Skip to content

P
Paddle

PaddlePaddle / Paddle
接近 2 年 前同步成功

通知 2323
Star 20933
Fork 5424
P
Paddle
提交 ee3aae56 编写于 作者: D dzhwinter
浏览文件
操作

merge develop branch. test=develop

上级 d6d3e6af b62b756b
展开全部 隐藏空白更改
内联 并排
Showing with 10559 addition and 4141 deletion
CMakeLists.txt
浏览文件 @ ee3aae56
......@@ -33,9 +33,7 @@ if(WIN32)
set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
endif(WIN32)
if(NOT CMAKE_CROSSCOMPILING)
find_package(CUDA QUIET)
endif(NOT CMAKE_CROSSCOMPILING)
find_package(CUDA QUIET)
find_package(Git REQUIRED)
find_package(Threads REQUIRED)
......@@ -49,7 +47,6 @@ option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FO
option(WITH_NGRAPH "Compile PaddlePaddle with nGraph support." OFF)
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)
option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON)
option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF)
option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF)
......@@ -60,11 +57,9 @@ option(WITH_DOC "Compile PaddlePaddle with documentation" OFF)
option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF)
option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF)
option(ON_TRAVIS "Exclude special unit test on Travis CI" OFF)
option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF)
option(WITH_FLUID_ONLY "Compile PaddlePaddle fluid only" OFF)
option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF)
option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF)
option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(WITH_PSLIB "Compile with pslib support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
......@@ -96,37 +91,6 @@ if(NOT CMAKE_BUILD_TYPE)
FORCE)
endif()
if(ANDROID OR IOS)
if(ANDROID)
if(${CMAKE_SYSTEM_VERSION} VERSION_LESS "16")
message(FATAL_ERROR "Unsupport standalone toolchains with Android API level lower than 16")
endif()
endif()
set(WITH_GPU OFF CACHE STRING
"Disable GPU when cross-compiling for Android and iOS" FORCE)
set(WITH_AVX OFF CACHE STRING
"Disable AVX when cross-compiling for Android and iOS" FORCE)
set(WITH_PYTHON OFF CACHE STRING
"Disable PYTHON when cross-compiling for Android and iOS" FORCE)
set(WITH_RDMA OFF CACHE STRING
"Disable RDMA when cross-compiling for Android and iOS" FORCE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL when cross-compiling for Android and iOS" FORCE)
set(WITH_NGRAPH OFF CACHE STRING
"Disable nGraph when cross-compiling for Android and iOS" FORCE)
set(WITH_GOLANG OFF CACHE STRING
"Disable golang when cross-compiling for Android and iOS" FORCE)
# Compile PaddlePaddle mobile inference library
if (NOT WITH_C_API)
set(WITH_C_API ON CACHE STRING
"Always compile the C_API when cross-compiling for Android and iOS" FORCE)
endif()
set(MOBILE_INFERENCE ON)
add_definitions(-DPADDLE_MOBILE_INFERENCE)
endif()
if (APPLE)
set(WITH_MKL OFF CACHE STRING
"Disable MKL for building on mac" FORCE)
......@@ -135,8 +99,6 @@ endif()
if (WIN32)
set(WITH_DISTRIBUTE OFF CACHE STRING
"Disable DISTRIBUTE when compiling for Windows" FORCE)
set(WITH_C_API OFF CACHE STRING
"Disable C_API when compiling for Windows" FORCE)
set(WITH_FLUID_ONLY ON CACHE STRING
"Enable FLUID_ONLY when compiling for Windows" FORCE)
endif()
......@@ -150,21 +112,7 @@ set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING
set(FLUID_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_inference_install_dir" CACHE STRING
"A path setting fluid inference shared and static libraries")
if (WITH_C_API AND WITH_PYTHON)
message(WARNING "It is suggest not embedded a python interpreter in Paddle "
"when using C-API. It will give an unpredictable behavior when using a "
"different Python interpreter from compiling.")
endif()
if (WITH_C_API)
set(WITH_FLUID_ONLY OFF CACHE STRING "Disable install fluid when compile the C_API" FORCE)
endif()
if(MOBILE_INFERENCE)
set(THIRD_PARTY_BUILD_TYPE MinSizeRel)
else()
set(THIRD_PARTY_BUILD_TYPE Release)
endif()
set(THIRD_PARTY_BUILD_TYPE Release)
set(WITH_MKLML ${WITH_MKL})
if (NOT DEFINED WITH_MKLDNN)
......@@ -193,7 +141,6 @@ include(external/python) # download, build, install python
include(external/openblas) # download, build, install openblas
include(external/mkldnn) # download, build, install mkldnn
include(external/ngraph) # download, build, install nGraph
include(external/swig) # download, build, install swig
include(external/boost) # download boost
include(external/any) # download libn::any
include(external/eigen) # download eigen3
......@@ -312,11 +259,6 @@ if(WITH_MKLDNN)
list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
endif()
if(USE_NNPACK)
include(external/nnpack)
list(APPEND EXTERNAL_LIBS ${NNPACK_LIBS})
endif(USE_NNPACK)
set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
......
Dockerfile.android 已删除 100644 → 0
浏览文件 @ d6d3e6af
FROM ubuntu:16.04
MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
ARG UBUNTU_MIRROR
RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ubuntu.com/ubuntu#${UBUNTU_MIRROR}#g' /etc/apt/sources.list; fi'
# ENV variables
ARG ANDROID_ABI
ARG ANDROID_API
ENV ANDROID_ABI=${ANDROID_ABI:-"armeabi-v7a"}
ENV ANDROID_API=${ANDROID_API:-21}
ENV HOME=/root \
ANDROID_NDK_HOME=/opt/android-ndk-linux \
ANDROID_TOOLCHAINS_DIR=/opt/toolchains
RUN apt-get update && \
apt-get install -y \
git python-dev python-pip python-numpy \
wget curl tar unzip gcc g++ locales clang-format-3.8 swig cmake && \
apt-get clean -y
# git credential to skip password typing
RUN git config --global credential.helper store
# Fix locales to en_US.UTF-8
RUN localedef -i en_US -f UTF-8 en_US.UTF-8
RUN pip install --upgrade pip==9.0.3 && \
pip install -U 'protobuf==3.1.0' && \
pip install -U wheel sphinx && \
pip install pre-commit
# Android NDK
RUN mkdir -p ${ANDROID_TOOLCHAINS_DIR} && \
mkdir -p /opt/android-ndk-tmp && \
cd /opt/android-ndk-tmp && \
wget -q https://dl.google.com/android/repository/android-ndk-r14b-linux-x86_64.zip && \
unzip -q android-ndk-r14b-linux-x86_64.zip && \
mv android-ndk-r14b ${ANDROID_NDK_HOME} && \
rm -rf /opt/android-ndk-tmp
cmake/cblas.cmake
浏览文件 @ ee3aae56
......@@ -64,24 +64,18 @@ endif()
## Then find the reference-cblas. www.netlib.org/blas/
set(REFERENCE_CBLAS_ROOT $ENV{REFERENCE_CBLAS_ROOT} CACHE PATH
"Folder contains reference-cblas")
if(NOT CMAKE_CROSSCOMPILING)
set(REFERENCE_CBLAS_INCLUDE_SEARCH_PATHS
${REFERENCE_CBLAS_ROOT}/include
/usr/include
/usr/include/cblas
)
set(REFERENCE_CBLAS_INCLUDE_SEARCH_PATHS
${REFERENCE_CBLAS_ROOT}/include
/usr/include
/usr/include/cblas
)
set(REFERENCE_CBLAS_LIB_SEARCH_PATHS
${REFERENCE_CBLAS_ROOT}/lib
/usr/lib
/usr/lib/blas/reference/
/usr/lib/reference/
)
else()
# Disable the finding of reference cblas under host's system path
set(REFERENCE_CBLAS_INCLUDE_SEARCH_PATHS ${REFERENCE_CBLAS_ROOT}/include)
set(REFERENCE_CBLAS_LIB_SEARCH_PATHS ${REFERENCE_CBLAS_ROOT}/lib)
endif()
set(REFERENCE_CBLAS_LIB_SEARCH_PATHS
${REFERENCE_CBLAS_ROOT}/lib
/usr/lib
/usr/lib/blas/reference/
/usr/lib/reference/
)
if(WITH_SYSTEM_BLAS)
find_path(REFERENCE_CBLAS_INCLUDE_DIR NAMES cblas.h PATHS
......@@ -98,10 +92,3 @@ if(WITH_SYSTEM_BLAS)
message(STATUS "Found reference-cblas (include: ${CBLAS_INC_DIR}, library: ${CBLAS_LIBRARIES})")
endif()
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/configure.cmake
浏览文件 @ ee3aae56
......@@ -49,12 +49,10 @@ if(NOT WITH_PROFILER)
add_definitions(-DPADDLE_DISABLE_PROFILER)
endif(NOT WITH_PROFILER)
if(NOT CMAKE_CROSSCOMPILING)
if(WITH_AVX AND AVX_FOUND)
set(SIMD_FLAG ${AVX_FLAG})
elseif(SSE3_FOUND)
set(SIMD_FLAG ${SSE3_FLAG})
endif()
if(WITH_AVX AND AVX_FOUND)
set(SIMD_FLAG ${AVX_FLAG})
elseif(SSE3_FOUND)
set(SIMD_FLAG ${SSE3_FLAG})
endif()
if(WIN32)
......
cmake/cross_compiling/android.cmake 已删除 100644 → 0
浏览文件 @ d6d3e6af
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# 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 Android, and the
# configuration refers to the open-source resposity:
# https://github.com/taka-no-me/android-cmake
# Most of the variables are compatible with that used in
# https://developer.android.com/ndk/guides/cmake.html
# The supported variables are listed belows:
#
# ANDROID_STANDALONE_TOOLCHAIN
# ANDROID_TOOLCHAIN
# ANDROID_ABI
# ANDROID_NATIVE_API_LEVEL
# ANDROID_ARM_MODE
# ANDROID_ARM_NEON
#
# For CMake >= 3.7.0, all the settings will be delivered to CMake system
# variables to let CMake do the cross-compiling configurations itself.
# More detail of cross-compiling settings
# https://cmake.org/cmake/help/v3.7/manual/cmake-toolchains.7.html
IF(NOT ANDROID)
return()
ENDIF()
# check the exist of android standalone toolchain
IF(NOT DEFINED ANDROID_STANDALONE_TOOLCHAIN)
SET(ANDROID_STANDALONE_TOOLCHAIN $ENV{ANDROID_STANDALONE_TOOLCHAIN}
CACHE PATH "Folder holds the standalone toolchain of Android NDK")
ENDIF()
IF(NOT ANDROID_STANDALONE_TOOLCHAIN)
MESSAGE(WARNING "It is recommended to set ANDROID_STANDALONE_TOOLCHAIN to "
"use a standalone toolchain.\n"
"To cross-compile for Android, you need to:\n"
"1. Download an Android NDK from"
" https://developer.android.com/ndk/downloads/index.html\n"
"2. Setup a standalone toolchain"
"https://developer.android.google.cn/ndk/guides/standalone_toolchain.html?hl=zh-cn\n")
ENDIF()
IF(NOT DEFINED CMAKE_SYSTEM_VERSION AND ANDROID_NATIVE_API_LEVEL)
IF(ANDROID_NATIVE_API_LEVEL MATCHES "^android-[0-9]+$")
STRING(REPLACE "android-" "" CMAKE_SYSTEM_VERSION "${CMAKE_MATCH_0}")
ELSEIF(ANDROID_NATIVE_API_LEVEL MATCHES "^[0-9]+$")
SET(CMAKE_SYSTEM_VERSION ${ANDROID_NATIVE_API_LEVEL})
ENDIF()
ENDIF()
IF(NOT DEFINED ANDROID_TOOLCHAIN)
SET(ANDROID_TOOLCHAIN clang)
ENDIF()
IF(NOT DEFINED ANDROID_ABI)
SET(ANDROID_ABI "armeabi-v7a")
ENDIF()
IF(NOT DEFINED ANDROID_ARM_MODE)
SET(ANDROID_ARM_MODE ON)
ENDIF()
IF(ANDROID_ARM_MODE)
SET(ANDROID_ARM_MODE_NAME "arm")
ELSE(ANDROID_ARM_MODE)
SET(ANDROID_ARM_MODE_NAME "thumb")
ENDIF(ANDROID_ARM_MODE)
IF(NOT DEFINED ANDROID_ARM_NEON)
SET(ANDROID_ARM_NEON ON)
ENDIF()
IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
IF("${CMAKE_VERSION}" VERSION_LESS "3.1.0")
SET(CMAKE_SYSTEM_NAME "Linux")
ENDIF()
MESSAGE(WARNING "It is recommended to use CMake >= 3.7.0 (current version: "
"${CMAKE_VERSION}), when cross-compiling for Android.")
IF(ANDROID_STANDALONE_TOOLCHAIN)
# Use standalone toolchain
SET(CMAKE_SYSROOT "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot")
IF(NOT CMAKE_SYSTEM_VERSION)
SET(ANDROID_STANDALONE_TOOLCHAIN_API "")
SET(ANDROID_API_LEVEL_H_REGEX "^[\t ]*#[\t ]*define[\t ]+__ANDROID_API__[\t ]+([0-9]+)")
FILE(STRINGS "${ANDROID_STANDALONE_TOOLCHAIN}/sysroot/usr/include/android/api-level.h"
ANDROID_API_LEVEL_H_CONTENT REGEX "${ANDROID_API_LEVEL_H_REGEX}")
IF(ANDROID_API_LEVEL_H_CONTENT MATCHES "${ANDROID_API_LEVEL_H_REGEX}")
SET(ANDROID_STANDALONE_TOOLCHAIN_API "${CMAKE_MATCH_1}")
ENDIF()
SET(CMAKE_SYSTEM_VERSION ${ANDROID_STANDALONE_TOOLCHAIN_API})
ENDIF()
# Toolchain
SET(ANDROID_TOOLCHAIN_ROOT ${ANDROID_STANDALONE_TOOLCHAIN})
ELSE(ANDROID_NDK)
# TODO: use android ndk
ENDIF()
IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
SET(ANDROID_TOOLCHAIN_NAME arm-linux-androideabi)
IF(ANDROID_ABI STREQUAL "armeabi")
SET(CMAKE_SYSTEM_PROCESSOR armv5te)
SET(ANDROID_CLANG_TRIPLE armv5te-none-linux-androideabi)
ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
SET(ANDROID_CLANG_TRIPLE armv7-none-linux-androideabi)
ENDIF()
ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
SET(CMAKE_SYSTEM_PROCESSOR aarch64)
SET(ANDROID_CLANG_TRIPLE aarch64-none-linux-android)
ELSE()
MESSAGE(FATAL_ERROR "Invalid Android ABI: ${ANDROID_ABI}.")
ENDIF()
SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
IF(ANDROID_TOOLCHAIN STREQUAL clang)
SET(ANDROID_C_COMPILER_NAME clang)
SET(ANDROID_CXX_COMPILER_NAME clang++)
SET(CMAKE_C_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE})
SET(CMAKE_CXX_COMPILER_TARGET ${ANDROID_CLANG_TRIPLE})
ELSEIF(ANDROID_TOOLCHAIN STREQUAL gcc)
SET(ANDROID_C_COMPILER_NAME gcc)
SET(ANDROID_CXX_COMPILER_NAME g++)
ELSE()
MESSAGE(FATAL_ERROR "Invalid Android toolchain: ${ANDROID_TOOLCHAIN}")
ENDIF()
# C compiler
IF(NOT CMAKE_C_COMPILER)
SET(ANDROID_C_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_C_COMPILER_NAME}")
ELSE()
GET_FILENAME_COMPONENT(ANDROID_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
ENDIF()
IF(NOT EXISTS ${ANDROID_C_COMPILER})
MESSAGE(FATAL_ERROR "Cannot find C compiler: ${ANDROID_C_COMPILER}")
ENDIF()
# CXX compiler
IF(NOT CMAKE_CXX_COMPILER)
SET(ANDROID_CXX_COMPILER "${ANDROID_TOOLCHAIN_PREFIX}${ANDROID_CXX_COMPILER_NAME}")
ELSE()
GET_FILENAME_COMPONENT(ANDROID_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
ENDIF()
IF(NOT EXISTS ${ANDROID_CXX_COMPILER})
MESSAGE(FATAL_ERROR "Cannot find CXX compiler: ${ANDROID_CXX_COMPILER}")
ENDIF()
SET(CMAKE_C_COMPILER ${ANDROID_C_COMPILER} CACHE PATH "C compiler" FORCE)
SET(CMAKE_CXX_COMPILER ${ANDROID_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
# Toolchain and ABI specific flags.
SET(ANDROID_COMPILER_FLAGS "-ffunction-sections -fdata-sections")
SET(ANDROID_LINKER_FLAGS "-Wl,--gc-sections")
IF(ANDROID_ABI STREQUAL "armeabi")
LIST(APPEND ANDROID_COMPILER_FLAGS
-march=armv5te
-mtune=xscale
-msoft-float)
ELSEIF(ANDROID_ABI STREQUAL "armeabi-v7a")
LIST(APPEND ANDROID_COMPILER_FLAGS
-march=armv7-a
-mfloat-abi=softfp)
IF(ANDROID_ARM_NEON)
LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=neon)
ELSE()
LIST(APPEND ANDROID_COMPILER_FLAGS -mfpu=vfpv3-d16)
ENDIF()
LIST(APPEND ANDROID_LINKER_FLAGS -Wl,--fix-cortex-a8)
ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
ENDIF()
IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
IF(ANDROID_ARM_MODE)
LIST(APPEND ANDROID_COMPILER_FLAGS -marm)
ELSE()
LIST(APPEND ANDROID_COMPILER_FLAGS -mthumb)
ENDIF()
IF(ANDROID_TOOLCHAIN STREQUAL clang)
# Disable integrated-as for better compatibility.
LIST(APPEND ANDROID_COMPILER_FLAGS -fno-integrated-as)
ENDIF()
ENDIF()
IF(ANDROID_TOOLCHAIN STREQUAL clang)
# CMake automatically forwards all compiler flags to the linker,
# and clang doesn't like having -Wa flags being used for linking.
# To prevent CMake from doing this would require meddling with
# the CMAKE_<LANG>_COMPILE_OBJECT rules, which would get quite messy.
LIST(APPEND ANDROID_LINKER_FLAGS -Qunused-arguments)
ENDIF()
STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}")
STRING(REPLACE ";" " " ANDROID_LINKER_FLAGS "${ANDROID_LINKER_FLAGS}")
SET(CMAKE_C_FLAGS "${ANDROID_COMPILER_FLAGS} ${CMAKE_C_FLAGS}"
CACHE STRING "C flags")
SET(CMAKE_CXX_FLAGS "${ANDROID_COMPILER_FLAGS} ${CMAKE_CXX_FLAGS}"
CACHE STRING "CXX flags")
SET(CMAKE_SHARED_LINKER_FLAGS "${ANDROID_LINKER_FLAGS} ${CMAKE_SHARED_LINKER_FLAGS}"
CACHE STRING "shared linker flags")
SET(CMAKE_POSITION_INDEPENDENT_CODE TRUE)
SET(CMAKE_EXE_LINKER_FLAGS "-pie -fPIE ${ANDROID_LINKER_FLAGS} ${CMAKE_EXE_LINKER_FLAGS}"
CACHE STRING "executable linker flags")
MESSAGE(STATUS "Android: Targeting API '${CMAKE_SYSTEM_VERSION}' "
"with architecture '${ANDROID_ARM_MODE_NAME}', "
"ABI '${ANDROID_ABI}', and processor '${CMAKE_SYSTEM_PROCESSOR}'")
MESSAGE(STATUS "System CMAKE_C_FLAGS: " ${CMAKE_C_FLAGS})
MESSAGE(STATUS "System CMAKE_CXX_FLAGS: " ${CMAKE_CXX_FLAGS})
ELSE()
IF(ANDROID_STANDALONE_TOOLCHAIN)
SET(CMAKE_ANDROID_STANDALONE_TOOLCHAIN ${ANDROID_STANDALONE_TOOLCHAIN})
ENDIF()
SET(CMAKE_ANDROID_ARCH_ABI ${ANDROID_ABI})
IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
SET(CMAKE_ANDROID_ARM_MODE ${ANDROID_ARM_MODE})
IF(ANDROID_ABI STREQUAL "armeabi-v7a")
SET(CMAKE_ANDROID_ARM_NEON ${ANDROID_ARM_NEON})
ENDIF()
ENDIF()
ENDIF()
cmake/cross_compiling/host.cmake 已删除 100644 → 0
浏览文件 @ d6d3e6af
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# 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.
# find host C compiler
IF(HOST_C_COMPILER)
SET(HOST_C_COMPILER_NAME ${HOST_C_COMPILER})
ELSEIF(NOT $ENV{CC} STREQUAL "")
SET(HOST_C_COMPILER_NAME $ENV{CC})
ELSE()
SET(HOST_C_COMPILER_NAME cc)
ENDIF()
GET_FILENAME_COMPONENT(HOST_C_COMPILER_PATH ${HOST_C_COMPILER_NAME} PROGRAM)
IF(NOT HOST_C_COMPILER_PATH OR NOT EXISTS ${HOST_C_COMPILER_PATH})
MESSAGE(FATAL_ERROR "Cannot find host C compiler, set host C compiler:\n"
"\tcmake .. -DHOST_C_COMPILER=...")
ENDIF()
# find host CXX compiler
IF(HOST_CXX_COMPILER)
SET(HOST_CXX_COMPILER_NAME ${HOST_CXX_COMPILER})
ELSEIF(NOT $ENV{CXX} STREQUAL "")
SET(HOST_CXX_COMPILER_NAME $ENV{CXX})
ELSE()
SET(HOST_CXX_COMPILER_NAME c++)
ENDIF()
GET_FILENAME_COMPONENT(HOST_CXX_COMPILER_PATH ${HOST_CXX_COMPILER_NAME} PROGRAM)
IF(NOT HOST_CXX_COMPILER_PATH OR NOT EXISTS ${HOST_CXX_COMPILER_PATH})
MESSAGE(FATAL_ERROR "Cannot find host CXX compiler, set host CXX compiler:\n"
"\tcmake .. -DHOST_CXX_COMPILER=...")
ENDIF()
SET(HOST_C_COMPILER ${HOST_C_COMPILER_PATH} CACHE PATH "Host C compiler")
SET(HOST_CXX_COMPILER ${HOST_CXX_COMPILER_PATH} CACHE PATH "Host CXX compiler")
MESSAGE(STATUS "Found host C compiler: " ${HOST_C_COMPILER})
MESSAGE(STATUS "Found host CXX compiler: " ${HOST_CXX_COMPILER})
cmake/cross_compiling/ios.cmake 已删除 100644 → 0
浏览文件 @ d6d3e6af
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# 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")
set(IOS_ARCH "armv7;armv7s;arm64")
elseif(IOS_PLATFORM STREQUAL "SIMULATOR")
set(IOS_ARCH "i386;x86_64")
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=hidden -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/cross_compiling/raspberry_pi.cmake 已删除 100644 → 0
浏览文件 @ d6d3e6af
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# 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 Raspberry Pi.
# The supported variables are listed belows:
#
# RPI_TOOLCHAIN
# RPI_ARM_NEON
#
# Also you can set CMAKE_C/CXX_COMPILER yourself, through cmake arguments.
IF(NOT RPI)
return()
ENDIF()
SET(CMAKE_SYSTEM_NAME Linux)
SET(CMAKE_SYSTEM_VERSION 1)
SET(CMAKE_SYSTEM_PROCESSOR arm)
# check the exist of raspberry pi toolchain
IF(NOT DEFINED RPI_TOOLCHAIN)
SET(RPI_TOOLCHAIN $ENV{RPI_TOOLCHAIN}
CACHE PATH "Folder holds the toolchain of Raspberr Pi")
ENDIF()
IF(NOT RPI_TOOLCHAIN)
MESSAGE(WARNING "It is recommended to set RPI_TOOLCHAIN to use toolchain.\n"
"To cross-compile for Raspberry Pi, you need to download the tools using:\n"
" git clone https://github.com/raspberrypi/tools\n")
ENDIF()
IF(NOT DEFINED RPI_ARM_NEON)
SET(RPI_ARM_NEON ON)
ENDIF()
IF(RPI_TOOLCHAIN)
SET(RPI_TOOLCHAIN_ROOT ${RPI_TOOLCHAIN})
IF(RPI_TOOLCHAIN_ROOT MATCHES "gcc-linaro-arm-linux-gnueabihf-raspbian(-x64)?$")
# gcc-linaro-arm-linux-gnueabihf-raspbian
# gcc-linaro-arm-linux-gnueabihf-raspbian-x64
SET(RPI_TOOLCHAIN_NAME arm-linux-gnueabihf)
ENDIF()
SET(RPI_TOOLCHAIN_PREFIX "${RPI_TOOLCHAIN_ROOT}/bin/${RPI_TOOLCHAIN_NAME}-")
ENDIF()
# C compiler
IF(NOT CMAKE_C_COMPILER)
SET(RPI_C_COMPILER "${RPI_TOOLCHAIN_PREFIX}gcc")
ELSE()
GET_FILENAME_COMPONENT(RPI_C_COMPILER ${CMAKE_C_COMPILER} PROGRAM)
ENDIF()
IF(NOT EXISTS ${RPI_C_COMPILER})
MESSAGE(FATAL_ERROR "Cannot find C compiler: ${RPI_C_COMPILER}")
ENDIF()
# CXX compiler
IF(NOT CMAKE_CXX_COMPILER)
SET(RPI_CXX_COMPILER "${RPI_TOOLCHAIN_PREFIX}g++")
ELSE()
GET_FILENAME_COMPONENT(RPI_CXX_COMPILER ${CMAKE_CXX_COMPILER} PROGRAM)
ENDIF()
IF(NOT EXISTS ${RPI_CXX_COMPILER})
MESSAGE(FATAL_ERROR "Cannot find CXX compiler: ${RPI_CXX_COMPILER}")
ENDIF()
SET(CMAKE_C_COMPILER ${RPI_C_COMPILER} CACHE PATH "C compiler" FORCE)
SET(CMAKE_CXX_COMPILER ${RPI_CXX_COMPILER} CACHE PATH "CXX compiler" FORCE)
IF(RPI_ARM_NEON)
SET(RPI_C_FLAGS "${RPI_C_FLAGS} -mfpu=neon")
ENDIF()
SET(CMAKE_C_FLAGS "${RPI_C_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags")
SET(CMAKE_CXX_FLAGS "${RPI_C_FLAGS} ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags")
cmake/cuda.cmake
浏览文件 @ ee3aae56
......@@ -63,9 +63,7 @@ function(select_nvcc_arch_flags out_variable)
# List of arch names
set(archs_names "Kepler" "Maxwell" "Pascal" "Volta" "Turing" "All" "Manual")
set(archs_name_default "All")
if(NOT CMAKE_CROSSCOMPILING)
list(APPEND archs_names "Auto")
endif()
list(APPEND archs_names "Auto")
# set CUDA_ARCH_NAME strings (so it will be seen as dropbox in CMake-Gui)
set(CUDA_ARCH_NAME ${archs_name_default} CACHE STRING "Select target NVIDIA GPU achitecture.")
......
cmake/external/cares.cmake
浏览文件 @ ee3aae56
......@@ -13,7 +13,7 @@
# limitations under the License.
#
IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE)
IF(NOT WITH_DISTRIBUTE)
return()
ENDIF()
......
cmake/external/gflags.cmake
浏览文件 @ ee3aae56
......@@ -71,13 +71,3 @@ if (WIN32)
set_property(GLOBAL PROPERTY OS_DEPENDENCY_MODULES shlwapi.lib)
endif(HAVE_SHLWAPI)
endif (WIN32)
IF(WITH_C_API)
INSTALL(DIRECTORY ${GFLAGS_INCLUDE_DIR} DESTINATION third_party/gflags)
IF(ANDROID)
INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${GFLAGS_LIBRARIES} DESTINATION third_party/gflags/lib)
ENDIF()
ENDIF()
cmake/external/glog.cmake
浏览文件 @ ee3aae56
......@@ -26,14 +26,8 @@ ENDIF(WIN32)
INCLUDE_DIRECTORIES(${GLOG_INCLUDE_DIR})
IF(ANDROID AND ${CMAKE_SYSTEM_VERSION} VERSION_LESS "21")
# Using the unofficial glog for Android API < 21
SET(GLOG_REPOSITORY "https://github.com/Xreki/glog.git")
SET(GLOG_TAG "8a547150548b284382ccb6582408e9140ff2bea8")
ELSE()
SET(GLOG_REPOSITORY "https://github.com/google/glog.git")
SET(GLOG_TAG "v0.3.5")
ENDIF()
SET(GLOG_REPOSITORY "https://github.com/google/glog.git")
SET(GLOG_TAG "v0.3.5")
ExternalProject_Add(
extern_glog
......@@ -78,12 +72,3 @@ ADD_DEPENDENCIES(glog extern_glog gflags)
LINK_LIBRARIES(glog gflags)
LIST(APPEND external_project_dependencies glog)
IF(WITH_C_API)
INSTALL(DIRECTORY ${GLOG_INCLUDE_DIR} DESTINATION third_party/glog)
IF(ANDROID)
INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${GLOG_LIBRARIES} DESTINATION third_party/glog/lib)
ENDIF()
ENDIF()
cmake/external/grpc.cmake
浏览文件 @ ee3aae56
......@@ -13,7 +13,7 @@
# limitations under the License.
#
IF(MOBILE_INFERENCE OR NOT WITH_DISTRIBUTE)
IF(NOT WITH_DISTRIBUTE)
return()
ENDIF()
......
cmake/external/gzstream.cmake
浏览文件 @ ee3aae56
......@@ -13,10 +13,6 @@
# limitations under the License.
#
IF(MOBILE_INFERENCE)
return()
ENDIF()
include (ExternalProject)
# NOTE: gzstream is needed when linking with ctr reader.
......
cmake/external/libxsmm.cmake
浏览文件 @ ee3aae56
......@@ -19,8 +19,8 @@ IF(NOT WITH_LIBXSMM)
return()
ENDIF()
IF(WIN32 OR APPLE OR ANDROID OR IOS)
MESSAGE(WARNING "Windows, Mac or Mobile are not supported with libxsmm in Paddle yet.")
IF(WIN32 OR APPLE)
MESSAGE(WARNING "Windows, Mac are not supported with libxsmm in Paddle yet.")
SET(WITH_LIBXSMM OFF CACHE STRING "Disable LIBXSMM" FORCE)
return()
ENDIF()
......
cmake/external/mkldnn.cmake
浏览文件 @ ee3aae56
......@@ -110,7 +110,3 @@ else(WIN32)
endif(WIN32)
ADD_CUSTOM_TARGET(mkldnn_shared_lib ALL DEPENDS ${MKLDNN_SHARED_LIB})
ADD_DEPENDENCIES(mkldnn_shared_lib ${MKLDNN_PROJECT} mkldnn)
IF(WITH_C_API)
INSTALL(FILES ${MKLDNN_SHARED_LIB} DESTINATION lib)
ENDIF()
cmake/external/mklml.cmake
浏览文件 @ ee3aae56
......@@ -74,7 +74,3 @@ ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
LIST(APPEND external_project_dependencies mklml)
IF(WITH_C_API)
INSTALL(FILES ${MKLML_LIB} ${MKLML_IOMP_LIB} DESTINATION lib)
ENDIF()
cmake/external/nnpack.cmake 已删除 100644 → 0
浏览文件 @ d6d3e6af
# Find the NNPACK library
# NNPACK_ROOT - where to find NNPACK include and library.
#
set(NNPACK_FOUND OFF)
set(NNPACK_ROOT $ENV{NNPACK_ROOT} CACHE PATH "Folder contains NNPACK")
find_path(NNPACK_INC_DIR nnpack.h PATHS ${NNPACK_ROOT}/include)
find_library(NNPACK_LIB NAMES nnpack PATHS ${NNPACK_ROOT}/lib)
find_library(PTHREADPOOL_LIB NAMES pthreadpool PATHS ${NNPACK_ROOT}/lib)
find_library(NNPACK_UKERNELS_LIB NAMES nnpack_ukernels PATHS ${NNPACK_ROOT}/lib)
find_library(NNPACK_CPUFEATURES_LIB NAMES cpufeatures PATHS ${NNPACK_ROOT}/lib)
if(NNPACK_INC_DIR AND NNPACK_LIB AND PTHREADPOOL_LIB)
set(NNPACK_FOUND ON)
INCLUDE_DIRECTORIES(${NNPACK_INC_DIR})
set(NNPACK_LIBS)
list(APPEND NNPACK_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB})
if (NNPACK_UKERNELS_LIB)
list(APPEND NNPACK_LIBS ${NNPACK_UKERNELS_LIB})
endif()
if (NNPACK_CPUFEATURES_LIB)
list(APPEND NNPACK_LIBS ${NNPACK_CPUFEATURES_LIB})
endif()
if(NOT ANDROID)
list(APPEND NNPACK_LIBS "rt")
endif()
else()
message(FATAL_ERROR "Cannot find NNPACK in (${NNPACK_ROOT})")
endif()
cmake/external/openblas.cmake
浏览文件 @ ee3aae56
......@@ -40,38 +40,12 @@ IF(NOT ${CBLAS_FOUND})
SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -Wno-unused-but-set-variable -Wno-unused-variable")
SET(OPENBLAS_COMMIT "v0.2.20")
IF(CMAKE_CROSSCOMPILING)
SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER})
GET_FILENAME_COMPONENT(CROSS_SUFFIX ${CMAKE_C_COMPILER} DIRECTORY)
SET(CROSS_SUFFIX ${CROSS_SUFFIX}/)
IF(ANDROID)
IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
# use softfp
SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0)
ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0)
ENDIF()
ELSEIF(IOS)
IF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}")
SET(OPENBLAS_CC "${OPENBLAS_CC} -arch arm64")
SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0 CROSS_SUFFIX=${CROSS_SUFFIX})
ELSE()
MESSAGE(FATAL_ERROR "OpenBLAS only support arm64 architectures on iOS. "
"You can set IOS_USE_VECLIB_FOR_BLAS=ON or USE_EIGEN_FOR_BLAS=ON to use other blas library instead.")
ENDIF()
ELSEIF(RPI)
# use hardfp
SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 USE_THREAD=0)
ENDIF()
ELSE()
IF(APPLE)
SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
ENDIF()
SET(OPTIONAL_ARGS "")
IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)
ENDIF()
IF(APPLE)
SET(OPENBLAS_CC "${CMAKE_C_COMPILER} -isysroot ${CMAKE_OSX_SYSROOT}")
ENDIF()
SET(OPTIONAL_ARGS "")
IF(CMAKE_SYSTEM_PROCESSOR MATCHES "^x86(_64)?$")
SET(OPTIONAL_ARGS DYNAMIC_ARCH=1 NUM_THREADS=64)
ENDIF()
SET(COMMON_ARGS CC=${OPENBLAS_CC} NO_SHARED=1 NO_LAPACK=1 libs)
......@@ -92,25 +66,6 @@ IF(NOT ${CBLAS_FOUND})
ELSE()
ENDIF(NOT WIN32)
SET(CBLAS_PROVIDER openblas)
IF(WITH_C_API)
INSTALL(DIRECTORY ${CBLAS_INC_DIR} DESTINATION third_party/openblas)
# Because libopenblas.a is a symbolic link of another library, thus need to
# install the whole directory.
IF(ANDROID)
SET(TMP_INSTALL_DIR third_party/openblas/lib/${ANDROID_ABI})
ELSE()
SET(TMP_INSTALL_DIR third_party/openblas/lib)
ENDIF()
INSTALL(CODE "execute_process(
COMMAND ${CMAKE_COMMAND} -E copy_directory ${CBLAS_INSTALL_DIR}/lib
${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}
)"
)
INSTALL(CODE "MESSAGE(STATUS \"Installing: \"
\"${CBLAS_INSTALL_DIR}/lib -> ${CMAKE_INSTALL_PREFIX}/${TMP_INSTALL_DIR}\"
)"
)
ENDIF()
ENDIF(NOT ${CBLAS_FOUND})
MESSAGE(STATUS "BLAS library: ${CBLAS_LIBRARIES}")
......
cmake/external/protobuf.cmake
浏览文件 @ ee3aae56
......@@ -204,15 +204,6 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
SET(PROTOBUF_REPO "https://github.com/google/protobuf.git")
SET(PROTOBUF_TAG "9f75c5aa851cd877fb0d93ccc31b8567a6706546")
IF(MOBILE_INFERENCE)
# The reason why the official version is not used is described in
# https://github.com/PaddlePaddle/Paddle/issues/6114
SET(PROTOBUF_REPO "https://github.com/qingqing01/protobuf.git")
SET(PROTOBUF_TAG "v3.2.0")
IF(NOT BUILD_FOR_HOST)
SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} "-Dprotobuf_BUILD_PROTOC_BINARIES=OFF")
ENDIF()
ENDIF()
ExternalProject_Add(
${TARGET_NAME}
......@@ -240,19 +231,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
)
ENDFUNCTION()
IF(NOT MOBILE_INFERENCE)
SET(PROTOBUF_VERSION 3.1)
ELSE()
SET(PROTOBUF_VERSION 3.2)
ENDIF()
IF(CMAKE_CROSSCOMPILING)
build_protobuf(protobuf_host TRUE)
LIST(APPEND external_project_dependencies protobuf_host)
SET(PROTOBUF_PROTOC_EXECUTABLE ${protobuf_host_PROTOC_EXECUTABLE}
CACHE FILEPATH "protobuf executable." FORCE)
ENDIF()
SET(PROTOBUF_VERSION 3.1)
IF(NOT PROTOBUF_FOUND)
build_protobuf(extern_protobuf FALSE)
......@@ -266,20 +245,7 @@ IF(NOT PROTOBUF_FOUND)
SET(PROTOBUF_PROTOC_LIBRARY ${extern_protobuf_PROTOC_LIBRARY}
CACHE FILEPATH "protoc library." FORCE)
IF(WITH_C_API)
INSTALL(DIRECTORY ${PROTOBUF_INCLUDE_DIR} DESTINATION third_party/protobuf)
IF(ANDROID)
INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${PROTOBUF_LITE_LIBRARY} DESTINATION third_party/protobuf/lib)
ENDIF()
ENDIF()
IF(CMAKE_CROSSCOMPILING)
PROMPT_PROTOBUF_LIB(protobuf_host extern_protobuf)
ELSE()
SET(PROTOBUF_PROTOC_EXECUTABLE ${extern_protobuf_PROTOC_EXECUTABLE}
CACHE FILEPATH "protobuf executable." FORCE)
PROMPT_PROTOBUF_LIB(extern_protobuf)
ENDIF()
SET(PROTOBUF_PROTOC_EXECUTABLE ${extern_protobuf_PROTOC_EXECUTABLE}
CACHE FILEPATH "protobuf executable." FORCE)
PROMPT_PROTOBUF_LIB(extern_protobuf)
ENDIF(NOT PROTOBUF_FOUND)
cmake/external/pslib.cmake
浏览文件 @ ee3aae56
......@@ -71,7 +71,3 @@ ADD_LIBRARY(pslib SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET pslib PROPERTY IMPORTED_LOCATION ${PSLIB_LIB})
ADD_DEPENDENCIES(pslib ${PSLIB_PROJECT})
LIST(APPEND external_project_dependencies pslib)
IF(WITH_C_API)
INSTALL(FILES ${PSLIB_LIB} ${PSLIB_IOMP_LIB} DESTINATION lib)
ENDIF()
cmake/external/pslib_brpc.cmake
浏览文件 @ ee3aae56
......@@ -71,7 +71,3 @@ ADD_LIBRARY(pslib_brpc SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET pslib_brpc PROPERTY IMPORTED_LOCATION ${PSLIB_BRPC_LIB})
ADD_DEPENDENCIES(pslib_brpc ${PSLIB_BRPC_PROJECT})
LIST(APPEND external_project_dependencies pslib_brpc)
IF(WITH_C_API)
INSTALL(FILES ${PSLIB_BRPC_LIB} ${PSLIB_BRPC_IOMP_LIB} DESTINATION lib)
ENDIF()
cmake/external/snappy.cmake
浏览文件 @ ee3aae56
......@@ -12,10 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
if(MOBILE_INFERENCE OR RPI)
return()
endif()
include (ExternalProject)
# NOTE: snappy is needed when linking with recordio
......
cmake/external/snappystream.cmake
浏览文件 @ ee3aae56
......@@ -12,10 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
IF(MOBILE_INFERENCE OR RPI)
return()
ENDIF()
include (ExternalProject)
set(SNAPPYSTREAM_SOURCES_DIR ${THIRD_PARTY_PATH}/snappy_stream)
......
cmake/external/swig.cmake 已删除 100644 → 0
浏览文件 @ d6d3e6af
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# 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(NOT WITH_SWIG_PY)
return()
ENDIF()
FIND_PACKAGE(SWIG)
IF(NOT SWIG_FOUND)
# build swig as an external project
INCLUDE(ExternalProject)
SET(SWIG_SOURCES_DIR ${THIRD_PARTY_PATH}/swig)
SET(SWIG_INSTALL_DIR ${THIRD_PARTY_PATH}/install/swig)
SET(SWIG_TARGET_VERSION "3.0.2")
SET(SWIG_DOWNLOAD_SRC_MD5 "62f9b0d010cef36a13a010dc530d0d41")
SET(SWIG_DOWNLOAD_WIN_MD5 "3f18de4fc09ab9abb0d3be37c11fbc8f")
IF(WIN32)
# swig.exe available as pre-built binary on Windows:
ExternalProject_Add(swig
URL http://prdownloads.sourceforge.net/swig/swigwin-${SWIG_TARGET_VERSION}.zip
URL_MD5 ${SWIG_DOWNLOAD_WIN_MD5}
SOURCE_DIR ${SWIG_SOURCES_DIR}
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
INSTALL_COMMAND ""
UPDATE_COMMAND ""
)
SET(SWIG_DIR ${SWIG_SOURCES_DIR} CACHE FILEPATH "SWIG Directory" FORCE)
SET(SWIG_EXECUTABLE ${SWIG_SOURCES_DIR}/swig.exe CACHE FILEPATH "SWIG Executable" FORCE)
ELSE(WIN32)
# swig uses bison find it by cmake and pass it down
FIND_PACKAGE(BISON)
# From SWIG configure
ExternalProject_Add(swig
GIT_REPOSITORY https://github.com/swig/swig.git
GIT_TAG rel-3.0.10
PREFIX ${SWIG_SOURCES_DIR}
CONFIGURE_COMMAND cd <SOURCE_DIR> && ./autogen.sh && ./configure
--prefix=${SWIG_INSTALL_DIR} --without-pcre
BUILD_COMMAND cd <SOURCE_DIR> && make
INSTALL_COMMAND cd <SOURCE_DIR> && make install
UPDATE_COMMAND ""
)
SET(SWIG_DIR ${SWIG_INSTALL_DIR}/share/swig/${SWIG_TARGET_VERSION})
SET(SWIG_EXECUTABLE ${SWIG_INSTALL_DIR}/bin/swig)
ENDIF(WIN32)
LIST(APPEND external_project_dependencies swig)
ENDIF(NOT SWIG_FOUND)
cmake/external/warpctc.cmake
浏览文件 @ ee3aae56
......@@ -12,10 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
IF(MOBILE_INFERENCE)
return()
ENDIF()
INCLUDE(ExternalProject)
SET(WARPCTC_SOURCES_DIR ${THIRD_PARTY_PATH}/warpctc)
......
cmake/external/xxhash.cmake
浏览文件 @ ee3aae56
......@@ -73,12 +73,3 @@ include_directories(${XXHASH_INCLUDE_DIR})
add_dependencies(xxhash extern_xxhash)
LIST(APPEND external_project_dependencies xxhash)
IF(WITH_C_API)
INSTALL(DIRECTORY ${XXHASH_INCLUDE_DIR} DESTINATION third_party/xxhash)
IF(ANDROID)
INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib)
ENDIF()
ENDIF()
cmake/external/zlib.cmake
浏览文件 @ ee3aae56
......@@ -59,12 +59,3 @@ SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
ADD_DEPENDENCIES(zlib extern_zlib)
LIST(APPEND external_project_dependencies zlib)
IF(WITH_C_API)
INSTALL(DIRECTORY ${ZLIB_INCLUDE_DIR} DESTINATION third_party/zlib)
IF(ANDROID)
INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib/${ANDROID_ABI})
ELSE()
INSTALL(FILES ${ZLIB_LIBRARIES} DESTINATION third_party/zlib/lib)
ENDIF()
ENDIF()
cmake/flags.cmake
浏览文件 @ ee3aae56
......@@ -156,10 +156,8 @@ set(GPU_COMMON_FLAGS
endif(NOT WIN32)
if (APPLE)
if(NOT CMAKE_CROSSCOMPILING)
# On Mac OS X build fat binaries with x86_64 architectures by default.
set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
endif()
# On Mac OS X build fat binaries with x86_64 architectures by default.
set (CMAKE_OSX_ARCHITECTURES "x86_64" CACHE STRING "Build architectures for OSX" FORCE)
# On Mac OS X register class specifier is deprecated and will cause warning error on latest clang 10.0
set (COMMON_FLAGS -Wno-deprecated-register)
endif(APPLE)
......
cmake/generic.cmake
浏览文件 @ ee3aae56
......@@ -90,11 +90,11 @@
# including binary directory for generated headers.
include_directories(${CMAKE_CURRENT_BINARY_DIR})
if(NOT APPLE AND NOT ANDROID)
if(NOT APPLE)
find_package(Threads REQUIRED)
link_libraries(${CMAKE_THREAD_LIBS_INIT})
set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt")
endif(NOT APPLE AND NOT ANDROID)
endif(NOT APPLE)
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
......@@ -388,6 +388,7 @@ function(cc_test TARGET_NAME)
endif()
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_limit_of_tmp_allocation=4294967296) # 4G
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
# No unit test should exceed 10 minutes.
set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
......@@ -460,6 +461,7 @@ function(nv_test TARGET_NAME)
endif()
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_limit_of_tmp_allocation=4294967296) # 4G
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
endif()
endfunction(nv_test)
......@@ -655,12 +657,6 @@ function(paddle_protobuf_generate_cpp SRCS HDRS)
set(${SRCS})
set(${HDRS})
if (MOBILE_INFERENCE)
set(EXTRA_FLAG "lite:")
else()
set(EXTRA_FLAG "")
endif()
foreach(FIL ${ARGN})
get_filename_component(ABS_FIL ${FIL} ABSOLUTE)
get_filename_component(FIL_WE ${FIL} NAME_WE)
......@@ -677,7 +673,7 @@ function(paddle_protobuf_generate_cpp SRCS HDRS)
COMMAND ${CMAKE_COMMAND} -E make_directory "${CMAKE_CURRENT_BINARY_DIR}"
COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
-I${CMAKE_CURRENT_SOURCE_DIR}
--cpp_out "${EXTRA_FLAG}${CMAKE_CURRENT_BINARY_DIR}" ${ABS_FIL}
--cpp_out "${CMAKE_CURRENT_BINARY_DIR}" ${ABS_FIL}
DEPENDS ${ABS_FIL} protoc
COMMENT "Running C++ protocol buffer compiler on ${FIL}"
VERBATIM )
......@@ -714,9 +710,10 @@ function(py_test TARGET_NAME)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS ARGS ENVS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_test(NAME ${TARGET_NAME}
COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
FLAGS_cpu_deterministic=true
FLAGS_cpu_deterministic=true FLAGS_limit_of_tmp_allocation=4294967296 # 4G
PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
......
cmake/inference_lib.cmake
浏览文件 @ ee3aae56
......@@ -149,25 +149,23 @@ if (WITH_NGRAPH)
)
endif ()
if (NOT MOBILE_INFERENCE AND NOT RPI)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
copy(snappy_lib
SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappy)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
copy(snappy_lib
SRCS ${SNAPPY_INCLUDE_DIR} ${SNAPPY_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappy)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
copy(snappystream_lib
SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappystream)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappystream")
copy(snappystream_lib
SRCS ${SNAPPYSTREAM_INCLUDE_DIR} ${SNAPPYSTREAM_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappystream)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
copy(zlib_lib
SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS zlib)
endif ()
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
copy(zlib_lib
SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS zlib)
# paddle fluid module
set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
......
cmake/system.cmake
浏览文件 @ ee3aae56
......@@ -74,21 +74,6 @@ MARK_AS_ADVANCED(HOST_SYSTEM CPU_CORES)
MESSAGE(STATUS "Found Paddle host system: ${HOST_SYSTEM}, version: ${HOST_SYSTEM_VERSION}")
MESSAGE(STATUS "Found Paddle host system's CPU: ${CPU_CORES} cores")
# configuration for cross-compiling
IF(DEFINED CMAKE_SYSTEM_NAME)
INCLUDE(cross_compiling/host)
IF(${CMAKE_SYSTEM_NAME} STREQUAL "Android")
SET(ANDROID TRUE)
INCLUDE(cross_compiling/android)
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()
# external dependencies log output
SET(EXTERNAL_PROJECT_LOG_ARGS
LOG_DOWNLOAD 0 # Wrap download in script to log output
......
cmake/util.cmake
浏览文件 @ ee3aae56
......@@ -53,118 +53,3 @@ function(target_circle_link_libraries TARGET_NAME)
"-Wl,--end-group")
endif()
endfunction()
# compile_cu_as_cpp
# Make a cu file compiled as C++
# Arguments: Source files
macro(compile_cu_as_cpp)
foreach(s ${ARGN})
set_source_files_properties(${s} PROPERTIES LANGUAGE CXX)
set_source_files_properties(${s} PROPERTIES COMPILE_FLAGS "-x c++")
endforeach()
endmacro()
# link_paddle_exe
# add paddle library for a paddle executable, such as trainer, pserver.
#
# It will handle WITH_PYTHON etc.
function(link_paddle_exe TARGET_NAME)
if(WITH_RDMA)
generate_rdma_links()
endif()
if(MOBILE_INFERENCE)
target_circle_link_libraries(${TARGET_NAME}
ARCHIVE_START
paddle_gserver
paddle_function
ARCHIVE_END
paddle_math
paddle_utils
paddle_parameter
paddle_proto
paddle_cuda
${EXTERNAL_LIBS}
${CMAKE_THREAD_LIBS_INIT}
${CMAKE_DL_LIBS}
${RDMA_LD_FLAGS}
${RDMA_LIBS})
else()
target_circle_link_libraries(${TARGET_NAME}
ARCHIVE_START
paddle_gserver
paddle_function
ARCHIVE_END
paddle_pserver
paddle_trainer_lib
paddle_network
paddle_math
paddle_utils
paddle_parameter
paddle_proto
paddle_cuda
paddle_optimizer
${EXTERNAL_LIBS}
${CMAKE_THREAD_LIBS_INIT}
${CMAKE_DL_LIBS}
${RDMA_LD_FLAGS}
${RDMA_LIBS})
endif()
if(ANDROID)
target_link_libraries(${TARGET_NAME} log)
endif(ANDROID)
if(WITH_MKLML AND MKLML_LIB_DIR AND MKLML_IOMP_LIB)
target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
endif()
add_dependencies(${TARGET_NAME} ${external_project_dependencies})
endfunction()
# link_paddle_test
# Link a paddle unittest for target
# TARGET_NAME: the unittest target name
# Rest Arguemnts: not used.
function(link_paddle_test TARGET_NAME)
link_paddle_exe(${TARGET_NAME})
target_link_libraries(${TARGET_NAME}
paddle_test_main
paddle_test_util
${GTEST_LIBRARIES})
endfunction()
# add_unittest_without_exec
#
# create a paddle unittest. not specifically define how to run this unittest.
# TARGET_NAME: the unittest target name, same as executable file name
# Rest Arguments: the source files to compile this unittest.
macro(add_unittest_without_exec TARGET_NAME)
add_executable(${TARGET_NAME} ${ARGN})
link_paddle_test(${TARGET_NAME})
endmacro()
# add_unittest
# create a paddle unittest and just to execute this binary to make unittest.
#
# TARGET_NAME: the unittest target name, same as executable file name
# Rest Arguments: the source files to compile this unittest.
macro(add_unittest TARGET_NAME)
add_unittest_without_exec(${TARGET_NAME} ${ARGN})
add_test(${TARGET_NAME} ${TARGET_NAME})
endmacro()
# add_simple_unittest
# create a paddle unittest with file name. It just compile ${TARGET_NAME}.cpp to
# ${TARGET_NAME} and then execute it.
macro(add_simple_unittest TARGET_NAME)
add_unittest(${TARGET_NAME} ${TARGET_NAME}.cpp)
endmacro()
# Creates C resources file from files in given resource file
function(create_resources res_file output_file)
add_custom_command(
OUTPUT ${output_file}
COMMAND python ARGS ${PADDLE_SOURCE_DIR}/cmake/make_resource.py ${res_file} ${output_file}
DEPENDS ${res_file} ${PADDLE_SOURCE_DIR}/cmake/make_resource.py)
endfunction()
paddle/fluid/API.spec
浏览文件 @ ee3aae56
......@@ -45,6 +45,7 @@ paddle.fluid.AsyncExecutor.save_model ArgSpec(args=['self', 'save_path'], vararg
paddle.fluid.AsyncExecutor.stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.CompiledProgram.__init__ ArgSpec(args=['self', 'program'], varargs=None, keywords=None, defaults=None)
paddle.fluid.CompiledProgram.with_data_parallel ArgSpec(args=['self', 'loss_name', 'build_strategy', 'exec_strategy', 'share_vars_from'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.CompiledProgram.with_inference_optimize ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=None)
paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.ExecutionStrategy) -> None
paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.GradientScaleStrategy, arg0: int) -> None
paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.ReduceStrategy, arg0: int) -> None
......@@ -66,6 +67,7 @@ paddle.fluid.initializer.BilinearInitializer.__init__ ArgSpec(args=['self'], var
paddle.fluid.initializer.MSRAInitializer.__init__ ArgSpec(args=['self', 'uniform', 'fan_in', 'seed'], varargs=None, keywords=None, defaults=(True, None, 0))
paddle.fluid.initializer.force_init_on_cpu ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.initializer.init_on_cpu ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
paddle.fluid.initializer.NumpyArrayInitializer.__init__ ArgSpec(args=['self', 'value'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.fc ArgSpec(args=['input', 'size', 'num_flatten_dims', 'param_attr', 'bias_attr', 'act', 'is_test', 'name'], varargs=None, keywords=None, defaults=(1, None, None, None, False, None))
paddle.fluid.layers.embedding ArgSpec(args=['input', 'size', 'is_sparse', 'is_distributed', 'padding_idx', 'param_attr', 'dtype'], varargs=None, keywords=None, defaults=(False, False, None, None, 'float32'))
paddle.fluid.layers.dynamic_lstm ArgSpec(args=['input', 'size', 'h_0', 'c_0', 'param_attr', 'bias_attr', 'use_peepholes', 'is_reverse', 'gate_activation', 'cell_activation', 'candidate_activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, True, False, 'sigmoid', 'tanh', 'tanh', 'float32', None))
......@@ -120,7 +122,7 @@ paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None,
paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None))
paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name', 'sampler', 'custom_dist', 'seed', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 'uniform', None, 0, False))
paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name', 'path_table', 'path_code', 'is_custom', 'is_sparse'], varargs=None, keywords=None, defaults=(None, None, None, None, None, False, False))
paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None))
paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'is_accumulated', 'name'], varargs=None, keywords=None, defaults=(0, True, None))
paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None))
......@@ -196,7 +198,7 @@ paddle.fluid.layers.clip ArgSpec(args=['x', 'min', 'max', 'name'], varargs=None,
paddle.fluid.layers.clip_by_norm ArgSpec(args=['x', 'max_norm', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None))
paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'ignore_index', 'name'], varargs=None, keywords=None, defaults=(-100, None))
paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'ignore_index', 'name', 'normalize'], varargs=None, keywords=None, defaults=(-100, None, False))
paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
......@@ -211,6 +213,7 @@ paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act
paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
paddle.fluid.layers.shuffle_channel ArgSpec(args=['x', 'group', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.py_func ArgSpec(args=['func', 'x', 'out', 'backward_func', 'skip_vars_in_backward_input'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.teacher_student_sigmoid_loss ArgSpec(args=['input', 'label', 'soft_max_up_bound', 'soft_max_lower_bound'], varargs=None, keywords=None, defaults=(15.0, -15.0))
......@@ -317,6 +320,7 @@ paddle.fluid.layers.anchor_generator ArgSpec(args=['input', 'anchor_sizes', 'asp
paddle.fluid.layers.roi_perspective_transform ArgSpec(args=['input', 'rois', 'transformed_height', 'transformed_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1.0,))
paddle.fluid.layers.generate_proposal_labels ArgSpec(args=['rpn_rois', 'gt_classes', 'is_crowd', 'gt_boxes', 'im_info', 'batch_size_per_im', 'fg_fraction', 'fg_thresh', 'bg_thresh_hi', 'bg_thresh_lo', 'bbox_reg_weights', 'class_nums', 'use_random'], varargs=None, keywords=None, defaults=(256, 0.25, 0.25, 0.5, 0.0, [0.1, 0.1, 0.2, 0.2], None, True))
paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'im_info', 'anchors', 'variances', 'pre_nms_top_n', 'post_nms_top_n', 'nms_thresh', 'min_size', 'eta', 'name'], varargs=None, keywords=None, defaults=(6000, 1000, 0.5, 0.1, 1.0, None))
paddle.fluid.layers.generate_mask_labels ArgSpec(args=['im_info', 'gt_classes', 'is_crowd', 'gt_segms', 'rois', 'labels_int32', 'num_classes', 'resolution'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None))
paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
......@@ -356,6 +360,7 @@ paddle.fluid.contrib.QuantizeTranspiler.__init__ ArgSpec(args=['self', 'weight_b
paddle.fluid.contrib.QuantizeTranspiler.convert_to_int8 ArgSpec(args=['self', 'program', 'place', 'scope'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.QuantizeTranspiler.freeze_program ArgSpec(args=['self', 'program', 'place', 'fuse_bn', 'scope'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.contrib.QuantizeTranspiler.training_transpile ArgSpec(args=['self', 'program', 'startup_program'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.contrib.reader.ctr_reader.ctr_reader ArgSpec(args=['feed_dict', 'file_type', 'file_format', 'dense_slot_index', 'sparse_slot_index', 'capacity', 'thread_num', 'batch_size', 'file_list', 'slots', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.contrib.build_compressor ArgSpec(args=['place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'config'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
paddle.fluid.contrib.CompressPass.__init__ ArgSpec(args=['self', 'place', 'data_reader', 'data_feeder', 'scope', 'metrics', 'epoch', 'program_exe'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None))
paddle.fluid.contrib.CompressPass.add_strategy ArgSpec(args=['self', 'strategy'], varargs=None, keywords=None, defaults=None)
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ ee3aae56
#windows treat symbolic file as a real file, which is different with unix
#We create a hidden file and compile it instead of origin source file.
function(windows_symbolic TARGET)
......@@ -129,12 +128,6 @@ cc_test(version_test SRCS version_test.cc DEPS version)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
if(WITH_NGRAPH)
cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler)
endif(WITH_NGRAPH)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
......@@ -171,13 +164,12 @@ if(WITH_DISTRIBUTE)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
else()
if(WITH_NGRAPH)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator variable_helper)
else(WITH_NGRAPH)
if (WITH_NGRAPH)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper ngraph_engine)
else ()
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif(WITH_NGRAPH)
endif()
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif()
......@@ -215,3 +207,24 @@ endif (NOT WIN32)
cc_library(dlpack_tensor SRCS dlpack_tensor.cc DEPS tensor dlpack)
cc_test(dlpack_tensor_test SRCS dlpack_tensor_test.cc DEPS dlpack_tensor glog)
# Get the current working branch
execute_process(
COMMAND git rev-parse --abbrev-ref HEAD
WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_BRANCH
OUTPUT_STRIP_TRAILING_WHITESPACE
)
# Get the latest abbreviated commit hash of the working branch
execute_process(
COMMAND git log -1 --format=%h
WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}
OUTPUT_VARIABLE PADDLE_COMMIT
OUTPUT_STRIP_TRAILING_WHITESPACE
)
message(STATUS "commit: ${PADDLE_COMMIT}")
message(STATUS "branch: ${PADDLE_BRANCH}")
configure_file(commit.h.in commit.h)
paddle/fluid/framework/commit.h.in 0 → 100644
浏览文件 @ ee3aae56
#pragma once
#include <string>
namespace paddle {
namespace framework {
static std::string paddle_commit() {
return "@PADDLE_COMMIT@";
}
static std::string paddle_compile_branch() {
return "@PADDLE_BRANCH@";
}
static std::string paddle_version() {
return "@PADDLE_VERSION@";
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ ee3aae56
......@@ -25,6 +25,7 @@ limitations under the License. */
#include "paddle/fluid/framework/details/sequential_execution_pass.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/ir/graph_to_program_pass.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
namespace paddle {
......@@ -282,3 +283,4 @@ USE_PASS(modify_op_lock_and_record_event_pass);
USE_PASS(inplace_pass);
USE_PASS(lock_free_optimize_pass);
USE_PASS(graph_print_pass);
USE_PASS(graph_to_program_pass);
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ ee3aae56
......@@ -93,7 +93,7 @@ struct BuildStrategy {
int num_trainers_{1};
int trainer_id_{0};
std::vector<std::string> trainers_endpoints_;
bool remove_unnecessary_lock_{false};
bool remove_unnecessary_lock_{true};
// NOTE:
// Before you add new options, think if it's a general strategy that works
......
paddle/fluid/framework/details/execution_strategy.h
浏览文件 @ ee3aae56
......@@ -25,6 +25,9 @@ struct ExecutionStrategy {
size_t num_threads_{0};
bool use_cuda_{true};
bool allow_op_delay_{false};
// If we set this to 1, we will delete all variables when finish a batch. and
// this will loss 15%+ performance.
// Please be aware about this parameters.
size_t num_iteration_per_drop_scope_{1};
ExecutorType type_{kDefault};
bool dry_run_{false};
......
paddle/fluid/framework/executor.cc
浏览文件 @ ee3aae56
......@@ -27,7 +27,7 @@ limitations under the License. */
#include "paddle/fluid/platform/profiler.h"
#ifdef PADDLE_WITH_NGRAPH
#include "paddle/fluid/framework/ngraph_operator.h"
#include "paddle/fluid/operators/ngraph/ngraph_engine.h"
#endif
DECLARE_bool(benchmark);
......@@ -133,24 +133,6 @@ static void DeleteUnusedTensors(
}
}
static void EnableFusedOp(ExecutorPrepareContext* ctx) {
#ifdef PADDLE_WITH_NGRAPH
VLOG(3) << "use_ngraph=True";
auto intervals = NgraphOperator::NgraphOpIntervals(&ctx->ops_);
for (auto& interval : intervals) {
auto* ng_op = new NgraphOperator(ctx->prog_, ctx->block_id_, interval.at(0),
interval.at(1));
*interval[0] = std::unique_ptr<OperatorBase>(ng_op);
}
for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) {
ctx->ops_.erase(it->at(0) + 1, it->at(1));
}
#else
LOG(WARNING)
<< "'NGRAPH' is not supported, Please re-compile with WITH_NGRAPH option";
#endif
}
Executor::Executor(const platform::Place& place) : place_(place) {}
void Executor::Close() {
......@@ -204,6 +186,9 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
bool create_local_scope, bool create_vars) {
platform::RecordBlock b(block_id);
if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
#ifdef PADDLE_WITH_NGRAPH
if (FLAGS_use_ngraph) operators::NgraphEngine::EnableNgraph(pdesc);
#endif
auto ctx = Prepare(pdesc, block_id);
RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars);
}
......@@ -379,7 +364,6 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
for (auto& op_desc : block.AllOps()) {
ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
}
if (FLAGS_use_ngraph) EnableFusedOp(ctx.get());
return ctx;
}
......
paddle/fluid/framework/ir/graph_traits.cc
浏览文件 @ ee3aae56
......@@ -14,6 +14,7 @@
#include "paddle/fluid/framework/ir/graph_traits.h"
#include <set>
#include <vector>
namespace paddle {
......@@ -79,7 +80,7 @@ NodesTSIterator::NodesTSIterator(const std::vector<Node *> &source) {
}
std::unordered_set<Node *> visited;
std::unordered_set<Node *> to_visit{source.begin(), source.end()};
std::set<Node *> to_visit{source.begin(), source.end()};
std::vector<Node *> inlink_visited;
while (!to_visit.empty()) {
......
paddle/fluid/framework/ir/pass.cc
浏览文件 @ ee3aae56
......@@ -28,10 +28,14 @@ std::unique_ptr<Graph> Pass::Apply(std::unique_ptr<Graph> graph) const {
PADDLE_ENFORCE(graph->Has(attr), "Required graph atrribute %s not set.",
attr);
}
auto* native_graph = graph.get();
auto applied_graph = ApplyImpl(std::move(graph));
// TODO(panyx0718): Add more verifications.
PADDLE_ENFORCE(!HasCircle(*applied_graph),
"Illegal Pass. Generated graph shouldn't has cycle.");
PADDLE_ENFORCE(applied_graph.get() == native_graph,
"Pass::Apply() cannot delete the passed graph and shouldn't "
"return a new graph.(For the need of pybind11)");
applied_ = true;
return applied_graph;
}
......
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ ee3aae56
......@@ -54,13 +54,14 @@ std::ostream &operator<<(std::ostream &os, const LoD &lod) {
std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
if (!platform::is_cpu_place(t.place())) {
LoDTensor tt;
framework::TensorCopy(t, platform::CPUPlace(), &tt);
LoDTensor cpu_tensor;
cpu_tensor.set_lod(t.lod());
framework::TensorCopy(t, platform::CPUPlace(), &cpu_tensor);
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto &dev_ctx = *pool.Get(t.place());
dev_ctx.Wait();
os << tt;
os << cpu_tensor;
return os;
}
......
paddle/fluid/framework/mixed_vector.h
浏览文件 @ ee3aae56
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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
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
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. */
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
......
paddle/fluid/framework/ngraph_operator.h 已删除 100644 → 0
浏览文件 @ d6d3e6af
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
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 <algorithm>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/attribute.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_kernel_type.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/variant.h"
#include "ngraph/type/element_type.hpp"
namespace paddle {
namespace framework {
class NgraphOperator : public OperatorBase {
public:
static std::vector<
std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
NgraphOpIntervals(
std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops);
explicit NgraphOperator(
const ProgramDesc& prog, size_t block_id,
std::vector<std::unique_ptr<OperatorBase>>::iterator start,
std::vector<std::unique_ptr<OperatorBase>>::iterator end,
const std::string& type = "fused_op", const VariableNameMap& inputs = {},
const VariableNameMap& outputs = {}, const AttributeMap& attrs = {});
void RunImpl(const Scope& scope, const platform::Place& place) const final;
private:
const ProgramDesc pdesc_;
size_t block_;
std::vector<std::shared_ptr<OperatorBase>> fused_ops_;
std::unordered_map<std::string, ngraph::element::Type> var_type_map_;
std::unordered_set<std::string> persistables_;
std::unordered_set<std::string> fetches_;
std::unordered_set<std::string> post_op_inputs_;
bool is_full_ = false;
void Process();
};
} // namespace framework
} // namespace paddle
paddle/fluid/framework/operator.cc
浏览文件 @ ee3aae56
......@@ -19,8 +19,6 @@ limitations under the License. */
#include <sstream>
#include <string>
#include <vector>
#include "gflags/gflags.h"
#include "glog/logging.h"
#include "paddle/fluid/framework/data_transform.h"
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/lod_tensor.h"
......@@ -1075,7 +1073,9 @@ Scope* OperatorWithKernel::PrepareData(
proto::VarType::Type OperatorWithKernel::IndicateDataType(
const ExecutionContext& ctx) const {
int data_type = -1;
proto::VarType::Type dafault_data_type =
static_cast<proto::VarType::Type>(-1);
proto::VarType::Type data_type = dafault_data_type;
for (auto& input : this->inputs_) {
const std::vector<const Variable*> vars = ctx.MultiInputVar(input.first);
for (size_t i = 0; i < vars.size(); ++i) {
......@@ -1092,18 +1092,19 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
if (t != nullptr) {
PADDLE_ENFORCE(t->IsInitialized(), "Input %s(%lu)is not initialized",
input.first, i);
int tmp = static_cast<int>(t->type());
proto::VarType::Type tmp = t->type();
PADDLE_ENFORCE(
tmp == data_type || data_type == -1,
tmp == data_type || data_type == dafault_data_type,
"DataType of Paddle Op %s must be the same. Get (%d) != (%d)",
Type(), data_type, tmp);
Type(), DataTypeToString(data_type), DataTypeToString(tmp));
data_type = tmp;
}
}
}
}
PADDLE_ENFORCE(data_type != -1, "DataType should be indicated by input");
return static_cast<proto::VarType::Type>(data_type);
PADDLE_ENFORCE(data_type != dafault_data_type,
"DataType should be indicated by input");
return data_type;
}
OpKernelType OperatorWithKernel::GetExpectedKernelType(
......
paddle/fluid/framework/tensor_impl.h
浏览文件 @ ee3aae56
......@@ -25,7 +25,8 @@ inline const T* Tensor::data() const {
check_memory_size();
bool valid =
std::is_same<T, void>::value || type_ == DataTypeTrait<T>::DataType;
PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %d", type_);
PADDLE_ENFORCE(valid, "Tensor holds the wrong type, it holds %d",
DataTypeToString(type_));
return reinterpret_cast<const T*>(
reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
......
paddle/fluid/imperative/CMakeLists.txt
浏览文件 @ ee3aae56
if(WITH_PYTHON)
cc_library(layer SRCS layer.cc DEPS proto_desc operator)
cc_library(tracer SRCS tracer.cc DEPS proto_desc)
cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas)
cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context)
cc_library(engine SRCS engine.cc)
endif()
paddle/fluid/imperative/layer.cc
浏览文件 @ ee3aae56
......@@ -13,6 +13,7 @@
// limitations under the License.
#include "paddle/fluid/imperative/layer.h"
#include <deque>
#include <limits>
#include <map>
......@@ -22,6 +23,9 @@
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/string/printf.h"
namespace paddle {
......@@ -34,22 +38,66 @@ std::map<int, py::object> py_funcs_;
using framework::Variable;
void AddTo(Variable* src, Variable* dst) {
framework::LoDTensor* dst_tensor = dst->GetMutable<framework::LoDTensor>();
framework::LoDTensor* src_tensor = src->GetMutable<framework::LoDTensor>();
namespace detail {
template <typename T>
class TensorAddToFunctor : public boost::static_visitor<> {
public:
TensorAddToFunctor(int64_t numel, const T* x, T* y)
: numel_(numel), x_(x), y_(y) {}
void operator()(const platform::CPUPlace& place) {
platform::CPUDeviceContext* ctx = dynamic_cast<platform::CPUDeviceContext*>(
platform::DeviceContextPool::Instance().Get(place));
auto blas = operators::math::GetBlas<platform::CPUDeviceContext, T>(*ctx);
blas.AXPY(numel_, 1., x_, y_);
}
#ifdef PADDLE_WITH_CUDA
void operator()(const platform::CUDAPlace& place) {
platform::CUDADeviceContext* ctx =
dynamic_cast<platform::CUDADeviceContext*>(
platform::DeviceContextPool::Instance().Get(place));
auto blas = operators::math::GetBlas<platform::CUDADeviceContext, T>(*ctx);
blas.AXPY(numel_, 1., x_, y_);
}
#else
void operator()(const platform::CUDAPlace& place) {
PADDLE_THROW("Do NOT support gradient merge in place %s", place);
}
#endif
// there is NO blas in CUDAPinnedPlace
void operator()(const platform::CUDAPinnedPlace& place) {
PADDLE_THROW("Do NOT support gradient merge in place %s", place);
}
private:
int64_t numel_;
const T* x_;
T* y_;
};
} // namespace detail
void AddTo(Variable* src, Variable* dst, platform::Place place) {
framework::Tensor* dst_tensor = dst->GetMutable<framework::LoDTensor>();
framework::Tensor* src_tensor = src->GetMutable<framework::LoDTensor>();
// FIXME(minqiyang): loss_grad op will pass a zero grad of label
// ugly fix for it
if (src_tensor->numel() == 0) {
return;
}
PADDLE_ENFORCE(dst_tensor->numel() == src_tensor->numel(),
"dst_numel %lld vs. src_numel %lld", dst_tensor->numel(),
src_tensor->numel());
float* dst_data = dst_tensor->mutable_data<float>(platform::CPUPlace());
const float* src_data = src_tensor->data<float>();
for (int64_t i = 0; i < src_tensor->numel(); ++i) {
dst_data[i] += src_data[i];
}
detail::TensorAddToFunctor<float> func(
src_tensor->numel(), src_tensor->data<float>(),
dst_tensor->mutable_data<float>(place));
boost::apply_visitor(func, place);
}
class Autograd {
......@@ -120,66 +168,104 @@ class Autograd {
}
};
std::unique_ptr<VarBase> VarBase::NewVarBase(const platform::Place& dst_place,
const bool blocking) const {
PADDLE_ENFORCE(var_->IsInitialized(),
"Variable must be initialized when getting numpy tensor");
std::unique_ptr<VarBase> new_var(new VarBase());
framework::LoDTensor* tensor =
new_var->var_->GetMutable<framework::LoDTensor>();
tensor->Resize(var_->Get<framework::LoDTensor>().dims());
tensor->set_lod(var_->Get<framework::LoDTensor>().lod());
if (blocking) {
platform::DeviceContext* dev_ctx =
platform::DeviceContextPool::Instance().Get(dst_place);
framework::TensorCopySync(var_->Get<framework::LoDTensor>(), dst_place,
tensor);
dev_ctx->Wait();
} else {
framework::TensorCopy(var_->Get<framework::LoDTensor>(), dst_place, tensor);
}
if (platform::is_gpu_place(dst_place)) {
VLOG(3) << "copy tensor " << var_desc_->Name() << " from gpu";
}
return new_var;
}
framework::LoDTensor& VarBase::GradValue() {
VLOG(3) << "get var grad " << var_desc_->Name();
return *(grads_->var_->GetMutable<framework::LoDTensor>());
}
std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
if (!grad_op_desc_ && backward_id_ <= 0) {
if (grad_op_descs_.empty() && backward_id_ <= 0) {
LOG(WARNING) << "op with no grad: " << op_desc_->Type();
return {};
}
std::map<std::string, std::vector<framework::Variable*>> grad_outputs;
std::vector<framework::VariableValueMap> grad_outputs;
if (backward_id_ > 0) {
VLOG(3) << "py_layer_grad";
grad_outputs[framework::GradVarName(PyLayer::kFwdOut)] = PyLayer::ApplyGrad(
backward_id_,
grad_input_vars_[framework::GradVarName(PyLayer::kFwdInp)]);
grad_outputs.resize(1);
grad_outputs[0][framework::GradVarName(PyLayer::kFwdOut)] =
PyLayer::ApplyGrad(
backward_id_,
grad_input_vars_[0][framework::GradVarName(PyLayer::kFwdInp)]);
} else {
VLOG(3) << "op grad " << grad_op_desc_->Type();
for (auto it : grad_output_vars_) {
auto& outputs = grad_outputs[it.first];
for (size_t i = 0; i < it.second.size(); ++i) {
// Allocate a new variable
Variable* tmp_var = new framework::Variable();
tmp_var->GetMutable<framework::LoDTensor>();
outputs.push_back(tmp_var);
grad_outputs.resize(grad_op_descs_.size());
for (size_t k = 0; k < grad_op_descs_.size(); ++k) {
framework::OpDesc* grad_op_desc = grad_op_descs_[k];
VLOG(3) << "op grad " << grad_op_desc->Type();
for (auto it : grad_output_vars_[k]) {
auto& outputs = grad_outputs[k][it.first];
for (size_t i = 0; i < it.second.size(); ++i) {
// Allocate a new variable
Variable* tmp_var = new framework::Variable();
tmp_var->GetMutable<framework::LoDTensor>();
outputs.push_back(tmp_var);
}
}
}
framework::RuntimeContext ctx(grad_input_vars_, grad_outputs);
framework::RuntimeContext ctx(grad_input_vars_[k], grad_outputs[k]);
// No need to do compile time infer shape here.
// grad_op_desc_->InferShape(*block_);
grad_op_desc_->InferVarType(block_);
// No need to do compile time infer shape here.
// grad_op_desc_->InferShape(*block_);
grad_op_desc->InferVarType(block_);
std::unique_ptr<framework::OperatorBase> opbase =
framework::OpRegistry::CreateOp(*grad_op_desc_);
framework::OperatorWithKernel* op_kernel =
dynamic_cast<framework::OperatorWithKernel*>(opbase.get());
PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
std::unique_ptr<framework::OperatorBase> opbase =
framework::OpRegistry::CreateOp(*grad_op_desc);
framework::OperatorWithKernel* op_kernel =
dynamic_cast<framework::OperatorWithKernel*>(opbase.get());
PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
framework::Scope scope;
platform::CPUPlace place;
PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place);
p.op.RuntimeInferShape(scope, place, ctx);
p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
framework::Scope scope;
PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place_);
p.op.RuntimeInferShape(scope, place_, ctx);
p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
}
}
for (auto it : grad_output_vars_) {
auto& outputs = grad_outputs[it.first];
auto& origin_outputs = it.second;
PADDLE_ENFORCE_EQ(outputs.size(), origin_outputs.size());
for (size_t i = 0; i < outputs.size(); ++i) {
framework::Variable* grad = outputs[i];
framework::Variable* orig_grad = origin_outputs[i];
AddTo(grad, orig_grad);
delete grad;
for (size_t k = 0; k < grad_output_vars_.size(); ++k) {
for (auto it : grad_output_vars_[k]) {
auto& outputs = grad_outputs[k][it.first];
auto& origin_outputs = it.second;
PADDLE_ENFORCE_EQ(outputs.size(), origin_outputs.size());
for (size_t i = 0; i < outputs.size(); ++i) {
framework::Variable* grad = outputs[i];
framework::Variable* orig_grad = origin_outputs[i];
AddTo(grad, orig_grad, place_);
delete grad;
}
}
}
return input_vars_;
}
......@@ -188,8 +274,10 @@ void VarBase::RunBackward() {
VLOG(3) << "start backward";
auto grads_t = grads_->var_->GetMutable<framework::LoDTensor>();
float* data = grads_t->mutable_data<float>(platform::CPUPlace());
std::fill(data, data + grads_t->numel(), 1.0);
operators::math::set_constant(
*(platform::DeviceContextPool::Instance().Get(
var_->GetMutable<framework::LoDTensor>()->place())),
grads_t, 1.0);
PADDLE_ENFORCE(
grads_ ==
......
paddle/fluid/imperative/layer.h
浏览文件 @ ee3aae56
......@@ -21,17 +21,21 @@
#include <map> // NOLINT
#include <string> // NOLINT
#include <vector> // NOLINT
#include <memory> // NOLINT
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/imperative/type_defs.h"
namespace paddle {
namespace imperative {
class VarBase;
namespace py = ::pybind11;
class PreparedOp {
......@@ -81,6 +85,8 @@ class PreparedOp {
return PreparedOp(op, ctx, kernel_iter->second, dev_ctx);
}
inline platform::DeviceContext* GetDeviceContext() const { return dev_ctx; }
const framework::OperatorBase& op;
const framework::RuntimeContext& ctx;
framework::OperatorWithKernel::OpKernelFunc func;
......@@ -148,6 +154,9 @@ class VarBase {
framework::LoDTensor& GradValue();
std::unique_ptr<VarBase> NewVarBase(const platform::Place& dst_place,
const bool blocking) const;
inline std::string GradName() const {
PADDLE_ENFORCE(
var_desc_,
......@@ -175,11 +184,13 @@ class OpBase {
OpBase()
: op_desc_(nullptr),
forward_id_(-1),
grad_op_desc_(nullptr),
backward_id_(-1) {}
backward_id_(-1),
place_(platform::CPUPlace()) {}
virtual ~OpBase() {
if (grad_op_desc_) delete grad_op_desc_;
for (framework::OpDesc* desc : grad_op_descs_) {
delete desc;
}
}
std::map<std::string, std::vector<VarBase*>> ApplyGrad();
......@@ -188,18 +199,25 @@ class OpBase {
// For pure python PyLayer, use `forward_id_`, otherwise, use op_desc_.
framework::OpDesc* op_desc_;
int forward_id_;
// When has backward, one of `grad_op_desc_` or `backward_id_` is set,
// When has backward, one of `grad_op_descs_` or `backward_id_` is set,
// not both.
framework::OpDesc* grad_op_desc_;
// Note: each fwd op corresponds to a vector of bwd ops.
std::vector<framework::OpDesc*> grad_op_descs_;
int backward_id_;
platform::Place place_;
VarBasePtrMap input_vars_;
VarBasePtrMap output_vars_;
OpBasePtrMap pre_ops_;
std::map<std::string, std::vector<int>> pre_ops_out_idx_;
framework::VariableValueMap grad_input_vars_;
framework::VariableValueMap grad_output_vars_;
// Inputs to a vector of bwd ops.
std::vector<framework::VariableValueMap> grad_input_vars_;
// Outputs to a vector of bwd ops.
std::vector<framework::VariableValueMap> grad_output_vars_;
framework::BlockDesc* block_;
};
......
paddle/fluid/imperative/tracer.cc
浏览文件 @ ee3aae56
......@@ -14,33 +14,60 @@
#include "paddle/fluid/imperative/tracer.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace imperative {
void CreateGradOp(const framework::OpDesc& op_desc,
const std::unordered_set<std::string>& no_grad_set,
const std::vector<framework::BlockDesc*>& grad_sub_block,
framework::OpDesc** grad_op_desc,
std::vector<framework::OpDesc*>* grad_op_descs,
std::unordered_map<std::string, std::string>* grad_to_var) {
std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
PADDLE_ENFORCE(grad_op_descs->empty());
std::vector<std::unique_ptr<framework::OpDesc>> descs =
framework::OpInfoMap::Instance()
.Get(op_desc.Type())
.GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
// TODO(panyx0718): Leak?
*grad_op_desc = grad_op_descs[0].release();
for (auto& desc : descs) {
grad_op_descs->emplace_back(desc.release());
}
}
void InitVar(framework::Variable* var, framework::Variable* grad_var) {
void InitVar(framework::Variable* var, framework::Variable* grad_var,
platform::DeviceContext* dev_ctx) {
PADDLE_ENFORCE_NOT_NULL(dev_ctx,
"Could not get valid device from forward op");
auto& var_t = var->Get<framework::LoDTensor>();
float* data =
grad_var->GetMutable<framework::LoDTensor>()->mutable_data<float>(
var_t.dims(), platform::CPUPlace());
std::fill(data, data + var_t.numel(), 0.0);
grad_var->GetMutable<framework::LoDTensor>()->mutable_data<float>(
var_t.dims(), dev_ctx->GetPlace());
operators::math::set_constant(
*dev_ctx, grad_var->GetMutable<framework::LoDTensor>(), 0.0);
}
platform::Place GetExpectedPlace(platform::Place place, VarBasePtrMap inputs) {
platform::Place result = place;
for (auto it : inputs) {
for (VarBase* var : it.second) {
platform::Place tmp_place =
var->var_->Get<framework::LoDTensor>().place();
if (!platform::is_same_place(tmp_place, result)) {
PADDLE_THROW(
"Input variable should keep in the same place: %s, but get place: "
"%s of input %s instead",
result, tmp_place, it.first);
}
}
}
return result;
}
void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
const VarBasePtrMap& outputs, framework::BlockDesc* block,
const platform::Place expected_place,
const bool stop_gradient) {
std::map<std::string, VarBase*> vars;
......@@ -105,51 +132,59 @@ void Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
PADDLE_ENFORCE_NOT_NULL(op_kernel, "only support op with kernel");
framework::Scope scope;
platform::CPUPlace place;
PreparedOp p = PreparedOp::Prepare(ctx, *op_kernel, place);
p.op.RuntimeInferShape(scope, place, ctx);
p.func(framework::ExecutionContext(p.op, scope, *p.dev_ctx, p.ctx));
op->place_ = GetExpectedPlace(expected_place, inputs);
PreparedOp prepared_op = PreparedOp::Prepare(ctx, *op_kernel, op->place_);
prepared_op.op.RuntimeInferShape(scope, op->place_, ctx);
prepared_op.func(framework::ExecutionContext(
prepared_op.op, scope, *prepared_op.dev_ctx, prepared_op.ctx));
if (!stop_gradient) {
framework::OpDesc* grad_op_desc;
// TODO(panyx): Is this leaked?
std::unique_ptr<std::unordered_map<std::string, std::string>> grad_to_var(
new std::unordered_map<std::string, std::string>());
CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var.get());
op->grad_op_desc_ = grad_op_desc;
for (auto it : grad_op_desc->Inputs()) {
auto& grad_in_vars = op->grad_input_vars_[it.first];
for (const std::string& grad_invar : it.second) {
block->FindRecursiveOrCreateVar(grad_invar);
auto var_it = grad_to_var->find(grad_invar);
if (var_it == grad_to_var->end()) {
auto fwd_var_it = vars.find(grad_invar);
PADDLE_ENFORCE(fwd_var_it != vars.end());
// Forward inputs or outputs.
grad_in_vars.push_back(fwd_var_it->second->var_);
} else {
VarBase* var = vars[var_it->second];
if (!var->grads_->var_->IsInitialized()) {
InitVar(var->var_, var->grads_->var_);
CreateGradOp(*op_desc, {}, {block}, &op->grad_op_descs_, grad_to_var.get());
op->grad_input_vars_.resize(op->grad_op_descs_.size());
op->grad_output_vars_.resize(op->grad_op_descs_.size());
for (size_t i = 0; i < op->grad_op_descs_.size(); ++i) {
framework::OpDesc* grad_op_desc = op->grad_op_descs_[i];
for (auto it : grad_op_desc->Inputs()) {
auto& grad_in_vars = op->grad_input_vars_[i][it.first];
for (const std::string& grad_invar : it.second) {
block->FindRecursiveOrCreateVar(grad_invar);
auto var_it = grad_to_var->find(grad_invar);
if (var_it == grad_to_var->end()) {
auto fwd_var_it = vars.find(grad_invar);
PADDLE_ENFORCE(fwd_var_it != vars.end());
// Forward inputs or outputs.
grad_in_vars.push_back(fwd_var_it->second->var_);
} else {
VarBase* var = vars[var_it->second];
if (!var->grads_->var_->IsInitialized()) {
InitVar(var->var_, var->grads_->var_,
prepared_op.GetDeviceContext());
}
// Douts.
grad_in_vars.push_back(var->grads_->var_);
}
// Douts.
grad_in_vars.push_back(var->grads_->var_);
}
}
}
for (auto it : grad_op_desc->Outputs()) {
auto& grad_out_vars = op->grad_output_vars_[it.first];
for (const std::string& grad_outvar : it.second) {
block->FindRecursiveOrCreateVar(grad_outvar);
auto var_it = grad_to_var->find(grad_outvar);
PADDLE_ENFORCE(var_it != grad_to_var->end());
VarBase* var = vars[var_it->second];
if (!var->grads_->var_->IsInitialized()) {
InitVar(var->var_, var->grads_->var_);
for (auto it : grad_op_desc->Outputs()) {
auto& grad_out_vars = op->grad_output_vars_[i][it.first];
for (const std::string& grad_outvar : it.second) {
block->FindRecursiveOrCreateVar(grad_outvar);
auto var_it = grad_to_var->find(grad_outvar);
PADDLE_ENFORCE(var_it != grad_to_var->end(),
"Could not found the grad op output var, should this "
"operator %s's stop gradient be True",
op_desc->Type());
VarBase* var = vars[var_it->second];
if (!var->grads_->var_->IsInitialized()) {
InitVar(var->var_, var->grads_->var_,
prepared_op.GetDeviceContext());
}
grad_out_vars.push_back(var->grads_->var_);
}
grad_out_vars.push_back(var->grads_->var_);
}
}
}
......@@ -178,10 +213,12 @@ std::vector<VarBase*> Tracer::PyTrace(OpBase* op,
out->TrackPreOp(op, PyLayer::kFwdOut, i, stop_gradient);
}
if (!stop_gradient) {
op->grad_input_vars_.resize(1);
op->grad_output_vars_.resize(1);
auto& grad_input_vars =
op->grad_input_vars_[framework::GradVarName(PyLayer::kFwdInp)];
op->grad_input_vars_[0][framework::GradVarName(PyLayer::kFwdInp)];
auto& grad_output_vars =
op->grad_output_vars_[framework::GradVarName(PyLayer::kFwdOut)];
op->grad_output_vars_[0][framework::GradVarName(PyLayer::kFwdOut)];
for (const VarBase* inp : inputs) {
grad_input_vars.push_back(inp->var_);
......@@ -189,16 +226,23 @@ std::vector<VarBase*> Tracer::PyTrace(OpBase* op,
for (VarBase* out : outputs) {
grad_input_vars.push_back(out->var_);
}
platform::CPUPlace place;
for (VarBase* out : outputs) {
grad_input_vars.push_back(out->grads_->var_);
if (!grad_input_vars.back()->IsInitialized()) {
InitVar(out->var_, grad_input_vars.back());
// TODO(minqiyang): Add GPU support for PyLayer, only support CPU now
InitVar(out->var_, grad_input_vars.back(),
platform::DeviceContextPool::Instance().Get(place));
}
}
for (const VarBase* inp : inputs) {
grad_output_vars.push_back(inp->grads_->var_);
if (!grad_output_vars.back()->IsInitialized()) {
InitVar(inp->var_, grad_output_vars.back());
// TODO(minqiyang): Add GPU support for PyLayer, only support CPU now
InitVar(inp->var_, grad_output_vars.back(),
platform::DeviceContextPool::Instance().Get(place));
}
}
}
......
paddle/fluid/imperative/tracer.h
浏览文件 @ ee3aae56
......@@ -22,6 +22,7 @@
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/imperative/engine.h"
#include "paddle/fluid/imperative/layer.h"
#include "paddle/fluid/platform/place.h"
namespace paddle {
namespace imperative {
......@@ -34,21 +35,25 @@ void CreateGradOp(const framework::OpDesc& op_desc,
void InitVar(framework::Variable* var, framework::Variable* grad_var);
platform::Place GetExpectedPlace(platform::Place place, VarBasePtrMap inputs);
class Tracer {
public:
explicit Tracer(framework::BlockDesc* root_block) : root_block_(root_block) {}
virtual ~Tracer() {}
void Trace(OpBase* op,
const std::map<std::string, std::vector<VarBase*>>& inputs,
const std::map<std::string, std::vector<VarBase*>>& outputs,
framework::BlockDesc* block, const bool stop_gradient = false);
void Trace(OpBase* op, const VarBasePtrMap& inputs,
const VarBasePtrMap& outputs, framework::BlockDesc* block,
const platform::Place expected_place,
const bool stop_gradient = false);
std::vector<VarBase*> PyTrace(OpBase* op, const std::vector<VarBase*>& inputs,
bool stop_gradient = false);
private:
platform::Place GetPlace(const VarBasePtrMap& inputs);
framework::BlockDesc* root_block_;
};
......
paddle/fluid/inference/analysis/argument.h
浏览文件 @ ee3aae56
......@@ -28,6 +28,7 @@
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/api/paddle_analysis_config.h"
#include "paddle/fluid/platform/variant.h"
namespace paddle {
......@@ -130,10 +131,14 @@ struct Argument {
DECL_ARGUMENT_FIELD(tensorrt_max_batch_size, TensorRtMaxBatchSize, int);
DECL_ARGUMENT_FIELD(tensorrt_workspace_size, TensorRtWorkspaceSize, int);
DECL_ARGUMENT_FIELD(tensorrt_min_subgraph_size, TensorRtMinSubgraphSize, int);
DECL_ARGUMENT_FIELD(tensorrt_precision_mode, TensorRtPrecisionMode,
contrib::AnalysisConfig::Precision);
// Memory optimized related.
DECL_ARGUMENT_FIELD(enable_memory_optim, EnableMemoryOptim, bool);
DECL_ARGUMENT_FIELD(memory_optim_force_update, MemoryOptimForceUpdate, bool);
DECL_ARGUMENT_FIELD(static_memory_optim, StaticMemoryOptim, bool);
DECL_ARGUMENT_FIELD(static_memory_optim_force_update,
StaticMemoryOptimForceUpdate, bool);
// Indicate which kind of sort algorithm is used for operators, the memory
// optimization relays on the sort algorithm.
DECL_ARGUMENT_FIELD(memory_optim_sort_kind, MemoryOptimSortKind, int);
......
paddle/fluid/inference/analysis/helper.cc
浏览文件 @ ee3aae56
......@@ -36,6 +36,14 @@ void SetAttr<int>(framework::proto::OpDesc *op, const std::string &name,
attr->set_i(data);
}
template <>
void SetAttr<bool>(framework::proto::OpDesc *op, const std::string &name,
const bool &data) {
auto *attr = op->add_attrs();
attr->set_name(name);
attr->set_type(paddle::framework::proto::AttrType::BOOLEAN);
attr->set_b(data);
}
template <>
void SetAttr<int64_t>(framework::proto::OpDesc *op, const std::string &name,
const int64_t &data) {
auto *attr = op->add_attrs();
......
paddle/fluid/inference/analysis/helper.h
浏览文件 @ ee3aae56
......@@ -17,6 +17,7 @@ limitations under the License. */
#include <sys/stat.h>
#include <cstdio>
#include <fstream>
#include <set>
#include <string>
#include <typeindex>
#include <unordered_map>
......@@ -29,9 +30,14 @@ limitations under the License. */
#include "paddle/fluid/platform/port.h"
#ifdef _WIN32
#include <direct.h>
#include <io.h>
#define GCC_ATTRIBUTE(attr__) ;
#define MKDIR(path) _mkdir(path)
#else
#include <unistd.h>
#define GCC_ATTRIBUTE(attr__) __attribute__((attr__));
#define MKDIR(path) mkdir(path, S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH)
#endif
#define __SHOULD_USE_RESULT__ GCC_ATTRIBUTE(warn_unused_result)
......@@ -163,6 +169,54 @@ static bool PathExists(const std::string &path) {
return false;
}
static std::string GetDirRoot(const std::string &path) {
char sep = '/';
#ifdef _WIN32
sep = '\\';
#endif
size_t i = path.rfind(sep, path.length());
if (i != std::string::npos) {
return (path.substr(0, i));
}
return path;
}
static std::string GetOrCreateModelOptCacheDir(const std::string &model_root) {
std::string opt_cache_dir = model_root + "/_opt_cache/";
if (!PathExists(opt_cache_dir)) {
PADDLE_ENFORCE(MKDIR(opt_cache_dir.c_str()) != -1,
"Can not create optimize cache directory: %s, Make sure you "
"have permission to write",
opt_cache_dir);
}
return opt_cache_dir;
}
static std::string GetTrtCalibPath(const std::string &model_root,
const std::string &engine_key) {
return model_root + "/trt_calib_" + engine_key;
}
// If there is no calib table data file in model_opt_cache_dir, return "".
static std::string GetTrtCalibTableData(const std::string &model_opt_cache_dir,
const std::string &engine_key,
bool enable_int8) {
std::string trt_calib_table_path =
GetTrtCalibPath(model_opt_cache_dir, engine_key);
if (enable_int8 && FileExists(trt_calib_table_path)) {
VLOG(3) << "Calibration table file: " << trt_calib_table_path
<< "is found here";
std::ifstream infile(trt_calib_table_path, std::ios::in);
std::stringstream buffer;
buffer << infile.rdbuf();
std::string calibration_data(buffer.str());
return calibration_data;
}
return "";
}
} // namespace analysis
} // namespace inference
} // namespace paddle
......
paddle/fluid/inference/analysis/ir_pass_manager.cc
浏览文件 @ ee3aae56
......@@ -67,6 +67,20 @@ void IRPassManager::CreatePasses(Argument *argument,
pass->Set("max_batch_size", new int(argument->tensorrt_max_batch_size()));
pass->Set("min_subgraph_size",
new int(argument->tensorrt_min_subgraph_size()));
pass->Set("program",
new framework::ProgramDesc *(&argument->main_program()));
bool enable_int8 = argument->tensorrt_precision_mode() ==
contrib::AnalysisConfig::Precision::kInt8;
pass->Set("enable_int8", new bool(enable_int8));
std::string model_opt_cache_dir =
argument->Has("model_dir")
? argument->model_dir()
: GetDirRoot(argument->model_program_path());
pass->Set(
"model_opt_cache_dir",
new std::string(GetOrCreateModelOptCacheDir(model_opt_cache_dir)));
}
// graph_ = pass->Apply(std::move(graph_));
......@@ -91,11 +105,14 @@ std::unique_ptr<Graph> IRPassManager::Apply(std::unique_ptr<Graph> graph) {
}
framework::proto::ProgramDesc IRPassManager::AcquireProgram(
std::unique_ptr<Graph> *graph, const ProgramDesc &program) const {
std::unique_ptr<Graph> *graph, ProgramDesc *program) const {
auto pass =
framework::ir::PassRegistry::Instance().Get("graph_to_program_pass");
ProgramDesc desc(program);
// Direct using ProgramDesc desc(argument->main_program()) may cause
// incomplete copies of information.
ProgramDesc desc;
desc.CopyFrom(*program->Proto());
pass->SetNotOwned("program", &desc);
auto *the_graph = graph->release();
*graph = pass->Apply(std::unique_ptr<Graph>(the_graph));
......
paddle/fluid/inference/analysis/ir_pass_manager.h
浏览文件 @ ee3aae56
......@@ -29,6 +29,7 @@
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/analysis/argument.h"
#include "paddle/fluid/inference/analysis/helper.h"
namespace paddle {
namespace inference {
......@@ -42,8 +43,8 @@ class IRPassManager final {
std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph);
framework::proto::ProgramDesc AcquireProgram(
std::unique_ptr<Graph> *graph, const ProgramDesc &program) const;
framework::proto::ProgramDesc AcquireProgram(std::unique_ptr<Graph> *graph,
ProgramDesc *program) const;
framework::ir::Graph &graph() const { return *graph_; }
......
paddle/fluid/inference/analysis/ir_passes/tensorrt_subgraph_pass.cc
浏览文件 @ ee3aae56
......@@ -13,6 +13,7 @@
// limitations under the License.
#include <algorithm>
#include <set>
#include <string>
#include <vector>
......@@ -67,12 +68,33 @@ std::unique_ptr<framework::ir::Graph> analysis::TensorRtSubgraphPass::ApplyImpl(
return graph;
}
std::string GenerateEngineKey(const std::set<std::string> &engine_inputs,
const std::set<std::string> &engine_outputs) {
std::string engine_hash_key = "";
for (auto name : engine_inputs) {
engine_hash_key += name;
}
for (auto name : engine_outputs) {
engine_hash_key += name;
}
auto engine_key = std::to_string(std::hash<std::string>()(engine_hash_key));
return engine_key;
}
void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
Graph *graph) const {
auto *op_desc = node->Op();
auto &subgraph = *Agent(node).subgraph();
PADDLE_ENFORCE(!subgraph.empty());
framework::ProgramDesc *program_desc =
Get<framework::ProgramDesc *>("program");
// Add new block for TensorRTEngineOP
const framework::BlockDesc &main_block =
program_desc->Block(framework::kRootBlockIndex);
// const framework::BlockDesc& main_block = program_desc->Block(0);
framework::BlockDesc *new_block = program_desc->AppendBlock(main_block);
// An fake block desc.
framework::proto::BlockDesc block_proto;
framework::BlockDesc block_desc(nullptr, &block_proto);
......@@ -82,13 +104,18 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
subgraph.size());
for (auto *node : subgraph) {
auto *new_block_op = new_block->AppendOp();
auto *op = block_desc.AppendOp();
*new_block_op->Proto() = *node->Op()->Proto();
*op->Proto() = *node->Op()->Proto();
}
// collect inputs
std::unordered_set<std::string> input_names;
std::unordered_set<std::string> input_names_with_id;
// Then, we will use the input_names_with_id and output_names_with_id to
// generate the eigine key.
// So, We use set instead of unordered_set here to ensure that the engine key
// is unique.
std::set<std::string> input_names;
std::set<std::string> input_names_with_id;
for (auto *x : node->inputs) {
input_names.insert(x->Name());
input_names_with_id.insert(x->Name() + std::to_string(x->id()));
......@@ -96,8 +123,8 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
op_desc->SetInput(
"Xs", std::vector<std::string>(input_names.begin(), input_names.end()));
std::unordered_set<std::string> output_names;
std::unordered_set<std::string> output_names_with_id;
std::set<std::string> output_names;
std::set<std::string> output_names_with_id;
for (auto *x : node->outputs) {
output_names.insert(x->Name());
output_names_with_id.insert(x->Name() + std::to_string(x->id()));
......@@ -182,7 +209,6 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
// to Tensor.
std::vector<std::string> output_mapping;
for (auto name : output_names) {
// LOG(INFO) << name << " " << output_name_map.size();
PADDLE_ENFORCE(output_name_map.count(name) != 0);
output_mapping.push_back(output_name_map[name]);
}
......@@ -193,16 +219,29 @@ void TensorRtSubgraphPass::CreateTensorRTOp(framework::ir::Node *node,
*vars->Add() = *node->Var()->Proto();
}
}
PADDLE_ENFORCE(!block_desc.Proto()->vars().empty(),
"the block has no var-desc");
PADDLE_ENFORCE(!output_mapping.empty());
// Set attrs
op_desc->SetBlockAttr("sub_block", new_block);
SetAttr(op_desc->Proto(), "subgraph",
block_desc.Proto()->SerializeAsString());
// Set attrs
SetAttr(op_desc->Proto(), "max_batch_size", Get<int>("max_batch_size"));
SetAttr(op_desc->Proto(), "workspace_size", Get<int>("workspace_size"));
SetAttr(op_desc->Proto(), "parameters", ExtractParameters(graph->Nodes()));
SetAttr(op_desc->Proto(), "output_name_mapping", output_mapping);
auto enable_int8 = Get<bool>("enable_int8");
auto engine_key =
GenerateEngineKey(input_names_with_id, output_names_with_id);
std::string calibration_data = GetTrtCalibTableData(
Get<std::string>("model_opt_cache_dir"), engine_key, enable_int8);
SetAttr(op_desc->Proto(), "calibration_data", calibration_data);
SetAttr(op_desc->Proto(), "enable_int8", enable_int8);
SetAttr(op_desc->Proto(), "engine_key", engine_key);
}
std::vector<std::string> ExtractParameters(
......
paddle/fluid/inference/analysis/passes/CMakeLists.txt
浏览文件 @ ee3aae56
cc_library(ir_graph_build_pass SRCS ir_graph_build_pass.cc DEPS analysis_pass argument ir_pass_manager)
cc_library(ir_analysis_pass SRCS ir_analysis_pass.cc DEPS analysis_pass argument ir_pass_manager)
cc_library(memory_optim_pass SRCS memory_optimize_pass.cc DEPS analysis_pass)
cc_library(memory_optim_pass SRCS memory_optimize_pass.cc DEPS analysis_pass zero_copy_tensor)
cc_library(ir_params_sync_among_devices_pass SRCS ir_params_sync_among_devices_pass.cc DEPS analysis_pass argument ir_pass_manager)
cc_library(ir_graph_to_program_pass SRCS ir_graph_to_program_pass.cc DEPS analysis_pass graph_to_program_pass)
......
paddle/fluid/inference/analysis/passes/ir_graph_to_program_pass.cc
浏览文件 @ ee3aae56
......@@ -31,7 +31,11 @@ void IrGraphToProgramPass::RunImpl(Argument *argument) {
}
std::unique_ptr<Graph> graph(argument->main_graph_ptr());
framework::ProgramDesc desc(argument->main_program());
// Direct using ProgramDesc desc(argument->main_program()) may cause
// incomplete copies of information.
framework::ProgramDesc desc;
desc.CopyFrom(*argument->main_program().Proto());
pass->SetNotOwned("program", &desc);
auto thegraph = pass->Apply(std::move(graph));
thegraph.release(); // the argument still own the graph.
......
paddle/fluid/inference/analysis/passes/memory_optimize_pass.cc
浏览文件 @ ee3aae56
......@@ -444,6 +444,26 @@ std::vector<std::map<std::string, std::vector<int>>> DeseralizeBatchVarShapes(
return batch_shapes;
}
// Replace the -1 in shape to a real number to fake the shape.
std::vector<std::map<std::string, std::vector<int>>> FakeBatchVarShapes(
const framework::ProgramDesc& program) {
std::vector<std::map<std::string, std::vector<int>>> res;
res.emplace_back();
auto& record = res.front();
const int fake_batch_size = 3;
for (auto* var : program.Block(0).AllVars()) {
if (var->GetType() ==
framework::proto::VarType::Type::VarType_Type_LOD_TENSOR) {
auto shape = var->GetShape();
for (auto& v : shape) {
if (v < 0) v = fake_batch_size;
}
record[var->Name()].assign(shape.begin(), shape.end());
}
}
return res;
}
// Calculate the average dim of each tensor from the batch shape cache.
std::unordered_map<std::string, size_t> GetBatchAverageSize(
const std::vector<std::map<std::string, std::vector<int>>>& batches) {
......@@ -478,6 +498,7 @@ std::vector<std::unordered_set<std::string>> AnalysisBatchShapesByBatchSize(
std::unordered_map<std::string, std::stringstream> var_batchsize_hashes;
for (auto& batch : batches) {
for (auto& ele : batch) {
PADDLE_ENFORCE(!ele.second.empty());
int batch_size = ele.second.front();
// TODO(Superjomn) might consume large memory here, use combine hash.
var_batchsize_hashes[ele.first] << batch_size;
......@@ -538,9 +559,21 @@ std::vector<std::unordered_set<std::string>> AnalysisBatchShapesBySimilarSize(
std::string MemoryOptimizePass::repr() const { return "memory optimize pass"; }
std::pair<size_t, size_t> GetRange(
const std::unordered_map<std::string, size_t>& ave_size) {
auto res = std::make_pair(std::numeric_limits<size_t>::max(),
std::numeric_limits<size_t>::min());
for (auto& item : ave_size) {
res.first = std::min(item.second, res.first);
res.second = std::max(item.second, res.second);
}
return res;
}
void MemoryOptimizePass::RunImpl(Argument* argument) {
// When force update, should not optimize memory.
if (!argument->enable_memory_optim() || argument->memory_optim_force_update())
if (!argument->enable_memory_optim() ||
argument->static_memory_optim_force_update())
return;
graph_ = argument->main_graph_ptr();
......@@ -549,21 +582,38 @@ void MemoryOptimizePass::RunImpl(Argument* argument) {
argument->model_program_path_valid() ? argument->model_program_path()
: "");
VLOG(3) << "Load memory cache from " << path;
if (inference::IsFileExists(path)) {
VLOG(4) << "Performing memory optimize";
auto batches = DeseralizeBatchVarShapes(path);
auto var_batch_ave_size = GetBatchAverageSize(batches);
std::vector<std::map<std::string, std::vector<int>>> batches;
if (argument->static_memory_optim() && inference::IsFileExists(path)) {
string::PrettyLogInfo("--- Performing static memory optimize");
batches = DeseralizeBatchVarShapes(path);
} else {
string::PrettyLogInfo("--- Performing dynamic memory optimize");
batches = FakeBatchVarShapes(argument->main_program());
}
auto var_batch_ave_size = GetBatchAverageSize(batches);
// Get min and max memory size.
const auto range = GetRange(var_batch_ave_size);
const int cluster_size = std::max(
static_cast<int>((range.second - range.first) / 100 /*cluster num*/),
1024);
const int cluster_size1 = std::max(
static_cast<int>((range.second - range.first) / 1000 /*cluster num*/),
1024);
std::unordered_map<std::string, Node*> tensor_nodes;
space_table_t space_table;
CollectVarMemorySize(var_batch_ave_size, &tensor_nodes, &space_table);
std::unordered_map<std::string, Node*> tensor_nodes;
space_table_t space_table;
CollectVarMemorySize(var_batch_ave_size, &tensor_nodes, &space_table);
std::unordered_map<std::string, std::string> reuse_table;
double max_saving_ratio = 0.;
std::unordered_map<std::string, std::string> reuse_table;
double max_saving_ratio = 0.;
std::vector<std::function<MemoryAllocation()>> strategies;
std::vector<std::function<MemoryAllocation()>> strategies;
for (int sort_kind = 0; sort_kind < 2; sort_kind++) {
for (int sort_kind = 0; sort_kind < 2; sort_kind++) {
if (argument->static_memory_optim()) {
// This strategy only make scene in static memory optimize.
strategies.emplace_back([&, sort_kind] {
auto clustered_vars_by_batch_size =
AnalysisBatchShapesByBatchSize(batches);
......@@ -572,71 +622,67 @@ void MemoryOptimizePass::RunImpl(Argument* argument) {
space_table, &reuse_table, sort_kind, &allocation);
return allocation;
});
}
strategies.emplace_back([&, sort_kind] {
auto clustered_vars_by_ave_size = AnalysisBatchShapesBySimilarSize(
space_table, batches, 1024); // interval 1kb
MemoryAllocation allocation;
MakeReusePlan(clustered_vars_by_ave_size, var_batch_ave_size,
space_table, &reuse_table, sort_kind, &allocation);
return allocation;
});
strategies.emplace_back([&, sort_kind] {
auto clustered_vars_by_ave_size =
AnalysisBatchShapesBySimilarSize(space_table, batches, cluster_size);
MemoryAllocation allocation;
MakeReusePlan(clustered_vars_by_ave_size, var_batch_ave_size, space_table,
&reuse_table, sort_kind, &allocation);
return allocation;
});
strategies.emplace_back([&, sort_kind] {
auto clustered_vars_by_ave_size =
AnalysisBatchShapesBySimilarSize(space_table, batches, cluster_size1);
MemoryAllocation allocation;
MakeReusePlan(clustered_vars_by_ave_size, var_batch_ave_size, space_table,
&reuse_table, sort_kind, &allocation);
return allocation;
});
strategies.emplace_back([&, sort_kind] {
auto clustered_vars_by_ave_size = AnalysisBatchShapesBySimilarSize(
space_table, batches,
std::numeric_limits<int>::max()); // no intervals
MemoryAllocation allocation;
MakeReusePlan(clustered_vars_by_ave_size, var_batch_ave_size, space_table,
&reuse_table, sort_kind, &allocation);
return allocation;
});
}
strategies.emplace_back([&, sort_kind] {
auto clustered_vars_by_ave_size = AnalysisBatchShapesBySimilarSize(
space_table, batches, 1024 * 1024); // interval 1MB
MemoryAllocation allocation;
MakeReusePlan(clustered_vars_by_ave_size, var_batch_ave_size,
space_table, &reuse_table, sort_kind, &allocation);
return allocation;
});
std::function<MemoryAllocation()>* best_strategy{nullptr};
strategies.emplace_back([&, sort_kind] {
auto clustered_vars_by_ave_size = AnalysisBatchShapesBySimilarSize(
space_table, batches,
std::numeric_limits<int>::max()); // no intervals
MemoryAllocation allocation;
MakeReusePlan(clustered_vars_by_ave_size, var_batch_ave_size,
space_table, &reuse_table, sort_kind, &allocation);
return allocation;
});
// Try all strategies to get the best result.
for (auto& strategy : strategies) {
auto allocation = strategy();
string::PrettyLogDetail("--- get strategy saving %f memory for workspace",
allocation.GetSavingRatio());
if (allocation.GetSavingRatio() > max_saving_ratio) {
max_saving_ratio = allocation.GetSavingRatio();
best_strategy = &strategy;
}
}
if (!best_strategy) {
LOG(ERROR) << "This model makes poor memory optimize, skip memory optimize";
return;
}
auto memory_allocation = (*best_strategy)();
std::function<MemoryAllocation()>* best_strategy{nullptr};
string::PrettyLogInfo(
"--- Saved %.2f%s memory for workspace(temporary variables)",
memory_allocation.GetSavingRatio() * 100, "%");
// Try all strategies to get the best result.
for (auto& strategy : strategies) {
auto allocation = strategy();
string::PrettyLogDetail("--- get strategy saving %f memory for workspace",
allocation.GetSavingRatio());
if (allocation.GetSavingRatio() > max_saving_ratio) {
max_saving_ratio = allocation.GetSavingRatio();
best_strategy = &strategy;
}
}
if (!best_strategy) {
LOG(ERROR)
<< "This model makes poor memory optimize, skip memory optimize";
return;
}
auto memory_allocation = (*best_strategy)();
string::PrettyLogH2(
"--- Saved %.2f%s memory for workspace(temporary variables)",
memory_allocation.GetSavingRatio() * 100, "%");
string::PrettyLogDetail("--- Allocated %d MB",
memory_allocation.allocated / 1024. / 1024.);
string::PrettyLogDetail("--- Saved %d MB",
memory_allocation.saved / 1024. / 1024.);
argument->main_graph().Set(framework::ir::kGraphToProgramVarsToRemove,
new std::unordered_set<std::string>);
auto& vars2remove =
argument->main_graph().Get<std::unordered_set<std::string>>(
framework::ir::kGraphToProgramVarsToRemove);
PerformReusePlan(reuse_table, memory_allocation.sort_kind, &vars2remove);
argument->SetMemoryOptimSortKind(memory_allocation.sort_kind);
}
argument->main_graph().Set(framework::ir::kGraphToProgramVarsToRemove,
new std::unordered_set<std::string>);
auto& vars2remove =
argument->main_graph().Get<std::unordered_set<std::string>>(
framework::ir::kGraphToProgramVarsToRemove);
PerformReusePlan(reuse_table, memory_allocation.sort_kind, &vars2remove);
argument->SetMemoryOptimSortKind(memory_allocation.sort_kind);
}
float MemoryOptimizePass::MemoryAllocation::GetSavingRatio() const {
......
paddle/fluid/inference/analysis/passes/memory_optimize_pass.h
浏览文件 @ ee3aae56
......@@ -15,7 +15,7 @@
#pragma once
#include "paddle/fluid/inference/analysis/analysis_pass.h"
#include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
#include "paddle/fluid/platform/port.h"
namespace paddle {
namespace inference {
......
paddle/fluid/inference/api/analysis_config.cc
浏览文件 @ ee3aae56
......@@ -95,12 +95,14 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) {
CP_MEMBER(memory_pool_init_size_mb_);
CP_MEMBER(enable_memory_optim_);
CP_MEMBER(memory_optim_force_update_);
CP_MEMBER(static_memory_optim_);
CP_MEMBER(static_memory_optim_force_update_);
// TensorRT releated.
CP_MEMBER(use_tensorrt_);
CP_MEMBER(tensorrt_workspace_size_);
CP_MEMBER(tensorrt_max_batchsize_);
CP_MEMBER(tensorrt_min_subgraph_size_);
CP_MEMBER(tensorrt_precision_mode_);
// MKLDNN releated.
CP_MEMBER(use_mkldnn_);
CP_MEMBER(mkldnn_enabled_op_types_);
......@@ -140,9 +142,9 @@ void contrib::AnalysisConfig::EnableMKLDNN() {
Update();
}
void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size,
int max_batch_size,
int min_subgraph_size) {
void contrib::AnalysisConfig::EnableTensorRtEngine(
int workspace_size, int max_batch_size, int min_subgraph_size,
contrib::AnalysisConfig::Precision precision_mode) {
#ifdef PADDLE_WITH_CUDA
if (!use_gpu()) {
LOG(ERROR) << "To use TensorRT engine, please call EnableGpu() first";
......@@ -153,6 +155,7 @@ void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size,
tensorrt_workspace_size_ = workspace_size;
tensorrt_max_batchsize_ = max_batch_size;
tensorrt_min_subgraph_size_ = min_subgraph_size;
tensorrt_precision_mode_ = precision_mode;
Update();
#else
......@@ -238,7 +241,8 @@ std::string contrib::AnalysisConfig::SerializeInfoCache() {
ss << tensorrt_min_subgraph_size_;
ss << enable_memory_optim_;
ss << memory_optim_force_update_;
ss << static_memory_optim_;
ss << static_memory_optim_force_update_;
ss << use_mkldnn_;
for (auto &item : mkldnn_enabled_op_types_) ss << item;
......@@ -278,9 +282,11 @@ float contrib::AnalysisConfig::fraction_of_gpu_memory_for_pool() const {
#endif
}
void contrib::AnalysisConfig::EnableMemoryOptim(bool force_update_cache) {
void contrib::AnalysisConfig::EnableMemoryOptim(
bool static_optim, bool force_update_static_cache) {
enable_memory_optim_ = true;
memory_optim_force_update_ = force_update_cache;
static_memory_optim_ = static_optim;
static_memory_optim_force_update_ = force_update_static_cache;
Update();
}
......@@ -300,4 +306,16 @@ void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer,
Update();
}
NativeConfig contrib::AnalysisConfig::ToNativeConfig() const {
NativeConfig config;
config.model_dir = model_dir_;
config.prog_file = prog_file_;
config.param_file = params_file_;
config.use_gpu = use_gpu_;
config.device = device_id_;
config.fraction_of_gpu_memory = fraction_of_gpu_memory_for_pool();
config.specify_input_name = specify_input_name_;
return config;
}
} // namespace paddle
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @ ee3aae56
......@@ -15,6 +15,7 @@
#include "paddle/fluid/inference/api/analysis_predictor.h"
#include <glog/logging.h>
#include <algorithm>
#include <fstream>
#include <memory>
#include <string>
#include <vector>
......@@ -25,6 +26,7 @@
#include "paddle/fluid/framework/naive_executor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/var_type_traits.h"
#include "paddle/fluid/inference/analysis/helper.h"
#include "paddle/fluid/inference/analysis/passes/memory_optimize_pass.h"
#include "paddle/fluid/inference/api/helper.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
......@@ -37,6 +39,8 @@
#if PADDLE_WITH_TENSORRT
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
#endif
DECLARE_bool(profile);
......@@ -44,6 +48,12 @@ DECLARE_bool(profile);
namespace paddle {
using contrib::AnalysisConfig;
using inference::Singleton;
#if PADDLE_WITH_TENSORRT
using inference::tensorrt::TRTInt8Calibrator;
using inference::tensorrt::TRTCalibratorEngine;
using inference::tensorrt::TRTCalibratorEngineManager;
#endif
namespace {
bool IsPersistable(const framework::VarDesc *var) {
......@@ -298,15 +308,15 @@ void AnalysisPredictor::GetFetchOne(const framework::LoDTensor &fetch,
bool AnalysisPredictor::GetFetch(std::vector<PaddleTensor> *outputs,
framework::Scope *scope) {
VLOG(3) << "Predictor::get_fetch";
outputs->resize(fetchs_.size());
for (size_t i = 0; i < fetchs_.size(); ++i) {
int idx = boost::get<int>(fetchs_[i]->GetAttr("col"));
outputs->resize(fetches_.size());
for (size_t i = 0; i < fetches_.size(); ++i) {
int idx = boost::get<int>(fetches_[i]->GetAttr("col"));
PADDLE_ENFORCE((size_t)idx == i);
framework::LoDTensor &fetch =
framework::GetFetchVariable(*scope, "fetch", idx);
auto type = fetch.type();
auto output = &(outputs->at(i));
output->name = fetchs_[idx]->Input("X")[0];
output->name = fetches_[idx]->Input("X")[0];
if (type == framework::proto::VarType::FP32) {
GetFetchOne<float>(fetch, output);
output->dtype = PaddleDType::FLOAT32;
......@@ -327,7 +337,9 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
argument_.SetUseGPU(config_.use_gpu());
argument_.SetGPUDeviceId(config_.gpu_device_id());
argument_.SetEnableMemoryOptim(config_.enable_memory_optim());
argument_.SetMemoryOptimForceUpdate(config_.memory_optim_force_update_);
argument_.SetStaticMemoryOptim(config_.static_memory_optim_);
argument_.SetStaticMemoryOptimForceUpdate(
config_.static_memory_optim_force_update_);
argument_.SetModelFromMemory(config_.model_from_memory_);
// Analyze inference_program
if (!config_.model_dir().empty()) {
......@@ -337,6 +349,8 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
!config_.params_file().empty(),
"Either model_dir or (param_file, prog_file) should be set.");
PADDLE_ENFORCE(!config_.prog_file().empty());
std::string dir = inference::analysis::GetDirRoot(config_.prog_file());
argument_.SetModelProgramPath(config_.prog_file());
argument_.SetModelParamsPath(config_.params_file());
}
......@@ -347,6 +361,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
argument_.SetTensorRtWorkspaceSize(config_.tensorrt_workspace_size_);
argument_.SetTensorRtMaxBatchSize(config_.tensorrt_max_batchsize_);
argument_.SetTensorRtMinSubgraphSize(config_.tensorrt_min_subgraph_size_);
argument_.SetTensorRtPrecisionMode(config_.tensorrt_precision_mode_);
}
if (config_.use_mkldnn_) {
......@@ -422,10 +437,10 @@ void AnalysisPredictor::PrepareFeedFetch() {
feed_names_[op->Output("Out")[0]] = idx;
} else if (op->Type() == "fetch") {
int idx = boost::get<int>(op->GetAttr("col"));
if (fetchs_.size() <= static_cast<size_t>(idx)) {
fetchs_.resize(idx + 1);
if (fetches_.size() <= static_cast<size_t>(idx)) {
fetches_.resize(idx + 1);
}
fetchs_[idx] = op;
fetches_[idx] = op;
}
}
}
......@@ -567,7 +582,67 @@ bool AnalysisPredictor::LoadParameters() {
return true;
}
#if PADDLE_WITH_TENSORRT
bool AnalysisPredictor::SaveTrtCalibToDisk() {
PADDLE_ENFORCE(config_.tensorrt_engine_enabled(),
"This func can be invoked only in trt mode");
auto &block = inference_program_->Block(0);
for (auto &op_desc : block.AllOps()) {
if (op_desc->Type() == "tensorrt_engine") {
std::string engine_name =
boost::get<std::string>(op_desc->GetAttr("engine_key"));
if (!Singleton<TRTCalibratorEngineManager>::Global().Has(engine_name)) {
LOG(ERROR) << "You should run the predictor(with trt) on the real data "
"to generate calibration info";
return false;
}
TRTCalibratorEngine *calib_engine =
Singleton<TRTCalibratorEngineManager>::Global().Get(engine_name);
LOG(INFO) << "Wait for calib threads done.";
calib_engine->calib_->waitAndSetDone();
LOG(INFO) << "Generating TRT Calibration table data, this may cost a lot "
"of time...";
calib_engine->thr_->join();
std::string calibration_table_data =
calib_engine->calib_->getCalibrationTableAsString();
if (calibration_table_data.empty()) {
LOG(ERROR) << "the calibration table is empty.";
return false;
}
std::string model_opt_cache_dir =
argument_.Has("model_dir")
? argument_.model_dir()
: inference::analysis::GetDirRoot(argument_.model_program_path());
std::string calibration_table_data_path =
inference::analysis::GetTrtCalibPath(
inference::analysis::GetOrCreateModelOptCacheDir(
model_opt_cache_dir),
engine_name);
std::ofstream ofile(calibration_table_data_path, std::ios::out);
LOG(INFO) << "Write Paddle-TRT INT8 calibration table data to file "
<< calibration_table_data_path;
ofile << calibration_table_data;
ofile.close();
}
}
// Free all calibrator resources.
Singleton<TRTCalibratorEngineManager>::Global().DeleteALL();
return true;
}
#endif
AnalysisPredictor::~AnalysisPredictor() {
#if PADDLE_WITH_TENSORRT
if (config_.tensorrt_engine_enabled() &&
config_.tensorrt_precision_mode_ == AnalysisConfig::Precision::kInt8 &&
Singleton<TRTCalibratorEngineManager>::Global().Has()) {
SaveTrtCalibToDisk();
}
#endif
if (FLAGS_profile) {
platform::DisableProfiler(platform::EventSortingKey::kTotal,
"./profile.log");
......@@ -638,12 +713,12 @@ bool AnalysisPredictor::need_collect_var_shapes_for_memory_optim() {
// check if the cache exists
if (!config_.enable_memory_optim()) {
need = false;
} else if (config_.enable_memory_optim() &&
} else if (config_.static_memory_optim_ &&
!inference::IsFileExists(inference::analysis::GetMemoryCachePath(
config_.model_dir(), config_.prog_file()))) {
need = true;
} else if (config_.enable_memory_optim() &&
config_.memory_optim_force_update_) {
} else if (config_.static_memory_optim_ &&
config_.static_memory_optim_force_update_) {
need = true;
}
......@@ -651,6 +726,10 @@ bool AnalysisPredictor::need_collect_var_shapes_for_memory_optim() {
return need;
}
std::string AnalysisPredictor::GetSeriazlizedProgram() const {
return inference_program_->Proto()->SerializeAsString();
}
template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<contrib::AnalysisConfig>(
const contrib::AnalysisConfig &config) {
......
paddle/fluid/inference/api/analysis_predictor.h
浏览文件 @ ee3aae56
......@@ -45,6 +45,7 @@ using contrib::AnalysisConfig;
class AnalysisPredictor : public PaddlePredictor {
public:
explicit AnalysisPredictor(const AnalysisConfig &config) : config_(config) {}
~AnalysisPredictor();
bool Init(const std::shared_ptr<framework::Scope> &parent_scope,
const std::shared_ptr<framework::ProgramDesc> &program = nullptr);
......@@ -74,6 +75,8 @@ class AnalysisPredictor : public PaddlePredictor {
void SetMkldnnThreadID(int tid);
std::string GetSeriazlizedProgram() const override;
protected:
// For memory optimization.
bool need_collect_var_shapes_for_memory_optim();
......@@ -95,7 +98,21 @@ class AnalysisPredictor : public PaddlePredictor {
template <typename T>
void GetFetchOne(const framework::LoDTensor &fetchs,
PaddleTensor *output_data);
~AnalysisPredictor();
#if PADDLE_WITH_TENSORRT
// When we use Paddle-TRT INT8 engine, we need to generate calibration table
// data first,
// the calibration table contains the range for each op's input and output,
// this whole process can be divided into several steps:
//
// 1. Builds a 32-bit engine, runs it on the calibration set, and records a
// histogram for each
// tensor of the distribution of activation values.
// 2. Builds a calibration table from the histograms.
//
// After step 2, we need to store the calibration table on disk
bool SaveTrtCalibToDisk();
#endif
// Some more detailed tests, they are made the friends of the predictor, so that
// the all the details can be tested.
......@@ -115,7 +132,7 @@ class AnalysisPredictor : public PaddlePredictor {
std::shared_ptr<framework::ProgramDesc> inference_program_;
std::vector<framework::OpDesc *> feeds_;
std::map<std::string, size_t> feed_names_;
std::vector<framework::OpDesc *> fetchs_;
std::vector<framework::OpDesc *> fetches_;
// Memory buffer for feed inputs. The temporary LoDTensor will cause serious
// concurrency problems, wrong results and memory leak, so cache them.
std::vector<framework::LoDTensor> feed_tensors_;
......
paddle/fluid/inference/api/analysis_predictor_tester.cc
浏览文件 @ ee3aae56
......@@ -215,6 +215,8 @@ TEST(AnalysisPredictor, memory_optim) {
{
// The first predictor help to cache the memory optimize strategy.
auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);
LOG(INFO) << "serialized program: " << predictor->GetSeriazlizedProgram();
ASSERT_FALSE(predictor->GetSeriazlizedProgram().empty());
// Run several times to check the parameters are not reused by mistake.
for (int i = 0; i < 5; i++) {
......
paddle/fluid/inference/api/api.cc
浏览文件 @ ee3aae56
......@@ -12,6 +12,8 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <sstream>
#include "paddle/fluid/framework/commit.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
......@@ -97,4 +99,12 @@ void PaddleBuf::Free() {
}
}
std::string get_version() {
std::stringstream ss;
ss << "version: " << framework::paddle_version() << "\n";
ss << "commit: " << framework::paddle_commit() << "\n";
ss << "branch: " << framework::paddle_compile_branch() << "\n";
return ss.str();
}
} // namespace paddle
paddle/fluid/inference/api/api_tester.cc
浏览文件 @ ee3aae56
......@@ -61,4 +61,10 @@ TEST(paddle_inference_api, demo) {
predictor->Run({}, &outputs);
}
TEST(paddle_inference_api, get_version) {
LOG(INFO) << "paddle version:\n" << get_version();
auto version = get_version();
ASSERT_FALSE(version.empty());
}
} // namespace paddle
paddle/fluid/inference/api/paddle_analysis_config.h
浏览文件 @ ee3aae56
......@@ -42,6 +42,10 @@ struct AnalysisConfig {
explicit AnalysisConfig(const std::string& model_dir);
explicit AnalysisConfig(const std::string& prog_file,
const std::string& params_file);
enum class Precision {
kFloat32 = 0,
kInt8,
};
/** Set model with a directory.
*/
......@@ -135,7 +139,8 @@ struct AnalysisConfig {
* subgraph is less than this, it will not transfer to TensorRT engine.
*/
void EnableTensorRtEngine(int workspace_size = 1 << 20,
int max_batch_size = 1, int min_subgraph_size = 3);
int max_batch_size = 1, int min_subgraph_size = 3,
Precision precision = Precision::kFloat32);
/** A boolean state telling whether the TensorRT engine is used.
*/
bool tensorrt_engine_enabled() const { return use_tensorrt_; }
......@@ -162,17 +167,7 @@ struct AnalysisConfig {
/** Transform the AnalysisConfig to NativeConfig.
*/
NativeConfig ToNativeConfig() const {
NativeConfig config;
config.model_dir = model_dir_;
config.prog_file = prog_file_;
config.param_file = params_file_;
config.use_gpu = use_gpu_;
config.device = device_id_;
config.fraction_of_gpu_memory = fraction_of_gpu_memory_for_pool();
config.specify_input_name = specify_input_name_;
return config;
}
NativeConfig ToNativeConfig() const;
/** Specify the operator type list to use MKLDNN acceleration.
* @param op_list the operator type list.
*/
......@@ -195,7 +190,8 @@ struct AnalysisConfig {
/** Turn on memory optimize
* NOTE still in development, will release latter.
*/
void EnableMemoryOptim(bool force_update_cache = false);
void EnableMemoryOptim(bool static_optim = false,
bool force_update_static_cache = false);
/** Tell whether the memory optimization is activated. */
bool enable_memory_optim() const;
......@@ -238,10 +234,12 @@ struct AnalysisConfig {
// We set this variable to control the minimum number of nodes in the
// subgraph, 3 as default value.
int tensorrt_min_subgraph_size_{3};
Precision tensorrt_precision_mode_;
// memory reuse related.
bool enable_memory_optim_{false};
bool memory_optim_force_update_{false};
bool static_memory_optim_{false};
bool static_memory_optim_force_update_{false};
bool use_mkldnn_{false};
std::unordered_set<std::string> mkldnn_enabled_op_types_;
......
paddle/fluid/inference/api/paddle_api.h
浏览文件 @ ee3aae56
......@@ -215,6 +215,14 @@ class PaddlePredictor {
*/
virtual ~PaddlePredictor() = default;
/** \brief Get the serialized model program that executes in inference phase.
* Its data type is ProgramDesc, which is a protobuf message.
*/
virtual std::string GetSeriazlizedProgram() const {
assert(false); // Force raise error.
return "NotImplemented";
};
/** The common configs for all the predictors.
*/
struct Config {
......@@ -288,4 +296,6 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
int PaddleDtypeSize(PaddleDType dtype);
std::string get_version();
} // namespace paddle
paddle/fluid/inference/api/paddle_pass_builder.h
浏览文件 @ ee3aae56
......@@ -154,13 +154,16 @@ class GpuPassStrategy : public PassStrategy {
public:
GpuPassStrategy() : PassStrategy({}) {
passes_.assign({
"infer_clean_graph_pass", //
"conv_affine_channel_fuse_pass", //
"conv_eltwiseadd_affine_channel_fuse_pass", //
"conv_bn_fuse_pass", //
"conv_elementwise_add_act_fuse_pass", //
"conv_elementwise_add2_act_fuse_pass", //
"conv_elementwise_add_fuse_pass", //
"infer_clean_graph_pass", //
"conv_affine_channel_fuse_pass", //
"conv_eltwiseadd_affine_channel_fuse_pass", //
"conv_bn_fuse_pass", //
#if CUDNN_VERSION >= 7100 // To run conv_fusion, the version of cudnn must be
// guaranteed at least v7
"conv_elementwise_add_act_fuse_pass", //
"conv_elementwise_add2_act_fuse_pass", //
"conv_elementwise_add_fuse_pass", //
#endif
});
for (int i = 6; i >= 3; i--) {
......
paddle/fluid/inference/tensorrt/CMakeLists.txt
浏览文件 @ ee3aae56
nv_library(tensorrt_engine SRCS engine.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context)
nv_library(tensorrt_engine SRCS engine.cc trt_int8_calibrator.cc DEPS ${GLOB_OPERATOR_DEPS} framework_proto device_context)
nv_library(tensorrt_op_teller SRCS op_teller.cc DEPS framework_proto)
nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)
......
paddle/fluid/inference/tensorrt/convert/test_op_converter.cc
浏览文件 @ ee3aae56
......@@ -29,9 +29,9 @@ TEST(OpConverter, ConvertBlock) {
// init trt engine
cudaStream_t stream_;
std::unique_ptr<TensorRTEngine> engine_;
engine_.reset(new TensorRTEngine(5, 1 << 15, &stream_));
engine_->InitNetwork();
PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
engine_.reset(new TensorRTEngine(5, 1 << 15, stream_));
engine_->InitNetwork();
engine_->DeclareInput("conv2d-X", nvinfer1::DataType::kFLOAT,
nvinfer1::Dims3(2, 5, 5));
......
paddle/fluid/inference/tensorrt/convert/ut_helper.h
浏览文件 @ ee3aae56
......@@ -78,11 +78,9 @@ class TRTConvertValidation {
scope_(scope),
if_add_batch_(if_add_batch),
max_batch_size_(max_batch_size) {
// create engine.
engine_.reset(new TensorRTEngine(max_batch_size, workspace_size, &stream_));
engine_->InitNetwork();
PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
engine_.reset(new TensorRTEngine(max_batch_size, workspace_size, stream_));
engine_->InitNetwork();
}
// Declare a Variable as input with random initialization.
......@@ -175,7 +173,7 @@ class TRTConvertValidation {
op_->Run(scope_, place);
// Execute TRT.
engine_->Execute(batch_size);
cudaStreamSynchronize(*engine_->stream());
cudaStreamSynchronize(engine_->stream());
ASSERT_FALSE(op_desc_->OutputArgumentNames().empty());
const size_t output_space_size = 3000;
......@@ -184,7 +182,7 @@ class TRTConvertValidation {
std::vector<float> fluid_out;
std::vector<float> trt_out(output_space_size);
engine_->GetOutputInCPU(output, &trt_out[0], output_space_size);
cudaStreamSynchronize(*engine_->stream());
cudaStreamSynchronize(engine_->stream());
auto* var = scope_.FindVar(output);
auto tensor = var->GetMutable<framework::LoDTensor>();
......
paddle/fluid/inference/tensorrt/engine.cc
浏览文件 @ ee3aae56
......@@ -42,14 +42,13 @@ void TensorRTEngine::Execute(int batch_size) {
PADDLE_ENFORCE(buf.device == DeviceType::GPU);
buffers.push_back(buf.buffer);
}
PADDLE_ENFORCE_NOT_NULL(stream_);
infer_context_->enqueue(batch_size, buffers.data(), *stream_, nullptr);
cudaStreamSynchronize(*stream_);
infer_context_->enqueue(batch_size, buffers.data(), stream_, nullptr);
cudaStreamSynchronize(stream_);
SetRuntimeBatch(batch_size);
}
TensorRTEngine::~TensorRTEngine() {
cudaStreamSynchronize(*stream_);
cudaStreamSynchronize(stream_);
// clean buffer
for (auto &buf : buffers_) {
if (buf.device == DeviceType::GPU && buf.buffer != nullptr) {
......@@ -70,6 +69,13 @@ void TensorRTEngine::FreezeNetwork() {
// build engine.
infer_builder_->setMaxBatchSize(max_batch_);
infer_builder_->setMaxWorkspaceSize(max_workspace_);
if (enable_int8_) {
infer_builder_->setInt8Mode(true);
PADDLE_ENFORCE(
calibrator_ != nullptr,
"The precision mode is 'INT8', the calibrator should not be nullptr");
infer_builder_->setInt8Calibrator(calibrator_);
}
infer_engine_.reset(infer_builder_->buildCudaEngine(*infer_network_));
PADDLE_ENFORCE(infer_engine_ != nullptr, "build cuda engine failed!");
......@@ -173,7 +179,7 @@ void TensorRTEngine::GetOutputInGPU(const std::string &name, void *dst,
auto &buf = buffer(name);
PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before");
PADDLE_ENFORCE_EQ(cudaMemcpyAsync(dst, buf.buffer, dst_size,
cudaMemcpyDeviceToDevice, *stream_),
cudaMemcpyDeviceToDevice, stream_),
0);
}
......@@ -194,7 +200,7 @@ void TensorRTEngine::GetOutputInCPU(const std::string &name, void *dst,
auto &buf = buffer(name);
PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before");
PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buf.buffer, dst_size,
cudaMemcpyDeviceToHost, *stream_));
cudaMemcpyDeviceToHost, stream_));
}
Buffer &TensorRTEngine::buffer(const std::string &name) {
......@@ -211,12 +217,11 @@ void TensorRTEngine::SetInputFromCPU(const std::string &name, const void *data,
auto &buf = buffer(name);
PADDLE_ENFORCE_NOT_NULL(buf.buffer);
PADDLE_ENFORCE_NOT_NULL(data);
PADDLE_ENFORCE_NOT_NULL(stream_);
PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small");
PADDLE_ENFORCE(buf.device == DeviceType::GPU);
buf.size = size;
PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size,
cudaMemcpyHostToDevice, *stream_));
cudaMemcpyHostToDevice, stream_));
}
void TensorRTEngine::SetInputFromGPU(const std::string &name, const void *data,
......@@ -227,7 +232,7 @@ void TensorRTEngine::SetInputFromGPU(const std::string &name, const void *data,
PADDLE_ENFORCE_LE(size, buf.max_size, "buffer is too small");
PADDLE_ENFORCE(buf.device == DeviceType::GPU);
PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(buf.buffer, data, size,
cudaMemcpyDeviceToDevice, *stream_));
cudaMemcpyDeviceToDevice, stream_));
}
void TensorRTEngine::SetITensor(const std::string &name,
......
paddle/fluid/inference/tensorrt/engine.h
浏览文件 @ ee3aae56
......@@ -23,12 +23,14 @@ limitations under the License. */
#include "paddle/fluid/inference/engine.h"
#include "paddle/fluid/inference/tensorrt/helper.h"
#include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
#include "paddle/fluid/inference/tensorrt/trt_int8_calibrator.h"
#include "paddle/fluid/inference/utils/singleton.h"
namespace paddle {
namespace inference {
namespace tensorrt {
class TRTInt8Calibrator;
/*
* TensorRT Engine.
*
......@@ -54,17 +56,17 @@ class TensorRTEngine : public EngineBase {
nvinfer1::Weights w_;
};
TensorRTEngine(int max_batch, int max_workspace,
cudaStream_t* stream = nullptr, int device = 0,
TensorRTEngine(int max_batch, int max_workspace, cudaStream_t stream,
int device = 0, bool enable_int8 = false,
TRTInt8Calibrator* calibrator = nullptr,
nvinfer1::ILogger& logger = NaiveLogger::Global())
: max_batch_(max_batch),
max_workspace_(max_workspace),
stream_(stream ? stream : &default_stream_),
logger_(logger),
device_(device) {
freshDeviceId();
cudaStreamCreate(stream_);
}
stream_(stream),
device_(device),
enable_int8_(enable_int8),
calibrator_(calibrator),
logger_(logger) {}
virtual ~TensorRTEngine();
......@@ -102,7 +104,7 @@ class TensorRTEngine : public EngineBase {
// NOTE this should be used after calling `FreezeNetwork`.
Buffer& buffer(const std::string& name) override;
cudaStream_t* stream() { return stream_; }
cudaStream_t stream() { return stream_; }
// Fill an input from CPU memory with name and size.
void SetInputFromCPU(const std::string& name, const void* data, size_t size);
......@@ -142,8 +144,8 @@ class TensorRTEngine : public EngineBase {
// In the normal case, the paddle-trt exists bug when runing the googlenet.
// When there are more than two convolutions of 1 * 1 with the same input, the
// paddle-tensorrt will do the merging optimization, which fuse those conv
// into
// one conv, and then trigger bug. So, We should use strategy to avoid this
// into one conv, and then trigger bug. So, We should use strategy to avoid
// this
// optimization for the time being. This bug will be fixed in the future.
std::unordered_map<std::string /*name*/, int /*ITensor_quote_num*/>
itensor_quote_num;
......@@ -156,11 +158,15 @@ class TensorRTEngine : public EngineBase {
// the max memory size the engine uses
int max_workspace_;
cudaStream_t stream_;
// The specific GPU id that the TensorRTEngine bounded to.
int device_;
bool enable_int8_;
TRTInt8Calibrator* calibrator_;
// batch size of the current data, will be updated each Executation.
int batch_size_{-1};
cudaStream_t* stream_;
// If stream_ is not set from outside, hold its own stream.
cudaStream_t default_stream_;
nvinfer1::ILogger& logger_;
std::vector<Buffer> buffers_;
......@@ -169,8 +175,6 @@ class TensorRTEngine : public EngineBase {
std::unordered_map<std::string /*name*/, nvinfer1::ITensor* /*ITensor*/>
itensor_map_;
// The specific GPU id that the TensorRTEngine bounded to.
int device_;
std::vector<std::unique_ptr<plugin::PluginTensorRT>> owned_plugin_;
// TensorRT related internal members
......@@ -208,38 +212,6 @@ class TensorRTEngine : public EngineBase {
#define TRT_ENGINE_ADD_LAYER(engine__, layer__, ARGS...) \
engine__->network()->add##layer__(ARGS);
/*
* Helper to control the TensorRT engine's creation and deletion.
*/
class TRT_EngineManager {
public:
bool HasEngine(const std::string& name) const {
return engines_.count(name) != 0;
}
// Get an engine called `name`.
TensorRTEngine* Get(const std::string& name) const {
return engines_.at(name).get();
}
// Create or get an engine called `name`
TensorRTEngine* Create(int max_batch, int max_workspace, cudaStream_t* stream,
const std::string& name, int gpu_device = 0) {
auto* p = new TensorRTEngine(max_batch, max_workspace, stream, gpu_device);
engines_[name].reset(p);
return p;
}
void DeleteALl() {
for (auto& item : engines_) {
item.second.reset(nullptr);
}
}
private:
std::unordered_map<std::string, std::unique_ptr<TensorRTEngine>> engines_;
};
} // namespace tensorrt
} // namespace inference
} // namespace paddle
paddle/fluid/inference/tensorrt/test_engine.cc
浏览文件 @ ee3aae56
......@@ -27,8 +27,8 @@ namespace tensorrt {
class TensorRTEngineTest : public ::testing::Test {
protected:
void SetUp() override {
// ASSERT_EQ(0, cudaStreamCreate(&stream_));
engine_ = new TensorRTEngine(10, 1 << 10, &stream_);
ASSERT_EQ(0, cudaStreamCreate(&stream_));
engine_ = new TensorRTEngine(10, 1 << 10, stream_);
engine_->InitNetwork();
}
......
paddle/fluid/inference/tensorrt/trt_int8_calibrator.cc 0 → 100644
浏览文件 @ ee3aae56
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// 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/fluid/inference/tensorrt/trt_int8_calibrator.h"
#include "glog/logging.h"
namespace paddle {
namespace inference {
namespace tensorrt {
// set the batch size before constructing the thread to execute engine
int TRTInt8Calibrator::getBatchSize() const { return batch_size_; }
TRTInt8Calibrator::TRTInt8Calibrator(
const std::unordered_map<std::string, size_t>& buffers, int batch_size,
std::string engine_name, const platform::Place place)
: batch_size_(batch_size), engine_name_(engine_name) {
int i = 0;
VLOG(4) << "Init a new calibrator: " << engine_name_;
for (const auto it : buffers) {
framework::Tensor temp_tensor;
std::string input_name = it.first;
int data_size = it.second;
int num_ele = data_size / sizeof(int16_t);
framework::DDim data_shape = framework::make_ddim({num_ele});
temp_tensor.Resize(data_shape);
data_tensors_.push_back(temp_tensor);
data_buffers_[input_name] = std::pair<void*, size_t>(
static_cast<void*>(temp_tensor.mutable_data<int16_t>(place)), num_ele);
i += 1;
}
}
TRTInt8Calibrator::TRTInt8Calibrator(const std::string& calib_data)
: batch_size_(0),
calib_running_(false),
data_is_set_(false),
done_(true),
calibration_table_(calib_data) {}
void TRTInt8Calibrator::waitAndSetDone() {
std::unique_lock<std::mutex> lk(mut_);
while ((calib_running_ || data_is_set_) && !done_) cond_.wait(lk);
if (!done_) {
done_ = true;
cond_.notify_all();
}
}
// There might be more than one input for trt subgraph,
// So, we use a map to store input information.
bool TRTInt8Calibrator::setBatch(
const std::unordered_map<std::string, void*>& data) {
VLOG(3) << "set batch: " << engine_name_;
std::unique_lock<std::mutex> lk(mut_);
// There is a producer and a consumer. The producer set the batch data and
// the consumer get the batch data. The size of the data pool is one.
// So, the producer has to wait for the consumer to finish processing before
// they can set the data.
while ((calib_running_ || data_is_set_) && (!done_)) cond_.wait(lk);
// The done_ is set to true using waitAndSetDone, When all calibration data
// are processed.
if (done_) return false;
// Sets the batch.
for (const auto& it : data) {
auto dataptr = data_buffers_.find(it.first);
if (dataptr == data_buffers_.end()) {
LOG(FATAL) << "FATAL " << engine_name_ << " input name '" << it.first
<< "' does not match with the buffer names";
}
const auto& d = dataptr->second;
PADDLE_ENFORCE(
cudaMemcpy(d.first, it.second, d.second, cudaMemcpyDeviceToDevice),
"Fail to cudaMemcpy %s for %s", engine_name_, it.first);
}
data_is_set_ = true;
cond_.notify_all();
return true;
}
bool TRTInt8Calibrator::getBatch(void** bindings, const char** names,
int num_bindings) {
VLOG(4) << "get batch: " << engine_name_;
std::unique_lock<std::mutex> lk(mut_);
// The consumer has just finished processing a data.
// The producer can set the data again.
calib_running_ = false;
cond_.notify_all();
// As long as there is data in the pool, the consumer can get it.
while (!data_is_set_ && !done_) cond_.wait(lk);
if (done_) return false;
// Gets the batch
for (int i = 0; i < num_bindings; i++) {
auto it = data_buffers_.find(names[i]);
if (it == data_buffers_.end()) {
LOG(FATAL) << "Calibration engine asked for unknown tensor name '"
<< names[i] << "' at position " << i;
}
bindings[i] = it->second.first;
}
data_is_set_ = false;
calib_running_ = true;
VLOG(4) << "get batch done: " << engine_name_;
return true;
}
void TRTInt8Calibrator::setDone() {
std::unique_lock<std::mutex> lk(mut_);
done_ = true;
cond_.notify_all();
}
const void* TRTInt8Calibrator::readCalibrationCache(size_t& length) {
if (calibration_table_.empty()) return nullptr;
length = calibration_table_.size();
return calibration_table_.data();
}
void TRTInt8Calibrator::writeCalibrationCache(const void* ptr,
std::size_t length) {
calibration_table_ = std::string((const char*)ptr, length);
VLOG(4) << "Got calibration data for " << engine_name_ << " " << ptr
<< " length=" << length;
}
TRTInt8Calibrator::~TRTInt8Calibrator() {
VLOG(4) << "Destroying calibrator for " << engine_name_;
}
} // namespace tensorrt
} // namespace inference
} // namespace paddle
paddle/fluid/inference/tensorrt/trt_int8_calibrator.h 0 → 100644
浏览文件 @ ee3aae56
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// 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 <atomic>
#include <memory>
#include <mutex>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
#include <NvInfer.h>
#include <cuda_runtime_api.h>
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/inference/tensorrt/engine.h"
#include "paddle/fluid/platform/place.h"
namespace paddle {
namespace inference {
namespace tensorrt {
class TensorRTEngine;
struct TRTInt8Calibrator : public nvinfer1::IInt8EntropyCalibrator {
public:
TRTInt8Calibrator(const std::unordered_map<std::string, size_t>& buffers,
int batch_size, std::string engine_name,
const platform::Place place);
explicit TRTInt8Calibrator(const std::string& calibration_data);
~TRTInt8Calibrator();
int getBatchSize() const override;
bool getBatch(void* bindings[], const char* names[],
int num_bindings) override;
bool setBatch(const std::unordered_map<std::string, void*>& data);
void setDone();
void waitAndSetDone();
const void* readCalibrationCache(std::size_t& length) override;
void writeCalibrationCache(const void* ptr, std::size_t length) override;
const std::string& getCalibrationTableAsString() {
return calibration_table_;
}
private:
const int batch_size_;
bool calib_running_{true};
bool data_is_set_{false};
bool done_{false};
std::mutex mut_;
std::condition_variable cond_;
std::unordered_map<std::string, std::pair<void*, size_t>> data_buffers_;
std::vector<framework::Tensor> data_tensors_;
std::string engine_name_;
std::string calibration_table_;
};
class TRTCalibratorEngine {
public:
TRTCalibratorEngine() {}
std::unique_ptr<TRTInt8Calibrator> calib_;
std::unique_ptr<std::thread> thr_;
std::unique_ptr<TensorRTEngine> engine_;
};
/*
* Manager to control the TensorRT Int8 calibration creation and deltetion.
*/
class TRTCalibratorEngineManager {
public:
bool Has() const { return res_.size() > 0; }
bool Has(const std::string& name) const {
if (res_.count(name) == 0) return false;
return res_.at(name).get() != nullptr;
}
// Get Int8Calibrator via name
TRTCalibratorEngine* Get(const std::string& name) const {
return res_.at(name).get();
}
// Look up or create a calibrator.
TRTCalibratorEngine* LookupOrCreate(const std::string& engine_name) {
if (res_.count(engine_name) == 0) {
auto* p = new TRTCalibratorEngine;
res_[engine_name].reset(p);
}
return res_.at(engine_name).get();
}
// Create an Int8Calibrator
TRTCalibratorEngine* Create(const std::string& engine_name) {
auto* p = new TRTCalibratorEngine;
res_[engine_name].reset(p);
return p;
}
void DeleteALL() {
for (auto& item : res_) {
item.second.reset(nullptr);
}
}
private:
std::unordered_map<std::string, std::unique_ptr<TRTCalibratorEngine>> res_;
};
} // namespace tensorrt
} // namespace inference
} // namespace paddle
paddle/fluid/inference/tests/api/CMakeLists.txt
浏览文件 @ ee3aae56
......@@ -54,6 +54,7 @@ else()
message(WARNING "These tests has been disabled in OSX or WITH_MKL=OFF before being fixed: \n test_analyzer_seq_pool1")
endif()
# RNN2
set(RNN2_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn2")
download_model_and_data(${RNN2_INSTALL_DIR} "rnn2_model.tar.gz" "rnn2_data.txt.tar.gz")
......@@ -115,6 +116,10 @@ if (NOT EXISTS ${MOBILENET_INSTALL_DIR})
endif()
inference_analysis_api_test_with_refer_result(test_analyzer_mobilenet_transpose ${MOBILENET_INSTALL_DIR} analyzer_vis_tester.cc SERIAL)
# googlenet
inference_analysis_api_test_with_fake_data(test_analyzer_googlenet
"${INFERENCE_DEMO_INSTALL_DIR}/googlenet" analyzer_resnet50_tester.cc "googlenet.tar.gz" SERIAL)
# resnet50
inference_analysis_api_test_with_fake_data(test_analyzer_resnet50
"${INFERENCE_DEMO_INSTALL_DIR}/resnet50" analyzer_resnet50_tester.cc "resnet50_model.tar.gz" SERIAL)
......
paddle/fluid/inference/tests/api/analyzer_dam_tester.cc
浏览文件 @ ee3aae56
......@@ -253,7 +253,7 @@ void compare(bool use_mkldnn = false) {
}
// Compare result of NativeConfig and AnalysisConfig with memory optimization.
TEST(Analyzer_dam, compare_with_memory_optim) {
TEST(Analyzer_dam, compare_with_static_memory_optim) {
// The small dam will core in CI, but works in local.
if (FLAGS_max_turn_num == 9) {
contrib::AnalysisConfig cfg, cfg1;
......@@ -263,7 +263,7 @@ TEST(Analyzer_dam, compare_with_memory_optim) {
SetInput(&input_slots_all);
// Run the first time to force to update memory cache
SetConfig(&cfg);
cfg.EnableMemoryOptim(true);
cfg.EnableMemoryOptim(true, true /*force update*/);
CompareNativeAndAnalysis(
reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
......@@ -271,7 +271,7 @@ TEST(Analyzer_dam, compare_with_memory_optim) {
// Run second time to use the memory cache and perform memory optimization.
SetConfig(&cfg1);
cfg1.EnableMemoryOptim();
cfg1.EnableMemoryOptim(true, false /*do not force update*/);
CompareNativeAndAnalysis(
reinterpret_cast<const PaddlePredictor::Config *>(&cfg1),
......@@ -279,6 +279,24 @@ TEST(Analyzer_dam, compare_with_memory_optim) {
}
}
TEST(Analyzer_dam, compare_with_dynamic_memory_optim) {
// The small dam will core in CI, but works in local.
if (FLAGS_max_turn_num == 9) {
contrib::AnalysisConfig cfg, cfg1;
DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
std::vector<std::vector<PaddleTensor>> input_slots_all;
SetInput(&input_slots_all);
// Run the first time to force to update memory cache
SetConfig(&cfg);
cfg.EnableMemoryOptim();
CompareNativeAndAnalysis(
reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
input_slots_all);
}
}
TEST(Analyzer_dam, compare) { compare(); }
#ifdef PADDLE_WITH_MKLDNN
......
paddle/fluid/inference/tests/api/tester_helper.h
浏览文件 @ ee3aae56
......@@ -56,6 +56,13 @@ DECLARE_int32(paddle_num_threads);
namespace paddle {
namespace inference {
float Random(float low, float high) {
static std::random_device rd;
static std::mt19937 mt(rd());
std::uniform_real_distribution<double> dist(low, high);
return dist(mt);
}
void PrintConfig(const PaddlePredictor::Config *config, bool use_analysis) {
const auto *analysis_config =
reinterpret_cast<const contrib::AnalysisConfig *>(config);
......@@ -176,7 +183,7 @@ void SetFakeImageInput(std::vector<std::vector<PaddleTensor>> *inputs,
float *input_data = static_cast<float *>(input.data.data());
// fill input data, for profile easily, do not use random data here.
for (size_t j = 0; j < len; ++j) {
*(input_data + j) = static_cast<float>(j) / len;
*(input_data + j) = Random(0.0, 1.0) / 10.;
}
}
(*inputs).emplace_back(input_slots);
......@@ -344,6 +351,16 @@ void CompareNativeAndAnalysis(
CompareResult(analysis_outputs, native_outputs);
}
void CompareNativeAndAnalysis(
PaddlePredictor *native_pred, PaddlePredictor *analysis_pred,
const std::vector<std::vector<PaddleTensor>> &inputs) {
int batch_size = FLAGS_batch_size;
std::vector<PaddleTensor> native_outputs, analysis_outputs;
native_pred->Run(inputs[0], &native_outputs, batch_size);
analysis_pred->Run(inputs[0], &analysis_outputs, batch_size);
CompareResult(analysis_outputs, native_outputs);
}
template <typename T>
std::string LoDTensorSummary(const framework::LoDTensor &tensor) {
std::stringstream ss;
......
paddle/fluid/inference/tests/api/trt_models_tester.cc
浏览文件 @ ee3aae56
......@@ -107,6 +107,27 @@ void compare(std::string model_dir, bool use_tensorrt) {
inputs_all);
}
void compare_continuous_input(std::string model_dir, bool use_tensorrt) {
contrib::AnalysisConfig analysis_config;
SetConfig<contrib::AnalysisConfig>(&analysis_config, model_dir, true,
use_tensorrt, FLAGS_batch_size);
auto config =
reinterpret_cast<const PaddlePredictor::Config*>(&analysis_config);
auto native_pred = CreateTestPredictor(config, false);
auto analysis_pred = CreateTestPredictor(config, true);
for (int i = 0; i < 100; i++) {
std::vector<std::vector<PaddleTensor>> inputs_all;
if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
SetFakeImageInput(&inputs_all, model_dir, true, FLAGS_prog_filename,
FLAGS_param_filename);
} else {
SetFakeImageInput(&inputs_all, model_dir, false, "__model__", "");
}
CompareNativeAndAnalysis(native_pred.get(), analysis_pred.get(),
inputs_all);
}
}
TEST(TensorRT_mobilenet, compare) {
std::string model_dir = FLAGS_infer_model + "/mobilenet";
compare(model_dir, /* use_tensorrt */ true);
......@@ -162,5 +183,15 @@ TEST(TensorRT_mobilenet, profile) {
profile(model_dir, true, false);
}
TEST(resnet50, compare_continuous_input) {
std::string model_dir = FLAGS_infer_model + "/resnet50";
compare_continuous_input(model_dir, true);
}
TEST(resnet50, compare_continuous_input_native) {
std::string model_dir = FLAGS_infer_model + "/resnet50";
compare_continuous_input(model_dir, false);
}
} // namespace inference
} // namespace paddle
paddle/fluid/memory/allocation/legacy_allocator.cc
浏览文件 @ ee3aae56
......@@ -14,6 +14,7 @@
#include "paddle/fluid/memory/allocation/legacy_allocator.h"
#include <string>
#include <utility>
#include <vector>
#include "glog/logging.h"
#include "paddle/fluid/memory/detail/buddy_allocator.h"
......@@ -37,7 +38,7 @@ template <typename Place>
void *Alloc(const Place &place, size_t size);
template <typename Place>
void Free(const Place &place, void *p);
void Free(const Place &place, void *p, size_t size);
template <typename Place>
size_t Used(const Place &place);
......@@ -52,6 +53,11 @@ size_t memory_usage(const platform::Place &p);
using BuddyAllocator = detail::BuddyAllocator;
std::unordered_map</*device id*/ int,
std::pair</*current memory usage*/ uint64_t,
/*peak memory usage*/ uint64_t>>
gpu_mem_info;
BuddyAllocator *GetCPUBuddyAllocator() {
// We tried thread_local for inference::RNN1 model, but that not works much
// for multi-thread test.
......@@ -98,7 +104,8 @@ void *Alloc<platform::CPUPlace>(const platform::CPUPlace &place, size_t size) {
}
template <>
void Free<platform::CPUPlace>(const platform::CPUPlace &place, void *p) {
void Free<platform::CPUPlace>(const platform::CPUPlace &place, void *p,
size_t size) {
VLOG(10) << "Free pointer=" << p << " on " << platform::Place(place);
GetCPUBuddyAllocator()->Free(p);
}
......@@ -177,9 +184,16 @@ void *Alloc<platform::CUDAPlace>(const platform::CUDAPlace &place,
LOG(WARNING) << "GPU memory used: "
<< string::HumanReadableSize(Used<platform::CUDAPlace>(place));
platform::SetDeviceId(cur_dev);
}
if (FLAGS_init_allocated_mem) {
cudaMemset(ptr, 0xEF, size);
} else {
gpu_mem_info[place.device].first += size;
if (gpu_mem_info[place.device].first > gpu_mem_info[place.device].second) {
gpu_mem_info[place.device].second = gpu_mem_info[place.device].first;
VLOG(3) << "device: " << place.device << " peak memory usage : "
<< (gpu_mem_info[place.device].second >> 20) << " MiB";
}
if (FLAGS_init_allocated_mem) {
cudaMemset(ptr, 0xEF, size);
}
}
return ptr;
#else
......@@ -188,9 +202,11 @@ void *Alloc<platform::CUDAPlace>(const platform::CUDAPlace &place,
}
template <>
void Free<platform::CUDAPlace>(const platform::CUDAPlace &place, void *p) {
void Free<platform::CUDAPlace>(const platform::CUDAPlace &place, void *p,
size_t size) {
#ifdef PADDLE_WITH_CUDA
GetGPUBuddyAllocator(place.device)->Free(p);
gpu_mem_info[place.device].first -= size;
#else
PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
#endif
......@@ -243,7 +259,7 @@ void *Alloc<platform::CUDAPinnedPlace>(const platform::CUDAPinnedPlace &place,
template <>
void Free<platform::CUDAPinnedPlace>(const platform::CUDAPinnedPlace &place,
void *p) {
void *p, size_t size) {
#ifdef PADDLE_WITH_CUDA
GetCUDAPinnedBuddyAllocator()->Free(p);
#else
......@@ -264,15 +280,17 @@ struct AllocVisitor : public boost::static_visitor<void *> {
};
struct FreeVisitor : public boost::static_visitor<void> {
inline explicit FreeVisitor(void *ptr) : ptr_(ptr) {}
inline explicit FreeVisitor(void *ptr, size_t size)
: ptr_(ptr), size_(size) {}
template <typename Place>
inline void operator()(const Place &place) const {
Free<Place>(place, ptr_);
Free<Place>(place, ptr_, size_);
}
private:
void *ptr_;
size_t size_;
};
size_t Usage::operator()(const platform::CPUPlace &cpu) const {
......@@ -304,8 +322,9 @@ Allocation *LegacyAllocator::AllocateImpl(size_t size, Allocator::Attr attr) {
}
void LegacyAllocator::Free(Allocation *allocation) {
boost::apply_visitor(legacy::FreeVisitor(allocation->ptr()),
allocation->place());
boost::apply_visitor(
legacy::FreeVisitor(allocation->ptr(), allocation->size()),
allocation->place());
delete allocation;
}
} // namespace allocation
......
paddle/fluid/operators/CMakeLists.txt
浏览文件 @ ee3aae56
......@@ -13,6 +13,7 @@ add_subdirectory(detection)
add_subdirectory(elementwise)
add_subdirectory(fused)
add_subdirectory(metrics)
add_subdirectory(ngraph)
add_subdirectory(optimizers)
add_subdirectory(reduce_ops)
add_subdirectory(sequence_ops)
......@@ -66,7 +67,7 @@ set(COMMON_OP_DEPS ${OP_HEADER_DEPS})
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} selected_rows_functor selected_rows lod_tensor maxouting unpooling pooling lod_rank_table context_project sequence_pooling executor)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dynload_warpctc)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_functor memory jit_kernel_helper concat_and_split cross_entropy softmax vol2col im2col sampler tree2col)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions beam_search)
if (WITH_GPU)
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} depthwise_conv prelu)
endif()
......@@ -86,7 +87,6 @@ set(GLOB_OPERATOR_DEPS ${OPERATOR_DEPS} CACHE INTERNAL "Global Op dependencies")
cc_test(gather_test SRCS gather_test.cc DEPS tensor)
cc_test(scatter_test SRCS scatter_test.cc DEPS tensor math_function)
cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
cc_test(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_search_op)
cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory)
cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op)
cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op)
......
paddle/fluid/operators/affine_channel_op.cu
浏览文件 @ ee3aae56
......@@ -83,7 +83,7 @@ __global__ void AffineChannelScaleBiasGradientCUDAKernel(
T* dbias) {
const int outer_size = C;
const int inner_size = N * HxW;
typedef cub::BlockReduce<T, BlockDim> BlockReduce;
typedef cub::BlockReduce<double, BlockDim> BlockReduce;
__shared__ typename BlockReduce::TempStorage ds_storage;
__shared__ typename BlockReduce::TempStorage db_storage;
......@@ -97,13 +97,16 @@ __global__ void AffineChannelScaleBiasGradientCUDAKernel(
ds_sum += dy[index] * x[index];
db_sum += dy[index];
}
ds_sum = BlockReduce(ds_storage).Reduce(ds_sum, cub::Sum());
db_sum = BlockReduce(db_storage).Reduce(db_sum, cub::Sum());
__syncthreads();
auto ds_out =
BlockReduce(ds_storage).Reduce(static_cast<double>(ds_sum), cub::Sum());
auto db_out =
BlockReduce(db_storage).Reduce(static_cast<double>(db_sum), cub::Sum());
__syncthreads();
if (threadIdx.x == 0) {
dscale[i] = ds_sum;
dbias[i] = db_sum;
dscale[i] = ds_out;
dbias[i] = db_out;
}
__syncthreads();
}
}
......
paddle/fluid/operators/beam_search_op.cc
浏览文件 @ ee3aae56
......@@ -12,205 +12,15 @@ 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 <algorithm>
#include <map>
#include "paddle/fluid/operators/beam_search_op.h"
#include <string>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/beam_search_op.h"
namespace paddle {
namespace operators {
void BeamSearch::operator()(const framework::LoDTensor &pre_ids,
const framework::LoDTensor &pre_scores,
framework::LoDTensor *selected_ids,
framework::LoDTensor *selected_scores) {
auto abs_lod = framework::ToAbsOffset(ids_->lod());
auto &high_level = abs_lod[lod_level_];
auto items = SelectTopBeamSizeItems(pre_ids, pre_scores);
auto selected_items = ToMap(items, high_level.back());
VLOG(3) << "selected_items:";
for (size_t i = 0; i < selected_items.size(); ++i) {
VLOG(3) << "offset:" << i;
for (auto &item : selected_items[i]) {
VLOG(3) << ItemToString(item);
}
}
PruneEndBeams(pre_ids, &selected_items);
// calculate the output tensor's height
size_t num_instances = std::accumulate(
std::begin(selected_items), std::end(selected_items), 0,
[](size_t a, std::vector<Item> &b) { return a + b.size(); });
// the output tensor shape should be [num_instances, 1]
auto dims = framework::make_ddim(
std::vector<int64_t>({static_cast<int>(num_instances), 1}));
selected_ids->Resize(dims);
selected_scores->Resize(dims);
std::map<size_t /*offset*/, std::vector<Item>> hash;
framework::LoD new_lod;
auto *ids_data = selected_ids->mutable_data<int64_t>(platform::CPUPlace());
auto *scores_data =
selected_scores->mutable_data<float>(platform::CPUPlace());
// fill in data
std::vector<size_t> low_level;
size_t low_offset = 0;
for (auto &items : selected_items) {
low_level.push_back(low_offset);
for (auto &item : items) {
ids_data[low_offset] = item.id;
scores_data[low_offset] = item.score;
low_offset++;
}
}
low_level.push_back(low_offset);
// fill lod
framework::LoD lod(2);
lod[0].assign(high_level.begin(), high_level.end());
lod[1].assign(low_level.begin(), low_level.end());
if (!framework::CheckLoD(lod)) {
PADDLE_THROW("lod %s is not right", framework::LoDToString(lod));
}
selected_ids->set_lod(lod);
selected_scores->set_lod(lod);
}
void BeamSearch::PruneEndBeams(const framework::LoDTensor &pre_ids,
std::vector<std::vector<Item>> *items) {
auto *pre_ids_data = pre_ids.data<int64_t>();
auto abs_lod = framework::ToAbsOffset(ids_->lod());
auto &high_level = abs_lod[lod_level_];
for (size_t src_idx = 0; src_idx < high_level.size() - 1; ++src_idx) {
size_t src_prefix_start = high_level[src_idx];
size_t src_prefix_end = high_level[src_idx + 1];
bool finish_flag = true;
for (size_t offset = src_prefix_start; offset < src_prefix_end; offset++) {
for (auto &item : items->at(offset)) {
if (item.id != static_cast<size_t>(end_id_) ||
pre_ids_data[offset] != end_id_) {
finish_flag = false;
break;
}
}
if (!finish_flag) break;
}
if (finish_flag) { // all branchs of the beam (source sentence) end and
// prune this beam
for (size_t offset = src_prefix_start; offset < src_prefix_end; offset++)
items->at(offset).clear();
}
}
}
std::vector<std::vector<BeamSearch::Item>> BeamSearch::ToMap(
const std::vector<std::vector<Item>> &items, size_t element_num) {
std::vector<std::vector<Item>> result;
result.resize(element_num);
for (auto &entries : items) {
for (const auto &item : entries) {
result[item.offset].push_back(item);
}
}
return result;
}
std::vector<std::vector<BeamSearch::Item>> BeamSearch::SelectTopBeamSizeItems(
const framework::LoDTensor &pre_ids,
const framework::LoDTensor &pre_scores) {
std::vector<std::vector<Item>> result;
std::vector<Item> items;
// for each source sentence, select the top beam_size items across all
// candidate sets.
while (NextItemSet(pre_ids, pre_scores, &items)) {
std::nth_element(
std::begin(items), std::begin(items) + beam_size_, std::end(items),
[](const Item &a, const Item &b) { return a.score > b.score; });
// prune the top beam_size items.
if (items.size() > beam_size_) {
items.resize(beam_size_);
}
result.emplace_back(items);
}
VLOG(3) << "SelectTopBeamSizeItems result size " << result.size();
for (auto &items : result) {
VLOG(3) << "item set:";
for (auto &item : items) {
VLOG(3) << ItemToString(item);
}
}
return result;
}
// the candidates of a source
bool BeamSearch::NextItemSet(const framework::LoDTensor &pre_ids,
const framework::LoDTensor &pre_scores,
std::vector<BeamSearch::Item> *items) {
if (sent_offset_ >= ids_->NumElements(lod_level_)) {
return false;
}
// find the current candidates
auto ids = *ids_;
auto scores = *scores_;
auto abs_lod = framework::ToAbsOffset(ids.lod());
auto *ids_data = ids.data<int64_t>();
auto *scores_data = scores.data<float>();
size_t instance_dim = 1;
for (int i = 1; i < ids.dims().size(); i++) {
instance_dim *= ids.dims()[i];
}
auto *pre_ids_data = pre_ids.data<int64_t>();
auto *pre_scores_data = pre_scores.data<float>();
items->clear();
items->reserve(framework::product(ids.dims()));
for (size_t offset = abs_lod[lod_level_][sent_offset_];
offset < abs_lod[lod_level_][sent_offset_ + 1]; offset++) {
auto pre_id = pre_ids_data[offset];
auto pre_score = pre_scores_data[offset];
if (pre_id == end_id_) {
// Allocate all probability mass to eos_id for finished branchs and the
// other candidate ids can be ignored.
items->emplace_back(offset, end_id_, pre_score);
} else {
for (size_t d = 0; d < instance_dim; d++) {
const size_t dim_offset = offset * instance_dim + d;
items->emplace_back(offset, ids_data[dim_offset],
scores_data[dim_offset]);
}
}
}
sent_offset_++;
return true;
}
std::ostream &operator<<(std::ostream &os, const BeamSearch::Item &item) {
os << "{";
os << "offset: " << item.offset << ", ";
os << "id: " << item.id << ", ";
os << "score: " << item.score << "";
os << "}";
return os;
}
std::string ItemToString(const BeamSearch::Item &item) {
std::ostringstream stream;
stream << item;
return stream.str();
}
class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
......@@ -219,18 +29,23 @@ class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker {
"(LoDTensor) The LoDTensor containing the selected ids at the "
"previous step. It should be a tensor with shape (batch_size, 1) "
"and lod `[[0, 1, ... , batch_size], [0, 1, ..., batch_size]]` at "
"thefirst step.");
"the first step.");
AddInput("pre_scores",
"(LoDTensor) The LoDTensor containing the accumulated "
"scores corresponding to the selected ids at the previous step.");
AddInput("ids",
"(LoDTensor) The LoDTensor containing the candidates ids. Its "
"shape should be (batch_size * beam_size, K), where K supposed to "
"be beam_size.");
"shape should be (batch_size * beam_size, W). If not set, it will "
"be calculated out according to Input(scores) in this operator.")
.AsDispensable();
AddInput("scores",
"(LoDTensor) The LodTensor containing the accumulated scores "
"corresponding to Input(ids) and its shape is the same as the "
"shape of Input(ids).");
"(LoDTensor) The LoDTensor containing the current scores "
"corresponding to Input(ids). If Input(ids) is not nullptr, its "
"shape is the same as that of Input(ids)."
"If is_accumulated is true, Input(scores) is accumulated scores "
"and will be used derectedly. Else, each score will be "
"transformed to the log field and accumulate Input(pre_sores) "
"first.");
AddOutput("selected_ids",
"A LodTensor that stores the IDs selected by beam search.");
AddOutput("selected_scores",
......@@ -242,6 +57,9 @@ class BeamSearchOpMaker : public framework::OpProtoAndCheckerMaker {
AddAttr<int>("beam_size", "beam size for beam search");
AddAttr<int>("end_id",
"the token id which indicates the end of a sequence");
AddAttr<bool>("is_accumulated",
"Whether the Input(scores) is accumulated scores.")
.SetDefault(true);
AddComment(R"DOC(
This operator does the search in beams for one time step.
......@@ -265,10 +83,9 @@ class BeamSearchOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext *ctx) const override {
for (const std::string &arg :
std::vector<std::string>({"pre_ids", "ids", "scores"})) {
std::vector<std::string>({"pre_ids", "scores"})) {
PADDLE_ENFORCE(ctx->HasInput(arg), "BeamSearch need input argument '%s'",
arg);
}
......@@ -279,12 +96,22 @@ class BeamSearchOp : public framework::OperatorWithKernel {
}
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext &ctx) const override {
framework::OpKernelType kt = framework::OpKernelType(
ctx.Input<framework::LoDTensor>("pre_ids")->type(),
platform::CPUPlace());
return kt;
auto *scores = ctx.Input<framework::LoDTensor>("scores");
size_t level = ctx.Attr<int>("level");
size_t batch_size = scores->lod()[level].size() - 1;
// The current CUDA kernel only support cases with batch_size < 4.
// Compute on CPU for cases with batch_size > 4.
if (batch_size <= 4) {
return framework::OpKernelType(
ctx.Input<framework::LoDTensor>("pre_ids")->type(), ctx.GetPlace());
} else {
return framework::OpKernelType(
ctx.Input<framework::LoDTensor>("pre_ids")->type(),
platform::CPUPlace());
}
}
};
......
paddle/fluid/operators/beam_search_op.cu.cc 0 → 100644
浏览文件 @ ee3aae56
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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/fluid/operators/beam_search_op.h"
#include "paddle/fluid/framework/op_registry.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
beam_search,
ops::BeamSearchOpKernel<paddle::platform::CUDADeviceContext, float>,
ops::BeamSearchOpKernel<paddle::platform::CUDADeviceContext, double>,
ops::BeamSearchOpKernel<paddle::platform::CUDADeviceContext, int>,
ops::BeamSearchOpKernel<paddle::platform::CUDADeviceContext, int64_t>);
paddle/fluid/operators/beam_search_op.h
浏览文件 @ ee3aae56
......@@ -4,7 +4,7 @@ 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
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,
......@@ -14,187 +14,12 @@ limitations under the License. */
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/operators/math/beam_search.h"
namespace paddle {
namespace operators {
/*
* This is an implementation of beam search.
*
* To explain the details, lets take machine translation task for example, in
* this task, one source sentence is translated to multiple target sentences,
* during this period, one sentence will be translated to multiple translation
* prefixes(target sentence that have not ended), in each time step a prefix
* will have some candidates, input the candidate ids and their corresponding
* scores (probabilities), it will sort and select the top beam_size candidates
* for each source sentence, and store the selected candidates's score and their
* corresponding ids to LoDTensors.
*
* A detailed example:
*
* Input
*
* ids:
* LoD (should have 2 levels)
* first level: [0, 1, 4]
* second level: [0, 1, 2, 3, 4]
*
* tensor's data
* [
* [4, 2, 5]
* [2, 1, 3]
* [3, 5, 2]
* [8, 2, 1]
* ]
*
* scores:
* LoD same as `ids`
* tensor's data
* [
* [0.5, 0.3, 0.2]
* [0.6, 0.3, 0.1]
* [0.9, 0.5, 0.1]
* [0.7, 0.5, 0.1]
* ]
*
* the inputs means that there are 2 source sentences to translate, and the
* first source has 1 prefix, the second source has 2 prefix.
*
* lets assume beam size is 2, and the beam search's output should be
* LoD
* first level:
* [0, 1, 2]
* second level:
* [0, 2, 4]
*
* id tensor's data
* [[
* 4,
* 1,
* 3,
* 8,
* ]]
*
* score tensor's data
* [[
* 0.5,
* 0.3,
* 0.9,
* 0.7
* ]]
*
* TODO all the prune operations should be in the beam search, so it is better
* to split the beam search algorithm into a sequence of smaller operators, and
* the prune operators can be inserted in this sequence.
*/
class BeamSearch {
public:
// TODO(superjom) make type customizable
using id_t = size_t;
using score_t = float;
/*
* Input the arguments that needed by this class.
*/
BeamSearch(const framework::LoDTensor& ids,
const framework::LoDTensor& scores, size_t level, size_t beam_size,
int end_id)
: beam_size_(beam_size),
ids_(&ids),
scores_(&scores),
lod_level_(level),
end_id_(end_id) {}
/*
* The main function of beam search.
*
* @selected_ids: a [None, 1]-shaped tensor with LoD.
* In a machine translation model, it might be the candidate term id sets,
* each set stored as a varience-length sequence.
* The format might be described with a two-level LoD
* - [[0 1]
* - [0 1 2]]
* - [[]
* - [0 1]]
* the first level of LoD tells that there are two source sentences. The
* second level describes the details of the candidate id set's offsets in
* the
* source sentences.
*
* @selected_scores: a LoD tensor with the same shape and LoD with
* selected_ids.
* It stores the corresponding scores of candidate ids in selected_ids.
*
* Return false if all the input tensor is empty, in machine translation task
* that means no candidates is provided, and the task will stop running.
*/
void operator()(const framework::LoDTensor& pre_ids,
const framework::LoDTensor& pre_scores,
framework::LoDTensor* selected_ids,
framework::LoDTensor* selected_scores);
/*
* The basic items help to sort.
*/
struct Item {
Item() {}
Item(size_t offset, size_t id, float score)
: offset(offset), id(id), score(score) {}
// offset in the higher lod level.
size_t offset;
// // prefix id in the lower lod level.
// size_t prefix;
// the candidate id
id_t id;
// the corresponding score
score_t score;
};
protected:
/*
* Prune the source sentences all branchs finished, and it is optional.
* Pruning must one step later than finishing (thus pre_ids is needed here),
* since the end tokens must be writed out.
*/
void PruneEndBeams(const framework::LoDTensor& pre_ids,
std::vector<std::vector<Item>>* items);
/*
* Transform the items into a map whose key is offset, value is the items.
* NOTE low performance.
*/
std::vector<std::vector<Item>> ToMap(
const std::vector<std::vector<Item>>& inputs, size_t element_num);
/*
* For each source, select top beam_size records.
*/
std::vector<std::vector<Item>> SelectTopBeamSizeItems(
const framework::LoDTensor& pre_ids,
const framework::LoDTensor& pre_scores);
/*
* Get the items of next source sequence, return false if no remaining items.
*/
bool NextItemSet(const framework::LoDTensor& pre_ids,
const framework::LoDTensor& pre_scores,
std::vector<Item>* items);
private:
size_t beam_size_;
const framework::LoDTensor* ids_;
const framework::LoDTensor* scores_;
size_t lod_level_{0};
size_t sent_offset_{0};
int end_id_{0};
};
std::ostream& operator<<(std::ostream& os, const BeamSearch::Item& item);
std::string ItemToString(const BeamSearch::Item& item);
template <typename DeviceContext, typename T>
class BeamSearchOpKernel : public framework::OpKernel<T> {
public:
......@@ -203,7 +28,7 @@ class BeamSearchOpKernel : public framework::OpKernel<T> {
auto* scores = context.Input<framework::LoDTensor>("scores");
auto* pre_ids = context.Input<framework::LoDTensor>("pre_ids");
auto* pre_scores = context.Input<framework::LoDTensor>("pre_scores");
PADDLE_ENFORCE_NOT_NULL(ids);
PADDLE_ENFORCE_NOT_NULL(scores);
PADDLE_ENFORCE_NOT_NULL(pre_ids);
PADDLE_ENFORCE_NOT_NULL(pre_scores);
......@@ -211,14 +36,20 @@ class BeamSearchOpKernel : public framework::OpKernel<T> {
size_t level = context.Attr<int>("level");
size_t beam_size = context.Attr<int>("beam_size");
int end_id = context.Attr<int>("end_id");
BeamSearch alg(*ids, *scores, level, beam_size, end_id);
bool is_accumulated = context.Attr<bool>("is_accumulated");
auto selected_ids = context.Output<framework::LoDTensor>("selected_ids");
auto selected_scores =
context.Output<framework::LoDTensor>("selected_scores");
PADDLE_ENFORCE_NOT_NULL(selected_ids);
PADDLE_ENFORCE_NOT_NULL(selected_scores);
alg(*pre_ids, *pre_scores, selected_ids, selected_scores);
math::BeamSearchFunctor<DeviceContext, T> alg;
alg(context.template device_context<DeviceContext>(), pre_ids, pre_scores,
ids, scores, selected_ids, selected_scores, level, beam_size, end_id,
is_accumulated);
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/beam_search_op_test.cc 已删除 100644 → 0
浏览文件 @ d6d3e6af
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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/fluid/operators/beam_search_op.h"
#include <gtest/gtest.h>
#include <vector>
namespace paddle {
namespace test {
using std::vector;
using framework::LoDTensor;
using framework::LoD;
using operators::BeamSearch;
using paddle::platform::CPUPlace;
using std::cout;
using std::endl;
void CreateInput(LoDTensor* ids, LoDTensor* scores) {
LoD lod;
vector<size_t> level0({0, 2, 4});
vector<size_t> level1({0, 1, 2, 3, 4});
lod.push_back(level0);
lod.push_back(level1);
ids->set_lod(lod);
scores->set_lod(lod);
auto dims = framework::make_ddim(vector<int64_t>({4, 3}));
ids->Resize(dims);
scores->Resize(dims);
CPUPlace place;
auto* ids_data = ids->mutable_data<int64_t>(place);
auto* scores_data = scores->mutable_data<float>(place);
vector<int64_t> _ids({4, 2, 5, 2, 1, 3, 3, 5, 2, 8, 2, 1});
vector<float> _scores(
{0.5f, 0.3f, 0.2f, 0.6f, 0.3f, 0.1f, 0.9f, 0.5f, 0.1f, 0.7f, 0.5f, 0.1f});
for (int i = 0; i < 12; i++) {
ids_data[i] = _ids[i];
scores_data[i] = _scores[i];
}
}
// It seems that beam_search_op has bugs.
TEST(DISABLED_beam_search_op, run) {
CPUPlace place;
LoDTensor ids, scores;
CreateInput(&ids, &scores);
LoDTensor pre_ids;
pre_ids.Resize(framework::make_ddim(vector<int64_t>(4, 1)));
for (int i = 0; i < 4; i++) {
pre_ids.mutable_data<int64_t>(place)[i] = i + 1;
}
LoDTensor pre_scores;
pre_scores.Resize(framework::make_ddim(vector<int64_t>(4, 1)));
for (int i = 0; i < 4; i++) {
pre_scores.mutable_data<float>(place)[i] = 0.1 * (i + 1);
}
BeamSearch beamsearch(ids, scores, (size_t)0, (size_t)2, 0);
LoDTensor sids, sscores;
beamsearch(pre_ids, pre_scores, &sids, &sscores);
LOG(INFO) << "score: " << sscores << endl;
ASSERT_EQ(sids.lod(), sscores.lod());
vector<int> tids({4, 2, 3, 8});
vector<float> tscores({0.5f, 0.6f, 0.9f, 0.7f});
for (int i = 0; i < 4; i++) {
ASSERT_EQ(tids[i], sids.data<int64_t>()[i]);
ASSERT_EQ(tscores[i], sscores.data<float>()[i]);
}
}
} // namespace test
} // namespace paddle
paddle/fluid/operators/bpr_loss_op.h
浏览文件 @ ee3aae56
......@@ -87,8 +87,8 @@ class BprLossGradientOpKernel : public framework::OpKernel<T> {
auto* label = ctx.Input<Tensor>("Label");
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
const int step_size = x->dims()[0];
const int num_classes = x->dims()[1];
const size_t step_size = static_cast<size_t>(x->dims()[0]);
const size_t num_classes = static_cast<size_t>(x->dims()[1]);
T* dx_data = dx->mutable_data<T>(ctx.GetPlace());
const T* dy_data = dy->data<T>();
const T* x_data = x->data<T>();
......
paddle/fluid/operators/detection/CMakeLists.txt
浏览文件 @ ee3aae56
......@@ -45,3 +45,7 @@ detection_library(roi_perspective_transform_op SRCS roi_perspective_transform_op
foreach(src ${LOCAL_DETECTION_LIBS})
set(OP_LIBRARY ${src} ${OP_LIBRARY} CACHE INTERNAL "op libs")
endforeach()
cc_library(mask_util SRCS mask_util.cc DEPS memory)
cc_test(mask_util_test SRCS mask_util_test.cc DEPS memory mask_util)
detection_library(generate_mask_labels_op SRCS generate_mask_labels_op.cc DEPS mask_util)
paddle/fluid/operators/detection/bbox_util.h
浏览文件 @ ee3aae56
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
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 <algorithm>
#include "paddle/fluid/framework/eigen.h"
......@@ -88,7 +92,9 @@ void BboxOverlaps(const framework::Tensor& r_boxes,
inter_w = std::max(x_max - x_min + 1, zero);
inter_h = std::max(y_max - y_min + 1, zero);
inter_area = inter_w * inter_h;
overlaps_et(i, j) = inter_area / (r_box_area + c_box_area - inter_area);
overlaps_et(i, j) =
(inter_area == 0.) ? 0 : inter_area /
(r_box_area + c_box_area - inter_area);
}
}
}
......
paddle/fluid/operators/detection/mask_util.cc 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/detection/mask_util.h 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/detection/mask_util_test.cc 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/gather_op.cc
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/math/beam_search.cc 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/math/beam_search.cu 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/math/beam_search.h 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/math/beam_search_test.cc 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/nce_op.h
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/ngraph/CMakeLists.txt 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/ngraph/ngraph_engine.h 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/ngraph/ngraph_engine_op.cc 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/ngraph/ngraph_engine_op.h 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/shuffle_channel_op.cc 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/shuffle_channel_op.cu 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/operators/shuffle_channel_op.h 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/platform/gpu_info.cc
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/pybind/imperative.cc
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/pybind/inference_api.cc 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/pybind/inference_api.h 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/pybind/ir.cc
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/pybind/protobuf.cc
浏览文件 @ ee3aae56
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ ee3aae56
此差异已折叠。
paddle/py_paddle/.gitignore 已删除 100644 → 0
浏览文件 @ d6d3e6af
此差异已折叠。
paddle/py_paddle/dataprovider_converter.py 已删除 100644 → 0
浏览文件 @ d6d3e6af
此差异已折叠。
paddle/py_paddle/util.py 已删除 100644 → 0
浏览文件 @ d6d3e6af
此差异已折叠。
paddle/scripts/README.md
浏览文件 @ ee3aae56
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ ee3aae56
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/compiler.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/contrib/int8_inference/__init__.py 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/contrib/int8_inference/utility.py 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/contrib/reader/README.md 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/contrib/tests/test_calibration.py 0 → 100644
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/data_feeder.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/initializer.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ ee3aae56
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ ee3aae56
此差异已折叠。
python/setup.py.in
浏览文件 @ ee3aae56
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