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

3f3848cd · wanghaoshuang · 2cde56c5 · d865b047 · 3f3848cd · 3f3848cd
42 changed file
--- 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:

--- 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(ANDROID OR IOS)
-    if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
+    if(ANDROID)
-        message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
+        if(AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
-    elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
+            message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
-        # TODO: support glog for Android api 16 ~ 19 in the future
+        elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
-        message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
+            # TODO: support glog for Android api 16 ~ 19 in the future
+            message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
+        endif()
    endif()
    set(WITH_GPU OFF CACHE STRING
-        "Disable GPU when cross-compiling for Android" FORCE)
+        "Disable GPU when cross-compiling for Android and iOS" FORCE)
    set(WITH_AVX OFF CACHE STRING
-        "Disable AVX when cross-compiling for Android" FORCE)
+        "Disable AVX when cross-compiling for Android and iOS" FORCE)
    set(WITH_PYTHON OFF CACHE STRING
-        "Disable PYTHON when cross-compiling for Android" FORCE)
+        "Disable PYTHON when cross-compiling for Android and iOS" FORCE)
    set(WITH_RDMA OFF CACHE STRING
-        "Disable RDMA when cross-compiling for Android" FORCE)
+        "Disable RDMA when cross-compiling for Android and iOS" FORCE)
    set(WITH_MKLDNN OFF CACHE STRING
-        "Disable MKLDNN when cross-compiling for Android" FORCE)
+        "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
    set(WITH_MKLML OFF CACHE STRING
-        "Disable MKLML package when cross-compiling for Android" FORCE)
+        "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
-endif(ANDROID)
+endif()
 set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
  "A path setting third party libraries download & build directories.")

--- 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()
--- a/cmake/cross_compiling/ios.cmake
+++ b/cmake/cross_compiling/ios.cmake
+# 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)
--- 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=${CMAKE_C_COMPILER}
+                    -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-    CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
+                    -DCMAKE_INSTALL_PREFIX=${GFLAGS_INSTALL_DIR}
-    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-    CMAKE_ARGS      -DBUILD_TESTING=OFF
+                    -DBUILD_TESTING=OFF
-    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                    -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

--- 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=${CMAKE_C_COMPILER}
+                    -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-    CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
+                    -DCMAKE_INSTALL_PREFIX=${GLOG_INSTALL_DIR}
-    CMAKE_ARGS      -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib
+                    -DCMAKE_INSTALL_LIBDIR=${GLOG_INSTALL_DIR}/lib
-    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-    CMAKE_ARGS      -DWITH_GFLAGS=ON
+                    -DWITH_GFLAGS=ON
-    CMAKE_ARGS      -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags
+                    -Dgflags_DIR=${GFLAGS_INSTALL_DIR}/lib/cmake/gflags
-    CMAKE_ARGS      -DBUILD_TESTING=OFF
+                    -DBUILD_TESTING=OFF
-    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                    -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

--- 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=${CMAKE_C_COMPILER}
+                        -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-        CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                        -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-        CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                        -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-        CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${GTEST_INSTALL_DIR}
+                        -DCMAKE_INSTALL_PREFIX=${GTEST_INSTALL_DIR}
-        CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-        CMAKE_ARGS      -DBUILD_GMOCK=ON
+                        -DBUILD_GMOCK=ON
-        CMAKE_ARGS      -Dgtest_disable_pthreads=ON
+                        -Dgtest_disable_pthreads=ON
-        CMAKE_ARGS      -Dgtest_force_shared_crt=ON
+                        -Dgtest_force_shared_crt=ON
-        CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                        -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

--- 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 "${CMAKE_C_COMPILER}")
-        SET(OPENBLAS_CC "${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=${CMAKE_C_COMPILER} NO_SHARED=1 NO_LAPACK=1 libs)
-    ENDIF()
    IF(CMAKE_CROSSCOMPILING)
+        SET(OPTIONAL_ARGS HOSTCC=${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}

--- 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()

--- 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)
-    FIND_PACKAGE(PythonLibs 2.7)
+# Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE.
-    # Fixme: Maybe find a static library. Get SHARED/STATIC by FIND_PACKAGE.
+ADD_LIBRARY(python SHARED IMPORTED GLOBAL)
-    ADD_LIBRARY(python SHARED IMPORTED GLOBAL)
+SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES})
-    SET_PROPERTY(TARGET python PROPERTY IMPORTED_LOCATION ${PYTHON_LIBRARIES})
-ENDIF(WITH_PYTHON)
 SET(py_env "")
 IF(PYTHONINTERP_FOUND)
@@ -36,9 +37,5 @@ IF(PYTHONINTERP_FOUND)
    ENDIF()
 ENDIF(PYTHONINTERP_FOUND)
-IF(WITH_PYTHON)
+INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR})
-    INCLUDE_DIRECTORIES(${PYTHON_INCLUDE_DIR})
+INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR})
-    INCLUDE_DIRECTORIES(${PYTHON_NUMPY_INCLUDE_DIR})
-ELSE()
-    SET(PYTHON_LIBRARIES "")
-ENDIF()
--- 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)

--- 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(${WARPCTC_INCLUDE_DIR})
+SET(WARPCTC_INCLUDE_DIR "${WARPCTC_INSTALL_DIR}/include"
+    CACHE PATH "Warp-ctc Directory" FORCE)
-SET(WARPCTC_LIB_DIR "${WARPCTC_INSTALL_DIR}/lib" CACHE PATH "Warp-ctc Library Directory" FORCE)
+# Used in unit test test_WarpCTCLayer
+SET(WARPCTC_LIB_DIR "${WARPCTC_INSTALL_DIR}/lib"
-IF(WIN32)
+    CACHE PATH "Warp-ctc Library Directory" FORCE)
-    SET(WARPCTC_LIBRARIES
+SET(WARPCTC_LIBRARIES "${WARPCTC_INSTALL_DIR}/lib/libwarpctc${CMAKE_SHARED_LIBRARY_SUFFIX}"
-        "${WARPCTC_INSTALL_DIR}/lib/warpctc.dll" CACHE FILEPATH "Warp-ctc Library" FORCE)
+    CACHE FILEPATH "Warp-ctc Library" FORCE)
-ELSE(WIN32)
-    IF(APPLE)
-        SET(_warpctc_SHARED_SUFFIX dylib)
-    ELSE(APPLE)
-        SET(_warpctc_SHARED_SUFFIX so)
-    ENDIF(APPLE)
-    SET(WARPCTC_LIBRARIES
-        "${WARPCTC_INSTALL_DIR}/lib/libwarpctc.${_warpctc_SHARED_SUFFIX}" CACHE FILEPATH "Warp-ctc Library" FORCE)
-ENDIF(WIN32)
 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=${CMAKE_C_COMPILER}
+                    -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-    CMAKE_ARGS      -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-    CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR}
+                    -DCMAKE_INSTALL_PREFIX=${WARPCTC_INSTALL_DIR}
-    CMAKE_ARGS      -DWITH_GPU=${WITH_GPU}
+                    -DWITH_GPU=${WITH_GPU}
-    CMAKE_ARGS      -DWITH_OMP=${USE_OMP}
+                    -DWITH_OMP=${USE_OMP}
-    CMAKE_ARGS      -DWITH_TORCH=OFF
+                    -DWITH_TORCH=OFF
-    CMAKE_ARGS      -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON
+                    -DCMAKE_DISABLE_FIND_PACKAGE_Torch=ON
-    CMAKE_ARGS      -DBUILD_SHARED=ON
+                    -DBUILD_SHARED=ON
-    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                    -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)

--- 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}
+                    -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-    CMAKE_ARGS      -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
+                    -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-    CMAKE_ARGS      -DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR}
+                    -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-    CMAKE_ARGS      -DBUILD_SHARED_LIBS=OFF
+                    -DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR}
-    CMAKE_ARGS      -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+                    -DBUILD_SHARED_LIBS=OFF
-    CMAKE_ARGS      -DCMAKE_MACOSX_RPATH=ON
+                    -DCMAKE_POSITION_INDEPENDENT_CODE=ON
-    CMAKE_ARGS      -DCMAKE_BUILD_TYPE=Release
+                    -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

--- 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.
+    if(NOT CMAKE_CROSSCOMPILING)
-    set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
+        # 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

--- 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()

--- 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}

--- 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)

--- 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)
+set(PADDLE_CAPI_INFER_LIBS
-add_custom_target(paddle_capi_whole ALL
+    paddle_utils
-        COMMAND mkdir -p o_files/capi && cd o_files/capi/ && ar -x $<TARGET_FILE:paddle_capi>
+    paddle_parameter
-        COMMAND mkdir -p o_files/utils && cd o_files/utils/ && ar -x $<TARGET_FILE:paddle_utils>
+    paddle_math
-        COMMAND mkdir -p o_files/parameter && cd o_files/parameter/ && ar -x $<TARGET_FILE:paddle_parameter>
+    paddle_cuda
-        COMMAND mkdir -p o_files/math && cd o_files/math/  && ar -x $<TARGET_FILE:paddle_math>
+    paddle_function
-        COMMAND mkdir -p o_files/cuda && cd o_files/cuda/ && ar -x $<TARGET_FILE:paddle_cuda>
+    paddle_gserver
-        COMMAND mkdir -p o_files/function && cd o_files/function/ && ar -x $<TARGET_FILE:paddle_function>
+    paddle_proto
-        COMMAND mkdir -p o_files/gserver && cd o_files/gserver/ && ar -x $<TARGET_FILE:paddle_gserver>
+    paddle_pserver
-        COMMAND mkdir -p o_files/proto && cd o_files/proto/ && ar -x $<TARGET_FILE:paddle_proto>
+    paddle_network)
-        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>
+cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
-        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(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
+# No shared library for iOS
-# TODO: merge mkl into paddle_capi_shared
+if(NOT IOS)
-add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
+  set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
-set_target_properties(paddle_capi_shared	PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
+  # TODO: merge mkl into paddle_capi_shared
-target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
+  add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
-link_paddle_exe(paddle_capi_shared)
+  set_target_properties(paddle_capi_shared	PROPERTIES LINK_FLAGS "${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_whole
-  install(TARGETS paddle_capi_shared DESTINATION lib)
+          ARCHIVE DESTINATION lib)
+  if(NOT IOS)
+    install(TARGETS paddle_capi_shared DESTINATION lib)
+  endif()
 endif(ANDROID)
 # this variable used for unittest

--- 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());

--- 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);
 }

--- a/paddle/gserver/activations/MKLDNNActivation.cpp
+++ b/paddle/gserver/activations/MKLDNNActivation.cpp
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#include "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
--- a/paddle/gserver/activations/MKLDNNActivation.h
+++ b/paddle/gserver/activations/MKLDNNActivation.h
+/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+    http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+#pragma once
+#include "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
--- a/paddle/gserver/layers/MKLDNNConvLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNConvLayer.cpp
@@ -294,12 +294,9 @@ 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_};

--- a/paddle/gserver/layers/MKLDNNFcLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNFcLayer.cpp
@@ -172,12 +172,10 @@ void MKLDNNFcLayer::resetWgtBiasValue(MKLDNNMatrixPtr& wgt,
 void MKLDNNFcLayer::resetOutValue(MKLDNNMatrixPtr& out) {
  out = MKLDNNMatrix::create(output_.value, {bs_, oc_}, format::nc, engine_);
-  // change original output value to mkldnn output value
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
  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());
  }
 }

--- 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;
      }

--- a/paddle/gserver/layers/MKLDNNPoolLayer.cpp
+++ b/paddle/gserver/layers/MKLDNNPoolLayer.cpp
@@ -134,7 +134,6 @@ void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
  memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
  out = MKLDNNMatrix::create(
      output_.value, outDims, inVal_->getFormat(), engine_);
-  output_.value = std::dynamic_pointer_cast<Matrix>(out);
  // create reorder if output value has cpu device and pd do not match
  cpuOutVal_ = nullptr;

--- 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()),
+  CHECK_EQ(
-           inputSeq.hasSubseq() ? inputSeq.getNumSubSequences()
+      indices1->getHeight(),
-                                : inputSeq.getNumSequences())
+      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)

--- 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]);
+  dnnLayer_ = testLayers_[DNN];
-  CHECK(dnnLayer_);
-  // for comparison with Paddle reference results,
-  // need manually add cpu device output for test
-  dnnLayer_->addOutputArgument(CPU_DEVICE);
  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()
+  CHECK(dnn.layerConfig.type().compare(0, 7, "mkldnn_") == 0 ||
-                     << " vs " << ref.layerConfig.type();
+        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;

--- 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_;
+  LayerPtr refLayer_, dnnLayer_;
-  MKLDNNLayerPtr dnnLayer_;
  /// run some iterations, all the result should pass
  size_t iter_;

--- 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
@@ -162,7 +163,6 @@ void testPoolLayer(const testPoolDesc& pm) {
       0});
  LayerInputConfig* input = cfg.layerConfig.add_inputs();
  PoolConfig* pool = input->mutable_pool_conf();
-  // pool->set_pool_type(poolType);
  pool->set_channels(pm.ch);
  pool->set_img_size(pm.iw);
  pool->set_img_size_y(pm.ih);
@@ -191,7 +191,7 @@ void testPoolLayer(const testPoolDesc& pm) {
  }
 }
-TEST(MkldnnLayer, PoolLayer) {
+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});
@@ -203,6 +203,49 @@ TEST(MkldnnLayer, PoolLayer) {
  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) {

--- 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>

--- a/paddle/operators/prelu_op.h
+++ b/paddle/operators/prelu_op.h
@@ -54,8 +54,9 @@ class PReluKernel : public framework::OpKernel {
    int numel = x->numel();
-    Transform(context.device_context(), x_ptr, x_ptr + numel, o_ptr,
+    Transform<Place> trans;
-              PReluFunctor<T>(alpha_ptr));
+    trans(context.device_context(), x_ptr, x_ptr + numel, o_ptr,
+          PReluFunctor<T>(alpha_ptr));
  }
 };
@@ -91,8 +92,9 @@ class PReluGradKernel : public framework::OpKernel {
    const T* out_ptr = out->data<T>();
    int numel = dx->numel();
-    Transform(context.device_context(), out_ptr, out_ptr + numel, dout_ptr,
+    Transform<Place> trans;
-              dx_ptr, PReluGradFunctor<T>(alpha_ptr));
+    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
  }

--- a/paddle/platform/transform.h
+++ b/paddle/platform/transform.h
@@ -29,45 +29,71 @@
 namespace paddle {
 namespace platform {
 // Transform on host or device. It provides the same API in std library.
-template <typename InputIter, typename OutputIter, typename UnaryOperation>
+template <typename Place>
-void Transform(const DeviceContext& context, InputIter first, InputIter last,
+struct Transform {
-               OutputIter result, UnaryOperation op) {
+  template <typename InputIter, typename OutputIter, typename UnaryOperation>
-  auto place = context.GetPlace();
+  void operator()(const DeviceContext& context, InputIter first, InputIter last,
-  if (is_cpu_place(place)) {
+                  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);
-  } else {
-#ifdef __NVCC__
-    auto& ctx = reinterpret_cast<const CUDADeviceContext&>(context);
-    using namespace details;
-    thrust::transform(thrust::cuda::par.on(ctx.stream()), DevPtrCast(first),
-                      DevPtrCast(last), DevPtrCast(result), op);
-#else
-    PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
-#endif
  }
-}
-template <typename InputIter1, typename InputIter2, typename OutputIter,
+  template <typename InputIter1, typename InputIter2, typename OutputIter,
-          typename BinaryOperation>
+            typename BinaryOperation>
-void Transform(const DeviceContext& context, InputIter1 first1,
+  void operator()(const DeviceContext& context, InputIter1 first1,
-               InputIter1 last1, InputIter2 first2, OutputIter result,
+                  InputIter1 last1, InputIter2 first2, OutputIter result,
-               BinaryOperation op) {
+                  BinaryOperation op) {
-  auto place = context.GetPlace();
+    auto place = context.GetPlace();
-  if (is_cpu_place(place)) {
+    PADDLE_ENFORCE(is_cpu_place(place), "It must use CPU place.");
    std::transform(first1, last1, first2, result, op);
-  } else {
+  }
+};
 #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);
-    using namespace details;
+    thrust::transform(thrust::cuda::par.on(ctx.stream()),
-    thrust::transform(thrust::cuda::par.on(ctx.stream()), DevPtrCast(first1),
+                      details::DevPtrCast(first), details::DevPtrCast(last),
-                      DevPtrCast(last1), DevPtrCast(first2), DevPtrCast(result),
+                      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);
-#else
-    PADDLE_THROW("Do not invoke `Transform<GPUPlace>` in .cc file");
-#endif
  }
 };
+#endif
 }  // namespace platform
 }  // namespace paddle
--- a/paddle/platform/transform_test.cu
+++ b/paddle/platform/transform_test.cu
@@ -15,6 +15,7 @@
 #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>
@@ -38,7 +39,8 @@ TEST(Transform, CPUUnary) {
  using namespace paddle::platform;
  CPUDeviceContext ctx;
  float buf[4] = {0.1, 0.2, 0.3, 0.4};
-  Transform(ctx, buf, buf + 4, buf, Scale<float>(10));
+  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);
  }
@@ -52,7 +54,8 @@ TEST(Transform, GPUUnary) {
  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(ctx, gpu_buf, gpu_buf + 4, gpu_buf, Scale<float>(10));
+  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);
@@ -65,7 +68,9 @@ TEST(Transform, CPUBinary) {
  using namespace paddle::platform;
  using namespace paddle::memory;
  int buf[4] = {1, 2, 3, 4};
-  Transform(CPUDeviceContext(), buf, buf + 4, buf, buf, Multiply<int>());
+  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]);
  }
@@ -79,11 +84,12 @@ TEST(Transform, GPUBinary) {
  CUDADeviceContext ctx(gpu0);
  int* gpu_buf = static_cast<int*>(Alloc(gpu0, sizeof(buf)));
  Copy(gpu0, gpu_buf, CPUPlace(), buf, sizeof(buf));
-  Transform(ctx, gpu_buf, gpu_buf + 4, gpu_buf, gpu_buf, Multiply<int>());
+  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]);
  }
 }
\ No newline at end of file
--- 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
+if(NOT WITH_C_API)
-    PERMISSIONS OWNER_EXECUTE OWNER_WRITE OWNER_READ
+  add_executable(paddle_pserver_main ${PSERVER_MAIN_SOURCES})
-        GROUP_EXECUTE GROUP_READ WORLD_EXECUTE WORLD_READ)
+  link_paddle_exe(paddle_pserver_main)
-set_target_properties(paddle_pserver_main PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE)
+  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()
--- a/paddle/scripts/travis/build_ios.sh
+++ b/paddle/scripts/travis/build_ios.sh
+#!/bin/bash
+set -e
+# Create the build directory for CMake.
+mkdir -p $TRAVIS_BUILD_DIR/build_ios
+cd $TRAVIS_BUILD_DIR/build_ios
+# Compile paddle binaries
+cmake -DCMAKE_SYSTEM_NAME=iOS \
+      -DIOS_PLATFORM=OS \
+      -DCMAKE_OSX_ARCHITECTURES="arm64" \
+      -DWITH_C_API=ON \
+      -DUSE_EIGEN_FOR_BLAS=ON \
+      -DWITH_TESTING=OFF \
+      -DWITH_SWIG_PY=OFF \
+      -DWITH_STYLE_CHECK=OFF \
+      -DCMAKE_BUILD_TYPE=Release \
+      ..
+make -j 2
--- a/paddle/scripts/travis/check_style.sh
+++ b/paddle/scripts/travis/check_style.sh
@@ -8,6 +8,12 @@ function abort(){
 trap 'abort' 0
 set -e
+# install glide
+curl https://glide.sh/get | bash
+eval "$(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

--- 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)
+if(NOT WITH_C_API)
-set_target_properties(paddle_merge_model PROPERTIES INSTALL_RPATH_USE_LINK_PATH TRUE)
+  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)
--- 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);

--- 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(); }

--- 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;

--- a/python/paddle/v2/framework/tests/test_prelu_op.py
+++ b/python/paddle/v2/framework/tests/test_prelu_op.py
@@ -17,10 +17,10 @@ class PReluTest(OpTest):
        assert out_np is not self.inputs['X']
        self.outputs = {'Out': out_np}
-    def not_test_check_output(self):
+    def test_check_output(self):
        self.check_output()
-    def not_test_check_grad(self):
+    def test_check_grad(self):
        self.check_grad(['X'], 'Out')