Merge pull request #5714 from tensor-tang/refine

add option WITH_MKL

Author: luotao1

CMakeLists.txt

@@ -36,8 +36,7 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
 option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
-option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." ${AVX_FOUND})
-option(WITH_MKLML "Compile PaddlePaddle with mklml package." ${AVX_FOUND})
+option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FOUND})
 option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING "Compile PaddlePaddle with unit testing" ON)
 option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)
@@ -82,10 +81,8 @@ if(ANDROID OR IOS)
  "Disable PYTHON when cross-compiling for Android and iOS" FORCE)
  set(WITH_RDMA OFF CACHE STRING
  "Disable RDMA when cross-compiling for Android and iOS" FORCE)
- set(WITH_MKLDNN OFF CACHE STRING
- "Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
- set(WITH_MKLML OFF CACHE STRING
- "Disable MKLML package when cross-compiling for Android and iOS" FORCE)
+ set(WITH_MKL OFF CACHE STRING
+ "Disable MKL when cross-compiling for Android and iOS" FORCE)
 
  # Compile PaddlePaddle mobile inference library
  if (NOT WITH_C_API)
@@ -111,6 +108,17 @@ else()
  set(THIRD_PARTY_BUILD_TYPE Release)
 endif()
 
+if(WITH_MKL)
+ set(WITH_MKLML ON)
+ set(WITH_MKLDNN ${AVX2_FOUND})
+ if(NOT WITH_MKLDNN)
+ message(WARNING "Do not have AVX2 intrinsics and disabled MKL-DNN")
+ endif()
+else()
+ set(WITH_MKLML OFF)
+ set(WITH_MKLDNN OFF)
+endif()
+
 ########################################################################################
 
 include(external/mklml) # download mklml package
@@ -164,8 +172,12 @@ if(WITH_GPU)
  endif(NOT WITH_DSO)
 endif(WITH_GPU)
 
+if(WITH_MKLML)
+ list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
+endif()
+
 if(WITH_MKLDNN)
- list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
+ list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
 endif()
 
 if(USE_NNPACK)

cmake/configure.cmake

@@ -76,27 +76,14 @@ else()
  include_directories(${CUDA_TOOLKIT_INCLUDE})
 endif(NOT WITH_GPU)
 
-if(WITH_MKLDNN)
- add_definitions(-DPADDLE_USE_MKLDNN)
- if (WITH_MKLML AND MKLDNN_IOMP_DIR)
- message(STATUS "Enable Intel OpenMP at ${MKLDNN_IOMP_DIR}")
- set(OPENMP_FLAGS "-fopenmp")
- set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
- set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
- set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")
- set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
- else()
- find_package(OpenMP)
- if(OPENMP_FOUND)
- set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
- set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
- else()
- message(WARNING "Can not find OpenMP."
- "Some performance features in MKLDNN may not be available")
- endif()
- endif()
-
-endif(WITH_MKLDNN)
+if (WITH_MKLML AND MKLML_IOMP_LIB)
+ message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}")
+ set(OPENMP_FLAGS "-fopenmp")
+ set(CMAKE_C_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
+ set(CMAKE_CXX_CREATE_SHARED_LIBRARY_FORBIDDEN_FLAGS ${OPENMP_FLAGS})
+ set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OPENMP_FLAGS}")
+ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OPENMP_FLAGS}")
+endif()
 
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}")

cmake/external/mkldnn.cmake

@@ -40,10 +40,9 @@ INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR})
 
 IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
  SET(MKLDNN_DEPENDS ${MKLML_PROJECT})
- SET(MKLDNN_MKLROOT ${MKLML_ROOT})
- SET(MKLDNN_IOMP_LIB ${MKLML_IOMP_LIB})
- SET(MKLDNN_IOMP_DIR ${MKLML_LIB_DIR})
- MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
+ MESSAGE(STATUS "Build MKLDNN with MKLML ${MKLML_ROOT}")
+ELSE()
+ MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
 ENDIF()
 
 SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} -Wno-error=strict-overflow")
@@ -57,15 +56,16 @@ ExternalProject_Add(
  PREFIX ${MKLDNN_SOURCES_DIR}
  UPDATE_COMMAND ""
  CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
- CMAKE_ARGS -DMKLROOT=${MKLDNN_MKLROOT}
+ CMAKE_ARGS -DMKLROOT=${MKLML_ROOT}
  CMAKE_ARGS -DCMAKE_C_FLAGS=${MKLDNN_CFLAG}
  CMAKE_ARGS -DCMAKE_CXX_FLAGS=${MKLDNN_CXXFLAG}
  CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
- -DMKLROOT:PATH=${MKLDNN_MKLROOT}
+ -DMKLROOT:PATH=${MKLML_ROOT}
 )
 
 ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
 SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
 ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
-MESSAGE(STATUS "Mkldnn library: ${MKLDNN_LIB}")
+MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
+add_definitions(-DPADDLE_USE_MKLDNN)
 LIST(APPEND external_project_dependencies mkldnn)

cmake/util.cmake

@@ -115,8 +115,8 @@ function(link_paddle_exe TARGET_NAME)
  target_link_libraries(${TARGET_NAME} log)
  endif(ANDROID)
 
- if(WITH_MKLDNN AND WITH_MKLML AND MKLDNN_IOMP_DIR)
- target_link_libraries(${TARGET_NAME} "-L${MKLDNN_IOMP_DIR} -liomp5 -Wl,--as-needed")
+ if(WITH_MKLML AND MKLML_LIB_DIR AND MKLML_IOMP_LIB)
+ target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
  endif()
 
  add_dependencies(${TARGET_NAME} ${external_project_dependencies})

doc/design/mkldnn/README.MD

@@ -36,13 +36,13 @@ Figure 1. PaddlePaddle on IA.
 我们把集成方案大致分为了如下几个方面。
 
 ### CMake
-我们会在`CMakeLists.txt`中会添加`WITH_MKLDNN`的选项，当设置这个值为`ON`的时候会启用编译MKL-DNN功能。同时会自动开启OpenMP用于提高MKL-DNN的性能。
+我们会在`CMakeLists.txt`中会给用户添加一个`WITH_MKL`的开关，他是负责`WITH_MKLML`和`WITH_MKLDNN`的总开关。
 
-同时，我们会引入`WITH_MKLML`选项，用于选择是否使用MKL-DNN自带的MKLML安装包。这个安装包可以独立于MKL-DNN使用，但是建议在开启MKL-DNN的同时也打开MKLML的开关，这样才能发挥最好的性能。
+当打开`WITH_MKL`时，会开启MKLML的功能，作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。 如果系统支持AVX2指令集及以上，同时会开启MKL-DNN功能。
 
-所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。
+当关闭`WITH_MKL`时，MKLML和MKL-DNN功能会同时关闭。
 
-**备注**：当`WITH_MKLML=ON`的时候，会优先使用这个包作为PaddlePaddle的CBLAS和LAPACK库，所以会稍微改动`cmake/cblas.cmake`中的逻辑。
+所以，我们会在`cmake/external`目录新建`mkldnn.cmake`和`mklml.cmake`文件，它们会在编译PaddlePaddle的时候下载对应的软件包，并放到PaddlePaddle的third party目录中。
 
 ### Layers
 所有MKL-DNN相关的C++ layers，都会按照PaddlePaddle的目录结构存放在

doc/howto/dev/write_docs_cn.rst

@@ -34,7 +34,7 @@ PaddlePaddle的文档构建有两种方式。
  cd TO_YOUR_PADDLE_CLONE_PATH
  mkdir -p build
  cd build
- cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
+ cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
  make gen_proto_py
  make paddle_docs paddle_docs_cn
 

paddle/gserver/layers/MKLDNNLayer.cpp

@@ -22,7 +22,7 @@ namespace paddle {
 bool MKLDNNLayer::init(const LayerMap& layerMap,
  const ParameterMap& parameterMap) {
  CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
- << "Please set WITH_MKLDNN=ON "
+ << "Please set WITH_MKL=ON "
  << "and set use_mkldnn=True";
  CHECK(!useGpu_) << "Do not support GPU yet";
 

paddle/scripts/docker/README.md

@@ -57,8 +57,7 @@ Users can specify the following Docker build arguments with either "ON" or "OFF"
 | `WITH_GPU` | OFF | Generates NVIDIA CUDA GPU code and relies on CUDA libraries. |
 | `WITH_AVX` | OFF | Set to "ON" to enable AVX support. |
 | `WITH_TESTING` | ON | Build unit tests binaries. |
-| `WITH_MKLDNN` | ON | Build with [Intel® MKL DNN](https://github.com/01org/mkl-dnn) support. |
-| `WITH_MKLML` | ON | Build with [Intel® MKL](https://software.intel.com/en-us/mkl) support. |
+| `WITH_MKL` | ON | Build with [Intel® MKL](https://software.intel.com/en-us/mkl) and [Intel® MKL-DNN](https://github.com/01org/mkl-dnn) support. |
 | `WITH_GOLANG` | ON | Build fault-tolerant parameter server written in go. |
 | `WITH_SWIG_PY` | ON | Build with SWIG python API support. |
 | `WITH_C_API` | OFF | Build capi libraries for inference. |

paddle/scripts/docker/build.sh

@@ -34,8 +34,7 @@ function cmake_gen() {
  ${PYTHON_FLAGS}
  -DWITH_DOC=OFF
  -DWITH_GPU=${WITH_GPU:-OFF}
- -DWITH_MKLDNN=${WITH_MKLDNN:-ON}
- -DWITH_MKLML=${WITH_MKLML:-ON}
+ -DWITH_MKL=${WITH_MKL:-ON}
  -DWITH_AVX=${WITH_AVX:-OFF}
  -DWITH_GOLANG=${WITH_GOLANG:-ON}
  -DWITH_SWIG_PY=ON
@@ -56,8 +55,7 @@ EOF
  ${PYTHON_FLAGS} \
  -DWITH_DOC=OFF \
  -DWITH_GPU=${WITH_GPU:-OFF} \
- -DWITH_MKLDNN=${WITH_MKLDNN:-ON} \
- -DWITH_MKLML=${WITH_MKLML:-ON} \
+ -DWITH_MKL=${WITH_MKL:-ON} \
  -DWITH_AVX=${WITH_AVX:-OFF} \
  -DWITH_GOLANG=${WITH_GOLANG:-ON} \
  -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} \

paddle/scripts/submit_local.sh.in

@@ -18,8 +18,8 @@ function version(){
  echo "PaddlePaddle @PADDLE_VERSION@, compiled with"
  echo " with_avx: @WITH_AVX@"
  echo " with_gpu: @WITH_GPU@"
+ echo " with_mkl: @WITH_MKL@"
  echo " with_mkldnn: @WITH_MKLDNN@"
- echo " with_mklml: @WITH_MKLML@"
  echo " with_double: @WITH_DOUBLE@"
  echo " with_python: @WITH_PYTHON@"
  echo " with_rdma: @WITH_RDMA@"
@@ -45,8 +45,8 @@ function ver2num() {
 
 function cpu_config() {
  # auto set KMP_AFFINITY and OMP_DYNAMIC from Hyper Threading Status
- # only when MKLDNN or MKLML enabled
- if [ "@WITH_MKLDNN@" == "OFF" ] && [ "@WITH_MKLML@" == "OFF"]; then
+ # only when MKL enabled
+ if [ "@WITH_MKL@" == "OFF" ]; then
  return 0
  fi
  ht=`lscpu |grep "per core"|awk -F':' '{print $2}'|xargs`
@@ -70,8 +70,8 @@ function cpu_config() {
 function threads_config() {
  # auto set OMP_NUM_THREADS and MKL_NUM_THREADS
  # according to trainer_count and total processors
- # only when MKLDNN or MKLML enabled
- if [ "@WITH_MKLDNN@" == "OFF" ] && [ "@WITH_MKLML@" == "OFF"]; then
+ # only when MKL enabled
+ if [ "@WITH_MKL@" == "OFF" ]; then
  return 0
  fi
  processors=`grep "processor" /proc/cpuinfo|sort -u|wc -l`