1. call maker class to complete `proto` and `checker`
+	2. with the completed `proto` and `checker`, build a new key-value pair in the `OpInfoMap`
+
+4. Invoke `USE` macro in where the Op is used to make sure it is linked.
+
+---
+# Backward Module (1/2)
+### Create Backward Operator
+- Mapping from forwarding Op to backward Op
+![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png)
+
+---
+# Backward Module (2/2)
+### Build Backward Network
+- **Input** graph of forwarding operators
+- **Output** graph of backward operators
+- **corner case in construction**
+	- shared variable => insert `Add` operator
+	- no gradient => insert `fill_zero_grad` operator
+	- recursive netOp => call `Backward` recursively
+	- RNN Op => recursively call `Backward` on stepnet
+
+
+---
+# Scope, Variable, Tensor
+
+* `Tensor` is an n-dimension array with type.
+	* Only dims and data pointers are stored in `Tensor`.
+	* All operators on `Tensor` is written in `Operator` or global functions.
+	* variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md)
+* `Variable` is the inputs and outputs of an operator. Not just `Tensor`.
+	* step_scopes in RNN is a variable and not a tensor.
+* `Scope` is where variables store at.
+	* map<string/*var name */, Variable>
+	* `Scope` has a hierarchical structure. The local scope can get variable from its parent scope.
+
+---
+# Block (in design)
+## the difference with original RNNOp
+- as an operator is more intuitive than `RNNOp`,
+- offers new interface `Eval(targets)` to deduce the minimal block to `Run`,
+- fits the compile-time/ runtime separation design.
+  - during the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc`
+  - when graph executes, a Block with `BlockDesc` passed in creates `Op` and `Var` then `Run`
+
+---
+# Milestone
+- take Paddle/books as the main line, the requirement of the models motivates framework refactoring,
+- model migration
+  - framework development gives **priority support** to model migration, for example,
+    - the MNIST demo needs a Python interface,
+    - the RNN models require the framework to support `LoDTensor`.
+  - determine some timelines,
+  - heavily-relied Ops need to be migrated first,
+  - different models can be migrated parallelly.
+- improve the framework at the same time
+- accept imperfection, concentrated on solving the specific problem at the right price.
+
+---
+# Control the migration quality
+- compare the performance of migrated models with old ones.
+- follow google C style
+- build the automatic workflow of generating Python/C++ documentations
+  - the documentation of layers and ops should be written inside the code
+  - take the documentation quality into account when doing PR
+  - preview the documentations, read and improve them from users' perspective
diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt
index ec866b2907d4623e8a94a249bc9af624071ade97..b435de80a224571d16efdee168541aa301c3f73a 100644
--- a/paddle/CMakeLists.txt
+++ b/paddle/CMakeLists.txt
@@ -19,7 +19,7 @@ if(Boost_FOUND)
 endif()
 
 if(WITH_C_API)
-    add_subdirectory(capi)
+  add_subdirectory(capi)
 endif()
 
 if(WITH_SWIG_PY)
diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt
index 3af111eb5738c3f2f399ff4e5c06c8d2ecd8973e..dd9e4f1cbd636e29a6934d1119fc93ebc9d0ecee 100644
--- a/paddle/capi/CMakeLists.txt
+++ b/paddle/capi/CMakeLists.txt
@@ -28,42 +28,38 @@ add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER}
 
 add_dependencies(paddle_capi paddle_proto)
 
-
 # combine all paddle static libraries together, into libpaddle_capi_whole.a
 # user should use PaddleCAPI as -lpaddle_capi_whole
-set(capi_whole_library libpaddle_capi_whole.a)
-add_custom_target(paddle_capi_whole ALL
-        COMMAND mkdir -p o_files/capi && cd o_files/capi/ && ar -x $<TARGET_FILE:paddle_capi>
-        COMMAND mkdir -p o_files/utils && cd o_files/utils/ && ar -x $<TARGET_FILE:paddle_utils>
-        COMMAND mkdir -p o_files/parameter && cd o_files/parameter/ && ar -x $<TARGET_FILE:paddle_parameter>
-        COMMAND mkdir -p o_files/math && cd o_files/math/  && ar -x $<TARGET_FILE:paddle_math>
-        COMMAND mkdir -p o_files/cuda && cd o_files/cuda/ && ar -x $<TARGET_FILE:paddle_cuda>
-        COMMAND mkdir -p o_files/function && cd o_files/function/ && ar -x $<TARGET_FILE:paddle_function>
-        COMMAND mkdir -p o_files/gserver && cd o_files/gserver/ && ar -x $<TARGET_FILE:paddle_gserver>
-        COMMAND mkdir -p o_files/proto && cd o_files/proto/ && ar -x $<TARGET_FILE:paddle_proto>
-        COMMAND mkdir -p o_files/network && cd o_files/network/ && ar -x $<TARGET_FILE:paddle_network>
-        COMMAND mkdir -p o_files/pserver && cd o_files/pserver/ && ar -x $<TARGET_FILE:paddle_pserver>
-        COMMAND ar crs ${capi_whole_library} `find ./o_files -name '*.o'`
-        COMMAND rm -rf o_files
-        WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
-        DEPENDS paddle_capi paddle_utils paddle_parameter paddle_math
-                paddle_cuda paddle_function paddle_gserver
-                paddle_proto paddle_pserver paddle_network
-        )
-set_target_properties(paddle_capi_whole
-  PROPERTIES IMPORTED_LOCATION ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library})
+set(PADDLE_CAPI_INFER_LIBS
+    paddle_utils
+    paddle_parameter
+    paddle_math
+    paddle_cuda
+    paddle_function
+    paddle_gserver
+    paddle_proto
+    paddle_pserver
+    paddle_network)
+
+cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
 
-set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
-# TODO: merge mkl into paddle_capi_shared
-add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
-set_target_properties(paddle_capi_shared	PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
-target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
-link_paddle_exe(paddle_capi_shared)
+# No shared library for iOS
+if(NOT IOS)
+  set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
+  # TODO: merge mkl into paddle_capi_shared
+  add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
+  set_target_properties(paddle_capi_shared	PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+  target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
+  link_paddle_exe(paddle_capi_shared)
+endif()
 
 # install library & headers.
 install(FILES ${CAPI_HEADERS} DESTINATION include/paddle)
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/config.h DESTINATION include/paddle)
 if(ANDROID)
+  install(TARGETS paddle_capi_whole paddle_capi_shared
+          ARCHIVE DESTINATION lib/${ANDROID_ABI}
+          LIBRARY DESTINATION lib/${ANDROID_ABI})
   execute_process(
     COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1
     OUTPUT_VARIABLE GIT_COMMITS_LIST
@@ -72,9 +68,6 @@ if(ANDROID)
   if(${GIT_COMMITS_LIST_RESULT})
     set(GIT_COMMITS_LIST "No commits.")
   endif()
-  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}
-          DESTINATION lib/${ANDROID_ABI})
-  install(TARGETS paddle_capi_shared DESTINATION lib/${ANDROID_ABI})
   install(CODE "FILE(WRITE ${CMAKE_INSTALL_PREFIX}/lib/${ANDROID_ABI}/BUILD.txt
           \"Compiler:\n\"
           \"\\t${CMAKE_C_COMPILER}\\n\"
@@ -88,8 +81,11 @@ if(ANDROID)
       )"
   )
 else(ANDROID)
-  install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library} DESTINATION lib)
-  install(TARGETS paddle_capi_shared DESTINATION lib)
+  install(TARGETS paddle_capi_whole
+          ARCHIVE DESTINATION lib)
+  if(NOT IOS)
+    install(TARGETS paddle_capi_shared DESTINATION lib)
+  endif()
 endif(ANDROID)
 
 # this variable used for unittest
diff --git a/paddle/cuda/include/hl_cuda_cudnn.h b/paddle/cuda/include/hl_cuda_cudnn.h
index 3f68c62de6d9b3aaadc9180d86159089dc728ea9..b44b071bd1b3b6e9e5539d5dc0c2b155c524fd57 100644
--- a/paddle/cuda/include/hl_cuda_cudnn.h
+++ b/paddle/cuda/include/hl_cuda_cudnn.h
@@ -22,10 +22,10 @@ limitations under the License. */
  */
 typedef enum {
   HL_POOLING_MAX = 0,
-  // average includes padded values
-  HL_POOLING_AVERAGE = 1,
   // average does not include padded values
-  HL_POOLING_AVERAGE_EXCLUDE_PADDING = 2,
+  HL_POOLING_AVERAGE = 1,
+  // average includes padded values
+  HL_POOLING_AVERAGE_INCLUDE_PADDING = 2,
   HL_POOLING_END
 } hl_pooling_mode_t;
 
diff --git a/paddle/cuda/include/hl_tensor_ops.h b/paddle/cuda/include/hl_tensor_ops.h
index 93d38b7d2299d994cde0934213668a525bffa80c..b2bf334dab9799153fe1d4fe2c74cce9d57168b9 100644
--- a/paddle/cuda/include/hl_tensor_ops.h
+++ b/paddle/cuda/include/hl_tensor_ops.h
@@ -461,7 +461,7 @@ class add<float32x4_t> {
 public:
   INLINE float32x4_t operator()(const float32x4_t a,
                                 const float32x4_t b) const {
-    return vmulq_f32(a, b);
+    return vaddq_f32(a, b);
   }
 };
 
diff --git a/paddle/cuda/src/hl_cuda_cnn.cu b/paddle/cuda/src/hl_cuda_cnn.cu
index 9ba3d142617537c0160f6dccb86ddca43ada15a5..58674febdc4a094c95ff03701e4586c32729847d 100644
--- a/paddle/cuda/src/hl_cuda_cnn.cu
+++ b/paddle/cuda/src/hl_cuda_cnn.cu
@@ -211,13 +211,11 @@ __global__ void KeAvgPoolForward(const int nthreads,
 
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (hend - hstart) * (wend - wstart);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * height * width;
@@ -299,12 +297,14 @@ __global__ void KeAvgPoolBackward(const int nthreads,
     outGrad += (frameNum * outStride + offsetC * pooledH * pooledW);
 
     for (int ph = phstart; ph < phend; ++ph) {
+      int hstart = ph * strideH - padH;
+      int hend = min(hstart + sizeY, height);
+      hstart = max(hstart, 0);
       for (int pw = pwstart; pw < pwend; ++pw) {
         // figure out the pooling size
-        int hstart = ph * strideH - padH;
         int wstart = pw * strideW - padW;
-        int hend = min(hstart + sizeY, height + padH);
-        int wend = min(wstart + sizeX, width + padW);
+        int wend = min(wstart + sizeX, width);
+        wstart = max(wstart, 0);
         int poolsize = (hend - hstart) * (wend - wstart);
         gradient += outGrad[ph * pooledW + pw] / poolsize;
       }
@@ -600,16 +600,13 @@ __global__ void KeAvgPool3DForward(const int nthreads,
     int dstart = pd * strideD - padD;
     int hstart = ph * strideH - padH;
     int wstart = pw * strideW - padW;
-    int dend = min(dstart + sizeZ, depth + padD);
-    int hend = min(hstart + sizeY, height + padH);
-    int wend = min(wstart + sizeX, width + padW);
-    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
+    int dend = min(dstart + sizeZ, depth);
+    int hend = min(hstart + sizeY, height);
+    int wend = min(wstart + sizeX, width);
     dstart = max(dstart, 0);
     hstart = max(hstart, 0);
     wstart = max(wstart, 0);
-    dend = min(dend, depth);
-    hend = min(hend, height);
-    wend = min(wend, width);
+    int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart);
 
     real aveval = 0;
     inputData += (frameNum * channels + c) * depth * height * width;
@@ -712,15 +709,18 @@ __global__ void KeAvgPool3DBackward(const int nthreads,
     outGrad += (frameNum * channels + offsetC) * pooledD * pooledH * pooledW;
 
     for (int pd = pdstart; pd < pdend; ++pd) {
+      int dstart = pd * strideD - padD;
+      int dend = min(dstart + sizeZ, depth);
+      dstart = max(dstart, 0);
       for (int ph = phstart; ph < phend; ++ph) {
+        int hstart = ph * strideH - padH;
+        int hend = min(hstart + sizeY, height);
+        hstart = max(hstart, 0);
         for (int pw = pwstart; pw < pwend; ++pw) {
           // figure out the pooling size
-          int dstart = pd * strideD - padD;
-          int hstart = ph * strideH - padH;
           int wstart = pw * strideW - padW;
-          int dend = min(dstart + sizeZ, depth + padD);
-          int hend = min(hstart + sizeY, height + padH);
-          int wend = min(wstart + sizeX, width + padW);
+          int wend = min(wstart + sizeX, width);
+          wstart = max(wstart, 0);
           int poolsize = (dend - dstart) * (hend - hstart) * (wend - wstart);
           gradient += outGrad[(pd * pooledH + ph) * pooledW + pw] / poolsize;
         }
diff --git a/paddle/cuda/src/hl_cuda_cudnn.cc b/paddle/cuda/src/hl_cuda_cudnn.cc
index f38ef692558b908ed65d2c84821bbb7c3b439742..b8caf48f9c06094e85765f7aa5a3f4195d0ca931 100644
--- a/paddle/cuda/src/hl_cuda_cudnn.cc
+++ b/paddle/cuda/src/hl_cuda_cudnn.cc
@@ -432,11 +432,11 @@ void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc,
       cudnn_mode = CUDNN_POOLING_MAX;
       break;
     case HL_POOLING_AVERAGE:
-      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
-      break;
-    case HL_POOLING_AVERAGE_EXCLUDE_PADDING:
       cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
       break;
+    case HL_POOLING_AVERAGE_INCLUDE_PADDING:
+      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
+      break;
     default:
       LOG(FATAL) << "parameter mode error";
   }
diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index c57537be4bf67a8db6a49669ab8d2ed1b1324bdc..f8a64a786611ef872dbbfced10919e00c4d46715 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -22,14 +22,14 @@ namespace framework {
 template <>
 Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
     platform::CPUPlace, Eigen::DefaultDevice>() const {
-  return *device_context_->get_eigen_device<Eigen::DefaultDevice>();
+  return *device_context_.get_eigen_device<Eigen::DefaultDevice>();
 }
 
 #ifndef PADDLE_ONLY_CPU
 template <>
 Eigen::GpuDevice&
 ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
-  return *device_context_->get_eigen_device<Eigen::GpuDevice>();
+  return *device_context_.get_eigen_device<Eigen::GpuDevice>();
 }
 #endif
 
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index adae7bfc3d7d31b1ed0373f01db4ef80343a08f7..b7c9c39402d57daf0aec97d98535ac8a8d9c0150 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -366,7 +366,7 @@ struct EigenDeviceConverter<platform::GPUPlace> {
 class ExecutionContext : public InferShapeContext {
  public:
   ExecutionContext(const OperatorBase& op, const Scope& scope,
-                   const platform::DeviceContext* device_context)
+                   const platform::DeviceContext& device_context)
       : InferShapeContext(op, scope), device_context_(device_context) {}
 
   template <typename PlaceType,
@@ -374,9 +374,9 @@ class ExecutionContext : public InferShapeContext {
                 typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
   DeviceType& GetEigenDevice() const;
 
-  platform::Place GetPlace() const { return device_context_->GetPlace(); }
+  platform::Place GetPlace() const { return device_context_.GetPlace(); }
 
-  const platform::DeviceContext* device_context() const {
+  const platform::DeviceContext& device_context() const {
     return device_context_;
   }
 
@@ -401,7 +401,8 @@ class ExecutionContext : public InferShapeContext {
     return res;
   }
 
-  const platform::DeviceContext* device_context_;
+ private:
+  const platform::DeviceContext& device_context_;
 };
 
 template <>
@@ -461,7 +462,7 @@ class OperatorWithKernel : public OperatorBase {
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const final {
     auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
-    opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
+    opKernel->Compute(ExecutionContext(*this, scope, dev_ctx));
   }
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&
diff --git a/paddle/framework/tensor_impl.h b/paddle/framework/tensor_impl.h
index 642b53efc7095d25712ca324638f5fe9b8316c0c..ed166935f76be9d25062b5e69536c7b7ac19045d 100644
--- a/paddle/framework/tensor_impl.h
+++ b/paddle/framework/tensor_impl.h
@@ -22,7 +22,7 @@ namespace framework {
 template <typename T>
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
-      holder_, "Tenosr holds no memory. Call Tensor::mutable_data first.");
+      holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
   PADDLE_ENFORCE_GE(
       holder_->size(), numel() * sizeof(T) + offset_,
       "Tensor's dims_ is out of bound. Call Tensor::mutable_data "
diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc
index 55302ea47120f420e952b26830c8ea4cbcce6435..e2ec738de35c90c6a06c9a46b062d4cce55f5eda 100644
--- a/paddle/framework/tensor_test.cc
+++ b/paddle/framework/tensor_test.cc
@@ -36,7 +36,7 @@ TEST(Tensor, DataAssert) {
   } catch (paddle::platform::EnforceNotMet err) {
     caught = true;
     std::string msg =
-        "holder_ should not be null\nTenosr holds no memory. Call "
+        "holder_ should not be null\nTensor holds no memory. Call "
         "Tensor::mutable_data first.";
     const char* what = err.what();
     for (size_t i = 0; i < msg.length(); ++i) {
@@ -112,7 +112,7 @@ TEST(Tensor, ShareDataWith) {
     } catch (paddle::platform::EnforceNotMet err) {
       caught = true;
       std::string msg =
-          "holder_ should not be null\nTenosr holds no memory. Call "
+          "holder_ should not be null\nTensor holds no memory. Call "
           "Tensor::mutable_data first.";
       const char* what = err.what();
       for (size_t i = 0; i < msg.length(); ++i) {
@@ -274,4 +274,4 @@ TEST(Tensor, ReshapeToMatrix) {
   Tensor res = ReshapeToMatrix<int>(src, 2);
   ASSERT_EQ(res.dims()[0], 2 * 3);
   ASSERT_EQ(res.dims()[1], 4 * 9);
-}
\ No newline at end of file
+}
diff --git a/paddle/function/neon/NeonDepthwiseConv.cpp b/paddle/function/neon/NeonDepthwiseConv.cpp
index 18126152ea0b4ebfe4ec5c8084479787814ed173..38aa6670612b0771cdd8f1805a6d1bd9f281bdc1 100644
--- a/paddle/function/neon/NeonDepthwiseConv.cpp
+++ b/paddle/function/neon/NeonDepthwiseConv.cpp
@@ -52,7 +52,7 @@ public:
     int outputHeight = output[2];
     int outputWidth = output[3];
     int filterMultiplier = outputChannels / groups_;
-    CHECK_EQ(inputChannels, groups_);
+    CHECK_EQ(static_cast<size_t>(inputChannels), groups_);
 
     // only support strideH() == strideW() and filterHeight == filterWidth.
     CHECK_EQ(strideH(), strideW());
diff --git a/paddle/gserver/activations/ActivationFunction.cpp b/paddle/gserver/activations/ActivationFunction.cpp
index 78e958e06fac84fa956abc9faea60157bf6132eb..8b7b2e9b65898950e036ebc023cd28990cef303f 100644
--- a/paddle/gserver/activations/ActivationFunction.cpp
+++ b/paddle/gserver/activations/ActivationFunction.cpp
@@ -22,9 +22,12 @@ limitations under the License. */
 #include <type_traits>
 #include "paddle/parameter/Argument.h"
 #include "paddle/utils/ClassRegistrar.h"
-
 #include "paddle/utils/Logging.h"
 
+#ifdef PADDLE_USE_MKLDNN
+#include "MKLDNNActivation.h"
+#endif
+
 namespace paddle {
 
 static ClassRegistrar<ActivationFunction> gActivationRegistrar;
@@ -456,6 +459,12 @@ Error __must_check backward(Argument& act) {
 END_DEFINE_ACTIVATION(log)
 
 ActivationFunction* ActivationFunction::create(const std::string& type) {
+#ifdef PADDLE_USE_MKLDNN
+  if (!type.empty() && type.compare(0, 7, "mkldnn_") == 0) {
+    return MKLDNNActivation::create(type);
+  }
+#endif
+
   return gActivationRegistrar.createByType(type);
 }
 
diff --git a/paddle/gserver/activations/MKLDNNActivation.cpp b/paddle/gserver/activations/MKLDNNActivation.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..ac50937ef3e28c1ac5aae651f9cf266ad07abcc4
--- /dev/null
+++ b/paddle/gserver/activations/MKLDNNActivation.cpp
@@ -0,0 +1,87 @@
+/* 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 "MKLDNNActivation.h"
+#include "mkldnn.hpp"
+#include "paddle/utils/ClassRegistrar.h"
+
+namespace paddle {
+
+static ClassRegistrar<ActivationFunction> gMKLDNNActivationRegistrar;
+/**
+ * @def MKLDNN_ACTIVATION_CLASS_NAME
+ * @note MKLDNN_ACTIVATION_CLASS_NAME(relu) relu_;
+ * means mkldnn_reluActivation relu_;
+ */
+#define MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE) mkldnn_##ACT_TYPE##Activation
+
+/**
+ * @def DEFINE_MKLDNN_ELTWISE_ACTIVATION
+ */
+#define DEFINE_MKLDNN_ELTWISE_ACTIVATION(ACT_TYPE, ALPHA, BWD_ALPHA)        \
+  class MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)                              \
+      : public MKLDNNEltwiseActivation {                                    \
+  private:                                                                  \
+    static const std::string name;                                          \
+    static const float alpha;                                               \
+    static const float bwdAlpha;                                            \
+                                                                            \
+  public:                                                                   \
+    const std::string& getName() const { return name; }                     \
+    float getAlpha() const { return alpha; }                                \
+    float getBwdAlpha() const { return bwdAlpha; }                          \
+  };                                                                        \
+  const std::string MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::name =          \
+      "mkldnn_" #ACT_TYPE;                                                  \
+  const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::alpha = ALPHA;        \
+  const float MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)::bwdAlpha = BWD_ALPHA; \
+  static InitFunction __reg_activation__mkldnn_##ACT_TYPE([] {              \
+    gMKLDNNActivationRegistrar                                              \
+        .registerClass<MKLDNN_ACTIVATION_CLASS_NAME(ACT_TYPE)>(             \
+            "mkldnn_" #ACT_TYPE);                                           \
+  });
+
+/**
+ * @brief MKLDNN Relu Activation.
+ * Actually mkldnn_relu is Leaky Relu.
+ *  f(x) = x                   (x >= 0)
+ *  f(x) = negative_slope * x  (x <  0)
+ * @note the negative_slope should be -0.f in forward
+ */
+DEFINE_MKLDNN_ELTWISE_ACTIVATION(relu, -0.f, 0.f)
+
+/**
+ * @brief MKLDNN Tanh Activation.
+ */
+DEFINE_MKLDNN_ELTWISE_ACTIVATION(tanh, 0.f, 0.f)
+
+/**
+ * @brief MKLDNN ELU(Exponential Linear Unit) Activation.
+ *  f(x) = x                              (x >= 0)
+ *  f(x) = negative_slope * (exp(x) - 1)  (x <  0)
+ */
+DEFINE_MKLDNN_ELTWISE_ACTIVATION(elu, 0.f, 0.f)
+
+ActivationFunction* MKLDNNActivation::create(const std::string& type) {
+  return gMKLDNNActivationRegistrar.createByType(type);
+}
+
+std::vector<std::string> MKLDNNActivation::getAllRegisteredTypes() {
+  std::vector<std::string> types;
+  gMKLDNNActivationRegistrar.forEachType(
+      [&](const std::string& type) { types.push_back(type); });
+  return types;
+}
+
+}  // namespace paddle
diff --git a/paddle/gserver/activations/MKLDNNActivation.h b/paddle/gserver/activations/MKLDNNActivation.h
new file mode 100644
index 0000000000000000000000000000000000000000..bda9bbebe5600dbe26d11ff32058f7b2647b763e
--- /dev/null
+++ b/paddle/gserver/activations/MKLDNNActivation.h
@@ -0,0 +1,182 @@
+/* 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 "ActivationFunction.h"
+#include "mkldnn.hpp"
+#include "paddle/gserver/layers/MKLDNNBase.h"
+#include "paddle/math/MKLDNNMatrix.h"
+#include "paddle/parameter/Argument.h"
+
+namespace paddle {
+
+/**
+ * @brief Base class of MKLDNN Activation.
+ * Common activation function are provieded,
+ * including mkldnn_relu, mkldnn_elu, mkldnn_tanh, mkldnn_softmax
+ */
+class MKLDNNActivation : public ActivationFunction {
+protected:
+  // input value element count
+  size_t cnt_;
+  // should not merge the resetBwd into resetFwd,
+  // because the grad data would be changing before backward.
+  bool needResetBwd_;
+  // mkldnn matrix, primitive, stream and pipeline
+  MKLDNNMatrixPtr val_;
+  MKLDNNMatrixPtr grad_;
+  std::shared_ptr<MKLDNNStream> stream_;
+  std::shared_ptr<mkldnn::primitive> fwd_;
+  std::shared_ptr<mkldnn::primitive> bwd_;
+  std::vector<mkldnn::primitive> pipelineFwd_;
+  std::vector<mkldnn::primitive> pipelineBwd_;
+
+public:
+  MKLDNNActivation() : cnt_(0), needResetBwd_(true) {}
+  ~MKLDNNActivation() {}
+  static ActivationFunction* create(const std::string& type);
+  static std::vector<std::string> getAllRegisteredTypes();
+  virtual const std::string& getName() const = 0;
+  virtual Error __must_check forward(Argument& act) = 0;
+  virtual Error __must_check backward(Argument& act) = 0;
+};
+
+/**
+ * @brief Base class of MKLDNN Eltwise Activation,
+ * includes mkldnn_relu, mkldnn_elu and mkldnn_tanh.
+ */
+class MKLDNNEltwiseActivation : public MKLDNNActivation {
+  typedef mkldnn::eltwise_forward eltwise_fwd;
+  typedef mkldnn::eltwise_backward eltwise_bwd;
+
+protected:
+  // save the forward primitive desc, which can be used backward
+  std::shared_ptr<eltwise_fwd::primitive_desc> fwdPD_;
+  // eltwise_bwd need src input value
+  MKLDNNMatrixPtr inVal_;
+  // use for copy data
+  std::shared_ptr<mkldnn::reorder> copyInVal_;
+
+public:
+  MKLDNNEltwiseActivation() {}
+
+  ~MKLDNNEltwiseActivation() {}
+
+  virtual const std::string& getName() const = 0;
+
+  // in common, the alpha of forward and backward should be equal.
+  // but for relu, to avoid negative value, they should be opposite
+  virtual float getAlpha() const = 0;
+  virtual float getBwdAlpha() const = 0;
+  virtual float getBeta() const { return 0.f; }
+  virtual mkldnn::algorithm getAlgo(const std::string& type) const {
+    if (type == "mkldnn_relu") {
+      return mkldnn::algorithm::eltwise_relu;
+    } else if (type == "mkldnn_tanh") {
+      return mkldnn::algorithm::eltwise_tanh;
+    } else if (type == "mkldnn_elu") {
+      return mkldnn::algorithm::eltwise_elu;
+    } else {
+      LOG(FATAL) << "Unkown eltwise activation type: " << type;
+    }
+    return (mkldnn::algorithm)0;
+  }
+
+  /**
+   * reshape and reset the forward primitives
+   */
+  void resetFwd(Argument& act) {
+    if (cnt_ == act.value->getElementCnt()) {
+      return;
+    }
+    cnt_ = act.value->getElementCnt();
+    stream_.reset(new MKLDNNStream());
+    auto eng = CPUEngine::Instance().getEngine();
+
+    // get algo setting
+    mkldnn::algorithm algo = getAlgo(this->getName());
+    // note: alpha represents the NegativeSlope when used in relu.
+    float alpha = getAlpha();
+    float beta = getBeta();
+
+    /// forward
+    pipelineFwd_.clear();
+    val_ = std::dynamic_pointer_cast<MKLDNNMatrix>(act.value);
+    if (val_ == nullptr) {
+      int bs = act.getBatchSize();
+      int ih = act.getFrameHeight() > 0 ? act.getFrameHeight() : 1;
+      int iw = act.getFrameWidth() > 0 ? act.getFrameWidth() : 1;
+      int ic = cnt_ / bs / ih / iw;
+      CHECK_EQ(cnt_, (size_t)bs * ic * ih * iw);
+      val_ = MKLDNNMatrix::create(
+          act.value, {bs, ic, ih, iw}, mkldnn::memory::format::nchw, eng);
+      CHECK(val_);
+    }
+    auto fwdDesc = eltwise_fwd::desc(mkldnn::prop_kind::forward_training,
+                                     algo,
+                                     val_->getMemoryDesc(),
+                                     alpha,
+                                     beta);
+    fwdPD_.reset(new eltwise_fwd::primitive_desc(fwdDesc, eng));
+    // use inplace for forward but save input value before submit
+    inVal_ = val_;
+    if (act.grad) {
+      // only copy when need do backward
+      inVal_ = MKLDNNMatrix::create(nullptr, val_->getPrimitiveDesc());
+      copyInVal_ = std::make_shared<mkldnn::reorder>(*val_, *inVal_);
+      CHECK(copyInVal_) << "should not be emptry";
+      pipelineFwd_.push_back(*copyInVal_);
+    }
+    fwd_.reset(new eltwise_fwd(*fwdPD_, *val_, *val_));
+    pipelineFwd_.push_back(*fwd_);
+    needResetBwd_ = true;
+  }
+
+  /**
+   * reset the backward primitives, can not merge into resetFwd as the grad data
+   * would be changing before backward.
+   */
+  void resetBwd(Argument& act) {
+    if (!needResetBwd_) {
+      return;
+    }
+    needResetBwd_ = false;
+    mkldnn::algorithm algo = getAlgo(this->getName());
+    float alpha = getBwdAlpha();
+    float beta = getBeta();
+    grad_ = MKLDNNMatrix::create(act.grad, val_->getPrimitiveDesc());
+    auto eng = CPUEngine::Instance().getEngine();
+    auto bwdDesc = eltwise_bwd::desc(
+        algo, grad_->getMemoryDesc(), val_->getMemoryDesc(), alpha, beta);
+    auto bwdPD = eltwise_bwd::primitive_desc(bwdDesc, eng, *fwdPD_);
+    CHECK(inVal_);
+    bwd_.reset(new eltwise_bwd(bwdPD, *inVal_, *grad_, *grad_));
+    pipelineBwd_.clear();
+    pipelineBwd_.push_back(*bwd_);
+  }
+
+  Error __must_check forward(Argument& act) {
+    resetFwd(act);
+    stream_->submit(pipelineFwd_);
+    return Error();
+  }
+
+  Error __must_check backward(Argument& act) {
+    resetBwd(act);
+    stream_->submit(pipelineBwd_);
+    return Error();
+  }
+};
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/CudnnPoolLayer.cpp b/paddle/gserver/layers/CudnnPoolLayer.cpp
index 4adb2d4709e585a6fec052435c33714d6e3a3f0e..810a1af2d09c63c3787a1ac225c2c7de4238d609 100644
--- a/paddle/gserver/layers/CudnnPoolLayer.cpp
+++ b/paddle/gserver/layers/CudnnPoolLayer.cpp
@@ -29,9 +29,9 @@ bool CudnnPoolLayer::typeCheck(const std::string &poolType,
     if (mode) {
       *mode = HL_POOLING_AVERAGE;
     }
-  } else if (poolType == "cudnn-avg-excl-pad-pool") {
+  } else if (poolType == "cudnn-avg-incl-pad-pool") {
     if (mode) {
-      *mode = HL_POOLING_AVERAGE_EXCLUDE_PADDING;
+      *mode = HL_POOLING_AVERAGE_INCLUDE_PADDING;
     }
   } else {
     return false;
diff --git a/paddle/gserver/layers/DetectionOutputLayer.cpp b/paddle/gserver/layers/DetectionOutputLayer.cpp
index 0cf0a92bf4bd8f9b8eba2016b2377d9dfb18c70a..f9040f7ae746f9ae1736cd477d3a69a2c49e9d34 100644
--- a/paddle/gserver/layers/DetectionOutputLayer.cpp
+++ b/paddle/gserver/layers/DetectionOutputLayer.cpp
@@ -143,7 +143,7 @@ void DetectionOutputLayer::forward(PassType passType) {
     resetOutput(numKept, 7);
   } else {
     MatrixPtr outV = getOutputValue();
-    outV = NULL;
+    if (outV) outV->resize(0, 0);
     return;
   }
   MatrixPtr outV = getOutputValue();
diff --git a/paddle/gserver/layers/ExpandConvBaseLayer.cpp b/paddle/gserver/layers/ExpandConvBaseLayer.cpp
deleted file mode 100644
index 2b7bef0a757d7c706be3815c539b036b094596cf..0000000000000000000000000000000000000000
--- a/paddle/gserver/layers/ExpandConvBaseLayer.cpp
+++ /dev/null
@@ -1,124 +0,0 @@
-/* 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 "ExpandConvBaseLayer.h"
-
-#include "paddle/utils/Logging.h"
-namespace paddle {
-
-bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
-                               const ParameterMap &parameterMap) {
-  /* Initialize the basic convolutional parent class */
-  ConvBaseLayer::init(layerMap, parameterMap);
-
-  int index = 0;
-  for (auto &inputConfig : config_.inputs()) {
-    const ConvConfig &conf = inputConfig.conv_conf();
-    /* Consistent caffe mode for multiple input */
-    caffeMode_ = conf.caffe_mode();
-
-    // create a new weight
-    size_t height, width;
-    height = filterPixels_[index] * filterChannels_[index];
-    width = (!isDeconv_) ? numFilters_ : channels_[index];
-    CHECK_EQ(parameters_[index]->getSize(), width * height);
-    Weight *w = new Weight(height, width, parameters_[index]);
-    weights_.emplace_back(w);
-    index++;
-  }
-  if (biasParameter_.get()) {
-    if (sharedBiases_) {
-      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(numFilters_, 1, biasParameter_));
-    } else {
-      biases_ =
-          std::unique_ptr<Weight>(new Weight(getSize(), 1, biasParameter_));
-    }
-  }
-  getOutputSize();
-
-  return true;
-}
-
-size_t ExpandConvBaseLayer::getOutputSize() {
-  CHECK_NE(inputLayers_.size(), 0UL);
-  size_t layerSize = ConvBaseLayer::calOutputSize();
-  return layerSize;
-}
-
-void ExpandConvBaseLayer::addSharedBias() {
-  size_t mapW = getOutputSize() / numFilters_;
-  size_t mapH = getOutputValue()->getElementCnt() / mapW;
-  MatrixPtr out =
-      Matrix::create(getOutputValue()->getData(), mapH, mapW, false, useGpu_);
-
-  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
-
-  out->transpose(transOutValue_, false);  // false means no memory allocation
-  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
-                          numFilters_);
-
-  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
-                                  1,
-                                  biases_->getW()->getElementCnt(),
-                                  false,
-                                  useGpu_);
-  transOutValue_->addBias(*bias, 1.0f);
-
-  transOutValue_->reshape(mapW, mapH);
-  transOutValue_->transpose(out, false);  // false means no memory allocation
-
-  out->clear();
-  bias->clear();
-}
-
-void ExpandConvBaseLayer::addUnsharedBias() {
-  MatrixPtr outValue = getOutputValue();
-  MatrixPtr bias = Matrix::create(biases_->getW()->getData(),
-                                  1,
-                                  biases_->getW()->getElementCnt(),
-                                  false,
-                                  useGpu_);
-  outValue->addBias(*bias, 1.0f);
-}
-
-void ExpandConvBaseLayer::bpropSharedBias(MatrixPtr biases, MatrixPtr v) {
-  size_t mapW = getOutputSize() / numFilters_;
-  size_t mapH = v->getElementCnt() / mapW;
-  MatrixPtr vTmp = Matrix::create(v->getData(), mapH, mapW, false, useGpu_);
-
-  Matrix::resizeOrCreate(transOutValue_, mapW, mapH, false, useGpu_);
-
-  vTmp->transpose(transOutValue_, false);  // false means no memory allocation
-  transOutValue_->reshape(transOutValue_->getElementCnt() / numFilters_,
-                          numFilters_);
-  biases->collectBias(*transOutValue_, 1.0f);
-}
-
-void ExpandConvBaseLayer::bpropBiases(MatrixPtr v) {
-  MatrixPtr biases = Matrix::create(biases_->getWGrad()->getData(),
-                                    1,
-                                    biases_->getWGrad()->getElementCnt(),
-                                    false,
-                                    useGpu_);
-  if (sharedBiases_) {
-    bpropSharedBias(biases, v);
-  } else {
-    biases->collectBias(*v, 1.0f);
-  }
-  biases->clear();
-}
-
-}  // namespace paddle
diff --git a/paddle/gserver/layers/ExpandConvBaseLayer.h b/paddle/gserver/layers/ExpandConvBaseLayer.h
deleted file mode 100644
index 01c699d2344443a1887ec0b5005125f617cbe279..0000000000000000000000000000000000000000
--- a/paddle/gserver/layers/ExpandConvBaseLayer.h
+++ /dev/null
@@ -1,57 +0,0 @@
-/* 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 <vector>
-#include "ConvBaseLayer.h"
-#include "paddle/math/Matrix.h"
-
-namespace paddle {
-
-/**
- * @brief A subclass of ConvBaseLayer that is a superclass of both
- * ExpandConvLayer and ExpandConvTransLayer
- */
-class ExpandConvBaseLayer : public ConvBaseLayer {
-protected:
-  /// The transpose of output, which is an auxiliary matrix.
-  MatrixPtr transOutValue_;
-
-public:
-  explicit ExpandConvBaseLayer(const LayerConfig& config)
-      : ConvBaseLayer(config) {}
-
-  ~ExpandConvBaseLayer() {}
-
-  bool init(const LayerMap& layerMap,
-            const ParameterMap& parameterMap) override;
-
-  size_t getOutputSize();
-
-  /**
-   * Add shared bias.
-   */
-  void addSharedBias();
-
-  /**
-   * Add unshared bias.
-   */
-  void addUnsharedBias();
-
-  void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
-  void bpropBiases(MatrixPtr v);
-};
-
-}  // namespace paddle
diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp
index 20de475fc3f6b6f3c05ac26bea8363daff0cf110..48dfcb49a4c2c46891bb5236fc1f8e644c03f327 100644
--- a/paddle/gserver/layers/ExpandConvLayer.cpp
+++ b/paddle/gserver/layers/ExpandConvLayer.cpp
@@ -36,7 +36,36 @@ inline bool isDepthwiseConv(int channels, int groups) {
 bool ExpandConvLayer::init(const LayerMap &layerMap,
                            const ParameterMap &parameterMap) {
   /* Initialize the basic convolutional parent class */
-  ExpandConvBaseLayer::init(layerMap, parameterMap);
+  ConvBaseLayer::init(layerMap, parameterMap);
+
+  int index = 0;
+  for (auto &inputConfig : config_.inputs()) {
+    const ConvConfig &conf = inputConfig.conv_conf();
+    /* Consistent caffe mode for multiple input */
+    caffeMode_ = conf.caffe_mode();
+
+    // create a new weight
+    size_t height, width;
+    height = filterPixels_[index] * filterChannels_[index];
+    width = (!isDeconv_) ? numFilters_ : channels_[index];
+    CHECK_EQ(parameters_[index]->getSize(), width * height);
+    Weight *w = new Weight(height, width, parameters_[index]);
+    weights_.emplace_back(w);
+    index++;
+  }
+
+  if (biasParameter_.get()) {
+    if (sharedBiases_) {
+      CHECK_EQ((size_t)numFilters_, biasParameter_->getSize());
+      biases_ = std::unique_ptr<Weight>(
+          new Weight(1, numFilters_, biasParameter_, 0));
+    } else {
+      biases_ =
+          std::unique_ptr<Weight>(new Weight(1, getSize(), biasParameter_, 0));
+    }
+  }
+
+  getOutputSize();
 
   size_t numInputs = config_.inputs_size();
   inputShape_.resize(numInputs);
@@ -108,6 +137,12 @@ bool ExpandConvLayer::init(const LayerMap &layerMap,
   return true;
 }
 
+size_t ExpandConvLayer::getOutputSize() {
+  CHECK_NE(inputLayers_.size(), 0UL);
+  size_t layerSize = ConvBaseLayer::calOutputSize();
+  return layerSize;
+}
+
 // i is the index of input layers
 #define BACKWARD_INPUT(i, inputs, outputs) \
   backward_[2 * i]->calc(inputs, outputs)
@@ -155,11 +190,7 @@ void ExpandConvLayer::forward(PassType passType) {
 
   /* add the bias-vector */
   if (biases_.get()) {
-    if (sharedBiases_) {
-      addSharedBias();
-    } else {
-      addUnsharedBias();
-    }
+    output_.value->addBias(*biases_->getW(), 1.0, sharedBiases_);
   }
 
   /* activation */
@@ -171,7 +202,7 @@ void ExpandConvLayer::backward(const UpdateCallback &callback) {
 
   MatrixPtr outGrad = getOutputGrad();
   if (biases_ && biases_->getWGrad()) {
-    bpropBiases(outGrad);
+    biases_->getWGrad()->collectBias(*getOutputGrad(), 1, sharedBiases_);
     /* Increasing the number of gradient */
     biases_->getParameterPtr()->incUpdate(callback);
   }
diff --git a/paddle/gserver/layers/ExpandConvLayer.h b/paddle/gserver/layers/ExpandConvLayer.h
index a1f943d1521547af0f82cec7da8a4efe9037cd71..a0873de19253f2496bc0c2fba550b3199dfc7486 100644
--- a/paddle/gserver/layers/ExpandConvLayer.h
+++ b/paddle/gserver/layers/ExpandConvLayer.h
@@ -15,7 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <vector>
-#include "ExpandConvBaseLayer.h"
+#include "ConvBaseLayer.h"
 #include "paddle/math/Matrix.h"
 
 namespace paddle {
@@ -28,10 +28,9 @@ namespace paddle {
  * The config file api is img_conv_layer.
  */
 
-class ExpandConvLayer : public ExpandConvBaseLayer {
+class ExpandConvLayer : public ConvBaseLayer {
 public:
-  explicit ExpandConvLayer(const LayerConfig& config)
-      : ExpandConvBaseLayer(config) {}
+  explicit ExpandConvLayer(const LayerConfig& config) : ConvBaseLayer(config) {}
 
   ~ExpandConvLayer() {}
 
@@ -41,6 +40,8 @@ public:
   void forward(PassType passType) override;
   void backward(const UpdateCallback& callback) override;
 
+  size_t getOutputSize();
+
 protected:
   std::vector<TensorShape> inputShape_;
   std::vector<TensorShape> filterShape_;
diff --git a/paddle/gserver/layers/MKLDNNConvLayer.cpp b/paddle/gserver/layers/MKLDNNConvLayer.cpp
index f8c06c5f868f8d48a9a222b92315ee0ef2cf265e..2647cb600653b4f43322016afb231a55f4db5642 100644
--- a/paddle/gserver/layers/MKLDNNConvLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNConvLayer.cpp
@@ -285,22 +285,18 @@ void MKLDNNConvLayer::resetWgtBiasValue(
   wgt = MKLDNNMatrix::create(weight_->getW(), pd->weights_primitive_desc());
   VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
 
-  bias = nullptr;
-  if (biases_ && biases_->getW()) {
-    bias = MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc());
-  }
+  bias = (biases_ && biases_->getW())
+             ? MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc())
+             : nullptr;
 }
 
 void MKLDNNConvLayer::resetOutValue(
     std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& out) {
   out = MKLDNNMatrix::create(output_.value, pd->dst_primitive_desc());
 
-  // change original output value from cpu matrix to mkldnn matrix
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
-
   // create reorder if output value has cpu device and pd do not match
   cpuOutVal_ = nullptr;
-  cpuOutVal_ = nullptr;
+  cvtOutVal_ = nullptr;
   if (!outputIsOnlyMKLDNN()) {
     const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
     memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
@@ -356,6 +352,7 @@ void MKLDNNConvLayer::resetBwdWgtPD(
 
 void MKLDNNConvLayer::resetBwdDataPD(
     std::shared_ptr<conv_bwdData::primitive_desc>& pd) {
+  pd = nullptr;
   if (inputLayers_[0]->getOutput().grad == nullptr) {
     return;
   }
@@ -476,6 +473,7 @@ void MKLDNNConvLayer::resetWgtBiasGrad(
       << "primitive desc of weight grad and value should be equal";
   VLOG(MKLDNN_FMTS) << "weight grad format: " << wgt->getFormat();
 
+  bias = nullptr;
   if (biasVal_ == nullptr) {
     return;
   }
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.cpp b/paddle/gserver/layers/MKLDNNFcLayer.cpp
index f70343251ad4fbb99f9614618f6d1bff1174f15e..66b358bcea53f61ddcc15323704fa9f154fb2a73 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp
@@ -17,9 +17,6 @@ limitations under the License. */
 
 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 {
 
@@ -93,82 +90,146 @@ void MKLDNNFcLayer::reshape(
   printSizeInfo();
 }
 
-void MKLDNNFcLayer::resetFwd(std::vector<mkldnn::primitive>& pipeline,
+void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
                              MKLDNNMatrixPtr& in,
                              MKLDNNMatrixPtr& wgt,
                              MKLDNNMatrixPtr& bias,
                              MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-  bool hasBias = biases_ && biases_->getW();
-  const MatrixPtr& wgtVal = weight_->getW();
-  const MatrixPtr& biasVal = hasBias ? biases_->getW() : nullptr;
-  const MatrixPtr& outVal = output_.value;
+  resetFwdBuffers(in, wgt, bias, out);
+
+  resetFwdPD(fwdPD_, in, wgt, bias, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
+                             MKLDNNMatrixPtr& in,
+                             MKLDNNMatrixPtr& wgt,
+                             MKLDNNMatrixPtr& bias,
+                             MKLDNNMatrixPtr& out) {
+  std::shared_ptr<fc_bwdWgt::primitive_desc> bwdWgtPD;
+  std::shared_ptr<fc_bwdData::primitive_desc> bwdDataPD;
+
+  resetBwdBuffers(in, wgt, bias, out);
+
+  resetBwdWgtPD(bwdWgtPD, wgt, bias, out);
+
+  resetBwdDataPD(bwdDataPD, in, out);
+
+  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNFcLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
 
+void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
+  weight_->getParameterPtr()->incUpdate(callback);
+  if (biases_ && biases_->getWGrad()) {
+    biases_->getParameterPtr()->incUpdate(callback);
+  }
+}
+
+void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
+                                    MKLDNNMatrixPtr& wgt,
+                                    MKLDNNMatrixPtr& bias,
+                                    MKLDNNMatrixPtr& out) {
+  resetInValue(in);
+
+  resetWgtBiasValue(wgt, bias);
+
+  resetOutValue(out);
+}
+
+void MKLDNNFcLayer::resetInValue(MKLDNNMatrixPtr& in) {
   if (inputIsOnlyMKLDNN()) {
-    const MatrixPtr& inVal = getInputValue(0);
-    in = std::dynamic_pointer_cast<MKLDNNMatrix>(inVal);
+    const MatrixPtr& dnnIn = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
     CHECK(in) << "Input should be MKLDNNMatrix";
   } else {
     CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
-    const MatrixPtr& inVal = getInputValue(0, CPU_DEVICE);
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
     in = MKLDNNMatrix::create(
-        inVal, memory::dims{bs_, ic_, ih_, iw_}, format::nchw, engine_);
+        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
   }
   in->downSpatial();
+}
+
+void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
+                                      MKLDNNMatrixPtr& bias) {
   wgt = MKLDNNMatrix::create(
-      wgtVal, memory::dims{oc_, ic_, ih_, iw_}, format::oihw, engine_);
+      weight_->getW(), {oc_, ic_, ih_, iw_}, format::oihw, engine_);
   wgt->downSpatial();
-  bias = hasBias ? MKLDNNMatrix::create(biasVal, {oc_}, format::x, engine_)
-                 : nullptr;
-  out = MKLDNNMatrix::create(outVal, {bs_, oc_}, format::nc, engine_);
 
-  // change original output value to mkldnn output value
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
+  bias = (biases_ && biases_->getW())
+             ? MKLDNNMatrix::create(biases_->getW(), {oc_}, format::x, engine_)
+             : nullptr;
+}
+
+void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
+  out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
   if (!outputIsOnlyMKLDNN()) {
     // fc cpu output value do not need create convert
     // just share point
-    getOutput(CPU_DEVICE).value->setData(output_.value->getData());
+    getOutput(CPU_DEVICE).value->setData(out->getData());
   }
+}
 
-  // create forward handle
+void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                               MKLDNNMatrixPtr in,
+                               MKLDNNMatrixPtr wgt,
+                               MKLDNNMatrixPtr bias,
+                               MKLDNNMatrixPtr out) {
+  CHECK(in);
+  CHECK(wgt);
+  CHECK(out);
   prop_kind pk = prop_kind::forward;
-  fc_fwd::desc fwdDesc = hasBias ? fc_fwd::desc(pk,
-                                                in->getMemoryDesc(),
-                                                wgt->getMemoryDesc(),
-                                                bias->getMemoryDesc(),
-                                                out->getMemoryDesc())
-                                 : fc_fwd::desc(pk,
-                                                in->getMemoryDesc(),
-                                                wgt->getMemoryDesc(),
-                                                out->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  if (hasBias) {
-    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *bias, *out));
+  fc_fwd::desc fwdDesc = bias != nullptr ? fc_fwd::desc(pk,
+                                                        in->getMemoryDesc(),
+                                                        wgt->getMemoryDesc(),
+                                                        bias->getMemoryDesc(),
+                                                        out->getMemoryDesc())
+                                         : fc_fwd::desc(pk,
+                                                        in->getMemoryDesc(),
+                                                        wgt->getMemoryDesc(),
+                                                        out->getMemoryDesc());
+  pd.reset(new fc_fwd::primitive_desc(fwdDesc, engine_));
+}
+
+void MKLDNNFcLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<fc_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+
+  if (bias) {
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *bias, *out));
   } else {
-    fwd_.reset(new fc_fwd(fwdPD, *in, *wgt, *out));
+    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
   }
-  printValueFormatFlow();
 
   pipeline.push_back(*fwd_);
 }
 
-void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
-                             MKLDNNMatrixPtr& in,
-                             MKLDNNMatrixPtr& wgt,
-                             MKLDNNMatrixPtr& bias,
-                             MKLDNNMatrixPtr& out) {
-  pipeline.clear();
-  if (!needResetBwd_) {
-    return;
-  }
-  needResetBwd_ = false;
-  bool hasBias = biases_ && biases_->getWGrad();
+void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
+                                    MKLDNNMatrixPtr& wgt,
+                                    MKLDNNMatrixPtr& bias,
+                                    MKLDNNMatrixPtr& out) {
+  resetOutGrad(out);
+
+  resetWgtBiasGrad(wgt, bias);
 
-  /// backward weight
-  CHECK(inVal_) << "Should have input value";
-  const MatrixPtr& wgtGrad = weight_->getWGrad();
-  const MatrixPtr& biasGrad = hasBias ? biases_->getWGrad() : nullptr;
+  resetInGrad(in);
+}
 
+void MKLDNNFcLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
   // TODO(TJ): merge outgrad
   int device = outputIsOnlyMKLDNN() ? MKLDNN_DEVICE : CPU_DEVICE;
   // for MKLDNN device:
@@ -178,66 +239,88 @@ void MKLDNNFcLayer::resetBwd(std::vector<mkldnn::primitive>& pipeline,
   // for CPU device:
   // fc do not need to convert from cpu device since output is always nc format
   // only need create from cpu device
-  const MatrixPtr& outGrad = getOutput(device).grad;
-  out = MKLDNNMatrix::create(outGrad, outVal_->getPrimitiveDesc());
-  wgt = MKLDNNMatrix::create(wgtGrad, wgtVal_->getPrimitiveDesc());
-  bias = hasBias ? MKLDNNMatrix::create(biasGrad, biasVal_->getPrimitiveDesc())
-                 : nullptr;
-
-  // create memory primitive desc
-  fc_fwd::desc fwdDesc = fc_fwd::desc(prop_kind::forward,
-                                      inVal_->getMemoryDesc(),
-                                      wgt->getMemoryDesc(),
-                                      out->getMemoryDesc());
-  fc_fwd::primitive_desc fwdPD = fc_fwd::primitive_desc(fwdDesc, engine_);
-  fc_bwdWgt::desc bwdWgtDesc = hasBias
-                                   ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgt->getMemoryDesc(),
-                                                     bias->getMemoryDesc(),
-                                                     out->getMemoryDesc())
-                                   : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
-                                                     wgt->getMemoryDesc(),
-                                                     out->getMemoryDesc());
-  fc_bwdWgt::primitive_desc bwdWgtPD =
-      fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, fwdPD);
-
-  if (hasBias) {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt, *bias));
-  } else {
-    bwdWgt_.reset(new fc_bwdWgt(bwdWgtPD, *inVal_, *out, *wgt));
+  CHECK(outVal_);
+  out =
+      MKLDNNMatrix::create(getOutput(device).grad, outVal_->getPrimitiveDesc());
+}
+
+void MKLDNNFcLayer::resetWgtBiasGrad(MKLDNNMatrixPtr& wgt,
+                                     MKLDNNMatrixPtr& bias) {
+  CHECK(wgtVal_);
+  wgt = MKLDNNMatrix::create(weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
+
+  bias = nullptr;
+  if (biasVal_ == nullptr) {
+    return;
   }
-  pipeline.push_back(*bwdWgt_);
+  bias =
+      MKLDNNMatrix::create(biases_->getWGrad(), biasVal_->getPrimitiveDesc());
+}
 
-  /// backward data
+void MKLDNNFcLayer::resetInGrad(MKLDNNMatrixPtr& in) {
+  in = nullptr;
   const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
   if (inGrad == nullptr) {
     return;
   }
-  if (getInput(0, MKLDNN_DEVICE).getAllCount() > 1) {
-    // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
-  } else {
-    in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
-  }
-
-  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
-      inVal_->getMemoryDesc(), wgt->getMemoryDesc(), out->getMemoryDesc());
-  fc_bwdData::primitive_desc bwdDataPD =
-      fc_bwdData::primitive_desc(bwdDataDesc, engine_, fwdPD);
+  // TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
+  CHECK(inVal_);
+  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+}
 
-  CHECK(wgtVal_) << "Should have weight memory";
-  bwdData_.reset(new fc_bwdData(bwdDataPD, *out, *wgtVal_, *in));
-  printGradFormatFlow();
-  pipeline.push_back(*bwdData_);
+void MKLDNNFcLayer::resetBwdWgtPD(
+    std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  CHECK(inVal_);
+  fc_bwdWgt::desc bwdWgtDesc = bias ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                      wgt->getMemoryDesc(),
+                                                      bias->getMemoryDesc(),
+                                                      out->getMemoryDesc())
+                                    : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
+                                                      wgt->getMemoryDesc(),
+                                                      out->getMemoryDesc());
+  pd.reset(new fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
 }
 
-void MKLDNNFcLayer::updateInputData() {
-  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+void MKLDNNFcLayer::resetBwdDataPD(
+    std::shared_ptr<fc_bwdData::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pd = nullptr;
+  if (in == nullptr) {
+    return;
+  }
+  CHECK(wgtVal_);
+  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
+      in->getMemoryDesc(), wgtVal_->getMemoryDesc(), out->getMemoryDesc());
+  pd.reset(new fc_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
 }
 
-void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
-  weight_->getParameterPtr()->incUpdate(callback);
-  if (biases_ && biases_->getWGrad()) {
-    biases_->getParameterPtr()->incUpdate(callback);
+void MKLDNNFcLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
+    std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& wgt,
+    MKLDNNMatrixPtr& bias,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  CHECK(inVal_);
+  if (bias) {
+    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt, *bias));
+  } else {
+    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt));
+  }
+  pipeline.push_back(*bwdWgt_);
+
+  if (bwdDataPD == nullptr) {
+    return;
   }
+  CHECK(wgtVal_) << "Should have weight memory";
+  bwdData_.reset(new fc_bwdData(*bwdDataPD, *out, *wgtVal_, *in));
+  pipeline.push_back(*bwdData_);
 }
+
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNFcLayer.h b/paddle/gserver/layers/MKLDNNFcLayer.h
index 3119f863496df092da13c08bf733f13c42e53780..c76878aafab7e986d2bf478eaba02f2f0aced293 100644
--- a/paddle/gserver/layers/MKLDNNFcLayer.h
+++ b/paddle/gserver/layers/MKLDNNFcLayer.h
@@ -18,6 +18,9 @@ limitations under the License. */
 #include "mkldnn.hpp"
 
 namespace paddle {
+typedef mkldnn::inner_product_forward fc_fwd;
+typedef mkldnn::inner_product_backward_weights fc_bwdWgt;
+typedef mkldnn::inner_product_backward_data fc_bwdData;
 
 /**
  * @brief A subclass of MKLDNNLayer fc layer.
@@ -32,6 +35,9 @@ protected:
   // if has already init the weight
   bool hasInitedWgt_;
 
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<fc_fwd::primitive_desc> fwdPD_;
+
   // fc weight and bias
   std::unique_ptr<Weight> weight_;
   std::unique_ptr<Weight> biases_;
@@ -67,6 +73,59 @@ public:
   void convertWeightsFromPaddle() override;
 
   void convertWeightsToPaddle() override;
+
+protected:
+  /**
+   * Forward functions: reset buffers(input, output, weight and bias),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  void resetFwdBuffers(MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetInValue(MKLDNNMatrixPtr& in);
+  void resetWgtBiasValue(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
+  void resetOutValue(MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr in,
+                  MKLDNNMatrixPtr wgt,
+                  MKLDNNMatrixPtr bias,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<fc_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * Backward functions: reset buffers(input, output, weight and bias),
+   *                     reset primitive descriptor for backward weight,
+   *                     reset primitive descriptor for backward data,
+   *                     reset pipeline.
+   */
+  void resetBwdBuffers(MKLDNNMatrixPtr& in,
+                       MKLDNNMatrixPtr& wgt,
+                       MKLDNNMatrixPtr& bias,
+                       MKLDNNMatrixPtr& out);
+  void resetOutGrad(MKLDNNMatrixPtr& out);
+  void resetWgtBiasGrad(MKLDNNMatrixPtr& wgt, MKLDNNMatrixPtr& bias);
+  void resetInGrad(MKLDNNMatrixPtr& in);
+  void resetBwdWgtPD(std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
+                     MKLDNNMatrixPtr& wgt,
+                     MKLDNNMatrixPtr& bias,
+                     MKLDNNMatrixPtr& out);
+  void resetBwdDataPD(std::shared_ptr<fc_bwdData::primitive_desc>& pd,
+                      MKLDNNMatrixPtr& in,
+                      MKLDNNMatrixPtr& out);
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
+                        std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& wgt,
+                        MKLDNNMatrixPtr& bias,
+                        MKLDNNMatrixPtr& out);
 };
 
 }  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNLayer.h b/paddle/gserver/layers/MKLDNNLayer.h
index 169679c8297542cac4a43f5a8e1af311ad9282df..c4e4a6874e6fdb491c344c70dfea422dc0924cd9 100644
--- a/paddle/gserver/layers/MKLDNNLayer.h
+++ b/paddle/gserver/layers/MKLDNNLayer.h
@@ -119,6 +119,10 @@ public:
         inputElemenCnt_ = elemenCnt;
         reshape(bs_, ic_, ih_, iw_, oc_, oh_, ow_);
         resetFwd(pipelineFwd_, inVal_, wgtVal_, biasVal_, outVal_);
+        if (outVal_) {
+          // change original output value to mkldnn output value
+          output_.value = std::dynamic_pointer_cast<Matrix>(outVal_);
+        }
         convertWeightsFromPaddle();
         needResetBwd_ = true;
       }
diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.cpp b/paddle/gserver/layers/MKLDNNPoolLayer.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..b62dfb7c54258a593aa50d5b30096423f375c69d
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNPoolLayer.cpp
@@ -0,0 +1,276 @@
+/* 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 "MKLDNNPoolLayer.h"
+#include "paddle/math/MathUtils.h"
+#include "paddle/utils/Logging.h"
+
+using namespace mkldnn;  // NOLINT
+typedef memory::format format;
+
+namespace paddle {
+
+REGISTER_LAYER(mkldnn_pool, MKLDNNPoolLayer);
+
+bool MKLDNNPoolLayer::init(const LayerMap& layerMap,
+                           const ParameterMap& parameterMap) {
+  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
+    return false;
+  }
+
+  /* the size of inputs for pool-layer is 1 */
+  CHECK_EQ(config_.inputs_size(), 1);
+  const PoolConfig& conf = config_.inputs(0).pool_conf();
+  ic_ = conf.channels();
+  ih_ = conf.img_size_y();
+  iw_ = conf.img_size();
+  oc_ = ic_;
+  oh_ = conf.output_y();
+  ow_ = conf.output_x();
+  fh_ = conf.size_y();
+  fw_ = conf.size_x();
+  ph_ = conf.padding_y();
+  pw_ = conf.padding();
+  sh_ = conf.stride_y();
+  sw_ = conf.stride();
+
+  const std::string& type = conf.pool_type();
+  if (type == "max-projection") {
+    poolAlgo_ = algorithm::pooling_max;
+  } else if (type == "avg-projection") {
+    // paddle only use exclude_padding
+    poolAlgo_ = algorithm::pooling_avg_exclude_padding;
+  } else {
+    LOG(FATAL) << "unknow pooling type!";
+  }
+  return true;
+}
+
+void MKLDNNPoolLayer::reshape(
+    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
+  reshapeInput(bs, ih, iw);
+  // ic_ and oc can not be changed
+  CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
+      << "Input channel can not be changed";
+
+  // cal output sizes
+  // paddle used false caffeMode for pooling
+  oh = outputSize(ih, fh_, ph_, sh_, false);
+  ow = outputSize(iw, fw_, pw_, sw_, false);
+  reshapeOutput(oh, ow);
+
+  resizeOutput(bs, oc * oh * ow);
+
+  printSizeInfo();
+}
+
+void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  resetFwdBuffers(in, out);
+
+  resetFwdPD(fwdPD_, in, out);
+
+  resetFwdPipeline(pipeline, fwdPD_, in, out);
+
+  printValueFormatFlow();
+}
+
+void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
+                               MKLDNNMatrixPtr& in,
+                               MKLDNNMatrixPtr& wgt,
+                               MKLDNNMatrixPtr& bias,
+                               MKLDNNMatrixPtr& out) {
+  std::shared_ptr<pool_bwd::primitive_desc> pd;
+
+  resetBwdBuffers(in, out);
+
+  resetBwdPD(pd, in, out);
+
+  resetBwdPipeline(pipeline, pd, in, out);
+
+  printGradFormatFlow();
+}
+
+void MKLDNNPoolLayer::updateInputData() {
+  inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
+}
+
+void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
+                                      MKLDNNMatrixPtr& out) {
+  resetInValue(in);
+
+  resetOutValue(out);
+}
+
+void MKLDNNPoolLayer::resetInValue(MKLDNNMatrixPtr& in) {
+  if (inputIsOnlyMKLDNN()) {
+    const MatrixPtr& dnnIn = getInputValue(0);
+    in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
+    CHECK(in) << "Input should be MKLDNNMatrix";
+  } else {
+    CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
+    const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
+    in = MKLDNNMatrix::create(
+        cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
+  }
+}
+
+void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
+  CHECK(inVal_) << "Should reset input value first";
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  out = MKLDNNMatrix::create(
+      output_.value, outDims, inVal_->getFormat(), engine_);
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutVal_ = nullptr;
+  cvtOutVal_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
+    cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
+    if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
+      CHECK(cvtOutVal_) << "should not be emptry";
+    } else {
+      // CPU output share the same data of MKLDNN output
+      cpuOut->setData(out->getData());
+      cpuOutVal_ = out;
+    }
+  }
+}
+
+void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                                 MKLDNNMatrixPtr in,
+                                 MKLDNNMatrixPtr out) {
+  memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
+  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
+  memory::dims kernels = memory::dims{fh_, fw_};
+  memory::dims strides = memory::dims{sh_, sw_};
+  memory::dims padL = memory::dims{ph_, pw_};
+  memory::dims padR = getPaddingR();
+  padding_kind padKind = padding_kind::zero;
+  prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
+                                        : prop_kind::forward_training;
+  auto fwdDesc = pool_fwd::desc(pk,
+                                poolAlgo_,
+                                in->getMemoryDesc(),
+                                out->getMemoryDesc(),
+                                strides,
+                                kernels,
+                                padL,
+                                padR,
+                                padKind);
+  pd.reset(new pool_fwd::primitive_desc(fwdDesc, engine_));
+
+  // prepare workspace if necessary
+  workspace_ =
+      (passType_ != PASS_TEST && poolAlgo_ == algorithm::pooling_max)
+          ? std::make_shared<memory>(memory(pd->workspace_primitive_desc()))
+          : nullptr;
+}
+
+void MKLDNNPoolLayer::resetFwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<pool_fwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  fwd_ = workspace_
+             ? std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out, *workspace_))
+             : std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out));
+  pipeline.push_back(*fwd_);
+
+  if (cvtOutVal_) {
+    pipeline.push_back(*cvtOutVal_);
+  }
+}
+
+void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
+                                      MKLDNNMatrixPtr& out) {
+  resetOutGrad(out);
+
+  resetInGrad(in);
+}
+void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
+  CHECK(outVal_) << "Should have output value";
+  out = MKLDNNMatrix::create(output_.grad, outVal_->getPrimitiveDesc());
+
+  // create reorder if output value has cpu device and pd do not match
+  cpuOutGrad_ = nullptr;
+  cvtOutGrad_ = nullptr;
+  if (!outputIsOnlyMKLDNN()) {
+    const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
+    cpuOutGrad_ = MKLDNNMatrix::create(
+        cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
+    if (cpuOutGrad_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
+      cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
+      CHECK(cvtOutGrad_) << "should not be emptry";
+    } else {
+      // share the same data of CPU output
+      output_.grad->setData(cpuOut->getData());
+      out = cpuOutGrad_;
+    }
+  }
+}
+
+void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
+  in = nullptr;
+  const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
+  if (inGrad == nullptr) {
+    return;
+  }
+  CHECK(inVal_);
+  in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
+}
+
+void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                                 MKLDNNMatrixPtr& in,
+                                 MKLDNNMatrixPtr& out) {
+  memory::dims kernels = memory::dims{fh_, fw_};
+  memory::dims strides = memory::dims{sh_, sw_};
+  memory::dims padL = memory::dims{ph_, pw_};
+  memory::dims padR = getPaddingR();
+  CHECK(in);
+  CHECK(out);
+  auto bwdDesc = pool_bwd::desc(poolAlgo_,
+                                in->getMemoryDesc(),
+                                out->getMemoryDesc(),
+                                strides,
+                                kernels,
+                                padL,
+                                padR,
+                                padding_kind::zero);
+  pd.reset(new pool_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_));
+}
+
+void MKLDNNPoolLayer::resetBwdPipeline(
+    std::vector<primitive>& pipeline,
+    std::shared_ptr<pool_bwd::primitive_desc>& pd,
+    MKLDNNMatrixPtr& in,
+    MKLDNNMatrixPtr& out) {
+  pipeline.clear();
+  if (cvtOutGrad_) {
+    pipeline.push_back(*cvtOutGrad_);
+  }
+
+  bwdData_ =
+      workspace_
+          ? std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *workspace_, *in))
+          : std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *in));
+  pipeline.push_back(*bwdData_);
+}
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/MKLDNNPoolLayer.h b/paddle/gserver/layers/MKLDNNPoolLayer.h
new file mode 100644
index 0000000000000000000000000000000000000000..891e15a7efcdd2e54f61352efc1ba7345b91c76b
--- /dev/null
+++ b/paddle/gserver/layers/MKLDNNPoolLayer.h
@@ -0,0 +1,138 @@
+/* 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 {
+typedef mkldnn::pooling_forward pool_fwd;
+typedef mkldnn::pooling_backward pool_bwd;
+
+/**
+ * @brief A subclass of MKLDNNLayer pool layer.
+ *
+ * The config file api is mkldnn_pool
+ */
+class MKLDNNPoolLayer : public MKLDNNLayer {
+protected:
+  // padding height and width
+  int ph_, pw_;
+  // stride height and width
+  int sh_, sw_;
+  // filter(kenerl) height and width
+  int fh_, fw_;
+
+  // pooling_avg or pooling_max
+  mkldnn::algorithm poolAlgo_;
+
+  // MKLDNNMatrixPtr which should be created from CPU Device
+  MKLDNNMatrixPtr cpuOutVal_;
+  MKLDNNMatrixPtr cpuOutGrad_;
+  // convert handle between CPU device and MKLDNN device
+  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
+  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
+
+  // save forward primitive_desc, which can be used backward
+  std::shared_ptr<pool_fwd::primitive_desc> fwdPD_;
+  // according to https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+  // test_pooling_forward.cpp, pool need workspace for backward
+  std::shared_ptr<mkldnn::memory> workspace_;
+
+public:
+  explicit MKLDNNPoolLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
+
+  ~MKLDNNPoolLayer() {}
+
+  bool init(const LayerMap& layerMap,
+            const ParameterMap& parameterMap) override;
+
+  void reshape(
+      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
+
+  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
+                MKLDNNMatrixPtr& in,
+                MKLDNNMatrixPtr& wgt,
+                MKLDNNMatrixPtr& bias,
+                MKLDNNMatrixPtr& out) override;
+
+  void updateInputData() override;
+
+  void printSizeInfo() override {
+    MKLDNNLayer::printSizeInfo();
+    VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
+                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", sw: " << sw_;
+  }
+
+protected:
+  /**
+   * Forward functions: reset buffers(input, output),
+   *                    reset primitive descriptor,
+   *                    reset pipeline.
+   */
+  void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
+  void resetInValue(MKLDNNMatrixPtr& in);
+  void resetOutValue(MKLDNNMatrixPtr& out);
+  void resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr in,
+                  MKLDNNMatrixPtr out);
+  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<pool_fwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * Backward functions: reset buffers(input, output),
+   *                     reset primitive descriptor,
+   *                     reset pipeline.
+   */
+  void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
+  void resetOutGrad(MKLDNNMatrixPtr& out);
+  void resetInGrad(MKLDNNMatrixPtr& in);
+  void resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                  MKLDNNMatrixPtr& in,
+                  MKLDNNMatrixPtr& out);
+  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
+                        std::shared_ptr<pool_bwd::primitive_desc>& pd,
+                        MKLDNNMatrixPtr& in,
+                        MKLDNNMatrixPtr& out);
+
+  /**
+   * get padding_r according to
+   * https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
+   * test_pooling_forward.cpp
+   */
+  mkldnn::memory::dims getPaddingR() const {
+    mkldnn::memory::dims padR = {ph_, pw_};
+    for (int i = 0; i < 2; ++i) {
+      if ((ih_ + ph_ + padR[0] - fh_) / sh_ + 1 < oh_) {
+        ++padR[0];
+      }
+      if ((iw_ + pw_ + padR[1] - fw_) / sw_ + 1 < ow_) {
+        ++padR[1];
+      }
+    }
+    return padR;
+  }
+};
+
+}  // namespace paddle
diff --git a/paddle/gserver/layers/SequenceSliceLayer.cpp b/paddle/gserver/layers/SequenceSliceLayer.cpp
index d3a83fad276a384ab3fddd5349912c56be6f3cc0..ce68ca449429711eeee692be750a4a2f1dac61a6 100644
--- a/paddle/gserver/layers/SequenceSliceLayer.cpp
+++ b/paddle/gserver/layers/SequenceSliceLayer.cpp
@@ -73,9 +73,10 @@ void SequenceSliceLayer::checkInputs() {
   CHECK(inputSeq.hasSeq()) << "The first input of sequence slice layer "
                            << "must be a sequence.";
   const MatrixPtr indices1 = getInputValue(1);
-  CHECK_EQ(static_cast<size_t>(indices1->getHeight()),
-           inputSeq.hasSubseq() ? inputSeq.getNumSubSequences()
-                                : inputSeq.getNumSequences())
+  CHECK_EQ(
+      indices1->getHeight(),
+      static_cast<size_t>(inputSeq.hasSubseq() ? inputSeq.getNumSubSequences()
+                                               : inputSeq.getNumSequences()))
       << "Height of the second input should be equal to number of sequence "
       << "in the first input.";
   if (inputLayers_.size() == 3) {
@@ -151,7 +152,7 @@ void SequenceSliceLayer::calSelectedRows(const MatrixPtr starts,
         if (ends) endPos = inputSeqInfoVec_[i][j] + ends->getElement(rowIdx, k);
 
         int seqLen = endPos - begPos + 1;
-        CHECK_GT(seqLen, 0U);
+        CHECK_GT(seqLen, 0);
         for (int m = begPos; m <= endPos; ++m) selectedRows_.push_back(m);
         hasSubseq
             ? outSubSeqStartPos_.push_back(outSubSeqStartPos_.back() + seqLen)
diff --git a/paddle/gserver/tests/MKLDNNTester.cpp b/paddle/gserver/tests/MKLDNNTester.cpp
index 2f48e5b2d3ffc9337ed1314f6db6549e56263fdd..f59618be9d09d146be52fb51cae84f4d24c15ef1 100644
--- a/paddle/gserver/tests/MKLDNNTester.cpp
+++ b/paddle/gserver/tests/MKLDNNTester.cpp
@@ -64,15 +64,17 @@ void MKLDNNTester::reset(const TestConfig& dnn,
         configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i]));
   }
   refLayer_ = testLayers_[REF];
-  dnnLayer_ = std::dynamic_pointer_cast<MKLDNNLayer>(testLayers_[DNN]);
-  CHECK(dnnLayer_);
-  // for comparison with Paddle reference results,
-  // need manually add cpu device output for test
-  dnnLayer_->addOutputArgument(CPU_DEVICE);
+  dnnLayer_ = testLayers_[DNN];
   EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size());
   EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
-
   setInputImgSize();
+
+  // for comparison with Paddle reference results,
+  // need manually add cpu device output for test
+  MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
+  if (dnnLayer) {
+    dnnLayer->addOutputArgument(CPU_DEVICE);
+  }
 }
 
 void MKLDNNTester::setInputImgSize() {
@@ -122,7 +124,7 @@ void MKLDNNTester::randomTopDiffs() {
 void MKLDNNTester::checkForward() {
   VLOG(MKLDNN_ALL) << "Check Forward";
   printTopDatas();
-  double delta = compareMatrix(dnnLayer_->getOutput(-1).value,
+  double delta = compareMatrix(dnnLayer_->getOutput(CPU_DEVICE).value,
                                refLayer_->getOutputValue());
   EXPECT_LE(fabs(delta), eps_);
 }
@@ -155,7 +157,10 @@ void MKLDNNTester::checkBackwardWgts() {
   vector<VectorPtr> dnnWgts;  // used to temply save mkldnn weights
   saveWgt(parameters_[DNN], dnnWgts);
 
-  dnnLayer_->convertWeightsToPaddle();
+  MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
+  if (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);
@@ -322,6 +327,10 @@ void MKLDNNTester::runOnce() {
   // and clearTopDatas(REF) should be coverd by ref layers
   clearBotDiffs(REF);
   clearWgtDiffs(REF);
+  // it is necessary to clear bottom diffs when only activation is dnn type
+  if (configs_[DNN].layerConfig.active_type().compare(0, 7, "mkldnn_") == 0) {
+    clearBotDiffs(DNN);
+  }
 }
 
 void MKLDNNTester::run(const TestConfig& dnn,
@@ -333,8 +342,19 @@ void MKLDNNTester::run(const TestConfig& dnn,
                        float epsilon,
                        bool log,
                        int level) {
-  VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " << dnn.layerConfig.type()
-                     << " vs " << ref.layerConfig.type();
+  CHECK(dnn.layerConfig.type().compare(0, 7, "mkldnn_") == 0 ||
+        dnn.layerConfig.active_type().compare(0, 7, "mkldnn_") == 0)
+      << "should be MKLDNN layer or MKLDNN activation";
+  if (dnn.layerConfig.type() == ref.layerConfig.type()) {
+    VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: "
+                       << dnn.layerConfig.active_type() << " vs "
+                       << ref.layerConfig.active_type();
+  } else {
+    VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: "
+                       << dnn.layerConfig.type() << " vs "
+                       << ref.layerConfig.type();
+  }
+
   ih_ = inputImgH;
   iw_ = inputImgW;
   iter_ = iter;
diff --git a/paddle/gserver/tests/MKLDNNTester.h b/paddle/gserver/tests/MKLDNNTester.h
index 5ac885638cde7693a0c847733e7a6149c1b7e6c2..171d176ee757f1164c38d86273bdf9e5aefeda06 100644
--- a/paddle/gserver/tests/MKLDNNTester.h
+++ b/paddle/gserver/tests/MKLDNNTester.h
@@ -41,8 +41,7 @@ protected:
   vector<LayerMap> layerMaps_;
   vector<vector<ParameterPtr>> parameters_;
   vector<LayerPtr> testLayers_;
-  LayerPtr refLayer_;
-  MKLDNNLayerPtr dnnLayer_;
+  LayerPtr refLayer_, dnnLayer_;
 
   /// run some iterations, all the result should pass
   size_t iter_;
diff --git a/paddle/gserver/tests/test_MKLDNN.cpp b/paddle/gserver/tests/test_MKLDNN.cpp
index e70802881e3f22160a87b7a4babda07ffbcf9d6f..406181370faf90d29167b62173ce4c8af44d243e 100644
--- a/paddle/gserver/tests/test_MKLDNN.cpp
+++ b/paddle/gserver/tests/test_MKLDNN.cpp
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <vector>
 #include "MKLDNNTester.h"
 #include "ModelConfig.pb.h"
+#include "paddle/gserver/activations/MKLDNNActivation.h"
 #include "paddle/math/MathUtils.h"
 
 using namespace paddle;  // NOLINT
@@ -141,6 +142,110 @@ TEST(MKLDNNLayer, ConvLayer) {
   testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1});
 }
 
+struct testPoolDesc {
+  int bs, ch;  // input channel and output channel are the same
+  int ih, iw;
+  int oh, ow;
+  int fh, fw;
+  int ph, pw;
+  int sh, sw;
+};
+
+void testPoolLayer(const testPoolDesc& pm) {
+  const std::string compareTypes[] = {"mkldnn_pool", "pool"};
+  TestConfig cfg;
+  cfg.layerConfig.set_type(compareTypes[0]);
+  cfg.layerConfig.set_size(pm.ch * pm.oh * pm.ow);
+  cfg.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       /* size of input layer= */ size_t(pm.ch * pm.ih * pm.iw),
+       0});
+  LayerInputConfig* input = cfg.layerConfig.add_inputs();
+  PoolConfig* pool = input->mutable_pool_conf();
+  pool->set_channels(pm.ch);
+  pool->set_img_size(pm.iw);
+  pool->set_img_size_y(pm.ih);
+  pool->set_output_x(pm.ow);
+  pool->set_output_y(pm.oh);
+  pool->set_size_x(pm.fw);
+  pool->set_size_y(pm.fh);
+  pool->set_padding(pm.pw);
+  pool->set_padding_y(pm.ph);
+  pool->set_stride(pm.sw);
+  pool->set_stride_y(pm.sh);
+
+  int oh = outputSize(pm.ih, pm.fh, pm.ph, pm.sh, false);
+  int ow = outputSize(pm.iw, pm.fw, pm.pw, pm.sw, false);
+  CHECK_EQ(ow, pm.ow) << "output size check failed";
+  CHECK_EQ(oh, pm.oh) << "output size check failed";
+
+  MKLDNNTester tester;
+  for (auto type : {"max-projection", "avg-projection"}) {
+    pool->set_pool_type(type);
+    TestConfig ref = cfg;
+    ref.layerConfig.set_type(compareTypes[1]);
+    for (auto bs : {pm.bs, 1}) {
+      tester.run(cfg, ref, bs, pm.ih, pm.iw);
+    }
+  }
+}
+
+TEST(MKLDNNLayer, PoolLayer) {
+  /* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw*/
+  testPoolLayer({2, 1, 4, 4, 2, 2, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({10, 8, 16, 16, 8, 8, 2, 2, 0, 0, 2, 2});
+  testPoolLayer({4, 2, 5, 5, 3, 3, 3, 3, 1, 1, 2, 2});
+  testPoolLayer({8, 16, 56, 56, 28, 28, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({8, 16, 14, 14, 7, 7, 3, 3, 0, 0, 2, 2});
+  testPoolLayer({4, 16, 7, 7, 1, 1, 7, 7, 0, 0, 1, 1});
+  testPoolLayer({4, 2, 5, 5, 3, 3, 5, 5, 1, 1, 1, 1});
+  testPoolLayer({2, 8, 56, 56, 29, 29, 3, 3, 1, 1, 2, 2});
+}
+
+struct testActDesc {
+  int bs, ch;
+  int ih, iw;
+};
+
+static void getAddtoConfig(TestConfig& cfg, const testActDesc& pm) {
+  cfg.biasSize = 0;
+  cfg.layerConfig.set_type("addto");
+  cfg.layerConfig.set_size(pm.ch * pm.ih * pm.iw);
+  cfg.inputDefs.push_back(
+      {INPUT_DATA,
+       "layer_0",
+       /* size of input layer= */ size_t(pm.ch * pm.ih * pm.iw),
+       0});
+  cfg.layerConfig.add_inputs();
+}
+
+void testActivation(std::string& type, const testActDesc& pm) {
+  const std::string compareTypes[] = {type, type.erase(0, 7)};
+  TestConfig cfg;
+  getAddtoConfig(cfg, pm);
+
+  TestConfig ref = cfg;
+  cfg.layerConfig.set_active_type(compareTypes[0]);
+  ref.layerConfig.set_active_type(compareTypes[1]);
+  MKLDNNTester tester;
+  for (auto bs : {pm.bs, 1}) {
+    tester.run(cfg, ref, bs, pm.ih, pm.iw);
+  }
+}
+
+TEST(MKLDNNActivation, Activations) {
+  auto types = MKLDNNActivation::getAllRegisteredTypes();
+  // TODO(TJ): mkldnn_softmax not implemented, paddle do not have elu activation
+  std::set<string> excluded{"mkldnn_softmax", "mkldnn_elu"};
+  for (auto type : types) {
+    if (excluded.count(type)) {
+      continue;
+    }
+    testActivation(type, {16, 64, 32, 32});
+  }
+}
+
 // TODO(TJ): add branch test
 
 int main(int argc, char** argv) {
diff --git a/paddle/math/BaseMatrix.cu b/paddle/math/BaseMatrix.cu
index 5435808fb7f70fdf1ac98815f7fe8890fb85527c..53dd5383601782231e6e742784007d1c9154dc6b 100644
--- a/paddle/math/BaseMatrix.cu
+++ b/paddle/math/BaseMatrix.cu
@@ -17,6 +17,7 @@ limitations under the License. */
 #include <cmath>
 #include "BaseMatrix.h"
 #include "MathFunctions.h"
+#include "NEONFunctions.h"
 #include "SIMDFunctions.h"
 #include "hl_matrix_apply.cuh"
 #include "hl_matrix_base.cuh"
@@ -666,6 +667,13 @@ void BaseMatrixT<T>::relu(BaseMatrixT& b) {
   applyBinary(binary::Relu<T>(), b);
 }
 
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+template <>
+void BaseMatrixT<float>::relu(BaseMatrixT& b) {
+  neon::relu(data_, b.data_, height_ * width_);
+}
+#endif
+
 DEFINE_MATRIX_BINARY_OP(ReluDerivative, a *= (b > 0.0f ? 1.0f : 0.0f));
 template <class T>
 void BaseMatrixT<T>::reluDerivative(BaseMatrixT& b) {
diff --git a/paddle/math/MKLDNNMatrix.h b/paddle/math/MKLDNNMatrix.h
index 0aa130b4a0d458ad78d5d1330164af9e73b22a44..c843115eb9a5be50d6ff873f1510844228c9d89f 100644
--- a/paddle/math/MKLDNNMatrix.h
+++ b/paddle/math/MKLDNNMatrix.h
@@ -66,11 +66,12 @@ public:
   /**
    * Create reorder primitive.
    * Create a mkldnn::reorder handle for converting src MKLDNNMatrix to dst.
-   * checkData: for whether to check the data handle of src and dst is the same.
-   *            if true, means check it and do not want support inplace reorder;
-   *            otherwise do not check data which means the created reorder
-   *            maybe inplace buffer and do not guarantee the logical is correct
-   *            since not all format or conversion support inplace.
+   * checkData: whether to check the data handle of src and dst.
+   *            if true, it will check the data and do not allow them equal;
+   *            otherwise, it will not check them, then the reorder created
+   *            may have inplace buffer.
+   *            Do not set false, if you can not guarantee the inplace logical
+   *            would work with your reorder.
    */
   static std::shared_ptr<mkldnn::reorder> createReorder(
       const MKLDNNMatrixPtr& src,
diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h
index e8ea6e37ac527a19c529d1731b94bed970211755..8193aa4adffc0409d8ea68417c68fa153a2942d8 100644
--- a/paddle/math/MathFunctions.h
+++ b/paddle/math/MathFunctions.h
@@ -26,7 +26,7 @@ limitations under the License. */
 #include <mkl_lapacke.h>
 #endif
 
-#ifdef PADDLE_USE_ATLAS
+#if defined(PADDLE_USE_ATLAS) || defined(PADDLE_USE_VECLIB)
 extern "C" {
 #include <cblas.h>
 #include <clapack.h>
diff --git a/paddle/math/Matrix.cpp b/paddle/math/Matrix.cpp
index 4a2132c8d1bfa329ced575f9b78052bdbfe3e4d5..0023b4d0f5da500f380ecb836b7c54e050b13d67 100644
--- a/paddle/math/Matrix.cpp
+++ b/paddle/math/Matrix.cpp
@@ -1033,17 +1033,15 @@ void GpuMatrix::maxPoolForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputH * outputW * channels);
 
   hl_maxpool_forward(frameNum,
                      inputData,
                      channels,
-                     height,
-                     width,
+                     imgSizeH,
+                     imgSizeW,
                      outputH,
                      outputW,
                      sizeX,
@@ -1080,11 +1078,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat,
   real* outDiff = outGrad.getData();
   size_t frameNum = inputMat.getHeight();
   size_t channels = outV.getWidth() / outputH / outputW;
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
-  CHECK(width_ == width * height * channels);
   CHECK(outGrad.getHeight() == outV.getHeight() &&
         outGrad.getWidth() == outV.getWidth());
 
@@ -1093,8 +1088,8 @@ void GpuMatrix::maxPoolBackward(Matrix& inputMat,
                       outData,
                       outDiff,
                       channels,
-                      height,
-                      width,
+                      imgSizeH,
+                      imgSizeW,
                       outputH,
                       outputW,
                       sizeX,
@@ -1125,17 +1120,15 @@ void GpuMatrix::avgPoolForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  CHECK(height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputH * outputW * channels);
 
   hl_avgpool_forward(frameNum,
                      inputData,
                      channels,
-                     height,
-                     width,
+                     imgSizeH,
+                     imgSizeW,
                      outputH,
                      outputW,
                      sizeX,
@@ -1166,17 +1159,15 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad,
   real* outDiff = outGrad.getData();
   size_t frameNum = outGrad.getHeight();
   size_t channels = outGrad.getWidth() / outputH / outputW;
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  CHECK(height * width * channels == width_);
+  CHECK(imgSizeH * imgSizeW * channels == width_);
   CHECK(height_ == outGrad.getHeight());
   CHECK(outGrad.getWidth() == outputH * outputW * channels);
 
   hl_avgpool_backward(frameNum,
                       outDiff,
                       channels,
-                      height,
-                      width,
+                      imgSizeH,
+                      imgSizeW,
                       outputH,
                       outputW,
                       sizeX,
@@ -1214,19 +1205,16 @@ void GpuMatrix::maxPool3DForward(Matrix& inputMat,
   real* inputData = inputMat.getData();
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t num = inputMat.getHeight();
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputD * outputH * outputW * channels);
 
   hl_maxpool3D_forward(num,
                        inputData,
                        channels,
-                       depth,
-                       height,
-                       width,
+                       imgSizeD,
+                       imgSizeH,
+                       imgSizeW,
                        outputD,
                        outputH,
                        outputW,
@@ -1269,20 +1257,16 @@ void GpuMatrix::maxPool3DBackward(Matrix& outGrad,
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t frameNum = getHeight();
   size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
-  size_t width = imgSizeW;
-  size_t height = imgSizeH;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == getWidth());
-  CHECK(width_ == depth * width * height * channels);
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
   CHECK(outGrad.getHeight() == maxPoolIdx.getHeight() &&
         outGrad.getWidth() == maxPoolIdx.getWidth());
 
   hl_maxpool3D_backward(frameNum,
                         outDiff,
                         channels,
-                        depth,
-                        height,
-                        width,
+                        imgSizeD,
+                        imgSizeH,
+                        imgSizeW,
                         outputD,
                         outputH,
                         outputW,
@@ -1323,19 +1307,16 @@ void GpuMatrix::avgPool3DForward(Matrix& inputMat,
 
   real* inputData = inputMat.getData();
   size_t frameNum = inputMat.getHeight();
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == inputMat.getWidth());
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == inputMat.getWidth());
   CHECK(height_ == inputMat.getHeight());
   CHECK(width_ == outputD * outputH * outputW * channels);
 
   hl_avgpool3D_forward(frameNum,
                        inputData,
                        channels,
-                       depth,
-                       height,
-                       width,
+                       imgSizeD,
+                       imgSizeH,
+                       imgSizeW,
                        outputD,
                        outputH,
                        outputW,
@@ -1375,19 +1356,16 @@ void GpuMatrix::avgPool3DBackward(Matrix& outGrad,
   real* outDiff = outGrad.getData();
   size_t frameNum = outGrad.getHeight();
   size_t channels = outGrad.getWidth() / outputD / outputH / outputW;
-  size_t height = imgSizeH;
-  size_t width = imgSizeW;
-  size_t depth = imgSizeD;
-  CHECK(depth * height * width * channels == width_);
+  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == width_);
   CHECK(height_ == outGrad.getHeight());
   CHECK(outGrad.getWidth() == outputD * outputH * outputW * channels);
 
   hl_avgpool3D_backward(frameNum,
                         outDiff,
                         channels,
-                        depth,
-                        height,
-                        width,
+                        imgSizeD,
+                        imgSizeH,
+                        imgSizeW,
                         outputD,
                         outputH,
                         outputW,
@@ -1999,11 +1977,11 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat,
   real* inputData = inputMat.getData();
   real* outData = data_;
   size_t num = inputMat.getHeight();
-  size_t inWidth = imgSizeW;
-  size_t inHeight = imgSizeH;
-  CHECK(inHeight * inWidth == inputMat.getWidth() / channels);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength == inputMat.getWidth() / channels);
   CHECK_EQ(num, this->getHeight());
-  CHECK_EQ(channels * outputH * outputW, this->getWidth());
+  CHECK_EQ(channels * outLength, this->getWidth());
   size_t outStride = getStride();
 
   /* initialize the data_ */
@@ -2020,24 +1998,24 @@ void CpuMatrix::maxPoolForward(Matrix& inputMat,
     }
     for (size_t c = 0; c < channels; ++c) {  // channel by channel
       for (size_t ph = 0; ph < outputH; ++ph) {
+        int hstart = ph * strideH - paddingH;
+        int hend = std::min(hstart + sizeY, imgSizeH);
+        hstart = std::max(hstart, 0);
         for (size_t pw = 0; pw < outputW; ++pw) {
-          int hstart = ph * strideH - paddingH;
           int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, inHeight);
-          int wend = std::min(wstart + sizeX, inWidth);
-          hstart = std::max(hstart, 0);
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
-              outData[ph * outputW + pw] = std::max(outData[ph * outputW + pw],
-                                                    inputData[h * inWidth + w]);
+              outData[ph * outputW + pw] = std::max(
+                  outData[ph * outputW + pw], inputData[h * imgSizeW + w]);
             }
           }
         }
       }
       // compute offset
-      inputData += inHeight * inWidth;
-      outData += outputH * outputW;
+      inputData += inLength;
+      outData += outLength;
     }
   }
 }
@@ -2058,8 +2036,10 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
                                 size_t paddingH,
                                 size_t paddingW) {
   size_t num = image.getHeight();
-  size_t channels = size_t(width_ / imgSizeH / imgSizeW);
-  CHECK(image.getWidth() == imgSizeH * imgSizeW * channels);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  size_t channels = size_t(width_ / inLength);
+  CHECK(image.getWidth() == inLength * channels);
   CHECK(image.getHeight() == height_ && image.getWidth() == width_);
   CHECK(outV.getHeight() == outGrad.getHeight() &&
         outV.getWidth() == outGrad.getWidth());
@@ -2080,12 +2060,12 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
+        int hstart = ph * strideH - paddingH;
+        int hend = std::min(hstart + sizeY, imgSizeH);
+        hstart = std::max(hstart, 0);
         for (size_t pw = 0; pw < outputW; ++pw) {
-          int hstart = ph * strideH - paddingH;
           int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, imgSizeH);
           int wend = std::min(wstart + sizeX, imgSizeW);
-          hstart = std::max(hstart, 0);
           wstart = std::max(wstart, 0);
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
@@ -2098,10 +2078,10 @@ void CpuMatrix::maxPoolBackward(Matrix& image,
         }
       }
       // offset
-      inData += imgSizeH * imgSizeW;
-      tgtGrad += imgSizeH * imgSizeW;
-      otData += outputH * outputW;
-      otGrad += outputH * outputW;
+      inData += inLength;
+      tgtGrad += inLength;
+      otData += outLength;
+      otGrad += outLength;
     }
   }
 }
@@ -2120,10 +2100,10 @@ void CpuMatrix::avgPoolForward(Matrix& input,
                                size_t paddingW) {
   // The main loop
   size_t num = input.getHeight();
-  size_t inHeight = imgSizeH;
-  size_t inWidth = imgSizeW;
-  CHECK(inHeight * inWidth * channels == input.getWidth());
-  CHECK(outputH * outputW * channels * num == height_ * width_);
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength * channels == input.getWidth());
+  CHECK(outLength * channels * num == height_ * width_);
   real* tgtData = data_;
   real* inData = input.getData();
 
@@ -2133,30 +2113,27 @@ void CpuMatrix::avgPoolForward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
+        int hstart = ph * strideH - paddingH;
+        int hend = std::min(hstart + sizeY, imgSizeH);
+        hstart = std::max(hstart, 0);
         for (size_t pw = 0; pw < outputW; ++pw) {
-          int hstart = ph * strideH - paddingH;
           int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, inHeight + paddingH);
-          int wend = std::min(wstart + sizeX, inWidth + paddingW);
-          int poolSize = (hend - hstart) * (wend - wstart);
-          hstart = std::max(hstart, 0);
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
-          hend = std::min(hend, static_cast<int>(inHeight));
-          wend = std::min(wend, static_cast<int>(inWidth));
-
-          CHECK(poolSize);
           tgtData[ph * outputW + pw] = 0;  // clear
           for (int h = hstart; h < hend; ++h) {
             for (int w = wstart; w < wend; ++w) {
-              tgtData[ph * outputW + pw] += inData[h * inWidth + w];
+              tgtData[ph * outputW + pw] += inData[h * imgSizeW + w];
             }
           }
+          int poolSize = (hend - hstart) * (wend - wstart);
+          CHECK(poolSize);
           tgtData[ph * outputW + pw] /= poolSize;
         }
       }
       // compute offset
-      inData += inHeight * inWidth;
-      tgtData += outputH * outputW;
+      inData += inLength;
+      tgtData += outLength;
     }
   }
 }
@@ -2176,7 +2153,9 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
                                 size_t paddingW) {
   size_t num = input.getHeight();
   size_t channels = input.getWidth() / outputH / outputW;
-  CHECK(imgSizeH * imgSizeW * channels == getWidth());
+  size_t inLength = imgSizeH * imgSizeW;
+  size_t outLength = outputH * outputW;
+  CHECK(inLength * channels == getWidth());
   real* inData = input.getData();
   real* outData = getData();
 
@@ -2186,16 +2165,14 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t ph = 0; ph < outputH; ++ph) {
+        int hstart = ph * strideH - paddingH;
+        int hend = std::min(hstart + sizeY, imgSizeH);
+        hstart = std::max(hstart, 0);
         for (size_t pw = 0; pw < outputW; ++pw) {
-          int hstart = ph * strideH - paddingH;
           int wstart = pw * strideW - paddingW;
-          int hend = std::min(hstart + sizeY, imgSizeH + paddingH);
-          int wend = std::min(wstart + sizeX, imgSizeW + paddingW);
-          int poolSize = (hend - hstart) * (wend - wstart);
-          hstart = std::max(hstart, 0);
+          int wend = std::min(wstart + sizeX, imgSizeW);
           wstart = std::max(wstart, 0);
-          hend = std::min(hend, static_cast<int>(imgSizeH));
-          wend = std::min(wend, static_cast<int>(imgSizeW));
+          int poolSize = (hend - hstart) * (wend - wstart);
           CHECK(poolSize);
 
           for (int h = hstart; h < hend; ++h) {
@@ -2206,8 +2183,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
         }
       }
       // offset
-      outData += imgSizeH * imgSizeW;
-      inData += outputH * outputW;
+      outData += inLength;
+      inData += outLength;
     }
   }
 }
@@ -2234,12 +2211,11 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
   real* outData = getData();
   real* maxPoolIdxData = maxPoolIdx.getData();
   size_t num = inputMat.getHeight();
-  size_t inWidth = imgSizeW;
-  size_t inHeight = imgSizeH;
-  size_t inDepth = imgSizeD;
-  CHECK(inHeight * inWidth * inDepth == inputMat.getWidth() / channels);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  CHECK(inLength == inputMat.getWidth() / channels);
   CHECK_EQ(num, this->getHeight());
-  CHECK_EQ(channels * outputH * outputW * outputD, this->getWidth());
+  CHECK_EQ(channels * outLength, this->getWidth());
   size_t outStride = getStride();
 
   /* initialize the data_ */
@@ -2258,16 +2234,16 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
     }
     for (size_t c = 0; c < channels; ++c) {  // channel by channel
       for (size_t pd = 0; pd < outputD; ++pd) {
+        int dstart = pd * strideD - paddingD;
+        int dend = std::min(dstart + sizeZ, imgSizeD);
+        dstart = std::max(dstart, 0);
         for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
+          hstart = std::max(hstart, 0);
           for (size_t pw = 0; pw < outputW; ++pw) {
-            int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
             int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, inDepth);
-            int hend = std::min(hstart + sizeY, inHeight);
-            int wend = std::min(wstart + sizeX, inWidth);
-            dstart = std::max(dstart, 0);
-            hstart = std::max(hstart, 0);
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
             int maxIdx = -1;
             real maxOutData = outData[(pd * outputH + ph) * outputW + pw];
@@ -2275,9 +2251,9 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
               for (int h = hstart; h < hend; ++h) {
                 for (int w = wstart; w < wend; ++w) {
                   if (maxOutData <
-                      inputData[(d * inHeight + h) * inWidth + w]) {
-                    maxOutData = inputData[(d * inHeight + h) * inWidth + w];
-                    maxIdx = (d * inHeight + h) * inWidth + w;
+                      inputData[(d * imgSizeH + h) * imgSizeW + w]) {
+                    maxOutData = inputData[(d * imgSizeH + h) * imgSizeW + w];
+                    maxIdx = (d * imgSizeH + h) * imgSizeW + w;
                   }
                 }
               }
@@ -2288,9 +2264,9 @@ void CpuMatrix::maxPool3DForward(Matrix& inputMat,
         }
       }
       // compute offset
-      inputData += inDepth * inHeight * inWidth;
-      outData += outputD * outputH * outputW;
-      maxPoolIdxData += outputD * outputH * outputW;
+      inputData += inLength;
+      outData += outLength;
+      maxPoolIdxData += outLength;
     }
   }
 }
@@ -2315,7 +2291,9 @@ void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
                                   real scaleTargets,
                                   real scaleOutput) {
   size_t num = getHeight();
-  size_t channels = size_t(width_ / imgSizeD / imgSizeH / imgSizeW);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  size_t channels = size_t(width_ / inLength);
   CHECK(maxPoolIdx.getHeight() == outGrad.getHeight() &&
         maxPoolIdx.getWidth() == outGrad.getWidth());
 
@@ -2341,9 +2319,9 @@ void CpuMatrix::maxPool3DBackward(Matrix& outGrad,
         }
       }
       // offset
-      tgtGrad += imgSizeD * imgSizeH * imgSizeW;
-      otGrad += outputD * outputH * outputW;
-      maxPoolIdxData += outputD * outputH * outputW;
+      tgtGrad += inLength;
+      otGrad += outLength;
+      maxPoolIdxData += outLength;
     }
   }
 }
@@ -2367,11 +2345,10 @@ void CpuMatrix::avgPool3DForward(Matrix& input,
                                  size_t paddingW) {
   // The main loop
   size_t num = input.getHeight();
-  size_t inDepth = imgSizeD;
-  size_t inHeight = imgSizeH;
-  size_t inWidth = imgSizeW;
-  CHECK(inDepth * inHeight * inWidth * channels == input.getWidth());
-  CHECK(outputD * outputH * outputW * channels * num == height_ * width_);
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  CHECK(inLength * channels == input.getWidth());
+  CHECK(outLength * channels * num == height_ * width_);
   real* tgtData = getData();
   real* inData = input.getData();
 
@@ -2381,39 +2358,36 @@ void CpuMatrix::avgPool3DForward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t pd = 0; pd < outputD; ++pd) {
+        int dstart = pd * strideD - paddingD;
+        int dend = std::min(dstart + sizeZ, imgSizeD);
+        dstart = std::max(dstart, 0);
         for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
+          hstart = std::max(hstart, 0);
           for (size_t pw = 0; pw < outputW; ++pw) {
-            int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
             int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, inDepth + paddingD);
-            int hend = std::min(hstart + sizeY, inHeight + paddingH);
-            int wend = std::min(wstart + sizeX, inWidth + paddingW);
-            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
-            dstart = std::max(dstart, 0);
-            hstart = std::max(hstart, 0);
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
-            dend = std::min(dend, static_cast<int>(inDepth));
-            hend = std::min(hend, static_cast<int>(inHeight));
-            wend = std::min(wend, static_cast<int>(inWidth));
 
-            CHECK(poolSize);
             tgtData[(pd * outputH + ph) * outputW + pw] = 0;  // clear
             for (int d = dstart; d < dend; ++d) {
               for (int h = hstart; h < hend; ++h) {
                 for (int w = wstart; w < wend; ++w) {
                   tgtData[(pd * outputH + ph) * outputW + pw] +=
-                      inData[(d * inHeight + h) * inWidth + w];
+                      inData[(d * imgSizeH + h) * imgSizeW + w];
                 }
               }
             }
+            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
+            CHECK(poolSize);
             tgtData[(pd * outputH + ph) * outputW + pw] /= poolSize;
           }
         }
       }
       // compute offset
-      inData += inDepth * inHeight * inWidth;
-      tgtData += outputD * outputH * outputW;
+      inData += inLength;
+      tgtData += outLength;
     }
   }
 }
@@ -2437,8 +2411,10 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
                                   real scaleTargets,
                                   real scaleOutput) {
   size_t num = input.getHeight();
-  size_t channels = input.getWidth() / outputD / outputH / outputW;
-  CHECK(imgSizeD * imgSizeH * imgSizeW * channels == getWidth());
+  size_t inLength = imgSizeH * imgSizeW * imgSizeD;
+  size_t outLength = outputH * outputW * outputD;
+  size_t channels = input.getWidth() / outLength;
+  CHECK(inLength * channels == getWidth());
   real* inData = input.getData();
   real* outData = getData();
 
@@ -2448,21 +2424,18 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
     }
     for (size_t c = 0; c < channels; ++c) {
       for (size_t pd = 0; pd < outputD; ++pd) {
+        int dstart = pd * strideD - paddingD;
+        int dend = std::min(dstart + sizeZ, imgSizeD);
+        dstart = std::max(dstart, 0);
         for (size_t ph = 0; ph < outputH; ++ph) {
+          int hstart = ph * strideH - paddingH;
+          int hend = std::min(hstart + sizeY, imgSizeH);
+          hstart = std::max(hstart, 0);
           for (size_t pw = 0; pw < outputW; ++pw) {
-            int dstart = pd * strideD - paddingD;
-            int hstart = ph * strideH - paddingH;
             int wstart = pw * strideW - paddingW;
-            int dend = std::min(dstart + sizeZ, imgSizeD + paddingD);
-            int hend = std::min(hstart + sizeY, imgSizeH + paddingH);
-            int wend = std::min(wstart + sizeX, imgSizeW + paddingW);
-            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
-            dstart = std::max(dstart, 0);
-            hstart = std::max(hstart, 0);
+            int wend = std::min(wstart + sizeX, imgSizeW);
             wstart = std::max(wstart, 0);
-            dend = std::min(dend, static_cast<int>(imgSizeD));
-            hend = std::min(hend, static_cast<int>(imgSizeH));
-            wend = std::min(wend, static_cast<int>(imgSizeW));
+            int poolSize = (dend - dstart) * (hend - hstart) * (wend - wstart);
             CHECK(poolSize);
             for (int d = dstart; d < dend; ++d) {
               for (int h = hstart; h < hend; ++h) {
@@ -2476,8 +2449,8 @@ void CpuMatrix::avgPool3DBackward(Matrix& input,
         }
       }
       // offset
-      outData += imgSizeD * imgSizeH * imgSizeW;
-      inData += outputD * outputH * outputW;
+      outData += inLength;
+      inData += outLength;
     }
   }
 }
diff --git a/paddle/math/NEONFunctions.cpp b/paddle/math/NEONFunctions.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..3bf47901f1069ac228fa1b877e29848d8cc130e8
--- /dev/null
+++ b/paddle/math/NEONFunctions.cpp
@@ -0,0 +1,55 @@
+/* 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. */
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+
+#include "NEONFunctions.h"
+#include <arm_neon.h>
+
+namespace paddle {
+namespace neon {
+
+// b[i] = a[i] > 0.0f ? a[i] : 0.0f
+void relu(const float* a, float* b, int len) {
+  int offset = len % 16;
+  float32x4_t ma0, ma1, ma2, ma3;
+  float32x4_t mb0, mb1, mb2, mb3;
+
+  float32x4_t zero = vdupq_n_f32(0.f);
+  for (int k = 0; k < len / 16; k++, a += 16, b += 16) {
+    ma0 = vld1q_f32(a);
+    ma1 = vld1q_f32(a + 4);
+    ma2 = vld1q_f32(a + 8);
+    ma3 = vld1q_f32(a + 12);
+
+    mb0 = vmaxq_f32(ma0, zero);
+    mb1 = vmaxq_f32(ma1, zero);
+    mb2 = vmaxq_f32(ma2, zero);
+    mb3 = vmaxq_f32(ma3, zero);
+
+    vst1q_f32(b, mb0);
+    vst1q_f32(b + 4, mb1);
+    vst1q_f32(b + 8, mb2);
+    vst1q_f32(b + 12, mb3);
+  }
+
+  for (int i = 0; i < offset; i++) {
+    b[i] = a[i] > 0.0f ? a[i] : 0.0f;
+  }
+}
+
+}  // namespace neon
+}  // namespace paddle
+
+#endif
diff --git a/paddle/math/NEONFunctions.h b/paddle/math/NEONFunctions.h
new file mode 100644
index 0000000000000000000000000000000000000000..69085e333547a31a341fbfde247f1e30adb957ee
--- /dev/null
+++ b/paddle/math/NEONFunctions.h
@@ -0,0 +1,23 @@
+/* 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
+
+namespace paddle {
+namespace neon {
+
+void relu(const float* a, float* b, int len);
+
+}  // namespace neon
+}  // namespace paddle
diff --git a/paddle/math/tests/test_matrixCompare.cpp b/paddle/math/tests/test_matrixCompare.cpp
index 103f06acc57d7a23f019f5e713f6cacf2179e9e0..061fb22e3fd744d9d9895fd1008089e4a6ce6a0f 100644
--- a/paddle/math/tests/test_matrixCompare.cpp
+++ b/paddle/math/tests/test_matrixCompare.cpp
@@ -825,9 +825,8 @@ void testMaxPoolFwdBwd(int numSamples,
                        int strideW,
                        int padH,
                        int padW) {
-  int outH = 0, outW = 0;
-  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
-  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
 
   int inWidth = imgSizeH * imgSizeW * channels;
   MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
@@ -927,9 +926,8 @@ void testAvgPoolFwdBwd(int numSamples,
                        int strideW,
                        int padH,
                        int padW) {
-  int outH = 0, outW = 0;
-  outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
-  outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
+  int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
+  int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
 
   int inWidth = imgSizeH * imgSizeW * channels;
   MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
diff --git a/paddle/memory/memcpy.cc b/paddle/memory/memcpy.cc
index a19a3e3675e3e2e7cc0c3594f21191f932d6379f..19ec9ba9b26f5919796181a19a048b7edb508bdd 100644
--- a/paddle/memory/memcpy.cc
+++ b/paddle/memory/memcpy.cc
@@ -62,6 +62,24 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
   }
 }
 
+template <>
+void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
+                                                  void* dst,
+                                                  platform::GPUPlace src_place,
+                                                  const void* src, size_t num) {
+  platform::SetDeviceId(src_place.device);
+  platform::GpuMemcpySync(dst, src, num, cudaMemcpyDeviceToHost);
+}
+
+template <>
+void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
+                                                  void* dst,
+                                                  platform::CPUPlace src_place,
+                                                  const void* src, size_t num) {
+  platform::SetDeviceId(dst_place.device);
+  platform::GpuMemcpySync(dst, src, num, cudaMemcpyHostToDevice);
+}
+
 #endif  // PADDLE_ONLY_CPU
 
 }  // namespace memory
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index 5b65584327e8afca383d2171cb42e3984baa8654..e3e934bcccd1a5f34d88a2f33f3708a46ddabe05 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -80,9 +80,11 @@ endfunction()
 add_subdirectory(math)
 
 set(DEPS_OPS
-    recurrent_op)
+    recurrent_op
+    cond_op)
 op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
   DEPS framework_proto tensor net_op)
+op_library(cond_op SRCS cond_op.cc DEPS framework_proto tensor operator net_op)
 
 list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS})
 foreach(src ${GENERAL_OPS})
diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc
index 4a6c6381b0341dd3531aa4c09024530ee67bb4f9..0c813748b2989a8f0c00a359345747242dd21dd8 100644
--- a/paddle/operators/accuracy_op.cc
+++ b/paddle/operators/accuracy_op.cc
@@ -23,10 +23,15 @@ class AccuracyOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Inference"),
-                            "Input of Inference must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Inference"),
+        "Input(Inference) of AccuracyOp should not be null.");
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
-                            "Input of Inference must be initialized.");
+                            "Input(Label) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Accuracy"),
+        "Output(Accuracy) of AccuracyOp should not be null.");
+
     auto *inference = ctx.Input<framework::Tensor>("Inference");
     auto *label = ctx.Input<framework::Tensor>("Label");
 
diff --git a/paddle/operators/accuracy_op.cu b/paddle/operators/accuracy_op.cu
index 4e6d1ef9654012ce6355cbd7561c4fdc1785c11a..0a6a0fd15c73330902552f7a9aa6339de24c1a18 100644
--- a/paddle/operators/accuracy_op.cu
+++ b/paddle/operators/accuracy_op.cu
@@ -12,26 +12,38 @@ 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 <thrust/execution_policy.h>
+#include <thrust/reduce.h>
 #include "paddle/operators/accuracy_op.h"
+#include "paddle/platform/cuda_helper.h"
 
 namespace paddle {
 namespace operators {
+using platform::PADDLE_CUDA_NUM_THREADS;
 
-__global__ void AccuracySingleKernel(const int N, const int D, const int top_k,
-                                     const int* Xdata, const int* labelData,
-                                     float* accuracy) {
-  int correct = 0;
-  for (int row = 0; row < N; row++) {
-    const int label = labelData[row];
-    for (int col = 0; col < D; col++) {
-      const int pred = Xdata[row * D + col];
-      if (pred == label) {
-        ++correct;
+template <int BlockSize>
+__global__ void AccuracyCudaKernel(const int N, const int D, const int* Xdata,
+                                   const int* labeldata, float* accuracy) {
+  int count = 0;
+  __shared__ int total[BlockSize];
+
+  // support only 1 block
+  for (int i = threadIdx.x; i < (N); i += BlockSize) {
+    for (int j = 0; j < D; ++j) {
+      if (Xdata[i * D + j] == labeldata[i]) {
+        ++count;
         break;
       }
     }
   }
-  *accuracy = static_cast<float>(correct) / static_cast<float>(N);
+  total[threadIdx.x] = count;
+  __syncthreads();
+
+  // reduce the count with init value 0, and output accuracy.
+  int result = thrust::reduce(thrust::device, total, total + BlockSize, 0);
+  if (threadIdx.x == 0) {
+    *accuracy = static_cast<float>(result) / static_cast<float>(N);
+  }
 }
 
 template <typename T>
@@ -57,8 +69,8 @@ class AccuracyOpCUDAKernel : public framework::OpKernel {
       return;
     }
 
-    AccuracySingleKernel<<<1, 1>>>(num_samples, infer_width, 1, inference_data,
-                                   label_data, accuracy_data);
+    AccuracyCudaKernel<PADDLE_CUDA_NUM_THREADS><<<1, PADDLE_CUDA_NUM_THREADS>>>(
+        num_samples, infer_width, inference_data, label_data, accuracy_data);
   }
 };
 
diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc
index b43c09d4f09c7f87cc60290bdd2a99cbe46f0d5c..e83c1efeaf897889d18a37a6bd2ca2f8f012db25 100644
--- a/paddle/operators/add_op.cc
+++ b/paddle/operators/add_op.cc
@@ -23,6 +23,13 @@ class AddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of AddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of AddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of AddOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
                       ctx.Input<Tensor>("Y")->dims(),
                       "Two input of Add Op's dimension must be same.");
diff --git a/paddle/operators/concat_op.cc b/paddle/operators/concat_op.cc
index 72fd179354a4be76a37e4571da168d844f7ce384..223bb0ffe6e75ce71919eb5f4cca06bedbb00764 100644
--- a/paddle/operators/concat_op.cc
+++ b/paddle/operators/concat_op.cc
@@ -25,6 +25,9 @@ class ConcatOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ConcatOp should not be null.");
+
     auto ins = ctx.MultiInput<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
diff --git a/paddle/operators/cond_op.cc b/paddle/operators/cond_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..8262a7a5c8c13c86c5f6c123a14fa89696358c57
--- /dev/null
+++ b/paddle/operators/cond_op.cc
@@ -0,0 +1,229 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/cond_op.h"
+
+#include <cstring>
+#include <sstream>
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/gather.h"
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/scatter.h"
+
+namespace paddle {
+namespace operators {
+
+using Scope = framework::Scope;
+using Variable = framework::Variable;
+using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
+using DDim = framework::DDim;
+
+void CondOp::CreateScope(const Scope& scope) const {
+  auto sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto sub_scopes = sub_scopes_var->GetMutable<std::vector<Scope*>>();
+  auto& sub_scope = scope.NewScope();
+  sub_scopes->push_back(&sub_scope);
+}
+
+void CondOp::CreateIndexTensor(const Scope& scope) const {
+  auto index_tensors_var = scope.FindVar("IndexTensors");
+  PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
+                          "Output(IndexTensors) of CondOp should not be null.");
+  auto& index_tensors =
+      *index_tensors_var->GetMutable<std::vector<LoDTensor>>();
+  index_tensors.push_back(LoDTensor());
+}
+
+void CondOp::InferShape(const Scope& scope) const {
+  auto sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto& sub_scopes = *sub_scopes_var->GetMutable<std::vector<Scope*>>();
+
+  for (int i = 0; i < 2; ++i) {
+    // Create two sub scopes for true and false branches
+    // sub_scopes[0] for the true branch and sub_scopes[1] for the false
+    // branch
+    CreateScope(scope);
+
+    // Create two tensors for true and false indices
+    // index_tensors[0] for the true branch and index_tensors[1] for the false
+    // branch
+    CreateIndexTensor(scope);
+
+    PADDLE_ENFORCE(!Inputs("Xs").empty(),
+                   "Inputs(Xs) of CondOp can't be empty.");
+    for (auto& input : Inputs("Xs")) {
+      // Create a new tensor in sub-scope for input-type tensor
+      Variable* v = sub_scopes[i]->NewVar(input);
+      LoDTensor* sub_input = v->GetMutable<LoDTensor>();
+      sub_input->Resize(scope.FindVar(input)->GetMutable<LoDTensor>()->dims());
+    }
+
+    for (auto& output : (*sub_net_op_[i]).Outputs()) {
+      for (auto& var_name : output.second) {
+        sub_scopes[i]->NewVar(var_name);
+      }
+    }
+
+    // each net calls InferShape
+    sub_net_op_[i]->InferShape(*sub_scopes[i]);
+  }
+
+  for (auto& output : Outputs("Outs")) {
+    LoDTensor* tensor_t_out =
+        sub_scopes[0]->FindVar(output)->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_t_out, "True output should not be NULL");
+    LoDTensor* tensor_f_out =
+        sub_scopes[1]->FindVar(output)->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_f_out, "False output should not be NULL");
+
+    auto* tensor_out_var = scope.FindVar(output);
+    PADDLE_ENFORCE_NOT_NULL(tensor_out_var, "Output not found");
+    LoDTensor* tensor_out = tensor_out_var->GetMutable<LoDTensor>();
+    PADDLE_ENFORCE_NOT_NULL(tensor_t_out,
+                            "True output tensor should not be NULL");
+
+    // check output size should be same
+    PADDLE_ENFORCE_EQ(tensor_t_out->dims(), tensor_f_out->dims(),
+                      "Outputs not of the same shape");
+    tensor_out->Resize(tensor_t_out->dims());
+    // tensor_out->mutable_data<float>(tensor_out->dims(),
+    // platform::CPUPlace());
+    tensor_out->mutable_data<float>(platform::CPUPlace());
+  }
+}
+
+void CondOp::Run(const Scope& scope,
+                 const platform::DeviceContext& dev_ctx) const {
+  auto* sub_scopes_var = scope.FindVar("SubScopes");
+  PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
+                          "Output(SubScopes) of CondOp should not be null.");
+  auto sub_scopes = sub_scopes_var->Get<std::vector<Scope*>>();
+  auto* index_tensors_var = scope.FindVar("IndexTensors");
+  PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
+                          "Output(IndexTensors) of CondOp should not be null.");
+  auto index_tensors = index_tensors_var->Get<std::vector<LoDTensor>>();
+
+  std::string cond_name = Input("Cond");
+  Variable* cond_var = scope.FindVar(cond_name);
+  PADDLE_ENFORCE_NOT_NULL(cond_var,
+                          "Input(Cond) of CondOp should not be null.");
+  const LoDTensor* cond = cond_var->GetMutable<LoDTensor>();
+
+  // Step 1: get the true/false index at runtime
+  // index_[0]: vector<int>, contains all index for cond[i] == true
+  // index_[1]: vector<int>, contains all index for cond[i] == false
+  for (int i = 0; i < 2; ++i) index_[i].clear();
+
+  const int* cond_data = cond->data<int>();
+  for (int i = 0; i < cond->dims()[0]; ++i) {
+    if (cond_data[i])
+      index_[0].push_back(i);
+    else
+      index_[1].push_back(i);
+  }
+
+  // put index_[0] and index_[1] into two tensors:
+  // index_tensor_[0] and index_tensor_[1]
+  DDim dim = paddle::framework::make_ddim({0});
+  for (int i = 0; i < 2; ++i) {
+    dim[0] = index_[i].size();
+    int* tmp_ptr =
+        index_tensors[i].mutable_data<int>(dim, platform::CPUPlace());
+    index_tensors[i].Resize(dim);
+    memcpy(tmp_ptr, index_[i].data(), dim[0] * sizeof(int));
+  }
+
+  // Step 2: collect data by calling gather
+  for (int i = 0; i < 2; ++i) {
+    // i= 0/i for True and False branches respectively
+    for (auto& input : Inputs("Xs")) {
+      // find Tensor
+      Variable* v = scope.FindVar(input);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_parent = v->GetMutable<LoDTensor>();
+
+      v = sub_scopes[i]->FindVar(input);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_child = v->GetMutable<LoDTensor>();
+
+      // Resize child
+      DDim dim = tensor_child->dims();
+      dim[0] = index_[i].size();
+      tensor_child->Resize(dim);
+      tensor_child->mutable_data<float>(dim, platform::CPUPlace());
+
+      Gather<float>(dev_ctx.GetPlace(), tensor_parent, &index_tensors[i],
+                    tensor_child);
+    }
+  }
+
+  // Step 3: run
+  for (int i = 0; i < 2; ++i) {
+    sub_net_op_[i]->Run(*sub_scopes[i], dev_ctx);
+  }
+
+  // Step 4: merge output results
+  PADDLE_ENFORCE(!Outputs("Outs").empty(),
+                 "Outputs(Outs) of CondOp can't be empty.");
+  for (int i = 0; i < 2; ++i) {
+    // i= 0/i for True and False branches respectively
+    for (auto& output : Outputs("Outs")) {
+      // find Tensor
+      Variable* v = scope.FindVar(output);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_parent = v->GetMutable<LoDTensor>();
+
+      v = sub_scopes[i]->FindVar(output);
+      PADDLE_ENFORCE_NOT_NULL(v);
+      LoDTensor* tensor_child = v->GetMutable<LoDTensor>();
+
+      ScatterUpdate<float>(dev_ctx.GetPlace(), tensor_child, &index_tensors[i],
+                           tensor_parent);
+    }
+  }
+}
+
+class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  CondOpProtoAndCheckerMaker(framework::OpProto* proto,
+                             framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Cond", "The condition, which is a bool vector");
+    AddInput("Xs", "Inputs of Subnets").AsDuplicable();
+    AddOutput("Outs", "Outputs of Cond_Op after merge").AsDuplicable();
+
+    AddOutput("SubScopes", "sub scopes for true and false branches");
+    AddOutput("IndexTensors", "Index Tensors contains indices for true/false");
+
+    AddComment(R"DOC(
+Sample dependent Cond Operator:
+Given Cond[i] as a 1/0 vector to indicate true/false
+The equation is: 
+Out[i] = subnet_t[i], if Cond[i] == true
+Out[i] = subnet_t[i], if Cond[i] == false
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OP_WITHOUT_GRADIENT(cond, paddle::operators::CondOp,
+                             paddle::operators::CondOpProtoAndCheckerMaker);
diff --git a/paddle/operators/cond_op.h b/paddle/operators/cond_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..b09e32331e66c53555c88c06d7b1456276050eaa
--- /dev/null
+++ b/paddle/operators/cond_op.h
@@ -0,0 +1,91 @@
+/* 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 <vector>
+#include "glog/logging.h"
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/operator.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+/*
+ * @brief CondOp is a dynamic if-else Operator
+ *
+ * It has a input tensor named cond indicating which netop each instance will
+ * run.
+ *
+ * if cond == 1, it will run true_net, which is a NetOp.
+ *
+ * if cond == 0, it will run false_net, which is another NetOp.
+ */
+class CondOp : public framework::OperatorBase {
+ public:
+  CondOp(const std::string& type, const framework::VariableNameMap& inputs,
+         const framework::VariableNameMap& outputs,
+         const framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {
+    index_.resize(2);
+    sub_net_op_.resize(2);
+  }
+
+  CondOp(const CondOp& o)
+      : framework::OperatorBase(
+            static_cast<const framework::OperatorBase&>(o)) {
+    // TODO(yuyang18): Implement copy ctor well.
+    PADDLE_THROW("Not implemented");
+  }
+
+  void CreateScope(const framework::Scope& scope) const;
+
+  void CreateIndexTensor(const framework::Scope& scope) const;
+
+  /*
+   * InferShape must be called before Run.
+   */
+  void InferShape(const framework::Scope& scope) const override;
+
+  /*
+   * Set True Block
+   */
+  void set_truenet(std::unique_ptr<OperatorBase>&& net) {
+    sub_net_op_[0] = std::move(net);
+  }
+
+  /*
+   * Set False Block
+   */
+  void set_falsenet(std::unique_ptr<OperatorBase>&& net) {
+    sub_net_op_[1] = std::move(net);
+  }
+
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override;
+
+ private:
+  // sub_net_op_[0]: subnet_t
+  // sub_net_op_[1]: subnet_f
+  std::vector<std::unique_ptr<framework::OperatorBase>> sub_net_op_;
+
+  // index_[0]: True_index;
+  // index_[1]: False_index;
+  mutable std::vector<std::vector<int>> index_;
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc
index 253b17d8a1b88eccc58fc458ae8274d2bbd1c323..72c446493684246959656dc048e7f0e761665423 100644
--- a/paddle/operators/cos_sim_op.cc
+++ b/paddle/operators/cos_sim_op.cc
@@ -26,8 +26,16 @@ class CosSimOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     // notnull check
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("XNorm"),
+                            "Output(XNorm) of CosSimOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("YNorm"),
+                            "Output(YNorm) of CosSimOp should not be null.");
 
     // shape check
     auto x_dims = ctx.Input<Tensor>("X")->dims();
diff --git a/paddle/operators/cos_sim_op.h b/paddle/operators/cos_sim_op.h
index 318b63f3707cf77755de773a39b00aa30d2296d3..bcf6f758cae561a2e22f5be6c7a242647ef1c144 100644
--- a/paddle/operators/cos_sim_op.h
+++ b/paddle/operators/cos_sim_op.h
@@ -56,7 +56,7 @@ class CosSimKernel : public framework::OpKernel {
     x_norm.device(place) = x.square().sum(row_along).sqrt();
     y_norm.device(place) = y.square().sum(row_along).sqrt();
     if (rows_x == rows_y) {
-      auto xy = (x * y).sum(Eigen::array<int, 1>({1}));
+      auto xy = (x * y).sum(Eigen::array<int, 1>({{1}}));
       z.device(place) = xy / x_norm / y_norm;
     } else {
       Eigen::DSizes<int, 2> bcast(rows_x, 1);
@@ -134,7 +134,7 @@ class CosSimGradKernel : public framework::OpKernel {
         out_grad_y->mutable_data<T>(context.GetPlace());
         auto dy = EigenMatrix<T>::Reshape(*out_grad_y, 1);
         auto grad = x / norm_prod_bcast - z_bcast * y_bcast / y_snorm_bcast;
-        dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({0}));
+        dy.device(place) = (dz_bcast * grad).sum(Eigen::array<int, 1>({{0}}));
       }
     }
   }
diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..953367eb8bcd1282ab6c7e1189d778f0ce3da541
--- /dev/null
+++ b/paddle/operators/cross_entropy_op.cc
@@ -0,0 +1,147 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/cross_entropy_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::LoDTensor;
+
+class CrossEntropyOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
+                            "Input(Label) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"), "Output(Y) must not be null.");
+
+    auto x = ctx.Input<Tensor>("X");
+    auto label = ctx.Input<Tensor>("Label");
+    PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(label->dims().size(), 2,
+                      "Input(Label)'s rank must be 2.");
+    // TODO(xinghai-sun): remove this check after swtiching to bool
+    PADDLE_ENFORCE(ctx.Attr<int>("soft_label") == 0 ||
+                   ctx.Attr<int>("soft_label") == 1);
+    PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0],
+                      "The 1st dimension of Input(X) and Input(Label) must "
+                      "be equal.");
+    if (ctx.Attr<int>("soft_label") == 1) {
+      PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1],
+                        "If Attr(soft_label) == 1, The 2nd dimension of "
+                        "Input(X) and Input(Label) must be equal.");
+    } else {
+      PADDLE_ENFORCE_EQ(label->dims()[1], 1,
+                        "If Attr(soft_label) == 0, The 2nd dimension of "
+                        "Input(Label) must be 1.");
+    }
+
+    ctx.Output<LoDTensor>("Y")->Resize({x->dims()[0], 1});
+  }
+};
+
+class CrossEntropyGradientOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
+                            "Input(Label) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")),
+                            "Input(Y@GRAD) must not be null.");
+
+    auto x = ctx.Input<Tensor>("X");
+    auto label = ctx.Input<Tensor>("Label");
+    auto dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(dy->dims().size(), 2, "Input(Y@Grad)'s rank must be 2.");
+    PADDLE_ENFORCE_EQ(label->dims().size(), 2,
+                      "Input(Label)'s rank must be 2.");
+    // TODO(xinghai-sun): remove this check after swtiching to bool
+    PADDLE_ENFORCE(ctx.Attr<int>("soft_label") == 0 ||
+                   ctx.Attr<int>("soft_label") == 1);
+    PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0],
+                      "The 1st dimension of Input(X) and Input(Label) must "
+                      "be equal.");
+    PADDLE_ENFORCE_EQ(x->dims()[0], dy->dims()[0],
+                      "The 1st dimension of Input(X) and Input(Y@Grad) must "
+                      "be equal.");
+    PADDLE_ENFORCE_EQ(dy->dims()[1], 1,
+                      "The 2nd dimension of Input(Y@Grad) must be 1.");
+    if (ctx.Attr<int>("soft_label") == 1) {
+      PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1],
+                        "If Attr(soft_label) == 1, The 2nd dimension of "
+                        "Input(X) and Input(Label) must be equal.");
+    } else {
+      PADDLE_ENFORCE_EQ(label->dims()[1], 1,
+                        "If Attr(soft_label) == 0, The 2nd dimension of "
+                        "Input(Label) must be 1.");
+    }
+
+    auto dx = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    dx->Resize(x->dims());
+  }
+};
+
+class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  CrossEntropyOpMaker(framework::OpProto *proto,
+                      framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of CrossEntropyOp");
+    AddInput("Label", "The second input of CrossEntropyOp");
+    AddOutput("Y", "The output of CrossEntropyOp");
+    AddAttr<int>("soft_label", "Is soft label. Default zero.").SetDefault(0);
+
+    AddComment(R"DOC(
+CrossEntropy Operator.
+
+It supports both standard cross-entropy and soft-label cross-entropy loss
+computation.
+1) One-hot cross-entropy:
+    soft_label = 0, Label[i, 0] indicates the class index for sample i:
+
+                Y[i] = -log(X[i, Label[i]])
+
+2) Soft-label cross-entropy:
+    soft_label = 1, Label[i, j] indicates the soft label of class j
+    for sample i:
+
+                Y[i] = \sum_j{-Label[i, j] * log(X[i, j])}
+
+   Please make sure that in this case the summuation of each row of Label
+   equals one.
+
+3) One-hot cross-entropy with vecterized Input(Label):
+     As a special case of 2), when each row of Input(Label) has only one
+     non-zero element (equals 1), soft-label cross-entropy degenerates to a
+     one-hot cross-entropy with one-hot label representation.
+)DOC");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker,
+            cross_entropy_grad, ops::CrossEntropyGradientOp);
+REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<float>);
+REGISTER_OP_CPU_KERNEL(cross_entropy_grad,
+                       ops::CrossEntropyGradientOpKernel<float>);
diff --git a/paddle/operators/cross_entropy_op.cu b/paddle/operators/cross_entropy_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..ab6ad0e062269483948bf70e492c9431991221fb
--- /dev/null
+++ b/paddle/operators/cross_entropy_op.cu
@@ -0,0 +1,158 @@
+/* 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/op_registry.h"
+#include "paddle/operators/cross_entropy_op.h"
+#include "paddle/platform/assert.h"
+#include "paddle/platform/hostdevice.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label,
+                                   const int N, const int D) {
+  // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file.
+  // CUDA_1D_KERNEL_LOOP(i, N) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
+       i += blockDim.x * gridDim.x) {
+    PADDLE_ASSERT(label[i] >= 0 && label[i] < D);
+    Y[i] = -tolerable_value(log(X[i * D + label[i]]));
+  }
+}
+
+template <typename T>
+__global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label,
+                                       const int N, const int D) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
+       i += blockDim.x * gridDim.x) {
+    T sum = static_cast<T>(0);
+    for (int j = 0; j < D; j++) {
+      sum += label[i * D + j] * tolerable_value(log(X[i * D + j]));
+    }
+    Y[i] = -sum;
+  }
+}
+
+// TODO(qingqing): make zero setting an common function.
+template <typename T>
+__global__ void zero(T* X, const int N) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
+       i += blockDim.x * gridDim.x) {
+    X[i] = 0.0;
+  }
+}
+
+template <typename T>
+__global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X,
+                                           const int* label, const int N,
+                                           const int D) {
+  // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file.
+  // CUDA_1D_KERNEL_LOOP(i, N) {
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
+       i += blockDim.x * gridDim.x) {
+    int idx = i * D + label[i];
+    dX[idx] = -dY[i] / X[idx];
+  }
+}
+
+template <typename T>
+__global__ void SoftCrossEntropyGradientKernel(T* dX, const T* dY, const T* X,
+                                               const T* label, const int N,
+                                               const int D) {
+  // TOOD(qingqing): optimize for this kernel
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
+       i += blockDim.x * gridDim.x) {
+    for (int j = 0; j < D; ++j) {
+      int idx = i * D + j;
+      dX[idx] = -label[idx] * dY[i] / X[idx];
+    }
+  }
+}
+
+template <typename T>
+class CrossEntropyOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+
+    auto x = ctx.Input<Tensor>("X");
+    auto y = ctx.Output<Tensor>("Y");
+    auto label = ctx.Input<Tensor>("Label");
+
+    auto* x_data = x->data<T>();
+    y->mutable_data<T>(ctx.GetPlace());
+    auto* y_data = y->data<T>();
+
+    int n = x->dims()[0];
+    int d = x->dims()[1];
+    int block = 512;
+    int grid = (n + block - 1) / block;
+    // TODO(qingqing) launch kernel on specified stream
+    // base on ExecutionContext.
+    if (ctx.Attr<int>("soft_label") == 1) {
+      auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
+      SoftCrossEntropyKernel<T><<<grid, block>>>(y_data, x_data, label_data, n,
+                                                 d);
+    } else {
+      auto* label_data = ctx.Input<Tensor>("Label")->data<int>();
+      CrossEntropyKernel<T><<<grid, block>>>(y_data, x_data, label_data, n, d);
+    }
+  }
+};
+
+template <typename T>
+class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
+                   "It must use GPUPlace.");
+
+    auto x = ctx.Input<Tensor>("X");
+    auto dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    auto label = ctx.Input<Tensor>("Label");
+
+    auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());
+    auto* dy_data = dy->data<T>();
+    auto* x_data = x->data<T>();
+
+    int n = x->dims()[0];
+    int d = x->dims()[1];
+    int block = 512;
+    int grid = (n * d + block - 1) / block;
+    zero<T><<<grid, block>>>(dx_data, n * d);
+    grid = (n + block - 1) / block;
+    // TODO(qingqing): launch kernel on specified stream
+    // base on ExecutionContext.
+    if (ctx.Attr<int>("soft_label") == 1) {
+      auto* label_data = label->data<T>();
+      SoftCrossEntropyGradientKernel<T><<<grid, block>>>(
+          dx_data, dy_data, x_data, label_data, n, d);
+    } else {
+      auto* label_data = label->data<int>();
+      CrossEntropyGradientKernel<T><<<grid, block>>>(dx_data, dy_data, x_data,
+                                                     label_data, n, d);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel<float>);
+REGISTER_OP_GPU_KERNEL(cross_entropy_grad,
+                       ops::CrossEntropyGradientOpCUDAKernel<float>);
diff --git a/paddle/operators/cross_entropy_op.h b/paddle/operators/cross_entropy_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..1b4b23ac2029138afadef0168262203ac2e20430
--- /dev/null
+++ b/paddle/operators/cross_entropy_op.h
@@ -0,0 +1,117 @@
+/* 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/op_registry.h"
+#include "paddle/platform/hostdevice.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+HOSTDEVICE T tolerable_value(const T x) {
+  PADDLE_ASSERT(std::is_floating_point<T>::value);
+  const T kApproInf = 1e20;
+  if (x == INFINITY) {
+    return kApproInf;
+  }
+  if (x == -INFINITY) {
+    return -kApproInf;
+  }
+  return x;
+}
+
+template <typename T>
+class CrossEntropyOpKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+
+    auto x = ctx.Input<Tensor>("X");
+    auto y = ctx.Output<Tensor>("Y");
+
+    auto* x_data = x->data<T>();
+    y->mutable_data<T>(ctx.GetPlace());
+    auto* y_data = y->data<T>();
+
+    int batch_size = x->dims()[0];
+    int class_num = x->dims()[1];
+
+    if (ctx.Attr<int>("soft_label") == 1) {
+      auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
+      int index = 0;
+      for (int i = 0; i < batch_size; ++i) {
+        T sum = static_cast<T>(0);
+        for (int j = 0; j < class_num; ++j) {
+          sum += label_data[index] * tolerable_value(std::log(x_data[index]));
+          y_data[i] = -sum;
+          index++;
+        }
+      }
+    } else {
+      auto* label_data = ctx.Input<Tensor>("Label")->data<int>();
+      for (int i = 0; i < batch_size; ++i) {
+        int index = i * class_num + label_data[i];
+        y_data[i] = -tolerable_value(std::log(x_data[index]));
+      }
+    }
+  }
+};
+
+template <typename T>
+class CrossEntropyGradientOpKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
+                   "It must use CPUPlace.");
+
+    auto x = ctx.Input<Tensor>("X");
+    auto dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+    auto dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
+    auto label = ctx.Input<Tensor>("Label");
+
+    auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());
+    auto* dy_data = dy->data<T>();
+    auto* x_data = x->data<T>();
+
+    int batch_size = x->dims()[0];
+    int class_num = x->dims()[1];
+
+    // TODO(qingqing): make zero setting an common function.
+    if (ctx.Attr<int>("soft_label") == 1) {
+      auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
+      int index = 0;
+      for (int i = 0; i < batch_size; ++i) {
+        for (int j = 0; j < class_num; ++j) {
+          dx_data[index] = -label_data[index] * dy_data[i] / x_data[index];
+          index++;
+        }
+      }
+    } else {
+      auto* label_data = label->data<int>();
+      memset(dx_data, 0, sizeof(T) * batch_size * class_num);
+      for (int i = 0; i < batch_size; ++i) {
+        PADDLE_ASSERT(label_data[i] >= 0 || label_data[i] < class_num);
+        int index = i * class_num + label_data[i];
+        dx_data[index] = -dy_data[i] / x_data[index];
+      }
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/dropout_op.cc b/paddle/operators/dropout_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..b111b9fccb2310bd5fb92bda878a497c51f62ce0
--- /dev/null
+++ b/paddle/operators/dropout_op.cc
@@ -0,0 +1,113 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/dropout_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+using framework::LoDTensor;
+
+class DropoutOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_GE(ctx.Attr<float>("dropout_prob"), 0);
+    PADDLE_ENFORCE_LE(ctx.Attr<float>("dropout_prob"), 1);
+    // TODO(xinghai-sun): remove this check after swtiching to bool
+    PADDLE_ENFORCE(ctx.Attr<int>("is_training") == 0 ||
+                   ctx.Attr<int>("is_training") == 1);
+
+    auto dims = ctx.Input<Tensor>("X")->dims();
+    ctx.Output<LoDTensor>("Out")->Resize(dims);
+    if (ctx.Attr<int>("is_training") == 1) {
+      ctx.Output<LoDTensor>("Mask")->Resize(dims);
+    }
+  }
+};
+
+template <typename AttrType>
+class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  DropoutOpMaker(framework::OpProto *proto,
+                 framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddAttr<AttrType>("dropout_prob", "Probability of setting units to zero.")
+        .SetDefault(.5f);
+    // TODO(xinghai-sun): use bool for is_training after bool is supported.
+    AddAttr<int>("is_training", "Whether in training phase.").SetDefault(1);
+    AddAttr<int>("seed", "Dropout random seed.").SetDefault(0);
+    AddInput("X", "The input of dropout op.");
+    AddOutput("Out", "The output of dropout op.");
+    AddOutput("Mask", "The random sampled dropout mask.").AsIntermediate();
+
+    AddComment(R"DOC(
+Dropout Operator.
+
+"Dropout" refers to randomly dropping out units in a nerual network. It is a
+regularization technique for reducing overfitting by preventing neuron
+co-adaption during training. The dropout operator randomly set (according to
+the given dropout probability) the outputs of some units to zero, while others
+being set to their inputs.
+)DOC");
+  }
+};
+
+template <typename AttrType>
+class DropoutOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_EQ(ctx.Attr<int>("is_training"), 1,
+                      "GradOp is only callable when is_training is true");
+
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Mask"), "Mask must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) must not be null.");
+
+    PADDLE_ENFORCE_GE(ctx.Attr<AttrType>("dropout_prob"), 0);
+    PADDLE_ENFORCE_LE(ctx.Attr<AttrType>("dropout_prob"), 1);
+    // TODO(xinghai-sun): remove this check after swtiching to bool
+    PADDLE_ENFORCE(ctx.Attr<int>("is_training") == 0 ||
+                   ctx.Attr<int>("is_training") == 1);
+    auto x_dims = ctx.Input<Tensor>("X")->dims();
+    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    PADDLE_ENFORCE_EQ(x_dims, out_dims,
+                      "Dimensions of Input(X) and Out@Grad must be the same.");
+    auto mask_dims = ctx.Input<Tensor>("Mask")->dims();
+    PADDLE_ENFORCE_EQ(x_dims, mask_dims,
+                      "Dimensions of Input(X) and Mask must be the same.");
+
+    auto *x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
+    x_grad->Resize(x_dims);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(dropout, ops::DropoutOp, ops::DropoutOpMaker<float>, dropout_grad,
+            ops::DropoutOpGrad<float>);
+REGISTER_OP_CPU_KERNEL(
+    dropout, ops::CPUDropoutKernel<paddle::platform::CPUPlace, float, float>);
+REGISTER_OP_CPU_KERNEL(
+    dropout_grad, ops::DropoutGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/dropout_op.cu b/paddle/operators/dropout_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..186237fb238add37f32403309a0f7e8a9846d335
--- /dev/null
+++ b/paddle/operators/dropout_op.cu
@@ -0,0 +1,86 @@
+/* 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. */
+
+#define EIGEN_USE_GPU
+#include <thrust/device_ptr.h>
+#include <thrust/iterator/counting_iterator.h>
+#include <thrust/random.h>
+#include <thrust/transform.h>
+#include "paddle/operators/dropout_op.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T, typename AttrType>
+struct MaskGenerator {
+  AttrType dropout_prob;
+  int seed;
+
+  __host__ __device__ MaskGenerator(AttrType dropout_prob, int seed)
+      : dropout_prob(dropout_prob), seed(seed) {}
+
+  __host__ __device__ T operator()(const unsigned int n) const {
+    thrust::minstd_rand rng;
+    rng.seed(seed);
+    thrust::uniform_real_distribution<AttrType> dist(0, 1);
+    rng.discard(n);
+    if (dist(rng) < dropout_prob) {
+      return static_cast<T>(0);
+    } else {
+      return static_cast<T>(1);
+    }
+  }
+};
+
+// It seems that Eigen::Tensor::setRandom in GPU will SEGFAULT.
+// Use std::random and thrust::random(thrust is a std library in CUDA) to
+// implement uniform random.
+template <typename Place, typename T, typename AttrType>
+class GPUDropoutKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<Tensor>("X");
+    auto* y = context.Output<Tensor>("Out");
+    y->mutable_data<T>(context.GetPlace());
+    AttrType dropout_prob = context.Attr<AttrType>("dropout_prob");
+
+    auto X = EigenMatrix<T>::Reshape(*x, 1);
+    auto Y = EigenMatrix<T>::Reshape(*y, 1);
+
+    auto place = context.GetEigenDevice<Place>();
+    if (context.Attr<int>("is_training") == 1) {
+      auto* mask = context.Output<Tensor>("Mask");
+      auto* mask_data = mask->mutable_data<T>(context.GetPlace());
+      int size = framework::product(mask->dims());
+      int seed = context.Attr<int>("seed");
+      thrust::counting_iterator<unsigned int> index_sequence_begin(0);
+      thrust::transform(index_sequence_begin, index_sequence_begin + size,
+                        thrust::device_ptr<T>(mask_data),
+                        MaskGenerator<T, AttrType>(dropout_prob, seed));
+      auto M = EigenMatrix<T>::Reshape(*mask, 1);
+      Y.device(place) = X * M;
+    } else {
+      Y.device(place) = X * dropout_prob;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    dropout, ops::GPUDropoutKernel<paddle::platform::GPUPlace, float, float>);
+REGISTER_OP_GPU_KERNEL(
+    dropout_grad, ops::DropoutGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/dropout_op.h b/paddle/operators/dropout_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..82eafee0e0e7db7b4b4ae5405f37146d061aefd5
--- /dev/null
+++ b/paddle/operators/dropout_op.h
@@ -0,0 +1,86 @@
+/* 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 <random>
+#include "paddle/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+
+template <typename Place, typename T, typename AttrType>
+class CPUDropoutKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<Tensor>("X");
+    auto* y = context.Output<Tensor>("Out");
+    const auto* x_data = x->data<T>();
+    auto* y_data = y->mutable_data<T>(context.GetPlace());
+    AttrType dropout_prob = context.Attr<AttrType>("dropout_prob");
+
+    if (context.Attr<int>("is_training") == 1) {
+      auto* mask = context.Output<Tensor>("Mask");
+      auto* mask_data = mask->mutable_data<T>(context.GetPlace());
+      int seed = context.Attr<int>("seed");
+      std::minstd_rand engine;
+      engine.seed(seed);
+      std::uniform_real_distribution<AttrType> dist(0, 1);
+      size_t size = framework::product(mask->dims());
+      for (size_t i = 0; i < size; ++i) {
+        if (dist(engine) < dropout_prob) {
+          mask_data[i] = 0;
+          y_data[i] = 0;
+        } else {
+          mask_data[i] = 1;
+          y_data[i] = x_data[i];
+        }
+      }
+    } else {
+      auto X = EigenMatrix<T>::Reshape(*x, 1);
+      auto Y = EigenMatrix<T>::Reshape(*y, 1);
+      auto place = context.GetEigenDevice<Place>();
+      Y.device(place) = X * dropout_prob;
+    }
+  }
+};
+
+template <typename Place, typename T>
+class DropoutGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    PADDLE_ENFORCE_EQ(context.Attr<int>("is_training"), 1,
+                      "GradOp is only callable when is_training is true");
+
+    auto* grad_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* grad_y = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* mask = context.Input<Tensor>("Mask");
+    grad_x->mutable_data<T>(context.GetPlace());
+
+    auto M = EigenMatrix<T>::Reshape(*mask, 1);
+    auto dX = EigenMatrix<T>::Reshape(*grad_x, 1);
+    auto dY = EigenMatrix<T>::Reshape(*grad_y, 1);
+
+    auto place = context.GetEigenDevice<Place>();
+    dX.device(place) = dY * M;
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc
index e37c582adbe5b9e728f683d97cc51063ce80c3a2..ee6e975b443691bf71cec904565ced20406f3fba 100644
--- a/paddle/operators/elementwise_mul_op.cc
+++ b/paddle/operators/elementwise_mul_op.cc
@@ -25,8 +25,14 @@ class ElementWiseMulOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ElementWiseMulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of ElementWiseMulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of ElementWiseMulOp should not be null.");
+
     auto x_dim = ctx.Input<Tensor>("X")->dims();
     auto y_dim = ctx.Input<Tensor>("Y")->dims();
     PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
diff --git a/paddle/operators/elementwise_mul_op.h b/paddle/operators/elementwise_mul_op.h
index e9ed6791799240039f9af42c1a4339be7126ee65..6d58da580b81b9e0a8ae170eec1a73638b190df8 100644
--- a/paddle/operators/elementwise_mul_op.h
+++ b/paddle/operators/elementwise_mul_op.h
@@ -13,10 +13,8 @@
    limitations under the License. */
 
 #pragma once
-#include <iostream>
 #include "paddle/framework/eigen.h"
 #include "paddle/framework/op_registry.h"
-#include "paddle/operators/math/math_function.h"
 
 namespace paddle {
 namespace operators {
diff --git a/paddle/operators/fc_op.cc b/paddle/operators/fc_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..e5d0f3c3724262a60a463ef3beadd9906d3ebaf6
--- /dev/null
+++ b/paddle/operators/fc_op.cc
@@ -0,0 +1,197 @@
+/* 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/op_registry.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+class FCOp : public NetOp {
+ public:
+  FCOp(const std::string &type, const framework::VariableNameMap &inputs,
+       const framework::VariableNameMap &outputs,
+       const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    PADDLE_ENFORCE(!Inputs("X").empty(),
+                   "Inputs(X) of FCOp should not be null.");
+    PADDLE_ENFORCE(!Inputs("W").empty(),
+                   "Inputs(W) of FCOp should not be null.");
+    PADDLE_ENFORCE(!Outputs("MulOut").empty(),
+                   "Outputs(MulOut) of FCOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("Out"), framework::kEmptyVarName,
+                      "Output(Out) of FCOp should not be null.");
+
+    auto x = Inputs("X");
+    auto w = Inputs("W");
+    auto mul_out = Outputs("MulOut");
+    PADDLE_ENFORCE_EQ(
+        x.size(), w.size(),
+        "The size of inputs X(%d) should be the same as that of weights W(%d).",
+        x.size(), w.size());
+    PADDLE_ENFORCE_EQ(mul_out.size(), x.size(),
+                      "The size of intermediate mul_out(%d) should be the same "
+                      "as that of inputs X(%d).",
+                      mul_out.size(), x.size());
+
+    size_t n = x.size();
+    PADDLE_ENFORCE_GE(n, static_cast<size_t>(1),
+                      "The size of inputs X(%d) should be no less than 1.", n);
+
+    auto x_num_col_dims = Attr<std::vector<int>>("xNumColDims");
+
+    // Set all values or set no values (use the default value)
+    if (!x_num_col_dims.empty()) {
+      PADDLE_ENFORCE_EQ(x_num_col_dims.size(), n,
+                        "The size of attribute xNumColDims(%d) should be the "
+                        "same as that of inputs X(%d).",
+                        x_num_col_dims.size(), n);
+    } else {
+      x_num_col_dims.resize(n);
+      for (size_t i = 0; i < n; i++) {
+        x_num_col_dims[i] = 1;
+      }
+    }
+
+    // mul_out[i] = X[i] * W[i]
+    for (size_t i = 0; i < n; i++) {
+      framework::AttributeMap mul_attr;
+      mul_attr["x_num_col_dims"] = static_cast<int>(x_num_col_dims[i]);
+      mul_attr["y_num_col_dims"] = static_cast<int>(1);
+      AppendOp(
+          framework::OpRegistry::CreateOp("mul", {{"X", {x[i]}}, {"Y", {w[i]}}},
+                                          {{"Out", {mul_out[i]}}}, mul_attr));
+    }
+
+    // sum_out = X[0] * W[0] + ... + X[n-1] * W[n-1]
+    auto sum_out = mul_out[0];
+    if (n > 1) {
+      PADDLE_ENFORCE_NE(Output("SumOut"), framework::kEmptyVarName,
+                        "Output(SumOut) of FCOp should not be null when the "
+                        "size of Inputs(X) > 1.");
+
+      sum_out = Output("SumOut");
+      AppendOp(framework::OpRegistry::CreateOp("sum", {{"X", {mul_out}}},
+                                               {{"Out", {sum_out}}}, {}));
+    } else {
+      if (Output("SumOut") != framework::kEmptyVarName) {
+        this->Rename(Output("SumOut"), framework::kEmptyVarName);
+      }
+    }
+
+    // add_out = sum_out + b
+    auto b = Input("B");
+    auto add_out = sum_out;
+    if (b != framework::kEmptyVarName) {
+      PADDLE_ENFORCE_NE(
+          Output("AddOut"), framework::kEmptyVarName,
+          "Output(AddOut) of FCOp should not be null when Input(B) is set.");
+
+      add_out = Output("AddOut");
+      AppendOp(framework::OpRegistry::CreateOp(
+          "rowwise_add", {{"X", {sum_out}}, {"b", {Input("B")}}},
+          {{"Out", {add_out}}}, {}));
+    } else {
+      if (Output("AddOut") != framework::kEmptyVarName) {
+        this->Rename(Output("AddOut"), framework::kEmptyVarName);
+      }
+    }
+
+    auto activation = Attr<std::string>("activation");
+    AppendOp(framework::OpRegistry::CreateOp(activation, {{"X", {add_out}}},
+                                             {{"Y", {Output("Out")}}}, {}));
+    CompleteAddOp(false);
+  }
+};
+
+class FCOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  FCOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(A vector of Tensors) each input Tensor can be of arbitrary "
+             "dimension, and will be reshaped to a 2-D matrix of size "
+             "(minibatch, number_of_input_features) according to attribute "
+             "xNumColDims.")
+        .AsDuplicable();
+    AddInput("W",
+             "(A vector of Tensors) the weights of FC operator, a "
+             "vector of 2-D matrix of size "
+             "(number_of_input_features, number_of_neurons).")
+        .AsDuplicable();
+    AddInput("B",
+             "(Tensor) the bias of FC operator, a 1-D vector of size "
+             "number_of_neurons.");
+
+    AddOutput("Out",
+              "(Tensor) the activated output matrix of FC operator, a 2-D "
+              "matrix of size (minibatch, number_of_neurons).");
+    AddOutput("MulOut",
+              "(A vector of Tensors) the intermediate outputs of FC operator, "
+              "each Tensor saving the product of X_i * W_i.")
+        .AsIntermediate()
+        .AsDuplicable();
+    AddOutput(
+        "SumOut",
+        "(Tensor) the intermediate output of FC operator, "
+        "saving the sum of the products of X and W, that is sum{X_i * W_i}.")
+        .AsIntermediate();
+    AddOutput("AddOut",
+              "(Tensor) the non-actived output of FC operator, "
+              "saving sum{X_i * W_i} + B.")
+        .AsIntermediate();
+    AddAttr<std::string>(
+        "activation",
+        "(string, default identity) the activation type of FC operator.")
+        .SetDefault("identity")
+        .InEnum({"identity", "sigmoid", "softmax"});
+    AddAttr<std::vector<int>>(
+        "xNumColDims",
+        "(std::vector<int>) The inputs Tensors of FC operator can be of "
+        "more than 2 dimensions. In that case, each input Tensor `X_i` will be "
+        "reshaped to a 2-D matrix. The matrix's first dimension "
+        "(the length of column) will be the product of `X_i`'s last "
+        "`xNumColDims_i` dimensions, that is "
+        "`X_i.dims[0] x ... x X_i.dims[xNumColDims_i - 1]`. "
+        "The matrix's second dimension (the length of row) will be the product "
+        "of `X_i`'s first `rank - xNumColDims_i` dimensions, that is "
+        "`X_i.dims[xNumColDims_i] x ... x X_i.dims[rank - 1]`)")
+        .SetDefault(std::vector<int>{});
+
+    AddComment(R"DOC(
+Fully Connected Operator, known as Fully Connected Layer or Inner Product Layer
+in Convolutional Neural Networks. Neurons in a fully connected layer have
+full connections to all activations in the previous layer.
+It computes an inner product of a set of
+learned weights with a matrix multiplication followed by a bias offset
+(optionally).
+
+Equation:
+  Out = Act(sum_n{X_i * W_i} + B)
+
+where X_i is Tensor that will be reshaped to a 2-D matrix of size (M x K),
+usually M is the minibatch size and K is the number of input features.
+W_i is a 2-D matrix of size (K x N), where N means the number of neurons
+in the fully connected layer. B is a 1-D vector of size N.
+Thus, the output Out is a 2-D matrix of size (M x N).
+Activation type can be set to `identity` (default), `sigmoid` or `softmax`.
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_WITHOUT_GRADIENT(fc, ops::FCOp, ops::FCOpMaker);
diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc
index 0c9734892aac216709d380ec66acadf792761b14..ba7857cc65f6860a6156674c6addc2bfdce21a99 100644
--- a/paddle/operators/fill_zeros_like_op.cc
+++ b/paddle/operators/fill_zeros_like_op.cc
@@ -23,6 +23,13 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Src"),
+        "Input(Src) of FillZerosLikeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Dst"),
+        "Output(Dst) of FillZerosLikeOp should not be null.");
+
     ctx.Output<framework::LoDTensor>("Dst")->Resize(
         ctx.Input<framework::Tensor>("Src")->dims());
   }
diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc
index 8883d6d5fed2d8900364cf713a1e8d8b290ef83e..d445b61c1657356f2cdcf1e98d756607de2bd042 100644
--- a/paddle/operators/gather_op.cc
+++ b/paddle/operators/gather_op.cc
@@ -24,6 +24,13 @@ class GatherOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of GatherOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
+                            "Input(Index) of GatherOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of GatherOp should not be null.");
+
     int batch_size = ctx.Input<Tensor>("Index")->dims()[0];
     PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
     framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());
diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc
index 25b0776a37488e876b0cc88aa3f2aa68e33fb270..c0e161bbc0c5486eb10408e43e6388f1b287abf8 100644
--- a/paddle/operators/gaussian_random_op.cc
+++ b/paddle/operators/gaussian_random_op.cc
@@ -43,8 +43,12 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext& context) const override {
-    auto* tensor = context.Output<framework::LoDTensor>("Out");
+  void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of GaussianRandomOp should not be null.");
+
+    auto* tensor = ctx.Output<framework::LoDTensor>("Out");
     auto dims = Attr<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());
diff --git a/paddle/operators/identity_op.cc b/paddle/operators/identity_op.cc
index 7d9d4fa519d1c690feacbadc5175aeab49082282..2cc632205e63abbe412b09af4b894420ac512ec5 100644
--- a/paddle/operators/identity_op.cc
+++ b/paddle/operators/identity_op.cc
@@ -27,7 +27,7 @@ class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
                   framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input tensor of identity operator.");
-    AddOutput("Out", "The output tensor of identity operator.");
+    AddOutput("Y", "The output tensor of identity operator.");
     AddComment(R"DOC(
 The identity operator is an alias of the scale operator
 with the attribute scale fixed to 1.0.
@@ -42,9 +42,15 @@ class IdentityOp : public NetOp {
              const framework::VariableNameMap &outputs,
              const framework::AttributeMap &attrs)
       : NetOp(type, inputs, outputs, attrs) {
+    PADDLE_ENFORCE_NE(Input("X"), framework::kEmptyVarName,
+                      "Input(X) of IdentityOp should not be null.");
+    PADDLE_ENFORCE_NE(Output("Y"), framework::kEmptyVarName,
+                      "Output(Y) of IdentityOp should not be null.");
+
     AppendOp(framework::OpRegistry::CreateOp(
-        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
+        "scale", {{"X", {Input("X")}}}, {{"Out", {Output("Y")}}},
         {{"scale", static_cast<AttrType>(1)}}));
+    CompleteAddOp(false);
   }
 };
 
diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc
index b3d15f1ec99813d242c86c99faa7385795eef3b1..07f6dfabca5879e3de6004e59d2e87f7fa68d66c 100644
--- a/paddle/operators/lookup_table_op.cc
+++ b/paddle/operators/lookup_table_op.cc
@@ -22,10 +22,17 @@ class LookupTableOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &context) const override {
-    auto table_t = context.Input<Tensor>("W");
-    auto ids_t = context.Input<Tensor>("Ids");
-    auto output_t = context.Output<framework::LoDTensor>("Out");
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("W"),
+                            "Input(W) of LookupTableOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ids"),
+                            "Input(Ids) of LookupTableOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of LookupTableOp should not be null.");
+
+    auto table_t = ctx.Input<Tensor>("W");
+    auto ids_t = ctx.Input<Tensor>("Ids");
+    auto output_t = ctx.Output<framework::LoDTensor>("Out");
 
     output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
   }
diff --git a/paddle/operators/math/math_function.cc b/paddle/operators/math/math_function.cc
index 1e86fc3d166077265e0f433a6712b0665ea5a152..def4b01da098fc960ce7c0e497732fbcc2579945 100644
--- a/paddle/operators/math/math_function.cc
+++ b/paddle/operators/math/math_function.cc
@@ -19,12 +19,13 @@ namespace operators {
 namespace math {
 
 template <>
-void gemm<platform::CPUPlace, float>(const CBLAS_TRANSPOSE transA,
+void gemm<platform::CPUPlace, float>(const platform::DeviceContext& context,
+                                     const CBLAS_TRANSPOSE transA,
                                      const CBLAS_TRANSPOSE transB, const int M,
                                      const int N, const int K,
                                      const float alpha, const float* A,
-                                     const float* B, const float beta, float* C,
-                                     platform::DeviceContext* context) {
+                                     const float* B, const float beta,
+                                     float* C) {
   int lda = (transA == CblasNoTrans) ? K : M;
   int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
@@ -33,13 +34,13 @@ void gemm<platform::CPUPlace, float>(const CBLAS_TRANSPOSE transA,
 }
 
 template <>
-void gemm<platform::CPUPlace, double>(const CBLAS_TRANSPOSE transA,
+void gemm<platform::CPUPlace, double>(const platform::DeviceContext& context,
+                                      const CBLAS_TRANSPOSE transA,
                                       const CBLAS_TRANSPOSE transB, const int M,
                                       const int N, const int K,
                                       const double alpha, const double* A,
                                       const double* B, const double beta,
-                                      double* C,
-                                      platform::DeviceContext* context) {
+                                      double* C) {
   int lda = (transA == CblasNoTrans) ? K : M;
   int ldb = (transB == CblasNoTrans) ? N : K;
   int ldc = N;
@@ -48,13 +49,10 @@ void gemm<platform::CPUPlace, double>(const CBLAS_TRANSPOSE transA,
 }
 
 template <>
-void matmul<platform::CPUPlace, float>(const framework::Tensor& matrix_a,
-                                       bool trans_a,
-                                       const framework::Tensor& matrix_b,
-                                       bool trans_b, float alpha,
-                                       framework::Tensor* matrix_out,
-                                       float beta,
-                                       platform::DeviceContext* context) {
+void matmul<platform::CPUPlace, float>(
+    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
+    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha,
+    framework::Tensor* matrix_out, float beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -74,18 +72,15 @@ void matmul<platform::CPUPlace, float>(const framework::Tensor& matrix_a,
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
   gemm<platform::CPUPlace, float>(
-      transA, transB, M, N, K, alpha, matrix_a.data<float>(),
-      matrix_b.data<float>(), beta, matrix_out->data<float>(), context);
+      context, transA, transB, M, N, K, alpha, matrix_a.data<float>(),
+      matrix_b.data<float>(), beta, matrix_out->data<float>());
 }
 
 template <>
-void matmul<platform::CPUPlace, double>(const framework::Tensor& matrix_a,
-                                        bool trans_a,
-                                        const framework::Tensor& matrix_b,
-                                        bool trans_b, double alpha,
-                                        framework::Tensor* matrix_out,
-                                        double beta,
-                                        platform::DeviceContext* context) {
+void matmul<platform::CPUPlace, double>(
+    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
+    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha,
+    framework::Tensor* matrix_out, double beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -105,8 +100,8 @@ void matmul<platform::CPUPlace, double>(const framework::Tensor& matrix_a,
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
   gemm<platform::CPUPlace, double>(
-      transA, transB, M, N, K, alpha, matrix_a.data<double>(),
-      matrix_b.data<double>(), beta, matrix_out->data<double>(), context);
+      context, transA, transB, M, N, K, alpha, matrix_a.data<double>(),
+      matrix_b.data<double>(), beta, matrix_out->data<double>());
 }
 
 }  // namespace math
diff --git a/paddle/operators/math/math_function.cu b/paddle/operators/math/math_function.cu
index da40b27c948918e4997f4a046d2145552296158b..71563b77b4b262c3f1e17ae7c4381da56ba780a3 100644
--- a/paddle/operators/math/math_function.cu
+++ b/paddle/operators/math/math_function.cu
@@ -19,12 +19,13 @@ namespace operators {
 namespace math {
 
 template <>
-void gemm<platform::GPUPlace, float>(const CBLAS_TRANSPOSE transA,
+void gemm<platform::GPUPlace, float>(const platform::DeviceContext& context,
+                                     const CBLAS_TRANSPOSE transA,
                                      const CBLAS_TRANSPOSE transB, const int M,
                                      const int N, const int K,
                                      const float alpha, const float* A,
-                                     const float* B, const float beta, float* C,
-                                     platform::DeviceContext* context) {
+                                     const float* B, const float beta,
+                                     float* C) {
   // Note that cublas follows fortran order, so the order is different from
   // the cblas convention.
   int lda = (transA == CblasNoTrans) ? K : M;
@@ -35,18 +36,19 @@ void gemm<platform::GPUPlace, float>(const CBLAS_TRANSPOSE transA,
       (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
 
   PADDLE_ENFORCE(platform::dynload::cublasSgemm(
-      reinterpret_cast<platform::CUDADeviceContext*>(context)->cublas_handle(),
+      reinterpret_cast<const platform::CUDADeviceContext&>(context)
+          .cublas_handle(),
       cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
 }
 
 template <>
-void gemm<platform::GPUPlace, double>(const CBLAS_TRANSPOSE transA,
+void gemm<platform::GPUPlace, double>(const platform::DeviceContext& context,
+                                      const CBLAS_TRANSPOSE transA,
                                       const CBLAS_TRANSPOSE transB, const int M,
                                       const int N, const int K,
                                       const double alpha, const double* A,
                                       const double* B, const double beta,
-                                      double* C,
-                                      platform::DeviceContext* context) {
+                                      double* C) {
   // Note that cublas follows fortran order, so the order is different from
   // the cblas convention.
   int lda = (transA == CblasNoTrans) ? K : M;
@@ -56,18 +58,16 @@ void gemm<platform::GPUPlace, double>(const CBLAS_TRANSPOSE transA,
   cublasOperation_t cuTransB =
       (transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
   PADDLE_ENFORCE(platform::dynload::cublasDgemm(
-      reinterpret_cast<platform::CUDADeviceContext*>(context)->cublas_handle(),
+      reinterpret_cast<const platform::CUDADeviceContext&>(context)
+          .cublas_handle(),
       cuTransB, cuTransA, N, M, K, &alpha, B, ldb, A, lda, &beta, C, N));
 }
 
 template <>
-void matmul<platform::GPUPlace, float>(const framework::Tensor& matrix_a,
-                                       bool trans_a,
-                                       const framework::Tensor& matrix_b,
-                                       bool trans_b, float alpha,
-                                       framework::Tensor* matrix_out,
-                                       float beta,
-                                       platform::DeviceContext* context) {
+void matmul<platform::GPUPlace, float>(
+    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
+    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, float alpha,
+    framework::Tensor* matrix_out, float beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -87,18 +87,15 @@ void matmul<platform::GPUPlace, float>(const framework::Tensor& matrix_a,
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
   gemm<platform::GPUPlace, float>(
-      transA, transB, M, N, K, alpha, matrix_a.data<float>(),
-      matrix_b.data<float>(), beta, matrix_out->data<float>(), context);
+      context, transA, transB, M, N, K, alpha, matrix_a.data<float>(),
+      matrix_b.data<float>(), beta, matrix_out->data<float>());
 }
 
 template <>
-void matmul<platform::GPUPlace, double>(const framework::Tensor& matrix_a,
-                                        bool trans_a,
-                                        const framework::Tensor& matrix_b,
-                                        bool trans_b, double alpha,
-                                        framework::Tensor* matrix_out,
-                                        double beta,
-                                        platform::DeviceContext* context) {
+void matmul<platform::GPUPlace, double>(
+    const platform::DeviceContext& context, const framework::Tensor& matrix_a,
+    bool trans_a, const framework::Tensor& matrix_b, bool trans_b, double alpha,
+    framework::Tensor* matrix_out, double beta) {
   auto dim_a = matrix_a.dims();
   auto dim_b = matrix_b.dims();
   auto dim_out = matrix_out->dims();
@@ -118,8 +115,8 @@ void matmul<platform::GPUPlace, double>(const framework::Tensor& matrix_a,
   CBLAS_TRANSPOSE transB = (trans_b == false) ? CblasNoTrans : CblasTrans;
 
   gemm<platform::GPUPlace, double>(
-      transA, transB, M, N, K, alpha, matrix_a.data<double>(),
-      matrix_b.data<double>(), beta, matrix_out->data<double>(), context);
+      context, transA, transB, M, N, K, alpha, matrix_a.data<double>(),
+      matrix_b.data<double>(), beta, matrix_out->data<double>());
 }
 
 }  // namespace math
diff --git a/paddle/operators/math/math_function.h b/paddle/operators/math/math_function.h
index 155589fadb3ed9f59160a750d546dd8093a56cbe..d8518e77fa7b4abdbcf08b7983013c24806e14ca 100644
--- a/paddle/operators/math/math_function.h
+++ b/paddle/operators/math/math_function.h
@@ -66,16 +66,16 @@ namespace math {
 // For more detailed info, please refer to
 // http://www.netlib.org/lapack/explore-html/d4/de2/sgemm_8f.html
 template <typename Place, typename T>
-void gemm(const CBLAS_TRANSPOSE transA, const CBLAS_TRANSPOSE transB,
-          const int M, const int N, const int K, const T alpha, const T* A,
-          const T* B, const T beta, T* C, platform::DeviceContext* context);
+void gemm(const platform::DeviceContext& context, const CBLAS_TRANSPOSE transA,
+          const CBLAS_TRANSPOSE transB, const int M, const int N, const int K,
+          const T alpha, const T* A, const T* B, const T beta, T* C);
 
 // matrix multiply with continuous memory
 template <typename Place, typename T>
-void matmul(const framework::Tensor& matrix_a, bool trans_a,
+void matmul(const platform::DeviceContext& context,
+            const framework::Tensor& matrix_a, bool trans_a,
             const framework::Tensor& matrix_b, bool trans_b, T alpha,
-            framework::Tensor* matrix_out, T beta,
-            platform::DeviceContext* context);
+            framework::Tensor* matrix_out, T beta);
 
 }  // namespace math
 }  // namespace operators
diff --git a/paddle/operators/math/math_function_test.cc b/paddle/operators/math/math_function_test.cc
index 6c020c4ff7285b43bc5836d80c173d3a068e72b3..7e339457f7f08ff16162f399064a4b4dca594d7f 100644
--- a/paddle/operators/math/math_function_test.cc
+++ b/paddle/operators/math/math_function_test.cc
@@ -15,8 +15,7 @@ TEST(math_function, notrans_mul_trans) {
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
   auto* gpu_place = new paddle::platform::GPUPlace(0);
-  paddle::platform::DeviceContext* context =
-      new paddle::platform::CUDADeviceContext(*gpu_place);
+  paddle::platform::CUDADeviceContext context(*gpu_place);
 
   input1_gpu.CopyFrom<float>(input1, *gpu_place);
   input2_gpu.CopyFrom<float>(input1, *gpu_place);
@@ -24,7 +23,7 @@ TEST(math_function, notrans_mul_trans) {
   out_gpu.mutable_data<float>({2, 2}, *gpu_place);
 
   paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
-      input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0, context);
+      context, input1_gpu, false, input2_gpu, true, 1, &out_gpu, 0);
 
   out.CopyFrom<float>(out_gpu, *cpu_place);
 
@@ -33,6 +32,7 @@ TEST(math_function, notrans_mul_trans) {
   EXPECT_EQ(out_ptr[1], 14);
   EXPECT_EQ(out_ptr[2], 14);
   EXPECT_EQ(out_ptr[3], 50);
+  delete gpu_place;
 }
 
 TEST(math_function, trans_mul_notrans) {
@@ -48,8 +48,7 @@ TEST(math_function, trans_mul_notrans) {
   memcpy(input1_ptr, arr, 6 * sizeof(float));
 
   auto* gpu_place = new paddle::platform::GPUPlace(0);
-  paddle::platform::DeviceContext* context =
-      new paddle::platform::CUDADeviceContext(*gpu_place);
+  paddle::platform::CUDADeviceContext context(*gpu_place);
 
   input1_gpu.CopyFrom<float>(input1, *gpu_place);
   input2_gpu.CopyFrom<float>(input1, *gpu_place);
@@ -57,7 +56,7 @@ TEST(math_function, trans_mul_notrans) {
   out_gpu.mutable_data<float>({3, 3}, *gpu_place);
 
   paddle::operators::math::matmul<paddle::platform::GPUPlace, float>(
-      input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0, context);
+      context, input1_gpu, true, input2_gpu, false, 1, &out_gpu, 0);
 
   out.CopyFrom<float>(out_gpu, *cpu_place);
 
@@ -71,5 +70,6 @@ TEST(math_function, trans_mul_notrans) {
   EXPECT_EQ(out_ptr[6], 15);
   EXPECT_EQ(out_ptr[7], 22);
   EXPECT_EQ(out_ptr[8], 29);
+  delete gpu_place;
 }
 #endif
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index 3e523d31b682d70825d50c2a57b6e98cbf29dcd3..7d7eeb59a23435036dc33c1e4fe6dd1c4a1a2f62 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -24,7 +24,9 @@ class MeanOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of MeanOp must be initialized.");
+                            "Input(X) of MeanOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MeanOp should not be null.");
     ctx.Output<framework::LoDTensor>("Out")->Resize({1});
   }
 };
diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc
index 8a583f24edf40ce805ca216ab014bd169f9236df..a97bbecdca1779df330d1053cf359bb658aa75c2 100644
--- a/paddle/operators/minus_op.cc
+++ b/paddle/operators/minus_op.cc
@@ -27,6 +27,13 @@ class MinusOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of MinusOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of MinusOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MinusOp should not be null.");
+
     auto *left_tensor = ctx.Input<framework::Tensor>("X");
     auto *right_tensor = ctx.Input<framework::Tensor>("Y");
 
@@ -64,7 +71,7 @@ class MinusGradOp : public NetOp {
 
     // x_grad = out_grad
     AppendOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}},
-                                             {{"Out", {x_grad}}}, {}));
+                                             {{"Y", {x_grad}}}, {}));
 
     framework::AttributeMap scale_attr;
     scale_attr["scale"] = static_cast<AttrType>(-1);
@@ -77,8 +84,6 @@ class MinusGradOp : public NetOp {
 }  // namespace operators
 }  // namespace paddle
 
-USE_OP(scale);
-USE_NO_KERNEL_OP(identity);
 namespace ops = paddle::operators;
 REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad,
             ops::MinusGradOp<float>);
diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc
index 015e13de9a09bcd1931ccf91413b6a3f484f82bb..b6d320b415e02549e85cb36ab517b0b5433887d5 100644
--- a/paddle/operators/mul_op.cc
+++ b/paddle/operators/mul_op.cc
@@ -26,6 +26,13 @@ class MulOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of MulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
+                            "Input(Y) of MulOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of MulOp should not be null.");
+
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto y_dims = ctx.Input<Tensor>("Y")->dims();
     int x_num_col_dims = Attr<int>("x_num_col_dims");
diff --git a/paddle/operators/mul_op.h b/paddle/operators/mul_op.h
index 3c01f868bda8cba488b3403df456d63d6b082fa6..ac7136a76933d1f3ead86518c65d589747227631 100644
--- a/paddle/operators/mul_op.h
+++ b/paddle/operators/mul_op.h
@@ -46,10 +46,8 @@ class MulKernel : public framework::OpKernel {
             : *y;
 
     z->mutable_data<T>(context.GetPlace());
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(context.device_context_);
-    math::matmul<Place, T>(x_matrix, false, y_matrix, false, 1, z, 0,
-                           device_context);
+    math::matmul<Place, T>(context.device_context(), x_matrix, false, y_matrix,
+                           false, 1, z, 0);
   }
 };
 
@@ -71,16 +69,14 @@ class MulGradKernel : public framework::OpKernel {
 
     Tensor* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
     Tensor* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
-    auto* device_context =
-        const_cast<platform::DeviceContext*>(ctx.device_context_);
     if (dx) {
       dx->mutable_data<T>(ctx.GetPlace());
       Tensor dx_matrix = dx->dims().size() > 2 ? framework::ReshapeToMatrix<T>(
                                                      *dx, x_num_col_dims)
                                                : *dx;
       // dx = dout * y'. dx: M x K, dout : M x N, y : K x N
-      math::matmul<Place, T>(*dout, false, y_matrix, true, 1, &dx_matrix, 0,
-                             device_context);
+      math::matmul<Place, T>(ctx.device_context(), *dout, false, y_matrix, true,
+                             1, &dx_matrix, 0);
     }
     if (dy) {
       dy->mutable_data<T>(ctx.GetPlace());
@@ -88,8 +84,8 @@ class MulGradKernel : public framework::OpKernel {
                                                      *dy, y_num_col_dims)
                                                : *dy;
       // dy = x' * dout. dy K x N, dout : M x N, x : M x K
-      math::matmul<Place, T>(x_matrix, true, *dout, false, 1, &dy_matrix, 0,
-                             device_context);
+      math::matmul<Place, T>(ctx.device_context(), x_matrix, true, *dout, false,
+                             1, &dy_matrix, 0);
     }
   }
 };
diff --git a/paddle/operators/onehot_cross_entropy_op.cc b/paddle/operators/onehot_cross_entropy_op.cc
deleted file mode 100644
index a9baada1cd4cd3af793bf1b0af7b029417e62b08..0000000000000000000000000000000000000000
--- a/paddle/operators/onehot_cross_entropy_op.cc
+++ /dev/null
@@ -1,75 +0,0 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
-
-#include "paddle/operators/onehot_cross_entropy_op.h"
-
-namespace paddle {
-namespace operators {
-
-class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
- protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto *X = ctx.Input<Tensor>("X");
-    auto *label = ctx.Input<Tensor>("label");
-
-    PADDLE_ENFORCE_EQ(X->dims().size(), 2, "X's dimension must be 2.");
-    PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1.");
-    PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]);
-    ctx.Output<framework::LoDTensor>("Y")->Resize({X->dims()[0], 1});
-  }
-};
-
-class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
- protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto dX = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
-    auto X = ctx.Input<Tensor>("X");
-
-    dX->Resize(X->dims());
-  }
-};
-
-class OnehotCrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
- public:
-  OnehotCrossEntropyOpMaker(framework::OpProto *proto,
-                            framework::OpAttrChecker *op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The first input of OnehotCrossEntropyOp");
-    AddInput("label", "The second input of OnehotCrossEntropyOp");
-    AddOutput("Y", "The output of OnehotCrossEntropyOp");
-    AddComment(R"DOC(
-OnehotCrossEntropy Operator.
-
-                Y[i] = -log(X[i][j])
-
-)DOC");
-  }
-};
-}  // namespace operators
-}  // namespace paddle
-
-namespace ops = paddle::operators;
-REGISTER_OP(onehot_cross_entropy, ops::OnehotCrossEntropyOp,
-            ops::OnehotCrossEntropyOpMaker, onehot_cross_entropy_grad,
-            ops::OnehotCrossEntropyGradientOp);
-REGISTER_OP_CPU_KERNEL(onehot_cross_entropy,
-                       ops::OnehotCrossEntropyOpKernel<float>);
-REGISTER_OP_CPU_KERNEL(onehot_cross_entropy_grad,
-                       ops::OnehotCrossEntropyGradientOpKernel<float>);
diff --git a/paddle/operators/onehot_cross_entropy_op.cu b/paddle/operators/onehot_cross_entropy_op.cu
deleted file mode 100644
index d999bfce58c8a6db5c811aad677c07094b881841..0000000000000000000000000000000000000000
--- a/paddle/operators/onehot_cross_entropy_op.cu
+++ /dev/null
@@ -1,133 +0,0 @@
-/* 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/op_registry.h"
-#include "paddle/platform/assert.h"
-
-namespace paddle {
-namespace operators {
-
-using Tensor = framework::Tensor;
-
-template <typename T>
-__host__ __device__ T clipping_log(const T x) {
-  PADDLE_ASSERT(std::is_floating_point<T>::value);
-  const T kApproInf = 1e20;
-  T v = log(x);
-  if (v == INFINITY) {
-    return kApproInf;
-  }
-  if (v == -INFINITY) {
-    return -kApproInf;
-  }
-  return v;
-}
-
-template <typename T>
-__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label,
-                                   const int N, const int D) {
-  // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file.
-  // CUDA_1D_KERNEL_LOOP(i, N) {
-  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
-       i += blockDim.x * gridDim.x) {
-    PADDLE_ASSERT(label[i] >= 0 && label[i] < D);
-    Y[i] = -clipping_log(X[i * D + label[i]]);
-  }
-}
-
-// TODO(qingqing): make zero setting an common function.
-template <typename T>
-__global__ void zero(T* X, const int N) {
-  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
-       i += blockDim.x * gridDim.x) {
-    X[i] = 0.0;
-  }
-}
-
-template <typename T>
-__global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X,
-                                           const int* label, const int N,
-                                           const int D) {
-  // TOOD(qingqing) define CUDA_1D_KERNEL_LOOP macro in a common file.
-  // CUDA_1D_KERNEL_LOOP(i, N) {
-  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
-       i += blockDim.x * gridDim.x) {
-    int idx = i * D + label[i];
-    dX[idx] = -dY[i] / X[idx];
-  }
-}
-
-template <typename T>
-class OnehotCrossEntropyOpCUDAKernel : public framework::OpKernel {
- public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
-
-    auto X = ctx.Input<Tensor>("X");
-    const T* Xdata = X->data<T>();
-    const int* label_data = ctx.Input<Tensor>("label")->data<int>();
-    auto Y = ctx.Output<Tensor>("Y");
-    Y->mutable_data<T>(ctx.GetPlace());
-    T* Ydata = Y->data<T>();
-
-    int N = X->dims()[0];
-    int D = X->dims()[1];
-    int block = 512;
-    int grid = (N + block - 1) / block;
-    // TODO(qingqing) launch kernel on specified stream
-    // base on ExecutionContext.
-    CrossEntropyKernel<T><<<grid, block>>>(Ydata, Xdata, label_data, N, D);
-  }
-};
-
-template <typename T>
-class OnehotCrossEntropyGradientOpCUDAKernel : public framework::OpKernel {
- public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
-                   "It must use GPUPlace.");
-
-    auto X = ctx.Input<Tensor>("X");
-    auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto dY = ctx.Input<Tensor>(framework::GradVarName("Y"));
-    auto label = ctx.Input<Tensor>("label");
-
-    auto* dXdata = dX->template mutable_data<T>(ctx.GetPlace());
-    auto* dYdata = dY->template data<T>();
-    auto* Xdata = X->template data<T>();
-    auto* label_data = label->data<int>();
-
-    int N = X->dims()[0];
-    int D = X->dims()[1];
-    int block = 512;
-    int grid = (N * D + block - 1) / block;
-    zero<T><<<grid, block>>>(dXdata, N * D);
-
-    grid = (N + block - 1) / block;
-    // TODO(qingqing): launch kernel on specified stream
-    // base on ExecutionContext.
-    CrossEntropyGradientKernel<T><<<grid, block>>>(dXdata, dYdata, Xdata,
-                                                   label_data, N, D);
-  }
-};
-
-}  // namespace operators
-}  // namespace paddle
-
-namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
-                       ops::OnehotCrossEntropyOpCUDAKernel<float>);
-REGISTER_OP_GPU_KERNEL(onehot_cross_entropy_grad,
-                       ops::OnehotCrossEntropyGradientOpCUDAKernel<float>);
diff --git a/paddle/operators/onehot_cross_entropy_op.h b/paddle/operators/onehot_cross_entropy_op.h
deleted file mode 100644
index eb4d1348de1d940e2648c83c8ba94b289f10c5b2..0000000000000000000000000000000000000000
--- a/paddle/operators/onehot_cross_entropy_op.h
+++ /dev/null
@@ -1,98 +0,0 @@
-/* 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/op_registry.h"
-
-namespace paddle {
-namespace operators {
-
-using Tensor = framework::Tensor;
-
-template <typename T>
-inline T tolerable_value(const T x) {
-  static_assert(std::is_floating_point<T>::value,
-                "tolerable_value works only on float, "
-                "double and double double.");
-
-  const T kApproInf = 1e20;
-
-  if (x == INFINITY) {
-    return kApproInf;
-  }
-
-  if (x == -INFINITY) {
-    return -kApproInf;
-  }
-
-  return x;
-}
-
-template <typename T>
-class OnehotCrossEntropyOpKernel : public framework::OpKernel {
- public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
-                   "It must use CPUPlace.");
-
-    auto X = ctx.Input<Tensor>("X");
-    const T* Xdata = X->data<T>();
-    const int* label_data = ctx.Input<Tensor>("label")->data<int>();
-    auto Y = ctx.Output<Tensor>("Y");
-
-    Y->mutable_data<T>(ctx.GetPlace());
-
-    T* Ydata = Y->data<T>();
-
-    int batch_size = X->dims()[0];
-    int class_num = X->dims()[1];
-
-    for (int i = 0; i < batch_size; ++i) {
-      int index = i * class_num + label_data[i];
-      Ydata[i] = -tolerable_value(std::log(Xdata[index]));
-    }
-  }
-};
-
-template <typename T>
-class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel {
- public:
-  void Compute(const framework::ExecutionContext& ctx) const override {
-    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
-                   "It must use CPUPlace.");
-
-    auto X = ctx.Input<Tensor>("X");
-    auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
-    auto dY = ctx.Input<Tensor>(framework::GradVarName("Y"));
-    auto label = ctx.Input<Tensor>("label");
-
-    auto* dXdata = dX->template mutable_data<T>(ctx.GetPlace());
-    auto* dYdata = dY->template data<T>();
-    auto* Xdata = X->template data<T>();
-    auto* label_data = label->data<int>();
-
-    const int batch_size = X->dims()[0];
-    const int class_num = X->dims()[1];
-
-    // TODO(qingqing): make zero setting an common function.
-    memset(dXdata, 0, sizeof(T) * batch_size * class_num);
-    for (int i = 0; i < batch_size; ++i) {
-      int index = i * class_num + label_data[i];
-      dXdata[index] = -tolerable_value(dYdata[i] / Xdata[index]);
-    }
-  }
-};
-
-}  // namespace operators
-}  // namespace paddle
diff --git a/paddle/operators/pad_op.cc b/paddle/operators/pad_op.cc
index 6cf7bd6f35b7592c41983efd75c1628043070687..a0b1c6b631d97a40d774f7d2ff9550fda9c32db4 100644
--- a/paddle/operators/pad_op.cc
+++ b/paddle/operators/pad_op.cc
@@ -25,6 +25,11 @@ class PadOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of PadOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of PadOp should not be null.");
+
     auto x_dim = ctx.Input<Tensor>("X")->dims();
     auto paddings = Attr<std::vector<int>>("paddings");
     PADDLE_ENFORCE_EQ(x_dim.size() * 2, int64_t(paddings.size()),
diff --git a/paddle/operators/prelu_op.cc b/paddle/operators/prelu_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..7ae80b296850f2f433c89d904ebf32355b2a29c7
--- /dev/null
+++ b/paddle/operators/prelu_op.cc
@@ -0,0 +1,94 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/prelu_op.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+class PReluOp : public framework::OperatorWithKernel {
+ public:
+  PReluOp(const std::string &type, const framework::VariableNameMap &inputs,
+          const framework::VariableNameMap &outputs,
+          const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
+    auto *in = ctx.Input<framework::Tensor>("X");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Alpha"),
+                            "Input(Alpha) should not be null");
+    auto *alpha = ctx.Input<framework::Tensor>("Alpha");
+    PADDLE_ENFORCE(alpha->numel() == 1, "Size of weight Alpha must be one.");
+
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) should not be null");
+    auto *out = ctx.Output<framework::LoDTensor>("Out");
+    out->Resize(in->dims());
+  }
+};
+
+class PReluOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  PReluOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The input tensor of prelu operator.");
+    AddInput("Alpha", "The alpha weight of PRelu operator.");
+    AddOutput("Out", "The output tensor of PRelu operator.");
+    AddComment(R"DOC(PRelu operator
+
+The equation is:
+
+  f(x) = alpha * x , for x < 0
+  f(x) = x         , for x >= 0
+
+)DOC");
+  }
+};
+
+// The operator to calculate gradients of a prelu operator.
+class PReluGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
+                            "Input(Out@GRAD) should not be null");
+    auto *dx = ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
+    auto *x = ctx.Input<framework::Tensor>("X");
+
+    auto *dalpha =
+        ctx.Output<framework::LoDTensor>(framework::GradVarName("Alpha"));
+    auto *alpha = ctx.Input<framework::Tensor>("Alpha");
+
+    dx->Resize(x->dims());
+    dalpha->Resize(alpha->dims());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP(prelu, ops::PReluOp, ops::PReluOpMaker, prelu_grad,
+            ops::PReluGradOp);
+REGISTER_OP_CPU_KERNEL(prelu,
+                       ops::PReluKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(prelu_grad,
+                       ops::PReluGradKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/prelu_op.cu b/paddle/operators/prelu_op.cu
new file mode 100644
index 0000000000000000000000000000000000000000..9e391dabae735cc8a740b46b50d31d271f99b65d
--- /dev/null
+++ b/paddle/operators/prelu_op.cu
@@ -0,0 +1,21 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include "paddle/operators/prelu_op.h"
+
+REGISTER_OP_GPU_KERNEL(
+    prelu, paddle::operators::PReluKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    prelu_grad,
+    paddle::operators::PReluGradKernel<paddle::platform::GPUPlace, float>);
diff --git a/paddle/operators/prelu_op.h b/paddle/operators/prelu_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..3269116c112f115e1e8fbbee0dc3b81dbe736e69
--- /dev/null
+++ b/paddle/operators/prelu_op.h
@@ -0,0 +1,104 @@
+/* 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/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/platform/transform.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+using platform::Transform;
+
+template <typename T>
+class PReluFunctor {
+ public:
+  explicit PReluFunctor(const T* alpha) : alpha_(alpha) {}
+
+  HOSTDEVICE T operator()(const T& x) const {
+    if (x > 0)
+      return x;
+    else
+      return x * (*alpha_);
+  }
+
+ private:
+  const T* alpha_;
+};
+
+template <typename Place, typename T>
+class PReluKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<Tensor>("X");
+    auto* alpha = context.Input<Tensor>("Alpha");
+    auto* out = context.Output<Tensor>("Out");
+
+    const T* x_ptr = x->data<T>();
+    T* o_ptr = out->mutable_data<T>(context.GetPlace());
+
+    auto* alpha_ptr = alpha->data<T>();
+
+    int numel = x->numel();
+
+    Transform<Place> trans;
+    trans(context.device_context(), x_ptr, x_ptr + numel, o_ptr,
+          PReluFunctor<T>(alpha_ptr));
+  }
+};
+
+template <typename T>
+class PReluGradFunctor {
+ public:
+  explicit PReluGradFunctor(const T* alpha) : alpha_(alpha) {}
+
+  HOSTDEVICE T operator()(const T& out, const T& dout) const {
+    if (out > 0)
+      return dout;
+    else
+      return dout * (*alpha_);
+  }
+
+ private:
+  const T* alpha_;
+};
+
+template <typename Place, typename T>
+class PReluGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* dx = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* dout = context.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto* out = context.Input<Tensor>("Out");
+    auto* alpha = context.Input<Tensor>("Alpha");
+    auto* alpha_ptr = alpha->data<T>();
+
+    T* dx_ptr = dx->mutable_data<T>(context.GetPlace());
+    const T* dout_ptr = dout->data<T>();
+    const T* out_ptr = out->data<T>();
+    int numel = dx->numel();
+
+    Transform<Place> trans;
+    trans(context.device_context(), out_ptr, out_ptr + numel, dout_ptr, dx_ptr,
+          PReluGradFunctor<T>(alpha_ptr));
+
+    // TODO (Zhuoyuan): add dalpha upgrade when GPU kernels ready
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/reshape_op.cc b/paddle/operators/reshape_op.cc
index d2817020921dfd3c044922de6a0f2ae0307936bd..0d05e344148c68f5625dd819ec59c5991892e4ce 100644
--- a/paddle/operators/reshape_op.cc
+++ b/paddle/operators/reshape_op.cc
@@ -28,7 +28,11 @@ class ReshapeOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     // input check
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) shouldn't be null");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ReshapeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ReshapeOp should not be null.");
+
     auto shape = ctx.Attr<std::vector<int>>("shape");
     PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
     for (auto dim : shape) {
diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc
index c6101685a3205a7a7459347ea5b0cc8487656550..2a3fd3be941d91aaa6b014df91d3025f07767577 100644
--- a/paddle/operators/rowwise_add_op.cc
+++ b/paddle/operators/rowwise_add_op.cc
@@ -25,6 +25,13 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of RowwiseAddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("b"),
+                            "Input(b) of RowwiseAddOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of RowwiseAddOp should not be null.");
+
     auto x_dims = ctx.Input<Tensor>("X")->dims();
     auto b_dims = ctx.Input<Tensor>("b")->dims();
     PADDLE_ENFORCE_GT(
diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc
index 35e6b70ba94a645da2b99093b2354acfb0ef2771..d1f42e8662537d35e17429f9d436fdc0e5a1dc11 100644
--- a/paddle/operators/scale_op.cc
+++ b/paddle/operators/scale_op.cc
@@ -27,6 +27,11 @@ class ScaleOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of ScaleOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ScaleOp should not be null.");
+
     auto *in = ctx.Input<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc
index 0f7510983e2e34485110719c92d26cbc78cd850c..8820262732327306f4f807702751708bd1e2aa36 100644
--- a/paddle/operators/scatter_op.cc
+++ b/paddle/operators/scatter_op.cc
@@ -24,6 +24,15 @@ class ScatterOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Ref"),
+                            "Input(Ref) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
+                            "Input(Index) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Updates"),
+                            "Input(Updates) of ScatterOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of ScatterOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Index")->dims().size(), 1,
                       "Update Index should be 1-D.");
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("Ref")->dims().size(),
diff --git a/paddle/operators/sequence_avg_pool_op.cc b/paddle/operators/sequence_avg_pool_op.cc
index c15a5833deba2e198f6cb724bda7e3306c56e461..9815b8f3a8d813959949bbfedc79f404721a8216 100644
--- a/paddle/operators/sequence_avg_pool_op.cc
+++ b/paddle/operators/sequence_avg_pool_op.cc
@@ -23,9 +23,12 @@ class SequenceAvgPoolOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of SequenceAvgPoolOp"
-                            "must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"), "Input(X) of SequenceAvgPoolOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of SequenceAvgPoolOp should not be null.");
+
     auto* x = ctx.Input<framework::LoDTensor>("X");
     auto dims = x->dims();
     auto lod = x->lod();
@@ -60,7 +63,9 @@ class SequenceAvgPoolGradOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Gradient of Out should not be null");
+                            "Gradient of Out should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "The input X should not be null.");
     auto og_dims =
         ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->dims();
     auto x_dims = ctx.Input<framework::LoDTensor>("X")->dims();
diff --git a/paddle/operators/sequence_avg_pool_op.h b/paddle/operators/sequence_avg_pool_op.h
index 6e343b87e2938399409498407ac46b2416dc2231..ebe0956344eb71d0fb2836f1b4a989ac546d9f78 100644
--- a/paddle/operators/sequence_avg_pool_op.h
+++ b/paddle/operators/sequence_avg_pool_op.h
@@ -21,6 +21,9 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 using LoDTensor = framework::LoDTensor;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
@@ -43,8 +46,8 @@ class SequenceAvgPoolKernel : public framework::OpKernel {
                                  static_cast<int>(lod[0][i + 1]));
       Tensor out_t = out->Slice<T>(i, i + 1);
       int64_t h = static_cast<int64_t>(lod[0][i + 1] - lod[0][i]);
-      auto in_e = EigenMatrix<T>::From(in_t, {h, w});
-      auto out_e = EigenMatrix<T>::From(out_t, {h, w});
+      auto in_e = EigenMatrix<T>::From(in_t, framework::make_ddim({h, w}));
+      auto out_e = EigenVector<T>::Flatten(out_t);
       out_e.device(place) = in_e.mean(Eigen::array<int, 1>({{0}}));
     }
   }
@@ -54,9 +57,9 @@ template <typename Place, typename T>
 class SequenceAvgPoolGradKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* in = context.Output<LoDTensor>("X");
-    auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
+    auto* in = context.Input<LoDTensor>("X");
     auto* out_g = context.Input<LoDTensor>(framework::GradVarName("Out"));
+    auto* in_g = context.Output<LoDTensor>(framework::GradVarName("X"));
 
     auto dims = in->dims();
     auto lod = in->lod();
@@ -71,7 +74,7 @@ class SequenceAvgPoolGradKernel : public framework::OpKernel {
       int64_t h = static_cast<int64_t>(lod[0][i + 1] - lod[0][i]);
       auto in_g_e = EigenMatrix<T>::From(in_g_t, {h, w});
       auto out_g_e = EigenMatrix<T>::From(out_g_t, {1, w});
-      Eigen::DSizes<int, 2> bcast(h, w);
+      Eigen::DSizes<int, 2> bcast(h, 1);
       in_g_e.device(place) = (out_g_e / static_cast<T>(h)).broadcast(bcast);
     }
   }
diff --git a/paddle/operators/sgd_op.cc b/paddle/operators/sgd_op.cc
index 7997bf690710b3675f4014790db8cc7fc06946d3..1232e64c7f0132b9ea19b3d7e1ebe9531e1e25a5 100644
--- a/paddle/operators/sgd_op.cc
+++ b/paddle/operators/sgd_op.cc
@@ -23,6 +23,13 @@ class SGDOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("param"),
+                            "Input(param) of SGDOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("grad"),
+                            "Input(grad) of SGDOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("param_out"),
+                            "Output(param_out) of SGDOp should not be null.");
+
     PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("param")->dims(),
                       ctx.Input<Tensor>("grad")->dims(),
                       "Two input of SGD Op's dimension must be same.");
diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc
index de6a1ba77382306923ee238e5cf309c791f9f216..992b19965e0ca9ce7dba1b8b3c5b7780af06eb45 100644
--- a/paddle/operators/sigmoid_op.cc
+++ b/paddle/operators/sigmoid_op.cc
@@ -23,6 +23,11 @@ class SigmoidOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of SigmoidOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of SigmoidOp should not be null.");
+
     ctx.Output<framework::LoDTensor>("Y")->Resize(
         ctx.Input<Tensor>("X")->dims());
   }
diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc
index 239d3d141e1076a0a6a943f340311b17aa6f542a..c67eb028c882ed82ca4e6a4dd70cdea9f69cdc24 100644
--- a/paddle/operators/softmax_op.cc
+++ b/paddle/operators/softmax_op.cc
@@ -23,6 +23,11 @@ class SoftmaxOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of SoftmaxOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of SoftmaxOp should not be null.");
+
     PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
                    "The input of softmax op must be a matrix.");
     ctx.Output<framework::LoDTensor>("Y")->Resize(
diff --git a/paddle/operators/split_op.cc b/paddle/operators/split_op.cc
new file mode 100644
index 0000000000000000000000000000000000000000..61296f5c8122fdce7083e9a91dc313482875c805
--- /dev/null
+++ b/paddle/operators/split_op.cc
@@ -0,0 +1,118 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/operators/split_op.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+using framework::Tensor;
+
+class SplitOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    // infershape
+    auto *in = ctx.Input<framework::Tensor>("X");
+    auto outs = ctx.MultiOutput<framework::LoDTensor>("Out");
+    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    size_t num = static_cast<size_t>(ctx.Attr<int>("num"));
+    std::vector<int> sections =
+        static_cast<std::vector<int>>(ctx.Attr<std::vector<int>>("sections"));
+    const size_t n = outs.size();
+
+    if (num > 0) {
+      int64_t in_axis_dim = in->dims()[axis];
+      PADDLE_ENFORCE_EQ(in_axis_dim % num, 0,
+                        "tensor split does not result"
+                        " in an equal division");
+      size_t out_axis_dim = in_axis_dim / num;
+      for (size_t i = 0; i < n; ++i) {
+        auto dim = in->dims();
+        dim[axis] = out_axis_dim;
+        outs[i]->Resize(dim);
+      }
+    } else if (sections.size() > 0) {
+      PADDLE_ENFORCE_EQ(sections.size(), n,
+                        "tensor split sections size"
+                        "should be equal to output size.");
+      for (size_t i = 0; i < n; ++i) {
+        auto dim = in->dims();
+        dim[axis] = sections[i];
+        outs[i]->Resize(dim);
+      }
+    } else {
+      PADDLE_ENFORCE_NOT_NULL(nullptr, "split operator should",
+                              " specify indices or sections.");
+    }
+  }
+};
+
+class SplitOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SplitOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "the input tensor of split operator.");
+    AddOutput("Out", "the output tensors of split operator.").AsDuplicable();
+    AddComment(R"DOC(
+      Split the input tensor into multiple sub-tensors.
+      Example:
+        Input = [[1,2],
+                 [3,4],
+                 [5,6]]
+        sections = [2,1]
+        axis = 0
+        Output[0] = [[1,2],
+                     [3,4]]
+        Output[1] = [[5,6]]
+
+    )DOC");
+    AddAttr<std::vector<int>>("sections",
+                              "the length for each"
+                              "output along with the specify axis.")
+        .SetDefault(std::vector<int>{});
+    AddAttr<int>("num",
+                 "number of the sub-tensors, it must evenly divide "
+                 "Input.dims()[axis]")
+        .SetDefault(0);
+    AddAttr<int>("axis", "The axis which the input will be splited on.")
+        .SetDefault(0);
+  }
+};
+
+class SplitOpGrad : public NetOp {
+ public:
+  SplitOpGrad(const std::string &type, const framework::VariableNameMap &inputs,
+              const framework::VariableNameMap &outputs,
+              const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    auto out_grad = Inputs(framework::GradVarName("Out"));
+    auto x_grad = Output(framework::GradVarName("X"));
+    AppendOp(framework::OpRegistry::CreateOp("concat", {{"X", out_grad}},
+                                             {{"Out", {x_grad}}}, attrs));
+    CompleteAddOp(false);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+USE_CPU_ONLY_OP(concat);
+REGISTER_OP(split, ops::SplitOp, ops::SplitOpMaker, split_grad,
+            ops::SplitOpGrad);
+REGISTER_OP_CPU_KERNEL(split,
+                       ops::SplitKernel<paddle::platform::CPUPlace, float>);
diff --git a/paddle/operators/split_op.h b/paddle/operators/split_op.h
new file mode 100644
index 0000000000000000000000000000000000000000..860690ee895075fda9ddef08776a2102642efff9
--- /dev/null
+++ b/paddle/operators/split_op.h
@@ -0,0 +1,62 @@
+/* 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 <vector>
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename Place, typename T>
+class SplitKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* in = ctx.Input<framework::Tensor>("X");
+    auto outs = ctx.MultiOutput<framework::Tensor>("Out");
+    int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
+    size_t before = 1, after = 1;
+    const size_t n = outs.size();
+    size_t input_axis_dim = in->dims()[axis];
+
+    for (int64_t i = 0; i < in->dims().size(); ++i) {
+      if (i == axis) {
+        continue;
+      }
+      if (i < axis) {
+        before *= in->dims()[i];
+      } else {
+        after *= in->dims()[i];
+      }
+    }
+    size_t input_offset = 0;
+    for (size_t i = 0; i < n; i++) {
+      auto& out = outs[i];
+      size_t axis_dim = out->dims()[axis];
+      for (size_t j = 0; j < before; j++) {
+        size_t len = axis_dim * after * sizeof(T);
+        T* dest =
+            out->mutable_data<T>(platform::CPUPlace()) + axis_dim * after * j;
+        const T* src =
+            in->data<T>() + input_offset + input_axis_dim * after * j;
+        memcpy(dest, src, len);
+      }
+      input_offset += axis_dim * after;
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc
index ebe5bd352e99e298fb86355730feed77b236d2bd..39f4305877de20d451bc35fe698a0eabf9758d57 100644
--- a/paddle/operators/squared_l2_distance_op.cc
+++ b/paddle/operators/squared_l2_distance_op.cc
@@ -23,12 +23,18 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of SquaredL2DistanceOp "
-                            "must be initialized.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
-                            "Target of SquaredL2DistanceOp "
-                            "must be initialized.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("X"),
+        "Input(X) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.InputVar("Y"),
+        "Input(Y) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("sub_result"),
+        "Output(sub_result) of SquaredL2DistanceOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of SquaredL2DistanceOp should not be null.");
 
     auto* x = ctx.Input<Tensor>("X");
     auto x_dims = x->dims();
diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc
index 7170e7256c206d338ef1f6f94d5d1889ca92a3de..41e05c27f9029b2664685d3979fadcfd2bf6dbce 100644
--- a/paddle/operators/sum_op.cc
+++ b/paddle/operators/sum_op.cc
@@ -22,6 +22,11 @@ class SumOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE(!ctx.MultiInputVar("X").empty(),
+                   "Input(X) of SumOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of SumOp should not be null.");
+
     auto ins = ctx.MultiInput<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     int N = ins.size();
diff --git a/paddle/operators/top_k_op.cc b/paddle/operators/top_k_op.cc
index ff0e77a344ede7709a805d7dca4397eb49fa300c..169b815feffd86f9ff04c129ccc997230ce03a8c 100644
--- a/paddle/operators/top_k_op.cc
+++ b/paddle/operators/top_k_op.cc
@@ -24,7 +24,12 @@ class TopkOp : public framework::OperatorWithKernel {
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
     PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input of TopkOP must be initialized.");
+                            "Input(X) of TopkOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
+                            "Output(Out) of TopkOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Indices"),
+                            "Output(Indices) of TopkOp should not be null.");
+
     auto *input = ctx.Input<framework::Tensor>("X");
     const int k = static_cast<int>(ctx.Attr<int>("k"));
 
diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc
index ed7973693619e7765643dda824100afd82616470..184bcbc29c0d26a214345506f126f9cc0d406b07 100644
--- a/paddle/operators/uniform_random_op.cc
+++ b/paddle/operators/uniform_random_op.cc
@@ -48,6 +48,10 @@ class UniformRandomOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext& ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(
+        ctx.OutputVar("Out"),
+        "Output(Out) of UniformRandomOp should not be null.");
+
     PADDLE_ENFORCE(Attr<float>("min") < Attr<float>("max"),
                    "uniform_random's min must less then max");
     auto* tensor = ctx.Output<framework::LoDTensor>("Out");
diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt
index 17bdac8749e31565b119b2cb84aed199fac0f441..daf519b91d623d4369774dc4e37dcb7b1733666b 100644
--- a/paddle/platform/CMakeLists.txt
+++ b/paddle/platform/CMakeLists.txt
@@ -24,3 +24,4 @@ cc_library(device_context SRCS device_context.cc DEPS memory buddy_allocator
 nv_test(device_context_test SRCS device_context_test.cc DEPS device_context gpu_info)
 
 nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda)
+nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context)
diff --git a/paddle/platform/cuda_helper.h b/paddle/platform/cuda_helper.h
index 6feec0d7f8bd5d32d9e5eedee962fcbeff655f1c..a7d99cde106a0a66f122a8c43f49717c03e60dec 100644
--- a/paddle/platform/cuda_helper.h
+++ b/paddle/platform/cuda_helper.h
@@ -24,6 +24,11 @@ namespace platform {
 #define USE_CUDA_ATOMIC(op, T) \
   CUDA_ATOMIC_WRAPPER(op, T) { return atomic##op(address, val); }
 
+// Default thread count per block(or block size).
+// TODO(typhoonzero): need to benchmark against setting this value
+//                    to 1024.
+constexpr int PADDLE_CUDA_NUM_THREADS = 512;
+
 // For atomicAdd.
 USE_CUDA_ATOMIC(Add, float);
 
diff --git a/paddle/platform/details/device_ptr_cast.h b/paddle/platform/details/device_ptr_cast.h
new file mode 100644
index 0000000000000000000000000000000000000000..4015491fcdc3554029aa771ab7da1b2f3424321f
--- /dev/null
+++ b/paddle/platform/details/device_ptr_cast.h
@@ -0,0 +1,56 @@
+/* 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
+
+#ifndef __NVCC__
+#error device_ptr_cast must be include by .cu file
+#endif
+
+#include <thrust/device_ptr.h>
+
+namespace paddle {
+namespace platform {
+namespace details {
+template <typename T, bool is_ptr>
+struct DevicePtrCast;
+
+template <typename T>
+struct DevicePtrCast<T, true> {
+  using ELEM = typename std::remove_pointer<T>::type;
+  using RTYPE = thrust::device_ptr<ELEM>;
+
+  inline thrust::device_ptr<ELEM> operator()(ELEM* ele) const {
+    return thrust::device_pointer_cast(ele);
+  }
+};
+
+template <typename T>
+struct DevicePtrCast<T, false> {
+  using RTYPE = T;
+  inline RTYPE operator()(RTYPE it) const { return it; }
+};
+
+// Cast T to thrust::device_ptr if T is a pointer.
+// Otherwise, e.g., T is a iterator, return T itself.
+template <typename T>
+auto DevPtrCast(T t) ->
+    typename DevicePtrCast<T, std::is_pointer<T>::value>::RTYPE {
+  DevicePtrCast<T, std::is_pointer<T>::value> cast;
+  return cast(t);
+}
+
+}  // namespace details
+}  // namespace platform
+}  // namespace paddle
diff --git a/paddle/platform/device_context.cc b/paddle/platform/device_context.cc
index ad212c5b2c47312743362db4926c80bf056e100d..93b472b41c8a4c3a2bfada9d4fbf0e9e1b0cc736 100644
--- a/paddle/platform/device_context.cc
+++ b/paddle/platform/device_context.cc
@@ -101,19 +101,17 @@ CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) {
   eigen_stream_.reset(new EigenCudaStreamDevice());
   eigen_stream_->Reinitialize(&stream_, place);
   eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
+  PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_));
+  PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_));
+  PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
+  PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_));
 }
 
 CUDADeviceContext::~CUDADeviceContext() {
   SetDeviceId(place_.device);
   Wait();
-  if (cublas_handle_) {
-    PADDLE_ENFORCE(dynload::cublasDestroy(cublas_handle_));
-  }
-
-  if (cudnn_handle_) {
-    PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
-  }
-
+  PADDLE_ENFORCE(dynload::cublasDestroy(cublas_handle_));
+  PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
   eigen_stream_.reset();
   eigen_device_.reset();
   PADDLE_ENFORCE(cudaStreamDestroy(stream_));
@@ -129,25 +127,13 @@ Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
   return eigen_device_.get();
 }
 
-cublasHandle_t CUDADeviceContext::cublas_handle() {
-  if (!cublas_handle_) {
-    SetDeviceId(place_.device);
-    PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_));
-    PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_));
-  }
+cublasHandle_t CUDADeviceContext::cublas_handle() const {
   return cublas_handle_;
 }
 
-cudnnHandle_t CUDADeviceContext::cudnn_handle() {
-  if (!cudnn_handle_) {
-    SetDeviceId(place_.device);
-    PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
-    PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_));
-  }
-  return cudnn_handle_;
-}
+cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_handle_; }
 
-cudaStream_t CUDADeviceContext::stream() { return stream_; }
+cudaStream_t CUDADeviceContext::stream() const { return stream_; }
 
 #endif  // PADDLE_ONLY_CPU
 
diff --git a/paddle/platform/device_context.h b/paddle/platform/device_context.h
index 11528e1194e4516891034fa8febdac3ba6eed204..a106592e454e21c46cd2f87f1bbf6694955d6e23 100644
--- a/paddle/platform/device_context.h
+++ b/paddle/platform/device_context.h
@@ -67,16 +67,14 @@ class CUDADeviceContext : public DeviceContext {
   /*! \brief  Return eigen device in the device context. */
   Eigen::GpuDevice* eigen_device() const;
 
-  // clang-format off
   /*! \brief  Return cublas handle in the device context. */
-  cublasHandle_t    cublas_handle();
+  cublasHandle_t cublas_handle() const;
 
   /*! \brief  Return cudnn  handle in the device context. */
-  cudnnHandle_t     cudnn_handle();
+  cudnnHandle_t cudnn_handle() const;
 
   /*! \brief  Return cuda stream in the device context. */
-  cudaStream_t      stream();
-  // clang-format on
+  cudaStream_t stream() const;
 
  private:
   GPUPlace place_;
@@ -84,11 +82,9 @@ class CUDADeviceContext : public DeviceContext {
   std::unique_ptr<Eigen::GpuDevice> eigen_device_;
   std::unique_ptr<EigenCudaStreamDevice> eigen_stream_;
 
-  // clang-format off
-  cudaStream_t       stream_{nullptr};
-  cudnnHandle_t      cudnn_handle_{nullptr};
-  cublasHandle_t     cublas_handle_{nullptr};
-  // clang-format on
+  cudaStream_t stream_;
+  cudnnHandle_t cudnn_handle_;
+  cublasHandle_t cublas_handle_;
 };
 
 #endif
diff --git a/paddle/platform/transform.h b/paddle/platform/transform.h
new file mode 100644
index 0000000000000000000000000000000000000000..f196868c725cbb91b3df710260c5b60f14d53f37
--- /dev/null
+++ b/paddle/platform/transform.h
@@ -0,0 +1,99 @@
+/* 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/platform/device_context.h"
+#include "paddle/platform/enforce.h"
+#include "paddle/platform/hostdevice.h"
+#include "paddle/platform/place.h"
+
+#include <algorithm>
+#include <type_traits>
+#ifdef __NVCC__
+#include <thrust/execution_policy.h>
+#include <thrust/transform.h>
+#include "paddle/platform/details/device_ptr_cast.h"
+#endif
+
+namespace paddle {
+namespace platform {
+
+// Transform on host or device. It provides the same API in std library.
+template <typename Place>
+struct Transform {
+  template <typename InputIter, typename OutputIter, typename UnaryOperation>
+  void operator()(const DeviceContext& context, InputIter first, InputIter last,
+                  OutputIter result, UnaryOperation op);
+
+  template <typename InputIter1, typename InputIter2, typename OutputIter,
+            typename BinaryOperation>
+  void operator()(const DeviceContext& context, InputIter1 first1,
+                  InputIter1 last1, InputIter2 first2, OutputIter result,
+                  BinaryOperation op);
+};
+
+template <>
+struct Transform<platform::CPUPlace> {
+  template <typename InputIter, typename OutputIter, typename UnaryOperation>
+  void operator()(const DeviceContext& context, InputIter first, InputIter last,
+                  OutputIter result, UnaryOperation op) {
+    auto place = context.GetPlace();
+    PADDLE_ENFORCE(is_cpu_place(place), "It must use CPU place.");
+    std::transform(first, last, result, op);
+  }
+
+  template <typename InputIter1, typename InputIter2, typename OutputIter,
+            typename BinaryOperation>
+  void operator()(const DeviceContext& context, InputIter1 first1,
+                  InputIter1 last1, InputIter2 first2, OutputIter result,
+                  BinaryOperation op) {
+    auto place = context.GetPlace();
+    PADDLE_ENFORCE(is_cpu_place(place), "It must use CPU place.");
+    std::transform(first1, last1, first2, result, op);
+  }
+};
+
+#ifdef __NVCC__
+template <>
+struct Transform<platform::GPUPlace> {
+  template <typename InputIter, typename OutputIter, typename UnaryOperation>
+  void operator()(const DeviceContext& context, InputIter first, InputIter last,
+                  OutputIter result, UnaryOperation op) {
+    auto place = context.GetPlace();
+    PADDLE_ENFORCE(is_gpu_place(place), "It must use GPU place.");
+    auto& ctx = reinterpret_cast<const CUDADeviceContext&>(context);
+    thrust::transform(thrust::cuda::par.on(ctx.stream()),
+                      details::DevPtrCast(first), details::DevPtrCast(last),
+                      details::DevPtrCast(result), op);
+  }
+
+  template <typename InputIter1, typename InputIter2, typename OutputIter,
+            typename BinaryOperation>
+  void operator()(const DeviceContext& context, InputIter1 first1,
+                  InputIter1 last1, InputIter2 first2, OutputIter result,
+                  BinaryOperation op) {
+    auto place = context.GetPlace();
+    PADDLE_ENFORCE(is_gpu_place(place), "It must use GPU place.");
+    auto& ctx = reinterpret_cast<const CUDADeviceContext&>(context);
+    thrust::transform(thrust::cuda::par.on(ctx.stream()),
+                      details::DevPtrCast(first1), details::DevPtrCast(last1),
+                      details::DevPtrCast(first2), details::DevPtrCast(result),
+                      op);
+  }
+};
+#endif
+
+}  // namespace platform
+}  // namespace paddle
diff --git a/paddle/platform/transform_test.cu b/paddle/platform/transform_test.cu
new file mode 100644
index 0000000000000000000000000000000000000000..c76cab80e4b0e8df98a7be15f86699cfb6f93af2
--- /dev/null
+++ b/paddle/platform/transform_test.cu
@@ -0,0 +1,95 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
+
+#include <gtest/gtest.h>
+#include "paddle/memory/memcpy.h"
+#include "paddle/memory/memory.h"
+#include "paddle/platform/hostdevice.h"
+#include "paddle/platform/transform.h"
+
+template <typename T>
+class Scale {
+ public:
+  explicit Scale(const T& scale) : scale_(scale) {}
+
+  HOSTDEVICE T operator()(const T& a) const { return a * scale_; }
+
+ private:
+  T scale_;
+};
+
+template <typename T>
+class Multiply {
+ public:
+  HOSTDEVICE T operator()(const T& a, const T& b) const { return a * b; }
+};
+
+TEST(Transform, CPUUnary) {
+  using namespace paddle::platform;
+  CPUDeviceContext ctx;
+  float buf[4] = {0.1, 0.2, 0.3, 0.4};
+  Transform<paddle::platform::CPUPlace> trans;
+  trans(ctx, buf, buf + 4, buf, Scale<float>(10));
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_NEAR(buf[i], static_cast<float>(i + 1), 1e-5);
+  }
+}
+
+TEST(Transform, GPUUnary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  GPUPlace gpu0(0);
+  CUDADeviceContext ctx(gpu0);
+  float cpu_buf[4] = {0.1, 0.2, 0.3, 0.4};
+  float* gpu_buf = static_cast<float*>(Alloc(gpu0, sizeof(float) * 4));
+  Copy(gpu0, gpu_buf, CPUPlace(), cpu_buf, sizeof(cpu_buf));
+  Transform<paddle::platform::GPUPlace> trans;
+  trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
+  ctx.Wait();
+  Copy(CPUPlace(), cpu_buf, gpu0, gpu_buf, sizeof(cpu_buf));
+  Free(gpu0, gpu_buf);
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_NEAR(cpu_buf[i], static_cast<float>(i + 1), 1e-5);
+  }
+}
+
+TEST(Transform, CPUBinary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  int buf[4] = {1, 2, 3, 4};
+  Transform<paddle::platform::CPUPlace> trans;
+  CPUDeviceContext ctx;
+  trans(ctx, buf, buf + 4, buf, buf, Multiply<int>());
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_EQ((i + 1) * (i + 1), buf[i]);
+  }
+}
+
+TEST(Transform, GPUBinary) {
+  using namespace paddle::platform;
+  using namespace paddle::memory;
+  int buf[4] = {1, 2, 3, 4};
+  GPUPlace gpu0(0);
+  CUDADeviceContext ctx(gpu0);
+  int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
+  Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
+  Transform<paddle::platform::GPUPlace> trans;
+  trans(ctx, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
+  ctx.Wait();
+  Copy(CPUPlace(), buf, gpu0, gpu_buf, sizeof(buf));
+  Free(gpu0, gpu_buf);
+  for (int i = 0; i < 4; ++i) {
+    ASSERT_EQ((i + 1) * (i + 1), buf[i]);
+  }
+}
diff --git a/paddle/pserver/CMakeLists.txt b/paddle/pserver/CMakeLists.txt
index 2245c7d88ca74922f9919db91977dfa6cb3ca468..ccfc0e76020c7b4f54a493cc4048e7571379ec1a 100644
--- a/paddle/pserver/CMakeLists.txt
+++ b/paddle/pserver/CMakeLists.txt
@@ -45,14 +45,18 @@ add_dependencies(paddle_pserver paddle_proto ${external_project_dependencies})
 set(PSERVER_MAIN_SOURCES
     ParameterServer2Main.cpp)
 
-add_executable(paddle_pserver_main
-    ${PSERVER_MAIN_SOURCES})
-link_paddle_exe(paddle_pserver_main)
 if(WITH_TESTING)
   add_subdirectory(test)
 endif()
-install(TARGETS paddle_pserver_main
-    RUNTIME DESTINATION opt/paddle/bin
-    PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
-        GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)
-set_target_properties(paddle_pserver_main PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE)
+
+if(NOT WITH_C_API)
+  add_executable(paddle_pserver_main ${PSERVER_MAIN_SOURCES})
+  link_paddle_exe(paddle_pserver_main)
+
+  install(TARGETS paddle_pserver_main
+          RUNTIME DESTINATION opt/paddle/bin
+          PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
+          GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)
+
+  set_target_properties(paddle_pserver_main PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE)
+endif()
diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt
index 00030050700bfb2cee224124d090b0027d456ba0..4f05406c7f74113d8fb10aa6914166e553858338 100644
--- a/paddle/pybind/CMakeLists.txt
+++ b/paddle/pybind/CMakeLists.txt
@@ -1,5 +1,5 @@
 if(WITH_PYTHON)
-cc_library(paddle_pybind SHARED
+  cc_library(paddle_pybind SHARED
     SRCS pybind.cc
     DEPS pybind python backward
     ${GLOB_OP_LIB})
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index a7a38339fb2c8689778b0a86d3713f67e1447a80..c7009a604f60cda11434ad33b6c7d7caee1befdd 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -19,6 +19,7 @@ limitations under the License. */
 #include "paddle/framework/backward.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/cond_op.h"
 #include "paddle/operators/net_op.h"
 #include "paddle/operators/recurrent_op.h"
 #include "paddle/platform/enforce.h"
@@ -288,6 +289,28 @@ All parameter, weight, gradient are variables in Paddle.
            [](operators::RecurrentOp &self, const operators::NetOp &net)
                -> void { self.set_stepnet(net.Clone()); });
 
+  // cond_op
+  py::class_<operators::CondOp, OperatorBase>(m, "CondOp")
+      .def_static("create",
+                  [](py::bytes protobin) -> operators::CondOp * {
+                    OpDesc desc;
+                    PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
+                                   "Cannot parse user input to OpDesc");
+                    PADDLE_ENFORCE(desc.IsInitialized(),
+                                   "User OpDesc is not initialized, reason %s",
+                                   desc.InitializationErrorString());
+                    auto cond_op = OpRegistry::CreateOp(desc);
+                    return static_cast<operators::CondOp *>(cond_op.release());
+                  })
+      .def("set_truenet",
+           [](operators::CondOp &self, const operators::NetOp &net) -> void {
+             self.set_truenet(net.Clone());
+           })
+      .def("set_falsenet",
+           [](operators::CondOp &self, const operators::NetOp &net) -> void {
+             self.set_falsenet(net.Clone());
+           });
+
   m.def("unique_integer", UniqueIntegerGenerator);
 
   m.def("is_compile_gpu", IsCompileGPU);
diff --git a/paddle/scripts/travis/build_ios.sh b/paddle/scripts/travis/build_ios.sh
new file mode 100755
index 0000000000000000000000000000000000000000..dee7cf7cbbcccffd727002108ae7f6b6ee2fbba8
--- /dev/null
+++ b/paddle/scripts/travis/build_ios.sh
@@ -0,0 +1,20 @@
+#!/bin/bash
+set -e
+
+# Create the build directory for CMake.
+mkdir -p $TRAVIS_BUILD_DIR/build_ios
+cd $TRAVIS_BUILD_DIR/build_ios
+
+# Compile paddle binaries
+cmake -DCMAKE_SYSTEM_NAME=iOS \
+      -DIOS_PLATFORM=OS \
+      -DCMAKE_OSX_ARCHITECTURES="arm64" \
+      -DWITH_C_API=ON \
+      -DUSE_EIGEN_FOR_BLAS=ON \
+      -DWITH_TESTING=OFF \
+      -DWITH_SWIG_PY=OFF \
+      -DWITH_STYLE_CHECK=OFF \
+      -DCMAKE_BUILD_TYPE=Release \
+      ..
+
+make -j 2
diff --git a/paddle/scripts/travis/check_style.sh b/paddle/scripts/travis/check_style.sh
index ec499a839ac6593bac788f4cca5e33afbed73010..cb483b0ffc0a1d99978508bc16464a7716d2bac2 100755
--- a/paddle/scripts/travis/check_style.sh
+++ b/paddle/scripts/travis/check_style.sh
@@ -8,6 +8,12 @@ function abort(){
 trap 'abort' 0
 set -e
 
+# install glide
+curl https://glide.sh/get | bash
+eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
+go get -u github.com/alecthomas/gometalinter
+gometalinter --install
+
 cd $TRAVIS_BUILD_DIR
 export PATH=/usr/bin:$PATH
 pre-commit install
diff --git a/paddle/trainer/CMakeLists.txt b/paddle/trainer/CMakeLists.txt
index eac0584d30958ab78a935d89d217a4876fb07a19..3d471a0c01ca17cb98272159baf6d489c18824d5 100644
--- a/paddle/trainer/CMakeLists.txt
+++ b/paddle/trainer/CMakeLists.txt
@@ -50,22 +50,22 @@ macro(add_paddle_exe TARGET_NAME)
   link_paddle_exe(${TARGET_NAME})
 endmacro()
 
-add_paddle_exe(paddle_trainer
-    TrainerMain.cpp)
-
-add_paddle_exe(paddle_merge_model
-    MergeModel.cpp)
-
 if(WITH_TESTING)
-    add_subdirectory(tests)
+  add_subdirectory(tests)
 endif()
-install(TARGETS paddle_trainer paddle_merge_model
-    RUNTIME DESTINATION opt/paddle/bin
-    PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
-        GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)
 
-set_target_properties(paddle_trainer PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE)
-set_target_properties(paddle_merge_model PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE)
+if(NOT WITH_C_API)
+  add_paddle_exe(paddle_trainer TrainerMain.cpp)
+  add_paddle_exe(paddle_merge_model MergeModel.cpp)
+
+  install(TARGETS paddle_trainer paddle_merge_model
+          RUNTIME DESTINATION opt/paddle/bin
+          PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
+          GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)
+
+  set_target_properties(paddle_trainer PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE)
+  set_target_properties(paddle_merge_model PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE)
+endif()
 
 if(APPLE)
   set(CMAKE_EXE_LINKER_FLAGS "-framework CoreFoundation -framework Security")
@@ -73,6 +73,8 @@ endif()
 
 if(WITH_GOLANG)
   add_dependencies(paddle_trainer_lib paddle_pserver_cclient)
-  target_link_libraries(paddle_trainer paddle_pserver_cclient)
   target_link_libraries(paddle_trainer_lib paddle_pserver_cclient)
+  if(NOT WITH_C_API)
+    target_link_libraries(paddle_trainer paddle_pserver_cclient)
+  endif()
 endif(WITH_GOLANG)
diff --git a/paddle/utils/Excepts.h b/paddle/utils/Excepts.h
index 5c2c504f53a586f2991ccfae891991465fdb39b6..0add66da7464293795927431daf0e90359f40b52 100644
--- a/paddle/utils/Excepts.h
+++ b/paddle/utils/Excepts.h
@@ -17,7 +17,8 @@ limitations under the License. */
 
 #include <fenv.h>
 
-#if defined(__APPLE__) || defined(__OSX__)
+#if (defined(__APPLE__) || defined(__OSX__)) && !defined(__arm__) && \
+    !defined(__aarch64__)
 
 int fegetexcept(void);
 int feenableexcept(unsigned int excepts);
diff --git a/paddle/utils/arch/linux/Locks.cpp b/paddle/utils/arch/linux/Locks.cpp
index 3a0903d1f268cf0132da3de43396391219edf004..a4e6c8f7b8397adc262588612c250bac5ef5eaa6 100644
--- a/paddle/utils/arch/linux/Locks.cpp
+++ b/paddle/utils/arch/linux/Locks.cpp
@@ -40,6 +40,8 @@ void Semaphore::wait() { sem_wait(&m->sem); }
 
 void Semaphore::post() { sem_post(&m->sem); }
 
+/// SpinLockPrivate
+
 #ifdef PADDLE_USE_PTHREAD_SPINLOCK
 
 class SpinLockPrivate {
@@ -79,6 +81,8 @@ SpinLock::~SpinLock() { delete m; }
 void SpinLock::lock() { m->lock(); }
 void SpinLock::unlock() { m->unlock(); }
 
+/// ThreadBarrierPrivate
+
 #ifdef PADDLE_USE_PTHREAD_BARRIER
 
 class ThreadBarrierPrivate {
@@ -136,6 +140,8 @@ public:
 
 #endif
 
+/// ThreadBarrier
+
 ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate(count)) {}
 ThreadBarrier::~ThreadBarrier() { delete m; }
 void ThreadBarrier::wait() { m->wait(); }
diff --git a/paddle/utils/arch/osx/Excepts.cpp b/paddle/utils/arch/osx/Excepts.cpp
index c8e904d8f9fe29e51447994af43dc62bf3514306..42ecaa06d256c9d259a20c648626605d77ce0308 100644
--- a/paddle/utils/arch/osx/Excepts.cpp
+++ b/paddle/utils/arch/osx/Excepts.cpp
@@ -14,7 +14,8 @@ limitations under the License. */
 
 #include "paddle/utils/Excepts.h"
 
-#if defined(__APPLE__) || defined(__OSX__)
+#if (defined(__APPLE__) || defined(__OSX__)) && !defined(__arm__) && \
+    !defined(__aarch64__)
 
 int fegetexcept(void) {
   static fenv_t fenv;
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index a9e1d6d2e06d56f837690ec95fa8f8d41a90725f..7c32eb0069f4075d72cd4c3654c83e3d5c98fb1c 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -2286,8 +2286,15 @@ class NormLayer(LayerBase):
 
 @config_layer('pool')
 class PoolLayer(LayerBase):
+    layer_type = 'pool'
+
     def __init__(self, name, inputs, ceil_mode=True, **xargs):
-        super(PoolLayer, self).__init__(name, 'pool', 0, inputs=inputs, **xargs)
+        use_mkldnn = int(g_command_config_args.get("use_mkldnn", 0))
+        if self.layer_type == "mkldnn_pool":
+            config_assert(use_mkldnn, "mkldnn_pool only support MKLDNN")
+        self.layer_type = 'mkldnn_pool' if use_mkldnn else 'pool'
+        super(PoolLayer, self).__init__(
+            name, self.layer_type, 0, inputs=inputs, **xargs)
         for input_index in xrange(len(self.inputs)):
             input_layer = self.get_input_layer(input_index)
             pool_conf = self.config.inputs[input_index].pool_conf
@@ -2297,6 +2304,11 @@ class PoolLayer(LayerBase):
                                pool_conf.channels)
 
 
+@config_layer('mkldnn_pool')
+class MKLDNNPoolLayer(PoolLayer):
+    layer_type = 'mkldnn_pool'
+
+
 @config_layer('pool3d')
 class Pool3DLayer(LayerBase):
     def __init__(self, name, inputs, ceil_mode=True, **xargs):
diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index 8c7d1738ad9753eb7afb27e893f979f8bce70a0d..9bdcca1716fca73e87adee861e91b6b90d1ef70d 100644
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -1224,8 +1224,8 @@ def detection_output_layer(input_loc,
                            name=None):
     """
     Apply the NMS to the output of network and compute the predict bounding
-    box location. The output of this layer could be None if there is no valid
-    bounding box.
+    box location. The output's shape of this layer could be zero if there is
+    no valid bounding box.
 
     :param name: The Layer Name.
     :type name: basestring
diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py
index 34be203ee254584027c79cf93fe54f404b7235db..93e8ac173e721d9623fce91f30ac4642d273caba 100644
--- a/python/paddle/trainer_config_helpers/networks.py
+++ b/python/paddle/trainer_config_helpers/networks.py
@@ -11,10 +11,8 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-"""
 
-"""
-# from activations import *
+
 from activations import LinearActivation, ReluActivation, SoftmaxActivation, \
     IdentityActivation, TanhActivation, SequenceSoftmaxActivation
 from attrs import ExtraAttr
@@ -55,49 +53,49 @@ def sequence_conv_pool(input,
                        context_attr=None,
                        pool_attr=None):
     """
-    Text convolution pooling layers helper.
+    Text convolution pooling group.
 
     Text input => Context Projection => FC Layer => Pooling => Output.
 
-    :param name: name of output layer(pooling layer name)
+    :param name: group name.
     :type name: basestring
-    :param input: name of input layer
+    :param input: input layer.
     :type input: LayerOutput
     :param context_len: context projection length. See
                         context_projection's document.
     :type context_len: int
     :param hidden_size: FC Layer size.
     :type hidden_size: int
-    :param context_start: context projection length. See
+    :param context_start: context start position. See
                           context_projection's context_start.
-    :type context_start: int or None
+    :type context_start: int|None
     :param pool_type: pooling layer type. See pooling_layer's document.
-    :type pool_type: BasePoolingType.
+    :type pool_type: BasePoolingType
     :param context_proj_layer_name: context projection layer name.
                                     None if user don't care.
     :type context_proj_layer_name: basestring
-    :param context_proj_param_attr: context projection parameter attribute.
-                                    None if user don't care.
-    :type context_proj_param_attr: ParameterAttribute or None.
+    :param context_proj_param_attr: padding parameter attribute of context projection layer.
+                                    If false, it means padding always be zero.
+    :type context_proj_param_attr: ParameterAttribute|None
     :param fc_layer_name: fc layer name. None if user don't care.
     :type fc_layer_name: basestring
     :param fc_param_attr: fc layer parameter attribute. None if user don't care.
-    :type fc_param_attr: ParameterAttribute or None
+    :type fc_param_attr: ParameterAttribute|None
     :param fc_bias_attr: fc bias parameter attribute. False if no bias,
                          None if user don't care.
-    :type fc_bias_attr: ParameterAttribute or None
-    :param fc_act: fc layer activation type. None means tanh
+    :type fc_bias_attr: ParameterAttribute|False|None
+    :param fc_act: fc layer activation type. None means tanh.
     :type fc_act: BaseActivation
-    :param pool_bias_attr: pooling layer bias attr. None if don't care.
-                           False if no bias.
-    :type pool_bias_attr: ParameterAttribute or None.
+    :param pool_bias_attr: pooling layer bias attr. False if no bias.
+                           None if user don't care.
+    :type pool_bias_attr: ParameterAttribute|False|None
     :param fc_attr: fc layer extra attribute.
     :type fc_attr: ExtraLayerAttribute
     :param context_attr: context projection layer extra attribute.
     :type context_attr: ExtraLayerAttribute
     :param pool_attr: pooling layer extra attribute.
     :type pool_attr: ExtraLayerAttribute
-    :return: output layer name.
+    :return: layer's output.
     :rtype: LayerOutput
     """
     # Set Default Value to param
@@ -163,45 +161,45 @@ def simple_img_conv_pool(input,
     """
     Simple image convolution and pooling group.
 
-    Input => conv => pooling
+    Img input => Conv => Pooling => Output.
 
-    :param name: group name
+    :param name: group name.
     :type name: basestring
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
-    :param filter_size: see img_conv_layer for details
+    :param filter_size: see img_conv_layer for details.
     :type filter_size: int
-    :param num_filters: see img_conv_layer for details
+    :param num_filters: see img_conv_layer for details.
     :type num_filters: int
-    :param pool_size: see img_pool_layer for details
+    :param pool_size: see img_pool_layer for details.
     :type pool_size: int
-    :param pool_type: see img_pool_layer for details
+    :param pool_type: see img_pool_layer for details.
     :type pool_type: BasePoolingType
-    :param act: see img_conv_layer for details
+    :param act: see img_conv_layer for details.
     :type act: BaseActivation
-    :param groups: see img_conv_layer for details
+    :param groups: see img_conv_layer for details.
     :type groups: int
-    :param conv_stride: see img_conv_layer for details
+    :param conv_stride: see img_conv_layer for details.
     :type conv_stride: int
-    :param conv_padding: see img_conv_layer for details
+    :param conv_padding: see img_conv_layer for details.
     :type conv_padding: int
-    :param bias_attr: see img_conv_layer for details
+    :param bias_attr: see img_conv_layer for details.
     :type bias_attr: ParameterAttribute
-    :param num_channel: see img_conv_layer for details
+    :param num_channel: see img_conv_layer for details.
     :type num_channel: int
-    :param param_attr: see img_conv_layer for details
+    :param param_attr: see img_conv_layer for details.
     :type param_attr: ParameterAttribute
-    :param shared_bias: see img_conv_layer for details
+    :param shared_bias: see img_conv_layer for details.
     :type shared_bias: bool
-    :param conv_layer_attr: see img_conv_layer for details
+    :param conv_layer_attr: see img_conv_layer for details.
     :type conv_layer_attr: ExtraLayerAttribute
-    :param pool_stride: see img_pool_layer for details
+    :param pool_stride: see img_pool_layer for details.
     :type pool_stride: int
-    :param pool_padding: see img_pool_layer for details
+    :param pool_padding: see img_pool_layer for details.
     :type pool_padding: int
-    :param pool_layer_attr: see img_pool_layer for details
+    :param pool_layer_attr: see img_pool_layer for details.
     :type pool_layer_attr: ExtraLayerAttribute
-    :return: Layer's output
+    :return: layer's output
     :rtype: LayerOutput
     """
     _conv_ = img_conv_layer(
@@ -252,48 +250,52 @@ def img_conv_bn_pool(input,
                      pool_layer_attr=None):
     """
     Convolution, batch normalization, pooling group.
+    
+    Img input => Conv => BN => Pooling => Output.
 
-    :param name: group name
+    :param name: group name.
     :type name: basestring
-    :param input: layer's input
-    :type input: LayerOutput
-    :param filter_size: see img_conv_layer's document
+    :param input: input layer.
+    :type input: LayerOutput 
+    :param filter_size: see img_conv_layer for details.
     :type filter_size: int
-    :param num_filters: see img_conv_layer's document
+    :param num_filters: see img_conv_layer for details.
     :type num_filters: int
-    :param pool_size: see img_pool_layer's document.
+    :param pool_size: see img_pool_layer for details.
     :type pool_size: int
-    :param pool_type: see img_pool_layer's document.
+    :param pool_type: see img_pool_layer for details.
     :type pool_type: BasePoolingType
-    :param act: see batch_norm_layer's document.
+    :param act: see batch_norm_layer for details.
     :type act: BaseActivation
-    :param groups: see img_conv_layer's document
+    :param groups: see img_conv_layer for details.
     :type groups: int
-    :param conv_stride: see img_conv_layer's document.
+    :param conv_stride: see img_conv_layer for details.
     :type conv_stride: int
-    :param conv_padding: see img_conv_layer's document.
+    :param conv_padding: see img_conv_layer for details.
     :type conv_padding: int
-    :param conv_bias_attr: see img_conv_layer's document.
+    :param conv_bias_attr: see img_conv_layer for details.
     :type conv_bias_attr: ParameterAttribute
-    :param num_channel: see img_conv_layer's document.
+    :param num_channel: see img_conv_layer for details.
     :type num_channel: int
-    :param conv_param_attr: see img_conv_layer's document.
+    :param conv_param_attr: see img_conv_layer for details.
     :type conv_param_attr: ParameterAttribute
-    :param shared_bias: see img_conv_layer's document.
+    :param shared_bias: see img_conv_layer for details.
     :type shared_bias: bool
-    :param conv_layer_attr: see img_conv_layer's document.
+    :param conv_layer_attr: see img_conv_layer for details.
     :type conv_layer_attr: ExtraLayerOutput
-    :param bn_param_attr: see batch_norm_layer's document.
-    :type bn_param_attr: ParameterAttribute.
-    :param bn_bias_attr: see batch_norm_layer's document.
-    :param bn_layer_attr: ParameterAttribute.
-    :param pool_stride: see img_pool_layer's document.
+    :param bn_param_attr: see batch_norm_layer for details.
+    :type bn_param_attr: ParameterAttribute
+    :param bn_bias_attr: see batch_norm_layer for details.
+    :type bn_bias_attr: ParameterAttribute
+    :param bn_layer_attr: see batch_norm_layer for details.
+    :type bn_layer_attr: ExtraLayerAttribute
+    :param pool_stride: see img_pool_layer for details.
     :type pool_stride: int
-    :param pool_padding: see img_pool_layer's document.
+    :param pool_padding: see img_pool_layer for details.
     :type pool_padding: int
-    :param pool_layer_attr: see img_pool_layer's document.
+    :param pool_layer_attr: see img_pool_layer for details.
     :type pool_layer_attr: ExtraLayerAttribute
-    :return: Layer groups output
+    :return: layer's output
     :rtype: LayerOutput
     """
     __conv__ = img_conv_layer(
@@ -348,10 +350,10 @@ def img_conv_group(input,
     :param conv_batchnorm_drop_rate: if conv_with_batchnorm[i] is true,
         conv_batchnorm_drop_rate[i] represents the drop rate of each batch norm.
     :type conv_batchnorm_drop_rate: list
-    :param input: layer's input.
+    :param input: input layer.
     :type input: LayerOutput
-    :param conv_num_filter: output channels num.
-    :type conv_num_filter: int
+    :param conv_num_filter: list of output channels num.
+    :type conv_num_filter: list|tuple
     :param pool_size: pooling filter size.
     :type pool_size: int
     :param num_channels: input channels num.
@@ -362,18 +364,18 @@ def img_conv_group(input,
     :type conv_filter_size: int
     :param conv_act: activation funciton after convolution.
     :type conv_act: BaseActivation
-    :param conv_with_batchnorm: conv_with_batchnorm[i] represents
-        if there is a batch normalization after each convolution.
+    :param conv_with_batchnorm: if conv_with_batchnorm[i] is true,
+        there is a batch normalization operation after each convolution.
     :type conv_with_batchnorm: list
     :param pool_stride: pooling stride size.
     :type pool_stride: int
     :param pool_type: pooling type.
     :type pool_type: BasePoolingType
-    :param param_attr: Convolution param attribute.
-        None means default attribute.
+    :param param_attr: param attribute of convolution layer,
+                       None means default attribute.
     :type param_attr: ParameterAttribute
-    :return: Layer's output
-    :type: LayerOutput
+    :return: layer's output
+    :rtype: LayerOutput
     """
     tmp = input
 
@@ -466,12 +468,14 @@ def vgg_16_network(input_image, num_channels, num_classes=1000):
     """
     Same model from https://gist.github.com/ksimonyan/211839e770f7b538e2d8
 
-    :param num_classes:
-    :param input_image:
+    :param num_classes: number of class.
+    :type num_classes: int
+    :param input_image: input layer.
     :type input_image: LayerOutput
-    :param num_channels:
+    :param num_channels: input channels num.
     :type num_channels: int
-    :return:
+    :return: layer's output
+    :rtype: LayerOutput
     """
 
     tmp = img_conv_group(
@@ -560,8 +564,8 @@ def simple_lstm(input,
     """
     Simple LSTM Cell.
 
-    It just combine a mixed layer with fully_matrix_projection and a lstmemory
-    layer. The simple lstm cell was implemented as follow equations.
+    It just combines a mixed layer with fully_matrix_projection and a lstmemory
+    layer. The simple lstm cell was implemented with follow equations.
 
     ..  math::
 
@@ -575,37 +579,37 @@ def simple_lstm(input,
 
         h_t & = o_t tanh(c_t)
 
-    Please refer **Generating Sequences With Recurrent Neural Networks** if you
-    want to know what lstm is. Link_ is here.
+    Please refer to **Generating Sequences With Recurrent Neural Networks** for more
+    details about lstm. Link_ is here.
 
     .. _Link: http://arxiv.org/abs/1308.0850
 
     :param name: lstm layer name.
     :type name: basestring
-    :param input: input layer name.
+    :param input: layer's input.
     :type input: LayerOutput
     :param size: lstm layer size.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param mat_param_attr: mixed layer's matrix projection parameter attribute.
+    :param mat_param_attr: parameter attribute of matrix projection in mixed layer.
     :type mat_param_attr: ParameterAttribute
     :param bias_param_attr: bias parameter attribute. False means no bias, None
                             means default bias.
     :type bias_param_attr: ParameterAttribute|False
-    :param inner_param_attr: lstm cell parameter attribute.
+    :param inner_param_attr: parameter attribute of lstm cell.
     :type inner_param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
-    :param mixed_layer_attr: mixed layer's extra attribute.
+    :param mixed_layer_attr: extra attribute of mixed layer.
     :type mixed_layer_attr: ExtraLayerAttribute
-    :param lstm_cell_attr: lstm layer's extra attribute.
+    :param lstm_cell_attr: extra attribute of lstm.
     :type lstm_cell_attr: ExtraLayerAttribute
-    :return: lstm layer name.
+    :return: layer's output.
     :rtype: LayerOutput
     """
     fc_name = 'lstm_transform_%s' % name
@@ -643,9 +647,9 @@ def lstmemory_unit(input,
                    lstm_bias_attr=None,
                    lstm_layer_attr=None):
     """
-    Define calculations that a LSTM unit performs during a single time step.
-    This function itself is not a recurrent layer, so it can not be
-    directly used to process sequence inputs. This function is always used in
+    lstmemory_unit defines the caculation process of a LSTM unit during a 
+    single time step. This function is not a recurrent layer, so it can not be
+    directly used to process sequence input. This function is always used in
     recurrent_group (see layers.py for more details) to implement attention
     mechanism.
 
@@ -676,7 +680,7 @@ def lstmemory_unit(input,
                                    state_act=TanhActivation())
 
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param out_memory: output of previous time step
     :type out_memory: LayerOutput | None
@@ -684,15 +688,15 @@ def lstmemory_unit(input,
     :type name: basestring
     :param size: lstmemory unit size.
     :type size: int
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: parameter attribute, None means default attribute.
     :type param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
-    :param input_proj_bias_attr: bias attribute for input-to-hidden projection.
+    :param input_proj_bias_attr: bias attribute for input to hidden projection.
                 False means no bias, None means default bias.
     :type input_proj_bias_attr: ParameterAttribute|False|None
     :param input_proj_layer_attr: extra layer attribute for input to hidden
@@ -700,8 +704,8 @@ def lstmemory_unit(input,
     :type input_proj_layer_attr: ExtraLayerAttribute
     :param lstm_bias_attr: bias parameter attribute of lstm layer.
                 False means no bias, None means default bias.
-    :type lstm_bias_attr: ParameterAttribute|False
-    :param lstm_layer_attr: lstm layer's extra attribute.
+    :type lstm_bias_attr: ParameterAttribute|False|None
+    :param lstm_layer_attr: extra attribute of lstm layer.
     :type lstm_layer_attr: ExtraLayerAttribute
     :return: lstmemory unit name.
     :rtype: LayerOutput
@@ -758,9 +762,9 @@ def lstmemory_group(input,
     lstm_group is a recurrent_group version of Long Short Term Memory. It
     does exactly the same calculation as the lstmemory layer (see lstmemory in
     layers.py for the maths) does. A promising benefit is that LSTM memory
-    cell states, or hidden states in every time step are accessible to the
+    cell states(or hidden states) in every time step are accessible to the
     user. This is especially useful in attention model. If you do not need to
-    access the internal states of the lstm, but merely use its outputs,
+    access the internal states of the lstm and merely use its outputs,
     it is recommended to use the lstmemory, which is relatively faster than
     lstmemory_group.
 
@@ -781,28 +785,28 @@ def lstmemory_group(input,
                                     gate_act=SigmoidActivation(),
                                     state_act=TanhActivation())
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param size: lstmemory group size.
     :type size: int
-    :param name: name of the lstmemory group.
+    :param name: name of lstmemory group.
     :type name: basestring
-    :param out_memory: output of previous time step
+    :param out_memory: output of previous time step.
     :type out_memory: LayerOutput | None
-    :param reverse: is lstm reversed
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param param_attr: Parameter config, None if use default.
+    :param param_attr: parameter attribute, None means default attribute.
     :type param_attr: ParameterAttribute
-    :param act: lstm final activiation type
+    :param act: last activiation type of lstm.
     :type act: BaseActivation
-    :param gate_act: lstm gate activiation type
+    :param gate_act: gate activiation type of lstm.
     :type gate_act: BaseActivation
-    :param state_act: lstm state activiation type.
+    :param state_act: state activiation type of lstm.
     :type state_act: BaseActivation
     :param lstm_bias_attr: bias parameter attribute of lstm layer.
                            False means no bias, None means default bias.
-    :type lstm_bias_attr: ParameterAttribute|False
-    :param input_proj_bias_attr: bias attribute for input-to-hidden projection.
+    :type lstm_bias_attr: ParameterAttribute|False|None
+    :param input_proj_bias_attr: bias attribute for input to hidden projection.
                 False means no bias, None means default bias.
     :type input_proj_bias_attr: ParameterAttribute|False|None
     :param input_proj_layer_attr: extra layer attribute for input to hidden
@@ -848,15 +852,15 @@ def gru_unit(input,
              gru_layer_attr=None,
              naive=False):
     """
-    Define calculations that a gated recurrent unit performs in a single time
-    step. This function itself is not a recurrent layer, so it can not be
-    directly used to process sequence inputs. This function is always used in
+    gru_unit defines the calculation process of a gated recurrent unit during a single 
+    time step. This function is not a recurrent layer, so it can not be
+    directly used to process sequence input. This function is always used in
     the recurrent_group (see layers.py for more details) to implement attention
     mechanism.
 
     Please see grumemory in layers.py for the details about the maths.
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param memory_boot: the initialization state of the LSTM cell.
     :type memory_boot: LayerOutput | None
@@ -864,12 +868,12 @@ def gru_unit(input,
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param act: type of the activation
+    :param act: activation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activation
+    :param gate_act: gate activation type or gru
     :type gate_act: BaseActivation
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru output layer.
     :rtype: LayerOutput
     """
@@ -915,7 +919,7 @@ def gru_group(input,
     does exactly the same calculation as the grumemory layer does. A promising
     benefit is that gru hidden states are accessible to the user. This is
     especially useful in attention model. If you do not need to access
-    any internal state, but merely use the outputs of a GRU, it is recommended
+    any internal state and merely use the outputs of a GRU, it is recommended
     to use the grumemory, which is relatively faster.
 
     Please see grumemory in layers.py for more detail about the maths.
@@ -924,12 +928,12 @@ def gru_group(input,
 
     ..  code-block:: python
 
-        gru = gur_group(input=[layer1],
+        gru = gru_group(input=[layer1],
                         size=256,
                         act=TanhActivation(),
                         gate_act=SigmoidActivation())
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param memory_boot: the initialization state of the LSTM cell.
     :type memory_boot: LayerOutput | None
@@ -937,16 +941,17 @@ def gru_group(input,
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer,
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -986,11 +991,11 @@ def simple_gru(input,
                gru_layer_attr=None,
                naive=False):
     """
-    You maybe see gru_step_layer, grumemory in layers.py, gru_unit, gru_group,
+    You may see gru_step_layer, grumemory in layers.py, gru_unit, gru_group,
     simple_gru in network.py. The reason why there are so many interfaces is
     that we have two ways to implement recurrent neural network. One way is to
     use one complete layer to implement rnn (including simple rnn, gru and lstm)
-    with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But,
+    with multiple time steps, such as recurrent_layer, lstmemory, grumemory. But 
     the multiplication operation :math:`W x_t` is not computed in these layers.
     See details in their interfaces in layers.py.
     The other implementation is to use an recurrent group which can ensemble a
@@ -1018,22 +1023,23 @@ def simple_gru(input,
 
         gru = simple_gru(input=[layer1], size=256)
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param name: name of the gru group.
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer,
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -1071,8 +1077,8 @@ def simple_gru2(input,
                 mixed_layer_attr=None,
                 gru_cell_attr=None):
     """
-    simple_gru2 is the same with simple_gru, but using grumemory instead
-    Please see grumemory in layers.py for more detail about the maths.
+    simple_gru2 is the same with simple_gru, but using grumemory instead.
+    Please refer to grumemory in layers.py for more detail about the math.
     simple_gru2 is faster than simple_gru.
 
     The example usage is:
@@ -1081,22 +1087,23 @@ def simple_gru2(input,
 
         gru = simple_gru2(input=[layer1], size=256)
 
-    :param input: input layer name.
+    :param input: input layer.
     :type input: LayerOutput
     :param name: name of the gru group.
     :type name: basestring
     :param size: hidden size of the gru.
     :type size: int
-    :param reverse: whether to process the input data in a reverse order
+    :param reverse: process the input in a reverse order or not.
     :type reverse: bool
-    :param act: type of the activiation
+    :param act: activiation type of gru
     :type act: BaseActivation
-    :param gate_act: type of the gate activiation
+    :param gate_act: gate activiation type of gru
     :type gate_act: BaseActivation
-    :param gru_bias_attr: bias. False means no bias, None means default bias.
-    :type gru_bias_attr: ParameterAttribute|False
-    :param gru_layer_attr: Extra parameter attribute of the gru layer.
-    :type gru_layer_attr: ParameterAttribute|False
+    :param gru_bias_attr: bias parameter attribute of gru layer, 
+                          False means no bias, None means default bias.
+    :type gru_bias_attr: ParameterAttribute|False|None
+    :param gru_layer_attr: Extra attribute of the gru layer.
+    :type gru_layer_attr: ExtraLayerAttribute
     :return: the gru group.
     :rtype: LayerOutput
     """
@@ -1145,7 +1152,7 @@ def bidirectional_gru(input,
                       concat_act=None):
     """
     A bidirectional_gru is a recurrent unit that iterates over the input
-    sequence both in forward and bardward orders, and then concatenate two
+    sequence both in forward and backward orders, and then concatenate two
     outputs to form a final output. However, concatenation of two outputs
     is not the only way to form the final output, you can also, for example,
     just add them together.
@@ -1162,11 +1169,10 @@ def bidirectional_gru(input,
     :type input: LayerOutput
     :param size: gru layer size.
     :type size: int
-    :param return_seq: If set False, outputs of the last time step are
-                       concatenated and returned.
-                       If set True, the entire output sequences that are
-                       processed in forward and backward directions are
+    :param return_seq: If set False, the last time step of output are
                        concatenated and returned.
+                       If set True, the entire output sequences in forward 
+                       and backward directions are concatenated and returned.
     :type return_seq: bool
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -1230,8 +1236,8 @@ def bidirectional_lstm(input,
                        concat_act=None):
     """
     A bidirectional_lstm is a recurrent unit that iterates over the input
-    sequence both in forward and bardward orders, and then concatenate two
-    outputs form a final output. However, concatenation of two outputs
+    sequence both in forward and backward orders, and then concatenate two
+    outputs to form a final output. However, concatenation of two outputs
     is not the only way to form the final output, you can also, for example,
     just add them together.
 
@@ -1252,13 +1258,12 @@ def bidirectional_lstm(input,
     :type input: LayerOutput
     :param size: lstm layer size.
     :type size: int
-    :param return_seq: If set False, outputs of the last time step are
-                       concatenated and returned.
-                       If set True, the entire output sequences that are
-                       processed in forward and backward directions are
+    :param return_seq: If set False, the last time step of output are
                        concatenated and returned.
+                       If set True, the entire output sequences in forward 
+                       and backward directions are concatenated and returned.
     :type return_seq: bool
-    :return: LayerOutput object accroding to the return_seq.
+    :return: LayerOutput object.
     :rtype: LayerOutput
     """
     args = locals()
@@ -1303,7 +1308,7 @@ def simple_attention(encoded_sequence,
                      weight_act=None,
                      name=None):
     """
-    Calculate and then return a context vector by attention machanism.
+    Calculate and return a context vector with attention mechanism.
     Size of the context vector equals to size of the encoded_sequence.
 
     ..  math::
@@ -1336,10 +1341,10 @@ def simple_attention(encoded_sequence,
     :param name: name of the attention model.
     :type name: basestring
     :param softmax_param_attr: parameter attribute of sequence softmax
-                               that is used to produce attention weight
+                               that is used to produce attention weight.
     :type softmax_param_attr: ParameterAttribute
-    :param weight_act: activation of the attention model
-    :type weight_act: Activation
+    :param weight_act: activation of the attention model.
+    :type weight_act: BaseActivation
     :param encoded_sequence: output of the encoder
     :type encoded_sequence: LayerOutput
     :param encoded_proj: attention weight is computed by a feed forward neural
@@ -1411,7 +1416,7 @@ def inputs(layers, *args):
 
 def outputs(layers, *args):
     """
-    Declare the outputs of network. If user have not defined the inputs of
+    Declare the outputs of network. If user has not defined the inputs of
     network, this method will calculate the input order by dfs travel.
 
     :param layers: Output layers.
diff --git a/python/paddle/v2/framework/op.py b/python/paddle/v2/framework/op.py
index 15e0d125c495fbc0688d8dc4e66881cb9ab95a90..6cca41e43b38b8cccb65ff9b347ef226dddecd4d 100644
--- a/python/paddle/v2/framework/op.py
+++ b/python/paddle/v2/framework/op.py
@@ -215,5 +215,27 @@ class __RecurrentOp__(object):
         return core.RecurrentOp.create(proto.SerializeToString())
 
 
+class __CondOp__(object):
+    __proto__ = None
+    type = "cond"
+
+    def __init__(self):
+        # cache recurrent_op's proto
+        if self.__proto__ is None:
+            for op_proto in get_all_op_protos():
+                if op_proto.type == self.type:
+                    self.__proto__ = op_proto
+
+    def __call__(self, *args, **kwargs):
+        if self.type not in args and "type" not in kwargs:
+            kwargs["type"] = self.type
+        # create proto
+        create_method = OpDescCreationMethod(self.__proto__)
+        proto = create_method(*args, **kwargs)
+        # create condop
+        return core.CondOp.create(proto.SerializeToString())
+
+
 Operator = OperatorFactory()  # The default global factory
 RecurrentOp = __RecurrentOp__()
+CondOp = __CondOp__()
diff --git a/python/paddle/v2/framework/tests/op_test.py b/python/paddle/v2/framework/tests/op_test.py
index 9936fd76baf3e64aed01b8ae1d54e50b39793925..0a5673868c547d9e184e8ce05346c3ebabe06892 100644
--- a/python/paddle/v2/framework/tests/op_test.py
+++ b/python/paddle/v2/framework/tests/op_test.py
@@ -28,10 +28,10 @@ def create_op(scope, op_type, inputs, outputs, attrs):
         if out_name in outputs:
             kwargs[out_name] = []
             if out_dup:
-                sub_in = outputs[out_name]
-                for sub_in_name, _ in sub_in:
-                    var = scope.new_var(sub_in_name)
-                    kwargs[out_name].append(sub_in_name)
+                sub_out = outputs[out_name]
+                for sub_out_name, _ in sub_out:
+                    var = scope.new_var(sub_out_name)
+                    kwargs[out_name].append(sub_out_name)
             else:
                 var = scope.new_var(out_name)
                 kwargs[out_name].append(out_name)
@@ -39,6 +39,7 @@ def create_op(scope, op_type, inputs, outputs, attrs):
     for attr_name in Operator.get_op_attr_names(op_type):
         if attr_name in attrs:
             kwargs[attr_name] = attrs[attr_name]
+
     return Operator(op_type, **kwargs)
 
 
@@ -47,17 +48,24 @@ def set_input(scope, op, inputs, place):
         if in_name in inputs:
             if in_dup:
                 sub_in = inputs[in_name]
-                for sub_in_name, sub_in_array in sub_in:
+                for sub_in_name, sub_in_val in sub_in:
                     var = scope.find_var(sub_in_name)
                     tensor = var.get_tensor()
+                    sub_in_array = sub_in_val[0] \
+                        if isinstance(sub_in_val, tuple) else sub_in_val
                     tensor.set_dims(sub_in_array.shape)
                     tensor.set(sub_in_array, place)
+                    if isinstance(sub_in_val, tuple):
+                        tensor.set_lod(sub_in_val[1])
             else:
                 var = scope.find_var(in_name)
                 tensor = var.get_tensor()
-                arr = inputs[in_name]
-                tensor.set_dims(arr.shape)
-                tensor.set(arr, place)
+                in_val = inputs[in_name]
+                in_array = in_val[0] if isinstance(in_val, tuple) else in_val
+                tensor.set_dims(in_array.shape)
+                tensor.set(in_array, place)
+                if isinstance(in_val, tuple):
+                    tensor.set_lod(in_val[1])
 
 
 def set_output_grad(scope, op, outputs, place):
@@ -172,8 +180,9 @@ class OpTest(unittest.TestCase):
     def check_output_with_place(self, place):
         self.scope = core.Scope()
         op_inputs = self.inputs if hasattr(self, "inputs") else dict()
+        op_outputs = self.outputs if hasattr(self, "outputs") else dict()
         op_attrs = self.attrs if hasattr(self, "attrs") else dict()
-        self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs,
+        self.op = create_op(self.scope, self.op_type, op_inputs, op_outputs,
                             op_attrs)
         if isinstance(place, core.GPUPlace) and not self.op.support_gpu():
             return
@@ -183,23 +192,29 @@ class OpTest(unittest.TestCase):
         self.op.run(self.scope, ctx)
 
         for out_name, out_dup in Operator.get_op_outputs(self.op.type()):
+            if out_name not in self.outputs:
+                continue
+
             if out_dup:
                 sub_out = self.outputs[out_name]
-                for sub_out_name in sub_out:
+                if not isinstance(sub_out, list):
+                    raise AssertionError("sub_out type %s is not list",
+                                         type(sub_out))
+
+                for sub_out_name, expect in sub_out:
                     actual = np.array(
                         self.scope.find_var(sub_out_name).get_tensor())
-                    expect = sub_out[sub_out_name]
                     self.assertTrue(
                         np.allclose(
                             actual, expect, atol=1e-05),
-                        "output name: " + out_name + "has diff")
+                        "output name: " + out_name + " has diff")
             else:
                 actual = np.array(self.scope.find_var(out_name).get_tensor())
                 expect = self.outputs[out_name]
                 self.assertTrue(
                     np.allclose(
                         actual, expect, atol=1e-05),
-                    "output name: " + out_name + "has diff")
+                    "output name: " + out_name + " has diff")
 
     def check_output(self):
         places = [core.CPUPlace()]
@@ -234,8 +249,9 @@ class OpTest(unittest.TestCase):
                    max_relative_error=0.005):
         self.scope = core.Scope()
         op_inputs = self.inputs if hasattr(self, "inputs") else dict()
+        op_outputs = self.outputs if hasattr(self, "outputs") else dict()
         op_attrs = self.attrs if hasattr(self, "attrs") else dict()
-        self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs,
+        self.op = create_op(self.scope, self.op_type, op_inputs, op_outputs,
                             op_attrs)
         if no_grad_set is None:
             no_grad_set = set()
diff --git a/python/paddle/v2/framework/tests/test_accuracy_op.py b/python/paddle/v2/framework/tests/test_accuracy_op.py
index 43d60eb90d5edbd6944a11f7555f0291720dd2be..b6f3a35d6f58ba90b39e3f6296ae635220a2e965 100644
--- a/python/paddle/v2/framework/tests/test_accuracy_op.py
+++ b/python/paddle/v2/framework/tests/test_accuracy_op.py
@@ -6,16 +6,17 @@ from op_test import OpTest
 class TestAccuracyOp(OpTest):
     def setUp(self):
         self.op_type = "accuracy"
-        infer = np.random.randint(0, 2, (32, 1)).astype("int")
-        label = np.random.randint(0, 2, (32, )).astype("int")
+        n = 8192
+        infer = np.random.randint(0, 2, (n, 1)).astype("int")
+        label = np.random.randint(0, 2, (n, )).astype("int")
         self.inputs = {'Inference': infer, "Label": label}
         num_correct = 0
-        for rowid in xrange(32):
+        for rowid in xrange(n):
             for ele in infer[rowid]:
                 if ele == label[rowid]:
                     num_correct += 1
                     break
-        self.outputs = {'Accuracy': [num_correct / 32.0]}
+        self.outputs = {'Accuracy': [num_correct / float(n)]}
 
     def test_check_output(self):
         self.check_output()
diff --git a/python/paddle/v2/framework/tests/test_add_two_op.py b/python/paddle/v2/framework/tests/test_add_op.py
similarity index 100%
rename from python/paddle/v2/framework/tests/test_add_two_op.py
rename to python/paddle/v2/framework/tests/test_add_op.py
diff --git a/python/paddle/v2/framework/tests/test_cond_op.py b/python/paddle/v2/framework/tests/test_cond_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..37177ae0b2482517c4183969c8ef0670f2b3de89
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_cond_op.py
@@ -0,0 +1,116 @@
+import logging
+import paddle.v2.framework.core as core
+import unittest
+import numpy as np
+from paddle.v2.framework.op import Operator, CondOp
+
+
+class PySimpleCond(object):
+    '''
+    A simple implementation of dynamic if-else based on numpy
+    '''
+
+    def __init__(self):
+        array = [1] * 10
+        for i in range(1, 10, 2):
+            array[i] = 0
+        self.cond = np.array(array)
+        self.x = np.ones(shape=(10, 1))
+
+    def forward(self):
+        self.index_t = np.where(self.cond == 1)
+        self.index_f = np.where(self.cond == 0)
+        y_t = self.x[self.index_t]
+        y_f = self.x[self.index_f]
+        y_t = y_t * 2.
+        y_f = y_f * (-2.)
+        output = np.zeros(shape=(10, 1))
+        output[self.index_t] = y_t
+        output[self.index_f] = y_f
+        return output
+
+
+class PySimpleCondTest(unittest.TestCase):
+    def setUp(self):
+        self.condnn = PySimpleCond()
+
+    def test_forward(self):
+        output = self.condnn.forward()
+
+
+def create_tensor(scope, name, shape, np_data):
+    tensor = scope.new_var(name).get_tensor()
+    tensor.set_dims(shape)
+    tensor.set(np_data, core.CPUPlace())
+    return tensor
+
+
+class TestCondOp(unittest.TestCase):
+    '''
+    Test CondOp
+
+    equation:
+        cond = [True, False, True, False, ...]
+        y[index_t] = x[index_t] * 2.
+        y[index_f] = x[index_f] * -2.
+    outputs:
+        y
+    '''
+
+    def setUp(self):
+        self.py_cond = PySimpleCond()
+
+    def forward(self):
+        self.scope = core.Scope()
+        self.create_global_variables()
+        self.create_cond_op()
+        self.create_sub_net()
+        ctx = core.DeviceContext.create(core.CPUPlace())
+        self.condop.infer_shape(self.scope)
+        self.condop.run(self.scope, ctx)
+        return np.array(self.scope.find_var("Out").get_tensor())
+
+    def create_global_variables(self):
+        x_np_data = self.py_cond.x
+        create_tensor(self.scope, "X", [10, 1], x_np_data)
+        cond_np_data = self.py_cond.cond.astype("int32")
+        create_tensor(self.scope, "cond", [10, 1], cond_np_data)
+        self.scope.new_var("SubScopes")
+        self.scope.new_var("IndexTensors")
+        self.scope.new_var("Out")
+
+    def create_cond_op(self):
+        self.condop = CondOp(
+            Cond="cond",
+            Xs=["X"],
+            Outs=["Out"],
+            SubScopes="SubScopes",
+            IndexTensors="IndexTensors")
+
+    def create_sub_net(self):
+        truenet = core.Net.create()
+        scale_op_t = Operator("scale", X='X', Out='Out', scale=2.)
+        truenet.append_op(scale_op_t)
+        truenet.complete_add_op(True)
+        self.condop.set_truenet(truenet)
+
+        falsenet = core.Net.create()
+        scale_op_t = Operator("scale", X='X', Out='Out', scale=-2.)
+        falsenet.append_op(scale_op_t)
+        falsenet.complete_add_op(True)
+        self.condop.set_falsenet(falsenet)
+
+    def test_forward(self):
+        print 'test cond op forward'
+        pd_output = self.forward()
+        py_output = self.py_cond.forward()
+        print 'pd_output', pd_output
+        print
+        print 'py_output', py_output
+        self.assertEqual(pd_output.shape, py_output.shape)
+        print 'test passed'
+        return 0
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_cross_entropy_op.py b/python/paddle/v2/framework/tests/test_cross_entropy_op.py
index 253e7b8a24465da63a7eacd7983eb831251e6230..0206ca064be87afe204aa99021979b7ddc3c5d63 100644
--- a/python/paddle/v2/framework/tests/test_cross_entropy_op.py
+++ b/python/paddle/v2/framework/tests/test_cross_entropy_op.py
@@ -1,23 +1,25 @@
 import unittest
-import numpy
+import numpy as np
 from op_test import OpTest
 
 
-class TestCrossEntropy(OpTest):
+class TestCrossEntropyOp1(OpTest):
+    """Test standard cross-entropy, with index representation of labels.
+    """
+
     def setUp(self):
-        self.op_type = "onehot_cross_entropy"
+        self.op_type = "cross_entropy"
         batch_size = 30
         class_num = 10
-
-        X = numpy.random.uniform(0.1, 1.0,
-                                 [batch_size, class_num]).astype("float32")
-        labels = numpy.random.randint(0, class_num, batch_size, dtype="int32")
-
-        cross_entropy = numpy.asmatrix(
-            [[-numpy.log(X[i][labels[i]])] for i in range(X.shape[0])],
+        X = np.random.uniform(0.1, 1.0,
+                              [batch_size, class_num]).astype("float32")
+        label = np.random.randint(0, class_num, (batch_size, 1), dtype="int32")
+        cross_entropy = np.asmatrix(
+            [[-np.log(X[i][label[i][0]])] for i in range(X.shape[0])],
             dtype="float32")
-        self.inputs = {"X": X, "label": labels}
+        self.inputs = {"X": X, "Label": label}
         self.outputs = {"Y": cross_entropy}
+        self.attrs = {'soft_label': 0}
 
     def test_check_output(self):
         self.check_output()
@@ -26,5 +28,62 @@ class TestCrossEntropy(OpTest):
         self.check_grad(["X"], "Y")
 
 
+class TestCrossEntropyOp2(OpTest):
+    """Test soft-label cross-entropy, with vecterized soft labels.
+    """
+
+    def setUp(self):
+        self.op_type = "cross_entropy"
+        batch_size = 10
+        class_num = 5
+        X = np.random.uniform(0.1, 1.0,
+                              [batch_size, class_num]).astype("float32")
+        label = np.random.uniform(0.1, 1.0,
+                                  [batch_size, class_num]).astype("float32")
+        label /= label.sum(axis=1, keepdims=True)
+        cross_entropy = (-label * np.log(X)).sum(
+            axis=1, keepdims=True).astype("float32")
+        self.inputs = {'X': X, 'Label': label}
+        self.outputs = {'Y': cross_entropy}
+        self.attrs = {'soft_label': 1}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y')
+
+
+class TestCrossEntropyOp3(OpTest):
+    """Test one-hot cross-entropy, with vecterized one-hot representation of
+    labels.
+    """
+
+    def setUp(self):
+        self.op_type = "cross_entropy"
+        batch_size = 30
+        class_num = 10
+        X = np.random.uniform(0.1, 1.0,
+                              [batch_size, class_num]).astype("float32")
+        label_index = np.random.randint(
+            0, class_num, (batch_size), dtype="int32")
+        label = np.zeros(X.shape)
+        label[np.arange(batch_size), label_index] = 1
+        cross_entropy = np.asmatrix(
+            [[-np.log(X[i][label_index[i]])] for i in range(X.shape[0])],
+            dtype="float32")
+        cross_entropy2 = (-label * np.log(X)).sum(
+            axis=1, keepdims=True).astype("float32")
+        self.inputs = {'X': X, 'Label': label}
+        self.outputs = {'Y': cross_entropy}
+        self.attrs = {'soft_label': 1}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y')
+
+
 if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_dropout_op.py b/python/paddle/v2/framework/tests/test_dropout_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..3638fee1a1c26195791bc1f5a46dd749da0aee95
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_dropout_op.py
@@ -0,0 +1,59 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestDropoutOp(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {'dropout_prob': 0.0, 'is_training': 1}
+        self.outputs = {'Out': self.inputs['X'], 'Mask': np.ones((32, 64))}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad_normal(self):
+        self.check_grad(['X'], 'Out', max_relative_error=0.05)
+
+
+class TestDropoutOp2(TestDropoutOp):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {'dropout_prob': 1.0, 'is_training': 1}
+        self.outputs = {'Out': np.zeros((32, 64)), 'Mask': np.zeros((32, 64))}
+
+
+class TestDropoutOp3(TestDropoutOp):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64, 2)).astype("float32")}
+        self.attrs = {'dropout_prob': 0.0, 'is_training': 1}
+        self.outputs = {'Out': self.inputs['X'], 'Mask': np.ones((32, 64, 2))}
+
+
+class TestDropoutOp4(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {'dropout_prob': 0.35, 'is_training': 0}
+        self.outputs = {'Out': self.inputs['X'] * self.attrs['dropout_prob']}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestDropoutOp5(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64, 3)).astype("float32")}
+        self.attrs = {'dropout_prob': 0.75, 'is_training': 0}
+        self.outputs = {'Out': self.inputs['X'] * self.attrs['dropout_prob']}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_fc_op.py b/python/paddle/v2/framework/tests/test_fc_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..9f56fe5049c66aa5fce40ce815105e7871ebc3b2
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_fc_op.py
@@ -0,0 +1,62 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestFCOp1(OpTest):
+    def setUp(self):
+        x0 = np.random.random((16, 32)).astype("float32")
+        w0 = np.random.random((32, 10)).astype("float32")
+
+        mul_out0 = np.dot(x0, w0)
+        identity_out = mul_out0
+
+        self.op_type = "fc"
+        self.inputs = {"X": [("X0", x0)], "W": [("W0", w0)]}
+        self.outputs = {"MulOut": [("MulOut0", mul_out0)], "Out": identity_out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X0", "W0"], "Out", max_relative_error=0.01)
+
+
+class TestFCOp2(OpTest):
+    def setUp(self):
+        x0 = np.random.random((16, 4, 8)).astype("float32")
+        x1 = np.random.random((4, 4, 32)).astype("float32")
+        w0 = np.random.random((32, 10)).astype("float32")
+        w1 = np.random.random((32, 10)).astype("float32")
+        b = np.random.random(10).astype("float32")
+
+        mul_out0 = np.dot(x0.reshape(16, 4 * 8), w0)
+        mul_out1 = np.dot(x1.reshape(4 * 4, 32), w1)
+        sum_out = mul_out0 + mul_out1
+        add_out = np.add(sum_out, b)
+        sigmoid_out = 1 / (1 + np.exp(-add_out))
+
+        self.op_type = "fc"
+        self.inputs = {
+            "X": [("X0", x0), ("X1", x1)],
+            "W": [("W0", w0), ("W1", w1)],
+            "B": b
+        }
+        self.attrs = {"xNumColDims": [1, 2], "activation": "sigmoid"}
+        self.outputs = {
+            "MulOut": [("MulOut0", mul_out0), ("MulOut1", mul_out1)],
+            "SumOut": sum_out,
+            "AddOut": add_out,
+            "Out": sigmoid_out
+        }
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(
+            ["X0", "X1", "W0", "W1", "B"], "Out", max_relative_error=0.01)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_gaussian_random_op.py b/python/paddle/v2/framework/tests/test_gaussian_random_op.py
index 1f9e4db783c9907a22db72c8a6ff06c7ca0735da..1888ee28f92c66496ce756d8a4a33d3e9ba57d7b 100644
--- a/python/paddle/v2/framework/tests/test_gaussian_random_op.py
+++ b/python/paddle/v2/framework/tests/test_gaussian_random_op.py
@@ -4,7 +4,7 @@ from paddle.v2.framework.op import Operator
 import numpy
 
 
-class GaussianRandomTest(unittest.TestCase):
+class TestGaussianRandomOp(unittest.TestCase):
     def test_cpu(self):
         self.gaussian_random_test(place=core.CPUPlace())
 
diff --git a/python/paddle/v2/framework/tests/test_identity_op.py b/python/paddle/v2/framework/tests/test_identity_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..26cec1fcc3ad003281c9c41571d475b55bd30026
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_identity_op.py
@@ -0,0 +1,20 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestIdentityOp(OpTest):
+    def setUp(self):
+        self.op_type = "identity"
+        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
+        self.outputs = {'Y': self.inputs['X']}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Y')
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_lookup_table.py b/python/paddle/v2/framework/tests/test_lookup_table_op.py
similarity index 100%
rename from python/paddle/v2/framework/tests/test_lookup_table.py
rename to python/paddle/v2/framework/tests/test_lookup_table_op.py
diff --git a/python/paddle/v2/framework/tests/test_minus_op.py b/python/paddle/v2/framework/tests/test_minus_op.py
index dea797a1fea34265d0a32e097f413f421abf2521..c56d7cb548706880dd482bad750f2989c0e9a710 100644
--- a/python/paddle/v2/framework/tests/test_minus_op.py
+++ b/python/paddle/v2/framework/tests/test_minus_op.py
@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class MinusOpTest(OpTest):
+class TestMinusOp(OpTest):
     def setUp(self):
         self.op_type = "minus"
         self.inputs = {
diff --git a/python/paddle/v2/framework/tests/test_mnist.py b/python/paddle/v2/framework/tests/test_mnist.py
index f6f8f49b797fb6e5016a5e309f12f192d5096431..66452cb3965d28fd15e814833079621410775c17 100644
--- a/python/paddle/v2/framework/tests/test_mnist.py
+++ b/python/paddle/v2/framework/tests/test_mnist.py
@@ -128,7 +128,7 @@ def fc_layer(net, input, size, act="softmax", bias=True, param=None, name=None):
 def cross_entropy_layer(net, input, label):
     cost_name = "cross_entropy_%d" % uniq_id()
     cross_entropy_op = Operator(
-        "onehot_cross_entropy", X=input, label=label, Y=cost_name)
+        "cross_entropy", X=input, Label=label, Y=cost_name)
     net.append_op(cross_entropy_op)
     scope.new_var(cost_name)
     net.infer_shape(scope)
@@ -181,7 +181,7 @@ def error_rate(predict, label):
 
 
 images = data_layer(name="pixel", dims=[BATCH_SIZE, 784])
-labels = data_layer(name="label", dims=[BATCH_SIZE])
+labels = data_layer(name="label", dims=[BATCH_SIZE, 1])
 fc1 = fc_layer(net=forward_net, input=images, size=100, act="sigmoid")
 fc2 = fc_layer(net=forward_net, input=fc1, size=100, act="sigmoid")
 predict = fc_layer(net=forward_net, input=fc2, size=10, act="softmax")
@@ -215,6 +215,7 @@ def test(cost_name):
     for data in test_reader():
         image_data = numpy.array(map(lambda x: x[0], data)).astype("float32")
         label_data = numpy.array(map(lambda x: x[1], data)).astype("int32")
+        label_data = numpy.expand_dims(label_data, axis=1)
         feed_data(images, image_data)
         feed_data(labels, label_data)
 
@@ -235,6 +236,7 @@ for pass_id in range(PASS_NUM):
     for data in train_reader():
         image_data = numpy.array(map(lambda x: x[0], data)).astype("float32")
         label_data = numpy.array(map(lambda x: x[1], data)).astype("int32")
+        label_data = numpy.expand_dims(label_data, axis=1)
         feed_data(images, image_data)
         feed_data(labels, label_data)
 
diff --git a/python/paddle/v2/framework/tests/test_prelu_op.py b/python/paddle/v2/framework/tests/test_prelu_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..2b6b7db36808a4b68c55328a1eb9ac212c18b678
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_prelu_op.py
@@ -0,0 +1,28 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class PReluTest(OpTest):
+    def setUp(self):
+        self.op_type = "prelu"
+        x_np = np.random.normal(size=(10, 10)).astype("float32")
+        x_np_sign = np.sign(x_np)
+        x_np = x_np_sign * np.maximum(x_np, .005)
+        alpha_np = np.array([.1])
+        self.inputs = {'X': x_np, 'Alpha': alpha_np}
+        out_np = np.maximum(self.inputs['X'], 0.)
+        out_np = out_np + np.minimum(self.inputs['X'],
+                                     0.) * self.inputs['Alpha']
+        assert out_np is not self.inputs['X']
+        self.outputs = {'Out': out_np}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], 'Out')
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_scale_and_identity_op.py b/python/paddle/v2/framework/tests/test_scale_op.py
similarity index 56%
rename from python/paddle/v2/framework/tests/test_scale_and_identity_op.py
rename to python/paddle/v2/framework/tests/test_scale_op.py
index 05d76d428299c8176d1a6adf6da15a203fa7502a..2ea1e185470280730ae8c8c0ea9568bbeb43eaf5 100644
--- a/python/paddle/v2/framework/tests/test_scale_and_identity_op.py
+++ b/python/paddle/v2/framework/tests/test_scale_op.py
@@ -3,20 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class IdentityTest(OpTest):
-    def setUp(self):
-        self.op_type = "identity"
-        self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
-        self.outputs = {'Out': self.inputs['X']}
-
-    def test_check_output(self):
-        self.check_output()
-
-    def test_check_grad(self):
-        self.check_grad(['X'], 'Out')
-
-
-class ScaleTest(OpTest):
+class TestScaleOp(OpTest):
     def setUp(self):
         self.op_type = "scale"
         self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
diff --git a/python/paddle/v2/framework/tests/test_seq_pool.py b/python/paddle/v2/framework/tests/test_seq_pool.py
new file mode 100644
index 0000000000000000000000000000000000000000..cf864936af6361da1f16df3cfb759b468214b970
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_seq_pool.py
@@ -0,0 +1,51 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSeqAvgPool1D(OpTest):
+    def setUp(self):
+        self.op_type = 'sequence_avg_pool'
+        # one level, batch size is 4
+        x = np.random.uniform(0.1, 1, [11, 23]).astype('float32')
+        lod = [[0, 4, 5, 8, 11]]
+
+        out = np.zeros((4, 23)).astype('float32')
+        for i in range(4):
+            sub_x = x[lod[0][i]:lod[0][i + 1], :]
+            out[i] = sub_x.mean(axis=0)
+
+        self.inputs = {'X': (x, lod)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+class TestSeqAvgPool2D(OpTest):
+    def setUp(self):
+        self.op_type = 'sequence_avg_pool'
+        # one level, batch size is 4
+        x = np.random.uniform(0.1, 1, [13, 3, 17]).astype('float32')
+        lod = [[0, 4, 5, 8, 13]]
+
+        out = np.zeros((4, 3, 17)).astype('float32')
+        for i in range(4):
+            sub_x = np.reshape(x[lod[0][i]:lod[0][i + 1], :], (-1, 3 * 17))
+            out[i] = np.reshape(sub_x.mean(axis=0), (3, 17))
+
+        self.inputs = {'X': (x, lod)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(["X"], "Out")
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_sgd_op.py b/python/paddle/v2/framework/tests/test_sgd_op.py
index 557cf15ace63e336462c7dcdbbc10f30aeedc6f4..64e54d1500c1bc134cc1efe33d41a16dbc08f2d4 100644
--- a/python/paddle/v2/framework/tests/test_sgd_op.py
+++ b/python/paddle/v2/framework/tests/test_sgd_op.py
@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class TestSGD(OpTest):
+class TestSGDOp(OpTest):
     def setUp(self):
         self.op_type = "sgd"
         w = np.random.random((102, 105)).astype("float32")
diff --git a/python/paddle/v2/framework/tests/test_sigmoid_op.py b/python/paddle/v2/framework/tests/test_sigmoid_op.py
index 2316e49eff7bb1cdb53acb3889a6ef05060b59f3..d65d887db4af58c40e4e78fdbfd8e8ee668b7ee3 100644
--- a/python/paddle/v2/framework/tests/test_sigmoid_op.py
+++ b/python/paddle/v2/framework/tests/test_sigmoid_op.py
@@ -3,7 +3,7 @@ import numpy as np
 from op_test import OpTest
 
 
-class TestSigmoid(OpTest):
+class TestSigmoidOp(OpTest):
     def setUp(self):
         self.op_type = "sigmoid"
         self.inputs = {
diff --git a/python/paddle/v2/framework/tests/test_split_op.py b/python/paddle/v2/framework/tests/test_split_op.py
new file mode 100644
index 0000000000000000000000000000000000000000..b4420db9d71b99556e305104ac17ef5e4b4bd0f2
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_split_op.py
@@ -0,0 +1,26 @@
+import unittest
+import numpy as np
+from op_test import OpTest
+
+
+class TestSplitOp(OpTest):
+    def setUp(self):
+        self.op_type = "split"
+        axis = 0
+        num = 2
+        x = np.random.random((4, 2)).astype('float32')
+        out = np.split(x, num, axis)
+        self.inputs = {'X': x}
+        self.attrs = {'axis': axis, 'num': num}
+        self.outputs = {'Out': [('out%d' % i, out[i]) \
+            for i in xrange(len(out))]}
+
+    def test_check_output(self):
+        self.check_output()
+
+    def test_check_grad(self):
+        self.check_grad(['X'], ['out0', 'out1'])
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_top_k_op.py b/python/paddle/v2/framework/tests/test_top_k_op.py
index cab799256d791889c295aa7f9048080f5caaf2dc..694f37d612d4c46e673dc894b05a0a446190732c 100644
--- a/python/paddle/v2/framework/tests/test_top_k_op.py
+++ b/python/paddle/v2/framework/tests/test_top_k_op.py
@@ -21,6 +21,9 @@ class TestTopkOp(OpTest):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
+    def test_check_output(self):
+        self.check_output()
+
 
 class TestTopkOp3d(OpTest):
     def setUp(self):
@@ -42,6 +45,9 @@ class TestTopkOp3d(OpTest):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
+    def test_check_output(self):
+        self.check_output()
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_uniform_random_op.py b/python/paddle/v2/framework/tests/test_uniform_random_op.py
index 76a5e36e56ab08230bdc2597d209fcf5d1d2acb0..9e8898fb5920defdfaa361bf45def7666a88beea 100644
--- a/python/paddle/v2/framework/tests/test_uniform_random_op.py
+++ b/python/paddle/v2/framework/tests/test_uniform_random_op.py
@@ -4,7 +4,7 @@ import paddle.v2.framework.core as core
 import numpy
 
 
-class UniformRandomTest(unittest.TestCase):
+class TestUniformRandomOp(unittest.TestCase):
     def test_uniform_random_cpu(self):
         self.uniform_random_test(place=core.CPUPlace())