From defd7ec6412e0c9d4a5761a9500f22f5b58cf438 Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Thu, 16 Nov 2017 23:35:01 +0800 Subject: [PATCH 1/4] mkldnn only need one trainer --- paddle/trainer/Trainer.cpp | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/paddle/trainer/Trainer.cpp b/paddle/trainer/Trainer.cpp index b68e29cd5e..65ca217470 100644 --- a/paddle/trainer/Trainer.cpp +++ b/paddle/trainer/Trainer.cpp @@ -137,6 +137,10 @@ void Trainer::init(const std::shared_ptr& config, } } + if (FLAGS_trainer_count > 1) { + CHECK(!FLAGS_use_mkldnn) << "MKLDNN only need 1 trainer"; + } + if (testing) { LOG(INFO) << "trainer: in testing mode"; if (config_->getOptConfig().use_sparse_remote_updater() || From c808fbbfcbaaf5c08f6254bfdb860f5dac76a627 Mon Sep 17 00:00:00 2001 From: Yiqun Liu Date: Fri, 17 Nov 2017 10:15:40 +0800 Subject: [PATCH 2/4] Support the build for multiple architectures at one cmake command (iOS). (#5677) * Support the build for multiple architectures at one cmake command (iOS). * Update the documentations. --- cmake/cross_compiling/ios.cmake | 8 +++----- cmake/external/openblas.cmake | 13 ++++++------- cmake/external/warpctc.cmake | 4 ++++ doc/mobile/cross_compiling_for_android_cn.md | 2 +- doc/mobile/cross_compiling_for_ios_cn.md | 12 ++++++------ doc/mobile/cross_compiling_for_raspberry_cn.md | 2 +- paddle/cuda/include/hl_gpu.h | 2 ++ 7 files changed, 23 insertions(+), 20 deletions(-) diff --git a/cmake/cross_compiling/ios.cmake b/cmake/cross_compiling/ios.cmake index 310450f7d0..d3f5bf6852 100644 --- a/cmake/cross_compiling/ios.cmake +++ b/cmake/cross_compiling/ios.cmake @@ -76,11 +76,9 @@ set(IOS_PLATFORM ${IOS_PLATFORM} CACHE STRING "Type of iOS Platform") # Set the architecture for iOS if(NOT DEFINED IOS_ARCH) if(IOS_PLATFORM STREQUAL "OS") - # FIXME(liuyiqun): support "armv7;armv7s;arm64" future - set(IOS_ARCH "arm64") + set(IOS_ARCH "armv7;armv7s;arm64") elseif(IOS_PLATFORM STREQUAL "SIMULATOR") - # FIXME(liuyiqun): support "i386;x86_64" future - set(IOS_ARCH "x86_64") + set(IOS_ARCH "i386;x86_64") endif() endif() set(CMAKE_OSX_ARCHITECTURES ${IOS_ARCH} CACHE string "Build architecture for iOS") @@ -248,7 +246,7 @@ set(IOS_COMPILER_FLAGS "${XCODE_IOS_PLATFORM_VERSION_FLAGS} ${XCODE_IOS_BITCODE_ # Hidden visibilty is required for cxx on iOS set(CMAKE_C_FLAGS "${IOS_COMPILER_FLAGS} ${CMAKE_C_FLAGS}" CACHE STRING "C flags") -set(CMAKE_CXX_FLAGS "${IOS_COMPILER_FLAGS} -fvisibility-inlines-hidden ${CMAKE_CXX_FLAGS}" CACHE STRING "CXX flags") +set(CMAKE_CXX_FLAGS "${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") diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index 2253807981..4c4f59656d 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -45,15 +45,14 @@ IF(NOT ${CBLAS_FOUND}) SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV8 BINARY=64 USE_THREAD=0) ENDIF() ELSEIF(IOS) - # FIXME(liuyiqun): support multiple architectures - SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5") - SET(OPENBLAS_CC "${OPENBLAS_CC} ${CMAKE_C_FLAGS} -isysroot ${CMAKE_OSX_SYSROOT}") - IF(CMAKE_OSX_ARCHITECTURES MATCHES "armv7") - SET(OPENBLAS_CC "${OPENBLAS_CC} -arch armv7") - SET(OPTIONAL_ARGS ${OPTIONAL_ARGS} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0) - ELSEIF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64") + IF(CMAKE_OSX_ARCHITECTURES MATCHES "arm64") + SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5") + SET(OPENBLAS_CC "${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 diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake index 8bd0582228..a8e1aca49c 100644 --- a/cmake/external/warpctc.cmake +++ b/cmake/external/warpctc.cmake @@ -12,6 +12,10 @@ # 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) diff --git a/doc/mobile/cross_compiling_for_android_cn.md b/doc/mobile/cross_compiling_for_android_cn.md index 882066f237..424d7718c6 100644 --- a/doc/mobile/cross_compiling_for_android_cn.md +++ b/doc/mobile/cross_compiling_for_android_cn.md @@ -1,4 +1,4 @@ -# 构建Android平台上的PaddlePaddle库 +# Android平台编译指南 用户可通过如下两种方式，交叉编译Android平台上适用的PaddlePaddle库： - 基于Docker容器的编译方式 diff --git a/doc/mobile/cross_compiling_for_ios_cn.md b/doc/mobile/cross_compiling_for_ios_cn.md index cda636a67d..9da48e7f21 100644 --- a/doc/mobile/cross_compiling_for_ios_cn.md +++ b/doc/mobile/cross_compiling_for_ios_cn.md @@ -1,4 +1,4 @@ -# 构建iOS平台上的PaddlePaddle库 +# iOS平台编译指南 交叉编译iOS平台上适用的PaddlePaddle库，需要在MacOS系统上进行。本文的将介绍在MacOS上，从源码交叉编译iOS平台上适用的PaddlePaddle库。 ## 准备交叉编译环境 @@ -25,7 +25,7 @@ iOS平台可选配置参数： - `IOS_PLATFORM`，可设置为`OS/SIMULATOR`，默认值为`OS`。 - `OS`，构建目标为`arm`架构的iPhone或者iPad等物理设备。 - `SIMULATOR`，构建目标为`x86`架构的模拟器平台。 -- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示： +- `IOS_ARCH`，目标架构。针对不同的`IOS_PLATFORM`，可设置的目标架构如下表所示，默认编译所有架构： @@ -41,11 +41,11 @@ iOS平台可选配置参数： - + - +
OSarmv7, armv7s, arm64 (默认)armv7, armv7s, arm64
SIMULATORi386, x86_64 (默认)i386, x86_64
@@ -66,7 +66,7 @@ iOS平台可选配置参数： ```bash cmake -DCMAKE_SYSTEM_NAME=iOS \ -DIOS_PLATFORM=OS \ - -DIOS_ARCH="arm64" \ + -DIOS_ARCH="armv7;arm64" \ -DIOS_ENABLE_BITCODE=ON \ -DIOS_USE_VECLIB_FOR_BLAS=ON \ -DCMAKE_INSTALL_PREFIX=your/path/to/install \ @@ -112,6 +112,6 @@ $ make install - `lib`目录，其中包含PaddlePaddle的C-API静态库 - `third_party`目录，其中包含所依赖的所有第三方库 -注意，不同架构的PaddlePaddle库建议安装到不同的目录下，然后使用`lipo`工具将多个静态库合并成一个支持多个架构的fat库。 +注意，如果PaddlePaddle库需要同时支持真机和模拟器，则需要分别编译真机和模拟器版本，然后使用`lipo`工具合并fat库。 自此，PaddlePaddle库已经安装完成，用户可将合成的fat库用于深度学习相关的iOS App中，调用方法见C-API文档。 diff --git a/doc/mobile/cross_compiling_for_raspberry_cn.md b/doc/mobile/cross_compiling_for_raspberry_cn.md index 6e983645fa..f8ef9dc803 100644 --- a/doc/mobile/cross_compiling_for_raspberry_cn.md +++ b/doc/mobile/cross_compiling_for_raspberry_cn.md @@ -1,4 +1,4 @@ -# 构建Raspberry Pi平台上的PaddlePaddle库 +# Raspberry Pi平台编译指南 通常有两个方法来构建基于 Rasspberry Pi 的版本： diff --git a/paddle/cuda/include/hl_gpu.h b/paddle/cuda/include/hl_gpu.h index ede2670882..4ab8de80d1 100644 --- a/paddle/cuda/include/hl_gpu.h +++ b/paddle/cuda/include/hl_gpu.h @@ -25,7 +25,9 @@ limitations under the License. */ #include "hl_matrix.h" #include "hl_sequence.h" #include "hl_sparse.h" +#ifndef PADDLE_MOBILE_INFERENCE #include "hl_warpctc_wrap.h" +#endif #ifdef HPPL_STUB_FUNC #include "stub/hl_aggregate_stub.h" From 23a674c98aee5eaf00280d6952d2cc3dec40b495 Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Fri, 17 Nov 2017 10:33:19 +0800 Subject: [PATCH 3/4] switch the flag --- paddle/trainer/Trainer.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/paddle/trainer/Trainer.cpp b/paddle/trainer/Trainer.cpp index 65ca217470..88e684849d 100644 --- a/paddle/trainer/Trainer.cpp +++ b/paddle/trainer/Trainer.cpp @@ -137,8 +137,8 @@ void Trainer::init(const std::shared_ptr& config, } } - if (FLAGS_trainer_count > 1) { - CHECK(!FLAGS_use_mkldnn) << "MKLDNN only need 1 trainer"; + if (FLAGS_use_mkldnn) { + CHECK_EQ(FLAGS_trainer_count, 1UL) << "MKLDNN only need 1 trainer"; } if (testing) { From d13c3a98ceffa807a8fb4e8d2971acf0235afa06 Mon Sep 17 00:00:00 2001 From: Yancey1989 Date: Fri, 17 Nov 2017 10:36:03 +0800 Subject: [PATCH 4/4] fix no framework proto file --- paddle/operators/math/CMakeLists.txt | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/paddle/operators/math/CMakeLists.txt b/paddle/operators/math/CMakeLists.txt index b9417f1d7f..002b68fecf 100644 --- a/paddle/operators/math/CMakeLists.txt +++ b/paddle/operators/math/CMakeLists.txt @@ -1,7 +1,7 @@ add_subdirectory(detail) if(WITH_GPU) - nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu DEPS cblas device_context) + nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc im2col.cu DEPS cblas device_context framework_proto) nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function tensor) nv_library(selected_rows_functor SRCS selected_rows_functor.cc selected_rows_functor.cu DEPS selected_rows math_function) nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor) @@ -15,7 +15,7 @@ if(WITH_GPU) nv_library(lstm_compute SRCS lstm_compute.cc lstm_compute.cu DEPS device_context activation_functions) nv_library(gru_compute SRCS gru_compute.cc gru_compute.cu DEPS device_context activation_functions math_function) else() - cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context) + cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context framework_proto) cc_library(selected_rows_functor SRCS selected_rows_functor.cc DEPS selected_rows math_function) cc_library(softmax SRCS softmax.cc DEPS device_context) cc_library(cross_entropy SRCS cross_entropy.cc DEPS device_context)