diff --git a/CMakeLists.txt b/CMakeLists.txt index 9be75f4a7de4c6ecac461f9dcb62ab10e79cee6b..ed3c390066dfac2322d802c6039bc7155a36e38a 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -27,6 +27,7 @@ if(NOT CMAKE_CROSSCOMPILING) endif(NOT CMAKE_CROSSCOMPILING) find_package(Git REQUIRED) find_package(Threads REQUIRED) +find_package(Boost QUIET) include(simd) @@ -48,6 +49,7 @@ 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_GOLANG "Compile PaddlePaddle with GOLANG" OFF) +option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) # CMAKE_BUILD_TYPE if(NOT CMAKE_BUILD_TYPE) @@ -110,6 +112,7 @@ include_directories("${PROJ_ROOT}") include_directories("${PROJ_ROOT}/paddle/cuda/include") include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto") include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/cclient") +include_directories(${Boost_INCLUDE_DIRS}) set(EXTERNAL_LIBS ${GFLAGS_LIBRARIES} @@ -127,15 +130,21 @@ if(WITH_GPU) endif(NOT WITH_DSO) endif(WITH_GPU) +if(USE_NNPACK) + list(APPEND EXTERNAL_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB} "rt") +endif(USE_NNPACK) + add_subdirectory(proto) -add_subdirectory(paddle) -add_subdirectory(python) +# "add_subdirectory(paddle)" and "add_subdirectory(python)" should be +# placed after this block, because they depends on it. if(WITH_GOLANG) - #TODO (add go/master/c back when fixed) - add_subdirectory(go/pserver/cclient) + add_subdirectory(go/master/c) + add_subdirectory(go/pserver/cclient) endif(WITH_GOLANG) +add_subdirectory(paddle) +add_subdirectory(python) if(WITH_DOC) add_subdirectory(doc) endif() diff --git a/cmake/configure.cmake b/cmake/configure.cmake index e8425aedbdd269d54035a0457fa37e0ba834427a..f6dca6d575956fcb893f8e11970f84a739832d58 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -69,3 +69,27 @@ endif(NOT WITH_GPU) set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SIMD_FLAG}") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SIMD_FLAG}") + +if(WITH_GOLANG) + # we need to symlink Paddle directory into GOPATH. If we + # don't do it and we have code that depends on Paddle, go + # get ./... will download a new Paddle repo from Github, + # without the changes in our current Paddle repo that we + # want to build. + set(GOPATH "${CMAKE_CURRENT_BINARY_DIR}/go") + file(MAKE_DIRECTORY ${GOPATH}) + set(PADDLE_IN_GOPATH "${GOPATH}/src/github.com/PaddlePaddle/Paddle") + add_custom_target(go_path) + add_custom_command(TARGET go_path + # Symlink Paddle directory into GOPATH + COMMAND mkdir -p ${PADDLE_IN_GOPATH} + COMMAND rm -rf ${PADDLE_IN_GOPATH} + COMMAND ln -sf ${CMAKE_SOURCE_DIR} ${PADDLE_IN_GOPATH} + # Automatically get all dependencies specified in the source code + # We can't run `go get -d ./...` for every target, because + # multiple `go get` can not run concurrently, but make need to be + # able to run with multiple jobs. + COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} get -d ./go/... + WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} + ) +endif(WITH_GOLANG) diff --git a/cmake/external/eigen.cmake b/cmake/external/eigen.cmake index 253d436bcc04d8e0db78f6a4a2c67a050f456bba..45f44f617dcb46062355df4e35d537086215a46d 100644 --- a/cmake/external/eigen.cmake +++ b/cmake/external/eigen.cmake @@ -7,8 +7,17 @@ INCLUDE_DIRECTORIES(${EIGEN_SOURCE_DIR}/src/eigen3) ExternalProject_Add( eigen3 ${EXTERNAL_PROJECT_LOG_ARGS} - URL "https://bitbucket.org/eigen/eigen/get/3.3.4.tar.gz" - URL_MD5 "1a47e78efe365a97de0c022d127607c3" + # for latest version, please get from official website + # URL "https://bitbucket.org/eigen/eigen/get/3.3.4.tar.gz" + # URL_MD5 "1a47e78efe365a97de0c022d127607c3" + + # for no-ssl http support, please get from bazel's mirror + # URL "http://mirror.bazel.build/bitbucket.org/eigen/eigen/get/f3a22f35b044.tar.gz" + # URL_MD5 "4645c66075982da6fa0bcf6b20f3e8f7" + + # get from github mirror + GIT_REPOSITORY "https://github.com/RLovelett/eigen.git" + GIT_TAG "a46d2e7337c4656f00abe54a8115f6d76153a048" PREFIX ${EIGEN_SOURCE_DIR} UPDATE_COMMAND "" CONFIGURE_COMMAND "" diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake index d43badc1da50723d5d3dbd1f19f0bd4ef4d24737..3c74944bc21a131fe90e61777d3dce8b3f21900a 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -13,6 +13,10 @@ # limitations under the License. INCLUDE(ExternalProject) +# Always invoke `FIND_PACKAGE(Protobuf)` for importing function protobuf_generate_cpp +FIND_PACKAGE(Protobuf QUIET) +SET(PROTOBUF_FOUND "OFF") + # Print and set the protobuf library information, # finish this cmake process and exit from this file. @@ -39,12 +43,19 @@ macro(PROMPT_PROTOBUF_LIB) ADD_LIBRARY(protobuf_lite ${protobuf_LIBTYPE} IMPORTED GLOBAL) SET_PROPERTY(TARGET protobuf_lite PROPERTY IMPORTED_LOCATION ${PROTOBUF_LITE_LIBRARY}) - ADD_LIBRARY(protoc ${protobuf_LIBTYPE} IMPORTED GLOBAL) - SET_PROPERTY(TARGET protoc PROPERTY IMPORTED_LOCATION ${PROTOC_LIBRARY}) + ADD_LIBRARY(libprotoc ${protobuf_LIBTYPE} IMPORTED GLOBAL) + SET_PROPERTY(TARGET libprotoc PROPERTY IMPORTED_LOCATION ${PROTOC_LIBRARY}) + + ADD_EXECUTABLE(protoc IMPORTED GLOBAL) + SET_PROPERTY(TARGET protoc PROPERTY IMPORTED_LOCATION ${PROTOBUF_PROTOC_EXECUTABLE}) + # FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`. + # make `protobuf_generate_cpp` happy. + SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE}) FOREACH(dep ${protobuf_DEPS}) ADD_DEPENDENCIES(protobuf ${dep}) ADD_DEPENDENCIES(protobuf_lite ${dep}) + ADD_DEPENDENCIES(libprotoc ${dep}) ADD_DEPENDENCIES(protoc ${dep}) ENDFOREACH() @@ -133,18 +144,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST) ENDFUNCTION() SET(PROTOBUF_VERSION 3.1) -IF(NOT CMAKE_CROSSCOMPILING) - FIND_PACKAGE(Protobuf ${PROTOBUF_VERSION}) - - IF(PROTOBUF_FOUND) - SET_PROTOBUF_VERSION() - IF("${PROTOBUF_VERSION}" VERSION_LESS "3.1.0") - SET(PROTOBUF_FOUND OFF) - ELSE() - PROMPT_PROTOBUF_LIB() - ENDIF() - ENDIF(PROTOBUF_FOUND) -ELSE() +IF(CMAKE_CROSSCOMPILING) build_protobuf(protobuf_host TRUE) LIST(APPEND external_project_dependencies protobuf_host) diff --git a/cmake/external/python.cmake b/cmake/external/python.cmake index f4d0daab06c9fcf17f4af59c25f62b415074a52f..6546b2c83bc8f81f89e4018a2216f191bbeb0d21 100644 --- a/cmake/external/python.cmake +++ b/cmake/external/python.cmake @@ -32,193 +32,6 @@ IF(PYTHONINTERP_FOUND) MESSAGE(FATAL_ERROR "Found Python Protobuf ${PY_GOOGLE.PROTOBUF_VERSION} < 3.0.0, " "please use pip to upgrade protobuf. pip install -U protobuf") ENDIF() -ELSE(PYTHONINTERP_FOUND) - MESSAGE(FATAL_ERROR "Please install python 2.7 before building PaddlePaddle.") - ##################################### PYTHON ######################################## - SET(PYTHON_SOURCES_DIR ${THIRD_PARTY_PATH}/python) - SET(PYTHON_INSTALL_DIR ${THIRD_PARTY_PATH}/install/python) - SET(_python_DIR ${PYTHON_INSTALL_DIR}) - - IF(UNIX) - SET(PYTHON_FOUND ON) - SET(PYTHON_INCLUDE_DIR "${PYTHON_INSTALL_DIR}/include/python2.7" CACHE PATH "Python include dir" FORCE) - SET(PYTHON_LIBRARIES "${PYTHON_INSTALL_DIR}/lib/libpython2.7.a" CACHE FILEPATH "Python library" FORCE) - SET(PYTHON_EXECUTABLE ${PYTHON_INSTALL_DIR}/bin/python CACHE FILEPATH "Python executable" FORCE) - SET(PY_SITE_PACKAGES_PATH "${PYTHON_INSTALL_DIR}/lib/python2.7/site-packages" CACHE PATH "Python site-packages path" FORCE) - ELSEIF(WIN32) - SET(PYTHON_FOUND ON) - SET(PYTHON_INCLUDE_DIR "${PYTHON_INSTALL_DIR}/include" CACHE PATH "Python include dir" FORCE) - SET(PYTHON_LIBRARIES "${PYTHON_INSTALL_DIR}/libs/python27.lib" CACHE FILEPATH "Python library" FORCE) - SET(PYTHON_EXECUTABLE "${PYTHON_INSTALL_DIR}/bin/python.exe" CACHE FILEPATH "Python executable" FORCE) - SET(PY_SITE_PACKAGES_PATH "${PYTHON_INSTALL_DIR}/Lib/site-packages" CACHE PATH "Python site-packages path" FORCE) - ELSE() - MESSAGE(FATAL_ERROR "Unknown system !") - ENDIF() - - IF(APPLE) - LIST(APPEND EXTERNAL_PROJECT_OPTIONAL_CMAKE_ARGS - -DCMAKE_BUILD_WITH_INSTALL_RPATH:BOOL=ON - ) - ENDIF() - - SET(EXTERNAL_PROJECT_OPTIONAL_CMAKE_CACHE_ARGS) - - # Force Python build to "Release". - IF(CMAKE_CONFIGURATION_TYPES) - SET(SAVED_CMAKE_CFG_INTDIR ${CMAKE_CFG_INTDIR}) - SET(CMAKE_CFG_INTDIR "Release") - ELSE() - LIST(APPEND EXTERNAL_PROJECT_OPTIONAL_CMAKE_CACHE_ARGS - -DCMAKE_BUILD_TYPE:STRING=Release - ) - ENDIF() - - ExternalProject_Add(python - ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/python-cmake-buildsystem/python-cmake-buildsystem.git" - PREFIX ${PYTHON_SOURCES_DIR} - UPDATE_COMMAND "" - CMAKE_ARGS -DPYTHON_VERSION=2.7.12 - CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} - CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} - CMAKE_CACHE_ARGS - -DCMAKE_INSTALL_PREFIX:PATH=${PYTHON_INSTALL_DIR} - -DBUILD_LIBPYTHON_SHARED:BOOL=OFF - -DUSE_SYSTEM_LIBRARIES:BOOL=OFF - -DZLIB_ROOT:FILEPATH=${ZLIB_ROOT} - -DZLIB_INCLUDE_DIR:PATH=${ZLIB_INCLUDE_DIR} - -DZLIB_LIBRARY:FILEPATH=${ZLIB_LIBRARIES} - -DDOWNLOAD_SOURCES:BOOL=ON - -DINSTALL_WINDOWS_TRADITIONAL:BOOL=OFF - ${EXTERNAL_PROJECT_OPTIONAL_CMAKE_CACHE_ARGS} - ${EXTERNAL_PROJECT_OPTIONAL_CMAKE_ARGS} - DEPENDS zlib - ) - - SET(py_env - PATH=${PYTHON_INSTALL_DIR}/bin - PYTHONHOME=${PYTHON_INSTALL_DIR} - PYTHONPATH=${PYTHON_INSTALL_DIR}/lib:${PYTHON_INSTALL_DIR}/lib/python2.7:${PY_SITE_PACKAGES_PATH}) - #################################################################################### - - ##################################### SETUPTOOLS ################################### - SET(SETUPTOOLS_SOURCES_DIR ${PYTHON_SOURCES_DIR}/setuptools) - ExternalProject_Add(setuptools - ${EXTERNAL_PROJECT_LOG_ARGS} - PREFIX ${SETUPTOOLS_SOURCES_DIR} - URL "https://pypi.python.org/packages/source/s/setuptools/setuptools-18.3.2.tar.gz" - BUILD_IN_SOURCE 1 - PATCH_COMMAND "" - UPDATE_COMMAND "" - CONFIGURE_COMMAND "" - INSTALL_COMMAND "" - BUILD_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py install - DEPENDS python zlib - ) - ##################################################################################### - - ##################################### SIX ########################################### - SET(SIX_SOURCES_DIR ${PYTHON_SOURCES_DIR}/six) - ExternalProject_Add(six - ${EXTERNAL_PROJECT_LOG_ARGS} - PREFIX ${SIX_SOURCES_DIR} - URL https://pypi.python.org/packages/source/s/six/six-1.10.0.tar.gz - BUILD_IN_SOURCE 1 - PATCH_COMMAND "" - UPDATE_COMMAND "" - CONFIGURE_COMMAND "" - INSTALL_COMMAND "" - BUILD_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py install - DEPENDS python setuptools - ) - ##################################################################################### - - ##################################### CYTHON ######################################## - SET(CYTHON_SOURCES_DIR ${PYTHON_SOURCES_DIR}/cython) - ExternalProject_Add(cython - ${EXTERNAL_PROJECT_LOG_ARGS} - PREFIX ${CYTHON_SOURCES_DIR} - URL https://github.com/cython/cython/archive/0.25.2.tar.gz - GIT_TAG 0.25.2 - BUILD_IN_SOURCE 1 - CONFIGURE_COMMAND "" - PATCH_COMMAND "" - UPDATE_COMMAND "" - INSTALL_COMMAND "" - BUILD_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py install - DEPENDS python - ) - #################################################################################### - - ##################################### NUMPY ######################################## - SET(NUMPY_SOURCES_DIR ${PYTHON_SOURCES_DIR}/numpy) - SET(NUMPY_TAG_VERSION "v1.11.3") - SET(NUMPY_VERSION "1.11.3") - - SET(EGG_NAME "") - SET(PYTHON_NUMPY_INCLUDE_DIR "") - IF(WIN32) - SET(EGG_NAME "numpy-${NUMPY_VERSION}-py2.7-${HOST_SYSTEM}.egg") - ELSE(WIN32) - IF(APPLE) - SET(EGG_NAME "numpy-${NUMPY_VERSION}-py2.7-${HOST_SYSTEM}-${MACOS_VERSION}") - ELSE(APPLE) - SET(EGG_NAME "numpy-${NUMPY_VERSION}-py2.7-linux") - SET(EGG_NAME "numpy-${NUMPY_VERSION}-py2.7-linux") - ENDIF(APPLE) - - FOREACH(suffix x86_64 intel fat64 fat32 universal) - LIST(APPEND PYTHON_NUMPY_INCLUDE_DIR ${PY_SITE_PACKAGES_PATH}/${EGG_NAME}-${suffix}.egg/numpy/core/include) - ENDFOREACH() - ENDIF(WIN32) - - ExternalProject_Add(numpy - ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY https://github.com/numpy/numpy.git - GIT_TAG ${NUMPY_TAG_VERSION} - CONFIGURE_COMMAND "" - UPDATE_COMMAND "" - PREFIX ${NUMPY_SOURCES_DIR} - BUILD_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py build - INSTALL_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py install - BUILD_IN_SOURCE 1 - DEPENDS python setuptools cython - ) - #################################################################################### - - ##################################### WHEEL ######################################## - SET(WHEEL_SOURCES_DIR ${PYTHON_SOURCES_DIR}/wheel) - ExternalProject_Add(wheel - ${EXTERNAL_PROJECT_LOG_ARGS} - URL https://pypi.python.org/packages/source/w/wheel/wheel-0.29.0.tar.gz - PREFIX ${WHEEL_SOURCES_DIR} - CONFIGURE_COMMAND "" - UPDATE_COMMAND "" - BUILD_COMMAND "" - INSTALL_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py install - BUILD_IN_SOURCE 1 - DEPENDS python setuptools - ) - #################################################################################### - - ################################### PROTOBUF ####################################### - SET(PY_PROTOBUF_SOURCES_DIR ${PYTHON_SOURCES_DIR}/protobuf) - ExternalProject_Add(python-protobuf - ${EXTERNAL_PROJECT_LOG_ARGS} - URL https://pypi.python.org/packages/e0/b0/0a1b364fe8a7d177b4b7d4dca5b798500dc57a7273b93cca73931b305a6a/protobuf-3.1.0.post1.tar.gz - URL_MD5 38b5fb160c768d2f8444d0c6d637ff91 - PREFIX ${PY_PROTOBUF_SOURCES_DIR} - BUILD_IN_SOURCE 1 - PATCH_COMMAND "" - CONFIGURE_COMMAND "" - BUILD_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py build - INSTALL_COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py install - DEPENDS python setuptools six - ) - #################################################################################### - - LIST(APPEND external_project_dependencies python setuptools six cython wheel python-protobuf numpy) - ENDIF(PYTHONINTERP_FOUND) IF(WITH_PYTHON) diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 11c1f677ae5b308558b54bf49caf168cf6023444..88be13b2ac95172d5d9099d62a40449c6a01e98a 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -87,6 +87,9 @@ # go_library(example SHARED) # +# including binary directory for generated headers. +include_directories(${CMAKE_BINARY_DIR}) + if(NOT APPLE) find_package(Threads REQUIRED) link_libraries(${CMAKE_THREAD_LIBS_INIT}) @@ -98,23 +101,16 @@ function(merge_static_libs TARGET_NAME) # First get the file names of the libraries to be merged foreach(lib ${libs}) - get_target_property(libtype ${lib} TYPE) - if(NOT libtype STREQUAL "STATIC_LIBRARY") - message(FATAL_ERROR "merge_static_libs can only process static libraries") - endif() set(libfiles ${libfiles} $) endforeach() if(APPLE) # Use OSX's libtool to merge archives - add_custom_target(${TARGET_NAME}_archive - COMMAND libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles} - WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} - DEPENDS ${libs} - ) - add_library(${TARGET_NAME} STATIC IMPORTED GLOBAL) - set_property(TARGET ${TARGET_NAME} PROPERTY - IMPORTED_LOCATION "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a") - add_dependencies(${TARGET_NAME} ${TARGET_NAME}_archive) + set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c) + file(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";") + add_library(${TARGET_NAME} STATIC ${dummyfile}) + add_custom_command(TARGET ${TARGET_NAME} POST_BUILD + COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" + COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles}) else() # general UNIX: use "ar" to extract objects and re-add to a common lib foreach(lib ${libs}) set(objlistfile ${lib}.objlist) # list of objects in the input library @@ -143,9 +139,9 @@ function(merge_static_libs TARGET_NAME) set(outlibfile "$") foreach(lib ${libs}) - add_custom_command(TARGET ${TARGET_NAME} POST_BUILD - COMMAND ${CMAKE_AR} ru ${outlibfile} @"../${objlistfile}" - WORKING_DIRECTORY ${objdir}) + add_custom_command(TARGET ${TARGET_NAME} POST_BUILD + COMMAND ${CMAKE_AR} ru ${outlibfile} @"../${lib}.objlist" + WORKING_DIRECTORY ${lib}.objdir) endforeach() add_custom_command(TARGET ${TARGET_NAME} POST_BUILD @@ -253,10 +249,6 @@ function(nv_test TARGET_NAME) endif() endfunction(nv_test) -set(GOPATH "${CMAKE_CURRENT_BINARY_DIR}/go") -file(MAKE_DIRECTORY ${GOPATH}) -set(PADDLE_IN_GOPATH "${GOPATH}/src/github.com/PaddlePaddle/Paddle") - function(go_library TARGET_NAME) set(options STATIC static SHARED shared) set(oneValueArgs "") @@ -265,10 +257,10 @@ function(go_library TARGET_NAME) if (go_library_SHARED OR go_library_shared) set(BUILD_MODE "-buildmode=c-shared") - set(LIB_NAME "${CMAKE_SHARED_LIBRARY_PREFIX}${TARGET_NAME}${CMAKE_SHARED_LIBRARY_SUFFIX}") + set(${TARGET_NAME}_LIB_NAME "${CMAKE_SHARED_LIBRARY_PREFIX}${TARGET_NAME}${CMAKE_SHARED_LIBRARY_SUFFIX}" CACHE STRING "output library name for target ${TARGET_NAME}") else() set(BUILD_MODE "-buildmode=c-archive") - set(LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}${TARGET_NAME}${CMAKE_STATIC_LIBRARY_SUFFIX}") + set(${TARGET_NAME}_LIB_NAME "${CMAKE_STATIC_LIBRARY_PREFIX}${TARGET_NAME}${CMAKE_STATIC_LIBRARY_SUFFIX}" CACHE STRING "output library name for target ${TARGET_NAME}") endif() # Add dummy code to support `make target_name` under Terminal Command @@ -283,25 +275,17 @@ function(go_library TARGET_NAME) add_dependencies(${TARGET_NAME} ${go_library_DEPS}) endif(go_library_DEPS) - # we need to symlink Paddle directory into GOPATH. If we - # don't do it and we have code that depends on Paddle, go - # get ./... will download a new Paddle repo from Github, - # without the changes in our current Paddle repo that we - # want to build. + set(${TARGET_NAME}_LIB_PATH "${CMAKE_CURRENT_BINARY_DIR}/${${TARGET_NAME}_LIB_NAME}" CACHE STRING "output library path for target ${TARGET_NAME}") + file(GLOB GO_SOURCE RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "*.go") add_custom_command(TARGET ${TARGET_NAME} POST_BUILD - COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/${LIB_NAME}" - # Symlink Paddle directory into GOPATH - COMMAND mkdir -p ${PADDLE_IN_GOPATH} - COMMAND rm -rf ${PADDLE_IN_GOPATH} - COMMAND ln -sf ${CMAKE_SOURCE_DIR} ${PADDLE_IN_GOPATH} - # Automatically get all dependencies specified in the source code - COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} get -d ./... + COMMAND rm "${${TARGET_NAME}_LIB_PATH}" # Golang build source code COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} build ${BUILD_MODE} - -o "${CMAKE_CURRENT_BINARY_DIR}/${LIB_NAME}" + -o "${${TARGET_NAME}_LIB_PATH}" ${GO_SOURCE} WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}) + add_dependencies(${TARGET_NAME} go_path) endfunction(go_library) function(go_binary TARGET_NAME) @@ -331,3 +315,13 @@ function(go_test TARGET_NAME) add_custom_target(${TARGET_NAME} ALL DEPENDS ${TARGET_NAME}_timestamp ${go_test_DEPS}) add_test(${TARGET_NAME} ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}) endfunction(go_test) + +function(proto_library TARGET_NAME) + set(oneValueArgs "") + set(multiValueArgs SRCS) + cmake_parse_arguments(proto_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + set(proto_srcs) + set(proto_hdrs) + protobuf_generate_cpp(proto_srcs proto_hdrs ${proto_library_SRCS}) + cc_library(${TARGET_NAME} SRCS ${proto_srcs} DEPS protobuf) +endfunction() diff --git a/doc/CMakeLists.txt b/doc/CMakeLists.txt index 6fa42fd0c71e78cc2fa6b0fe2cb970baf4ac89ed..94dd3457fb5b513441c4c8e339e1862de9092517 100644 --- a/doc/CMakeLists.txt +++ b/doc/CMakeLists.txt @@ -27,10 +27,6 @@ sphinx_add_target(paddle_docs ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_EN}) -add_dependencies(paddle_docs - gen_proto_py) - - # configured documentation tools and intermediate build results set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build") @@ -51,6 +47,3 @@ sphinx_add_target(paddle_docs_cn ${SPHINX_CACHE_DIR_CN} ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_CN}) - -add_dependencies(paddle_docs_cn - gen_proto_py) diff --git a/doc/design/build_system/README.md b/doc/design/build_system/README.md index 310739f37ae48934afe1d042e87efef85b98f1fc..bf0e4dddc1b640ecbce489f65820aaf8a4b3b1e7 100644 --- a/doc/design/build_system/README.md +++ b/doc/design/build_system/README.md @@ -105,3 +105,48 @@ shared_library(api ### Implementation As above example CMakeLists.txt executes, each function invocation adds "nodes" to a dependency graph. It also use this graph to generate CMake commands including `add_executable`, `add_dependencies`, `target_link_libraries`, and `add_test`. + +### Using Package Manager For Go + +Building Go binaries and libraries need to satisfy their dependencies, generally +we can do `go get ./...` to download and compile all external dependencies. The +problems are: + +1. `go get` will always get the latest code from the default branch of the + remote repo, so changes of dependents might break the build. This is very + different with what we already have in `cmake/external` which download a + specific version or commit id of the dependency. +1. Some locations can not access external dependencies through the internet, as mentioned + in https://github.com/PaddlePaddle/Paddle/issues/2605. Using package management + tools can package the dependencies as a "vendor" package, which can be mirrored + at many cloud file hosting, so users what to compile paddle by themselves can + download this "vendor" package from a mirror site. + +#### Choose A Suitable Tool + +As mentioned by @wangkuiyi, [Here](https://github.com/golang/go/wiki/PackageManagementTools) +list dozens of Go package managers. We choose the tool using following principles: + +- Most "active" projects with more stars, more pull requests or commits +- Widely used project + +After comparing all these projects, we shall choose between the most popular +tools: Godep and Glide. + +Here's a brief comparison between Godep and Glide +: https://github.com/Masterminds/glide/wiki/Go-Package-Manager-Comparison. There are +also many complaints about using `Godep`. There's also a new "official" pakcage +management tool has been started at: https://github.com/golang/dep to resolve +such problems, but it's currently at Alpha stage. So the best choice now is +glide obviously. + +#### Manage Go Packages + +- Dependencies: `go/glide.yaml` will store the dependencies and their versions which + is directly imported by paddle. `go/glide.lock` will store all dependencies recursively + with their commit id. Builds will "lock" to these packages if we don't `glide up` + them +- Vendor package: `go/vendor` directory will generated when running `cmake` command. `cmake` + will download the code corresponding to `go/glide.lock`. If we put a vendor folder + under `go/`, cmake will just check the commit id to the packages under the folder, + if commit id matches, there will be no download at all. diff --git a/doc/design/cluster_train/save_model.md b/doc/design/cluster_train/save_model.md new file mode 100644 index 0000000000000000000000000000000000000000..b70f00176b6701ef487ef88ac0933b9b227037ea --- /dev/null +++ b/doc/design/cluster_train/save_model.md @@ -0,0 +1,110 @@ +# Design Doc: Save Model + +## Overview + +The model is the output of the training process. There are two +ways from which user can obtain a model: + +- Save model triggered by user code: user code asks PaddlePaddle to + save a model. +- Convert model from the checkpoint: model being converted from + pservers' periodic checkpoint. In this way, the user can cancel a + job at any time, and still have a relatively fresh model (we + checkpoint around every 5 minutes). + +### Trainer Saving Model vs. Pservers Saving Model + +Both trainers and pservers have access to the model. So the model can +be saved from a trainer or pservers. We need to decide where the model +is saved from. + +#### Dense Update vs. Sparse Update + +There are two types of model update methods: dense update and sparse +update (when the model parameter is configured to be sparse). + +- Dense update + + Every trainer has it's own full copy of the model. Every model + update will update the entire model. + +- Sparse update + + The training input is sparse, and the trainer does not have the + entire model. It will only download the sub-model necessary related + to the input. When updating the model, only the sub-model related to + the training input is updated. + + +#### Pservers Saving Model + +The benefit of letting pservers save model is they have the entire +model all the time. However, since pservers are on different nodes, it +requires a merging process to merge model shards into the same +model. Thus requires the pservers to write models to a distributed +filesystem, making the checkpoint shards visible to the merge program. + +#### Trainer Saving Model + +The benefit of letting one trainer to save the model is it does not +require a distributed filesystem. And it's reusing the same save model +logic when training locally - except when doing sparse update, the +trainer needs to download the entire model during the saving process. + +#### Conclusion + +Given trainer saving model does not require a distributed filesystem, +and is an intuitive extension to trainer saving model when training +locally, we decide to let the trainer save the model when doing +distributed training. + + +### Convert Model from Checkpoint + +TODO + + +## Timeline + +We first implement trainer save the model. Converting the latest +snapshot to a model will be a TODO for future. + + +## Trainer Save Model + +### Trainer Election + +One trainer will be elected as the one to save the model. When using +etcd, trainer ID is a randomly generated UUID, we will utilize etcd to +elect one trainer. When not using etcd, unique trainer IDs will be +given by the administrator, the trainer whose ID is "0" is elected to +save the model. + +### Model Save Path + +Each trainer will be given the directory to save the model. The +elected trainer will save the model to +`given-directory/trainerID`. Since the trainer ID is unique, this +would prevent concurrent save to the same file when multiple trainers +are elected to save the model when split-brain problem happens. + +### What Happens When Model Is Saving + +It takes some time to save model, we need to define what will happen +when save model is taking place. + +When doing dense update, the trainer uses the local model. Pservers +does not need to pause model update. + +When doing sparse update. The trainer needs to download the entire +model while saving. To get the most accurate model, the model update +needs to be paused before the download starts and resumed after the +download finishes. Otherwise, the trainer gets a model that is +"polluted": some part of the model is old, some part of the model is +new. + +It's unclear that the "polluted" model will be inferior due to the +stochastic nature of deep learning, and pausing the model update will +add more complexity to the system. Since supporting sparse update is a +TODO item. We defer the evaluation of pause the model update or not +during saving model to the future. diff --git a/doc/design/scope.md b/doc/design/scope.md index 2ff416f06e8ada48b1d4922f8869a106f35799e2..afe6bc028cafc5ee24b0041905857af58d3f5790 100644 --- a/doc/design/scope.md +++ b/doc/design/scope.md @@ -41,7 +41,7 @@ class Scope { const Variable* GetVariable(const std::string& name) const; private: - std::unordered_map> vars_; + std::unordered_map> vars_; }; ``` @@ -59,9 +59,9 @@ class Scope { Scope(const std::shared_ptr& scope): parent_(scope) {} Variable* GetVariable(const std::string& name) const { - Variable* var = GetVarLocally(name); - if (var != nullptr) { - return var; + auto it = vars_.find(name); + if (it != vars_.end()) { + return it->second.get(); } else if (parent_ != nullptr) { return parent_->GetVariable(name); } else { @@ -97,8 +97,8 @@ class Scope { // return nullptr if not found. Variable* GetVariable(const std::string& name) const; - // return Error if already contains same name variable. - Error CreateVariable(const std::string& name); + // return if already contains same name variable. + Variable* CreateVariable(const std::string& name); private: std::shared_ptr parent_; diff --git a/doc/getstarted/concepts/src/train.py b/doc/getstarted/concepts/src/train.py index 679d0a931a7d650108ea89a04080a55d2976f72e..7e604f23de38543a00f305d508af0791193f78ba 100644 --- a/doc/getstarted/concepts/src/train.py +++ b/doc/getstarted/concepts/src/train.py @@ -31,7 +31,7 @@ def event_handler(event): # define training dataset reader def train_reader(): train_x = np.array([[1, 1], [1, 2], [3, 4], [5, 2]]) - train_y = np.array([-2, -3, -7, -7]) + train_y = np.array([[-2], [-3], [-7], [-7]]) def reader(): for i in xrange(train_y.shape[0]): diff --git a/go/cmd/pserver/pserver.go b/go/cmd/pserver/pserver.go index 6c85b1804bb9c5f3a8bc46bb3f54cc62c56cca70..8a42d4f8af1713e246f9efaf5dc7ba878c3b271e 100644 --- a/go/cmd/pserver/pserver.go +++ b/go/cmd/pserver/pserver.go @@ -30,7 +30,13 @@ func main() { log.SetLevel(level) timeout := time.Second * time.Duration((*etcdTimeout)) - s, err := pserver.NewService(*etcdEndpoint, *numPservers, timeout) + e := pserver.NewEtcdClient(*etcdEndpoint, *numPservers, timeout) + idx, err := e.Register() + if err != nil { + panic(err) + } + + s, err := pserver.NewService(idx) if err != nil { panic(err) } diff --git a/go/master/c/CMakeLists.txt b/go/master/c/CMakeLists.txt index 3eb598a877f3597336b509d3bcc2ea9589af3342..a4e92635bab8f0dbc5d91c7131fb0f8d1f5a6363 100644 --- a/go/master/c/CMakeLists.txt +++ b/go/master/c/CMakeLists.txt @@ -1,21 +1,3 @@ cmake_minimum_required(VERSION 3.0) -get_filename_component(PARENT_DIR ${CMAKE_CURRENT_SOURCE_DIR} DIRECTORY) -get_filename_component(PARENT_DIR ${PARENT_DIR} DIRECTORY) -set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${PARENT_DIR}/cmake") - -project(cxx_go C Go) - -#include(golang) -include(flags) - -set(MASTER_LIB_NAME "paddle_master") -go_library(${MASTER_LIB_NAME} SHARED) - -if(PROJ_ROOT) - add_custom_command(OUTPUT ${PROJ_ROOT}/python/paddle/v2/master/lib${MASTER_LIB_NAME}.so - COMMAND rm ${CMAKE_CURRENT_BINARY_DIR}/lib${MASTER_LIB_NAME}.h - COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/lib${MASTER_LIB_NAME}.so ${PROJ_ROOT}/python/paddle/v2/master/ - DEPENDS ${MASTER_LIB_NAME}) - add_custom_target(paddle_master_shared ALL DEPENDS ${PROJ_ROOT}/python/paddle/v2/master/lib${MASTER_LIB_NAME}.so) -endif(PROJ_ROOT) +go_library(paddle_master SHARED) diff --git a/go/master/c/client.go b/go/master/c/client.go index 13ed3b7680489d9a6d7e82ce38d2c76ef0e0bbde..635688f196b5014da95ca6107312bf1c15fe6a60 100644 --- a/go/master/c/client.go +++ b/go/master/c/client.go @@ -13,10 +13,13 @@ typedef int paddle_master_client; import "C" import ( + "strings" "sync" + "time" "unsafe" "github.com/PaddlePaddle/Paddle/go/master" + "github.com/coreos/etcd/clientv3" log "github.com/sirupsen/logrus" ) @@ -48,16 +51,33 @@ func remove(client C.paddle_master_client) *master.Client { return h } -type addresser string - -func (a addresser) Address() string { - return string(a) +//export paddle_new_etcd_master_client +func paddle_new_etcd_master_client(etcdEndpoints *C.char, timeout int, bufSize int) C.paddle_master_client { + p := C.GoString(etcdEndpoints) + cli, err := clientv3.New(clientv3.Config{ + Endpoints: strings.Split(p, ","), + DialTimeout: time.Second * time.Duration(timeout), + }) + if err != nil { + panic(err) + } + ch := make(chan string, 1) + a, err := master.GetKey(cli, master.DefaultAddrPath, timeout) + if err != nil { + panic(err) + } + ch <- a + go master.WatchKey(cli, master.DefaultAddrPath, ch) + c := master.NewClient(ch, bufSize) + return add(c) } //export paddle_new_master_client func paddle_new_master_client(addr *C.char, bufSize int) C.paddle_master_client { a := C.GoString(addr) - c := master.NewClient(addresser(a), bufSize) + ch := make(chan string, 1) + ch <- a + c := master.NewClient(ch, bufSize) return add(c) } diff --git a/go/master/client.go b/go/master/client.go index 985b96b0af4103beef9045db318fd6d3fe32a233..05383f1bf40c0e2b1f974e802ee8fd6aac109b00 100644 --- a/go/master/client.go +++ b/go/master/client.go @@ -2,18 +2,12 @@ package master import ( "os" - "time" "github.com/PaddlePaddle/Paddle/go/connection" "github.com/PaddlePaddle/recordio" log "github.com/sirupsen/logrus" ) -// Addresser provide the address of the master server. -type Addresser interface { - Address() string -} - // Client is the client of the master server. type Client struct { conn *connection.Conn @@ -29,11 +23,11 @@ type record struct { // // bufSize is the record buffer size. NextRecord will read from this // buffer. -func NewClient(addr Addresser, bufSize int) *Client { +func NewClient(addrCh <-chan string, bufSize int) *Client { c := &Client{} c.conn = connection.New() c.ch = make(chan record, bufSize) - go c.monitorMaster(addr) + go c.monitorMaster(addrCh) go c.getRecords() return c } @@ -78,12 +72,10 @@ func (c *Client) getRecords() { } } -func (c *Client) monitorMaster(addr Addresser) { +func (c *Client) monitorMaster(addrCh <-chan string) { lastMaster := "" - monitor := func() { - // get the lastest address of the master server, + for curMaster := range addrCh { // connect to the new address once address changed. - curMaster := addr.Address() if curMaster != lastMaster { if curMaster == "" { err := c.conn.Close() @@ -100,18 +92,10 @@ func (c *Client) monitorMaster(addr Addresser) { // to retry next time. curMaster = lastMaster } - } } - lastMaster = curMaster } - - monitor() - ticker := time.NewTicker(10 * time.Second) - for _ = range ticker.C { - monitor() - } } // SetDataset set dataset for the master server to dispatch. diff --git a/go/master/client_internal_test.go b/go/master/client_internal_test.go index 251225780ae3077f90655b4e874d03b4f3794525..364dce7b58cf6366af711bde9107559a762563a4 100644 --- a/go/master/client_internal_test.go +++ b/go/master/client_internal_test.go @@ -26,12 +26,6 @@ func init() { log.SetLevel(log.ErrorLevel) } -type TestAddresser string - -func (a TestAddresser) Address() string { - return string(a) -} - func TestGetFinishTask(t *testing.T) { const path = "/tmp/master_client_test_0" @@ -45,7 +39,6 @@ func TestGetFinishTask(t *testing.T) { if err != nil { panic(err) } - go func(l net.Listener) { s, err := NewService(&InMemStore{}, chunkPerTask, time.Second, 1) if err != nil { @@ -82,9 +75,11 @@ func TestGetFinishTask(t *testing.T) { // Manually intialize client to avoid calling c.getRecords() c := &Client{} c.conn = connection.New() - go c.monitorMaster(TestAddresser(fmt.Sprintf(":%d", p))) + addr := fmt.Sprintf(":%d", p) + ch := make(chan string, 1) + ch <- addr + go c.monitorMaster(ch) c.SetDataset([]string{path}) - checkOnePass := func(i int) { var tasks []Task for idx := 0; idx < totalTask; idx++ { diff --git a/go/master/client_test.go b/go/master/client_test.go index 0a401d8a43c7616e7052f8cb15709b411e18755a..6666d3860c412daa8711fbfa2d729a261b3fd887 100644 --- a/go/master/client_test.go +++ b/go/master/client_test.go @@ -20,7 +20,6 @@ func TestNextRecord(t *testing.T) { path = "/tmp/master_client_TestFull" total = 50 ) - l, err := net.Listen("tcp", ":0") if err != nil { panic(err) @@ -31,7 +30,6 @@ func TestNextRecord(t *testing.T) { if err != nil { panic(err) } - go func(l net.Listener) { s, err := master.NewService(&master.InMemStore{}, 10, time.Second, 1) if err != nil { @@ -63,10 +61,10 @@ func TestNextRecord(t *testing.T) { } w.Close() f.Close() - - c := master.NewClient(master.TestAddresser(fmt.Sprintf(":%d", p)), 10) + curAddr := make(chan string, 1) + curAddr <- fmt.Sprintf(":%d", p) + c := master.NewClient(curAddr, 10) c.SetDataset([]string{path}) - for pass := 0; pass < 50; pass++ { received := make(map[byte]bool) for i := 0; i < total; i++ { diff --git a/go/master/etcd_client.go b/go/master/etcd_client.go index b7293a759896f113d630d57d14b4b4ac8963f54a..e27c014792f31ca27fe1a1636d69acccc4206ea3 100644 --- a/go/master/etcd_client.go +++ b/go/master/etcd_client.go @@ -18,8 +18,8 @@ const ( DefaultAddrPath = "/master/addr" ) -// EtcdClient is the etcd client that master uses for fault tolerance -// and service registry. +// EtcdClient is the etcd client that the master uses for fault +// tolerance and service registry. type EtcdClient struct { lockPath string statePath string @@ -142,3 +142,31 @@ func (e *EtcdClient) Load() ([]byte, error) { state := kvs[0].Value return state, nil } + +// GetKey gets the value by the specify key. +func GetKey(c *clientv3.Client, key string, timeout int) (string, error) { + ctx, cancel := context.WithTimeout(context.Background(), time.Second*time.Duration(timeout)) + resp, err := c.Get(ctx, key) + cancel() + if err != nil { + return "", err + } + kvs := resp.Kvs + if len(kvs) == 0 { + return "", nil + } + v := kvs[0].Value + return string(v), nil +} + +// WatchKey watches the specify key and send to valChan if there is some event. +func WatchKey(c *clientv3.Client, key string, valChan chan<- string) { + rch := c.Watch(context.Background(), key) + for wresp := range rch { + for _, ev := range wresp.Events { + // if received event is DELETE, the value will be an empty string + log.Infof("received event %s, %q : %q\n", ev.Type, ev.Kv.Key, ev.Kv.Value) + valChan <- string(ev.Kv.Value) + } + } +} diff --git a/go/pserver/cclient/test/CMakeLists.txt b/go/pserver/cclient/test/CMakeLists.txt index 916e4e99a24ea7f76f1935fc7d281cd158ac5061..170730ccebbae9c99ebafe360261c32f5b2f4e08 100644 --- a/go/pserver/cclient/test/CMakeLists.txt +++ b/go/pserver/cclient/test/CMakeLists.txt @@ -1,3 +1,3 @@ -cc_library(main SRCS main.c DEPS paddle_pserver_cclient) +cc_binary(main SRCS main.c DEPS paddle_pserver_cclient) cc_test(test_cclient SRCS test_cclient.c DEPS paddle_pserver_cclient) diff --git a/go/pserver/client.go b/go/pserver/client.go index dda915977282d4880ddcc8c18ef6fd80ede9e01b..6938b9d5ce6f6d73c05bd6e3154777023965c319 100644 --- a/go/pserver/client.go +++ b/go/pserver/client.go @@ -1,6 +1,7 @@ package pserver import ( + "errors" "hash/fnv" "sort" "time" @@ -123,6 +124,9 @@ func (c *Client) FinishInitParams() error { // SendGrads sends gradients to parameter servers for updating // parameters. func (c *Client) SendGrads(grads []Gradient) error { + if len(grads) == 0 { + return errors.New("no gradient received") + } errCh := make(chan error, len(grads)) for _, g := range grads { go func(g Gradient) { diff --git a/go/pserver/client_test.go b/go/pserver/client_test.go index 6ecf1fa08a02ed2ce04fae0903cebd46a7b768a4..5bd16118a7f70b766016abfce55f6bb2adf8cc60 100644 --- a/go/pserver/client_test.go +++ b/go/pserver/client_test.go @@ -7,7 +7,6 @@ import ( "strconv" "strings" "testing" - "time" "github.com/PaddlePaddle/Paddle/go/pserver" ) @@ -31,7 +30,7 @@ func init() { port[i] = p go func(l net.Listener) { - s, err := pserver.NewService("", time.Second*5) + s, err := pserver.NewService(0) if err != nil { panic(err) } diff --git a/go/pserver/etcd_client.go b/go/pserver/etcd_client.go new file mode 100644 index 0000000000000000000000000000000000000000..4d88243edd4aa817ddc263ba316a3f6be9e1e67f --- /dev/null +++ b/go/pserver/etcd_client.go @@ -0,0 +1,181 @@ +package pserver + +import ( + "context" + "errors" + "strconv" + "strings" + "time" + + "github.com/PaddlePaddle/Paddle/go/utils/networkhelper" + "github.com/coreos/etcd/clientv3" + "github.com/coreos/etcd/clientv3/concurrency" + log "github.com/sirupsen/logrus" +) + +// EtcdClient is the etcd client that the pserver uses for fault +// tolerance, service registry and coordination. +type EtcdClient struct { + numPservers int + etcdEndpoints string + etcdClient *clientv3.Client + // etcdTimeout is also used as retry intervals. + etcdTimeout time.Duration + // FIXME: ensure GetExternalIP gets the correct ip for trainers to connect. + externalIP string + // desired number of pservers in the job. + // assume desired will not change during one training job. + desired int +} + +// NewEtcdClient creates an EtcdClient +func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *EtcdClient { + return &EtcdClient{ + etcdTimeout: timeout, + numPservers: numPservers, + etcdEndpoints: endpoints, + } +} + +// Register registers the pserver on etcd +// +// Register returns the index of the current pserver. +func (e *EtcdClient) Register() (int, error) { + + var err error + e.externalIP, err = networkhelper.GetExternalIP() + if err != nil { + return 0, err + } + + // initialize connection to etcd. + ep := strings.Split(e.etcdEndpoints, ",") + for { + cli, err := clientv3.New(clientv3.Config{ + Endpoints: ep, + DialTimeout: e.etcdTimeout, + }) + if err != nil { + log.Errorf("connect to etcd error: %v", err) + time.Sleep(e.etcdTimeout) + continue + } + e.etcdClient = cli + log.Debugf("inited client to %s", e.etcdEndpoints) + break + } + // init /ps_desired using transaction, for multiple pservers may want to write + // it at the same time. + for { + ctx, cancel := context.WithTimeout(context.Background(), time.Second) + _, err := e.initDesiredPsercers(ctx, e.numPservers) + cancel() + if err != nil { + log.Warn(err) + time.Sleep(e.etcdTimeout) + continue + } + break + } + // TODO: when implementing extending or reducing pservers, /ps_desired is + // changed, then we need to watch /ps_desired node for events. For now, just + // write once when init and read from it. + // wait and set s.desired init value + for { + ctx, cancel := context.WithTimeout(context.Background(), time.Second) + resp, err := e.etcdClient.Get(ctx, PsDesired) + cancel() + if err != nil { + log.Errorf("getting %s error: %v", PsDesired, err) + time.Sleep(e.etcdTimeout) + continue + } + if len(resp.Kvs) != 0 { + e.desired, err = strconv.Atoi(string(resp.Kvs[0].Value)) + if err != nil { + log.Errorf("value of %s invalid %v\n", PsDesired, err) + time.Sleep(e.etcdTimeout) + // NOTE: wait util ps_desired value change + continue + } + break + } + } + + var pserverIdx int + // try register pserver node on etcd + for { + ctx, cancel := context.WithTimeout(context.Background(), time.Second) + var err error + pserverIdx, err = e.registerPserverEtcd(ctx) + cancel() + if err != nil { + log.Warn(err) + time.Sleep(e.etcdTimeout) + continue + } + break + } + + return pserverIdx, nil +} + +func (e *EtcdClient) initDesiredPsercers(ctx context.Context, numPservers int) (*clientv3.TxnResponse, error) { + return concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error { + dsStr := c.Get(PsDesired) + if dsStr == "" { + c.Put(PsDesired, strconv.Itoa(numPservers)) + } + return nil + }, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads)) +} + +// registerPserverEtcd registers pserver node on etcd using transaction. +func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) { + var idx int + _, err := concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error { + registered := false + for i := 0; i < e.desired; i++ { + psKey := "/ps/" + strconv.Itoa(i) + log.Debugf("checking %s", psKey) + ps := c.Get(psKey) + log.Debugf("got value (%s) for key: %s", ps, psKey) + + if ps == "" { + resp, err := e.etcdClient.Grant(context.TODO(), 5) + if err != nil { + log.Fatal(err) + } + // find the first id and write info + c.Put(psKey, e.externalIP, clientv3.WithLease(resp.ID)) + log.Debugf("set pserver node %s with value %s", psKey, e.externalIP) + ch, kaerr := e.etcdClient.KeepAlive(context.TODO(), resp.ID) + if kaerr != nil { + log.Errorf("keepalive etcd node error: %v", kaerr) + return kaerr + } + + // Eat the keep alive message so etcd + // will not expire the lease. + go func(ch <-chan *clientv3.LeaseKeepAliveResponse) { + ka := <-ch + log.Debugf("keepalive: %d\n", ka.TTL) + }(ch) + log.Debug("register finished") + idx = i + registered = true + break + } + } + if registered == true { + return nil + } + return errors.New("not registerd, may due to already have enough pservers") + }, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads)) + + if err != nil { + return 0, err + } + + return idx, nil +} diff --git a/go/pserver/service.go b/go/pserver/service.go index f966595fdccbf23e23f94a857503ce05815164ef..f386ebea1eb8659a988de2a807303bb6687fa429 100644 --- a/go/pserver/service.go +++ b/go/pserver/service.go @@ -1,18 +1,9 @@ package pserver import ( - "context" "errors" "fmt" - "strconv" - "strings" "sync" - "time" - - "github.com/PaddlePaddle/Paddle/go/utils/networkhelper" - "github.com/coreos/etcd/clientv3" - "github.com/coreos/etcd/clientv3/concurrency" - log "github.com/sirupsen/logrus" ) // ElementType is the type of elements of a Parameter. @@ -55,160 +46,25 @@ type Gradient Parameter // Service is the RPC service for pserver. type Service struct { initialized chan struct{} + idx int mu sync.Mutex opt *optimizer paramMap map[string]Parameter - - etcdEndpoints string - etcdClient *clientv3.Client - // etcdTimeout is also used as retry intervals. - etcdTimeout time.Duration - // desired number of pservers in the job. - // assume desired will not change during one training job. - desired int - // FIXME: ensure GetExternalIP gets the correct ip for trainers to connect. - externalIP string } // NewService creates a new service, will bypass etcd registration if no // endpoints specified. -func NewService(endpoints string, numPservers int, timeout time.Duration) (*Service, error) { - s := &Service{opt: newOptimizer(sgd, 0.005)} +func NewService(idx int) (*Service, error) { + s := &Service{ + idx: idx, + opt: newOptimizer(sgd, 0.005), + } s.paramMap = make(map[string]Parameter) s.initialized = make(chan struct{}) - s.etcdEndpoints = endpoints - s.etcdTimeout = timeout - - var err error - s.externalIP, err = networkhelper.GetExternalIP() - if err != nil { - return nil, err - } - - if endpoints != "" { - // initialize connection to etcd, try - ep := strings.Split(s.etcdEndpoints, ",") - for { - cli, err := clientv3.New(clientv3.Config{ - Endpoints: ep, - DialTimeout: s.etcdTimeout, - }) - if err != nil { - log.Errorf("connect to etcd error: %v", err) - time.Sleep(s.etcdTimeout) - continue - } - s.etcdClient = cli - log.Debugf("inited client to %s", s.etcdEndpoints) - break - } - // init /ps_desired using transaction, for multiple pservers may want to write - // it at the same time. - for { - ctx, cancel := context.WithTimeout(context.Background(), time.Second) - _, err := s.initDesiredPsercers(ctx, numPservers) - cancel() - if err != nil { - log.Warn(err) - time.Sleep(s.etcdTimeout) - continue - } - break - } - // TODO: when implementing extending or reducing pservers, /ps_desired is - // changed, then we need to watch /ps_desired node for events. For now, just - // write once when init and read from it. - // wait and set s.desired init value - for { - ctx, cancel := context.WithTimeout(context.Background(), time.Second) - resp, err := s.etcdClient.Get(ctx, PsDesired) - cancel() - if err != nil { - log.Errorf("getting %s error: %v", PsDesired, err) - time.Sleep(s.etcdTimeout) - continue - } - if len(resp.Kvs) != 0 { - s.desired, err = strconv.Atoi(string(resp.Kvs[0].Value)) - if err != nil { - log.Errorf("value of %s invalid %v\n", PsDesired, err) - time.Sleep(s.etcdTimeout) - // NOTE: wait util ps_desired value change - continue - } - break - } - } - // try register pserver node on etcd - for { - ctx, cancel := context.WithTimeout(context.Background(), time.Second) - _, err := s.registerPserverEtcd(ctx) - cancel() - if err != nil { - log.Warn(err) - time.Sleep(s.etcdTimeout) - continue - } - break - } - } // if endpoints != "" - // Bypass etcd registration if no endpoints specified return s, nil } -func (s *Service) initDesiredPsercers(ctx context.Context, numPservers int) (*clientv3.TxnResponse, error) { - return concurrency.NewSTM(s.etcdClient, func(c concurrency.STM) error { - dsStr := c.Get(PsDesired) - if dsStr == "" { - c.Put(PsDesired, strconv.Itoa(numPservers)) - } - return nil - }, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads)) -} - -// registerPserverEtcd registers pserver node on etcd using transaction. -func (s *Service) registerPserverEtcd(ctx context.Context) (*clientv3.TxnResponse, error) { - return concurrency.NewSTM(s.etcdClient, func(c concurrency.STM) error { - registered := false - for i := 0; i < s.desired; i++ { - psKey := "/ps/" + strconv.Itoa(i) - log.Debugf("checking %s", psKey) - ps := c.Get(psKey) - log.Debugf("got value (%s) for key: %s", ps, psKey) - - if ps == "" { - resp, err := s.etcdClient.Grant(context.TODO(), 5) - if err != nil { - log.Fatal(err) - } - // find the first id and write info - c.Put(psKey, s.externalIP, clientv3.WithLease(resp.ID)) - log.Debugf("set pserver node %s with value %s", psKey, s.externalIP) - ch, kaerr := s.etcdClient.KeepAlive(context.TODO(), resp.ID) - if kaerr != nil { - log.Errorf("keepalive etcd node error: %v", kaerr) - return kaerr - } - - // Eat the keep alive message so etcd - // will not expire the lease. - go func(ch <-chan *clientv3.LeaseKeepAliveResponse) { - ka := <-ch - log.Debugf("keepalive: %d\n", ka.TTL) - }(ch) - log.Debug("register finished") - registered = true - break - } - } - if registered == true { - return nil - } - return errors.New("not registerd, may due to already have enough pservers") - }, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads)) -} - // InitParam initializes a parameter. func (s *Service) InitParam(paramWithConfigs ParameterWithConfig, dummy *int) error { select { diff --git a/go/pserver/service_test.go b/go/pserver/service_test.go index f317535592165b921491120888badd30c6795c12..d9d887cffd462eed48b972466a7d83bae35d9a1c 100644 --- a/go/pserver/service_test.go +++ b/go/pserver/service_test.go @@ -10,7 +10,7 @@ import ( ) func TestFull(t *testing.T) { - s, err := pserver.NewService("", time.Second*5) + s, err := pserver.NewService(0) if err != nil { t.Error(err) } @@ -75,7 +75,7 @@ func TestFull(t *testing.T) { } func TestMultipleInit(t *testing.T) { - s, err := pserver.NewService("", time.Second*5) + s, err := pserver.NewService(0) if err != nil { t.Error(err) } @@ -91,7 +91,7 @@ func TestMultipleInit(t *testing.T) { } func TestUninitialized(t *testing.T) { - s, err := pserver.NewService("", time.Second*5) + s, err := pserver.NewService(0) err = s.SendGrad(pserver.Gradient{}, nil) if err.Error() != pserver.Uninitialized { t.FailNow() @@ -99,7 +99,7 @@ func TestUninitialized(t *testing.T) { } func TestBlockUntilInitialized(t *testing.T) { - s, err := pserver.NewService("", time.Second*5) + s, err := pserver.NewService(0) if err != nil { t.Error(err) } diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index 573bd937a351a6f308974e14f3bc92cbe1b541bc..307e99bbe3a833f1fe26057ec38d0b96e04bc0fe 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -9,17 +9,10 @@ add_subdirectory(pserver) add_subdirectory(trainer) add_subdirectory(scripts) add_subdirectory(optimizer) -add_subdirectory(strings) - -# Do not build go directory until go cmake is working smoothly. -# if(CMAKE_Go_COMPILER) -# add_subdirectory(go) -# endif() - -find_package(Boost QUIET) +add_subdirectory(string) if(Boost_FOUND) - include_directories(${Boost_INCLUDE_DIRS}) + add_subdirectory(memory) add_subdirectory(platform) add_subdirectory(framework) endif() diff --git a/paddle/api/CMakeLists.txt b/paddle/api/CMakeLists.txt index f2315e31cc06d8b5fea7a9fd203a697bac603a90..39d8aa075bc072d37dc8df67746f0d2b503418a6 100644 --- a/paddle/api/CMakeLists.txt +++ b/paddle/api/CMakeLists.txt @@ -16,7 +16,7 @@ set(API_HEADER Internal.h) add_library(paddle_api STATIC ${API_SOURCES}) -add_dependencies(paddle_api gen_proto_cpp paddle_trainer_lib) +add_dependencies(paddle_api paddle_proto paddle_trainer_lib) INCLUDE(${SWIG_USE_FILE}) INCLUDE_DIRECTORIES(${PROJ_ROOT}/paddle) diff --git a/paddle/capi/CMakeLists.txt b/paddle/capi/CMakeLists.txt index 206f512563466d40e9ad1db0ddb4753ffb6bf55a..11022d17541476c97a2b29be8eb8fecce7e39435 100644 --- a/paddle/capi/CMakeLists.txt +++ b/paddle/capi/CMakeLists.txt @@ -26,7 +26,7 @@ target_include_directories(paddle_capi PUBLIC ${CMAKE_CURRENT_BINARY_DIR}) add_style_check_target(paddle_capi ${CAPI_SOURCES} ${CAPI_HEADER} ${CAPI_PRIVATE_HEADER}) -add_dependencies(paddle_capi gen_proto_cpp) +add_dependencies(paddle_capi paddle_proto) # combine all paddle static libraries together, into libpaddle_capi_whole.a diff --git a/paddle/cuda/CMakeLists.txt b/paddle/cuda/CMakeLists.txt index f9061e96deb659dcf7bfb88b46e6509af0425199..73ffa690d9d91b673079fc0ecf91f17cbabfdb1e 100755 --- a/paddle/cuda/CMakeLists.txt +++ b/paddle/cuda/CMakeLists.txt @@ -83,7 +83,7 @@ else() ${CUDA_CXX_SOURCES}) endif() -add_dependencies(paddle_cuda ${external_project_dependencies}) +add_dependencies(paddle_cuda paddle_proto ${external_project_dependencies}) add_style_check_target(paddle_cuda ${CUDA_SOURCES} diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt index e3c3155aa902c941058ea1b15488360df6c06175..6aa6b9bc2db6a223dd8562b76ba9d777206bfd40 100644 --- a/paddle/framework/CMakeLists.txt +++ b/paddle/framework/CMakeLists.txt @@ -1,6 +1,7 @@ +# ddim lib cc_library(ddim SRCS ddim.cc) cc_test(ddim_test SRCS ddim_test.cc DEPS ddim) - nv_test(dim_test SRCS dim_test.cu DEPS ddim) - cc_test(variable_test SRCS variable_test.cc) +cc_test(scope_test SRCS scope_test.cc) +cc_test(enforce_test SRCS enforce_test.cc) diff --git a/paddle/framework/enforce.h b/paddle/framework/enforce.h new file mode 100644 index 0000000000000000000000000000000000000000..56cb7f95647e81efef58b156002d0d378ee22820 --- /dev/null +++ b/paddle/framework/enforce.h @@ -0,0 +1,69 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include + +namespace paddle { +namespace framework { + +/** + * @brief Enforce exception. Inherits std::exception + * + * All enforce condition not met, will throw an EnforceNotMet exception. + */ +class EnforceNotMet : public std::exception { + public: + EnforceNotMet(const std::string& msg, const char* file, int fileline) { + std::ostringstream sout; + sout << msg << " at [" << file << ":" << fileline << "];"; + all_msg_ = sout.str(); + } + + const char* what() const noexcept override { return all_msg_.c_str(); } + + private: + std::string all_msg_; +}; + +// From https://stackoverflow.com/questions/30130930/ +// __buildin_expect is in C++ 11 standard. Since the condition which enforced +// should be true in most situation, it will make the compiler generate faster +// code by adding `UNLIKELY` macro. +#define UNLIKELY(condition) __builtin_expect(static_cast(condition), 0) + +/** + * @brief Throw a EnforceNotMet exception, automatically filled __FILE__ & + * __LINE__ + * + * This macro take __VA_ARGS__, user can pass any type if that type can + * serialize to std::ostream + */ +#define PADDLE_THROW(...) \ + do { \ + throw ::paddle::framework::EnforceNotMet( \ + ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__); \ + } while (0) + +/** + * @brief Enforce a condition, otherwise throw an EnforceNotMet + */ +#define PADDLE_ENFORCE(condition, ...) \ + do { \ + if (UNLIKELY(!(condition))) { \ + PADDLE_THROW(__VA_ARGS__); \ + } \ + } while (0) + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/enforce_test.cc b/paddle/framework/enforce_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..f8da1a192f63a54324d80725c9d2f156fb11a481 --- /dev/null +++ b/paddle/framework/enforce_test.cc @@ -0,0 +1,35 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include + +TEST(ENFORCE, OK) { + PADDLE_ENFORCE(true, "Enforce is ok %d now %f", 123, 0.345); + size_t val = 1; + const size_t limit = 10; + PADDLE_ENFORCE(val < limit, "Enforce is OK too"); +} + +TEST(ENFORCE, FAILED) { + bool in_catch = false; + try { + PADDLE_ENFORCE(false, "Enforce is not ok %d at all", 123); + } catch (paddle::framework::EnforceNotMet err) { + in_catch = true; + std::string msg = "Enforce is not ok 123 at all"; + const char* what = err.what(); + for (size_t i = 0; i < msg.length(); ++i) { + ASSERT_EQ(what[i], msg[i]); + } + } + ASSERT_TRUE(in_catch); +} \ No newline at end of file diff --git a/paddle/framework/scope.h b/paddle/framework/scope.h new file mode 100644 index 0000000000000000000000000000000000000000..a4470f726fb0d59a82db29b3239c111ea1569c55 --- /dev/null +++ b/paddle/framework/scope.h @@ -0,0 +1,95 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +#include "paddle/framework/variable.h" + +namespace paddle { +namespace framework { + +/** + * @brief Scope that manage all variables. + * + * Scope is an association of a name to Variable. All variables belong to + * Scope. You need to specify a scope to run a Net, i.e., `net.Run(&scope)`. + * One net can run in different scopes and update different variable in the + * scope. + */ +class Scope { + public: + /** + * @brief Initialize s Scope without parent. + */ + Scope() {} + + /** + * @brief Initialize a Scope with parent. + */ + explicit Scope(const std::shared_ptr& parent) : parent_(parent) {} + + /** + * @brief Create Variable + * + * Create Variable in this Scope. Return the exist one if Variable already + * been created. + */ + Variable* CreateVariable(const std::string& name) { + auto var = GetVariable(name); + if (var) { + return var; + } else { + vars_[name] = std::unique_ptr(new Variable()); + return GetVariable(name); + } + } + + /** + * @brief Get Variable. + * + * Get Variable from this Scope, this function will recursive find Variable + * from it's parent scope. Return nullptr if not found. + */ + Variable* GetVariable(const std::string& name) const { + auto it = vars_.find(name); + if (it != vars_.end()) { + return it->second.get(); + } else if (parent_ != nullptr) { + return parent_->GetVariable(name); + } else { + return nullptr; + } + } + + /** + * @brief If this scope has a Var named name. + * + * Find if there is a Variable in this scope and it's parent scope + */ + bool HasVariable(const std::string& name) const { + return (vars_.find(name) != vars_.end() || + (parent_ && parent_->HasVariable(name))); + } + + private: + std::unordered_map> vars_; + std::shared_ptr parent_{nullptr}; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/scope_test.cc b/paddle/framework/scope_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..df1afb200ce9e75c5b1e40f2da56667487ae3576 --- /dev/null +++ b/paddle/framework/scope_test.cc @@ -0,0 +1,58 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/framework/scope.h" +#include "gtest/gtest.h" + +TEST(Scope, Create) { + using paddle::framework::Scope; + using paddle::framework::Variable; + + auto scope = std::make_shared(); + + Variable* var0 = scope->CreateVariable(""); + EXPECT_NE(var0, nullptr); + + /// GetVariable will return nullptr if not exist. + Variable* var1 = scope->GetVariable("a"); + EXPECT_EQ(var1, nullptr); + + /// CreateVariable will return one. + Variable* var2 = scope->CreateVariable("a"); + EXPECT_NE(var2, nullptr); + + /// Get the created variable. + Variable* var3 = scope->GetVariable("a"); + EXPECT_EQ(var2, var3); + + /// CreateVariable will just return the variable if it's + /// already exist. + Variable* var4 = scope->CreateVariable("a"); + EXPECT_EQ(var4, var2); +} + +TEST(Scope, Parent) { + using paddle::framework::Scope; + using paddle::framework::Variable; + + auto parent_scope = std::make_shared(); + auto scope = std::make_shared(parent_scope); + + Variable* var0 = parent_scope->CreateVariable("a"); + EXPECT_NE(var0, nullptr); + + /// GetVariable will get Variable from parent scope if exist. + Variable* var1 = scope->GetVariable("a"); + EXPECT_EQ(var0, var1); +} diff --git a/paddle/function/CMakeLists.txt b/paddle/function/CMakeLists.txt index 5e170714cf5b183fcf6e76d34746333397e6b060..1518a8a654cfb54376a49760dc5873733c916937 100644 --- a/paddle/function/CMakeLists.txt +++ b/paddle/function/CMakeLists.txt @@ -10,9 +10,17 @@ if(WITH_GPU) cuda_compile(cu_objs ${cu_files}) endif() +if(USE_NNPACK) + include(nnpack/nnpack.cmake) + list(APPEND cpp_files nnpack/NNPACKConvOp.cpp) + if(WITH_TESTING) + add_unittest(NNPACKConvOpTest nnpack/NNPACKConvOpTest.cpp) + endif() +endif() + add_library(paddle_function STATIC ${cpp_files} ${cu_objs}) add_dependencies(paddle_function ${external_project_dependencies}) -add_dependencies(paddle_function gen_proto_cpp) +add_dependencies(paddle_function paddle_proto) if(WITH_TESTING) if(WITH_GPU) diff --git a/paddle/function/nnpack/NNPACKConvOp.cpp b/paddle/function/nnpack/NNPACKConvOp.cpp new file mode 100644 index 0000000000000000000000000000000000000000..e8080c3d714b324f072a380f738b9764477dfe04 --- /dev/null +++ b/paddle/function/nnpack/NNPACKConvOp.cpp @@ -0,0 +1,238 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "nnpack.h" +#include "paddle/function/ConvOp.h" + +DEFINE_bool(nnpack_allocate_outside, + false, + "Allocate and free workspace memory outside the NNPACK interface."); +DEFINE_int32(nnpack_num_threads, + 0, + "The number of nnpack threads" + "default: 0; 0 to disable threadpool."); + +namespace paddle { + +nnp_convolution_algorithm get_nnp_convolution_algorithm( + const std::string& algorithm) { + if (algorithm == "auto") { + return nnp_convolution_algorithm_auto; + } else if (algorithm == "ft8x8") { + return nnp_convolution_algorithm_ft8x8; + } else if (algorithm == "ft16x16") { + return nnp_convolution_algorithm_ft16x16; + } else if (algorithm == "wt8x8") { + return nnp_convolution_algorithm_wt8x8; + } else if (algorithm == "implicit-gemm") { + return nnp_convolution_algorithm_implicit_gemm; + } else if (algorithm == "direct") { + return nnp_convolution_algorithm_direct; + } else { + return nnp_convolution_algorithm_auto; + } +} + +template +class NNPACKConvFunction : public ConvFunctionBase { +public: + void init(const FuncConfig& config) override { + ConvFunctionBase::init(config); + CHECK_EQ(groups_, (size_t)1); + algorithm_ = get_nnp_convolution_algorithm(config.get("algo")); + // algorithm_ = nnp_convolution_algorithm_auto; + transform_strategy_ = nnp_convolution_transform_strategy_compute; + nnp_status status = nnp_initialize(); + CHECK_EQ(status, nnp_status_success); + workspaceBuffer_ = nullptr; + workspaceSize_ = 0; + + threadpool_ = nullptr; + if (FLAGS_nnpack_num_threads) { + threadpool_ = pthreadpool_create(FLAGS_nnpack_num_threads); + VLOG(3) << "Number of threads " + << pthreadpool_get_threads_count(threadpool_); + } + } + + ~NNPACKConvFunction() { + if (threadpool_) { + pthreadpool_destroy(threadpool_); + } + if (workspaceBuffer_) { + free(workspaceBuffer_); + } + } + + virtual void check(const BufferArgs& inputs, + const BufferArgs& outputs) override { + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + checkShape(input, filter, output); + } + + void calc(const BufferArgs& inputs, const BufferArgs& outputs) override { + CHECK_EQ(numInputs_, inputs.size()); + CHECK_EQ(numOutputs_, outputs.size()); + CHECK_EQ(outputs[0].getArgType(), ASSIGN_TO); + check(inputs, outputs); + const TensorShape& input = inputs[0].shape(); + const TensorShape& filter = inputs[1].shape(); + const TensorShape& output = outputs[0].shape(); + + size_t batchSize = input[0]; + size_t inputChannels = input[1]; + size_t inputHeight = input[2]; + size_t inputWidth = input[3]; + size_t filterHeight = getFilterHeight(filter); + size_t filterWidth = getFilterWidth(filter); + size_t outputChannels = output[1]; + // size_t outputHeight = output[2]; + // size_t outputWidth = output[3]; + + nnp_size inputSize = {.width = inputWidth, .height = inputHeight}; + nnp_padding padding = {.top = (size_t)paddingH(), + .right = (size_t)paddingW(), + .bottom = (size_t)paddingH(), + .left = (size_t)paddingW()}; + nnp_size kernelSize = {.width = filterWidth, .height = filterHeight}; + nnp_size outputSubsampling = {.width = (size_t)strideW(), + .height = (size_t)strideH()}; + + float* inputData = inputs[0].data(); + float* filterData = inputs[1].data(); + float* outputData = outputs[0].data(); + + void* bufferPtr = nullptr; + size_t* sizePtr = nullptr; + size_t needSize; + if (FLAGS_nnpack_allocate_outside) { + if (batchSize == 1) { + nnp_status status = nnp_convolution_inference(algorithm_, + transform_strategy_, + inputChannels, + outputChannels, + inputSize, + padding, + kernelSize, + outputSubsampling, + nullptr, + nullptr, + nullptr, + nullptr, + nullptr, + &needSize, + nnp_activation_identity, + nullptr, + nullptr, + nullptr); + CHECK_EQ(status, nnp_status_success); + } else { + // only supports stride = 1 + CHECK_EQ(strideH(), 1); + CHECK_EQ(strideW(), 1); + nnp_status status = nnp_convolution_output(algorithm_, + batchSize, + inputChannels, + outputChannels, + inputSize, + padding, + kernelSize, + nullptr, + nullptr, + nullptr, + nullptr, + nullptr, + &needSize, + nnp_activation_identity, + nullptr, + nullptr, + nullptr); + CHECK_EQ(status, nnp_status_success); + } + + VLOG(3) << "workspace size is " << needSize; + if (needSize > workspaceSize_) { + workspaceSize_ = needSize; + if (workspaceBuffer_) { + free(workspaceBuffer_); + } else { + posix_memalign(&workspaceBuffer_, 64, needSize); + } + } + + if (needSize) { + bufferPtr = workspaceBuffer_; + sizePtr = &needSize; + } + } + + if (batchSize == 1) { + nnp_status status = + nnp_convolution_inference(algorithm_, + transform_strategy_, + inputChannels, + outputChannels, + inputSize, + padding, + kernelSize, + outputSubsampling, + inputData, + filterData, + nullptr, /* bias */ + outputData, + bufferPtr, + sizePtr, + nnp_activation_identity, + nullptr, + threadpool_, /* threadpool */ + nullptr); + CHECK_EQ(status, nnp_status_success); + } else { + // only supports stride = 1 + CHECK_EQ(strideH(), 1); + CHECK_EQ(strideW(), 1); + nnp_status status = nnp_convolution_output(algorithm_, + batchSize, + inputChannels, + outputChannels, + inputSize, + padding, + kernelSize, + inputData, + filterData, + nullptr, /* bias */ + outputData, + bufferPtr, + sizePtr, + nnp_activation_identity, + nullptr, + threadpool_, /* threadpool */ + nullptr); + CHECK_EQ(status, nnp_status_success); + } + } + +private: + nnp_convolution_algorithm algorithm_; + nnp_convolution_transform_strategy transform_strategy_; + void* workspaceBuffer_; + size_t workspaceSize_; + pthreadpool_t threadpool_; +}; + +REGISTER_TYPED_FUNC(NNPACKConv, CPU, NNPACKConvFunction); + +} // namespace paddle diff --git a/paddle/function/nnpack/NNPACKConvOpTest.cpp b/paddle/function/nnpack/NNPACKConvOpTest.cpp new file mode 100644 index 0000000000000000000000000000000000000000..48180112111c67f36ddd425008187201655089c9 --- /dev/null +++ b/paddle/function/nnpack/NNPACKConvOpTest.cpp @@ -0,0 +1,99 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/function/Function.h" +#include "paddle/function/FunctionTest.h" + +DEFINE_string(algo, + "auto", + "The algorithm (auto, ft8x8, ft16x16, wt8x8, " + "implicit-gemm, or direct) for computing convolution of NNPACK."); + +namespace paddle { + +#define IS_NNPACK_SUPPORT(algo, filterSize, stride) \ + if (algo == "direct" && filterSize != 1) continue; \ + if (algo == "direct" && batchSize != 1) continue; \ + if (algo == "wt8x8" && filterSize != 3) continue; \ + if (algo == "implicit-gemm" && batchSize != 1) continue; \ + if (algo != "auto" && algo != "implicit-gemm" && stride > 1) continue; + +class ConvolutionTest { +public: + ConvolutionTest(const std::string& conv1, + const std::string& conv2, + std::string algo = "auto") { + for (size_t batchSize : {1, 32}) { + for (size_t inputSize : {7, 14, 54}) { + for (size_t filterSize : {1, 3, 5}) { + for (size_t inputChannels : {3, 64}) { + for (size_t outputChannels : {3, 64, 128}) { + if (inputChannels < outputChannels) break; + for (size_t stride : {1, 2}) { + // if batchSize > 1 NNPACKConv only supports stride = 1 + if (batchSize > 1 && stride > 1) break; + for (size_t padding : {0, 1}) { + if (padding >= filterSize) break; + size_t outputSize = + (inputSize - filterSize + 2 * padding + stride) / stride; + IS_NNPACK_SUPPORT(algo, filterSize, stride); + LOG(INFO) << " batchSize=" << batchSize + << " inputChannels=" << inputChannels + << " inputHeight=" << inputSize + << " inputWidth=" << inputSize + << " outputChannels=" << outputChannels + << " filterHeight=" << filterSize + << " filterWidth=" << filterSize + << " outputHeight=" << outputSize + << " outputWidth=" << outputSize + << " stride=" << stride << " padding=" << padding; + + std::vector paddings = {padding, padding}; + std::vector strides = {stride, stride}; + Compare2Function test( + conv1, + conv2, + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", (size_t)1) + .set("algo", algo)); + + TensorShape shape0{ + batchSize, inputChannels, inputSize, inputSize}; + TensorShape shape1{ + outputChannels, inputChannels, filterSize, filterSize}; + TensorShape shape2{ + batchSize, outputChannels, outputSize, outputSize}; + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape0)); + test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape1)); + test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape2)); + test.run(); + } + } + } + } + } + } + } + } +}; + +TEST(Convolution, NNPACK) { + // NNPACK only supports stride = 1 + ConvolutionTest test("GemmConv-CPU", "NNPACKConv-CPU", FLAGS_algo); +} + +} // namespace paddle diff --git a/paddle/function/nnpack/nnpack.cmake b/paddle/function/nnpack/nnpack.cmake new file mode 100644 index 0000000000000000000000000000000000000000..7182730ae8f133bdc4f73bfc46fa8acbe5f3b603 --- /dev/null +++ b/paddle/function/nnpack/nnpack.cmake @@ -0,0 +1,16 @@ +# 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 ${NNPACK_ROOT}/lib) + +if(NNPACK_INC_DIR AND NNPACK_LIB AND PTHREADPOOL_LIB) + set(NNPACK_FOUND ON) + INCLUDE_DIRECTORIES(${NNPACK_INC_DIR}) +else() + message(FATAL_ERROR "Cannot find NNPACK in (${NNPACK_ROOT})") +endif() diff --git a/paddle/gserver/CMakeLists.txt b/paddle/gserver/CMakeLists.txt index 93a6a99848aa13bb36c9c5c7091fbaa891fc9823..0012636b8f618a1b45cfc801c04781e67694956f 100644 --- a/paddle/gserver/CMakeLists.txt +++ b/paddle/gserver/CMakeLists.txt @@ -58,7 +58,7 @@ endif() add_style_check_target(paddle_gserver ${GSERVER_SOURCES}) add_style_check_target(paddle_gserver ${GSERVER_HEADER}) -add_dependencies(paddle_gserver gen_proto_cpp) +add_dependencies(paddle_gserver paddle_proto ${external_project_dependencies}) if(WITH_TESTING) add_subdirectory(tests) endif() diff --git a/paddle/gserver/gradientmachines/MultiGradientMachine.cpp b/paddle/gserver/gradientmachines/MultiGradientMachine.cpp index 8ef5e9d0c116dd088b5c5c318dfb47c245b471fa..018da6c76dc27a74b074ec52c18347beba8164fc 100644 --- a/paddle/gserver/gradientmachines/MultiGradientMachine.cpp +++ b/paddle/gserver/gradientmachines/MultiGradientMachine.cpp @@ -601,7 +601,7 @@ void TrainerThread::backward() { void TrainerThread::backwardCallback(Parameter* para) { // CPU parameters are merged in the end - if (!para->useGpu()) return; + if (!para->useGpu() || para->isStatic()) return; int paramId = para->getID(); if (multiMachine_->getNumThreads() == 1) { diff --git a/paddle/gserver/layers/ExpandConvLayer.cpp b/paddle/gserver/layers/ExpandConvLayer.cpp index 914689e66cdb8947e886e17e75829183c1af1a42..af79e65a7c09e5a1b55febf1df1e8f5bb61bdcb8 100644 --- a/paddle/gserver/layers/ExpandConvLayer.cpp +++ b/paddle/gserver/layers/ExpandConvLayer.cpp @@ -16,6 +16,10 @@ limitations under the License. */ #include "paddle/utils/Logging.h" #include "paddle/utils/Stat.h" +DEFINE_bool(use_nnpack, + false, + "Whether to use nnpack for convolution calculation."); + namespace paddle { /* @@ -37,26 +41,38 @@ bool ExpandConvLayer::init(const LayerMap &layerMap, for (int i = 0; i < config_.inputs_size(); i++) { std::vector paddings = {(size_t)paddingY_[i], (size_t)padding_[i]}; std::vector strides = {(size_t)strideY_[i], (size_t)stride_[i]}; - createFunction(forward_, - !isDeconv_ ? "GemmConv" : "GemmConvGradInput", - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("groups", (size_t)groups_[i])); - - createFunction(backward_, - !isDeconv_ ? "GemmConvGradInput" : "GemmConv", - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("groups", (size_t)groups_[i])); - - createFunction(backward_, - "GemmConvGradFilter", - FuncConfig() - .set("paddings", paddings) - .set("strides", strides) - .set("groups", (size_t)groups_[i])); + + if (FLAGS_use_nnpack) { + CHECK_EQ(isDeconv_, false); + createFunction(forward_, + "NNPACKConv", + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", (size_t)groups_[i]) + .set("algo", std::string("auto"))); + } else { + createFunction(forward_, + !isDeconv_ ? "GemmConv" : "GemmConvGradInput", + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", (size_t)groups_[i])); + + createFunction(backward_, + !isDeconv_ ? "GemmConvGradInput" : "GemmConv", + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", (size_t)groups_[i])); + + createFunction(backward_, + "GemmConvGradFilter", + FuncConfig() + .set("paddings", paddings) + .set("strides", strides) + .set("groups", (size_t)groups_[i])); + } } return true; } diff --git a/paddle/math/CMakeLists.txt b/paddle/math/CMakeLists.txt index f5657c4690ca71200346efd4e2c5244c02c92eb1..9981de61606bda6baac103592125b929d4c12a3d 100644 --- a/paddle/math/CMakeLists.txt +++ b/paddle/math/CMakeLists.txt @@ -33,7 +33,7 @@ endif() add_style_check_target(paddle_math ${MATH_SOURCES}) add_style_check_target(paddle_math ${MATH_HEADERS}) -add_dependencies(paddle_math gen_proto_cpp) # depends +add_dependencies(paddle_math paddle_proto ${external_project_dependencies}) # depends if(WITH_TESTING) add_subdirectory(tests) endif() diff --git a/paddle/memory/.clang-format b/paddle/memory/.clang-format new file mode 100644 index 0000000000000000000000000000000000000000..29282dc87e2c499988c17d90d47d44cd5cf7f115 --- /dev/null +++ b/paddle/memory/.clang-format @@ -0,0 +1,5 @@ +--- +Language: Cpp +BasedOnStyle: Google +Standard: Cpp11 +... diff --git a/paddle/memory/CMakeLists.txt b/paddle/memory/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..3943c3cfad31d13a00645aba6fc153d3d13da987 --- /dev/null +++ b/paddle/memory/CMakeLists.txt @@ -0,0 +1 @@ +add_subdirectory(detail) diff --git a/paddle/memory/README.md b/paddle/memory/README.md index e5f7880e4cad346da5399815f5e76b7b9b99bdea..96a331a486f57d3e030408fee182199bad5b38c2 100644 --- a/paddle/memory/README.md +++ b/paddle/memory/README.md @@ -97,6 +97,7 @@ class BuddyAllocator { struct Block { size_t size; Block* left, right; + size_t index; // allocator id }; ... }; diff --git a/paddle/memory/detail/CMakeLists.txt b/paddle/memory/detail/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..72d3749ad789eca9a4b10944131171c0cf8dfe5a --- /dev/null +++ b/paddle/memory/detail/CMakeLists.txt @@ -0,0 +1,7 @@ +if(${WITH_GPU}) + nv_library(system_allocator SRCS system_allocator.cc DEPS gflags) + nv_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator gflags) +else(${WITH_GPU}) + cc_library(system_allocator SRCS system_allocator.cc DEPS gflags) + cc_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator gflags) +endif(${WITH_GPU}) diff --git a/paddle/memory/detail/buddy_allocator.cc b/paddle/memory/detail/buddy_allocator.cc new file mode 100644 index 0000000000000000000000000000000000000000..ebe680f5eea4948339fb8c5584a5b9f5d71c752e --- /dev/null +++ b/paddle/memory/detail/buddy_allocator.cc @@ -0,0 +1,35 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/memory/detail/buddy_allocator.h" + +namespace paddle { +namespace memory { +namespace detail { + +BuddyAllocator::BuddyAllocator(size_t pool_size, size_t max_pools, + SystemAllocator* system_allocator) + : pool_size_(pool_size), + max_pools_(max_pools), + system_allocator_(system_allocator) { + PADDLE_ASSERT(pool_size > 0); + PADDLE_ASSERT(max_pools > 0); + PADDLE_ASSERT(system_allocator != nullptr); +} + +} // namespace detail +} // namespace memory +} // namespace paddle diff --git a/paddle/memory/detail/buddy_allocator.h b/paddle/memory/detail/buddy_allocator.h new file mode 100644 index 0000000000000000000000000000000000000000..82e6aaedc719966b4074449ce1ef7193c73dc265 --- /dev/null +++ b/paddle/memory/detail/buddy_allocator.h @@ -0,0 +1,86 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include "paddle/memory/detail/system_allocator.h" + +#include +#include + +namespace paddle { +namespace memory { +namespace detail { + +class BuddyAllocator { + public: + BuddyAllocator(size_t pool_size, size_t max_pools, + SystemAllocator* system_allocator); + ~BuddyAllocator(); + + void* Alloc(size_t size); + void Free(void*); + size_t Used(); + + private: + struct Block { + size_t size_; + Block* left_; // left buddy + Block* right_; // right buddy + }; + + // Initially, there is only one pool. If a Alloc founds not enough + // memory from that pool, and there has not been max_num_pools_, + // create a new pool by calling system_allocator_.Alloc(pool_size_). + std::vector pools_; + + size_t pool_size_; // the size of each pool; + size_t max_num_pools_; // the size of all pools; + + SystemAllocator* system_allocator_; + + std::mutex mutex_; + + // Disable copy and assignment. + BuddyAllocator(const BuddyAllocator&) = delete; + BuddyAllocator& operator=(const BuddyAllocator&) = delete; +}; + +BuddyAllocator* GetCPUBuddyAllocator() { + static BuddyAllocator* a = nullptr; + if (a == nullptr) { + a = new BuddyAllocator(); + } + return a; +} + +#ifndef PADDLE_ONLY_CPU // The following code are for CUDA. + +BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { + static BuddyAllocator** as = NULL; + if (as == NULL) { + int gpu_num = platform::GetDeviceCount(); + as = new BuddyAllocator*[gpu_num]; + for (int gpu = 0; gpu < gpu_num; gpu++) { + as[gpu] = new BuddyAllocator(); + } + } + return as[gpu_id]; +} + +#endif // PADDLE_ONLY_CPU + +} // namespace detail +} // namespace memory +} // namespace paddle diff --git a/paddle/memory/detail/system_allocator.cc b/paddle/memory/detail/system_allocator.cc new file mode 100644 index 0000000000000000000000000000000000000000..50bec926f83dee8a4343d0b16aeb088f9d2a4871 --- /dev/null +++ b/paddle/memory/detail/system_allocator.cc @@ -0,0 +1,90 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/memory/detail/system_allocator.h" + +#include // for malloc and free +#include // for mlock and munlock + +#include "gflags/gflags.h" +#include "paddle/platform/assert.h" +#include "paddle/platform/cuda.h" + +// If use_pinned_memory is true, CPUAllocator calls mlock, which +// returns pinned and locked memory as staging areas for data exchange +// between host and device. Allocates too much would reduce the amount +// of memory available to the system for paging. So, by default, we +// should set false to use_pinned_memory. +DEFINE_bool(use_pinned_memory, false, + "If set, allocate cpu/gpu pinned memory."); + +namespace paddle { +namespace memory { +namespace detail { + +void* CPUAllocator::Alloc(size_t size) { + // According to http://www.cplusplus.com/reference/cstdlib/malloc/, + // malloc might not return nullptr if size is zero, but the returned + // pointer shall not be dereferenced -- so we make it nullptr. + if (size <= 0) return nullptr; + + void* p = malloc(size); + if (p != nullptr && FLAGS_use_pinned_memory) { + mlock(p, size); + } + return p; +} + +void CPUAllocator::Free(void* p, size_t size) { + if (p != nullptr && FLAGS_use_pinned_memory) { + munlock(p, size); + } + free(p); +} + +#ifndef PADDLE_ONLY_CPU + +void* GPUAllocator::Alloc(size_t size) { + // CUDA documentation doesn't explain if cudaMalloc returns nullptr + // if size is 0. We just make sure it does. + if (size <= 0) { + return nullptr; + } + + void* p = 0; + cudaError_t result = + FLAGS_use_pinned_memory ? cudaMallocHost(&p, size) : cudaMalloc(&p, size); + if (result != cudaSuccess) { + cudaGetLastError(); // clear error if there is any. + } + return result == cudaSuccess ? p : nullptr; +} + +void GPUAllocator::Free(void* p, size_t size) { + // Purposefully allow cudaErrorCudartUnloading, because + // that is returned if you ever call cudaFree after the + // driver has already shutdown. This happens only if the + // process is terminating, in which case we don't care if + // cudaFree succeeds. + cudaError_t err = FLAGS_use_pinned_memory ? cudaFreeHost(p) : cudaFree(p); + if (err != cudaErrorCudartUnloading) { + platform::throw_on_error(err, "cudaFree{Host} failed"); + } +} + +#endif // PADDLE_ONLY_CPU + +} // namespace detail +} // namespace memory +} // namespace paddle diff --git a/paddle/memory/detail/system_allocator.h b/paddle/memory/detail/system_allocator.h new file mode 100644 index 0000000000000000000000000000000000000000..184b383f7f78244fa6632a3bffb1a0a78b3aa664 --- /dev/null +++ b/paddle/memory/detail/system_allocator.h @@ -0,0 +1,53 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include // for size_t + +namespace paddle { +namespace memory { +namespace detail { + +// SystemAllocator is the parent class of CPUAllocator and +// GPUAllocator. A BuddyAllocator object uses a SystemAllocator* +// pointing to the underlying system allocator. An alternative to +// this class hierarchy is to pass a system allocator class to +// BuddyAllocator as a template parameter. This approach makes +// BuddyAllocator a class template, and it's very complicated +// algorithm would make the buddy_allocator.h messy. +class SystemAllocator { + public: + virtual ~SystemAllocator() {} + virtual void* Alloc(size_t size) = 0; + virtual void Free(void* p, size_t size) = 0; +}; + +class CPUAllocator : public SystemAllocator { + public: + virtual void* Alloc(size_t size); + virtual void Free(void* p, size_t size); +}; + +#ifndef PADDLE_ONLY_CPU +class GPUAllocator : public SystemAllocator { + public: + virtual void* Alloc(size_t size); + virtual void Free(void* p, size_t size); +}; +#endif // PADDLE_ONLY_CPU + +} // namespace detail +} // namespace memory +} // namespace paddle diff --git a/paddle/memory/detail/system_allocator_test.cc b/paddle/memory/detail/system_allocator_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..9bd5706a4e4d1546a8c879ebbac0f3349c9d59f6 --- /dev/null +++ b/paddle/memory/detail/system_allocator_test.cc @@ -0,0 +1,71 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/memory/detail/system_allocator.h" + +#include +#include + +#include "gflags/gflags.h" +#include "gtest/gtest.h" + +DECLARE_bool(use_pinned_memory); + +void TestAllocator(paddle::memory::detail::SystemAllocator& a, size_t size) { + bool freed = false; + { + void* p = a.Alloc(size); + if (size > 0) { + EXPECT_NE(p, nullptr); + } else { + EXPECT_EQ(p, nullptr); + } + + int* i = static_cast(p); + std::shared_ptr ptr(i, [&](void* p) { + freed = true; + a.Free(p, size); + }); + } + EXPECT_TRUE(freed); +} + +TEST(CPUAllocator, NoLockMem) { + FLAGS_use_pinned_memory = false; + paddle::memory::detail::CPUAllocator a; + TestAllocator(a, 2048); + TestAllocator(a, 0); +} + +TEST(CPUAllocator, LockMem) { + FLAGS_use_pinned_memory = true; + paddle::memory::detail::CPUAllocator a; + TestAllocator(a, 2048); + TestAllocator(a, 0); +} + +#ifndef PADDLE_ONLY_CPU +TEST(GPUAllocator, NoStaging) { + FLAGS_use_pinned_memory = false; + paddle::memory::detail::GPUAllocator a; + TestAllocator(a, 2048); + TestAllocator(a, 0); +} +TEST(GPUAllocator, Staging) { + FLAGS_use_pinned_memory = true; + paddle::memory::detail::GPUAllocator a; + TestAllocator(a, 2048); + TestAllocator(a, 0); +} +#endif // PADDLE_ONLY_CPU diff --git a/paddle/memory/memory.cc b/paddle/memory/memory.cc new file mode 100644 index 0000000000000000000000000000000000000000..0d123d99e234a378ee64850eebacece223e2b121 --- /dev/null +++ b/paddle/memory/memory.cc @@ -0,0 +1,59 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/memory/memory.h" +#include "paddle/memory/detail/buddy_allocator.h" +#include "paddle/memory/detail/system_allocator.h" +#include "paddle/platform/assert.h" + +#include + +namespace paddle { +namespace memory { + +void* Alloc(platform::Place pl, size_t size) { +#ifndef PADDLE_ONLY_CPU + if (paddle::platform::is_gpu_place(pl)) { + size_t gpu_id = boost::get(pl).device; + return detail::GetGPUBuddyAllocator(gpu_id)->Alloc(size); + } +#endif // PADDLE_ONLY_CPU + PADDLE_ASSERT(paddle::platform::is_cpu_place(pl)); + return detail::GetCPUBuddyAllocator()->Alloc(size); +} + +void Free(paddle::platform::Place pl, void* p) { +#ifndef PADDLE_ONLY_CPU + if (paddle::platform::is_gpu_place(pl)) { + size_t gpu_id = boost::get(pl).device; + detail::GetGPUBuddyAllocator(gpu_id)->Free(p); + } +#endif // PADDLE_ONLY_CPU + PADDLE_ASSERT(paddle::platform::is_cpu_place(pl)); + detail::GetCPUBuddyAllocator()->Free(p); +} + +size_t Used(paddle::platform::Place pl) { +#ifndef PADDLE_ONLY_CPU + if (paddle::platform::is_gpu_place(pl)) { + size_t gpu_id = boost::get(pl).device; + return detail::GetGPUBuddyAllocator(gpu_id)->Used(); + } +#endif // PADDLE_ONLY_CPU + PADDLE_ASSERT(paddle::platform::is_cpu_place(pl)); + return detail::GetCPUBuddyAllocator()->Used(); +} + +} // namespace memory +} // namespace paddle diff --git a/paddle/platform/must_check.h b/paddle/memory/memory.h similarity index 53% rename from paddle/platform/must_check.h rename to paddle/memory/memory.h index 4fcc62afc05b14949fc43266f0d05be1f1b7891a..a33092bade65e6df0faee226a8967c9fc9caa032 100644 --- a/paddle/platform/must_check.h +++ b/paddle/memory/memory.h @@ -1,8 +1,11 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. @@ -10,17 +13,15 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -/** - * __must_check macro. It make the function's return value must be used, - * otherwise it will raise a compile warning. And also Paddle treat all compile - * warnings as errors. - */ -#ifdef __GNUC__ -#if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) >= 30400 -#define __must_check __attribute__((warn_unused_result)) -#else -#define __must_check -#endif -#else -#define __must_check -#endif + +#include "paddle/platform/place.h" + +namespace paddle { +namespace memory { + +void* Alloc(paddle::platform::Place, size_t); +void Free(paddle::platform::Place, void*); +size_t Used(paddle::platform::Place); + +} // namespace memory +} // namespace paddle diff --git a/paddle/optimizer/CMakeLists.txt b/paddle/optimizer/CMakeLists.txt index 4536f62ec7c2c3423d91e309dee993d4212160fe..9996d01d18b1185e9b01f8b1e4aab325eb28c894 100644 --- a/paddle/optimizer/CMakeLists.txt +++ b/paddle/optimizer/CMakeLists.txt @@ -10,7 +10,7 @@ set(OPITMIZER_SRCS ) add_library(paddle_optimizer STATIC ${OPITMIZER_SRCS}) -add_dependencies(paddle_optimizer gen_proto_cpp) +add_dependencies(paddle_optimizer paddle_proto ${external_project_dependencies}) if(WITH_TESTING) add_simple_unittest(serialization_test) diff --git a/paddle/parameter/CMakeLists.txt b/paddle/parameter/CMakeLists.txt index a35e46997fb04e9378e106bf428a629b286c2e8c..d2ae1c16c6b7316f1a6facdef4b933693d6ba818 100644 --- a/paddle/parameter/CMakeLists.txt +++ b/paddle/parameter/CMakeLists.txt @@ -7,7 +7,7 @@ add_library(paddle_parameter STATIC ${PARAMETERS_SOURCES}) add_style_check_target(paddle_parameter ${PARAMETERS_SOURCES}) add_style_check_target(paddle_parameter ${PARAMETERS_HEADERS}) -add_dependencies(paddle_parameter gen_proto_cpp) +add_dependencies(paddle_parameter paddle_proto ${external_project_dependencies}) if(WITH_TESTING) add_subdirectory(tests) endif() diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt index 7abe2ab89e0798672149e28a8d02f7a58b6de3ea..c7d7b14518ebb8415014a78fc1a3bafa8c386191 100644 --- a/paddle/platform/CMakeLists.txt +++ b/paddle/platform/CMakeLists.txt @@ -2,4 +2,3 @@ nv_test(cuda_test SRCS cuda_test.cu) cc_library(place SRCS place.cc) cc_test(place_test SRCS place_test.cc DEPS place glog gflags) -cc_test(must_check_test SRCS must_check_test.cc) diff --git a/paddle/platform/cuda.h b/paddle/platform/cuda.h new file mode 100644 index 0000000000000000000000000000000000000000..8fe891f9ce6c3add1df48a8b1f79fd811c7a4362 --- /dev/null +++ b/paddle/platform/cuda.h @@ -0,0 +1,40 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#ifndef PADDLE_ONLY_CPU + +#include +#include + +namespace paddle { +namespace platform { + +inline void throw_on_error(cudaError_t e, const char* message) { + if (e) { + throw thrust::system_error(e, thrust::cuda_category(), message); + } +} + +int GetDeviceCount(void) { + int count; + throw_on_error(cudaGetDeviceCount(&count), "cudaGetDeviceCount failed"); + return count; +} + +} // namespace platform +} // namespace paddle + +#endif // PADDLE_ONLY_CPU diff --git a/paddle/platform/must_check_test.cc b/paddle/platform/must_check_test.cc deleted file mode 100644 index 6ee3ea49acdc4384b5d5df353bfa1290856e982c..0000000000000000000000000000000000000000 --- a/paddle/platform/must_check_test.cc +++ /dev/null @@ -1,10 +0,0 @@ -#include -#include - -int __must_check SomeFunctionMustCheck() { return 0; } - -TEST(MustCheck, all) { - // This line should not be compiled, because the - // return value of SomeFunctionMustCheck marked as __must_check - // SomeFunctionMustCheck(); -} \ No newline at end of file diff --git a/paddle/platform/place.cc b/paddle/platform/place.cc index 1afd03c01169d395b086c1da458ce25c66a12a51..0704820aa05079401eb56814d689d6e280311edb 100644 --- a/paddle/platform/place.cc +++ b/paddle/platform/place.cc @@ -8,8 +8,8 @@ namespace detail { class PlacePrinter : public boost::static_visitor<> { public: PlacePrinter(std::ostream &os) : os_(os) {} - void operator()(const CpuPlace &) { os_ << "CpuPlace"; } - void operator()(const GpuPlace &p) { os_ << "GpuPlace(" << p.device << ")"; } + void operator()(const CPUPlace &) { os_ << "CPUPlace"; } + void operator()(const GPUPlace &p) { os_ << "GPUPlace(" << p.device << ")"; } private: std::ostream &os_; @@ -22,14 +22,14 @@ static Place the_default_place; void set_place(const Place &place) { the_default_place = place; } const Place &get_place() { return the_default_place; } -const GpuPlace default_gpu() { return GpuPlace(0); } -const CpuPlace default_cpu() { return CpuPlace(); } +const GPUPlace default_gpu() { return GPUPlace(0); } +const CPUPlace default_cpu() { return CPUPlace(); } bool is_gpu_place(const Place &p) { - return boost::apply_visitor(IsGpuPlace(), p); + return boost::apply_visitor(IsGPUPlace(), p); } bool is_cpu_place(const Place &p) { - return !boost::apply_visitor(IsGpuPlace(), p); + return !boost::apply_visitor(IsGPUPlace(), p); } bool places_are_same_class(const Place &p1, const Place &p2) { diff --git a/paddle/platform/place.h b/paddle/platform/place.h index 489572c526e162500c8f747f0ec8df10da9d86a2..7cead183884bc9379355cd931921b40d6c11ce90 100644 --- a/paddle/platform/place.h +++ b/paddle/platform/place.h @@ -1,43 +1,58 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + #pragma once + #include #include namespace paddle { namespace platform { -struct CpuPlace { +struct CPUPlace { // WORKAROUND: for some reason, omitting this constructor // causes errors with boost 1.59 and OSX - CpuPlace() {} + CPUPlace() {} // needed for variant equality comparison - inline bool operator==(const CpuPlace &) const { return true; } - inline bool operator!=(const CpuPlace &) const { return false; } + inline bool operator==(const CPUPlace &) const { return true; } + inline bool operator!=(const CPUPlace &) const { return false; } }; -struct GpuPlace { - GpuPlace() : GpuPlace(0) {} - GpuPlace(int d) : device(d) {} +struct GPUPlace { + GPUPlace() : GPUPlace(0) {} + GPUPlace(int d) : device(d) {} // needed for variant equality comparison - inline bool operator==(const GpuPlace &o) const { return device == o.device; } - inline bool operator!=(const GpuPlace &o) const { return !(*this == o); } + inline bool operator==(const GPUPlace &o) const { return device == o.device; } + inline bool operator!=(const GPUPlace &o) const { return !(*this == o); } int device; }; -struct IsGpuPlace : public boost::static_visitor { - bool operator()(const CpuPlace &) const { return false; } - bool operator()(const GpuPlace &gpu) const { return true; } +struct IsGPUPlace : public boost::static_visitor { + bool operator()(const CPUPlace &) const { return false; } + bool operator()(const GPUPlace &gpu) const { return true; } }; -typedef boost::variant Place; +typedef boost::variant Place; void set_place(const Place &); const Place &get_place(); -const GpuPlace default_gpu(); -const CpuPlace default_cpu(); +const GPUPlace default_gpu(); +const CPUPlace default_cpu(); bool is_gpu_place(const Place &); bool is_cpu_place(const Place &); diff --git a/paddle/platform/place_test.cc b/paddle/platform/place_test.cc index 73fccceedf6918148a26100f64cf322305c3ac20..33e2e5a439ce6801c02daba4bcbd462a74d7a614 100644 --- a/paddle/platform/place_test.cc +++ b/paddle/platform/place_test.cc @@ -3,8 +3,8 @@ #include "gtest/gtest.h" TEST(Place, Equality) { - paddle::platform::CpuPlace cpu; - paddle::platform::GpuPlace g0(0), g1(1), gg0(0); + paddle::platform::CPUPlace cpu; + paddle::platform::GPUPlace g0(0), g1(1), gg0(0); EXPECT_EQ(cpu, cpu); EXPECT_EQ(g0, g0); @@ -22,19 +22,19 @@ TEST(Place, Default) { EXPECT_TRUE(paddle::platform::is_gpu_place(paddle::platform::default_gpu())); EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::default_cpu())); - paddle::platform::set_place(paddle::platform::CpuPlace()); + paddle::platform::set_place(paddle::platform::CPUPlace()); EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::get_place())); } TEST(Place, Print) { { std::stringstream ss; - ss << paddle::platform::GpuPlace(1); - EXPECT_EQ("GpuPlace(1)", ss.str()); + ss << paddle::platform::GPUPlace(1); + EXPECT_EQ("GPUPlace(1)", ss.str()); } { std::stringstream ss; - ss << paddle::platform::CpuPlace(); - EXPECT_EQ("CpuPlace", ss.str()); + ss << paddle::platform::CPUPlace(); + EXPECT_EQ("CPUPlace", ss.str()); } } diff --git a/paddle/pserver/CMakeLists.txt b/paddle/pserver/CMakeLists.txt index b7f85ea1a6dfda2a37c315ba15c6ca1979cf4131..2245c7d88ca74922f9919db91977dfa6cb3ca468 100644 --- a/paddle/pserver/CMakeLists.txt +++ b/paddle/pserver/CMakeLists.txt @@ -17,7 +17,7 @@ add_library(paddle_network STATIC add_style_check_target(paddle_network ${NETWORK_SOURCES}) add_style_check_target(paddle_network ${NETWORK_HEADERS}) -add_dependencies(paddle_network gen_proto_cpp) +add_dependencies(paddle_network paddle_proto ${external_project_dependencies}) ################### paddle_pserver ###################### set(PSERVER_SOURCES @@ -40,7 +40,7 @@ add_library(paddle_pserver STATIC add_style_check_target(paddle_pserver ${PSERVER_SOURCES}) add_style_check_target(paddle_pserver ${PSERVER_HEADERS}) -add_dependencies(paddle_pserver gen_proto_cpp) +add_dependencies(paddle_pserver paddle_proto ${external_project_dependencies}) set(PSERVER_MAIN_SOURCES ParameterServer2Main.cpp) diff --git a/paddle/py_paddle/dataprovider_converter.py b/paddle/py_paddle/dataprovider_converter.py index 218cb5ec560ed0717d96a50e9560492ee55b9f70..43614b9779d21795f1f274589ea93639e923ce75 100644 --- a/paddle/py_paddle/dataprovider_converter.py +++ b/paddle/py_paddle/dataprovider_converter.py @@ -144,7 +144,7 @@ class DenseScanner(IScanner): if len(self.__shape__) > 1: # The last-two dimenstions are the frame height and width. # For example, the layout is CHW for 3-D feature of image. - # The H and W are the fram height and width. + # The H and W are the frame height and width. h, w = self.__shape__[-2:] argument.setSlotFrameHeight(self.pos, h) argument.setSlotFrameWidth(self.pos, w) diff --git a/paddle/string/CMakeLists.txt b/paddle/string/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..5becf62672d0c606c98ea1a1a4383df97088ab05 --- /dev/null +++ b/paddle/string/CMakeLists.txt @@ -0,0 +1,4 @@ +cc_library(stringpiece SRCS piece.cc) +cc_test(stringpiece_test SRCS piece_test.cc DEPS stringpiece glog gflags) + +cc_test(stringprintf_test SRCS printf_test.cc DEPS glog gflags) diff --git a/paddle/string/piece.cc b/paddle/string/piece.cc new file mode 100644 index 0000000000000000000000000000000000000000..b80afdec82d642fd3a8245b96ce1bb2bea17cbae --- /dev/null +++ b/paddle/string/piece.cc @@ -0,0 +1,138 @@ +/* + Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. +*/ + +#include "paddle/string/piece.h" + +#include + +#include +#include +#include + +namespace paddle { +namespace string { + +Piece::Piece() : data_(NULL), size_(0) {} + +Piece::Piece(const char* d, size_t n) : data_(d), size_(n) { + if (d == NULL && n != 0) + throw std::invalid_argument("Piece requires len to be 0 for NULL data"); +} + +Piece::Piece(const char* s) : data_(s) { size_ = (s == NULL) ? 0 : strlen(s); } + +Piece::Piece(const std::string& s) : data_(s.data()), size_(s.size()) {} + +char Piece::operator[](size_t n) const { + if (n >= len()) throw std::invalid_argument("index out of Piece length"); + return data_[n]; +} + +int Compare(Piece a, Piece b) { + const size_t min_len = (a.len() < b.len()) ? a.len() : b.len(); + int r = memcmp(a.data(), b.data(), min_len); + if (r == 0) { + if (a.len() < b.len()) + return -1; + else if (a.len() > b.len()) + return 1; + } + return r; +} + +bool operator==(Piece x, Piece y) { + return ((x.len() == y.len()) && + (x.data() == y.data() || memcmp(x.data(), y.data(), x.len()) == 0)); +} + +bool operator!=(Piece x, Piece y) { return !(x == y); } + +bool operator<(Piece x, Piece y) { return Compare(x, y) < 0; } +bool operator>(Piece x, Piece y) { return Compare(x, y) > 0; } + +bool operator<=(Piece x, Piece y) { return Compare(x, y) <= 0; } +bool operator>=(Piece x, Piece y) { return Compare(x, y) >= 0; } + +bool HasPrefix(Piece s, Piece x) { + return ((s.len() >= x.len()) && (memcmp(s.data(), x.data(), x.len()) == 0)); +} + +bool HasSuffix(Piece s, Piece x) { + return ((s.len() >= x.len()) && + (memcmp(s.data() + (s.len() - x.len()), x.data(), x.len()) == 0)); +} + +Piece SkipPrefix(Piece s, size_t n) { + if (n > s.len()) + throw std::invalid_argument("Skip distance larger than Piece length"); + return Piece(s.data() + n, s.len() - n); +} + +Piece SkipSuffix(Piece s, size_t n) { + if (n > s.len()) + throw std::invalid_argument("Skip distance larger than Piece length"); + return Piece(s.data(), s.len() - n); +} + +Piece TrimPrefix(Piece s, Piece x) { + return HasPrefix(s, x) ? SkipPrefix(s, x.len()) : s; +} + +Piece TrimSuffix(Piece s, Piece x) { + return HasSuffix(s, x) ? SkipSuffix(s, x.len()) : s; +} + +bool Contains(Piece s, Piece sub) { + return std::search(s.begin(), s.end(), sub.begin(), sub.end()) != s.end(); +} + +size_t Index(Piece s, Piece sub) { + auto e = std::search(s.begin(), s.end(), sub.begin(), sub.end()); + return e != s.end() ? e - s.data() : Piece::npos; +} + +size_t Find(Piece s, char c, size_t pos) { + if (pos >= s.len()) { + return Piece::npos; + } + const char* result = + reinterpret_cast(memchr(s.data() + pos, c, s.len() - pos)); + return result != nullptr ? result - s.data() : Piece::npos; +} + +size_t RFind(Piece s, char c, size_t pos) { + if (s.len() == 0) return Piece::npos; + for (const char* p = s.data() + std::min(pos, s.len() - 1); p >= s.data(); + p--) { + if (*p == c) { + return p - s.data(); + } + } + return Piece::npos; +} + +Piece SubStr(Piece s, size_t pos, size_t n) { + if (pos > s.len()) pos = s.len(); + if (n > s.len() - pos) n = s.len() - pos; + return Piece(s.data() + pos, n); +} + +std::ostream& operator<<(std::ostream& o, Piece piece) { + return o << piece.ToString(); +} + +} // namespace string +} // namespace paddle diff --git a/paddle/strings/stringpiece.h b/paddle/string/piece.h similarity index 57% rename from paddle/strings/stringpiece.h rename to paddle/string/piece.h index adff713e86f49349b8f189c1d24584bfc1bb8aa7..db7c3e69804a6a8f0510ba376432fe560ae74442 100644 --- a/paddle/strings/stringpiece.h +++ b/paddle/string/piece.h @@ -20,33 +20,34 @@ #include namespace paddle { +namespace string { -// StringPiece points into a std::string object but doesn't own the +// Piece points into a std::string object but doesn't own the // string. It is for efficient access to strings. Like Go's string -// type. Not that StringPiece doesn't mutate the underlying string, +// type. Not that Piece doesn't mutate the underlying string, // so it is thread-safe given that the underlying string doesn't -// change. Because StringPiece contains a little data members, and +// change. Because Piece contains a little data members, and // its syntax is simple as it doesn't own/manage the string, it is -// cheap to construct StringPieces and pass them around. -class StringPiece { +// cheap to construct Pieces and pass them around. +class Piece { public: static const size_t npos = static_cast(-1); // We provide non-explicit singleton constructors so users can - // pass in a "const char*" or a "string" wherever a "StringPiece" + // pass in a "const char*" or a "string" wherever a "Piece" // is expected. These contructors ensure that if data_ is NULL, // size_ is 0. - StringPiece(); - StringPiece(const char* d, size_t n); - StringPiece(const char* d); - StringPiece(const std::string& s); + Piece(); + Piece(const char* d, size_t n); + Piece(const char* d); + Piece(const std::string& s); const char* data() const { return data_; } size_t len() const { return size_; } char operator[](size_t n) const; - // StringPiece doesn't own the string, so both iterator and const + // Piece doesn't own the string, so both iterator and const // iterator are const char* indeed. typedef const char* const_iterator; typedef const char* iterator; @@ -63,43 +64,44 @@ private: // Intentionally copyable }; -int Compare(StringPiece a, StringPiece b); +int Compare(Piece a, Piece b); -bool operator==(StringPiece x, StringPiece y); -bool operator!=(StringPiece x, StringPiece y); -bool operator<(StringPiece x, StringPiece y); -bool operator>(StringPiece x, StringPiece y); -bool operator<=(StringPiece x, StringPiece y); -bool operator>=(StringPiece x, StringPiece y); +bool operator==(Piece x, Piece y); +bool operator!=(Piece x, Piece y); +bool operator<(Piece x, Piece y); +bool operator>(Piece x, Piece y); +bool operator<=(Piece x, Piece y); +bool operator>=(Piece x, Piece y); -bool HasPrefix(StringPiece s, StringPiece prefix); -bool HasSuffix(StringPiece s, StringPiece suffix); +bool HasPrefix(Piece s, Piece prefix); +bool HasSuffix(Piece s, Piece suffix); -StringPiece SkipPrefix(StringPiece s, size_t n); -StringPiece SkipSuffix(StringPiece s, size_t n); +Piece SkipPrefix(Piece s, size_t n); +Piece SkipSuffix(Piece s, size_t n); // Skip the prefix (or suffix) if it matches with the string. -StringPiece TrimPrefix(StringPiece s, StringPiece prefix); -StringPiece TrimSuffix(StringPiece s, StringPiece suffix); +Piece TrimPrefix(Piece s, Piece prefix); +Piece TrimSuffix(Piece s, Piece suffix); // Returns if s contains sub. Any s except for empty s contains an // empty sub. -bool Contains(StringPiece s, StringPiece sub); +bool Contains(Piece s, Piece sub); // Return the first occurrence of sub in s, or npos. If both s and // sub is empty, it returns npos; otherwise, if only sub is empty, it // returns 0. -size_t Index(StringPiece s, StringPiece sub); +size_t Index(Piece s, Piece sub); // Return the first occurrence of c in s[pos:end], or npos. -size_t Find(StringPiece s, char c, size_t pos); +size_t Find(Piece s, char c, size_t pos); // Search range is [0..pos] inclusive. If pos == npos, search everything. -size_t RFind(StringPiece s, char c, size_t pos); +size_t RFind(Piece s, char c, size_t pos); -StringPiece SubStr(StringPiece s, size_t pos, size_t n); +Piece SubStr(Piece s, size_t pos, size_t n); -// allow StringPiece to be logged -std::ostream& operator<<(std::ostream& o, StringPiece piece); +// allow Piece to be logged +std::ostream& operator<<(std::ostream& o, Piece piece); +} // namespace string } // namespace paddle diff --git a/paddle/strings/stringpiece_test.cc b/paddle/string/piece_test.cc similarity index 77% rename from paddle/strings/stringpiece_test.cc rename to paddle/string/piece_test.cc index 2ba66a04f641c3457efa713383484491a213668f..cf5152ff5a3cb0a2afae0c90b787abf291122fa3 100644 --- a/paddle/strings/stringpiece_test.cc +++ b/paddle/string/piece_test.cc @@ -14,7 +14,7 @@ limitations under the License. */ -#include "paddle/strings/stringpiece.h" +#include "paddle/string/piece.h" #include @@ -22,42 +22,44 @@ TEST(StringPiece, Construct) { { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_EQ(NULL, s.data()); EXPECT_EQ(0U, s.len()); } - { EXPECT_THROW(paddle::StringPiece s(NULL, 10000U), std::invalid_argument); } { - paddle::StringPiece s(NULL); + EXPECT_THROW(paddle::string::Piece s(NULL, 10000U), std::invalid_argument); + } + { + paddle::string::Piece s(NULL); EXPECT_EQ(0U, s.len()); } { std::string a; EXPECT_EQ(0U, a.size()); - paddle::StringPiece s(a); + paddle::string::Piece s(a); EXPECT_EQ(0U, s.len()); } } TEST(StringPiece, CopyAndAssign) { - paddle::StringPiece empty; + paddle::string::Piece empty; EXPECT_EQ(0U, empty.len()); - paddle::StringPiece a("hello"); - paddle::StringPiece b = a; + paddle::string::Piece a("hello"); + paddle::string::Piece b = a; EXPECT_EQ(b.len(), strlen("hello")); EXPECT_EQ(a, b); std::string storage("hello"); - paddle::StringPiece c(storage); + paddle::string::Piece c(storage); EXPECT_EQ(a, c); EXPECT_NE(a.data(), c.data()); } TEST(StringPiece, Compare) { { - paddle::StringPiece a("hello"); - paddle::StringPiece b("world"); + paddle::string::Piece a("hello"); + paddle::string::Piece b("world"); EXPECT_TRUE(a != b); EXPECT_FALSE(a == b); EXPECT_TRUE(a < b); @@ -68,7 +70,7 @@ TEST(StringPiece, Compare) { EXPECT_GT(Compare(b, a), 0); } { - paddle::StringPiece a, b; + paddle::string::Piece a, b; EXPECT_TRUE(a == b); EXPECT_FALSE(a != b); EXPECT_FALSE(a < b); @@ -82,31 +84,31 @@ TEST(StringPiece, Compare) { TEST(StringPiece, ToString) { { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_EQ(std::string(""), s.ToString()); } { - paddle::StringPiece s(NULL); + paddle::string::Piece s(NULL); EXPECT_EQ(std::string(""), s.ToString()); } { - paddle::StringPiece s("hello"); + paddle::string::Piece s("hello"); EXPECT_EQ(std::string("hello"), s.ToString()); } } TEST(StringPiece, HasPrefixSuffix) { - using paddle::HasPrefix; - using paddle::HasSuffix; + using paddle::string::HasPrefix; + using paddle::string::HasSuffix; { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_FALSE(HasPrefix(s, "something")); EXPECT_TRUE(HasPrefix(s, "")); EXPECT_FALSE(HasSuffix(s, "something")); EXPECT_TRUE(HasSuffix(s, "")); } { - paddle::StringPiece s("app"); + paddle::string::Piece s("app"); EXPECT_TRUE(HasPrefix(s, "")); EXPECT_TRUE(HasPrefix(s, "a")); EXPECT_TRUE(HasPrefix(s, "ap")); @@ -120,10 +122,10 @@ TEST(StringPiece, HasPrefixSuffix) { } TEST(StringPiece, SkipPrefixSuffix) { - using paddle::SkipPrefix; - using paddle::SkipSuffix; + using paddle::string::SkipPrefix; + using paddle::string::SkipSuffix; { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_EQ("", SkipPrefix(s, 0)); EXPECT_THROW(SkipPrefix(s, 1), std::invalid_argument); @@ -131,7 +133,7 @@ TEST(StringPiece, SkipPrefixSuffix) { EXPECT_THROW(SkipSuffix(s, 1), std::invalid_argument); } { - paddle::StringPiece s("app"); + paddle::string::Piece s("app"); EXPECT_EQ("app", SkipPrefix(s, 0)); EXPECT_EQ("pp", SkipPrefix(s, 1)); EXPECT_EQ("p", SkipPrefix(s, 2)); @@ -147,10 +149,10 @@ TEST(StringPiece, SkipPrefixSuffix) { } TEST(StringPiece, TrimPrefixSuffix) { - using paddle::TrimPrefix; - using paddle::TrimSuffix; + using paddle::string::TrimPrefix; + using paddle::string::TrimSuffix; { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_EQ("", TrimPrefix(s, "")); EXPECT_EQ("", TrimPrefix(s, "something")); @@ -158,7 +160,7 @@ TEST(StringPiece, TrimPrefixSuffix) { EXPECT_EQ("", TrimSuffix(s, "something")); } { - paddle::StringPiece s("app"); + paddle::string::Piece s("app"); EXPECT_EQ("app", TrimPrefix(s, "")); EXPECT_EQ("pp", TrimPrefix(s, "a")); EXPECT_EQ("p", TrimPrefix(s, "ap")); @@ -174,14 +176,14 @@ TEST(StringPiece, TrimPrefixSuffix) { } TEST(StringPiece, Contains) { - using paddle::Contains; + using paddle::string::Contains; { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_FALSE(Contains(s, "")); EXPECT_FALSE(Contains(s, "something")); } { - paddle::StringPiece s("app"); + paddle::string::Piece s("app"); EXPECT_TRUE(Contains(s, "")); EXPECT_TRUE(Contains(s, "a")); EXPECT_TRUE(Contains(s, "p")); @@ -193,15 +195,15 @@ TEST(StringPiece, Contains) { } TEST(StringPiece, Index) { - using paddle::Index; - auto npos = paddle::StringPiece::npos; + using paddle::string::Index; + auto npos = paddle::string::Piece::npos; { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_EQ(npos, Index(s, "")); EXPECT_EQ(npos, Index(s, "something")); } { - paddle::StringPiece s("app"); + paddle::string::Piece s("app"); EXPECT_EQ(0U, Index(s, "")); EXPECT_EQ(0U, Index(s, "a")); EXPECT_EQ(1U, Index(s, "p")); @@ -213,14 +215,14 @@ TEST(StringPiece, Index) { } TEST(StringPiece, Find) { - using paddle::Find; - auto npos = paddle::StringPiece::npos; + using paddle::string::Find; + auto npos = paddle::string::Piece::npos; { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_EQ(npos, Find(s, 'a', 0U)); } { - paddle::StringPiece s("app"); + paddle::string::Piece s("app"); EXPECT_EQ(0U, Find(s, 'a', 0U)); EXPECT_EQ(1U, Find(s, 'p', 0U)); EXPECT_EQ(1U, Find(s, 'p', 1U)); @@ -230,14 +232,14 @@ TEST(StringPiece, Find) { } TEST(StringPiece, RFind) { - using paddle::RFind; - auto npos = paddle::StringPiece::npos; + using paddle::string::RFind; + auto npos = paddle::string::Piece::npos; { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_EQ(npos, RFind(s, 'a', 0U)); } { - paddle::StringPiece s("app"); + paddle::string::Piece s("app"); EXPECT_EQ(2U, RFind(s, 'p', 2U)); EXPECT_EQ(0U, RFind(s, 'a', 2U)); EXPECT_EQ(1U, RFind(s, 'p', 1U)); @@ -247,15 +249,15 @@ TEST(StringPiece, RFind) { } TEST(StringPiece, SubStr) { - using paddle::SubStr; + using paddle::string::SubStr; { - paddle::StringPiece s; + paddle::string::Piece s; EXPECT_EQ("", SubStr(s, 0, 0)); EXPECT_EQ("", SubStr(s, 0, 1)); EXPECT_EQ("", SubStr(s, 1, 0)); } { - paddle::StringPiece s("app"); + paddle::string::Piece s("app"); EXPECT_EQ("", SubStr(s, 0, 0)); EXPECT_EQ("", SubStr(s, 1, 0)); EXPECT_EQ("", SubStr(s, 2, 0)); @@ -279,15 +281,15 @@ TEST(StringPiece, SubStr) { } TEST(StringPiece, StreamOutput) { - using paddle::StringPiece; + using paddle::string::Piece; std::stringstream o; - o << StringPiece(); + o << paddle::string::Piece(); EXPECT_EQ("", o.str()); - o << StringPiece("hello"); + o << paddle::string::Piece("hello"); EXPECT_EQ("hello", o.str()); - o << StringPiece(); + o << paddle::string::Piece(); EXPECT_EQ("hello", o.str()); } diff --git a/paddle/string/printf.h b/paddle/string/printf.h new file mode 100644 index 0000000000000000000000000000000000000000..8b5ce63a8e8dfe77962ff1e7415911d381a28aac --- /dev/null +++ b/paddle/string/printf.h @@ -0,0 +1,99 @@ +/* + Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. +*/ + +// Compared with std::stringstream, there are primary purpose of +// string::Printf: +// +// 1. Type-safe printing, with why and how explained in +// http://www.drdobbs.com/stringprintf-a-typesafe-printf-family-fo/184401999. +// Implementation includes +// +// https://github.com/c42f/tinyformat +// boost::format +// std::stringstream +// +// std::stringstream is not convenient enough in many cases. For example: +// +// std::cout << std::setprecision(2) << std::fixed << 1.23456 << "\n"; +// +// boost::format is the most convenient one. We can have +// +// std::cout << format("%2% %1%") % 36 % 77; +// +// or +// +// format fmter("%2% %1%"); +// fmter % 36; fmter % 77; +// std::cout << fmter.c_str(); +// +// But the overloading of % might be overkilling and it would be +// more efficient if it can write to std::cout directly. +// +// tinyformat has an interface compatible with the C-printf style, +// and it can writes to a stream or returns a std::string: +// +// std::cout << tfm::printf( +// "%s, %s %d, %.2d:%.2d\n", +// weekday, month, day, hour, min); +// +// or +// +// tfm::format(std::cout, +// "%s, %s %d, %.2d:%.2d\n", +// weekday, month, day, hour, min); +// +// 2. High-performance -- most printed strings are not too long and +// doens't need dynamic memory allocation. Many StringPrintf +// implementations doesn't enforce type-safe, but are +// high-performance, including +// +// https://developers.google.com/optimization/reference/base/stringprintf/ +// https://github.com/adobe/chromium/blob/master/base/stringprintf.h +// https://github.com/google/protobuf/blob/master/src/google/protobuf/stubs/stringprintf.h +// +// According to +// https://github.com/c42f/tinyformat#compile-time-and-code-bloat, +// boost::format runs too slow and results in large executable binary +// files. So here we port tinyformat. + +#pragma once + +#include +#include +#include "paddle/string/tinyformat/tinyformat.h" // https://github.com/c42f/tinyformat + +namespace paddle { +namespace string { + +template +void Fprintf(std::ostream& out, const char* fmt, const Args&... args) { + tinyformat::vformat(out, fmt, tinyformat::makeFormatList(args...)); +} + +template +std::string Sprintf(const char* fmt, const Args&... args) { + std::ostringstream oss; + Fprintf(oss, fmt, args...); + return oss.str(); +} + +template +void Printf(const char* fmt, const Args&... args) { + Fprintf(std::cout, fmt, args...); +} + +} // namespace string +} // namespace paddle diff --git a/paddle/string/printf_test.cc b/paddle/string/printf_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..d8f2454165d741b3937f908dcfd87501940750d5 --- /dev/null +++ b/paddle/string/printf_test.cc @@ -0,0 +1,16 @@ +#include "paddle/string/printf.h" + +#include + +#include "gtest/gtest.h" + +TEST(StringPrintf, StringPrintf) { + std::string weekday = "Wednesday"; + const char* month = "July"; + size_t day = 27; + long hour = 14; + int min = 44; + EXPECT_EQ(std::string("Wednesday, July 27, 14:44"), + paddle::string::Sprintf( + "%s, %s %d, %.2d:%.2d", weekday, month, day, hour, min)); +} diff --git a/paddle/string/tinyformat/tinyformat.h b/paddle/string/tinyformat/tinyformat.h new file mode 100644 index 0000000000000000000000000000000000000000..f0e5e0160fb018b813c1dade727da2861a295147 --- /dev/null +++ b/paddle/string/tinyformat/tinyformat.h @@ -0,0 +1,902 @@ +// tinyformat.h +// Copyright (C) 2011, Chris Foster [chris42f (at) gmail (d0t) com] +// +// Boost Software License - Version 1.0 +// +// Permission is hereby granted, free of charge, to any person or organization +// obtaining a copy of the software and accompanying documentation covered by +// this license (the "Software") to use, reproduce, display, distribute, +// execute, and transmit the Software, and to prepare derivative works of the +// Software, and to permit third-parties to whom the Software is furnished to +// do so, all subject to the following: +// +// The copyright notices in the Software and this entire statement, including +// the above license grant, this restriction and the following disclaimer, +// must be included in all copies of the Software, in whole or in part, and +// all derivative works of the Software, unless such copies or derivative +// works are solely in the form of machine-executable object code generated by +// a source language processor. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT +// SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE +// FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE, +// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +// DEALINGS IN THE SOFTWARE. + +//------------------------------------------------------------------------------ +// Tinyformat: A minimal type safe printf replacement +// +// tinyformat.h is a type safe printf replacement library in a single C++ +// header file. Design goals include: +// +// * Type safety and extensibility for user defined types. +// * C99 printf() compatibility, to the extent possible using std::ostream +// * Simplicity and minimalism. A single header file to include and distribute +// with your projects. +// * Augment rather than replace the standard stream formatting mechanism +// * C++98 support, with optional C++11 niceties +// +// +// Main interface example usage +// ---------------------------- +// +// To print a date to std::cout: +// +// std::string weekday = "Wednesday"; +// const char* month = "July"; +// size_t day = 27; +// long hour = 14; +// int min = 44; +// +// tfm::printf("%s, %s %d, %.2d:%.2d\n", weekday, month, day, hour, min); +// +// The strange types here emphasize the type safety of the interface; it is +// possible to print a std::string using the "%s" conversion, and a +// size_t using the "%d" conversion. A similar result could be achieved +// using either of the tfm::format() functions. One prints on a user provided +// stream: +// +// tfm::format(std::cerr, "%s, %s %d, %.2d:%.2d\n", +// weekday, month, day, hour, min); +// +// The other returns a std::string: +// +// std::string date = tfm::format("%s, %s %d, %.2d:%.2d\n", +// weekday, month, day, hour, min); +// std::cout << date; +// +// These are the three primary interface functions. There is also a +// convenience function printfln() which appends a newline to the usual result +// of printf() for super simple logging. +// +// +// User defined format functions +// ----------------------------- +// +// Simulating variadic templates in C++98 is pretty painful since it requires +// writing out the same function for each desired number of arguments. To make +// this bearable tinyformat comes with a set of macros which are used +// internally to generate the API, but which may also be used in user code. +// +// The three macros TINYFORMAT_ARGTYPES(n), TINYFORMAT_VARARGS(n) and +// TINYFORMAT_PASSARGS(n) will generate a list of n argument types, +// type/name pairs and argument names respectively when called with an integer +// n between 1 and 16. We can use these to define a macro which generates the +// desired user defined function with n arguments. To generate all 16 user +// defined function bodies, use the macro TINYFORMAT_FOREACH_ARGNUM. For an +// example, see the implementation of printf() at the end of the source file. +// +// Sometimes it's useful to be able to pass a list of format arguments through +// to a non-template function. The FormatList class is provided as a way to do +// this by storing the argument list in a type-opaque way. Continuing the +// example from above, we construct a FormatList using makeFormatList(): +// +// FormatListRef formatList = tfm::makeFormatList(weekday, month, day, hour, +// min); +// +// The format list can now be passed into any non-template function and used +// via a call to the vformat() function: +// +// tfm::vformat(std::cout, "%s, %s %d, %.2d:%.2d\n", formatList); +// +// +// Additional API information +// -------------------------- +// +// Error handling: Define TINYFORMAT_ERROR to customize the error handling for +// format strings which are unsupported or have the wrong number of format +// specifiers (calls assert() by default). +// +// User defined types: Uses operator<< for user defined types by default. +// Overload formatValue() for more control. + +#pragma once + +#include +#include +#include +#include + +namespace paddle { +namespace string { +namespace tinyformat { + +#ifndef TINYFORMAT_ERROR +#define TINYFORMAT_ERROR(reason) assert(0 && reason) +#endif + +//------------------------------------------------------------------------------ +namespace detail { + +// Test whether type T1 is convertible to type T2 +template +struct is_convertible { +private: + // two types of different size + struct fail { + char dummy[2]; + }; + struct succeed { + char dummy; + }; + // Try to convert a T1 to a T2 by plugging into tryConvert + static fail tryConvert(...); + static succeed tryConvert(const T2 &); + static const T1 &makeT1(); + +public: + // Standard trick: the (...) version of tryConvert will be chosen from + // the overload set only if the version taking a T2 doesn't match. + // Then we compare the sizes of the return types to check which + // function matched. Very neat, in a disgusting kind of way :) + static const bool value = sizeof(tryConvert(makeT1())) == sizeof(succeed); +}; + +// Format the value by casting to type fmtT. This default implementation +// should never be called. +template ::value> +struct formatValueAsType { + static void invoke(std::ostream & /*out*/, const T & /*value*/) { assert(0); } +}; +// Specialized version for types that can actually be converted to fmtT, as +// indicated by the "convertible" template parameter. +template +struct formatValueAsType { + static void invoke(std::ostream &out, const T &value) { + out << static_cast(value); + } +}; + +// Convert an arbitrary type to integer. The version with convertible=false +// throws an error. +template ::value> +struct convertToInt { + static int invoke(const T & /*value*/) { + TINYFORMAT_ERROR( + "tinyformat: Cannot convert from argument type to " + "integer for use as variable width or precision"); + return 0; + } +}; +// Specialization for convertToInt when conversion is possible +template +struct convertToInt { + static int invoke(const T &value) { return static_cast(value); } +}; + +// Format at most ntrunc characters to the given stream. +template +inline void formatTruncated(std::ostream &out, const T &value, int ntrunc) { + std::ostringstream tmp; + tmp << value; + std::string result = tmp.str(); + out.write(result.c_str(), + (std::min)(ntrunc, static_cast(result.size()))); +} +#define TINYFORMAT_DEFINE_FORMAT_TRUNCATED_CSTR(type) \ + inline void formatTruncated(std::ostream &out, type *value, int ntrunc) { \ + std::streamsize len = 0; \ + while (len < ntrunc && value[len] != 0) ++len; \ + out.write(value, len); \ + } +// Overload for const char* and char*. Could overload for signed & unsigned +// char too, but these are technically unneeded for printf compatibility. +TINYFORMAT_DEFINE_FORMAT_TRUNCATED_CSTR(const char) +TINYFORMAT_DEFINE_FORMAT_TRUNCATED_CSTR(char) +#undef TINYFORMAT_DEFINE_FORMAT_TRUNCATED_CSTR + +} // namespace detail + +//------------------------------------------------------------------------------ +// Variable formatting functions. May be overridden for user-defined types if +// desired. + +/// Format a value into a stream, delegating to operator<< by default. +/// +/// Users may override this for their own types. When this function is called, +/// the stream flags will have been modified according to the format string. +/// The format specification is provided in the range [fmtBegin, fmtEnd). For +/// truncating conversions, ntrunc is set to the desired maximum number of +/// characters, for example "%.7s" calls formatValue with ntrunc = 7. +/// +/// By default, formatValue() uses the usual stream insertion operator +/// operator<< to format the type T, with special cases for the %c and %p +/// conversions. +template +inline void formatValue(std::ostream &out, + const char * /*fmtBegin*/, + const char *fmtEnd, + int ntrunc, + const T &value) { + // The mess here is to support the %c and %p conversions: if these + // conversions are active we try to convert the type to a char or const + // void* respectively and format that instead of the value itself. For the + // %p conversion it's important to avoid dereferencing the pointer, which + // could otherwise lead to a crash when printing a dangling (const char*). + const bool canConvertToChar = detail::is_convertible::value; + const bool canConvertToVoidPtr = + detail::is_convertible::value; + if (canConvertToChar && *(fmtEnd - 1) == 'c') + detail::formatValueAsType::invoke(out, value); + else if (canConvertToVoidPtr && *(fmtEnd - 1) == 'p') + detail::formatValueAsType::invoke(out, value); + else if (ntrunc >= 0) { + // Take care not to overread C strings in truncating conversions like + // "%.4s" where at most 4 characters may be read. + detail::formatTruncated(out, value, ntrunc); + } else + out << value; +} + +// Overloaded version for char types to support printing as an integer +#define TINYFORMAT_DEFINE_FORMATVALUE_CHAR(charType) \ + inline void formatValue(std::ostream &out, \ + const char * /*fmtBegin*/, \ + const char *fmtEnd, \ + int /**/, \ + charType value) { \ + switch (*(fmtEnd - 1)) { \ + case 'u': \ + case 'd': \ + case 'i': \ + case 'o': \ + case 'X': \ + case 'x': \ + out << static_cast(value); \ + break; \ + default: \ + out << value; \ + break; \ + } \ + } +// per 3.9.1: char, signed char and unsigned char are all distinct types +TINYFORMAT_DEFINE_FORMATVALUE_CHAR(char) +TINYFORMAT_DEFINE_FORMATVALUE_CHAR(signed char) +TINYFORMAT_DEFINE_FORMATVALUE_CHAR(unsigned char) +#undef TINYFORMAT_DEFINE_FORMATVALUE_CHAR + +//------------------------------------------------------------------------------ +// Tools for emulating variadic templates in C++98. The basic idea here is +// stolen from the boost preprocessor metaprogramming library and cut down to +// be just general enough for what we need. + +#define TINYFORMAT_ARGTYPES(n) TINYFORMAT_ARGTYPES_##n +#define TINYFORMAT_VARARGS(n) TINYFORMAT_VARARGS_##n +#define TINYFORMAT_PASSARGS(n) TINYFORMAT_PASSARGS_##n +#define TINYFORMAT_PASSARGS_TAIL(n) TINYFORMAT_PASSARGS_TAIL_##n + +// To keep it as transparent as possible, the macros below have been generated +// using python via the excellent cog.py code generation script. This avoids +// the need for a bunch of complex (but more general) preprocessor tricks as +// used in boost.preprocessor. +// +// To rerun the code generation in place, use `cog.py -r tinyformat.h` +// (see http://nedbatchelder.com/code/cog). Alternatively you can just create +// extra versions by hand. + +/*[[[cog +maxParams = 16 + +def makeCommaSepLists(lineTemplate, elemTemplate, startInd=1): + for j in range(startInd,maxParams+1): + list = ', '.join([elemTemplate % {'i':i} for i in range(startInd,j+1)]) + cog.outl(lineTemplate % {'j':j, 'list':list}) + +makeCommaSepLists('#define TINYFORMAT_ARGTYPES_%(j)d %(list)s', + 'class T%(i)d') + +cog.outl() +makeCommaSepLists('#define TINYFORMAT_VARARGS_%(j)d %(list)s', + 'const T%(i)d& v%(i)d') + +cog.outl() +makeCommaSepLists('#define TINYFORMAT_PASSARGS_%(j)d %(list)s', 'v%(i)d') + +cog.outl() +cog.outl('#define TINYFORMAT_PASSARGS_TAIL_1') +makeCommaSepLists('#define TINYFORMAT_PASSARGS_TAIL_%(j)d , %(list)s', + 'v%(i)d', startInd = 2) + +cog.outl() +cog.outl('#define TINYFORMAT_FOREACH_ARGNUM(m) \\\n ' + + ' '.join(['m(%d)' % (j,) for j in range(1,maxParams+1)])) +]]]*/ +#define TINYFORMAT_ARGTYPES_1 class T1 +#define TINYFORMAT_ARGTYPES_2 class T1, class T2 +#define TINYFORMAT_ARGTYPES_3 class T1, class T2, class T3 +#define TINYFORMAT_ARGTYPES_4 class T1, class T2, class T3, class T4 +#define TINYFORMAT_ARGTYPES_5 class T1, class T2, class T3, class T4, class T5 +#define TINYFORMAT_ARGTYPES_6 \ + class T1, class T2, class T3, class T4, class T5, class T6 +#define TINYFORMAT_ARGTYPES_7 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7 +#define TINYFORMAT_ARGTYPES_8 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, class T8 +#define TINYFORMAT_ARGTYPES_9 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, \ + class T8, class T9 +#define TINYFORMAT_ARGTYPES_10 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, \ + class T8, class T9, class T10 +#define TINYFORMAT_ARGTYPES_11 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, \ + class T8, class T9, class T10, class T11 +#define TINYFORMAT_ARGTYPES_12 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, \ + class T8, class T9, class T10, class T11, class T12 +#define TINYFORMAT_ARGTYPES_13 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, \ + class T8, class T9, class T10, class T11, class T12, class T13 +#define TINYFORMAT_ARGTYPES_14 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, \ + class T8, class T9, class T10, class T11, class T12, class T13, \ + class T14 +#define TINYFORMAT_ARGTYPES_15 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, \ + class T8, class T9, class T10, class T11, class T12, class T13, \ + class T14, class T15 +#define TINYFORMAT_ARGTYPES_16 \ + class T1, class T2, class T3, class T4, class T5, class T6, class T7, \ + class T8, class T9, class T10, class T11, class T12, class T13, \ + class T14, class T15, class T16 + +#define TINYFORMAT_VARARGS_1 const T1 &v1 +#define TINYFORMAT_VARARGS_2 const T1 &v1, const T2 &v2 +#define TINYFORMAT_VARARGS_3 const T1 &v1, const T2 &v2, const T3 &v3 +#define TINYFORMAT_VARARGS_4 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4 +#define TINYFORMAT_VARARGS_5 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5 +#define TINYFORMAT_VARARGS_6 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6 +#define TINYFORMAT_VARARGS_7 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7 +#define TINYFORMAT_VARARGS_8 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8 +#define TINYFORMAT_VARARGS_9 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8, const T9 &v9 +#define TINYFORMAT_VARARGS_10 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8, const T9 &v9, const T10 &v10 +#define TINYFORMAT_VARARGS_11 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8, const T9 &v9, const T10 &v10, \ + const T11 &v11 +#define TINYFORMAT_VARARGS_12 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8, const T9 &v9, const T10 &v10, \ + const T11 &v11, const T12 &v12 +#define TINYFORMAT_VARARGS_13 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8, const T9 &v9, const T10 &v10, \ + const T11 &v11, const T12 &v12, const T13 &v13 +#define TINYFORMAT_VARARGS_14 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8, const T9 &v9, const T10 &v10, \ + const T11 &v11, const T12 &v12, const T13 &v13, const T14 &v14 +#define TINYFORMAT_VARARGS_15 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8, const T9 &v9, const T10 &v10, \ + const T11 &v11, const T12 &v12, const T13 &v13, const T14 &v14, \ + const T15 &v15 +#define TINYFORMAT_VARARGS_16 \ + const T1 &v1, const T2 &v2, const T3 &v3, const T4 &v4, const T5 &v5, \ + const T6 &v6, const T7 &v7, const T8 &v8, const T9 &v9, const T10 &v10, \ + const T11 &v11, const T12 &v12, const T13 &v13, const T14 &v14, \ + const T15 &v15, const T16 &v16 + +#define TINYFORMAT_PASSARGS_1 v1 +#define TINYFORMAT_PASSARGS_2 v1, v2 +#define TINYFORMAT_PASSARGS_3 v1, v2, v3 +#define TINYFORMAT_PASSARGS_4 v1, v2, v3, v4 +#define TINYFORMAT_PASSARGS_5 v1, v2, v3, v4, v5 +#define TINYFORMAT_PASSARGS_6 v1, v2, v3, v4, v5, v6 +#define TINYFORMAT_PASSARGS_7 v1, v2, v3, v4, v5, v6, v7 +#define TINYFORMAT_PASSARGS_8 v1, v2, v3, v4, v5, v6, v7, v8 +#define TINYFORMAT_PASSARGS_9 v1, v2, v3, v4, v5, v6, v7, v8, v9 +#define TINYFORMAT_PASSARGS_10 v1, v2, v3, v4, v5, v6, v7, v8, v9, v10 +#define TINYFORMAT_PASSARGS_11 v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11 +#define TINYFORMAT_PASSARGS_12 v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12 +#define TINYFORMAT_PASSARGS_13 \ + v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13 +#define TINYFORMAT_PASSARGS_14 \ + v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14 +#define TINYFORMAT_PASSARGS_15 \ + v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15 +#define TINYFORMAT_PASSARGS_16 \ + v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16 + +#define TINYFORMAT_PASSARGS_TAIL_1 +#define TINYFORMAT_PASSARGS_TAIL_2 , v2 +#define TINYFORMAT_PASSARGS_TAIL_3 , v2, v3 +#define TINYFORMAT_PASSARGS_TAIL_4 , v2, v3, v4 +#define TINYFORMAT_PASSARGS_TAIL_5 , v2, v3, v4, v5 +#define TINYFORMAT_PASSARGS_TAIL_6 , v2, v3, v4, v5, v6 +#define TINYFORMAT_PASSARGS_TAIL_7 , v2, v3, v4, v5, v6, v7 +#define TINYFORMAT_PASSARGS_TAIL_8 , v2, v3, v4, v5, v6, v7, v8 +#define TINYFORMAT_PASSARGS_TAIL_9 , v2, v3, v4, v5, v6, v7, v8, v9 +#define TINYFORMAT_PASSARGS_TAIL_10 , v2, v3, v4, v5, v6, v7, v8, v9, v10 +#define TINYFORMAT_PASSARGS_TAIL_11 , v2, v3, v4, v5, v6, v7, v8, v9, v10, v11 +#define TINYFORMAT_PASSARGS_TAIL_12 \ + , v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12 +#define TINYFORMAT_PASSARGS_TAIL_13 \ + , v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13 +#define TINYFORMAT_PASSARGS_TAIL_14 \ + , v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14 +#define TINYFORMAT_PASSARGS_TAIL_15 \ + , v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15 +#define TINYFORMAT_PASSARGS_TAIL_16 \ + , v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16 + +#define TINYFORMAT_FOREACH_ARGNUM(m) \ + m(1) m(2) m(3) m(4) m(5) m(6) m(7) m(8) m(9) m(10) m(11) m(12) m(13) m(14) \ + m(15) m(16) +//[[[end]]] + +namespace detail { + +// Type-opaque holder for an argument to format(), with associated actions on +// the type held as explicit function pointers. This allows FormatArg's for +// each argument to be allocated as a homogenous array inside FormatList +// whereas a naive implementation based on inheritance does not. +class FormatArg { +public: + FormatArg() {} + + template + FormatArg(const T &value) + : m_value(static_cast(&value)), + m_formatImpl(&formatImpl), + m_toIntImpl(&toIntImpl) {} + + void format(std::ostream &out, + const char *fmtBegin, + const char *fmtEnd, + int ntrunc) const { + m_formatImpl(out, fmtBegin, fmtEnd, ntrunc, m_value); + } + + int toInt() const { return m_toIntImpl(m_value); } + +private: + template + static void formatImpl(std::ostream &out, + const char *fmtBegin, + const char *fmtEnd, + int ntrunc, + const void *value) { + formatValue(out, fmtBegin, fmtEnd, ntrunc, *static_cast(value)); + } + + template + static int toIntImpl(const void *value) { + return convertToInt::invoke(*static_cast(value)); + } + + const void *m_value; + void (*m_formatImpl)(std::ostream &out, + const char *fmtBegin, + const char *fmtEnd, + int ntrunc, + const void *value); + int (*m_toIntImpl)(const void *value); +}; + +// Parse and return an integer from the string c, as atoi() +// On return, c is set to one past the end of the integer. +inline int parseIntAndAdvance(const char *&c) { + int i = 0; + for (; *c >= '0' && *c <= '9'; ++c) i = 10 * i + (*c - '0'); + return i; +} + +// Print literal part of format string and return next format spec +// position. +// +// Skips over any occurrences of '%%', printing a literal '%' to the +// output. The position of the first % character of the next +// nontrivial format spec is returned, or the end of string. +inline const char *printFormatStringLiteral(std::ostream &out, + const char *fmt) { + const char *c = fmt; + for (;; ++c) { + switch (*c) { + case '\0': + out.write(fmt, c - fmt); + return c; + case '%': + out.write(fmt, c - fmt); + if (*(c + 1) != '%') return c; + // for "%%", tack trailing % onto next literal section. + fmt = ++c; + break; + default: + break; + } + } +} + +// Parse a format string and set the stream state accordingly. +// +// The format mini-language recognized here is meant to be the one from C99, +// with the form "%[flags][width][.precision][length]type". +// +// Formatting options which can't be natively represented using the ostream +// state are returned in spacePadPositive (for space padded positive numbers) +// and ntrunc (for truncating conversions). argIndex is incremented if +// necessary to pull out variable width and precision . The function returns a +// pointer to the character after the end of the current format spec. +inline const char *streamStateFromFormat(std::ostream &out, + bool &spacePadPositive, + int &ntrunc, + const char *fmtStart, + const detail::FormatArg *formatters, + int &argIndex, + int numFormatters) { + if (*fmtStart != '%') { + TINYFORMAT_ERROR( + "tinyformat: Not enough conversion specifiers in format string"); + return fmtStart; + } + // Reset stream state to defaults. + out.width(0); + out.precision(6); + out.fill(' '); + // Reset most flags; ignore irrelevant unitbuf & skipws. + out.unsetf(std::ios::adjustfield | std::ios::basefield | + std::ios::floatfield | std::ios::showbase | std::ios::boolalpha | + std::ios::showpoint | std::ios::showpos | std::ios::uppercase); + bool precisionSet = false; + bool widthSet = false; + int widthExtra = 0; + const char *c = fmtStart + 1; + // 1) Parse flags + for (;; ++c) { + switch (*c) { + case '#': + out.setf(std::ios::showpoint | std::ios::showbase); + continue; + case '0': + // overridden by left alignment ('-' flag) + if (!(out.flags() & std::ios::left)) { + // Use internal padding so that numeric values are + // formatted correctly, eg -00010 rather than 000-10 + out.fill('0'); + out.setf(std::ios::internal, std::ios::adjustfield); + } + continue; + case '-': + out.fill(' '); + out.setf(std::ios::left, std::ios::adjustfield); + continue; + case ' ': + // overridden by show positive sign, '+' flag. + if (!(out.flags() & std::ios::showpos)) spacePadPositive = true; + continue; + case '+': + out.setf(std::ios::showpos); + spacePadPositive = false; + widthExtra = 1; + continue; + default: + break; + } + break; + } + // 2) Parse width + if (*c >= '0' && *c <= '9') { + widthSet = true; + out.width(parseIntAndAdvance(c)); + } + if (*c == '*') { + widthSet = true; + int width = 0; + if (argIndex < numFormatters) + width = formatters[argIndex++].toInt(); + else + TINYFORMAT_ERROR( + "tinyformat: Not enough arguments to read variable width"); + if (width < 0) { + // negative widths correspond to '-' flag set + out.fill(' '); + out.setf(std::ios::left, std::ios::adjustfield); + width = -width; + } + out.width(width); + ++c; + } + // 3) Parse precision + if (*c == '.') { + ++c; + int precision = 0; + if (*c == '*') { + ++c; + if (argIndex < numFormatters) + precision = formatters[argIndex++].toInt(); + else + TINYFORMAT_ERROR( + "tinyformat: Not enough arguments to read variable precision"); + } else { + if (*c >= '0' && *c <= '9') + precision = parseIntAndAdvance(c); + else if (*c == '-') // negative precisions ignored, treated as zero. + parseIntAndAdvance(++c); + } + out.precision(precision); + precisionSet = true; + } + // 4) Ignore any C99 length modifier + while (*c == 'l' || *c == 'h' || *c == 'L' || *c == 'j' || *c == 'z' || + *c == 't') + ++c; + // 5) We're up to the conversion specifier character. + // Set stream flags based on conversion specifier (thanks to the + // boost::format class for forging the way here). + bool intConversion = false; + switch (*c) { + case 'u': + case 'd': + case 'i': + out.setf(std::ios::dec, std::ios::basefield); + intConversion = true; + break; + case 'o': + out.setf(std::ios::oct, std::ios::basefield); + intConversion = true; + break; + case 'X': + out.setf(std::ios::uppercase); + case 'x': + case 'p': + out.setf(std::ios::hex, std::ios::basefield); + intConversion = true; + break; + case 'E': + out.setf(std::ios::uppercase); + case 'e': + out.setf(std::ios::scientific, std::ios::floatfield); + out.setf(std::ios::dec, std::ios::basefield); + break; + case 'F': + out.setf(std::ios::uppercase); + case 'f': + out.setf(std::ios::fixed, std::ios::floatfield); + break; + case 'G': + out.setf(std::ios::uppercase); + case 'g': + out.setf(std::ios::dec, std::ios::basefield); + // As in boost::format, let stream decide float format. + out.flags(out.flags() & ~std::ios::floatfield); + break; + case 'a': + case 'A': + TINYFORMAT_ERROR( + "tinyformat: the %a and %A conversion specs " + "are not supported"); + break; + case 'c': + // Handled as special case inside formatValue() + break; + case 's': + if (precisionSet) ntrunc = static_cast(out.precision()); + // Make %s print booleans as "true" and "false" + out.setf(std::ios::boolalpha); + break; + case 'n': + // Not supported - will cause problems! + TINYFORMAT_ERROR("tinyformat: %n conversion spec not supported"); + break; + case '\0': + TINYFORMAT_ERROR( + "tinyformat: Conversion spec incorrectly " + "terminated by end of string"); + return c; + default: + break; + } + if (intConversion && precisionSet && !widthSet) { + // "precision" for integers gives the minimum number of digits (to be + // padded with zeros on the left). This isn't really supported by the + // iostreams, but we can approximately simulate it with the width if + // the width isn't otherwise used. + out.width(out.precision() + widthExtra); + out.setf(std::ios::internal, std::ios::adjustfield); + out.fill('0'); + } + return c + 1; +} + +//------------------------------------------------------------------------------ +inline void formatImpl(std::ostream &out, + const char *fmt, + const detail::FormatArg *formatters, + int numFormatters) { + // Saved stream state + std::streamsize origWidth = out.width(); + std::streamsize origPrecision = out.precision(); + std::ios::fmtflags origFlags = out.flags(); + char origFill = out.fill(); + + for (int argIndex = 0; argIndex < numFormatters; ++argIndex) { + // Parse the format string + fmt = printFormatStringLiteral(out, fmt); + bool spacePadPositive = false; + int ntrunc = -1; + const char *fmtEnd = streamStateFromFormat(out, + spacePadPositive, + ntrunc, + fmt, + formatters, + argIndex, + numFormatters); + if (argIndex >= numFormatters) { + // Check args remain after reading any variable width/precision + TINYFORMAT_ERROR("tinyformat: Not enough format arguments"); + return; + } + const FormatArg &arg = formatters[argIndex]; + // Format the arg into the stream. + if (!spacePadPositive) + arg.format(out, fmt, fmtEnd, ntrunc); + else { + // The following is a special case with no direct correspondence + // between stream formatting and the printf() behaviour. Simulate + // it crudely by formatting into a temporary string stream and + // munging the resulting string. + std::ostringstream tmpStream; + tmpStream.copyfmt(out); + tmpStream.setf(std::ios::showpos); + arg.format(tmpStream, fmt, fmtEnd, ntrunc); + std::string result = tmpStream.str(); // allocates... yuck. + for (size_t i = 0, iend = result.size(); i < iend; ++i) + if (result[i] == '+') result[i] = ' '; + out << result; + } + fmt = fmtEnd; + } + + // Print remaining part of format string. + fmt = printFormatStringLiteral(out, fmt); + if (*fmt != '\0') + TINYFORMAT_ERROR( + "tinyformat: Too many conversion specifiers in format string"); + + // Restore stream state + out.width(origWidth); + out.precision(origPrecision); + out.flags(origFlags); + out.fill(origFill); +} + +} // namespace detail + +/// List of template arguments format(), held in a type-opaque way. +/// +/// A const reference to FormatList (typedef'd as FormatListRef) may be +/// conveniently used to pass arguments to non-template functions: All type +/// information has been stripped from the arguments, leaving just enough of a +/// common interface to perform formatting as required. +class FormatList { +public: + FormatList(detail::FormatArg *formatters, int N) + : m_formatters(formatters), m_N(N) {} + + friend void vformat(std::ostream &out, + const char *fmt, + const FormatList &list); + +private: + const detail::FormatArg *m_formatters; + int m_N; +}; + +/// Reference to type-opaque format list for passing to vformat() +typedef const FormatList &FormatListRef; + +namespace detail { + +// Format list subclass with fixed storage to avoid dynamic allocation +template +class FormatListN : public FormatList { +public: + template + FormatListN(const Args &... args) + : FormatList(&m_formatterStore[0], N), + m_formatterStore{FormatArg(args)...} { + static_assert(sizeof...(args) == N, "Number of args must be N"); + } + +private: + FormatArg m_formatterStore[N]; +}; + +// Special 0-arg version - MSVC says zero-sized C array in struct is nonstandard +template <> +class FormatListN<0> : public FormatList { +public: + FormatListN() : FormatList(0, 0) {} +}; + +} // namespace detail + +//------------------------------------------------------------------------------ +// Primary API functions + +/// Make type-agnostic format list from list of template arguments. +/// +/// The exact return type of this function is an implementation detail and +/// shouldn't be relied upon. Instead it should be stored as a FormatListRef: +/// +/// FormatListRef formatList = makeFormatList( /*...*/ ); +template +detail::FormatListN makeFormatList(const Args &... args) { + return detail::FormatListN(args...); +} + +/// Format list of arguments to the stream according to the given format string. +/// +/// The name vformat() is chosen for the semantic similarity to vprintf(): the +/// list of format arguments is held in a single function argument. +inline void vformat(std::ostream &out, const char *fmt, FormatListRef list) { + detail::formatImpl(out, fmt, list.m_formatters, list.m_N); +} + +/// Format list of arguments to the stream according to given format string. +template +void format(std::ostream &out, const char *fmt, const Args &... args) { + vformat(out, fmt, makeFormatList(args...)); +} + +/// Format list of arguments according to the given format string and return +/// the result as a string. +template +std::string format(const char *fmt, const Args &... args) { + std::ostringstream oss; + format(oss, fmt, args...); + return oss.str(); +} + +/// Format list of arguments to std::cout, according to the given format string +template +void printf(const char *fmt, const Args &... args) { + format(std::cout, fmt, args...); +} + +template +void printfln(const char *fmt, const Args &... args) { + format(std::cout, fmt, args...); + std::cout << '\n'; +} + +} // namespace tinyformat +} // namespace string +} // namespace paddle diff --git a/paddle/strings/CMakeLists.txt b/paddle/strings/CMakeLists.txt deleted file mode 100644 index 4e55eecd484c0e218ecd51bbd19b3eb4f6f92a25..0000000000000000000000000000000000000000 --- a/paddle/strings/CMakeLists.txt +++ /dev/null @@ -1,2 +0,0 @@ -cc_library(stringpiece SRCS stringpiece.cc) -cc_test(stringpiece_test SRCS stringpiece_test.cc DEPS stringpiece glog gflags) diff --git a/paddle/strings/stringpiece.cc b/paddle/strings/stringpiece.cc deleted file mode 100644 index 415b3558d5dfffde26275bcb16ea3922424ca9f3..0000000000000000000000000000000000000000 --- a/paddle/strings/stringpiece.cc +++ /dev/null @@ -1,141 +0,0 @@ -/* - Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. -*/ - -#include "paddle/strings/stringpiece.h" - -#include - -#include -#include -#include - -namespace paddle { - -StringPiece::StringPiece() : data_(NULL), size_(0) {} - -StringPiece::StringPiece(const char* d, size_t n) : data_(d), size_(n) { - if (d == NULL && n != 0) - throw std::invalid_argument( - "StringPiece requires len to be 0 for NULL data"); -} - -StringPiece::StringPiece(const char* s) : data_(s) { - size_ = (s == NULL) ? 0 : strlen(s); -} - -StringPiece::StringPiece(const std::string& s) - : data_(s.data()), size_(s.size()) {} - -char StringPiece::operator[](size_t n) const { - if (n >= len()) - throw std::invalid_argument("index out of StringPiece length"); - return data_[n]; -} - -int Compare(StringPiece a, StringPiece b) { - const size_t min_len = (a.len() < b.len()) ? a.len() : b.len(); - int r = memcmp(a.data(), b.data(), min_len); - if (r == 0) { - if (a.len() < b.len()) - return -1; - else if (a.len() > b.len()) - return 1; - } - return r; -} - -bool operator==(StringPiece x, StringPiece y) { - return ((x.len() == y.len()) && - (x.data() == y.data() || memcmp(x.data(), y.data(), x.len()) == 0)); -} - -bool operator!=(StringPiece x, StringPiece y) { return !(x == y); } - -bool operator<(StringPiece x, StringPiece y) { return Compare(x, y) < 0; } -bool operator>(StringPiece x, StringPiece y) { return Compare(x, y) > 0; } - -bool operator<=(StringPiece x, StringPiece y) { return Compare(x, y) <= 0; } -bool operator>=(StringPiece x, StringPiece y) { return Compare(x, y) >= 0; } - -bool HasPrefix(StringPiece s, StringPiece x) { - return ((s.len() >= x.len()) && (memcmp(s.data(), x.data(), x.len()) == 0)); -} - -bool HasSuffix(StringPiece s, StringPiece x) { - return ((s.len() >= x.len()) && - (memcmp(s.data() + (s.len() - x.len()), x.data(), x.len()) == 0)); -} - -StringPiece SkipPrefix(StringPiece s, size_t n) { - if (n > s.len()) - throw std::invalid_argument("Skip distance larger than StringPiece length"); - return StringPiece(s.data() + n, s.len() - n); -} - -StringPiece SkipSuffix(StringPiece s, size_t n) { - if (n > s.len()) - throw std::invalid_argument("Skip distance larger than StringPiece length"); - return StringPiece(s.data(), s.len() - n); -} - -StringPiece TrimPrefix(StringPiece s, StringPiece x) { - return HasPrefix(s, x) ? SkipPrefix(s, x.len()) : s; -} - -StringPiece TrimSuffix(StringPiece s, StringPiece x) { - return HasSuffix(s, x) ? SkipSuffix(s, x.len()) : s; -} - -bool Contains(StringPiece s, StringPiece sub) { - return std::search(s.begin(), s.end(), sub.begin(), sub.end()) != s.end(); -} - -size_t Index(StringPiece s, StringPiece sub) { - auto e = std::search(s.begin(), s.end(), sub.begin(), sub.end()); - return e != s.end() ? e - s.data() : StringPiece::npos; -} - -size_t Find(StringPiece s, char c, size_t pos) { - if (pos >= s.len()) { - return StringPiece::npos; - } - const char* result = - reinterpret_cast(memchr(s.data() + pos, c, s.len() - pos)); - return result != nullptr ? result - s.data() : StringPiece::npos; -} - -size_t RFind(StringPiece s, char c, size_t pos) { - if (s.len() == 0) return StringPiece::npos; - for (const char* p = s.data() + std::min(pos, s.len() - 1); p >= s.data(); - p--) { - if (*p == c) { - return p - s.data(); - } - } - return StringPiece::npos; -} - -StringPiece SubStr(StringPiece s, size_t pos, size_t n) { - if (pos > s.len()) pos = s.len(); - if (n > s.len() - pos) n = s.len() - pos; - return StringPiece(s.data() + pos, n); -} - -std::ostream& operator<<(std::ostream& o, StringPiece piece) { - return o << piece.ToString(); -} - -} // namespace paddle diff --git a/paddle/testing/CMakeLists.txt b/paddle/testing/CMakeLists.txt index c47add04b081cbdf78b5a5d3bca3a71025b3d9ac..4245df5ab72bf0fd67261818b307f0babdb5d685 100644 --- a/paddle/testing/CMakeLists.txt +++ b/paddle/testing/CMakeLists.txt @@ -2,7 +2,7 @@ if(WITH_TESTING) add_library(paddle_test_main STATIC TestMain.cpp) - add_dependencies(paddle_test_main gen_proto_cpp) + add_dependencies(paddle_test_main paddle_proto ${external_project_dependencies}) add_library(paddle_test_util STATIC TestUtil.cpp) - add_dependencies(paddle_test_util gen_proto_cpp) + add_dependencies(paddle_test_util paddle_proto ${external_project_dependencies}) endif() diff --git a/paddle/trainer/CMakeLists.txt b/paddle/trainer/CMakeLists.txt index f34d53ae99f913a8aed8767b7271a538efce4778..eac0584d30958ab78a935d89d217a4876fb07a19 100644 --- a/paddle/trainer/CMakeLists.txt +++ b/paddle/trainer/CMakeLists.txt @@ -41,7 +41,8 @@ add_style_check_target(paddle_trainer_lib add_style_check_target(paddle_trainer_lib ${TRAINER_HEADERS}) add_dependencies(paddle_trainer_lib - gen_proto_cpp) + paddle_proto + ${external_project_dependencies}) macro(add_paddle_exe TARGET_NAME) add_executable(${TARGET_NAME} ${ARGN}) @@ -72,6 +73,6 @@ endif() if(WITH_GOLANG) add_dependencies(paddle_trainer_lib paddle_pserver_cclient) - target_link_libraries(paddle_trainer ${CMAKE_BINARY_DIR}/go/pserver/cclient/libpaddle_pserver_cclient.a) - target_link_libraries(paddle_trainer_lib ${CMAKE_BINARY_DIR}/go/pserver/cclient/libpaddle_pserver_cclient.a) + target_link_libraries(paddle_trainer paddle_pserver_cclient) + target_link_libraries(paddle_trainer_lib paddle_pserver_cclient) endif(WITH_GOLANG) diff --git a/paddle/utils/CMakeLists.txt b/paddle/utils/CMakeLists.txt index af59951752d1799c95e293d3eae233e6aa26e5f3..7a4977935ede4878c07f4fb6ba0dd76bf50acd42 100644 --- a/paddle/utils/CMakeLists.txt +++ b/paddle/utils/CMakeLists.txt @@ -17,7 +17,7 @@ add_library(paddle_utils STATIC add_style_check_target(paddle_utils ${UTIL_HEADERS}) add_style_check_target(paddle_utils ${UTIL_SOURCES} ${UTIL_ARCH_SOURCES}) -add_dependencies(paddle_utils gen_proto_cpp) +add_dependencies(paddle_utils paddle_proto ${external_project_dependencies}) if(WITH_TESTING) add_subdirectory(tests) endif() diff --git a/paddle/utils/Error.h b/paddle/utils/Error.h index f3d535c69c53fa350612459560dd9ac7c279aa72..27ddaab3f003110a2684a871a2de17afb473d660 100644 --- a/paddle/utils/Error.h +++ b/paddle/utils/Error.h @@ -19,7 +19,21 @@ limitations under the License. */ #include #include #include -#include "paddle/platform/must_check.h" + +/** + * __must_check macro. It make the function's return value must be used, + * otherwise it will raise a compile warning. And also Paddle treat all compile + * warnings as errors. + */ +#ifdef __GNUC__ +#if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) >= 30400 +#define __must_check __attribute__((warn_unused_result)) +#else +#define __must_check +#endif +#else +#define __must_check +#endif namespace paddle { diff --git a/proto/CMakeLists.txt b/proto/CMakeLists.txt index c942620990765832f21c887d30f85a2d211a5f32..18584cafe7971bad281b498908c54780250791b7 100644 --- a/proto/CMakeLists.txt +++ b/proto/CMakeLists.txt @@ -1,43 +1,23 @@ -set(proto_filenames - DataConfig.proto - DataFormat.proto - ModelConfig.proto - ParameterConfig.proto - ParameterService.proto - TrainerConfig.proto - OptimizerConfig.proto - ParameterServerConfig.proto) +file(GLOB proto_filenames . *.proto) +include_directories(${CMAKE_CURRENT_BINARY_DIR}) +proto_library(paddle_proto SRCS ${proto_filenames}) set(PROTO_GEN) set(PROTO_GEN_PY) foreach(filename ${proto_filenames}) - get_filename_component(base_filename ${filename} NAME_WE) - set(CUR_PROTO_GEN - ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.pb.h - ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.pb.cc) - set(PROTO_GEN - ${PROTO_GEN} - ${CUR_PROTO_GEN}) - add_custom_command(OUTPUT ${CUR_PROTO_GEN} - COMMAND env ${py_env} ${PROTOBUF_PROTOC_EXECUTABLE} - --cpp_out ${CMAKE_CURRENT_BINARY_DIR} - --proto_path ${PROJ_ROOT}/proto ${PROJ_ROOT}/proto/${filename} - DEPENDS ${filename} ${external_project_dependencies}) - + get_filename_component(ABS_FIL ${filename} ABSOLUTE) + get_filename_component(FIL_WE ${filename} NAME_WE) set(CUR_PROTO_GEN_PY - ${PROJ_ROOT}/paddle/python/paddle/proto/${base_filename}_pb2.py) + ${PROJ_ROOT}/paddle/python/paddle/proto/${FIL_WE}_pb2.py) set(PROTO_GEN_PY - ${CUR_PROTO_GEN_PY} - ${PROTO_GEN_PY}) + ${CUR_PROTO_GEN_PY} + ${PROTO_GEN_PY}) add_custom_command(OUTPUT ${CUR_PROTO_GEN_PY} - COMMAND env ${py_env} ${PROTOBUF_PROTOC_EXECUTABLE} --python_out ${PROJ_ROOT}/python/paddle/proto - --proto_path ${PROJ_ROOT}/proto ${PROJ_ROOT}/proto/${filename} - DEPENDS ${filename} ${external_project_dependencies}) + COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} + ARGS "--python_out=${PROJ_ROOT}/python/paddle/proto" + "-I" ${CMAKE_CURRENT_SOURCE_DIR} ${ABS_FIL} + DEPENDS ${ABS_FIL} ${external_project_dependencies}) endforeach() -add_custom_target(gen_proto_cpp ALL DEPENDS ${PROTO_GEN}) add_custom_target(gen_proto_py ALL DEPENDS ${PROTO_GEN_PY}) - -add_library(paddle_proto STATIC ${PROTO_GEN}) -target_include_directories(paddle_proto PUBLIC ${CMAKE_CURRENT_BINARY_DIR}) diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt index 3640dd3a75ea212a84255ea7f6369b63606482ab..361e764e25ba1801bd22f785bc282e51f058aae6 100644 --- a/python/CMakeLists.txt +++ b/python/CMakeLists.txt @@ -7,10 +7,21 @@ file(GLOB UTILS_PY_FILES . ./paddle/utils/*.py) file(GLOB_RECURSE V2_PY_FILES ./paddle/v2/ *.py) set(PY_FILES paddle/__init__.py - ${TRAINER_PY_FILES} - ${HELPERS_PY_FILES} - ${UTILS_PY_FILES} - ${V2_PY_FILES}) + ${TRAINER_PY_FILES} + ${HELPERS_PY_FILES} + ${UTILS_PY_FILES} + ${V2_PY_FILES}) + +add_custom_target(copy_paddle_master) + +SET(COPY_PADDLE_MASTER "") +if(WITH_GOLANG) + SET(COPY_PADDLE_MASTER "copy_paddle_master") + add_custom_command(TARGET ${COPY_PADDLE_MASTER} + COMMAND cp ${paddle_master_LIB_PATH} ${PROJ_ROOT}/python/paddle/v2/master/ + ) + add_dependencies(copy_paddle_master paddle_master) +endif(WITH_GOLANG) configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in ${CMAKE_CURRENT_BINARY_DIR}/setup.py) @@ -18,7 +29,7 @@ configure_file(${CMAKE_CURRENT_SOURCE_DIR}/setup.py.in add_custom_command(OUTPUT ${OUTPUT_DIR}/.timestamp COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel COMMAND ${CMAKE_COMMAND} -E touch ${OUTPUT_DIR}/.timestamp - DEPENDS gen_proto_py ${PY_FILES} ${external_project_dependencies}) + DEPENDS gen_proto_py ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER}) add_custom_target(paddle_python ALL DEPENDS ${OUTPUT_DIR}/.timestamp) diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py index 58e4902f57aa8018b820f48f6cbf659f1e5f5183..b7418101d83fde1b91781d3a42b056cc7708cba9 100644 --- a/python/paddle/trainer/config_parser.py +++ b/python/paddle/trainer/config_parser.py @@ -2082,10 +2082,10 @@ class MaxOutLayer(LayerBase): class RowConvLayer(LayerBase): def __init__(self, name, inputs, context_length, **xargs): super(RowConvLayer, self).__init__( - name, 'maxout', 0, inputs=inputs, **xargs) + name, 'row_conv', 0, inputs=inputs, **xargs) config_assert( len(self.inputs) == 1, - 'TransLayer must have one and only one input') + 'row convolution layer must have one and only one input.') input_layer = self.get_input_layer(0) row_conv_conf = self.config.inputs[0].row_conv_conf row_conv_conf.context_length = context_length diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 84ed160773065da15fc26bfb5c5882b068874f1c..a601d5c84ad222785e68b9fa81c51b1e120b4f29 100755 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -1149,10 +1149,10 @@ def pooling_layer(input, @layer_support(DROPOUT) def lstmemory(input, name=None, + size=None, reverse=False, act=None, gate_act=None, - size=None, state_act=None, bias_attr=None, param_attr=None, @@ -1194,6 +1194,8 @@ def lstmemory(input, :param name: The lstmemory layer name. :type name: basestring + :param size: DEPRECATED. size of the lstm cell + :type size: int :param input: input layer name. :type input: LayerOutput :param reverse: is sequence process reversed or not. @@ -1220,15 +1222,15 @@ def lstmemory(input, assert state_act.support_hppl assert act.support_hppl assert input.size is not None and input.size % 4 == 0 + if size is not None: if input.size / 4 == size: plog = logger.warning else: plog = logger.fatal - - plog("NOTE: The lstmemory layer[%s]'s size is set by previous input " - "layer. The lstm size should be equal with input layer size/4. The" - " size which is set explicitly will be ignored." % name) + plog("size of lstmemory layer: %s is automatically set to " + "size of input layer / 4. The parameter size passing to " + "this layer is ignored." % (name)) Layer( name=name, @@ -1255,11 +1257,11 @@ def lstmemory(input, @wrap_name_default("gru") @layer_support(DROPOUT) def grumemory(input, + size=None, name=None, reverse=False, act=None, gate_act=None, - size=None, bias_attr=None, param_attr=None, layer_attr=None): @@ -1318,6 +1320,8 @@ def grumemory(input, :type name: None|basestring :param input: input layer. :type input: LayerOutput. + :param size: DEPRECATED. size of the gru cell + :type size: int :param reverse: Whether sequence process is reversed or not. :type reverse: bool :param act: activation type, TanhActivation by default. This activation @@ -1334,9 +1338,6 @@ def grumemory(input, :type param_attr: ParameterAttribute|None|False :param layer_attr: Extra Layer attribute :type layer_attr: ExtraLayerAttribute|None - :param size: Stub parameter of size, but actually not used. If set this size - will get a warning. - :type size: None :return: LayerOutput object. :rtype: LayerOutput """ @@ -1348,9 +1349,9 @@ def grumemory(input, plog = logger.warning else: plog = logger.fatal - plog("NOTE: the gru memory layer's size is set by previous input layer," - " and should be input size / 3. Set size explicitly will be " - "ignored.") + plog("size of grumemory layer: %s is automatically set to " + "size of input layer / 3. The parameter size passing to this " + "layer is ignored." % (name)) Layer( name=name, @@ -2524,8 +2525,8 @@ def img_cmrnorm_layer(input, @wrap_bias_attr_default() -@wrap_param_attr_default(default_factory=lambda _: ParamAttr(initial_mean=1.0, - initial_std=0.)) +@wrap_param_attr_default( + default_factory=lambda _: ParamAttr(initial_mean=1.0, initial_std=0.)) @wrap_act_default(act=ReluActivation()) @wrap_name_default("batch_norm") @layer_support(DROPOUT) @@ -3013,25 +3014,25 @@ def lstm_step_layer(input, bias_attr=None, layer_attr=None): """ - LSTM Step Layer. It used in recurrent_group. The lstm equations are shown - as follow. + LSTM Step Layer. This function is used only in recurrent_group. + The lstm equations are shown as follows. .. math:: - i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i) + i_t & = \\sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + W_{c_i}c_{t-1} + b_i) - f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f) + f_t & = \\sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + W_{c_f}c_{t-1} + b_f) - c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c) + c_t & = f_tc_{t-1} + i_t tanh (W_{x_c}x_t+W_{h_c}h_{t-1} + b_c) - o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o) + o_t & = \\sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + W_{c_o}c_t + b_o) h_t & = o_t tanh(c_t) The input of lstm step is :math:`Wx_t + Wh_{t-1}`, and user should use :code:`mixed_layer` and :code:`full_matrix_projection` to calculate these - input vector. + input vectors. The state of lstm step is :math:`c_{t-1}`. And lstm step layer will do @@ -3042,14 +3043,14 @@ def lstm_step_layer(input, ... - This layer contains two outputs. Default output is :math:`h_t`. The other - output is :math:`o_t`, which name is 'state' and can use + This layer has two outputs. Default output is :math:`h_t`. The other + output is :math:`o_t`, whose name is 'state' and can use :code:`get_output_layer` to extract this output. :param name: Layer's name. :type name: basestring - :param size: Layer's size. NOTE: lstm layer's size, should be equal as - :code:`input.size/4`, and should be equal as + :param size: Layer's size. NOTE: lstm layer's size, should be equal to + :code:`input.size/4`, and should be equal to :code:`state.size`. :type size: int :param input: input layer. :math:`Wx_t + Wh_{t-1}` diff --git a/python/paddle/trainer_config_helpers/networks.py b/python/paddle/trainer_config_helpers/networks.py index 67154a8d7d366bd983b4426da87e0b33307fced4..b77932ce5f09470329a97cc0a6273942a9155c6a 100755 --- a/python/paddle/trainer_config_helpers/networks.py +++ b/python/paddle/trainer_config_helpers/networks.py @@ -614,6 +614,7 @@ def simple_lstm(input, @wrap_name_default('lstm_unit') def lstmemory_unit(input, + memory_boot=None, name=None, size=None, param_attr=None, @@ -626,9 +627,9 @@ def lstmemory_unit(input, lstm_layer_attr=None, get_output_layer_attr=None): """ - Define calculations that a LSTM unit performs in a single time step. - This function itself is not a recurrent layer, so that it can not be - directly applied to sequence input. This function is always used in + Define calculations that a LSTM unit performs during a single time step. + This function itself is not a recurrent layer, so it can not be + directly used to process sequence inputs. This function is always used in recurrent_group (see layers.py for more details) to implement attention mechanism. @@ -638,13 +639,13 @@ def lstmemory_unit(input, .. math:: - i_t & = \\sigma(W_{xi}x_{t} + W_{hi}h_{t-1} + W_{ci}c_{t-1} + b_i) + i_t & = \\sigma(W_{x_i}x_{t} + W_{h_i}h_{t-1} + W_{c_i}c_{t-1} + b_i) - f_t & = \\sigma(W_{xf}x_{t} + W_{hf}h_{t-1} + W_{cf}c_{t-1} + b_f) + f_t & = \\sigma(W_{x_f}x_{t} + W_{h_f}h_{t-1} + W_{c_f}c_{t-1} + b_f) - c_t & = f_tc_{t-1} + i_t tanh (W_{xc}x_t+W_{hc}h_{t-1} + b_c) + c_t & = f_tc_{t-1} + i_t tanh (W_{x_c}x_t+W_{h_c}h_{t-1} + b_c) - o_t & = \\sigma(W_{xo}x_{t} + W_{ho}h_{t-1} + W_{co}c_t + b_o) + o_t & = \\sigma(W_{x_o}x_{t} + W_{h_o}h_{t-1} + W_{c_o}c_t + b_o) h_t & = o_t tanh(c_t) @@ -661,6 +662,8 @@ def lstmemory_unit(input, :param input: input layer name. :type input: LayerOutput + :param memory_boot: the initialization state of the LSTM cell. + :type memory_boot: LayerOutput | None :param name: lstmemory unit name. :type name: basestring :param size: lstmemory unit size. @@ -692,7 +695,8 @@ def lstmemory_unit(input, assert input.size % 4 == 0 size = input.size / 4 out_mem = memory(name=name, size=size) - state_mem = memory(name="%s_state" % name, size=size) + state_mem = memory( + name="%s_state" % name, size=size, boot_layer=memory_boot) with mixed_layer( name="%s_input_recurrent" % name, @@ -726,6 +730,7 @@ def lstmemory_unit(input, def lstmemory_group(input, size=None, name=None, + memory_boot=None, reverse=False, param_attr=None, act=None, @@ -737,7 +742,7 @@ def lstmemory_group(input, lstm_layer_attr=None, get_output_layer_attr=None): """ - lstm_group is a recurrent layer group version of Long Short Term Memory. It + lstm_group is a recurrent_group version of Long Short Term Memory. It does exactly the same calculation as the lstmemory layer (see lstmemory in layers.py for the maths) does. A promising benefit is that LSTM memory cell states, or hidden states in every time step are accessible to the @@ -748,8 +753,8 @@ def lstmemory_group(input, NOTE: In PaddlePaddle's implementation, the following input-to-hidden multiplications: - :math:`W_{xi}x_{t}` , :math:`W_{xf}x_{t}`, - :math:`W_{xc}x_t`, :math:`W_{xo}x_{t}` are not done in lstmemory_unit to + :math:`W_{x_i}x_{t}` , :math:`W_{x_f}x_{t}`, + :math:`W_{x_c}x_t`, :math:`W_{x_o}x_{t}` are not done in lstmemory_unit to speed up the calculations. Consequently, an additional mixed_layer with full_matrix_projection must be included before lstmemory_unit is called. @@ -765,10 +770,12 @@ def lstmemory_group(input, :param input: input layer name. :type input: LayerOutput - :param name: lstmemory group name. - :type name: basestring :param size: lstmemory group size. :type size: int + :param name: name of the lstmemory group. + :type name: basestring + :param memory_boot: the initialization state of LSTM cell. + :type memory_boot: LayerOutput | None :param reverse: is lstm reversed :type reverse: bool :param param_attr: Parameter config, None if use default. @@ -798,6 +805,7 @@ def lstmemory_group(input, def __lstm_step__(ipt): return lstmemory_unit( input=ipt, + memory_boot=memory_boot, name=name, size=size, mixed_bias_attr=mixed_bias_attr, @@ -819,6 +827,7 @@ def lstmemory_group(input, @wrap_name_default('gru_unit') def gru_unit(input, + memory_boot=None, size=None, name=None, gru_bias_attr=None, @@ -829,8 +838,8 @@ def gru_unit(input, naive=False): """ Define calculations that a gated recurrent unit performs in a single time - step. This function itself is not a recurrent layer, so that it can not be - directly applied to sequence input. This function is almost always used in + step. This function itself is not a recurrent layer, so it can not be + directly used to process sequence inputs. This function is always used in the recurrent_group (see layers.py for more details) to implement attention mechanism. @@ -838,6 +847,8 @@ def gru_unit(input, :param input: input layer name. :type input: LayerOutput + :param memory_boot: the initialization state of the LSTM cell. + :type memory_boot: LayerOutput | None :param name: name of the gru group. :type name: basestring :param size: hidden size of the gru. @@ -856,7 +867,7 @@ def gru_unit(input, if size is None: size = input.size / 3 - out_mem = memory(name=name, size=size) + out_mem = memory(name=name, size=size, boot_layer=memory_boot) if naive: __step__ = gru_step_naive_layer @@ -878,6 +889,7 @@ def gru_unit(input, @wrap_name_default('gru_group') def gru_group(input, + memory_boot=None, size=None, name=None, reverse=False, @@ -888,7 +900,7 @@ def gru_group(input, gru_layer_attr=None, naive=False): """ - gru_group is a recurrent layer group version of Gated Recurrent Unit. It + gru_group is a recurrent_group version of Gated Recurrent Unit. It does exactly the same calculation as the grumemory layer does. A promising benefit is that gru hidden states are accessible to the user. This is especially useful in attention model. If you do not need to access @@ -908,6 +920,8 @@ def gru_group(input, :param input: input layer name. :type input: LayerOutput + :param memory_boot: the initialization state of the LSTM cell. + :type memory_boot: LayerOutput | None :param name: name of the gru group. :type name: basestring :param size: hidden size of the gru. @@ -929,6 +943,7 @@ def gru_group(input, def __gru_step__(ipt): return gru_unit( input=ipt, + memory_boot=memory_boot, name=name, size=size, gru_bias_attr=gru_bias_attr, @@ -1083,7 +1098,6 @@ def simple_gru2(input, return grumemory( name=name, - size=size, input=m, reverse=reverse, bias_attr=gru_bias_attr, diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_row_conv.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_row_conv.protostr index 9ec15d2a19ec50a1729f9eeaa6dce8b1153c776b..19c9f16574ca6fb3a9e9dbfb2d1f52024e604239 100644 --- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_row_conv.protostr +++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_row_conv.protostr @@ -7,7 +7,7 @@ layers { } layers { name: "__row_conv_layer_0__" - type: "maxout" + type: "row_conv" size: 2560 active_type: "relu" inputs { diff --git a/python/paddle/v2/__init__.py b/python/paddle/v2/__init__.py index 6a1e23a343d6a8de9dbec573f257efb4fb658e92..3ba5c31871807027e452df5d889b3b403e1c6414 100644 --- a/python/paddle/v2/__init__.py +++ b/python/paddle/v2/__init__.py @@ -56,6 +56,7 @@ __all__ = [ 'plot', 'evaluator', 'image', + 'master', ] diff --git a/python/paddle/v2/dataset/__init__.py b/python/paddle/v2/dataset/__init__.py index 26252d5bbd77ddb70b4f03843679e4737e2f96d3..2e4beb6882789249db09705f3f4d6c5c19e492cd 100644 --- a/python/paddle/v2/dataset/__init__.py +++ b/python/paddle/v2/dataset/__init__.py @@ -25,8 +25,9 @@ import uci_housing import sentiment import wmt14 import mq2007 +import flowers __all__ = [ 'mnist', 'imikolov', 'imdb', 'cifar', 'movielens', 'conll05', 'sentiment' - 'uci_housing', 'wmt14', 'mq2007' + 'uci_housing', 'wmt14', 'mq2007', 'flowers' ] diff --git a/python/paddle/v2/dataset/flowers.py b/python/paddle/v2/dataset/flowers.py index 07c13cf719ae0c864c23fef51f0bd7d47f265759..e2a21e6e3e04e79fdfc225ce1b4496b6b69d1e89 100644 --- a/python/paddle/v2/dataset/flowers.py +++ b/python/paddle/v2/dataset/flowers.py @@ -13,30 +13,31 @@ # limitations under the License. """ This module will download dataset from -http://www.robots.ox.ac.uk/~vgg/data/flowers/102/index.html +http://www.robots.ox.ac.uk/~vgg/data/flowers/102/index.html and parse train/test set intopaddle reader creators. -This set contains images of flowers belonging to 102 different categories. +This set contains images of flowers belonging to 102 different categories. The images were acquired by searching the web and taking pictures. There are a minimum of 40 images for each category. The database was used in: Nilsback, M-E. and Zisserman, A. Automated flower classification over a large - number of classes.Proceedings of the Indian Conference on Computer Vision, -Graphics and Image Processing (2008) + number of classes.Proceedings of the Indian Conference on Computer Vision, +Graphics and Image Processing (2008) http://www.robots.ox.ac.uk/~vgg/publications/papers/nilsback08.{pdf,ps.gz}. """ import cPickle import itertools +import functools from common import download import tarfile import scipy.io as scio from paddle.v2.image import * +from paddle.v2.reader import * import os import numpy as np -import paddle.v2 as paddle from multiprocessing import cpu_count __all__ = ['train', 'test', 'valid'] @@ -46,39 +47,51 @@ SETID_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/setid.mat' DATA_MD5 = '52808999861908f626f3c1f4e79d11fa' LABEL_MD5 = 'e0620be6f572b9609742df49c70aed4d' SETID_MD5 = 'a5357ecc9cb78c4bef273ce3793fc85c' +# In official 'readme', tstid is the flag of test data +# and trnid is the flag of train data. But test data is more than train data. +# So we exchange the train data and test data. +TRAIN_FLAG = 'tstid' +TEST_FLAG = 'trnid' +VALID_FLAG = 'valid' -def default_mapper(sample): +def default_mapper(is_train, sample): ''' map image bytes data to type needed by model input layer ''' img, label = sample - img = paddle.image.load_image_bytes(img) - img = paddle.image.simple_transform(img, 256, 224, True) + img = load_image_bytes(img) + img = simple_transform( + img, 256, 224, is_train, mean=[103.94, 116.78, 123.68]) return img.flatten().astype('float32'), label +train_mapper = functools.partial(default_mapper, True) +test_mapper = functools.partial(default_mapper, False) + + def reader_creator(data_file, label_file, setid_file, dataset_name, - mapper=default_mapper, - buffered_size=1024): + mapper, + buffered_size=1024, + use_xmap=True): ''' - 1. read images from tar file and + 1. read images from tar file and merge images into batch files in 102flowers.tgz_batch/ 2. get a reader to read sample from batch file - - :param data_file: downloaded data file + + :param data_file: downloaded data file :type data_file: string - :param label_file: downloaded label file + :param label_file: downloaded label file :type label_file: string :param setid_file: downloaded setid file containing information about how to split dataset :type setid_file: string :param dataset_name: data set name (tstid|trnid|valid) :type dataset_name: string - :param mapper: a function to map image bytes data to type + :param mapper: a function to map image bytes data to type needed by model input layer :type mapper: callable :param buffered_size: the size of buffer used to process images @@ -105,15 +118,17 @@ def reader_creator(data_file, for sample, label in itertools.izip(data, batch['label']): yield sample, int(label) - return paddle.reader.xmap_readers(mapper, reader, - cpu_count(), buffered_size) + if use_xmap: + return xmap_readers(mapper, reader, cpu_count(), buffered_size) + else: + return map_readers(mapper, reader) -def train(mapper=default_mapper, buffered_size=1024): +def train(mapper=train_mapper, buffered_size=1024, use_xmap=True): ''' - Create flowers training set reader. - It returns a reader, each sample in the reader is - image pixels in [0, 1] and label in [1, 102] + Create flowers training set reader. + It returns a reader, each sample in the reader is + image pixels in [0, 1] and label in [1, 102] translated from original color image by steps: 1. resize to 256*256 2. random crop to 224*224 @@ -128,15 +143,15 @@ def train(mapper=default_mapper, buffered_size=1024): return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), 'trnid', mapper, - buffered_size) + download(SETID_URL, 'flowers', SETID_MD5), TRAIN_FLAG, mapper, + buffered_size, use_xmap) -def test(mapper=default_mapper, buffered_size=1024): +def test(mapper=test_mapper, buffered_size=1024, use_xmap=True): ''' - Create flowers test set reader. - It returns a reader, each sample in the reader is - image pixels in [0, 1] and label in [1, 102] + Create flowers test set reader. + It returns a reader, each sample in the reader is + image pixels in [0, 1] and label in [1, 102] translated from original color image by steps: 1. resize to 256*256 2. random crop to 224*224 @@ -151,15 +166,15 @@ def test(mapper=default_mapper, buffered_size=1024): return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), 'tstid', mapper, - buffered_size) + download(SETID_URL, 'flowers', SETID_MD5), TEST_FLAG, mapper, + buffered_size, use_xmap) -def valid(mapper=default_mapper, buffered_size=1024): +def valid(mapper=test_mapper, buffered_size=1024, use_xmap=True): ''' - Create flowers validation set reader. - It returns a reader, each sample in the reader is - image pixels in [0, 1] and label in [1, 102] + Create flowers validation set reader. + It returns a reader, each sample in the reader is + image pixels in [0, 1] and label in [1, 102] translated from original color image by steps: 1. resize to 256*256 2. random crop to 224*224 @@ -174,8 +189,8 @@ def valid(mapper=default_mapper, buffered_size=1024): return reader_creator( download(DATA_URL, 'flowers', DATA_MD5), download(LABEL_URL, 'flowers', LABEL_MD5), - download(SETID_URL, 'flowers', SETID_MD5), 'valid', mapper, - buffered_size) + download(SETID_URL, 'flowers', SETID_MD5), VALID_FLAG, mapper, + buffered_size, use_xmap) def fetch(): diff --git a/python/paddle/v2/dataset/tests/flowers_test.py b/python/paddle/v2/dataset/tests/flowers_test.py index cc0626f4feae287d18dfb227cc69a4174da055da..a8ae9a07acc22eb9d3c0cc5ebb07f8f11ed21233 100644 --- a/python/paddle/v2/dataset/tests/flowers_test.py +++ b/python/paddle/v2/dataset/tests/flowers_test.py @@ -31,13 +31,13 @@ class TestFlowers(unittest.TestCase): def test_train(self): instances, max_label_value = self.check_reader( paddle.v2.dataset.flowers.train()) - self.assertEqual(instances, 1020) + self.assertEqual(instances, 6149) self.assertEqual(max_label_value, 102) def test_test(self): instances, max_label_value = self.check_reader( paddle.v2.dataset.flowers.test()) - self.assertEqual(instances, 6149) + self.assertEqual(instances, 1020) self.assertEqual(max_label_value, 102) def test_valid(self): diff --git a/python/paddle/v2/dataset/uci_housing.py b/python/paddle/v2/dataset/uci_housing.py index c715ea96819659c60215d61e5701ca565bb5d3ff..ec10ce646ebf3eca2c2a6423b69ee11b6a2b99cf 100644 --- a/python/paddle/v2/dataset/uci_housing.py +++ b/python/paddle/v2/dataset/uci_housing.py @@ -14,7 +14,7 @@ """ UCI Housing dataset. -This module will paddle.v2.dataset.common.download dataset from +This module will download dataset from https://archive.ics.uci.edu/ml/machine-learning-databases/housing/ and parse training set and test set into paddle reader creators. """ diff --git a/python/paddle/v2/image.py b/python/paddle/v2/image.py index 0d648e9ae697ff0373c6cdc166608d395a8d8086..965d965335a56a97448bd8c738b03eceaee550e2 100644 --- a/python/paddle/v2/image.py +++ b/python/paddle/v2/image.py @@ -262,7 +262,12 @@ def left_right_flip(im): return im[:, ::-1, :] -def simple_transform(im, resize_size, crop_size, is_train, is_color=True): +def simple_transform(im, + resize_size, + crop_size, + is_train, + is_color=True, + mean=None): """ Simply data argumentation for training. These operations include resizing, croping and flipping. @@ -288,7 +293,19 @@ def simple_transform(im, resize_size, crop_size, is_train, is_color=True): im = left_right_flip(im) else: im = center_crop(im, crop_size) - im = to_chw(im) + if len(im.shape) == 3: + im = to_chw(im) + + im = im.astype('float32') + if mean is not None: + mean = np.array(mean, dtype=np.float32) + # mean value, may be one value per channel + if mean.ndim == 1: + mean = mean[:, np.newaxis, np.newaxis] + else: + # elementwise mean + assert len(mean.shape) == len(im) + im -= mean return im @@ -297,7 +314,8 @@ def load_and_transform(filename, resize_size, crop_size, is_train, - is_color=True): + is_color=True, + mean=None): """ Load image from the input file `filename` and transform image for data argumentation. Please refer to the `simple_transform` interface @@ -318,5 +336,5 @@ def load_and_transform(filename, :type is_train: bool """ im = load_image(filename) - im = simple_transform(im, resize_size, crop_size, is_train, is_color) + im = simple_transform(im, resize_size, crop_size, is_train, is_color, mean) return im diff --git a/python/paddle/v2/parameters.py b/python/paddle/v2/parameters.py index ad20241b98302f136326ae491c6723a6c12ae284..bbaf8bfa979fbbf460561ebf7077b75b9c41a11a 100644 --- a/python/paddle/v2/parameters.py +++ b/python/paddle/v2/parameters.py @@ -51,7 +51,7 @@ class Parameters(object): def __init__(self): self.__param_conf__ = dict() self.__gradient_machines__ = [] - self.__tmp_params__ = [] + self.__tmp_params__ = dict() def __append_config__(self, param_conf): """ @@ -128,13 +128,10 @@ class Parameters(object): if len(self.__gradient_machines__) == 0: # create new parameter in python numpy. - if len(self.__tmp_params__) != 0: - ret_list = [ - mat for name, mat in self.__tmp_params__ if name == key - ] - if len(ret_list) == 1: - return ret_list[0] - return np.ndarray(shape=shape, dtype=np.float32) + if key in self.__tmp_params__: + return self.__tmp_params__[key] + else: + return np.ndarray(shape=shape, dtype=np.float32) else: for each_gradient_machine in self.__gradient_machines__: param = __get_parameter_in_gradient_machine__( @@ -187,7 +184,7 @@ class Parameters(object): (shape, value.shape)) if len(self.__gradient_machines__) == 0: - self.__tmp_params__.append((key, value)) + self.__tmp_params__[key] = value else: for each_gradient_machine in self.__gradient_machines__: __copy_parameter_to_gradient_machine__(each_gradient_machine, @@ -231,7 +228,7 @@ class Parameters(object): raise ValueError("gradient_machine should be api.GradientMachine") if len(self.__tmp_params__) != 0: - for name, val in self.__tmp_params__: + for name, val in self.__tmp_params__.iteritems(): try: __copy_parameter_to_gradient_machine__(gradient_machine, name, val) @@ -287,6 +284,18 @@ class Parameters(object): @staticmethod def from_tar(f): + """ + Create a `Parameters` object from the given file. And + the `Parameters` only contains the parameters in this + file. It is adapted the parameters are same in the + defined network and the given file. For example, it + can be used in the inference. + + :param f: the initialized model file. + :type f: tar file + :return: A Parameters object. + :rtype: Parameters. + """ params = Parameters() tar = tarfile.TarFile(fileobj=f, mode='r') for finfo in tar: @@ -302,6 +311,21 @@ class Parameters(object): params.deserialize(param_name, f) return params + def init_from_tar(self, f): + """ + Different from `from_tar`, this interface can be used to + init partial network parameters from another saved model. + + :param f: the initialized model file. + :type f: tar file + :return: Nothing. + """ + + tar_param = Parameters.from_tar(f) + for pname in tar_param.names(): + if pname in self.names(): + self.set(pname, tar_param.get(pname)) + def __get_parameter_in_gradient_machine__(gradient_machine, name): """ diff --git a/python/paddle/v2/reader/decorator.py b/python/paddle/v2/reader/decorator.py index e432003129d2b8dea60138d08f13ec5e9d29a7ad..45a4288751e37b99dd1005ec78f30a98044926ff 100644 --- a/python/paddle/v2/reader/decorator.py +++ b/python/paddle/v2/reader/decorator.py @@ -166,12 +166,12 @@ def buffered(reader, size): The buffered data reader will read and save data entries into a buffer. Reading from the buffered data reader will proceed as long as the buffer is not empty. - + :param reader: the data reader to read from. :type reader: callable :param size: max buffer size. :type size: int - + :returns: the buffered data reader. """ @@ -238,7 +238,7 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False): :type mapper: callable :param reader: the data reader to read from :type reader: callable - :param process_num: process number to handle original sample + :param process_num: process number to handle original sample :type process_num: int :param buffer_size: max buffer size :type buffer_size: int @@ -248,9 +248,6 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False): :rtype: callable """ end = XmapEndSignal() - in_queue = Queue(buffer_size) - out_queue = Queue(buffer_size) - out_order = [0] # define a worker to read samples from reader to in_queue def read_worker(reader, in_queue): @@ -266,12 +263,6 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False): in_order += 1 in_queue.put(end) - # start a read worker in a thread - target = order_read_worker if order else read_worker - t = Thread(target=target, args=(reader, in_queue)) - t.daemon = True - t.start() - # define a worker to handle samples from in_queue by mapper # and put mapped samples into out_queue def handle_worker(in_queue, out_queue, mapper): @@ -298,19 +289,27 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False): in_queue.put(end) out_queue.put(end) - # start several handle_workers - target = order_handle_worker if order else handle_worker - args = (in_queue, out_queue, mapper, out_order) if order else ( - in_queue, out_queue, mapper) - workers = [] - for i in xrange(process_num): - worker = Thread(target=target, args=args) - worker.daemon = True - workers.append(worker) - for w in workers: - w.start() - def xreader(): + in_queue = Queue(buffer_size) + out_queue = Queue(buffer_size) + out_order = [0] + # start a read worker in a thread + target = order_read_worker if order else read_worker + t = Thread(target=target, args=(reader, in_queue)) + t.daemon = True + t.start() + # start several handle_workers + target = order_handle_worker if order else handle_worker + args = (in_queue, out_queue, mapper, out_order) if order else ( + in_queue, out_queue, mapper) + workers = [] + for i in xrange(process_num): + worker = Thread(target=target, args=args) + worker.daemon = True + workers.append(worker) + for w in workers: + w.start() + sample = out_queue.get() while not isinstance(sample, XmapEndSignal): yield sample diff --git a/python/paddle/v2/reader/tests/decorator_test.py b/python/paddle/v2/reader/tests/decorator_test.py index bb3c5d220b9ce1552d2fc429abb1863930cd4d17..5a92951b100fa51ab6df7039d9c6b54d1f9d963e 100644 --- a/python/paddle/v2/reader/tests/decorator_test.py +++ b/python/paddle/v2/reader/tests/decorator_test.py @@ -132,15 +132,17 @@ class TestXmap(unittest.TestCase): for order in orders: for tNum in thread_nums: for size in buffered_size: - result = [] - for i in paddle.v2.reader.xmap_readers(mapper, + reader = paddle.v2.reader.xmap_readers(mapper, reader_creator_10(0), - tNum, size, order)(): - result.append(i) - if not order: - result.sort() - for idx, e in enumerate(result): - self.assertEqual(e, mapper(idx)) + tNum, size, order) + for n in xrange(3): + result = [] + for i in reader(): + result.append(i) + if not order: + result.sort() + for idx, e in enumerate(result): + self.assertEqual(e, mapper(idx)) if __name__ == '__main__': diff --git a/python/paddle/v2/tests/test_parameters.py b/python/paddle/v2/tests/test_parameters.py index 45372e7dd0ec7cbdd6a2eb5c0397ef7e74284cd0..7ba8a939fbd1a949d61a007b40c054e7543c0cbc 100644 --- a/python/paddle/v2/tests/test_parameters.py +++ b/python/paddle/v2/tests/test_parameters.py @@ -20,14 +20,17 @@ import cStringIO import numpy -def __rand_param_config__(name): +def __rand_param_config__(name, psize=None): conf = ParameterConfig() conf.name = name size = 1 - for i in xrange(2): - dim = random.randint(1, 1000) - conf.dims.append(dim) - size *= dim + if psize is None: + for i in xrange(2): + dim = random.randint(1, 1000) + conf.dims.append(dim) + size *= dim + else: + size = psize conf.size = size assert conf.IsInitialized() return conf @@ -77,6 +80,50 @@ class TestParameters(unittest.TestCase): expected = numpy.array([[1, 1], [1, 2], [1, 1]], numpy.float32) assert numpy.logical_and.reduce(numpy.reshape(val == expected, 6)) + def test_init_from_tar(self): + def get_param(names, size): + p = parameters.Parameters() + for k, v in zip(names, size): + p.__append_config__(__rand_param_config__(k, v)) + for name in p.names(): + param = p.get(name) + param[:] = numpy.random.uniform( + -1.0, 1.0, size=p.get_shape(name)) + p.set(name, param) + return p + + def get_parames(): + name1 = ['param_0', 'param_1'] + size1 = [128, 256] + p1 = get_param(name1, size1) + file1 = cStringIO.StringIO() + p1.to_tar(file1) + file1.seek(0) + + name2 = ['param_0', 'param_1', 'param_2'] + size2 = [128, 256, 288] + p2 = get_param(name2, size2) + file2 = cStringIO.StringIO() + p2.to_tar(file2) + file2.seek(0) + return p1, file1, p2, file2 + + p1, file1, p2, file2 = get_parames() + p2.init_from_tar(file1) + for name in p1.names(): + self.assertEqual(p1.get_shape(name), p2.get_shape(name)) + v1 = p1.get(name) + v2 = p2.get(name) + self.assertTrue(numpy.isclose(v1, v2).all()) + + p1, file1, p2, file2 = get_parames() + p1.init_from_tar(file2) + for name in p1.names(): + self.assertEqual(p1.get_shape(name), p2.get_shape(name)) + v1 = p1.get(name) + v2 = p2.get(name) + self.assertTrue(numpy.isclose(v1, v2).all()) + if __name__ == '__main__': unittest.main() diff --git a/python/setup.py.in b/python/setup.py.in index 86fc0fc5c0318b03659bf84f8ad9e2a114467c74..dae01664876a913b49403d3f95001f009721f73b 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -8,6 +8,7 @@ packages=['paddle', 'paddle.v2', 'paddle.v2.dataset', 'paddle.v2.reader', + 'paddle.v2.master', 'paddle.v2.plot'] setup_requires=["requests", @@ -15,7 +16,8 @@ setup_requires=["requests", "protobuf==3.1", "recordio", "matplotlib", - "rarfile"] + "rarfile", + "scipy>=0.19.0"] if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']: setup_requires+=["opencv-python"] @@ -25,6 +27,7 @@ setup(name='paddle', description='Parallel Distributed Deep Learning', install_requires=setup_requires, packages=packages, + package_data={'paddle.v2.master': ['${paddle_master_LIB_NAME}'], }, package_dir={ '': '${CMAKE_CURRENT_SOURCE_DIR}' },