Skip to content

P
Paddle

GoAI / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 a2a0d6f8 编写于 作者: C caoying03
浏览文件
操作

Merge branch 'develop' into softmax_with_cross_entropy_op

上级 8f8ea005 414a7a1e
展开全部 隐藏空白更改
内联 并排
Showing with 5073 addition and 1297 deletion
.gitignore
浏览文件 @ a2a0d6f8
......@@ -27,3 +27,4 @@ CMakeFiles
cmake_install.cmake
paddle/.timestamp
python/paddlepaddle.egg-info/
paddle/pybind/pybind.h
.travis.yml
浏览文件 @ a2a0d6f8
......@@ -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' "$@"; }
script:
......
CMakeLists.txt
浏览文件 @ a2a0d6f8
......@@ -27,7 +27,7 @@ if(NOT CMAKE_CROSSCOMPILING)
endif(NOT CMAKE_CROSSCOMPILING)
find_package(Git REQUIRED)
find_package(Threads REQUIRED)
if(NOT ANDROID)
if(NOT ANDROID AND NOT IOS)
find_package(Boost QUIET)
endif()
......@@ -64,27 +64,29 @@ if(NOT CMAKE_BUILD_TYPE)
FORCE)
endif()
if(ANDROID)
if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
# TODO: support glog for Android api 16 ~ 19 in the future
message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
if(ANDROID OR IOS)
if(ANDROID)
if(AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
elseif(${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
# TODO: support glog for Android api 16 ~ 19 in the future
message(WARNING "Using the unofficial git repository <https://github.com/Xreki/glog.git> instead")
endif()
endif()
set(WITH_GPU OFF CACHE STRING
"Disable GPU when cross-compiling for Android" FORCE)
"Disable GPU when cross-compiling for Android and iOS" FORCE)
set(WITH_AVX OFF CACHE STRING
"Disable AVX when cross-compiling for Android" FORCE)
"Disable AVX when cross-compiling for Android and iOS" FORCE)
set(WITH_PYTHON OFF CACHE STRING
"Disable PYTHON when cross-compiling for Android" FORCE)
"Disable PYTHON when cross-compiling for Android and iOS" FORCE)
set(WITH_RDMA OFF CACHE STRING
"Disable RDMA when cross-compiling for Android" FORCE)
"Disable RDMA when cross-compiling for Android and iOS" FORCE)
set(WITH_MKLDNN OFF CACHE STRING
"Disable MKLDNN when cross-compiling for Android" FORCE)
"Disable MKLDNN when cross-compiling for Android and iOS" FORCE)
set(WITH_MKLML OFF CACHE STRING
"Disable MKLML package when cross-compiling for Android" FORCE)
endif(ANDROID)
"Disable MKLML package when cross-compiling for Android and iOS" FORCE)
endif()
set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
"A path setting third party libraries download & build directories.")
......
cmake/cblas.cmake
浏览文件 @ a2a0d6f8
......@@ -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/cross_compiling/ios.cmake 0 → 100644
浏览文件 @ a2a0d6f8
# 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 @ a2a0d6f8
......@@ -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
......
cmake/external/glog.cmake
浏览文件 @ a2a0d6f8
......@@ -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
......
cmake/external/gtest.cmake
浏览文件 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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
${EXTERNAL_PROJECT_LOG_ARGS}
......
cmake/external/protobuf.cmake
浏览文件 @ a2a0d6f8
......@@ -173,7 +173,8 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
"-DCMAKE_CXX_FLAGS=${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()
......
cmake/external/python.cmake
浏览文件 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -34,15 +34,16 @@ ExternalProject_Add(
GIT_TAG "v1.2.8"
PREFIX ${ZLIB_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
CMAKE_ARGS -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
CMAKE_ARGS -DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR}
CMAKE_ARGS -DBUILD_SHARED_LIBS=OFF
CMAKE_ARGS -DCMAKE_POSITION_INDEPENDENT_CODE=ON
CMAKE_ARGS -DCMAKE_MACOSX_RPATH=ON
CMAKE_ARGS -DCMAKE_BUILD_TYPE=Release
-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}
-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
-DCMAKE_INSTALL_PREFIX=${ZLIB_INSTALL_DIR}
-DBUILD_SHARED_LIBS=OFF
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
-DCMAKE_MACOSX_RPATH=ON
-DCMAKE_BUILD_TYPE=Release
${EXTERNAL_OPTIONAL_ARGS}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ZLIB_INSTALL_DIR}
-DCMAKE_POSITION_INDEPENDENT_CODE:BOOL=ON
-DCMAKE_BUILD_TYPE:STRING=Release
......
cmake/flags.cmake
浏览文件 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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(${CMAKE_SYSTEM_NAME} STREQUAL "iOS")
SET(IOS TRUE)
INCLUDE(cross_compiling/ios)
ENDIF()
ENDIF()
......
cmake/util.cmake
浏览文件 @ a2a0d6f8
......@@ -25,7 +25,9 @@ function(target_circle_link_libraries TARGET_NAME)
endif()
endforeach()
if("${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/block.md 0 → 100644
浏览文件 @ a2a0d6f8
# 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/refactorization.md 0 → 100644
浏览文件 @ a2a0d6f8
# Design Doc: Refactorization Overview
The goal of refactorizaiton include:
1. Make it easy for external contributors to write new elementory computaiton operations.
1. Make the codebase clean and readable.
1. Introduce a new design of computation representation -- a computation graph of operators and variables.
1. The graph representation helps implementing auto-scalable and auto fault recoverable distributed computing.
## Computation Graphs
1. PaddlePaddle represent the computation, training and inference of DL models, by computation graphs.
1. Please dig into [computation graphs](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/graph.md) for a solid example.
1. Users write Python programs to describe the graphs and run it (locally or remotely).
1. A graph is composed of *variabels* and *operators*.
1. The description of graphs must be able to be serialized/deserialized, so it
1. could to be sent to the cloud for distributed execution, and
1. be sent to clients for mobile or enterprise deployment.
1. The Python program do
1. *compilation*: runs a Python program to generate a protobuf message representation of the graph and send it to
1. the C++ library `libpaddle.so` for local execution,
1. the master process of a distributed training job for training, or
1. the server process of a Kubernetes serving job for distributed serving.
1. *execution*: according to the protobuf message, constructs instances of class `Variable` and `OperatorBase`, and run them.
## Description and Realization
At compile time, the Python program generates protobuf message representation of the graph, or the description of the graph.
At runtime, the C++ program realizes the graph and run it.
| | Representation (protobuf messages) | Realization (C++ class objects) |
|---|---|---|
|Data|[VarDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L107)|[Variable](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24)|
|Operation|[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35)|[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64)|
|Block|BlockDesc|Block|
The word *graph* is exchangable with *block* in this document. A graph represent computation steps and local variables as a C++/Java program block, or a pair of { and }.
## Compilation and Execution
1. Run an applicaton Python program to describe the graph. In particular,
1. create VarDesc to represent local/intermediate variables,
1. create operators and set attributes,
1. validate attribute values,
1. inference the type and the shape of variables,
1. plan for memory-reuse for variables,
1. generate backward and optimization part of the Graph.
1. possiblly split the graph for distributed training.
1. The invocation of `train` or `infer` in the application Python program:
1. create a new Scope instance in the [scope hierarchy](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/scope.md) for each run of a block,
1. realize local variables defined in the BlockDesc message in the new scope,
1. a scope is similar to the stack frame in programming languages,
1. create an instance of class `Block`, in which,
1. realize operators in the BlockDesc message,
1. run the Block by calling
1. `Block::Eval(vector<Variable>* targets)` for forward and backward computations, or
1. `Block::Eval(vector<Operator>* targets)` for optimization.
## Intermediate Representation (IR)
```text
Compile Time -> IR -> Runtime
```
### Benefit
- Optimization
```text
Compile Time -> IR -> Optimized IR -> Runtime
```
- Send automatically partitioned IR to different nodes.
- Automatic data parallel
```text
Compile Time
|-> Single GPU IR
|-> [trainer-IR-0, trainer-IR-1, pserver-IR]
|-> Node-0 (runs trainer-IR-0)
|-> Node-1 (runs trainer-IR-1)
|-> Node-2 (runs pserver-IR)
```
- Automatic model parallel (planned for future)
---
# Operator/OpWithKernel/OpKernel
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot)
---
# Operator
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot)
* `Operator` is the fundamental building block as the user interface.
* Operator stores input/output variable name, and attributes.
* The `InferShape` interface is used to infer output variable shapes by its input shapes.
* Use `Run` to compute `input variables` to `output variables`.
---
# OpWithKernel/Kernel
![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot)
* `OpWithKernel` inherits `Operator`.
* `OpWithKernel` contains a Kernel map.
* `OpWithKernel::Run` get device's kernel, and invoke `OpKernel::Compute`.
* `OpKernelKey` is the map key. Only device place now, but may be data type later.
---
# Why separate Kernel and Operator
* Separate GPU and CPU code.
* Make Paddle can run without GPU.
* Make one operator (which is user interface) can contain many implementations.
* Same mul op, different FP16, FP32 Kernel. different MKL, eigen kernel.
---
# Libraries for Kernel development
* `Eigen::Tensor` contains basic math and element-wise functions.
* Note that `Eigen::Tensor` has broadcast implementation.
* Limit number of `tensor.device(dev) = ` in your code.
* `thrust::tranform` and `std::transform`.
* `thrust` has the same API as C++ standard library. Using `transform` can quickly implement a customized elementwise kernel.
* `thrust` has more complex API, like `scan`, `reduce`, `reduce_by_key`.
* Hand-writing `GPUKernel` and `CPU` code
* Do not write `.h`. CPU Kernel should be in `.cc`. CPU kernel should be in `.cu`. (`GCC` cannot compile GPU code.)
---
# Operator Register
## Why register is necessary?
We need a method to build mappings between Op type names and Op classes.
## How to do the register?
Maintain a map, whose key is the type name and value is corresponding Op constructor.
---
# The Registry Map
### `OpInfoMap`
`op_type(string)` -> `OpInfo`
`OpInfo`:
- **`creator`**: The Op constructor.
- **`grad_op_type`**: The type of the gradient Op.
- **`proto`**: The Op's Protobuf, including inputs, outputs and required attributes.
- **`checker`**: Used to check attributes.
---
# Related Concepts
### Op_Maker
It's constructor takes `proto` and `checker`. They are compeleted during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37))
### Register Macros
```cpp
REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type, grad_op_class)
REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class)
```
### `USE` Macros
make sure the registration process is executed and linked.
---
# Register Process
1. Write Op class, as well as its gradient Op class if there is.
2. Write Op maker class. In the constructor, describe its inputs, outputs, and attributes.
3. Invoke macro `REGISTER_OP`. The macro will
1. call maker class to complete `proto` and `checker`
2. with the completed `proto` and `checker`, build a new key-value pair in the `OpInfoMap`
4. Invoke `USE` macro in where the Op is used to make sure it is linked.
---
# Backward Module (1/2)
### Create Backward Operator
- Mapping from forwarding Op to backward Op
![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png)
---
# Backward Module (2/2)
### Build Backward Network
- **Input** graph of forwarding operators
- **Output** graph of backward operators
- **corner case in construction**
- shared variable => insert `Add` operator
- no gradient => insert `fill_zero_grad` operator
- recursive netOp => call `Backward` recursively
- RNN Op => recursively call `Backward` on stepnet
---
# Scope, Variable, Tensor
* `Tensor` is an n-dimension array with type.
* Only dims and data pointers are stored in `Tensor`.
* All operators on `Tensor` is written in `Operator` or global functions.
* variable length Tensor design [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md)
* `Variable` is the inputs and outputs of an operator. Not just `Tensor`.
* step_scopes in RNN is a variable and not a tensor.
* `Scope` is where variables store at.
* map<string/*var name */, Variable>
* `Scope` has a hierarchical structure. The local scope can get variable from its parent scope.
---
# Block (in design)
## the difference with original RNNOp
- as an operator is more intuitive than `RNNOp`,
- offers new interface `Eval(targets)` to deduce the minimal block to `Run`,
- fits the compile-time/ runtime separation design.
- during the compilation, `SymbolTable` stores `VarDesc`s and `OpDesc`s and serialize to a `BlockDesc`
- when graph executes, a Block with `BlockDesc` passed in creates `Op` and `Var` then `Run`
---
# Milestone
- take Paddle/books as the main line, the requirement of the models motivates framework refactoring,
- model migration
- framework development gives **priority support** to model migration, for example,
- the MNIST demo needs a Python interface,
- the RNN models require the framework to support `LoDTensor`.
- determine some timelines,
- heavily-relied Ops need to be migrated first,
- different models can be migrated parallelly.
- improve the framework at the same time
- accept imperfection, concentrated on solving the specific problem at the right price.
---
# Control the migration quality
- compare the performance of migrated models with old ones.
- follow google C style
- build the automatic workflow of generating Python/C++ documentations
- the documentation of layers and ops should be written inside the code
- take the documentation quality into account when doing PR
- preview the documentations, read and improve them from users' perspective
doc/faq/index_cn.rst
浏览文件 @ a2a0d6f8
......@@ -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))
paddle/CMakeLists.txt
浏览文件 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -28,42 +28,38 @@ add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER}
add_dependencies(paddle_capi paddle_proto)
# combine all paddle static libraries together, into libpaddle_capi_whole.a
# user should use PaddleCAPI as -lpaddle_capi_whole
set(capi_whole_library libpaddle_capi_whole.a)
add_custom_target(paddle_capi_whole ALL
COMMAND mkdir -p o_files/capi && cd o_files/capi/ && ar -x $<TARGET_FILE:paddle_capi>
COMMAND mkdir -p o_files/utils && cd o_files/utils/ && ar -x $<TARGET_FILE:paddle_utils>
COMMAND mkdir -p o_files/parameter && cd o_files/parameter/ && ar -x $<TARGET_FILE:paddle_parameter>
COMMAND mkdir -p o_files/math && cd o_files/math/ && ar -x $<TARGET_FILE:paddle_math>
COMMAND mkdir -p o_files/cuda && cd o_files/cuda/ && ar -x $<TARGET_FILE:paddle_cuda>
COMMAND mkdir -p o_files/function && cd o_files/function/ && ar -x $<TARGET_FILE:paddle_function>
COMMAND mkdir -p o_files/gserver && cd o_files/gserver/ && ar -x $<TARGET_FILE:paddle_gserver>
COMMAND mkdir -p o_files/proto && cd o_files/proto/ && ar -x $<TARGET_FILE:paddle_proto>
COMMAND mkdir -p o_files/network && cd o_files/network/ && ar -x $<TARGET_FILE:paddle_network>
COMMAND mkdir -p o_files/pserver && cd o_files/pserver/ && ar -x $<TARGET_FILE:paddle_pserver>
COMMAND ar crs ${capi_whole_library} `find ./o_files -name '*.o'`
COMMAND rm -rf o_files
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
DEPENDS paddle_capi paddle_utils paddle_parameter paddle_math
paddle_cuda paddle_function paddle_gserver
paddle_proto paddle_pserver paddle_network
)
set_target_properties(paddle_capi_whole
PROPERTIES IMPORTED_LOCATION ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library})
set(PADDLE_CAPI_INFER_LIBS
paddle_utils
paddle_parameter
paddle_math
paddle_cuda
paddle_function
paddle_gserver
paddle_proto
paddle_pserver
paddle_network)
cc_library(paddle_capi_whole DEPS paddle_capi ${PADDLE_CAPI_INFER_LIBS})
set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
# TODO: merge mkl into paddle_capi_shared
add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
set_target_properties(paddle_capi_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
link_paddle_exe(paddle_capi_shared)
# No shared library for iOS
if(NOT IOS)
set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
# TODO: merge mkl into paddle_capi_shared
add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
set_target_properties(paddle_capi_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
link_paddle_exe(paddle_capi_shared)
endif()
# install library & headers.
install(FILES ${CAPI_HEADERS} DESTINATION include/paddle)
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/config.h DESTINATION include/paddle)
if(ANDROID)
install(TARGETS paddle_capi_whole paddle_capi_shared
ARCHIVE DESTINATION lib/${ANDROID_ABI}
LIBRARY DESTINATION lib/${ANDROID_ABI})
execute_process(
COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1
OUTPUT_VARIABLE GIT_COMMITS_LIST
......@@ -72,9 +68,6 @@ if(ANDROID)
if(${GIT_COMMITS_LIST_RESULT})
set(GIT_COMMITS_LIST "No commits.")
endif()
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library}
DESTINATION lib/${ANDROID_ABI})
install(TARGETS paddle_capi_shared DESTINATION lib/${ANDROID_ABI})
install(CODE "FILE(WRITE ${CMAKE_INSTALL_PREFIX}/lib/${ANDROID_ABI}/BUILD.txt
\"Compiler:\n\"
\"\\t${CMAKE_C_COMPILER}\\n\"
......@@ -88,8 +81,11 @@ if(ANDROID)
)"
)
else(ANDROID)
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library} DESTINATION lib)
install(TARGETS paddle_capi_shared DESTINATION lib)
install(TARGETS paddle_capi_whole
ARCHIVE DESTINATION lib)
if(NOT IOS)
install(TARGETS paddle_capi_shared DESTINATION lib)
endif()
endif(ANDROID)
# this variable used for unittest
......
paddle/cuda/include/hl_cuda_cudnn.h
浏览文件 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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.md
浏览文件 @ a2a0d6f8
......@@ -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 32x32xO-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;
```
- 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
......@@ -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/operator.cc
浏览文件 @ a2a0d6f8
......@@ -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
......
paddle/framework/operator.h
浏览文件 @ a2a0d6f8
......@@ -366,7 +366,7 @@ struct EigenDeviceConverter<platform::GPUPlace> {
class ExecutionContext : public InferShapeContext {
public:
ExecutionContext(const OperatorBase& op, const Scope& scope,
const platform::DeviceContext* device_context)
const platform::DeviceContext& device_context)
: InferShapeContext(op, scope), device_context_(device_context) {}
template <typename PlaceType,
......@@ -374,9 +374,9 @@ class ExecutionContext : public InferShapeContext {
typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
DeviceType& GetEigenDevice() const;
platform::Place GetPlace() const { return device_context_->GetPlace(); }
platform::Place GetPlace() const { return device_context_.GetPlace(); }
const platform::DeviceContext* device_context() const {
const platform::DeviceContext& device_context() const {
return device_context_;
}
......@@ -401,7 +401,8 @@ class ExecutionContext : public InferShapeContext {
return res;
}
const platform::DeviceContext* device_context_;
private:
const platform::DeviceContext& device_context_;
};
template <>
......@@ -461,7 +462,7 @@ class OperatorWithKernel : public OperatorBase {
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const final {
auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
opKernel->Compute(ExecutionContext(*this, scope, dev_ctx));
}
static std::unordered_map<std::string /* op_type */, OpKernelMap>&
......
paddle/function/neon/NeonDepthwiseConv.cpp
浏览文件 @ a2a0d6f8
......@@ -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/gserver/activations/ActivationFunction.cpp
浏览文件 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
/* 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 @ a2a0d6f8
/* 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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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 @ 8f8ea005
/* 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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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
浏览文件 @ a2a0d6f8
......@@ -285,22 +285,18 @@ void MKLDNNConvLayer::resetWgtBiasValue(
wgt = MKLDNNMatrix::create(weight_->getW(), pd->weights_primitive_desc());
VLOG(MKLDNN_FMTS) << "Weight value format: " << wgt->getFormat();
bias = nullptr;
if (biases_ && biases_->getW()) {
bias = MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc());
}
bias = (biases_ && biases_->getW())
? MKLDNNMatrix::create(biases_->getW(), pd->bias_primitive_desc())
: nullptr;
}
void MKLDNNConvLayer::resetOutValue(
std::shared_ptr<conv_fwd::primitive_desc>& pd, MKLDNNMatrixPtr& out) {
out = MKLDNNMatrix::create(output_.value, pd->dst_primitive_desc());
// change original output value from cpu matrix to mkldnn matrix
output_.value = std::dynamic_pointer_cast<Matrix>(out);
// create reorder if output value has cpu device and pd do not match
cpuOutVal_ = nullptr;
cpuOutVal_ = nullptr;
cvtOutVal_ = nullptr;
if (!outputIsOnlyMKLDNN()) {
const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
......@@ -356,6 +352,7 @@ void MKLDNNConvLayer::resetBwdWgtPD(
void MKLDNNConvLayer::resetBwdDataPD(
std::shared_ptr<conv_bwdData::primitive_desc>& pd) {
pd = nullptr;
if (inputLayers_[0]->getOutput().grad == nullptr) {
return;
}
......@@ -476,6 +473,7 @@ void MKLDNNConvLayer::resetWgtBiasGrad(
<< "primitive desc of weight grad and value should be equal";
VLOG(MKLDNN_FMTS) << "weight grad format: " << wgt->getFormat();
bias = nullptr;
if (biasVal_ == nullptr) {
return;
}
......
paddle/gserver/layers/MKLDNNFcLayer.cpp
浏览文件 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "MKLDNNPoolLayer.h"
#include "paddle/math/MathUtils.h"
#include "paddle/utils/Logging.h"
using namespace mkldnn; // NOLINT
typedef memory::format format;
namespace paddle {
REGISTER_LAYER(mkldnn_pool, MKLDNNPoolLayer);
bool MKLDNNPoolLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
if (!MKLDNNLayer::init(layerMap, parameterMap)) {
return false;
}
/* the size of inputs for pool-layer is 1 */
CHECK_EQ(config_.inputs_size(), 1);
const PoolConfig& conf = config_.inputs(0).pool_conf();
ic_ = conf.channels();
ih_ = conf.img_size_y();
iw_ = conf.img_size();
oc_ = ic_;
oh_ = conf.output_y();
ow_ = conf.output_x();
fh_ = conf.size_y();
fw_ = conf.size_x();
ph_ = conf.padding_y();
pw_ = conf.padding();
sh_ = conf.stride_y();
sw_ = conf.stride();
const std::string& type = conf.pool_type();
if (type == "max-projection") {
poolAlgo_ = algorithm::pooling_max;
} else if (type == "avg-projection") {
// paddle only use exclude_padding
poolAlgo_ = algorithm::pooling_avg_exclude_padding;
} else {
LOG(FATAL) << "unknow pooling type!";
}
return true;
}
void MKLDNNPoolLayer::reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
reshapeInput(bs, ih, iw);
// ic_ and oc can not be changed
CHECK_EQ(inputElemenCnt_ / bs / ih / iw, (size_t)ic)
<< "Input channel can not be changed";
// cal output sizes
// paddle used false caffeMode for pooling
oh = outputSize(ih, fh_, ph_, sh_, false);
ow = outputSize(iw, fw_, pw_, sw_, false);
reshapeOutput(oh, ow);
resizeOutput(bs, oc * oh * ow);
printSizeInfo();
}
void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
resetFwdBuffers(in, out);
resetFwdPD(fwdPD_, in, out);
resetFwdPipeline(pipeline, fwdPD_, in, out);
printValueFormatFlow();
}
void MKLDNNPoolLayer::resetBwd(std::vector<primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) {
std::shared_ptr<pool_bwd::primitive_desc> pd;
resetBwdBuffers(in, out);
resetBwdPD(pd, in, out);
resetBwdPipeline(pipeline, pd, in, out);
printGradFormatFlow();
}
void MKLDNNPoolLayer::updateInputData() {
inVal_->setData(getInputValue(0, CPU_DEVICE)->getData());
}
void MKLDNNPoolLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
resetInValue(in);
resetOutValue(out);
}
void MKLDNNPoolLayer::resetInValue(MKLDNNMatrixPtr& in) {
if (inputIsOnlyMKLDNN()) {
const MatrixPtr& dnnIn = getInputValue(0);
in = std::dynamic_pointer_cast<MKLDNNMatrix>(dnnIn);
CHECK(in) << "Input should be MKLDNNMatrix";
} else {
CHECK_EQ(getPrev(0)->getDeviceId(), CPU_DEVICE) << "Only support CPU yet";
const MatrixPtr& cpuIn = getInputValue(0, CPU_DEVICE);
in = MKLDNNMatrix::create(
cpuIn, {bs_, ic_, ih_, iw_}, format::nchw, engine_);
}
}
void MKLDNNPoolLayer::resetOutValue(MKLDNNMatrixPtr& out) {
CHECK(inVal_) << "Should reset input value first";
memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
out = MKLDNNMatrix::create(
output_.value, outDims, inVal_->getFormat(), engine_);
// create reorder if output value has cpu device and pd do not match
cpuOutVal_ = nullptr;
cvtOutVal_ = nullptr;
if (!outputIsOnlyMKLDNN()) {
const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).value;
cpuOutVal_ = MKLDNNMatrix::create(cpuOut, outDims, format::nchw, engine_);
if (cpuOutVal_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
cvtOutVal_ = MKLDNNMatrix::createReorder(out, cpuOutVal_);
CHECK(cvtOutVal_) << "should not be emptry";
} else {
// CPU output share the same data of MKLDNN output
cpuOut->setData(out->getData());
cpuOutVal_ = out;
}
}
}
void MKLDNNPoolLayer::resetFwdPD(std::shared_ptr<pool_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr in,
MKLDNNMatrixPtr out) {
memory::dims inDims = memory::dims{bs_, ic_, ih_, iw_};
memory::dims outDims = memory::dims{bs_, oc_, oh_, ow_};
memory::dims kernels = memory::dims{fh_, fw_};
memory::dims strides = memory::dims{sh_, sw_};
memory::dims padL = memory::dims{ph_, pw_};
memory::dims padR = getPaddingR();
padding_kind padKind = padding_kind::zero;
prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
: prop_kind::forward_training;
auto fwdDesc = pool_fwd::desc(pk,
poolAlgo_,
in->getMemoryDesc(),
out->getMemoryDesc(),
strides,
kernels,
padL,
padR,
padKind);
pd.reset(new pool_fwd::primitive_desc(fwdDesc, engine_));
// prepare workspace if necessary
workspace_ =
(passType_ != PASS_TEST && poolAlgo_ == algorithm::pooling_max)
? std::make_shared<memory>(memory(pd->workspace_primitive_desc()))
: nullptr;
}
void MKLDNNPoolLayer::resetFwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<pool_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
pipeline.clear();
fwd_ = workspace_
? std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out, *workspace_))
: std::make_shared<pool_fwd>(pool_fwd(*pd, *in, *out));
pipeline.push_back(*fwd_);
if (cvtOutVal_) {
pipeline.push_back(*cvtOutVal_);
}
}
void MKLDNNPoolLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
resetOutGrad(out);
resetInGrad(in);
}
void MKLDNNPoolLayer::resetOutGrad(MKLDNNMatrixPtr& out) {
CHECK(outVal_) << "Should have output value";
out = MKLDNNMatrix::create(output_.grad, outVal_->getPrimitiveDesc());
// create reorder if output value has cpu device and pd do not match
cpuOutGrad_ = nullptr;
cvtOutGrad_ = nullptr;
if (!outputIsOnlyMKLDNN()) {
const MatrixPtr& cpuOut = getOutput(CPU_DEVICE).grad;
cpuOutGrad_ = MKLDNNMatrix::create(
cpuOut, memory::dims{bs_, oc_, oh_, ow_}, format::nchw, engine_);
if (cpuOutGrad_->getPrimitiveDesc() != out->getPrimitiveDesc()) {
cvtOutGrad_ = MKLDNNMatrix::createReorder(cpuOutGrad_, out);
CHECK(cvtOutGrad_) << "should not be emptry";
} else {
// share the same data of CPU output
output_.grad->setData(cpuOut->getData());
out = cpuOutGrad_;
}
}
}
void MKLDNNPoolLayer::resetInGrad(MKLDNNMatrixPtr& in) {
in = nullptr;
const MatrixPtr& inGrad = inputLayers_[0]->getOutput().grad;
if (inGrad == nullptr) {
return;
}
CHECK(inVal_);
in = MKLDNNMatrix::create(inGrad, inVal_->getPrimitiveDesc());
}
void MKLDNNPoolLayer::resetBwdPD(std::shared_ptr<pool_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
memory::dims kernels = memory::dims{fh_, fw_};
memory::dims strides = memory::dims{sh_, sw_};
memory::dims padL = memory::dims{ph_, pw_};
memory::dims padR = getPaddingR();
CHECK(in);
CHECK(out);
auto bwdDesc = pool_bwd::desc(poolAlgo_,
in->getMemoryDesc(),
out->getMemoryDesc(),
strides,
kernels,
padL,
padR,
padding_kind::zero);
pd.reset(new pool_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_));
}
void MKLDNNPoolLayer::resetBwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<pool_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
pipeline.clear();
if (cvtOutGrad_) {
pipeline.push_back(*cvtOutGrad_);
}
bwdData_ =
workspace_
? std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *workspace_, *in))
: std::make_shared<pool_bwd>(pool_bwd(*pd, *out, *in));
pipeline.push_back(*bwdData_);
}
} // namespace paddle
paddle/gserver/layers/MKLDNNPoolLayer.h 0 → 100644
浏览文件 @ a2a0d6f8
/* 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 @ a2a0d6f8
......@@ -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.cpp
浏览文件 @ a2a0d6f8
......@@ -64,15 +64,17 @@ void MKLDNNTester::reset(const TestConfig& dnn,
configs_[i], &(layerMaps_[i]), &(parameters_[i]), &(testLayers_[i]));
}
refLayer_ = testLayers_[REF];
dnnLayer_ = std::dynamic_pointer_cast<MKLDNNLayer>(testLayers_[DNN]);
CHECK(dnnLayer_);
// for comparison with Paddle reference results,
// need manually add cpu device output for test
dnnLayer_->addOutputArgument(CPU_DEVICE);
dnnLayer_ = testLayers_[DNN];
EXPECT_EQ(dataLayers_[DNN].size(), dataLayers_[REF].size());
EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
setInputImgSize();
// for comparison with Paddle reference results,
// need manually add cpu device output for test
MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
if (dnnLayer) {
dnnLayer->addOutputArgument(CPU_DEVICE);
}
}
void MKLDNNTester::setInputImgSize() {
......@@ -122,7 +124,7 @@ void MKLDNNTester::randomTopDiffs() {
void MKLDNNTester::checkForward() {
VLOG(MKLDNN_ALL) << "Check Forward";
printTopDatas();
double delta = compareMatrix(dnnLayer_->getOutput(-1).value,
double delta = compareMatrix(dnnLayer_->getOutput(CPU_DEVICE).value,
refLayer_->getOutputValue());
EXPECT_LE(fabs(delta), eps_);
}
......@@ -155,7 +157,10 @@ void MKLDNNTester::checkBackwardWgts() {
vector<VectorPtr> dnnWgts; // used to temply save mkldnn weights
saveWgt(parameters_[DNN], dnnWgts);
dnnLayer_->convertWeightsToPaddle();
MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast<MKLDNNLayer>(dnnLayer_);
if (dnnLayer) {
dnnLayer->convertWeightsToPaddle();
}
for (size_t i = 0; i < parameters_[DNN].size(); ++i) {
const VectorPtr& dnn = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
const VectorPtr& ref = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
......@@ -322,6 +327,10 @@ void MKLDNNTester::runOnce() {
// and clearTopDatas(REF) should be coverd by ref layers
clearBotDiffs(REF);
clearWgtDiffs(REF);
// it is necessary to clear bottom diffs when only activation is dnn type
if (configs_[DNN].layerConfig.active_type().compare(0, 7, "mkldnn_") == 0) {
clearBotDiffs(DNN);
}
}
void MKLDNNTester::run(const TestConfig& dnn,
......@@ -333,8 +342,19 @@ void MKLDNNTester::run(const TestConfig& dnn,
float epsilon,
bool log,
int level) {
VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: " << dnn.layerConfig.type()
<< " vs " << ref.layerConfig.type();
CHECK(dnn.layerConfig.type().compare(0, 7, "mkldnn_") == 0 ||
dnn.layerConfig.active_type().compare(0, 7, "mkldnn_") == 0)
<< "should be MKLDNN layer or MKLDNN activation";
if (dnn.layerConfig.type() == ref.layerConfig.type()) {
VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: "
<< dnn.layerConfig.active_type() << " vs "
<< ref.layerConfig.active_type();
} else {
VLOG(MKLDNN_TESTS) << "Test MKLDNN functionality: "
<< dnn.layerConfig.type() << " vs "
<< ref.layerConfig.type();
}
ih_ = inputImgH;
iw_ = inputImgW;
iter_ = iter;
......
paddle/gserver/tests/MKLDNNTester.h
浏览文件 @ a2a0d6f8
......@@ -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/gserver/tests/test_MKLDNN.cpp
浏览文件 @ a2a0d6f8
......@@ -17,6 +17,7 @@ limitations under the License. */
#include <vector>
#include "MKLDNNTester.h"
#include "ModelConfig.pb.h"
#include "paddle/gserver/activations/MKLDNNActivation.h"
#include "paddle/math/MathUtils.h"
using namespace paddle; // NOLINT
......@@ -141,6 +142,110 @@ TEST(MKLDNNLayer, ConvLayer) {
testConvLayer({4, 4, 16, 3, 3, 16, 3, 3, 3, 3, 1, 1, 1, 1, 1, 1});
}
struct testPoolDesc {
int bs, ch; // input channel and output channel are the same
int ih, iw;
int oh, ow;
int fh, fw;
int ph, pw;
int sh, sw;
};
void testPoolLayer(const testPoolDesc& pm) {
const std::string compareTypes[] = {"mkldnn_pool", "pool"};
TestConfig cfg;
cfg.layerConfig.set_type(compareTypes[0]);
cfg.layerConfig.set_size(pm.ch * pm.oh * pm.ow);
cfg.inputDefs.push_back(
{INPUT_DATA,
"layer_0",
/* size of input layer= */ size_t(pm.ch * pm.ih * pm.iw),
0});
LayerInputConfig* input = cfg.layerConfig.add_inputs();
PoolConfig* pool = input->mutable_pool_conf();
pool->set_channels(pm.ch);
pool->set_img_size(pm.iw);
pool->set_img_size_y(pm.ih);
pool->set_output_x(pm.ow);
pool->set_output_y(pm.oh);
pool->set_size_x(pm.fw);
pool->set_size_y(pm.fh);
pool->set_padding(pm.pw);
pool->set_padding_y(pm.ph);
pool->set_stride(pm.sw);
pool->set_stride_y(pm.sh);
int oh = outputSize(pm.ih, pm.fh, pm.ph, pm.sh, false);
int ow = outputSize(pm.iw, pm.fw, pm.pw, pm.sw, false);
CHECK_EQ(ow, pm.ow) << "output size check failed";
CHECK_EQ(oh, pm.oh) << "output size check failed";
MKLDNNTester tester;
for (auto type : {"max-projection", "avg-projection"}) {
pool->set_pool_type(type);
TestConfig ref = cfg;
ref.layerConfig.set_type(compareTypes[1]);
for (auto bs : {pm.bs, 1}) {
tester.run(cfg, ref, bs, pm.ih, pm.iw);
}
}
}
TEST(MKLDNNLayer, PoolLayer) {
/* bs, ch, ih, iw, oh, ow, fh, fw, ph, pw, sh, sw*/
testPoolLayer({2, 1, 4, 4, 2, 2, 3, 3, 0, 0, 2, 2});
testPoolLayer({10, 8, 16, 16, 8, 8, 2, 2, 0, 0, 2, 2});
testPoolLayer({4, 2, 5, 5, 3, 3, 3, 3, 1, 1, 2, 2});
testPoolLayer({8, 16, 56, 56, 28, 28, 3, 3, 0, 0, 2, 2});
testPoolLayer({8, 16, 14, 14, 7, 7, 3, 3, 0, 0, 2, 2});
testPoolLayer({4, 16, 7, 7, 1, 1, 7, 7, 0, 0, 1, 1});
testPoolLayer({4, 2, 5, 5, 3, 3, 5, 5, 1, 1, 1, 1});
testPoolLayer({2, 8, 56, 56, 29, 29, 3, 3, 1, 1, 2, 2});
}
struct testActDesc {
int bs, ch;
int ih, iw;
};
static void getAddtoConfig(TestConfig& cfg, const testActDesc& pm) {
cfg.biasSize = 0;
cfg.layerConfig.set_type("addto");
cfg.layerConfig.set_size(pm.ch * pm.ih * pm.iw);
cfg.inputDefs.push_back(
{INPUT_DATA,
"layer_0",
/* size of input layer= */ size_t(pm.ch * pm.ih * pm.iw),
0});
cfg.layerConfig.add_inputs();
}
void testActivation(std::string& type, const testActDesc& pm) {
const std::string compareTypes[] = {type, type.erase(0, 7)};
TestConfig cfg;
getAddtoConfig(cfg, pm);
TestConfig ref = cfg;
cfg.layerConfig.set_active_type(compareTypes[0]);
ref.layerConfig.set_active_type(compareTypes[1]);
MKLDNNTester tester;
for (auto bs : {pm.bs, 1}) {
tester.run(cfg, ref, bs, pm.ih, pm.iw);
}
}
TEST(MKLDNNActivation, Activations) {
auto types = MKLDNNActivation::getAllRegisteredTypes();
// TODO(TJ): mkldnn_softmax not implemented, paddle do not have elu activation
std::set<string> excluded{"mkldnn_softmax", "mkldnn_elu"};
for (auto type : types) {
if (excluded.count(type)) {
continue;
}
testActivation(type, {16, 64, 32, 32});
}
}
// TODO(TJ): add branch test
int main(int argc, char** argv) {
......
paddle/math/BaseMatrix.cu
浏览文件 @ a2a0d6f8
......@@ -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.h
浏览文件 @ a2a0d6f8
......@@ -66,11 +66,12 @@ public:
/**
* Create reorder primitive.
* Create a mkldnn::reorder handle for converting src MKLDNNMatrix to dst.
* checkData: for whether to check the data handle of src and dst is the same.
* if true, means check it and do not want support inplace reorder;
* otherwise do not check data which means the created reorder
* maybe inplace buffer and do not guarantee the logical is correct
* since not all format or conversion support inplace.
* checkData: whether to check the data handle of src and dst.
* if true, it will check the data and do not allow them equal;
* otherwise, it will not check them, then the reorder created
* may have inplace buffer.
* Do not set false, if you can not guarantee the inplace logical
* would work with your reorder.
*/
static std::shared_ptr<mkldnn::reorder> createReorder(
const MKLDNNMatrixPtr& src,
......
paddle/math/MathFunctions.h
浏览文件 @ a2a0d6f8
......@@ -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 @ a2a0d6f8
此差异已折叠。
paddle/math/NEONFunctions.cpp 0 → 100644
浏览文件 @ a2a0d6f8
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
#include "NEONFunctions.h"
#include <arm_neon.h>
namespace paddle {
namespace neon {
// b[i] = a[i] > 0.0f ? a[i] : 0.0f
void relu(const float* a, float* b, int len) {
int offset = len % 16;
float32x4_t ma0, ma1, ma2, ma3;
float32x4_t mb0, mb1, mb2, mb3;
float32x4_t zero = vdupq_n_f32(0.f);
for (int k = 0; k < len / 16; k++, a += 16, b += 16) {
ma0 = vld1q_f32(a);
ma1 = vld1q_f32(a + 4);
ma2 = vld1q_f32(a + 8);
ma3 = vld1q_f32(a + 12);
mb0 = vmaxq_f32(ma0, zero);
mb1 = vmaxq_f32(ma1, zero);
mb2 = vmaxq_f32(ma2, zero);
mb3 = vmaxq_f32(ma3, zero);
vst1q_f32(b, mb0);
vst1q_f32(b + 4, mb1);
vst1q_f32(b + 8, mb2);
vst1q_f32(b + 12, mb3);
}
for (int i = 0; i < offset; i++) {
b[i] = a[i] > 0.0f ? a[i] : 0.0f;
}
}
} // namespace neon
} // namespace paddle
#endif
paddle/gserver/layers/ExpandConvBaseLayer.h → paddle/math/NEONFunctions.h
浏览文件 @ a2a0d6f8
......@@ -14,44 +14,10 @@ limitations under the License. */
#pragma once
#include <vector>
#include "ConvBaseLayer.h"
#include "paddle/math/Matrix.h"
namespace paddle {
namespace neon {
/**
* @brief A subclass of ConvBaseLayer that is a superclass of both
* ExpandConvLayer and ExpandConvTransLayer
*/
class ExpandConvBaseLayer : public ConvBaseLayer {
protected:
/// The transpose of output, which is an auxiliary matrix.
MatrixPtr transOutValue_;
public:
explicit ExpandConvBaseLayer(const LayerConfig& config)
: ConvBaseLayer(config) {}
~ExpandConvBaseLayer() {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
size_t getOutputSize();
/**
* Add shared bias.
*/
void addSharedBias();
/**
* Add unshared bias.
*/
void addUnsharedBias();
void bpropSharedBias(MatrixPtr biases, MatrixPtr v);
void bpropBiases(MatrixPtr v);
};
void relu(const float* a, float* b, int len);
} // namespace neon
} // namespace paddle
paddle/math/tests/test_matrixCompare.cpp
浏览文件 @ a2a0d6f8
......@@ -825,9 +825,8 @@ void testMaxPoolFwdBwd(int numSamples,
int strideW,
int padH,
int padW) {
int outH = 0, outW = 0;
outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
int inWidth = imgSizeH * imgSizeW * channels;
MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
......@@ -927,9 +926,8 @@ void testAvgPoolFwdBwd(int numSamples,
int strideW,
int padH,
int padW) {
int outH = 0, outW = 0;
outH = (imgSizeH - ksizeH + 2 * padH + strideH - 1) / strideH + 1;
outW = (imgSizeW - ksizeW + 2 * padW + strideW - 1) / strideW + 1;
int outH = outputSize(imgSizeH, ksizeH, padH, strideH, true);
int outW = outputSize(imgSizeW, ksizeW, padW, strideW, true);
int inWidth = imgSizeH * imgSizeW * channels;
MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
......
paddle/operators/accuracy_op.cc
浏览文件 @ a2a0d6f8
......@@ -23,10 +23,15 @@ class AccuracyOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Inference"),
"Input of Inference must be initialized.");
PADDLE_ENFORCE_NOT_NULL(
ctx.InputVar("Inference"),
"Input(Inference) of AccuracyOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
"Input of Inference must be initialized.");
"Input(Label) of AccuracyOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(
ctx.OutputVar("Accuracy"),
"Output(Accuracy) of AccuracyOp should not be null.");
auto *inference = ctx.Input<framework::Tensor>("Inference");
auto *label = ctx.Input<framework::Tensor>("Label");
......
paddle/operators/accuracy_op.cu
浏览文件 @ a2a0d6f8
......@@ -12,26 +12,38 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <thrust/execution_policy.h>
#include <thrust/reduce.h>
#include "paddle/operators/accuracy_op.h"
#include "paddle/platform/cuda_helper.h"
namespace paddle {
namespace operators {
using platform::PADDLE_CUDA_NUM_THREADS;
__global__ void AccuracySingleKernel(const int N, const int D, const int top_k,
const int* Xdata, const int* labelData,
float* accuracy) {
int correct = 0;
for (int row = 0; row < N; row++) {
const int label = labelData[row];
for (int col = 0; col < D; col++) {
const int pred = Xdata[row * D + col];
if (pred == label) {
++correct;
template <int BlockSize>
__global__ void AccuracyCudaKernel(const int N, const int D, const int* Xdata,
const int* labeldata, float* accuracy) {
int count = 0;
__shared__ int total[BlockSize];
// support only 1 block
for (int i = threadIdx.x; i < (N); i += BlockSize) {
for (int j = 0; j < D; ++j) {
if (Xdata[i * D + j] == labeldata[i]) {
++count;
break;
}
}
}
*accuracy = static_cast<float>(correct) / static_cast<float>(N);
total[threadIdx.x] = count;
__syncthreads();
// reduce the count with init value 0, and output accuracy.
int result = thrust::reduce(thrust::device, total, total + BlockSize, 0);
if (threadIdx.x == 0) {
*accuracy = static_cast<float>(result) / static_cast<float>(N);
}
}
template <typename T>
......@@ -57,8 +69,8 @@ class AccuracyOpCUDAKernel : public framework::OpKernel {
return;
}
AccuracySingleKernel<<<1, 1>>>(num_samples, infer_width, 1, inference_data,
label_data, accuracy_data);
AccuracyCudaKernel<PADDLE_CUDA_NUM_THREADS><<<1, PADDLE_CUDA_NUM_THREADS>>>(
num_samples, infer_width, inference_data, label_data, accuracy_data);
}
};
......
paddle/operators/add_op.cc
浏览文件 @ a2a0d6f8
......@@ -23,6 +23,13 @@ class AddOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
"Input(X) of AddOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
"Input(Y) of AddOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
"Output(Out) of AddOp should not be null.");
PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
ctx.Input<Tensor>("Y")->dims(),
"Two input of Add Op's dimension must be same.");
......
paddle/operators/concat_op.cc
浏览文件 @ a2a0d6f8
......@@ -25,6 +25,9 @@ class ConcatOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
"Output(Out) of ConcatOp should not be null.");
auto ins = ctx.MultiInput<framework::Tensor>("X");
auto *out = ctx.Output<framework::LoDTensor>("Out");
size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
......
paddle/operators/cond_op.cc
浏览文件 @ a2a0d6f8
......@@ -33,7 +33,8 @@ using DDim = framework::DDim;
void CondOp::CreateScope(const Scope& scope) const {
auto sub_scopes_var = scope.FindVar("SubScopes");
PADDLE_ENFORCE(sub_scopes_var != nullptr, "");
PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
"Output(SubScopes) of CondOp should not be null.");
auto sub_scopes = sub_scopes_var->GetMutable<std::vector<Scope*>>();
auto& sub_scope = scope.NewScope();
sub_scopes->push_back(&sub_scope);
......@@ -41,7 +42,8 @@ void CondOp::CreateScope(const Scope& scope) const {
void CondOp::CreateIndexTensor(const Scope& scope) const {
auto index_tensors_var = scope.FindVar("IndexTensors");
PADDLE_ENFORCE(index_tensors_var != nullptr, "");
PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
"Output(IndexTensors) of CondOp should not be null.");
auto& index_tensors =
*index_tensors_var->GetMutable<std::vector<LoDTensor>>();
index_tensors.push_back(LoDTensor());
......@@ -49,7 +51,8 @@ void CondOp::CreateIndexTensor(const Scope& scope) const {
void CondOp::InferShape(const Scope& scope) const {
auto sub_scopes_var = scope.FindVar("SubScopes");
PADDLE_ENFORCE_NOT_NULL(sub_scopes_var);
PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
"Output(SubScopes) of CondOp should not be null.");
auto& sub_scopes = *sub_scopes_var->GetMutable<std::vector<Scope*>>();
for (int i = 0; i < 2; ++i) {
......@@ -63,7 +66,8 @@ void CondOp::InferShape(const Scope& scope) const {
// branch
CreateIndexTensor(scope);
PADDLE_ENFORCE(!Inputs("Xs").empty(), "Inputs can't be empty");
PADDLE_ENFORCE(!Inputs("Xs").empty(),
"Inputs(Xs) of CondOp can't be empty.");
for (auto& input : Inputs("Xs")) {
// Create a new tensor in sub-scope for input-type tensor
Variable* v = sub_scopes[i]->NewVar(input);
......@@ -108,13 +112,18 @@ void CondOp::InferShape(const Scope& scope) const {
void CondOp::Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const {
auto* sub_scopes_var = scope.FindVar("SubScopes");
PADDLE_ENFORCE_NOT_NULL(sub_scopes_var,
"Output(SubScopes) of CondOp should not be null.");
auto sub_scopes = sub_scopes_var->Get<std::vector<Scope*>>();
auto* index_tensors_var = scope.FindVar("IndexTensors");
PADDLE_ENFORCE_NOT_NULL(index_tensors_var,
"Output(IndexTensors) of CondOp should not be null.");
auto index_tensors = index_tensors_var->Get<std::vector<LoDTensor>>();
std::string cond_name = Input("Cond");
Variable* cond_var = scope.FindVar(cond_name);
PADDLE_ENFORCE_NOT_NULL(cond_var);
PADDLE_ENFORCE_NOT_NULL(cond_var,
"Input(Cond) of CondOp should not be null.");
const LoDTensor* cond = cond_var->GetMutable<LoDTensor>();
// Step 1: get the true/false index at runtime
......@@ -171,6 +180,8 @@ void CondOp::Run(const Scope& scope,
}
// Step 4: merge output results
PADDLE_ENFORCE(!Outputs("Outs").empty(),
"Outputs(Outs) of CondOp can't be empty.");
for (int i = 0; i < 2; ++i) {
// i= 0/i for True and False branches respectively
for (auto& output : Outputs("Outs")) {
......
paddle/operators/cos_sim_op.cc
浏览文件 @ a2a0d6f8
......@@ -26,8 +26,16 @@ class CosSimOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
// notnull check
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
"Input(X) of CosSimOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
"Input(Y) of CosSimOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
"Output(Out) of CosSimOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("XNorm"),
"Output(XNorm) of CosSimOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("YNorm"),
"Output(YNorm) of CosSimOp should not be null.");
// shape check
auto x_dims = ctx.Input<Tensor>("X")->dims();
......
paddle/operators/cross_entropy_op.cc 0 → 100644
浏览文件 @ a2a0d6f8
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/cross_entropy_op.h"
namespace paddle {
namespace operators {
using framework::LoDTensor;
class CrossEntropyOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
"Input(Label) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"), "Output(Y) must not be null.");
auto x = ctx.Input<Tensor>("X");
auto label = ctx.Input<Tensor>("Label");
PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2.");
PADDLE_ENFORCE_EQ(label->dims().size(), 2,
"Input(Label)'s rank must be 2.");
// TODO(xinghai-sun): remove this check after swtiching to bool
PADDLE_ENFORCE(ctx.Attr<int>("soft_label") == 0 ||
ctx.Attr<int>("soft_label") == 1);
PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0],
"The 1st dimension of Input(X) and Input(Label) must "
"be equal.");
if (ctx.Attr<int>("soft_label") == 1) {
PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1],
"If Attr(soft_label) == 1, The 2nd dimension of "
"Input(X) and Input(Label) must be equal.");
} else {
PADDLE_ENFORCE_EQ(label->dims()[1], 1,
"If Attr(soft_label) == 0, The 2nd dimension of "
"Input(Label) must be 1.");
}
ctx.Output<LoDTensor>("Y")->Resize({x->dims()[0], 1});
}
};
class CrossEntropyGradientOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
"Input(Label) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")),
"Input(Y@GRAD) must not be null.");
auto x = ctx.Input<Tensor>("X");
auto label = ctx.Input<Tensor>("Label");
auto dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank must be 2.");
PADDLE_ENFORCE_EQ(dy->dims().size(), 2, "Input(Y@Grad)'s rank must be 2.");
PADDLE_ENFORCE_EQ(label->dims().size(), 2,
"Input(Label)'s rank must be 2.");
// TODO(xinghai-sun): remove this check after swtiching to bool
PADDLE_ENFORCE(ctx.Attr<int>("soft_label") == 0 ||
ctx.Attr<int>("soft_label") == 1);
PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0],
"The 1st dimension of Input(X) and Input(Label) must "
"be equal.");
PADDLE_ENFORCE_EQ(x->dims()[0], dy->dims()[0],
"The 1st dimension of Input(X) and Input(Y@Grad) must "
"be equal.");
PADDLE_ENFORCE_EQ(dy->dims()[1], 1,
"The 2nd dimension of Input(Y@Grad) must be 1.");
if (ctx.Attr<int>("soft_label") == 1) {
PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1],
"If Attr(soft_label) == 1, The 2nd dimension of "
"Input(X) and Input(Label) must be equal.");
} else {
PADDLE_ENFORCE_EQ(label->dims()[1], 1,
"If Attr(soft_label) == 0, The 2nd dimension of "
"Input(Label) must be 1.");
}
auto dx = ctx.Output<LoDTensor>(framework::GradVarName("X"));
dx->Resize(x->dims());
}
};
class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
public:
CrossEntropyOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of CrossEntropyOp");
AddInput("Label", "The second input of CrossEntropyOp");
AddOutput("Y", "The output of CrossEntropyOp");
AddAttr<int>("soft_label", "Is soft label. Default zero.").SetDefault(0);
AddComment(R"DOC(
CrossEntropy Operator.
It supports both standard cross-entropy and soft-label cross-entropy loss
computation.
1) One-hot cross-entropy:
soft_label = 0, Label[i, 0] indicates the class index for sample i:
Y[i] = -log(X[i, Label[i]])
2) Soft-label cross-entropy:
soft_label = 1, Label[i, j] indicates the soft label of class j
for sample i:
Y[i] = \sum_j{-Label[i, j] * log(X[i, j])}
Please make sure that in this case the summuation of each row of Label
equals one.
3) One-hot cross-entropy with vecterized Input(Label):
As a special case of 2), when each row of Input(Label) has only one
non-zero element (equals 1), soft-label cross-entropy degenerates to a
one-hot cross-entropy with one-hot label representation.
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker,
cross_entropy_grad, ops::CrossEntropyGradientOp);
REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<float>);
REGISTER_OP_CPU_KERNEL(cross_entropy_grad,
ops::CrossEntropyGradientOpKernel<float>);
paddle/operators/onehot_cross_entropy_op.cu → paddle/operators/cross_entropy_op.cu
浏览文件 @ a2a0d6f8
......@@ -13,14 +13,13 @@
limitations under the License. */
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/utils.h"
#include "paddle/operators/cross_entropy_op.h"
#include "paddle/platform/assert.h"
#include "paddle/platform/hostdevice.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T>
__global__ void CrossEntropyKernel(T* Y, const T* X, const int* label,
const int N, const int D) {
......@@ -29,7 +28,20 @@ __global__ void CrossEntropyKernel(T* Y, const T* X, const int* label,
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
i += blockDim.x * gridDim.x) {
PADDLE_ASSERT(label[i] >= 0 && label[i] < D);
Y[i] = -math::tolerable_value(log(X[i * D + label[i]]));
Y[i] = -tolerable_value(log(X[i * D + label[i]]));
}
}
template <typename T>
__global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label,
const int N, const int D) {
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
i += blockDim.x * gridDim.x) {
T sum = static_cast<T>(0);
for (int j = 0; j < D; j++) {
sum += label[i * D + j] * tolerable_value(log(X[i * D + j]));
}
Y[i] = -sum;
}
}
......@@ -56,58 +68,84 @@ __global__ void CrossEntropyGradientKernel(T* dX, const T* dY, const T* X,
}
template <typename T>
class OnehotCrossEntropyOpCUDAKernel : public framework::OpKernel {
__global__ void SoftCrossEntropyGradientKernel(T* dX, const T* dY, const T* X,
const T* label, const int N,
const int D) {
// TOOD(qingqing): optimize for this kernel
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N;
i += blockDim.x * gridDim.x) {
for (int j = 0; j < D; ++j) {
int idx = i * D + j;
dX[idx] = -label[idx] * dY[i] / X[idx];
}
}
}
template <typename T>
class CrossEntropyOpCUDAKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"This kernel only runs on GPU device.");
"It must use GPUPlace.");
auto X = ctx.Input<Tensor>("X");
const T* Xdata = X->data<T>();
const int* label_data = ctx.Input<Tensor>("label")->data<int>();
auto Y = ctx.Output<Tensor>("Y");
Y->mutable_data<T>(ctx.GetPlace());
T* Ydata = Y->data<T>();
auto x = ctx.Input<Tensor>("X");
auto y = ctx.Output<Tensor>("Y");
auto label = ctx.Input<Tensor>("Label");
int N = X->dims()[0];
int D = X->dims()[1];
int block = 512;
int grid = (N + block - 1) / block;
auto* x_data = x->data<T>();
y->mutable_data<T>(ctx.GetPlace());
auto* y_data = y->data<T>();
int n = x->dims()[0];
int d = x->dims()[1];
int block = 512;
int grid = (n + block - 1) / block;
// TODO(qingqing) launch kernel on specified stream
// base on ExecutionContext.
CrossEntropyKernel<T><<<grid, block>>>(Ydata, Xdata, label_data, N, D);
if (ctx.Attr<int>("soft_label") == 1) {
auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
SoftCrossEntropyKernel<T><<<grid, block>>>(y_data, x_data, label_data, n,
d);
} else {
auto* label_data = ctx.Input<Tensor>("Label")->data<int>();
CrossEntropyKernel<T><<<grid, block>>>(y_data, x_data, label_data, n, d);
}
}
};
template <typename T>
class OnehotCrossEntropyGradientOpCUDAKernel : public framework::OpKernel {
class CrossEntropyGradientOpCUDAKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
"It must use GPUPlace.");
auto X = ctx.Input<Tensor>("X");
auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
auto dY = ctx.Input<Tensor>(framework::GradVarName("Y"));
auto label = ctx.Input<Tensor>("label");
auto x = ctx.Input<Tensor>("X");
auto dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
auto label = ctx.Input<Tensor>("Label");
auto* dXdata = dX->template mutable_data<T>(ctx.GetPlace());
auto* dYdata = dY->template data<T>();
auto* Xdata = X->template data<T>();
auto* label_data = label->data<int>();
auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());
auto* dy_data = dy->data<T>();
auto* x_data = x->data<T>();
int N = X->dims()[0];
int D = X->dims()[1];
int n = x->dims()[0];
int d = x->dims()[1];
int block = 512;
int grid = (N * D + block - 1) / block;
zero<T><<<grid, block>>>(dXdata, N * D);
grid = (N + block - 1) / block;
int grid = (n * d + block - 1) / block;
zero<T><<<grid, block>>>(dx_data, n * d);
grid = (n + block - 1) / block;
// TODO(qingqing): launch kernel on specified stream
// base on ExecutionContext.
CrossEntropyGradientKernel<T><<<grid, block>>>(dXdata, dYdata, Xdata,
label_data, N, D);
if (ctx.Attr<int>("soft_label") == 1) {
auto* label_data = label->data<T>();
SoftCrossEntropyGradientKernel<T><<<grid, block>>>(
dx_data, dy_data, x_data, label_data, n, d);
} else {
auto* label_data = label->data<int>();
CrossEntropyGradientKernel<T><<<grid, block>>>(dx_data, dy_data, x_data,
label_data, n, d);
}
}
};
......@@ -115,7 +153,6 @@ class OnehotCrossEntropyGradientOpCUDAKernel : public framework::OpKernel {
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
ops::OnehotCrossEntropyOpCUDAKernel<float>);
REGISTER_OP_GPU_KERNEL(onehot_cross_entropy_grad,
ops::OnehotCrossEntropyGradientOpCUDAKernel<float>);
REGISTER_OP_GPU_KERNEL(cross_entropy, ops::CrossEntropyOpCUDAKernel<float>);
REGISTER_OP_GPU_KERNEL(cross_entropy_grad,
ops::CrossEntropyGradientOpCUDAKernel<float>);
paddle/operators/onehot_cross_entropy_op.h → paddle/operators/cross_entropy_op.h
浏览文件 @ a2a0d6f8
......@@ -14,7 +14,7 @@ limitations under the License. */
#pragma once
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/utils.h"
#include "paddle/platform/hostdevice.h"
namespace paddle {
namespace operators {
......@@ -22,56 +22,93 @@ namespace operators {
using Tensor = framework::Tensor;
template <typename T>
class OnehotCrossEntropyOpKernel : public framework::OpKernel {
HOSTDEVICE T tolerable_value(const T x) {
PADDLE_ASSERT(std::is_floating_point<T>::value);
const T kApproInf = 1e20;
if (x == INFINITY) {
return kApproInf;
}
if (x == -INFINITY) {
return -kApproInf;
}
return x;
}
template <typename T>
class CrossEntropyOpKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
auto X = ctx.Input<Tensor>("X");
const T* Xdata = X->data<T>();
const int* label_data = ctx.Input<Tensor>("label")->data<int>();
auto Y = ctx.Output<Tensor>("Y");
Y->mutable_data<T>(ctx.GetPlace());
T* Ydata = Y->data<T>();
const int batch_size = X->dims()[0];
const int class_num = X->dims()[1];
for (int i = 0; i < batch_size; ++i) {
int index = i * class_num + label_data[i];
Ydata[i] = -math::tolerable_value(std::log(Xdata[index]));
auto x = ctx.Input<Tensor>("X");
auto y = ctx.Output<Tensor>("Y");
auto* x_data = x->data<T>();
y->mutable_data<T>(ctx.GetPlace());
auto* y_data = y->data<T>();
int batch_size = x->dims()[0];
int class_num = x->dims()[1];
if (ctx.Attr<int>("soft_label") == 1) {
auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
int index = 0;
for (int i = 0; i < batch_size; ++i) {
T sum = static_cast<T>(0);
for (int j = 0; j < class_num; ++j) {
sum += label_data[index] * tolerable_value(std::log(x_data[index]));
y_data[i] = -sum;
index++;
}
}
} else {
auto* label_data = ctx.Input<Tensor>("Label")->data<int>();
for (int i = 0; i < batch_size; ++i) {
int index = i * class_num + label_data[i];
y_data[i] = -tolerable_value(std::log(x_data[index]));
}
}
}
};
template <typename T>
class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel {
class CrossEntropyGradientOpKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
auto X = ctx.Input<Tensor>("X");
auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
auto dY = ctx.Input<Tensor>(framework::GradVarName("Y"));
auto label = ctx.Input<Tensor>("label");
auto x = ctx.Input<Tensor>("X");
auto dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
auto label = ctx.Input<Tensor>("Label");
auto* dXdata = dX->template mutable_data<T>(ctx.GetPlace());
auto* dYdata = dY->template data<T>();
auto* Xdata = X->template data<T>();
auto* label_data = label->data<int>();
auto* dx_data = dx->mutable_data<T>(ctx.GetPlace());
auto* dy_data = dy->data<T>();
auto* x_data = x->data<T>();
const int batch_size = X->dims()[0];
const int class_num = X->dims()[1];
int batch_size = x->dims()[0];
int class_num = x->dims()[1];
// TODO(qingqing): make zero setting an common function.
memset(dXdata, 0, sizeof(T) * batch_size * class_num);
for (int i = 0; i < batch_size; ++i) {
int index = i * class_num + label_data[i];
dXdata[index] = -math::tolerable_value(dYdata[i] / Xdata[index]);
if (ctx.Attr<int>("soft_label") == 1) {
auto* label_data = ctx.Input<Tensor>("Label")->data<T>();
int index = 0;
for (int i = 0; i < batch_size; ++i) {
for (int j = 0; j < class_num; ++j) {
dx_data[index] = -label_data[index] * dy_data[i] / x_data[index];
index++;
}
}
} else {
auto* label_data = label->data<int>();
memset(dx_data, 0, sizeof(T) * batch_size * class_num);
for (int i = 0; i < batch_size; ++i) {
PADDLE_ASSERT(label_data[i] >= 0 || label_data[i] < class_num);
int index = i * class_num + label_data[i];
dx_data[index] = -dy_data[i] / x_data[index];
}
}
}
};
......
paddle/operators/dropout_op.cc 0 → 100644
浏览文件 @ a2a0d6f8
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/dropout_op.h"
namespace paddle {
namespace operators {
using framework::Tensor;
using framework::LoDTensor;
class DropoutOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
PADDLE_ENFORCE_GE(ctx.Attr<float>("dropout_prob"), 0);
PADDLE_ENFORCE_LE(ctx.Attr<float>("dropout_prob"), 1);
// TODO(xinghai-sun): remove this check after swtiching to bool
PADDLE_ENFORCE(ctx.Attr<int>("is_training") == 0 ||
ctx.Attr<int>("is_training") == 1);
auto dims = ctx.Input<Tensor>("X")->dims();
ctx.Output<LoDTensor>("Out")->Resize(dims);
if (ctx.Attr<int>("is_training") == 1) {
ctx.Output<LoDTensor>("Mask")->Resize(dims);
}
}
};
template <typename AttrType>
class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
public:
DropoutOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddAttr<AttrType>("dropout_prob", "Probability of setting units to zero.")
.SetDefault(.5f);
// TODO(xinghai-sun): use bool for is_training after bool is supported.
AddAttr<int>("is_training", "Whether in training phase.").SetDefault(1);
AddAttr<int>("seed", "Dropout random seed.").SetDefault(0);
AddInput("X", "The input of dropout op.");
AddOutput("Out", "The output of dropout op.");
AddOutput("Mask", "The random sampled dropout mask.").AsIntermediate();
AddComment(R"DOC(
Dropout Operator.
"Dropout" refers to randomly dropping out units in a nerual network. It is a
regularization technique for reducing overfitting by preventing neuron
co-adaption during training. The dropout operator randomly set (according to
the given dropout probability) the outputs of some units to zero, while others
being set to their inputs.
)DOC");
}
};
template <typename AttrType>
class DropoutOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_EQ(ctx.Attr<int>("is_training"), 1,
"GradOp is only callable when is_training is true");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Mask"), "Mask must not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
"Input(Out@GRAD) must not be null.");
PADDLE_ENFORCE_GE(ctx.Attr<AttrType>("dropout_prob"), 0);
PADDLE_ENFORCE_LE(ctx.Attr<AttrType>("dropout_prob"), 1);
// TODO(xinghai-sun): remove this check after swtiching to bool
PADDLE_ENFORCE(ctx.Attr<int>("is_training") == 0 ||
ctx.Attr<int>("is_training") == 1);
auto x_dims = ctx.Input<Tensor>("X")->dims();
auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
PADDLE_ENFORCE_EQ(x_dims, out_dims,
"Dimensions of Input(X) and Out@Grad must be the same.");
auto mask_dims = ctx.Input<Tensor>("Mask")->dims();
PADDLE_ENFORCE_EQ(x_dims, mask_dims,
"Dimensions of Input(X) and Mask must be the same.");
auto *x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
x_grad->Resize(x_dims);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(dropout, ops::DropoutOp, ops::DropoutOpMaker<float>, dropout_grad,
ops::DropoutOpGrad<float>);
REGISTER_OP_CPU_KERNEL(
dropout, ops::CPUDropoutKernel<paddle::platform::CPUPlace, float, float>);
REGISTER_OP_CPU_KERNEL(
dropout_grad, ops::DropoutGradKernel<paddle::platform::CPUPlace, float>);
paddle/operators/dropout_op.cu 0 → 100644
浏览文件 @ a2a0d6f8
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#define EIGEN_USE_GPU
#include <thrust/device_ptr.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/random.h>
#include <thrust/transform.h>
#include "paddle/operators/dropout_op.h"
namespace paddle {
namespace operators {
template <typename T, typename AttrType>
struct MaskGenerator {
AttrType dropout_prob;
int seed;
__host__ __device__ MaskGenerator(AttrType dropout_prob, int seed)
: dropout_prob(dropout_prob), seed(seed) {}
__host__ __device__ T operator()(const unsigned int n) const {
thrust::minstd_rand rng;
rng.seed(seed);
thrust::uniform_real_distribution<AttrType> dist(0, 1);
rng.discard(n);
if (dist(rng) < dropout_prob) {
return static_cast<T>(0);
} else {
return static_cast<T>(1);
}
}
};
// It seems that Eigen::Tensor::setRandom in GPU will SEGFAULT.
// Use std::random and thrust::random(thrust is a std library in CUDA) to
// implement uniform random.
template <typename Place, typename T, typename AttrType>
class GPUDropoutKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* x = context.Input<Tensor>("X");
auto* y = context.Output<Tensor>("Out");
y->mutable_data<T>(context.GetPlace());
AttrType dropout_prob = context.Attr<AttrType>("dropout_prob");
auto X = EigenMatrix<T>::Reshape(*x, 1);
auto Y = EigenMatrix<T>::Reshape(*y, 1);
auto place = context.GetEigenDevice<Place>();
if (context.Attr<int>("is_training") == 1) {
auto* mask = context.Output<Tensor>("Mask");
auto* mask_data = mask->mutable_data<T>(context.GetPlace());
int size = framework::product(mask->dims());
int seed = context.Attr<int>("seed");
thrust::counting_iterator<unsigned int> index_sequence_begin(0);
thrust::transform(index_sequence_begin, index_sequence_begin + size,
thrust::device_ptr<T>(mask_data),
MaskGenerator<T, AttrType>(dropout_prob, seed));
auto M = EigenMatrix<T>::Reshape(*mask, 1);
Y.device(place) = X * M;
} else {
Y.device(place) = X * dropout_prob;
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
dropout, ops::GPUDropoutKernel<paddle::platform::GPUPlace, float, float>);
REGISTER_OP_GPU_KERNEL(
dropout_grad, ops::DropoutGradKernel<paddle::platform::GPUPlace, float>);
paddle/operators/dropout_op.h 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/elementwise_mul_op.cc
浏览文件 @ a2a0d6f8
......@@ -25,8 +25,14 @@ class ElementWiseMulOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
"Input(X) of ElementWiseMulOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
"Input(Y) of ElementWiseMulOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(
ctx.OutputVar("Out"),
"Output(Out) of ElementWiseMulOp should not be null.");
auto x_dim = ctx.Input<Tensor>("X")->dims();
auto y_dim = ctx.Input<Tensor>("Y")->dims();
PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
......
paddle/operators/fc_op.cc 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/fill_zeros_like_op.cc
浏览文件 @ a2a0d6f8
......@@ -23,6 +23,13 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(
ctx.InputVar("Src"),
"Input(Src) of FillZerosLikeOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(
ctx.OutputVar("Dst"),
"Output(Dst) of FillZerosLikeOp should not be null.");
ctx.Output<framework::LoDTensor>("Dst")->Resize(
ctx.Input<framework::Tensor>("Src")->dims());
}
......
paddle/operators/gather_op.cc
浏览文件 @ a2a0d6f8
......@@ -24,6 +24,13 @@ class GatherOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
"Input(X) of GatherOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
"Input(Index) of GatherOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
"Output(Out) of GatherOp should not be null.");
int batch_size = ctx.Input<Tensor>("Index")->dims()[0];
PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());
......
paddle/operators/identity_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/lookup_table_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/mean_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/minus_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/mul_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/mul_op.h
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/onehot_cross_entropy_op.cc 已删除 100644 → 0
浏览文件 @ 8f8ea005
此差异已折叠。
paddle/operators/pad_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/prelu_op.cc 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/prelu_op.cu 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/prelu_op.h 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/reshape_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/rowwise_add_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/scale_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/scatter_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/sgd_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/sigmoid_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/softmax_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/split_op.cc 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/split_op.h 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/sum_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/operators/top_k_op.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/platform/CMakeLists.txt
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/platform/cuda_helper.h
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/platform/device_context.cc
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/platform/device_context.h
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/platform/transform.h
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/platform/transform_test.cu
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/pserver/CMakeLists.txt
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/pybind/CMakeLists.txt
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/scripts/travis/build_ios.sh 0 → 100755
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/trainer/CMakeLists.txt
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/utils/Excepts.h
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/utils/arch/linux/Locks.cpp
浏览文件 @ a2a0d6f8
此差异已折叠。
paddle/utils/arch/osx/Excepts.cpp
浏览文件 @ a2a0d6f8
此差异已折叠。
python/paddle/v2/framework/tests/test_dropout_op.py 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
python/paddle/v2/framework/tests/test_fc_op.py 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
python/paddle/v2/framework/tests/test_identity_op.py 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
python/paddle/v2/framework/tests/test_prelu_op.py 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
python/paddle/v2/framework/tests/test_seq_pool.py 0 → 100644
浏览文件 @ a2a0d6f8
此差异已折叠。
python/paddle/v2/framework/tests/test_split_op.py 0 → 100644
浏览文件 @ a2a0d6f8
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