Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 9367fa12 编写于 作者: Y yangyaming
浏览文件
操作

Add more test cases and refine doc. 

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix-4029
上级 4520afcf 6e358758
展开全部 隐藏空白更改
内联 并排
Showing with 9149 addition and 1767 deletion
.gitignore
浏览文件 @ 9367fa12
......@@ -22,7 +22,9 @@ cmake-build-*
# generated while compiling
python/paddle/v2/framework/core.so
paddle/pybind/pybind.h
CMakeFiles
cmake_install.cmake
paddle/.timestamp
python/paddlepaddle.egg-info/
paddle/pybind/pybind.h
.travis.yml
浏览文件 @ 9367fa12
......@@ -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' "$@"; }
script:
......
CMakeLists.txt
浏览文件 @ 9367fa12
......@@ -27,7 +27,7 @@ if(NOT CMAKE_CROSSCOMPILING)
endif(NOT CMAKE_CROSSCOMPILING)
find_package(Git REQUIRED)
find_package(Threads REQUIRED)
if(NOT ANDROID)
if(NOT ANDROID AND NOT IOS)
find_package(Boost QUIET)
endif()
......@@ -64,24 +64,29 @@ if(NOT CMAKE_BUILD_TYPE)
FORCE)
endif()
if(ANDROID)
if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
if(ANDROID OR IOS)
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")
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.")
......
Dockerfile.android
浏览文件 @ 9367fa12
......@@ -6,13 +6,14 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
# ENV variables
ARG ANDROID_ABI
ARG ANDROID_API
ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
ENV ANDROID_API=${ANDROID_API:-21}
ENV HOME=/root \
ANDROID_NDK_HOME=/opt/android-ndk-linux \
ANDROID_ARM_STANDALONE_TOOLCHAIN=/opt/arm-toolchain \
ANDROID_ARM64_STANDALONE_TOOLCHAIN=/opt/arm64-toolchain
ANDROID_TOOLCHAINS_DIR=/opt/toolchains
RUN apt-get update && \
apt-get install -y \
......@@ -42,14 +43,12 @@ RUN pip install --upgrade pip && \
pip install pre-commit
# Android NDK
RUN mkdir /opt/android-ndk-tmp && \
RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
mkdir -p /opt/android-ndk-tmp && \
cd /opt/android-ndk-tmp && \
wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
unzip -q android-ndk-r14b-linux-x86_64.zip && \
mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm --platform=android-23 --install-dir=${ANDROID_ARM_STANDALONE_TOOLCHAIN} && \
${ANDROID_NDK_HOME}/build/tools/make-standalone-toolchain.sh --arch=arm64 --platform=android-23 --install-dir=${ANDROID_ARM64_STANDALONE_TOOLCHAIN} && \
rm -rf /opt/android-ndk-tmp && \
rm -rf ${ANDROID_NDK_HOME}
rm -rf /opt/android-ndk-tmp
CMD ["bash", "/paddle/paddle/scripts/docker/build_android.sh"]
cmake/cblas.cmake
浏览文件 @ 9367fa12
......@@ -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()
cmake/cpplint.cmake
浏览文件 @ 9367fa12
......@@ -26,9 +26,9 @@ set(IGNORE_PATTERN
.*ImportanceSampler.*
.*cblas\\.h.*
.*\\.pb\\.txt
.*LtrDataProvider.*
.*MultiDataProvider.*
.*pb.*)
.*pb.*
.*pybind.h)
# add_style_check_target
#
......
cmake/cross_compiling/ios.cmake 0 → 100644
浏览文件 @ 9367fa12
# 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)
cmake/external/gflags.cmake
浏览文件 @ 9367fa12
......@@ -18,9 +18,9 @@ SET(GFLAGS_SOURCES_DIR ${THIRD_PARTY_PATH}/gflags)
SET(GFLAGS_INSTALL_DIR ${THIRD_PARTY_PATH}/install/gflags)
SET(GFLAGS_INCLUDE_DIR "${GFLAGS_INSTALL_DIR}/include" CACHE PATH "gflags include directory." FORCE)
IF(WIN32)
set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/gflags.lib" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
ELSE(WIN32)
set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
set(GFLAGS_LIBRARIES "${GFLAGS_INSTALL_DIR}/lib/libgflags.a" CACHE FILEPATH "GFLAGS_LIBRARIES" FORCE)
ENDIF(WIN32)
INCLUDE_DIRECTORIES(${GFLAGS_INCLUDE_DIR})
......@@ -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
......@@ -56,3 +57,12 @@ SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
ADD_DEPENDENCIES(gflags extern_gflags)
LIST(APPEND external_project_dependencies gflags)
IF(WITH_C_API)
INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
IF(ANDROID)
INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
ENDIF()
ENDIF()
cmake/external/glog.cmake
浏览文件 @ 9367fa12
......@@ -19,9 +19,9 @@ SET(GLOG_INSTALL_DIR ${THIRD_PARTY_PATH}/install/glog)
SET(GLOG_INCLUDE_DIR "${GLOG_INSTALL_DIR}/include" CACHE PATH "glog include directory." FORCE)
IF(WIN32)
SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE)
SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.lib" CACHE FILEPATH "glog library." FORCE)
ELSE(WIN32)
SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE)
SET(GLOG_LIBRARIES "${GLOG_INSTALL_DIR}/lib/libglog.a" CACHE FILEPATH "glog library." FORCE)
ENDIF(WIN32)
INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
......@@ -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
......@@ -56,3 +57,12 @@ ADD_DEPENDENCIES(glog extern_glog gflags)
LINK_LIBRARIES(glog gflags)
LIST(APPEND external_project_dependencies glog)
IF(WITH_C_API)
INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
IF(ANDROID)
INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib)
ENDIF()
ENDIF()
cmake/external/gtest.cmake
浏览文件 @ 9367fa12
......@@ -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
......
cmake/external/openblas.cmake
浏览文件 @ 9367fa12
......@@ -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}
......@@ -73,6 +86,26 @@ IF(NOT ${CBLAS_FOUND})
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
)
IF(WITH_C_API)
INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
# Because libopenblas.a is a symbolic link of another library, thus need to
# install the whole directory.
IF(ANDROID)
SET(TMP_INSTALL_DIR third_party/openblas/lib/${ANDROID_ABI})
ELSE()
SET(TMP_INSTALL_DIR third_party/openblas/lib)
ENDIF()
INSTALL(CODE "execute_process(
COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
destination ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
)"
)
INSTALL(CODE "MESSAGE(STATUS \"Installing: \"
\"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
)"
)
ENDIF()
ENDIF(NOT ${CBLAS_FOUND})
MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
......
cmake/external/protobuf.cmake
浏览文件 @ 9367fa12
......@@ -173,7 +173,8 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
"-DCMAKE_CXX_FLAGS=${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()
......@@ -223,6 +224,15 @@ IF(NOT PROTOBUF_FOUND)
SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
CACHE FILEPATH "protoc library." FORCE)
IF(WITH_C_API)
INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
IF(ANDROID)
INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${PROTOBUF_LIBRARY} DESTINATION third_party/protobuf/lib)
ENDIF()
ENDIF()
IF(CMAKE_CROSSCOMPILING)
PROMPT_PROTOBUF_LIB(protobuf_host extern_protobuf)
ELSE()
......
cmake/external/python.cmake
浏览文件 @ 9367fa12
......@@ -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})
cmake/external/swig.cmake
浏览文件 @ 9367fa12
......@@ -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)
......
cmake/external/warpctc.cmake
浏览文件 @ 9367fa12
......@@ -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)
......
cmake/external/zlib.cmake
浏览文件 @ 9367fa12
......@@ -34,18 +34,28 @@ 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
)
LIST(APPEND external_project_dependencies zlib)
IF(WITH_C_API)
INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)
IF(ANDROID)
INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib)
ENDIF()
ENDIF()
cmake/flags.cmake
浏览文件 @ 9367fa12
......@@ -128,8 +128,10 @@ set(GPU_COMMON_FLAGS
)
if (APPLE)
# On Mac OS X build fat binaries with x86_64 architectures by default.
set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
if(NOT CMAKE_CROSSCOMPILING)
# On Mac OS X build fat binaries with x86_64 architectures by default.
set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
endif()
else()
set(GPU_COMMON_FLAGS
-Wall
......
cmake/system.cmake
浏览文件 @ 9367fa12
......@@ -24,11 +24,10 @@ IF(WIN32)
SET(HOST_SYSTEM "win32")
ELSE(WIN32)
IF(APPLE)
EXEC_PROGRAM (sw_vers ARGS -productVersion OUTPUT_VARIABLE MACOSX_VERSION)
STRING(REGEX MATCH "[0-9]+.[0-9]+" VERSION "${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(${CMAKE_SYSTEM_NAME} STREQUAL "iOS")
SET(IOS TRUE)
INCLUDE(cross_compiling/ios)
ENDIF()
ENDIF()
......
cmake/util.cmake
浏览文件 @ 9367fa12
......@@ -25,7 +25,9 @@ function(target_circle_link_libraries TARGET_NAME)
endif()
endforeach()
if("${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}
......
doc/design/api.md
浏览文件 @ 9367fa12
......@@ -3,7 +3,7 @@
## Ingredients
As our design principle is starting from the essence: how could we
allow users to express and solve their problems at neural networks.
allow users to express and solve their problems as neural networks.
Some essential concepts that our API have to provide include:
1. A *topology* is an expression of *layers*.
......@@ -233,7 +233,7 @@ paddle.dist_train(model,
num_parameter_servers=15)
```
The pseudo code if `paddle.dist_train` is as follows:
The pseudo code of `paddle.dist_train` is as follows:
```python
def dist_train(topology, parameters, trainer, reader, ...):
......
doc/design/auto_gradient_check.md
浏览文件 @ 9367fa12
## Auto Gradient Checker Design
## Backgraound：
- Operator forward computing is easy to check if the result is right because it has a clear definition. **But** backpropagation is a notoriously difficult algorithm to debug and get right:
- 1. you should get the right backpropagation formula according to the forward computation.
- 2. you should implement it right in CPP.
- 3. it's difficult to prepare test data.
- Generally, it is easy to check whether the forward computation of an Operator is correct or not. However, backpropagation is a notoriously difficult algorithm to debug and get right:
1. you should get the right backpropagation formula according to the forward computation.
2. you should implement it right in CPP.
3. it's difficult to prepare test data.
- Auto gradient check gets a numeric gradient by forward Operator and use it as a reference of the backward Operator's result. It has several advantages:
- 1. numeric gradient checker only need forward operator.
- 2. user only need to prepare the input data for forward Operator.
- Auto gradient checking gets a numerical gradient by forward Operator and use it as a reference of the backward Operator's result. It has several advantages:
1. numerical gradient checker only need forward operator.
2. user only need to prepare the input data for forward Operator.
## Mathematical Theory
The following two document from stanford has a detailed explanation of how to get numeric gradient and why it's useful.
The following two document from Stanford has a detailed explanation of how to get numerical gradient and why it's useful.
- [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization)
- [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96)
......@@ -20,7 +20,7 @@ The following two document from stanford has a detailed explanation of how to ge
## Numeric Gradient Implementation
### Python Interface
```python
def get_numeric_gradient(op,
def get_numerical_gradient(op,
input_values,
output_name,
input_to_check,
......@@ -30,13 +30,13 @@ def get_numeric_gradient(op,
Get Numeric Gradient for an operator's input.
:param op: C++ operator instance, could be an network
:param input_values: The input variables. Should be an dictionary, key is
variable name. Value is numpy array.
:param input_values: The input variables. Should be an dictionary, whose key is
variable name, and value is numpy array.
:param output_name: The final output variable name.
:param input_to_check: The input variable need to get gradient.
:param input_to_check: The input variable with respect to which to compute the gradient.
:param delta: The perturbation value for numeric gradient method. The
smaller delta is, the more accurate result will get. But if that delta is
too small, it could occur numerical stability problem.
too small, it will suffer from numerical stability problem.
:param local_scope: The local scope used for get_numeric_gradient.
:return: The gradient array in numpy format.
"""
......@@ -45,28 +45,28 @@ def get_numeric_gradient(op,
### Explaination:
- Why need `output_name`
- One Operator may have multiple Output, you can get independent gradient from each Output. So user should set one output to calculate.
- An Operator may have multiple Output, one can get independent gradient from each Output. So caller should specify the name of the output variable.
- Why need `input_to_check`
- One operator may have multiple inputs. Gradient Op can calculate the gradient of these Inputs at the same time. But Numeric Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times.
- One operator may have multiple inputs. Gradient Op can calculate the gradient of these inputs at the same time. But Numeric Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times.
### Core Algorithm Implementation
```python
# we only compute gradient of one element each time.
# we use a for loop to compute the gradient of every element.
# we only compute gradient of one element a time.
# we use a for loop to compute the gradient of each element.
for i in xrange(tensor_size):
# get one input element throw it's index i.
# get one input element by its index i.
origin = tensor_to_check.get_float_element(i)
# add delta to it, run op and then get the sum of the result tensor.
# add delta to it, run op and then get the new value of the result tensor.
x_pos = origin + delta
tensor_to_check.set_float_element(i, x_pos)
y_pos = get_output()
# plus delta to this element, run op and get the sum of the result tensor.
# plus delta to this element, run op and get the new value of the result tensor.
x_neg = origin - delta
tensor_to_check.set_float_element(i, x_neg)
y_neg = get_output()
......@@ -85,15 +85,15 @@ def get_numeric_gradient(op,
Each Operator Kernel has three kinds of Gradient:
- 1. Numeric Gradient
- 2. CPU Operator Gradient
- 3. GPU Operator Gradient(if supported)
1. Numerical gradient
2. CPU kernel gradient
3. GPU kernel gradient (if supported)
Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as the reference value.
The numerical gradient only relies on forward Operator. So we use the numerical gradient as the reference value. And the gradient checking is performed in the following three steps:
- 1. calculate the numeric gradient.
- 2. calculate CPU kernel Gradient with the backward Operator and compare it with the numeric gradient.
- 3. calculate GPU kernel Gradient with the backward Operator and compare it with the numeric gradient.(if support GPU)
1. calculate the numerical gradient
2. calculate CPU kernel gradient with the backward Operator and compare it with the numerical gradient
3. calculate GPU kernel gradient with the backward Operator and compare it with the numeric gradient (if supported)
#### Python Interface
......@@ -110,8 +110,8 @@ Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as
:param forward_op: used to create backward_op
:param input_vars: numpy value of input variable. The following
computation will use these variables.
:param inputs_to_check: inputs var names that should check gradient.
:param output_name: output name that used to
:param inputs_to_check: the input variable with respect to which to compute the gradient.
:param output_name: The final output variable name.
:param max_relative_error: The relative tolerance parameter.
:param no_grad_set: used when create backward ops
:param only_cpu: only compute and check gradient on cpu kernel.
......@@ -120,24 +120,24 @@ Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as
```
### How to check if two numpy array is close enough?
if `abs_numeric_grad` is nearly zero, then use abs error for numeric_grad, not relative
if `abs_numerical_grad` is nearly zero, then use abs error for numerical_grad
```python
numeric_grad = ...
numerical_grad = ...
operator_grad = numpy.array(scope.find_var(grad_var_name(name)).get_tensor())
abs_numeric_grad = numpy.abs(numeric_grad)
# if abs_numeric_grad is nearly zero, then use abs error for numeric_grad, not relative
abs_numerical_grad = numpy.abs(numerical_grad)
# if abs_numerical_grad is nearly zero, then use abs error for numeric_grad, not relative
# error.
abs_numeric_grad[abs_numeric_grad < 1e-3] = 1
abs_numerical_grad[abs_numerical_grad < 1e-3] = 1
diff_mat = numpy.abs(abs_numeric_grad - operator_grad) / abs_numeric_grad
diff_mat = numpy.abs(abs_numerical_grad - operator_grad) / abs_numerical_grad
max_diff = numpy.max(diff_mat)
```
#### Notes：
1，The Input data for auto gradient checker should be reasonable to avoid numeric problem.
The Input data for auto gradient checker should be reasonable to avoid numerical stability problem.
#### Refs:
......
doc/design/block.md 0 → 100644
浏览文件 @ 9367fa12
# 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);
}
```
doc/design/functions_operators_layers.md
浏览文件 @ 9367fa12
......@@ -53,12 +53,12 @@ Let's explain using an example. Suppose that we are going to compose the FC usi
```python
def operator.mul(X1, X2):
O = Var()
paddle.cpp.create_operator("mul", input={X1, Y1], output=O)
paddle.cpp.create_operator("mul", input={X1, Y1}, output=O)
return O
def operator.add(X1, X2):
O = Var()
paddle.cpp.create_operator("add", input={X1, X2], output=O)
paddle.cpp.create_operator("add", input={X1, X2}, output=O)
return O
```
......
doc/design/graph.md
浏览文件 @ 9367fa12
......@@ -56,7 +56,7 @@ For each parameter, like W and b created by `layer.fc`, marked as double circles
## Block and Graph
The word block and graph are interchangable in the desgin of PaddlePaddle. A [Block[(https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions. A graph of operators and variables is a representation of the block.
The word block and graph are interchangable in the desgin of PaddlePaddle. A [Block](https://github.com/PaddlePaddle/Paddle/pull/3708) is a metaphore of the code and local variables in a pair of curly braces in programming languages, where operators are like statements or instructions. A graph of operators and variables is a representation of the block.
A Block keeps operators in an array `BlockDesc::ops`
......@@ -67,4 +67,4 @@ message BlockDesc {
}
```
in the order that there appear in user programs, like the Python program at the beginning of this article. We can imagine that in `ops`, we have some forward operators, followed by some gradient operators, and then some optimization operators.
in the order that they appear in user programs, like the Python program at the beginning of this article. We can imagine that in `ops`, we have some forward operators, followed by some gradient operators, and then some optimization operators.
doc/design/if_else_op.md
浏览文件 @ 9367fa12
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
......
doc/design/ops/images/2_level_rnn.dot 0 → 100644
浏览文件 @ 9367fa12
digraph G {
rnn [label="1-th level RNN" shape=box]
subgraph cluster0 {
label = "time step 0"
sent0 [label="sentence"]
sent1 [label="sentence"]
rnn1 [label="2-th level RNN" shape=box]
sent0 -> rnn1
sent1 -> rnn1
}
subgraph cluster1 {
label = "time step 1"
sent2 [label="sentence"]
sent3 [label="sentence"]
rnn2 [label="2-th level RNN" shape=box]
sent2 -> rnn2
sent3 -> rnn2
}
subgraph cluster2 {
label = "time step 2"
sent4 [label="sentence"]
sent5 [label="sentence"]
rnn3 [label="2-th level RNN" shape=box]
sent4 -> rnn3
sent5 -> rnn3
}
para0 [label="paragraph info 0"]
para1 [label="paragraph info 1"]
para2 [label="paragraph info 2"]
rnn1 -> para0
rnn2 -> para1
rnn3 -> para2
para0 -> rnn
para1 -> rnn
para2 -> rnn
chapter [label="chapter info"]
rnn -> chapter
}
doc/design/ops/images/rnn.dot 0 → 100644
浏览文件 @ 9367fa12
digraph G {
label = "simple RNN implementation"
ranksep=2;
//graph [nodesep=1, ranksep=1];
node[nodesep=1]
subgraph cluster0 {
label = "global scope"
rankdir = TB
W
boot_memory
input
output
}
subgraph cluster1 {
label = "step-scope 0"
rankdir = TB
memory0[label="memory"]
prememory0[label="pre-memory"]
step_input0[label="step input"]
step_output0[label="step output"]
}
subgraph cluster2 {
label = "step-scope 1"
rankdir = TB
memory1[label="memory"]
prememory1[label="pre-memory"]
step_input1[label="step input"]
step_output1[label="step output"]
}
subgraph cluster3 {
label = "step-scope 2"
rankdir = TB
memory2[label="memory"]
prememory2[label="pre-memory"]
step_input2[label="step input"]
step_output2[label="step output"]
}
stepnet [shape=box]
stepnet0 [shape=box, style=dashed]
stepnet1 [shape=box, style=dashed]
stepnet2 [shape=box, style=dashed]
edge[color=blue]
boot_memory -> prememory0 [label="init" color="blue"]
memory0 -> prememory1 [label="copy/reference" color="blue"]
memory1 -> prememory2 [label="copy/reference" color="blue"]
edge[color=black]
W -> stepnet0[constraint=false, style=dashed]
W -> stepnet1[constraint=false, style=dashed]
W -> stepnet2[constraint=false, style=dashed]
memory0 -> stepnet0[style=dashed]
prememory0 -> stepnet0 -> step_output0[style=dashed]
memory1 -> stepnet1[style=dashed]
prememory1 -> stepnet1 -> step_output1[style=dashed]
memory2 -> stepnet2[style=dashed]
prememory2 -> stepnet2 -> step_output2[style=dashed]
input -> step_input0
input -> step_input1
input -> step_input2
step_input0 -> stepnet0 [style=dashed]
step_input1 -> stepnet1[style=dashed]
step_input2 -> stepnet2[style=dashed]
step_output0 -> output
step_output1 -> output
step_output2 -> output
stepnet0 -> stepnet[style=dashed]
stepnet1 -> stepnet[style=dashed]
stepnet2 -> stepnet[style=dashed]
}
doc/design/ops/images/rnn_2level_data.dot 0 → 100644
浏览文件 @ 9367fa12
digraph G {
chapter [label="chapter"]
subgraph cluster0 {
label = "paragraph 0"
top_rnn0[label="top rnn step 0" shape=box]
p0 [label="paragraph 0"]
p1 [label="paragraph 1"]
}
subgraph cluster1{
label = "paragraph 1"
top_rnn1[label="top rnn step 1" shape=box]
p2 [label="paragraph 0"]
p3 [label="paragraph 1"]
}
subgraph cluster_p0 {
label = "sentence 0"
low_rnn0 [label="low rnn step 0" shape=box]
s00 [label="sentence 0"]
s01 [label="sentence 1"]
low_rnn0 -> s00
low_rnn0 -> s01
}
subgraph cluster_p1 {
label = "sentence 1"
low_rnn1 [label="low rnn step 1" shape=box]
s10 [label="sentence 0"]
s11 [label="sentence 1"]
low_rnn1 -> s10
low_rnn1 -> s11
}
subgraph cluster_p2 {
label = "sentence 1"
low_rnn2 [label="low rnn step 0" shape=box]
s20 [label="sentence 0"]
s21 [label="sentence 1"]
low_rnn2 -> s20
low_rnn2 -> s21
}
subgraph cluster_p3 {
label = "sentence 1"
low_rnn3 [label="low rnn step 1" shape=box]
s30 [label="sentence 0"]
s31 [label="sentence 1"]
low_rnn3 -> s30
low_rnn3 -> s31
}
chapter -> top_rnn0
chapter -> top_rnn1
top_rnn0 -> p0
top_rnn0 -> p1
top_rnn1 -> p2
top_rnn1 -> p3
p0 -> low_rnn0
p1 -> low_rnn1
p2 -> low_rnn2
p3 -> low_rnn3
}
doc/design/ops/rnn.md 0 → 100644
浏览文件 @ 9367fa12
# RNNOp design
This document is about an RNN operator which requires that instances in a mini-batch have the same length. We will have a more flexible RNN operator.
## RNN Algorithm Implementation
<p aligh="center">
<img src="./images/rnn.jpg"/>
</p>
The above diagram shows an RNN unrolled into a full network.
There are several important concepts:
- *step-net*: the sub-graph to run at each step,
- *memory*, $h_t$, the state of the current step,
- *ex-memory*, $h_{t-1}$, the state of the previous step,
- *initial memory value*, the ex-memory of the first step.
### Step-scope
There could be local variables defined in step-nets. PaddlePaddle runtime realizes these variables in *step-scopes* -- scopes created for each step.
<p aligh="center">
<img src="./images/rnn.png"/><br/>
Figure 2 the RNN's data flow
</p>
Please be aware that all steps run the same step-net. Each step
1. creates the step-scope,
2. realizes local variables, including step-outputs, in the step-scope, and
3. runs the step-net, which could use these variables.
The RNN operator will compose its output from step outputs in step scopes.
### Memory and Ex-memory
Let's give more details about memory and ex-memory via a simply example:
$$
h_t = U h_{t-1} + W x_t
$$,
where $h_t$ and $h_{t-1}$ are the memory and ex-memory of step $t$'s respectively.
In the implementation, we can make an ex-memory variable either "refers to" the memory variable of the previous step,
or copy the value of the previous memory value to the current ex-memory variable.
### Usage in Python
For more information on Block, please refer to the [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/block.md).
We can define an RNN's step-net using Block:
```python
import paddle as pd
X = some_op() # x is some operator's output, and is a LoDTensor
a = some_op()
# declare parameters
W = pd.Variable(shape=[20, 30])
U = pd.Variable(shape=[20, 30])
rnn = pd.create_rnn_op(output_num=1)
with rnn.stepnet():
x = rnn.add_input(X)
# declare a memory (rnn's step)
h = rnn.add_memory(init=a)
# h.pre_state() means previous memory of rnn
new_state = pd.add_two( pd.matmul(W, x) + pd.matmul(U, h.pre_state()))
# update current memory
h.update(new_state)
# indicate that h variables in all step scopes should be merged
rnn.add_outputs(h)
out = rnn()
```
Python API functions in above example:
- `rnn.add_input` indicates the parameter is a variable that will be segmented into step-inputs.
- `rnn.add_memory` creates a variable used as the memory.
- `rnn.add_outputs` mark the variables that will be concatenated across steps into the RNN output.
### Nested RNN and LoDTensor
An RNN whose step-net includes other RNN operators is known as an *nested RNN*.
For example, we could have a 2-level RNN, where the top level corresponds to paragraphs, and the lower level corresponds to sentences.
The following figure illustrates the feeding of text into the lower level, one sentence each step, and the feeding of step outputs to the top level. The final top level output is about the whole text.
<p aligh="center">
<img src="./images/2_level_rnn.png"/>
</p>
```python
import paddle as pd
W = pd.Variable(shape=[20, 30])
U = pd.Variable(shape=[20, 30])
W0 = pd.Variable(shape=[20, 30])
U0 = pd.Variable(shape=[20, 30])
# a is output of some op
a = some_op()
# chapter_data is a set of 128-dim word vectors
# the first level of LoD is sentence
# the second level of LoD is chapter
chapter_data = pd.Variable(shape=[None, 128], type=pd.lod_tensor, level=2)
def lower_level_rnn(paragraph):
'''
x: the input
'''
rnn = pd.create_rnn_op(output_num=1)
with rnn.stepnet():
sentence = rnn.add_input(paragraph, level=0)
h = rnn.add_memory(shape=[20, 30])
h.update(
pd.matmul(W, sentence) + pd.matmul(U, h.pre_state()))
# get the last state as sentence's info
rnn.add_outputs(h)
return rnn
top_level_rnn = pd.create_rnn_op(output_num=1)
with top_level_rnn.stepnet():
paragraph_data = rnn.add_input(chapter_data, level=1)
low_rnn = lower_level_rnn(paragraph_data)
paragraph_out = low_rnn()
h = rnn.add_memory(init=a)
h.update(
pd.matmul(W0, paragraph_data) + pd.matmul(U0, h.pre_state()))
top_level_rnn.add_outputs(h)
# just output the last step
chapter_out = top_level_rnn(output_all_steps=False)
```
in above example, the construction of the `top_level_rnn` calls `lower_level_rnn`. The input is a LoD Tensor. The top level RNN segments input text data into paragraphs, and the lower level RNN segments each paragraph into sentences.
By default, the `RNNOp` will concatenate the outputs from all the time steps,
if the `output_all_steps` set to False, it will only output the final time step.
<p align="center">
<img src="images/rnn_2level_data.png"/>
</p>
doc/design/parameters_in_cpp.md
浏览文件 @ 9367fa12
# Design Doc: The C++ Class `Parameters`
`Parameters` is a concept we designed in Paddle V2 API. `Parameters` is a container of parameters, and make Paddle can shared parameter between topologies. We described usages of `Parameter` in [api.md](./api.md).
`Parameters` is a concept we designed in PaddlePaddle V2 API. `Parameters` is a container of parameters, which makes PaddlePaddle capable of sharing parameter between topologies. We described usages of `Parameter` in [api.md](./api.md).
We used Python to implement Parameters when designing V2 API before. There are several defects for current implementation:
We used Python to implement Parameters when designing V2 API before. There are several defects for the current implementation:
* We just use `memcpy` to share Parameters between topologies, but this is very inefficient.
* We did not implement share Parameters while training. We just trigger `memcpy` when start training.
* We did not support sharing Parameters while training. We just trigger `memcpy` when start training.
It is necessary that we implement Parameters in CPP side. However, it could be a code refactoring for Paddle, because Paddle was designed for training only one topology before, i.e., each GradientMachine contains its Parameter as a data member. In current Paddle implementation, there are three concepts associated with `Parameters`:
It is necessary that we implement Parameters in CPP side. However, it could result a code refactoring for PaddlePaddle, because PaddlePaddle was designed for training only one topology before, i.e., each GradientMachine contains its Parameter as a data member. In current PaddlePaddle implementation, there are three concepts associated with `Parameters`:
1. `paddle::Parameter`. A `Parameters` is a container for `paddle::Parameter`.
It is evident that we should use `paddle::Parameter` when developing `Parameters`.
However, the `Parameter` class contains many functions and does not have a clear interface.
It contains `create/store Parameter`, `serialize/deserialize`, `optimize(i.e SGD)`, `randomize/zero`.
When we developing `Parameters`, we only use `create/store Parameter` functionality.
We should extract functionalities of Parameter into many classes to clean Paddle CPP implementation.
We should extract functionalities of Parameter into many classes to clean PaddlePaddle CPP implementation.
2. `paddle::GradientMachine` and its sub-classes, e.g., `paddle::MultiGradientMachine`, `paddle::NeuralNetwork`.
We should pass `Parameters` to `paddle::GradientMachine` when `forward/backward` to avoid `memcpy` between topologies.
......@@ -24,7 +24,7 @@ Also, we should handle multi-GPU/CPU training, because `forward` and `backward`
So `Parameters` should be used by `paddle::ParameterUpdater`, and `paddle::ParameterUpdater` should optimize `Parameters` (by SGD).
The step by step approach for implementation Parameters in Paddle C++ core is listed below. Each step should be a PR and could be merged into Paddle one by one.
The step by step approach for implementation Parameters in PaddlePaddle C++ core is listed below. Each step should be a PR and could be merged into PaddlePaddle one by one.
1. Clean `paddle::Parameter` interface. Extract the functionalities of `paddle::Parameter` to prepare for the implementation of Parameters.
......
doc/design/reader/README.md
浏览文件 @ 9367fa12
......@@ -52,7 +52,7 @@ Here are valid outputs:
# a mini batch of three data items, each data item is a list (single column).
[([1,1,1],),
([2,2,2],),
([3,3,3],),
([3,3,3],)]
```
Please note that each item inside the list must be a tuple, below is an invalid output:
......
doc/design/refactorization.md 0 → 100644
浏览文件 @ 9367fa12
# 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 *variables* 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`. GPU 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
doc/design/releasing_process.md
浏览文件 @ 9367fa12
# Paddle发行规范
# PaddlePaddle发行规范
Paddle使用git-flow branching model做分支管理，使用[Semantic Versioning](http://semver.org/)标准表示Paddle版本号。
PaddlePaddle使用git-flow branching model做分支管理，使用[Semantic Versioning](http://semver.org/)标准表示PaddlePaddle版本号。
Paddle每次发新的版本，遵循以下流程:
PaddlePaddle每次发新的版本，遵循以下流程:
1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0`
2. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
......@@ -27,14 +27,14 @@ Paddle每次发新的版本，遵循以下流程:
需要注意的是:
* `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试Paddle的行为。
* `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。
* 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。
# Paddle 分支规范
# PaddlePaddle 分支规范
Paddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。
PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。
* Paddle的主版本库遵循[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范。其中:
* PaddlePaddle的主版本库遵循[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范。其中:
* `master`分支为稳定(stable branch)版本分支。每一个`master`分支的版本都是经过单元测试和回归测试的版本。
* `develop`分支为开发(develop branch)版本分支。每一个`develop`分支的版本都经过单元测试，但并没有经过回归测试。
* `release/版本号`分支为每一次Release时建立的临时分支。在这个阶段的代码正在经历回归测试。
......@@ -42,18 +42,18 @@ Paddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branch
* 其他用户的fork版本库并不需要严格遵守[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，但所有fork的版本库的所有分支都相当于特性分支。
* 建议，开发者fork的版本库使用`develop`分支同步主版本库的`develop`分支
* 建议，开发者fork的版本库中，再基于`develop`版本fork出自己的功能分支。
* 当功能分支开发完毕后，向Paddle的主版本库提交`Pull Reuqest`，进而进行代码评审。
* 当功能分支开发完毕后，向PaddlePaddle的主版本库提交`Pull Reuqest`，进而进行代码评审。
* 在评审过程中，开发者修改自己的代码，可以继续在自己的功能分支提交代码。
* BugFix分支也是在开发者自己的fork版本库维护，与功能分支不同的是，BugFix分支需要分别给主版本库的`master`、`develop`与可能有的`release/版本号`分支，同时提起`Pull Request`。
# Paddle回归测试列表
# PaddlePaddle回归测试列表
本列表说明Paddle发版之前需要测试的功能点。
本列表说明PaddlePaddle发版之前需要测试的功能点。
## Paddle Book中所有章节
## PaddlePaddle Book中所有章节
Paddle每次发版本首先要保证Paddle Book中所有章节功能的正确性。功能的正确性包括验证Paddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。
PaddlePaddle每次发版本首先要保证PaddlePaddle Book中所有章节功能的正确性。功能的正确性包括验证PaddlePaddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。
| | 新手入门章节 | 识别数字 | 图像分类 | 词向量 | 情感分析 | 语意角色标注 | 机器翻译 | 个性化推荐 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
......
doc/design/scope.md
浏览文件 @ 9367fa12
......@@ -17,7 +17,7 @@ Scope is an association of a name to variable. All variables belong to `Scope`.
1. Scope only contains a map of a name to variable.
All parameters, data, states in a Net should be variables and stored inside a scope. Each op should get inputs and outputs to do computation from a scope, such as data buffer, state(momentum) etc.
All parameters, data, states in a Net should be variables and stored inside a scope. Each op should get inputs and outputs to do computation from a scope, such as data buffer, state (momentum) etc.
1. Variable can only be created by Scope and a variable can only be got from Scope. User cannot create or get a variable outside a scope. This is a constraints of our framework, and will keep our framework simple and clear.
......@@ -32,7 +32,7 @@ Scope is an association of a name to variable. All variables belong to `Scope`.
1. Scope should destruct all Variables inside it when itself is destructed. User can never store `Variable` pointer somewhere else.
Because Variable can only be got from Scope. When destroying Scope, we also need to destroy all the Variables in it. If user store `Variable` pointer to private data member or some global variable, the pointer will be a invalid pointer when associated `Scope` is destroyed.
Because Variable can only be got from Scope. When destroying Scope, we also need to destroy all the Variables in it. If user store `Variable` pointer to private data member or some global variable, the pointer will be an invalid pointer when associated `Scope` is destroyed.
```cpp
class Scope {
......@@ -50,7 +50,7 @@ class Scope {
Just like [scope](https://en.wikipedia.org/wiki/Scope_(computer_science)) in programming languages, `Scope` in the neural network can also be a local scope. There are two attributes about local scope.
1. We can create local variables in a local scope. When that local scope are destroyed, all local variables should also be destroyed.
1. We can create local variables in a local scope. When that local scope is destroyed, all local variables should also be destroyed.
2. Variables in a parent scope can be retrieved from local scopes of that parent scope, i.e., when user get a variable from a scope, it will try to search this variable in current scope. If there is no such variable in the local scope, `scope` will keep searching from its parent, until the variable is found or there is no parent.
```cpp
......@@ -121,4 +121,4 @@ Also, as the parent scope is a `shared_ptr`, we can only `Create()` a scope shar
## Orthogonal interface
`FindVar` will return `nullptr` when `name` is not found. It can be used as `Contains` method. `NewVar` will return a `Error` when there is a name conflict locally. Combine `FindVar` and `NewVar`, we can implement `NewVar` easily.
`FindVar` will return `nullptr` when `name` is not found. It can be used as `Contains` method. `NewVar` will return an `Error` when there is a name conflict locally. Combine `FindVar` and `NewVar`, we can implement `NewVar` easily.
doc/design/simple_op_design.md
浏览文件 @ 9367fa12
......@@ -6,9 +6,9 @@ The Interaction between Python and C++ can be simplified as two steps:
1. C++ tells Python how many Ops there are, and what parameter do users need to offer to initialize a new Op. Python then builds API for each Op at compile time.
2. Users invoke APIs built by Python and provide necessary parameters. These parameters will be sent to C++ fo finish Op construction task.
2. Users invoke APIs built by Python and provide necessary parameters. These parameters will be sent to C++ for finishing the Op construction task.
### Message form C++ to Python
### Message from C++ to Python
We define a Protobuf message class `OpProto` to hold message needed in the first step. What should an `OpProto` contain? This question is equivalent to “What message do we need to offer, to build a Python API which is legal and user oriented and can use to describe a whole Op.”
......@@ -193,7 +193,7 @@ def fc_layer(input, size, with_bias, activation):
elif:
# ...
return act_output;
```
```
### Low Leval API
......
doc/design/var_desc.md
浏览文件 @ 9367fa12
## Background
PaddlePaddle divides the description of neural network computation graph into two stages: compile time and runtime.
PaddlePaddle use proto message to describe compile time graph for
PaddlePaddle use proto message to describe compile time graph because
1. Computation graph should be able to be saved to a file.
1. In distributed training, the graph will be serialized and send to multiple workers.
......
doc/faq/index_cn.rst
浏览文件 @ 9367fa12
......@@ -321,3 +321,55 @@ pip uninstall py_paddle paddle
然后安装paddle的python环境, 在build目录下执行
pip install python/dist/paddle*.whl && pip install ../paddle/dist/py_paddle*.whl
16. PaddlePaddle存储的参数格式是什么，如何和明文进行相互转化
---------------------------------------------------------
PaddlePaddle保存的模型参数文件内容由16字节头信息和网络参数两部分组成。头信息中，1~4字节表示PaddlePaddle版本信息，请直接填充0；5~8字节表示每个参数占用的字节数，当保存的网络参数为float类型时为4，double类型时为8；9~16字节表示保存的参数总个数。
将PaddlePaddle保存的模型参数还原回明文时，可以使用相应数据类型的 :code:`numpy.array` 加载具体网络参数，此时可以跳过PaddlePaddle模型参数文件的头信息。若在PaddlePaddle编译时，未指定按照double精度编译，默认情况下按照float精度计算，保存的参数也是float类型。这时在使用 :code:`numpy.array` 时，一般设置 :code:`dtype=float32` 。示例如下：
.. code-block:: python
def read_parameter(fname, width):
s = open(fname).read()
# skip header
vec = np.fromstring(s[16:], dtype=np.float32)
# width is the size of the corresponding layer
np.savetxt(fname + ".csv", vec.reshape(width, -1),
fmt="%.6f", delimiter=",")
将明文参数转化为PaddlePaddle可加载的模型参数时，首先构造头信息，再写入网络参数。下面的代码将随机生成的矩阵转化为可以被PaddlePaddle加载的模型参数。
.. code-block:: python
def gen_rand_param(param_file, width, height, need_trans):
np.random.seed()
header = struct.pack("iil", 0, 4, height * width)
param = np.float32(np.random.rand(height, width))
with open(param_file, "w") as fparam:
fparam.write(header + param.tostring())
17. 如何加载预训练参数
------------------------------
* 对加载预训练参数的层，设置其参数属性 :code:`is_static=True`，使该层的参数在训练过程中保持不变。以embedding层为例，代码如下：
.. code-block:: python
emb_para = paddle.attr.Param(name='emb', is_static=True)
paddle.layer.embedding(size=word_dim, input=x, param_attr=emb_para)
* 从模型文件将预训练参数载入 :code:`numpy.array`，在创建parameters后，使用 :code:`parameters.set()` 加载预训练参数。PaddlePaddle保存的模型参数文件前16字节为头信息，用户将参数载入 :code:`numpy.array` 时须从第17字节开始。以embedding层为例，代码如下：
.. code-block:: python
def load_parameter(file_name, h, w):
with open(file_name, 'rb') as f:
f.read(16) # skip header.
return np.fromfile(f, dtype=np.float32).reshape(h, w)
parameters = paddle.parameters.create(my_cost)
parameters.set('emb', load_parameter(emb_param_file, 30000, 256))
doc/howto/dev/new_op_cn.md
浏览文件 @ 9367fa12
......@@ -34,7 +34,7 @@ Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU
注册Op | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中
实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。
实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
下面以矩阵乘操作，即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。
......@@ -224,45 +224,15 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
### 5. 编译
- 简单**无特殊依赖**的OP无需修改CMakeList.txt文件。[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt) 会自动将 `paddle/operators` 目录下新增的 `*_op.cc` 文件加入编译。
- 较为复杂、**有额外依赖** 的operator仍需要修改[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)。如，`mul_op` 依赖 `math_function`，需要在`CMakeLists.txt`中添加如下内容：
运行下面命令可以进行编译：
```
op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function) +
```
- 运行下面命令可以进行编译：
```
make mul_op
```
```
make mul_op
```
## 绑定Python
- 绑定Python
在 [`paddle/pybind/pybind.cc
`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc) 使用`USE_OP`告知编译器需要链接的Op，具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
```
USE_OP(mul);
```
如果只实现了CPU版本，则使用`USE_CPU_ONLY_OP`:
```
USE_CPU_ONLY_OP(gather);
```
如果OP不带Kernel，则使用`USE_NO_KENREL_OP`:
```
USE_NO_KENREL_OP(recurrent);
```
- 生成库
`paddle/operators` 目录下新增的 `*_op.cc` 文件会被自动添加链接到生成的lib库中。
系统会对新增的op自动绑定Python，并链接到生成的lib库中。
## 实现单元测试
......@@ -367,3 +337,10 @@ make test ARGS="-R test_mul_op -V"
```bash
ctest -R test_mul_op
```
## 注意事项
- 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，这将会导致编译出错。
- 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。
- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
- 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。
paddle/CMakeLists.txt
浏览文件 @ 9367fa12
......@@ -19,7 +19,7 @@ if(Boost_FOUND)
endif()
if(WITH_C_API)
add_subdirectory(capi)
add_subdirectory(capi)
endif()
if(WITH_SWIG_PY)
......
paddle/capi/CMakeLists.txt
浏览文件 @ 9367fa12
......@@ -28,48 +28,64 @@ 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(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}
DESTINATION lib/${ANDROID_ABI})
install(TARGETS paddle_capi_shared DESTINATION lib/${ANDROID_ABI})
install(TARGETS paddle_capi_whole paddle_capi_shared
ARCHIVE DESTINATION lib/${ANDROID_ABI}
LIBRARY DESTINATION lib/${ANDROID_ABI})
execute_process(
COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1
OUTPUT_VARIABLE GIT_COMMITS_LIST
RESULT_VARIABLE GIT_COMMITS_LIST_RESULT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
if(${GIT_COMMITS_LIST_RESULT})
set(GIT_COMMITS_LIST "No commits.")
endif()
install(CODE "FILE(WRITE ${CMAKE_INSTALL_PREFIX}/lib/${ANDROID_ABI}/BUILD.txt
\"Compiler:\n\"
\"\\t${CMAKE_C_COMPILER}\\n\"
\"\\t${CMAKE_CXX_COMPILER}\\n\"
\"Compiler Flags:\\n\"
\"\\t${CMAKE_F_FLAGS}\\n\"
\"\\t${CMAKE_CXX_FLAGS}\\n\"
\"Android API: ${CMAKE_SYSTEM_VERSION}\\n\"
\"Lastest commit:\\n\"
\"\\t${GIT_COMMITS_LIST}\\n\"
)"
)
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
......
paddle/cuda/include/hl_cuda_cudnn.h
浏览文件 @ 9367fa12
......@@ -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;
......
paddle/cuda/include/hl_tensor_ops.h
浏览文件 @ 9367fa12
......@@ -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);
}
};
......
paddle/cuda/src/hl_cuda_cnn.cu
浏览文件 @ 9367fa12
......@@ -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;
}
......
paddle/cuda/src/hl_cuda_cudnn.cc
浏览文件 @ 9367fa12
......@@ -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";
}
......
paddle/framework/lod_tensor.h
浏览文件 @ 9367fa12
......@@ -51,18 +51,15 @@ bool operator==(const LoD& a, const LoD& b);
* LoDTensor (Level of details Tensor)
* see https://en.wikipedia.org/wiki/Level_of_details for reference.
*/
class LoDTensor {
class LoDTensor : public Tensor {
public:
LoDTensor() {}
LoDTensor(const LoD& lod, Tensor* t) : lod_(lod), tensor_(t) {}
void set_lod(const LoD& lod) { lod_ = lod; }
void set_tensor(Tensor* tensor) { tensor_ = tensor; }
explicit LoDTensor(const LoD& lod) : lod_(lod) {}
Tensor& tensor() { return *tensor_; }
void set_lod(const LoD& lod) { lod_ = lod; }
LoD lod() { return lod_; }
LoD lod() const { return lod_; }
/*
* Get a element from LoD.
......@@ -104,7 +101,6 @@ class LoDTensor {
private:
LoD lod_;
Tensor* tensor_; // not owned
};
} // namespace framework
} // namespace paddle
paddle/framework/lod_tensor.md
浏览文件 @ 9367fa12
......@@ -4,13 +4,13 @@ PaddlePaddle's RNN doesn't require that all instances have the same length. To
## Challenge of Variable-length Inputs
People usually represent a mini-batch by a Tensor. For example, a mini-batch of 32 images, each of size 32x32, is a 10x32x32 Tensor. So a transformation, T, of all images can be a matrix multiplication of the 32x32xO-dimensional tensor T and the 10x32x32 Tensor.
People usually represent a mini-batch by a Tensor. For example, a mini-batch of 10 images, each of size 32x32, is a 10x32x32 Tensor. So a transformation, T, of all images can be a matrix multiplication of the 10xOx32-dimensional tensor T and the 10x32x32 Tensor.
Another example is that each mini-batch contains 32 sentences, where each word is a D-dimensional one-hot vector. If all sentences have the same length L, we can represent this mini-batch by a 32xLxD tensor. However, in most cases, sentences have variable lengths, and we will need an index data structure to record these variable lengths.
## LoD as a Solution
### Mini-Batch of variable-length sentenses
### Mini-Batch of variable-length sentences
Let's imagine a mini-batch of 3 variable lengths sentences, containing 3, 1, and 2 words respectively. We can represent it by a (3+1+2)xD tensor plus some index information:
......@@ -51,17 +51,17 @@ The many 1's on the second level seem duplicated. For this particular case of 2
In summary, as long as that the essential elements (words or images) have the same size, we can represent mini-batches by a LoD Tensor:
- The underlying tensor has size LxD1xD2x..., where D1xD2... is the size of the essential elements, and
- the first dimension size L has an additon property -- a LoD index as a nested vector:
- The first dimension size L has an additonal property -- a LoD index as a nested vector:
```c++
typedef std::vector<std::vector> > LoD;
typedef std::vector<std::<vector>> LoD;
```
- The LoD index can is not necessary when there are only two levels and all elements of the second level have length 1.
- The LoD index is not necessary when there are only two levels and all elements of the second level have length 1.
## Slicing of LoD Tensor
Consider that we have a network with three levels of RNN: the top level one handles articles, the second level one handles sentences, and the basic level one handles words. This network requires that mini-batches represented by 4 level LoD Tensor, for example,
Consider that we have a network with three levels of RNN: the top level one handles articles, the second level one handles sentences, and the basic level one handles words. This network requires that mini-batches represented by 3 level LoD Tensor, for example,
```
3
......@@ -90,8 +90,9 @@ and the <1,2>-slice of above example is
Let's go on slicing this slice. Its <1,1>-slice is
```
3
|||
1
1
|
```
### The Slicing Algorithm
......@@ -99,7 +100,7 @@ Let's go on slicing this slice. Its <1,1>-slice is
The algorithm, with over-simplified data structure, is defined as
```c++
typedef vector<vector<int> > LoD;
typedef std::vector<std::vector<int>> LoD;
struct LoDTensor {
LoD lod_;
......@@ -128,7 +129,7 @@ Suppose that we want to retrieve the <1,2>-slice
we will need to find out the starting position of this slice by summing over all leaf nodes in `LoD` to the left of the slice, i.e., 3 + 2 + 4 + 1 = 10.
To avoid the traversal of the LoD tree at slcing time, we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level. For example, above LoD Tensor can be transformed into
To avoid the traversal of the LoD tree at slicing time, we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level. For example, above LoD Tensor can be transformed into
```
0
......
paddle/framework/lod_tensor_test.cc
浏览文件 @ 9367fa12
......@@ -36,69 +36,64 @@ class LoDTensorTester : public ::testing::Test {
ASSERT_EQ(lod.size(), 3UL);
tensor.Resize({20 /*batch size*/, 128 /*dim*/});
lod_tensor_.Resize({20 /*batch size*/, 128 /*dim*/});
// malloc memory
tensor.mutable_data<float>(place);
lod_tensor_.mutable_data<float>(place);
lod_tensor.set_lod(lod);
lod_tensor.set_tensor(&tensor);
lod_tensor_.set_lod(lod);
}
protected:
platform::CPUPlace place;
Tensor tensor;
LoDTensor lod_tensor;
LoDTensor lod_tensor_;
};
TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor.NumLevels(), 3UL); }
TEST_F(LoDTensorTester, NumLevels) { ASSERT_EQ(lod_tensor_.NumLevels(), 3UL); }
TEST_F(LoDTensorTester, NumElements) {
ASSERT_EQ(lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(lod_tensor.NumElements(2), 8UL);
ASSERT_EQ(lod_tensor_.NumElements(0), 2UL);
ASSERT_EQ(lod_tensor_.NumElements(1), 4UL);
ASSERT_EQ(lod_tensor_.NumElements(2), 8UL);
}
TEST_F(LoDTensorTester, SliceLevels) {
// slice 1 level
for (size_t level = 0; level < 3UL; ++level) {
LoDTensor new_lod_tensor = lod_tensor;
LoDTensor new_lod_tensor = lod_tensor_;
new_lod_tensor.SliceLevels(level, level + 1);
ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
lod_tensor.tensor().data<float>());
ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
}
// slice 2 level
for (size_t level = 0; level < 2UL; ++level) {
LoDTensor new_lod_tensor = lod_tensor;
LoDTensor new_lod_tensor = lod_tensor_;
new_lod_tensor.SliceLevels(level, level + 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor.NumElements(level));
ASSERT_EQ(new_lod_tensor.NumElements(1), lod_tensor.NumElements(level + 1));
ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
lod_tensor.tensor().data<float>());
ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor_.NumElements(level));
ASSERT_EQ(new_lod_tensor.NumElements(1),
lod_tensor_.NumElements(level + 1));
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
}
}
TEST_F(LoDTensorTester, SliceInLevel) {
size_t level = 0;
LoDTensor new_lod_tensor = lod_tensor;
LoDTensor new_lod_tensor = lod_tensor_;
new_lod_tensor.SliceInLevel(level, 0, 2);
EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
lod_tensor.tensor().data<float>());
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
level = 1;
new_lod_tensor = lod_tensor;
new_lod_tensor = lod_tensor_;
new_lod_tensor.SliceInLevel(level, 0, 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.tensor().data<float>(),
lod_tensor.tensor().data<float>());
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
}
} // namespace framework
......
paddle/framework/lod_tensor_test.cu
浏览文件 @ 9367fa12
......@@ -26,18 +26,16 @@ __global__ void test(size_t* a, int size) {
}
TEST(LoDTensor, LoDInGPU) {
paddle::framework::Tensor tensor;
paddle::framework::LoDTensor lod_tensor;
paddle::platform::GPUPlace place(0);
paddle::framework::LoD src_lod;
src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});
tensor.Resize({14, 16});
tensor.mutable_data<float>(place);
lod_tensor.Resize({14, 16});
lod_tensor.mutable_data<float>(place);
lod_tensor.set_lod(src_lod);
lod_tensor.set_tensor(&tensor);
CHECK_EQ(lod_tensor.lod_element(0, 2), 4);
CHECK_EQ(lod_tensor.lod_element(0, 4), 8);
......
paddle/framework/operator.cc
浏览文件 @ 9367fa12
......@@ -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
......@@ -186,6 +186,48 @@ void OperatorBase::GenerateTemporaryNames() {
}
}
template <>
const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const {
auto* var = InputVar(name);
return var == nullptr ? nullptr : GetTensorFromVar(var);
}
template <>
const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
const std::string& name) const {
auto names = op().Inputs(name);
std::vector<const Tensor*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) {
auto var = scope_.FindVar(sub_name);
return var == nullptr ? nullptr : GetTensorFromVar(var);
});
return res;
}
template <>
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
auto* var = OutputVar(name);
return var == nullptr ? nullptr : const_cast<Tensor*>(GetTensorFromVar(var));
}
template <>
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
const std::string& name) const {
auto names = op().Outputs(name);
std::vector<Tensor*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) {
auto var = scope().FindVar(sub_name);
return var == nullptr
? nullptr
: const_cast<Tensor*>(GetTensorFromVar(var));
});
return res;
}
void OpProtoAndCheckerMaker::Validate() {
validated_ = true;
CheckNoDuplicatedInOutAttrs();
......
paddle/framework/operator.h
浏览文件 @ 9367fa12
......@@ -22,6 +22,7 @@ limitations under the License. */
#include "op_info.h"
#include "paddle/framework/attribute.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/scope.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/device_context.h"
......@@ -326,11 +327,27 @@ class InferShapeContext {
return res;
}
const Tensor* GetTensorFromVar(const Variable* var) const {
if (var->IsType<LoDTensor>()) {
return &var->Get<LoDTensor>();
}
PADDLE_ENFORCE(var->IsType<Tensor>(),
"The Input(%s) must be LoDTensor or Tensor.");
return &var->Get<Tensor>();
}
private:
const OperatorBase& op_;
const Scope& scope_;
};
template <>
const Tensor* InferShapeContext::Input<Tensor>(const std::string& name) const;
template <>
const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
const std::string& name) const;
template <typename T>
struct EigenDeviceConverter;
......@@ -349,7 +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,
......@@ -357,15 +374,44 @@ 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_;
}
const platform::DeviceContext* device_context_;
// redefine Output function,
// use Variable::Get instead of Variable::GetMutable
template <typename T>
T* Output(const std::string& name) const {
auto var = OutputVar(name);
return var == nullptr ? nullptr : const_cast<T*>(&var->Get<T>());
}
// redefine MultiOutput function.
// use Variable::Get instead of Variable::GetMutable
template <typename T>
std::vector<T*> MultiOutput(const std::string& name) const {
auto names = op().Outputs(name);
std::vector<T*> res;
res.reserve(names.size());
std::transform(
names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) { return Output<T>(sub_name); });
return res;
}
private:
const platform::DeviceContext& device_context_;
};
template <>
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
template <>
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
const std::string& name) const;
class OpKernel {
public:
/**
......@@ -416,7 +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>&
......
paddle/framework/tensor_impl.h
浏览文件 @ 9367fa12
......@@ -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 "
......
paddle/framework/tensor_test.cc
浏览文件 @ 9367fa12
......@@ -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
}
paddle/function/neon/NeonDepthwiseConv.cpp
浏览文件 @ 9367fa12
......@@ -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());
......
paddle/function/neon/NeonDepthwiseConv.h
浏览文件 @ 9367fa12
......@@ -594,7 +594,7 @@ struct StridePadding {
float32x4_t s1 = vdupq_n_f32(0.f);
for (int s = 0; s < step; s++) {
float32x4_t s0 = vld1q_f32(input);
float32x4x2_t v = {s0, s1};
float32x4x2_t v = {{s0, s1}};
vst2q_f32(inputPadding, v);
input += 4;
inputPadding += 8;
......
paddle/gserver/activations/ActivationFunction.cpp
浏览文件 @ 9367fa12
......@@ -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);
}
......
paddle/gserver/activations/MKLDNNActivation.cpp 0 → 100644
浏览文件 @ 9367fa12
/* 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
paddle/gserver/activations/MKLDNNActivation.h 0 → 100644
浏览文件 @ 9367fa12
/* 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
paddle/gserver/layers/CudnnPoolLayer.cpp
浏览文件 @ 9367fa12
......@@ -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;
......
paddle/gserver/layers/DetectionOutputLayer.cpp
浏览文件 @ 9367fa12
......@@ -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();
......
paddle/gserver/layers/ExpandConvBaseLayer.cpp 已删除 100644 → 0
浏览文件 @ 4520afcf
/* 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
paddle/gserver/layers/ExpandConvLayer.cpp
浏览文件 @ 9367fa12
......@@ -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);
}
......
paddle/gserver/layers/ExpandConvLayer.h
浏览文件 @ 9367fa12
......@@ -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_;
......
paddle/gserver/layers/Layer.cpp
浏览文件 @ 9367fa12
......@@ -14,26 +14,12 @@ limitations under the License. */
#include "paddle/utils/Util.h"
#include "CostLayer.h"
#include "ValidationLayer.h"
#include "paddle/math/SparseMatrix.h"
#include "paddle/utils/Error.h"
#include "paddle/utils/Logging.h"
#include "AddtoLayer.h"
#include "CRFLayer.h"
#include "CosSimLayer.h"
#include "CostLayer.h"
#include "DataLayer.h"
#include "ExpandConvLayer.h"
#include "FullyConnectedLayer.h"
#include "HierarchicalSigmoidLayer.h"
#include "MaxLayer.h"
#include "MixedLayer.h"
#include "NormLayer.h"
#include "PoolLayer.h"
#include "TensorLayer.h"
#include "TransLayer.h"
#include "ValidationLayer.h"
DEFINE_bool(log_error_clipping, false, "enable log error clipping or not");
namespace paddle {
......@@ -109,6 +95,10 @@ ClassRegistrar<Layer, LayerConfig> Layer::registrar_;
LayerPtr Layer::create(const LayerConfig& config) {
std::string type = config.type();
// NOTE: As following types have illegal character '-',
// they can not use REGISTER_LAYER to registrar.
// Besides, to fit with old training models,
// they can not use '_' instead.
if (type == "multi-class-cross-entropy")
return LayerPtr(new MultiClassCrossEntropy(config));
else if (type == "rank-cost")
......@@ -117,8 +107,6 @@ LayerPtr Layer::create(const LayerConfig& config) {
return LayerPtr(new AucValidation(config));
else if (type == "pnpair-validation")
return LayerPtr(new PnpairValidation(config));
// NOTE: stop adding "if" statements here.
// Instead, use REGISTER_LAYER to add more layer types
return LayerPtr(registrar_.createByType(config.type(), config));
}
......
paddle/gserver/layers/MKLDNNConvLayer.cpp 0 → 100644
浏览文件 @ 9367fa12
/* 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 "MKLDNNConvLayer.h"
#include "paddle/math/MathUtils.h"
#include "paddle/utils/Logging.h"
using namespace mkldnn; // NOLINT
typedef memory::format format;
namespace paddle {
REGISTER_LAYER(mkldnn_conv, MKLDNNConvLayer);
bool MKLDNNConvLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
if (!MKLDNNLayer::init(layerMap, parameterMap)) {
return false;
}
CHECK_EQ(inputLayers_.size(), 1) << "Only support one input layer yet";
CHECK_EQ(inputLayers_.size(), parameters_.size());
CHECK(config_.shared_biases()) << "Only support shared biases yet";
oc_ = config_.num_filters();
const ConvConfig& conf = config_.inputs(0).conv_conf();
ic_ = conf.channels();
fw_ = conf.filter_size();
fh_ = conf.filter_size_y();
pw_ = conf.padding();
ph_ = conf.padding_y();
dw_ = conf.dilation();
dh_ = conf.dilation_y();
sw_ = conf.stride();
sh_ = conf.stride_y();
gp_ = conf.groups();
oh_ = conf.output_y();
ow_ = conf.output_x();
ih_ = conf.img_size_y();
iw_ = conf.img_size();
caffeMode_ = conf.caffe_mode();
CHECK(caffeMode_) << "Only support caffe mode yet";
CHECK(dh_ == 1 && dw_ == 1) << "Only support dilation 1 yet";
// check group setting
CHECK_EQ((oc_ / gp_) * gp_, oc_) << "group is indivisible for oc";
CHECK_EQ((ic_ / gp_) * gp_, ic_) << "group is indivisible for ic";
// create weight
size_t height = oc_ / gp_;
size_t width = ic_ * fh_ * fw_;
CHECK_EQ(parameters_[0]->getSize(), height * width);
weight_ =
std::unique_ptr<Weight>(new Weight(height, width, parameters_[0], 0));
// create biases
if (biasParameter_.get() != NULL) {
biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_));
}
return true;
}
void MKLDNNConvLayer::convertWeightsFromPaddle() {
if (hasInitedWgt_) {
return;
}
CHECK(wgtVal_) << "should have been initialized";
// the paddle weight format is oihw or goihw
auto targetDim = wgtVal_->getDims();
auto srcFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw;
wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
hasInitedWgt_ = true;
}
void MKLDNNConvLayer::convertWeightsToPaddle() {
CHECK(wgtVal_) << "should have been initialized";
auto targetDim = wgtVal_->getDims();
auto dstFmt = (gp_ == 1) ? memory::format::oihw : memory::format::goihw;
wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
}
void MKLDNNConvLayer::reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
reshapeInput(bs, ih, iw);
// cal output sizes
// oc can not be changed
int fh = (fh_ - 1) * dh_ + 1;
int fw = (fw_ - 1) * dw_ + 1;
oh = outputSize(ih, fh, ph_, sh_, caffeMode_);
ow = outputSize(iw, fw, pw_, sw_, caffeMode_);
reshapeOutput(oh, ow);
resizeOutput(bs, oc * oh * ow);
printSizeInfo();
}
void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
resetFwdPD(fwdPD_);
resetFwdBuffers(fwdPD_, in, wgt, bias, out);
resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
printValueFormatFlow();
}
void MKLDNNConvLayer::resetBwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
std::shared_ptr<conv_bwdWgt::primitive_desc> bwdWgtPD;
std::shared_ptr<conv_bwdData::primitive_desc> bwdDataPD;
resetBwdWgtPD(bwdWgtPD);
resetBwdDataPD(bwdDataPD);
resetBwdBuffers(bwdWgtPD, bwdDataPD, in, wgt, bias, out);
resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
printGradFormatFlow();
}
void MKLDNNConvLayer::updateInputData() {
cpuInVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
}
void MKLDNNConvLayer::updateWeights(const UpdateCallback& callback) {
weight_->getParameterPtr()->incUpdate(callback);
if (biases_ && biases_->getWGrad()) {
biases_->getParameterPtr()->incUpdate(callback);
}
}
void MKLDNNConvLayer::loadConvSettings(memory::dims& wgt,
memory::dims& bias,
memory::dims& stride,
memory::dims& dilation,
memory::dims& padL,
memory::dims& padR) {
wgt = (gp_ == 1) ? memory::dims{oc_, ic_, fh_, fw_}
: memory::dims{gp_, oc_ / gp_, ic_ / gp_, fh_, fw_};
bias = memory::dims{oc_};
stride = memory::dims{sh_, sw_};
padL = memory::dims{ph_, pw_};
padR = getPaddingR();
// note: mkldnn dilation start from 0
dilation = memory::dims{dh_ - 1, dw_ - 1};
}
void MKLDNNConvLayer::resetFwdPD(
std::shared_ptr<conv_fwd::primitive_desc>& pd) {
// dims for conv
memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
: prop_kind::forward_training;
algorithm algo = algorithm::convolution_direct;
padding_kind padKind = padding_kind::zero;
conv_fwd::desc fwdDesc =
biases_ && biases_->getW()
? conv_fwd::desc(pk,
algo,
MKLDNNMatrix::createMemoryDesc(inDims),
MKLDNNMatrix::createMemoryDesc(wgtDims),
MKLDNNMatrix::createMemoryDesc(biasDims),
MKLDNNMatrix::createMemoryDesc(outDims),
strides,
dilations,
padL,
padR,
padKind)
: conv_fwd::desc(pk,
algo,
MKLDNNMatrix::createMemoryDesc(inDims),
MKLDNNMatrix::createMemoryDesc(wgtDims),
MKLDNNMatrix::createMemoryDesc(outDims),
strides,
dilations,
padL,
padR,
padKind);
pd.reset(new conv_fwd::primitive_desc(fwdDesc, engine_));
}
void MKLDNNConvLayer::resetFwdBuffers(
std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
CHECK(pd);
resetInValue(pd, in);
resetWgtBiasValue(pd, wgt, bias);
resetOutValue(pd, out);
}
void MKLDNNConvLayer::resetFwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
pipeline.clear();
if (cvtInVal_) {
pipeline.push_back(*cvtInVal_);
}
if (bias) {
fwd_.reset(new conv_fwd(*pd, *in, *wgt, *bias, *out));
} else {
fwd_.reset(new conv_fwd(*pd, *in, *wgt, *out));
}
pipeline.push_back(*fwd_);
if (cvtOutVal_) {
pipeline.push_back(*cvtOutVal_);
}
}
void MKLDNNConvLayer::resetInValue(
std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& in) {
const MatrixPtr& inMat = inputLayers_[0]->getOutput().value;
in = MKLDNNMatrix::create(inMat, pd->src_primitive_desc());
// create buffer and reorder if input value do not match
cpuInVal_ = nullptr;
cvtInVal_ = nullptr;
if (inputIsOnlyMKLDNN()) {
MKLDNNMatrixPtr dnnIn = std::dynamic_pointer_cast<MKLDNNMatrix>(inMat);
CHECK(dnnIn) << "Input should be MKLDNNMatrix";
if (dnnIn->getPrimitiveDesc() != in->getPrimitiveDesc()) {
CHECK_EQ(dnnIn->getFormat(), format::nc);
CHECK(ih_ == 1 && iw_ == 1) << "when input is nc format";
// create a new one with nchw format and same data
memory::dims inDims = memory::dims{bs_, ic_, 1, 1};
dnnIn = MKLDNNMatrix::create(inMat, inDims, format::nchw, engine_);
CHECK(dnnIn->getPrimitiveDesc() == in->getPrimitiveDesc());
}
in = dnnIn;
} else {
const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
cpuInVal_ = MKLDNNMatrix::create(cpuIn, inDims, format::nchw, engine_);
if (cpuInVal_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
// create new mkldnn matrix
in = MKLDNNMatrix::create(nullptr, pd->src_primitive_desc());
cvtInVal_ = MKLDNNMatrix::createReorder(cpuInVal_, in);
CHECK(cvtInVal_) << "should not be emptry";
} else {
in = cpuInVal_;
}
}
}
void MKLDNNConvLayer::resetWgtBiasValue(
std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias) {
wgt = MKLDNNMatrix::create(weight_->getW(), pd->weights_primitive_desc());
VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
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());
// 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;
memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
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 MKLDNNConvLayer::resetBwdWgtPD(
std::shared_ptr<conv_bwdWgt::primitive_desc>& pd) {
memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
// create backward weight using input, output and weight value memory desc
CHECK(inVal_) << "Should have input value";
CHECK(outVal_) << "Should have output value";
CHECK(wgtVal_) << "Should have weight value";
algorithm algo = algorithm::convolution_direct;
padding_kind padKind = padding_kind::zero;
auto bwdWgtDesc = biasVal_ != nullptr
? conv_bwdWgt::desc(algo,
inVal_->getMemoryDesc(),
wgtVal_->getMemoryDesc(),
biasVal_->getMemoryDesc(),
outVal_->getMemoryDesc(),
strides,
padL,
padR,
padKind)
: conv_bwdWgt::desc(algo,
inVal_->getMemoryDesc(),
wgtVal_->getMemoryDesc(),
outVal_->getMemoryDesc(),
strides,
padL,
padR,
padKind);
pd.reset(new conv_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
CHECK(pd->src_primitive_desc() == inVal_->getPrimitiveDesc())
<< "primitive desc of in value should equal";
CHECK(pd->diff_dst_primitive_desc() == outVal_->getPrimitiveDesc())
<< "primitive desc of out grad should equal the out value";
CHECK(pd->diff_weights_primitive_desc() == wgtVal_->getPrimitiveDesc())
<< "primitive desc of weight grad should equal the weight value";
}
void MKLDNNConvLayer::resetBwdDataPD(
std::shared_ptr<conv_bwdData::primitive_desc>& pd) {
pd = nullptr;
if (inputLayers_[0]->getOutput().grad == nullptr) {
return;
}
memory::dims wgtDims, biasDims, strides, dilations, padL, padR;
loadConvSettings(wgtDims, biasDims, strides, dilations, padL, padR);
CHECK(inVal_) << "Should have input value";
CHECK(outVal_) << "Should have output value";
// create backward data using input and output value memory desc
// but using weight memory desc with any format
auto bwdDataDesc = conv_bwdData::desc(algorithm::convolution_direct,
inVal_->getMemoryDesc(),
MKLDNNMatrix::createMemoryDesc(wgtDims),
outVal_->getMemoryDesc(),
strides,
padL,
padR,
padding_kind::zero);
pd.reset(new conv_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
CHECK(pd->diff_src_primitive_desc() == inVal_->getPrimitiveDesc())
<< "primitive desc of in grad should equal the in value";
CHECK(pd->diff_dst_primitive_desc() == outVal_->getPrimitiveDesc())
<< "primitive desc of out grad should equal";
}
void MKLDNNConvLayer::resetBwdBuffers(
std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
CHECK(wgtPD);
resetOutGrad(wgtPD, out);
resetWgtBiasGrad(wgtPD, wgt, bias);
resetInGrad(dataPD, in);
resetWgtValBwdData(dataPD, wgtValBwdData_);
}
void MKLDNNConvLayer::resetBwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
pipeline.clear();
if (cvtOutGrad_) {
pipeline.push_back(*cvtOutGrad_);
}
// add bwdWgt handle
if (bias) {
bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt, *bias));
} else {
bwdWgt_.reset(new conv_bwdWgt(*wgtPD, *inVal_, *out, *wgt));
}
pipeline.push_back(*bwdWgt_);
if (dataPD == nullptr) {
return;
}
if (cvtWgtVal_) {
pipeline.push_back(*cvtWgtVal_);
}
// add bwdData handle
CHECK(wgtValBwdData_) << "Should have weight memory";
bwdData_.reset(new conv_bwdData(*dataPD, *out, *wgtValBwdData_, *in));
pipeline.push_back(*bwdData_);
if (cvtInGrad_) {
pipeline.push_back(*cvtInGrad_);
}
}
void MKLDNNConvLayer::resetOutGrad(
std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD, MKLDNNMatrixPtr& out) {
const MatrixPtr& outMat = output_.grad;
out = MKLDNNMatrix::create(outMat, wgtPD->diff_dst_primitive_desc());
CHECK(outVal_ != nullptr &&
out->getPrimitiveDesc() == outVal_->getPrimitiveDesc())
<< "primitive desc of out grad and value should be equal";
// TODO(TJ): merge outgrad
// create reorder if has output grad does not match
cpuOutGrad_ = nullptr;
cvtOutGrad_ = nullptr;
if (!outputIsOnlyMKLDNN()) {
const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
// same PrimitiveDesc with cpuInVal_
CHECK(cpuOutVal_);
cpuOutGrad_ = MKLDNNMatrix::create(cpuOut, cpuOutVal_->getPrimitiveDesc());
if (cpuOutGrad_->getPrimitiveDesc() == out->getPrimitiveDesc()) {
outMat->setData(cpuOut->getData());
out = cpuOutGrad_;
} else {
cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
CHECK(cvtOutGrad_);
}
}
}
void MKLDNNConvLayer::resetWgtBiasGrad(
std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias) {
wgt = MKLDNNMatrix::create(weight_->getWGrad(),
wgtPD->diff_weights_primitive_desc());
CHECK(nullptr != wgtVal_ &&
wgt->getPrimitiveDesc() == wgtVal_->getPrimitiveDesc())
<< "primitive desc of weight grad and value should be equal";
VLOG(MKLDNN_FMTS) << "weight grad format: " << wgt->getFormat();
bias = nullptr;
if (biasVal_ == nullptr) {
return;
}
bias = MKLDNNMatrix::create(biases_->getWGrad(),
wgtPD->diff_bias_primitive_desc());
CHECK(bias->getPrimitiveDesc() == biasVal_->getPrimitiveDesc())
<< "primitive desc of bias grad should equal the bias value";
}
void MKLDNNConvLayer::resetInGrad(
std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& in) {
if (dataPD == nullptr) {
return;
}
// TODO(TJ): use outputMaps_ ways to get the inGrad_ when merge outgrad done
in = MKLDNNMatrix::create(inputLayers_[0]->getOutput().grad,
dataPD->diff_src_primitive_desc());
CHECK(nullptr != inVal_ &&
in->getPrimitiveDesc() == inVal_->getPrimitiveDesc())
<< "primitive desc of input grad and value should be equal";
// create reorder if has output grad does not match
cpuInGrad_ = nullptr;
cvtInGrad_ = nullptr;
if (!inputIsOnlyMKLDNN()) {
const MatrixPtr& cpuIn = getInputGrad(0, CPU_DEVICE);
// same PrimitiveDesc with cpuInVal_
CHECK(cpuInVal_);
cpuInGrad_ = MKLDNNMatrix::create(cpuIn, cpuInVal_->getPrimitiveDesc());
if (cpuInGrad_->getPrimitiveDesc() != in->getPrimitiveDesc()) {
const MatrixPtr& dnnIn = getInputGrad(0, MKLDNN_DEVICE);
in = MKLDNNMatrix::create(dnnIn, in->getPrimitiveDesc());
cvtInGrad_ = MKLDNNMatrix::createReorder(in, cpuInGrad_);
CHECK(cvtInGrad_);
} else {
in = cpuInGrad_;
}
}
}
void MKLDNNConvLayer::resetWgtValBwdData(
std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& wgt) {
if (dataPD == nullptr) {
return;
}
// create new weight value for backward data, and create reorder if necessary
// since the primitive_desc would be different with wgtVal_
CHECK(wgtVal_) << "should have weight value";
if (dataPD->weights_primitive_desc() != wgtVal_->getPrimitiveDesc()) {
wgtValBwdData_ =
MKLDNNMatrix::create(nullptr, dataPD->weights_primitive_desc());
cvtWgtVal_ = MKLDNNMatrix::createReorder(wgtVal_, wgtValBwdData_);
CHECK(cvtWgtVal_);
} else {
wgtValBwdData_ = wgtVal_;
}
VLOG(MKLDNN_FMTS) << "weight value format for backward data"
<< wgtValBwdData_->getFormat();
}
} // namespace paddle
paddle/gserver/layers/MKLDNNConvLayer.h 0 → 100644
浏览文件 @ 9367fa12
/* 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::convolution_forward conv_fwd;
typedef mkldnn::convolution_backward_weights conv_bwdWgt;
typedef mkldnn::convolution_backward_data conv_bwdData;
/**
* @brief A subclass of MKLDNNLayer conv layer.
*
* The config file api is mkldnn_conv
*/
class MKLDNNConvLayer : public MKLDNNLayer {
protected:
// padding height and width
int ph_, pw_;
// stride height and width
int sh_, sw_;
// dilation height and width
int dh_, dw_;
// filter(kenerl) height and width
int fh_, fw_;
// group number
int gp_;
// in resetBwdData, the format of wgtValBwdData_ is different with wgtVal_
MKLDNNMatrixPtr wgtValBwdData_;
// convert handle from wgtVal_ to wgtValBwdData_
std::shared_ptr<mkldnn::reorder> cvtWgtVal_;
// save forward primitive_desc, which can be used backward
std::shared_ptr<conv_fwd::primitive_desc> fwdPD_;
// MKLDNNMatrixPtr which should be created from CPU Device
MKLDNNMatrixPtr cpuInVal_;
MKLDNNMatrixPtr cpuInGrad_;
MKLDNNMatrixPtr cpuOutVal_;
MKLDNNMatrixPtr cpuOutGrad_;
// convert handle between CPU device and MKLDNN device
std::shared_ptr<mkldnn::reorder> cvtInVal_;
std::shared_ptr<mkldnn::reorder> cvtInGrad_;
std::shared_ptr<mkldnn::reorder> cvtOutVal_;
std::shared_ptr<mkldnn::reorder> cvtOutGrad_;
// whether the weight has been init
bool hasInitedWgt_;
// true by default, which impact the calculation of output image size.
// details can refer to mathUtil.h
bool caffeMode_;
// weight and bias
std::unique_ptr<Weight> weight_;
std::unique_ptr<Weight> biases_;
public:
explicit MKLDNNConvLayer(const LayerConfig& config)
: MKLDNNLayer(config), hasInitedWgt_(false), caffeMode_(true) {}
~MKLDNNConvLayer() {}
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 updateWeights(const UpdateCallback& callback) override;
void convertWeightsFromPaddle() override;
void convertWeightsToPaddle() override;
void printSizeInfo() override {
MKLDNNLayer::printSizeInfo();
VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
<< ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
<< ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
}
void printValueFormatFlow() override {
if (cpuInVal_) {
VLOG(MKLDNN_FMTS) << cpuInVal_->getFormat() << " >>>";
}
MKLDNNLayer::printValueFormatFlow();
if (cpuOutVal_) {
VLOG(MKLDNN_FMTS) << " >>> " << cpuOutVal_->getFormat();
}
}
void printGradFormatFlow() override {
if (cpuInGrad_) {
VLOG(MKLDNN_FMTS) << cpuInGrad_->getFormat() << " <<<";
}
MKLDNNLayer::printGradFormatFlow();
if (cpuOutGrad_) {
VLOG(MKLDNN_FMTS) << " <<< " << cpuOutGrad_->getFormat();
}
}
protected:
/**
* load the dims settings of this conv
*/
void loadConvSettings(mkldnn::memory::dims& wgt,
mkldnn::memory::dims& bias,
mkldnn::memory::dims& stride,
mkldnn::memory::dims& dilation,
mkldnn::memory::dims& padL,
mkldnn::memory::dims& padR);
/**
* reset the forward primitive descriptor.
*/
void resetFwdPD(std::shared_ptr<conv_fwd::primitive_desc>& pd);
/**
* reset the MKLDNNMatrix buffers used in forward.
*/
void resetFwdBuffers(std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
/**
* reset the forward pipeline.
*/
void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
/**
* reset MKLDNNMatrix of input value
*/
void resetInValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in);
/**
* reset MKLDNNMatrix of weight and bias value
*/
void resetWgtBiasValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias);
/**
* reset MKLDNNMatrix of output value
*/
void resetOutValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& out);
/**
* reset the backward weight primitive descriptor.
*/
void resetBwdWgtPD(std::shared_ptr<conv_bwdWgt::primitive_desc>& pd);
/**
* reset the backward data primitive descriptor.
*/
void resetBwdDataPD(std::shared_ptr<conv_bwdData::primitive_desc>& pd);
/**
* reset the MKLDNNMatrix buffers used in backward.
*/
void resetBwdBuffers(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
/**
* reset the backward pipeline.
*/
void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out);
/**
* reset MKLDNNMatrix of output grad
*/
void resetOutGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
MKLDNNMatrixPtr& out);
/**
* reset MKLDNNMatrix of weight and bias grad
*/
void resetWgtBiasGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias);
/**
* reset MKLDNNMatrix of input grad
*/
void resetInGrad(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& in);
/**
* reset MKLDNNMatrix of weight value for backward data
* since the primitive_desc would be different with wgtVal_
*/
void resetWgtValBwdData(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
MKLDNNMatrixPtr& wgt);
/**
* get padding_r according to
* https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
* test_convolution_forward_common.hpp
* @note: mkldnn dilation start from 0 while paddle start from 1
*/
mkldnn::memory::dims getPaddingR() const {
mkldnn::memory::dims padR = {ph_, pw_};
for (int i = 0; i < 2; ++i) {
if ((ih_ - ((fh_ - 1) * dh_ + 1) + ph_ + padR[0]) / sh_ + 1 != oh_) {
++padR[0];
}
if ((iw_ - ((fw_ - 1) * dw_ + 1) + pw_ + padR[1]) / sw_ + 1 != ow_) {
++padR[1];
}
}
return padR;
}
};
} // namespace paddle
paddle/gserver/layers/MKLDNNFcLayer.cpp
浏览文件 @ 9367fa12
......@@ -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
paddle/gserver/layers/MKLDNNFcLayer.h
浏览文件 @ 9367fa12
......@@ -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
paddle/gserver/layers/MKLDNNLayer.h
浏览文件 @ 9367fa12
......@@ -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;
}
......
paddle/gserver/layers/MKLDNNPoolLayer.cpp 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/gserver/layers/MKLDNNPoolLayer.h 0 → 100644
浏览文件 @ 9367fa12
/* 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
paddle/gserver/layers/SequenceSliceLayer.cpp
浏览文件 @ 9367fa12
......@@ -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)
......
paddle/gserver/tests/MKLDNNTester.h
浏览文件 @ 9367fa12
......@@ -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_;
......
paddle/math/BaseMatrix.cu
浏览文件 @ 9367fa12
......@@ -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) {
......
paddle/math/MKLDNNMatrix.cpp
浏览文件 @ 9367fa12
此差异已折叠。
paddle/math/MKLDNNMatrix.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/math/MathFunctions.h
浏览文件 @ 9367fa12
......@@ -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>
......
paddle/math/Matrix.cpp
浏览文件 @ 9367fa12
此差异已折叠。
paddle/memory/memcpy.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/CMakeLists.txt
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/accuracy_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/accuracy_op.cu 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/accuracy_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/add_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/concat_op.cc
浏览文件 @ 9367fa12
......@@ -25,8 +25,11 @@ 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::Tensor>("Out");
auto *out = ctx.Output<framework::LoDTensor>("Out");
size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
size_t n = ins.size();
......
paddle/operators/cond_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/cond_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/cos_sim_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/cos_sim_op.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/cross_entropy_op.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/dropout_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/dropout_op.cu 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/dropout_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/expand_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/expand_op.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/fc_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/gather_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/identity_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/lookup_table_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/mean_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/minus_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/modified_huber_loss_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/modified_huber_loss_op.cu 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/modified_huber_loss_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/mul_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/mul_op.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/name_convention.md
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/pad_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/pad_op.cu 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/pad_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/prelu_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/prelu_op.cu 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/prelu_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/recurrent_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/reshape_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/rowwise_add_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/scale_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/scatter_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/sequence_avg_pool_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/sequence_avg_pool_op.cu 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/sequence_avg_pool_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/sgd_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/sigmoid_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/smooth_l1_loss_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/smooth_l1_loss_op.cu 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/smooth_l1_loss_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/softmax_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/split_op.cc 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/split_op.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/sum_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/operators/top_k_op.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/platform/CMakeLists.txt
浏览文件 @ 9367fa12
此差异已折叠。
paddle/platform/cuda_helper.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/platform/details/device_ptr_cast.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/platform/device_context.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/platform/device_context.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/platform/enforce.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/platform/transform.h 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/platform/transform_test.cu 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
paddle/pserver/CMakeLists.txt
浏览文件 @ 9367fa12
此差异已折叠。
paddle/pybind/CMakeLists.txt
浏览文件 @ 9367fa12
此差异已折叠。
paddle/pybind/pybind.cc
浏览文件 @ 9367fa12
此差异已折叠。
paddle/scripts/travis/build_ios.sh 0 → 100755
浏览文件 @ 9367fa12
此差异已折叠。
paddle/trainer/CMakeLists.txt
浏览文件 @ 9367fa12
此差异已折叠。
paddle/utils/Excepts.h
浏览文件 @ 9367fa12
此差异已折叠。
paddle/utils/arch/linux/Locks.cpp
浏览文件 @ 9367fa12
此差异已折叠。
paddle/utils/arch/osx/Excepts.cpp
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/op.py
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_accuracy_op.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_cond_op.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_dropout_op.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_fc_op.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_identity_op.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_pad_op.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_prelu_op.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_seq_pool.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
python/paddle/v2/framework/tests/test_split_op.py 0 → 100644
浏览文件 @ 9367fa12
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或