diff --git a/CMakeLists.txt b/CMakeLists.txt index 61f5e63098c40f..8e7ffe72b5fb84 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -54,23 +54,12 @@ option(WITH_NGRAPH "Compile PaddlePaddle with nGraph support." OFF) option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON) option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF) option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON) -option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF) -option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF) -option(WITH_TIMER "Compile PaddlePaddle with stats timer" OFF) option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools" OFF) option(WITH_JEMALLOC "Compile PaddlePaddle with jemalloc" OFF) -option(WITH_DOC "Compile PaddlePaddle with documentation" OFF) option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF) option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF) -option(ON_TRAVIS "Exclude special unit test on Travis CI" OFF) -option(WITH_FLUID_ONLY "Compile PaddlePaddle fluid only" OFF) -option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) -option(GLIDE_INSTALL "Download and install go dependencies " ON) option(WITH_DISTRIBUTE "Compile with distributed support" OFF) option(WITH_PSLIB "Compile with pslib support" OFF) -option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) -option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF) -option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF) option(WITH_CONTRIB "Compile the third-party contributation" OFF) option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF) option(WITH_ANAKIN "Compile with Anakin library" OFF) @@ -105,8 +94,6 @@ endif() if (WIN32) set(WITH_DISTRIBUTE OFF CACHE STRING "Disable DISTRIBUTE when compiling for Windows" FORCE) - set(WITH_FLUID_ONLY ON CACHE STRING - "Enable FLUID_ONLY when compiling for Windows" FORCE) endif() set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING @@ -148,7 +135,6 @@ include(external/openblas) # download, build, install openblas include(external/mkldnn) # download, build, install mkldnn include(external/ngraph) # download, build, install nGraph include(external/boost) # download boost -include(external/any) # download libn::any include(external/eigen) # download eigen3 include(external/pybind11) # download pybind11 include(external/cares) @@ -225,7 +211,6 @@ include(generic) # simplify cmake module include(package) # set paddle packages include(ccache) # set ccache for compilation include(util) # set unittest and link libs -include(rdma) # set rdma libraries include(version) # set PADDLE_VERSION include(coveralls) # set code coverage include(inference_lib) # add paddle fluid inference libraries @@ -233,38 +218,11 @@ include(inference_lib) # add paddle fluid inference libraries include_directories("${PADDLE_SOURCE_DIR}") -set(EXTERNAL_LIBS - gflags - glog - ${CBLAS_LIBRARIES} - protobuf - zlib - ${PYTHON_LIBRARIES} -) - -if(WITH_PSLIB) - list(APPEND EXTERNAL_LIBS pslib) - list(APPEND EXTERNAL_LIBS pslib_brpc) - list(APPEND EXTERNAL_LIBS libmct) -endif(WITH_PSLIB) - if(WITH_AMD_GPU) find_package(HIP) include(hip) endif(WITH_AMD_GPU) -if(WITH_MKLML) - list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB}) -endif() - -if(WITH_LIBXSMM) - list(APPEND EXTERNAL_LIBS ${LIBXSMM_LIBS}) -endif() - -if(WITH_MKLDNN) - list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB}) -endif() - set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build") set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG") diff --git a/cmake/configure.cmake b/cmake/configure.cmake index b0f54bf49aafb6..93d74bb0a8f726 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -20,31 +20,10 @@ if(WITH_DSO) add_definitions(-DPADDLE_USE_DSO) endif(WITH_DSO) -if(WITH_DOUBLE) - add_definitions(-DPADDLE_TYPE_DOUBLE) -endif(WITH_DOUBLE) - -if(WITH_ARM_FP16) - add_definitions(-DPADDLE_ARM_FP16) - add_definitions("-march=armv8.2-a+fp16+simd") -endif(WITH_ARM_FP16) - if(WITH_TESTING) add_definitions(-DPADDLE_WITH_TESTING) endif(WITH_TESTING) -if(NOT WITH_TIMER) - add_definitions(-DPADDLE_DISABLE_TIMER) -endif(NOT WITH_TIMER) - -if(USE_EIGEN_FOR_BLAS) - add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS) -endif(USE_EIGEN_FOR_BLAS) - -if(EIGEN_USE_THREADS) - add_definitions(-DEIGEN_USE_THREADS) -endif(EIGEN_USE_THREADS) - if(NOT WITH_PROFILER) add_definitions(-DPADDLE_DISABLE_PROFILER) endif(NOT WITH_PROFILER) @@ -78,10 +57,6 @@ if(WIN32) endif(NOT MSVC) endif(WIN32) -if(NOT WITH_GOLANG) - add_definitions(-DPADDLE_WITHOUT_GOLANG) -endif(NOT WITH_GOLANG) - if(WITH_PSLIB) add_definitions(-DPADDLE_WITH_PSLIB) endif() @@ -171,55 +146,6 @@ if(WITH_DISTRIBUTE) add_definitions(-DPADDLE_WITH_DISTRIBUTE) endif() -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") - file(MAKE_DIRECTORY "${PADDLE_IN_GOPATH}") - set(PADDLE_GO_PATH "${CMAKE_SOURCE_DIR}/go") - - 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. - WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} - ) - - if (GLIDE_INSTALL) - if(EXISTS $ENV{GOPATH}/bin/glide) - set(GLIDE "$ENV{GOPATH}/bin/glide") - else() - message(FATAL_ERROR "no glide executeble found: $ENV{GOPATH}/bin/glide") - endif() - - # this command will only run when the file it depends is missing - # or has changed, or the output is missing. - add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide - COMMAND env GOPATH=${GOPATH} ${GLIDE} install - COMMAND touch ${CMAKE_BINARY_DIR}/glide - DEPENDS ${PADDLE_SOURCE_DIR}/go/glide.lock - WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go" - ) - - # depends on the custom command which outputs - # ${CMAKE_BINARY_DIR}/glide, the custom command does not need to - # run every time this target is built. - add_custom_target(go_vendor DEPENDS ${CMAKE_BINARY_DIR}/glide go_path) - endif() - -endif(WITH_GOLANG) - if(WITH_GRPC) add_definitions(-DPADDLE_WITH_GRPC) endif(WITH_GRPC) diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake index ef4192ecc98ea6..735846db1db04e 100644 --- a/cmake/cuda.cmake +++ b/cmake/cuda.cmake @@ -168,10 +168,7 @@ elseif (${CUDA_VERSION} LESS 11.0) # CUDA 10.x endif() include_directories(${CUDA_INCLUDE_DIRS}) -list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY}) if(NOT WITH_DSO) - # TODO(panyx0718): CUPTI only allows DSO? - list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY}) if(WIN32) set_property(GLOBAL PROPERTY CUDA_MODULES ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY}) endif(WIN32) diff --git a/cmake/external/anakin.cmake b/cmake/external/anakin.cmake index 06fc6061bc98ee..77f4b34537577c 100644 --- a/cmake/external/anakin.cmake +++ b/cmake/external/anakin.cmake @@ -74,5 +74,3 @@ add_dependencies(anakin_shared extern_anakin) add_library(anakin_saber SHARED IMPORTED GLOBAL) set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB}) add_dependencies(anakin_saber extern_anakin) - -list(APPEND external_project_dependencies anakin_shared anakin_saber) diff --git a/cmake/external/any.cmake b/cmake/external/any.cmake deleted file mode 100644 index 85cce80b70a1fc..00000000000000 --- a/cmake/external/any.cmake +++ /dev/null @@ -1,31 +0,0 @@ -INCLUDE(ExternalProject) - -SET(ANY_SOURCE_DIR ${THIRD_PARTY_PATH}/any) - -INCLUDE_DIRECTORIES(${ANY_SOURCE_DIR}/src/extern_lib_any) - -ExternalProject_Add( - extern_lib_any - ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/PaddlePaddle/any.git" - GIT_TAG "15595d8324be9e8a9a80d9ae442fdd12bd66df5d" - PREFIX ${ANY_SOURCE_DIR} - UPDATE_COMMAND "" - CONFIGURE_COMMAND "" - BUILD_COMMAND "" - INSTALL_COMMAND "" - TEST_COMMAND "" -) - -if (${CMAKE_VERSION} VERSION_LESS "3.3.0") - set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/lib_any_dummy.c) - file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";") - add_library(lib_any STATIC ${dummyfile}) -else() - add_library(lib_any INTERFACE) -endif() - -add_dependencies(lib_any extern_lib_any) - -add_definitions(-DANY_IMPL_ANY_CAST_MOVEABLE) -LIST(APPEND external_project_dependencies lib_any) diff --git a/cmake/external/boost.cmake b/cmake/external/boost.cmake index 12412a51a0fd1a..fc204dc9193bb2 100644 --- a/cmake/external/boost.cmake +++ b/cmake/external/boost.cmake @@ -57,5 +57,4 @@ else() endif() add_dependencies(boost ${BOOST_PROJECT}) -list(APPEND external_project_dependencies boost) set(Boost_INCLUDE_DIR ${BOOST_INCLUDE_DIR}) diff --git a/cmake/external/brpc.cmake b/cmake/external/brpc.cmake index 6b50cff7a66a33..989d1dbd4cf593 100644 --- a/cmake/external/brpc.cmake +++ b/cmake/external/brpc.cmake @@ -69,5 +69,3 @@ SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES}) ADD_DEPENDENCIES(brpc extern_brpc) add_definitions(-DBRPC_WITH_GLOG) - -LIST(APPEND external_project_dependencies brpc) diff --git a/cmake/external/cub.cmake b/cmake/external/cub.cmake index f06728de91e450..41ad8207743201 100644 --- a/cmake/external/cub.cmake +++ b/cmake/external/cub.cmake @@ -31,5 +31,3 @@ else() endif() add_dependencies(cub extern_cub) - -LIST(APPEND external_project_dependencies cub) diff --git a/cmake/external/dlpack.cmake b/cmake/external/dlpack.cmake index 4587475d7902a1..63dd16b28e40a0 100644 --- a/cmake/external/dlpack.cmake +++ b/cmake/external/dlpack.cmake @@ -27,5 +27,3 @@ else() endif() add_dependencies(dlpack extern_dlpack) - -LIST(APPEND external_project_dependencies dlpack) diff --git a/cmake/external/eigen.cmake b/cmake/external/eigen.cmake index 6aef97f21244ef..72441160f89d2c 100644 --- a/cmake/external/eigen.cmake +++ b/cmake/external/eigen.cmake @@ -52,5 +52,3 @@ else() endif() add_dependencies(eigen3 extern_eigen3) - -LIST(APPEND external_project_dependencies eigen3) diff --git a/cmake/external/gflags.cmake b/cmake/external/gflags.cmake index f3ca74faea3629..911920ed6212b8 100644 --- a/cmake/external/gflags.cmake +++ b/cmake/external/gflags.cmake @@ -61,8 +61,6 @@ ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES}) ADD_DEPENDENCIES(gflags extern_gflags) -LIST(APPEND external_project_dependencies gflags) - # On Windows (including MinGW), the Shlwapi library is used by gflags if available. if (WIN32) include(CheckIncludeFileCXX) diff --git a/cmake/external/glog.cmake b/cmake/external/glog.cmake index d3a4d69d3a0551..7fa17ce6b7b106 100644 --- a/cmake/external/glog.cmake +++ b/cmake/external/glog.cmake @@ -72,5 +72,3 @@ ADD_LIBRARY(glog STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES}) ADD_DEPENDENCIES(glog extern_glog gflags) LINK_LIBRARIES(glog gflags) - -LIST(APPEND external_project_dependencies glog) diff --git a/cmake/external/gtest.cmake b/cmake/external/gtest.cmake index 9be625b620287c..e459526583bd5e 100644 --- a/cmake/external/gtest.cmake +++ b/cmake/external/gtest.cmake @@ -79,5 +79,4 @@ IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC)) SET_PROPERTY(TARGET gtest_main PROPERTY IMPORTED_LOCATION ${GTEST_MAIN_LIBRARIES}) ADD_DEPENDENCIES(gtest_main extern_gtest) - LIST(APPEND external_project_dependencies gtest gtest_main) ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC)) diff --git a/cmake/external/leveldb.cmake b/cmake/external/leveldb.cmake index 0df61b01ab64c8..ac0febd076e659 100644 --- a/cmake/external/leveldb.cmake +++ b/cmake/external/leveldb.cmake @@ -39,6 +39,3 @@ ADD_DEPENDENCIES(extern_leveldb snappy) ADD_LIBRARY(leveldb STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET leveldb PROPERTY IMPORTED_LOCATION ${LEVELDB_LIBRARIES}) ADD_DEPENDENCIES(leveldb extern_leveldb) - -LIST(APPEND external_project_dependencies leveldb) - diff --git a/cmake/external/libmct.cmake b/cmake/external/libmct.cmake index 27cff8cfb6315c..b944f2945b7874 100644 --- a/cmake/external/libmct.cmake +++ b/cmake/external/libmct.cmake @@ -72,7 +72,4 @@ else() add_library(libmct INTERFACE) endif() -#ADD_LIBRARY(libmct SHARED IMPORTED GLOBAL) ADD_DEPENDENCIES(libmct ${LIBMCT_PROJECT}) -LIST(APPEND external_project_dependencies libmct) - diff --git a/cmake/external/libxsmm.cmake b/cmake/external/libxsmm.cmake index 39f49d210a20d4..69cdba7c5921f1 100644 --- a/cmake/external/libxsmm.cmake +++ b/cmake/external/libxsmm.cmake @@ -53,5 +53,3 @@ MESSAGE(STATUS "Libxsmm library: ${LIBXSMM_LIBS}") include_directories(${LIBXSMM_INCLUDE_DIR}) ADD_DEFINITIONS(-DPADDLE_WITH_LIBXSMM) ADD_DEPENDENCIES(libxsmm extern_libxsmm) -LIST(APPEND external_project_dependencies libxsmm) - diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index 92fe76d05c7507..94a266c50114a9 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -89,7 +89,6 @@ SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB}) ADD_DEPENDENCIES(shared_mkldnn ${MKLDNN_PROJECT}) MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}") add_definitions(-DPADDLE_WITH_MKLDNN) -LIST(APPEND external_project_dependencies shared_mkldnn) # generate a static dummy target to track mkldnn dependencies # for cc_library(xxx SRCS xxx.c DEPS mkldnn) diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index 2caff273576870..54826cedb87169 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -73,4 +73,3 @@ INCLUDE_DIRECTORIES(${MKLML_INC_DIR}) ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB}) ADD_DEPENDENCIES(mklml ${MKLML_PROJECT}) -LIST(APPEND external_project_dependencies mklml) diff --git a/cmake/external/ngraph.cmake b/cmake/external/ngraph.cmake index 14af98b2d74d4a..5812a61f0ddc3a 100644 --- a/cmake/external/ngraph.cmake +++ b/cmake/external/ngraph.cmake @@ -77,4 +77,3 @@ add_dependencies(ngraph ${NGRAPH_PROJECT}) target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH) target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR}) target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB}) -LIST(APPEND external_project_dependencies ngraph) diff --git a/cmake/external/openblas.cmake b/cmake/external/openblas.cmake index b347a592929836..d8a4a0be6f5aaa 100644 --- a/cmake/external/openblas.cmake +++ b/cmake/external/openblas.cmake @@ -11,11 +11,6 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - -IF(USE_EIGEN_FOR_BLAS) - return() -ENDIF(USE_EIGEN_FOR_BLAS) - INCLUDE(cblas) IF(NOT ${CBLAS_FOUND}) @@ -91,7 +86,6 @@ ENDIF() IF(NOT ${CBLAS_FOUND}) ADD_DEPENDENCIES(cblas extern_openblas) - LIST(APPEND external_project_dependencies cblas) ELSE() IF("${CBLAS_PROVIDER}" STREQUAL "MKLML") ADD_DEPENDENCIES(cblas mklml) diff --git a/cmake/external/protobuf.cmake b/cmake/external/protobuf.cmake index e05b7694ddf1e1..bc7fe5454f5883 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -129,7 +129,6 @@ macro(PROMPT_PROTOBUF_LIB) ADD_DEPENDENCIES(protoc ${dep}) ENDFOREACH() - LIST(APPEND external_project_dependencies protobuf) RETURN() endmacro() macro(SET_PROTOBUF_VERSION) @@ -231,7 +230,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST) ) ENDFUNCTION() -SET(PROTOBUF_VERSION 3.1) +SET(PROTOBUF_VERSION 3.1.0) IF(NOT PROTOBUF_FOUND) build_protobuf(extern_protobuf FALSE) diff --git a/cmake/external/pslib.cmake b/cmake/external/pslib.cmake index b4ea268e5a48e2..0287e5cf2a835e 100644 --- a/cmake/external/pslib.cmake +++ b/cmake/external/pslib.cmake @@ -70,4 +70,3 @@ ExternalProject_Add( ADD_LIBRARY(pslib SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET pslib PROPERTY IMPORTED_LOCATION ${PSLIB_LIB}) ADD_DEPENDENCIES(pslib ${PSLIB_PROJECT}) -LIST(APPEND external_project_dependencies pslib) diff --git a/cmake/external/pslib_brpc.cmake b/cmake/external/pslib_brpc.cmake index 8b43f2ef5c999f..22c8c1b463764b 100644 --- a/cmake/external/pslib_brpc.cmake +++ b/cmake/external/pslib_brpc.cmake @@ -70,4 +70,3 @@ ExternalProject_Add( ADD_LIBRARY(pslib_brpc SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET pslib_brpc PROPERTY IMPORTED_LOCATION ${PSLIB_BRPC_LIB}) ADD_DEPENDENCIES(pslib_brpc ${PSLIB_BRPC_PROJECT}) -LIST(APPEND external_project_dependencies pslib_brpc) diff --git a/cmake/external/threadpool.cmake b/cmake/external/threadpool.cmake index 0159815fed81bd..1f56bc7ab056ef 100644 --- a/cmake/external/threadpool.cmake +++ b/cmake/external/threadpool.cmake @@ -26,5 +26,3 @@ else() endif() add_dependencies(simple_threadpool extern_threadpool) - -LIST(APPEND external_project_dependencies simple_threadpool) diff --git a/cmake/external/warpctc.cmake b/cmake/external/warpctc.cmake index 7a25aaf15f2c7f..6f2af8670f25c0 100644 --- a/cmake/external/warpctc.cmake +++ b/cmake/external/warpctc.cmake @@ -83,5 +83,3 @@ INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include wa ADD_LIBRARY(warpctc SHARED IMPORTED GLOBAL) SET_PROPERTY(TARGET warpctc PROPERTY IMPORTED_LOCATION ${WARPCTC_LIBRARIES}) ADD_DEPENDENCIES(warpctc extern_warpctc) - -LIST(APPEND external_project_dependencies warpctc) diff --git a/cmake/external/xbyak.cmake b/cmake/external/xbyak.cmake index 384c2f9328296c..1d61154c0d45de 100644 --- a/cmake/external/xbyak.cmake +++ b/cmake/external/xbyak.cmake @@ -55,4 +55,3 @@ else() endif() add_dependencies(xbyak ${XBYAK_PROJECT}) -list(APPEND external_project_dependencies xbyak) diff --git a/cmake/external/xxhash.cmake b/cmake/external/xxhash.cmake index a0f300c2e8bab9..23b1e02108642d 100644 --- a/cmake/external/xxhash.cmake +++ b/cmake/external/xxhash.cmake @@ -71,5 +71,3 @@ add_library(xxhash STATIC IMPORTED GLOBAL) set_property(TARGET xxhash PROPERTY IMPORTED_LOCATION ${XXHASH_LIBRARIES}) include_directories(${XXHASH_INCLUDE_DIR}) add_dependencies(xxhash extern_xxhash) - -LIST(APPEND external_project_dependencies xxhash) diff --git a/cmake/external/zlib.cmake b/cmake/external/zlib.cmake index 6c8d79c25e6a26..5569fefe992d10 100644 --- a/cmake/external/zlib.cmake +++ b/cmake/external/zlib.cmake @@ -57,5 +57,3 @@ ENDIF(WIN32) ADD_LIBRARY(zlib STATIC IMPORTED GLOBAL) SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES}) ADD_DEPENDENCIES(zlib extern_zlib) - -LIST(APPEND external_project_dependencies zlib) diff --git a/cmake/hip.cmake b/cmake/hip.cmake index 4276bc5b08cd88..c3a748db502037 100644 --- a/cmake/hip.cmake +++ b/cmake/hip.cmake @@ -11,8 +11,6 @@ include_directories("/opt/rocm/rocrand/include") include_directories("/opt/rocm/rccl/include") include_directories("/opt/rocm/thrust") -list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc") - set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++11" ) if(WITH_DSO) @@ -31,22 +29,12 @@ if(WITH_GRPC) set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_GRPC") endif(WITH_GRPC) -if(NOT WITH_GOLANG) - set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITHOUT_GOLANG") -endif(NOT WITH_GOLANG) - if(WITH_MKLDNN) set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_MKLDNN") endif(WITH_MKLDNN) set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DANY_IMPL_ANY_CAST_MOVEABLE") -if(NOT WITH_RDMA) - set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_DISABLE_RDMA") -endif(NOT WITH_RDMA) - - - if(CMAKE_BUILD_TYPE STREQUAL "Debug") list(APPEND HIP_HCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG}) elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo") diff --git a/cmake/rdma.cmake b/cmake/rdma.cmake deleted file mode 100644 index b698f3bdc3ff58..00000000000000 --- a/cmake/rdma.cmake +++ /dev/null @@ -1,82 +0,0 @@ -# user should download rdma first from subversion repository - -# execute following instruction to download svn mannally -# svn co https://svn.baidu.com/sys/ip/trunk/rdma/sockrdmav1 rdma/ -# svn co https://svn.baidu.com/sys/ip/trunk/rdma/thirdparty rdma/ -# we use static output in svn repositories to avoid implict bugs from not standard runtime env. - -if(WITH_RDMA) - set(RDMA_ROOT $ENV{RDMA_ROOT} CACHE PATH "Folder contains RDMA sock library and thirdparty library") - - function(generate_rdma_links) - #redirect to current DIR to isolate the pollution from system runtime environment - #it can benifits unified control for different gcc environment. - #e.g, by default gcc48 did not refer /usr/lib64 which could contain low version - #runtime libraries that will crash process while loading it. That redirect trick - #can fix it. - execute_process( - COMMAND mkdir -p librdma - COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so.1 - COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so - COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so.1 - COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so - COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so.1 - COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so - WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR} - ) - endfunction(generate_rdma_links) - - #check and set headers - find_path(RDMA_INC_SXISOCK sxi_sock.h PATHS ${RDMA_ROOT}/sockrdmav1/output/include) - find_path(RDMA_INC_XIO libxio.h PATHS ${RDMA_ROOT}/thirdparty/output/accelio) - find_path(RDMA_INC_EVENT event2 PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_path(RDMA_INC_NUMA numa.h PATHS ${RDMA_ROOT}/thirdparty/output/libnuma) - - #check and set libs - find_library(RDMA_LIB_SXISOCK NAMES sxisock PATHS ${RDMA_ROOT}/sockrdmav1/output) - find_library(RDMA_LIB_XIO NAMES xio PATHS ${RDMA_ROOT}/thirdparty/output/accelio) - find_library(RDMA_LIB_EVENT NAMES event PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_library(RDMA_LIB_EVENT_CORE NAMES event_core PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_library(RDMA_LIB_EVENT_EXTRA NAMES event_extra PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_library(RDMA_LIB_EVENT_PTHREADS NAMES event_pthreads PATHS ${RDMA_ROOT}/thirdparty/output/libevent) - find_library(RDMA_LIB_NUMA NAMES numa PATHS ${RDMA_ROOT}/thirdparty/output/libnuma) - - if( - RDMA_INC_SXISOCK AND - RDMA_INC_XIO AND - RDMA_INC_EVENT AND - RDMA_INC_NUMA AND - RDMA_LIB_SXISOCK AND - RDMA_LIB_XIO AND - RDMA_LIB_EVENT AND - RDMA_LIB_EVENT_CORE AND - RDMA_LIB_EVENT_EXTRA AND - RDMA_LIB_EVENT_PTHREADS AND - RDMA_LIB_NUMA - ) - - set(RDMA_INC_DIR - ${RDMA_INC_SXISOCK} - ${RDMA_INC_XIO} - ${RDMA_INC_EVENT} - ${RDMA_INC_NUMA}) - set(RDMA_LIBS - ${RDMA_LIB_SXISOCK} - ${RDMA_LIB_XIO} - ${RDMA_LIB_EVENT} - ${RDMA_LIB_EVENT_CORE} - ${RDMA_LIB_EVENT_EXTRA} - ${RDMA_LIB_EVENT_PTHREADS} - ${RDMA_LIB_NUMA} - ) - set(RDMA_LD_FLAGS "-L./librdma -libverbs -lrdmacm -Xlinker -rpath ./librdma") - include_directories("${RDMA_INC_DIR}") - else() - #if this module is not called, RDMA_INC_DIR RDMA_LIBS will be null, so top module always refer this variable - message(FATAL_ERROR, "RDMA libraries are not found, try to set RDMA_ROOT or check all related libraries.") - endif() -else(WITH_RDMA) - set(RDMA_LIBS "") - set(RDMA_LD_FLAGS "") - add_definitions(-DPADDLE_DISABLE_RDMA) -endif(WITH_RDMA) diff --git a/cmake/tensorrt.cmake b/cmake/tensorrt.cmake index 3dc7171551bfb7..891ff222633741 100644 --- a/cmake/tensorrt.cmake +++ b/cmake/tensorrt.cmake @@ -33,6 +33,5 @@ if(TENSORRT_FOUND) message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. " "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ") include_directories(${TENSORRT_INCLUDE_DIR}) - list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY}) add_definitions(-DPADDLE_WITH_TENSORRT) endif() diff --git a/paddle/contrib/float16/run_float16_demo.sh b/paddle/contrib/float16/run_float16_demo.sh index 031225a85dabb2..34cb7a12db1719 100755 --- a/paddle/contrib/float16/run_float16_demo.sh +++ b/paddle/contrib/float16/run_float16_demo.sh @@ -14,9 +14,7 @@ cmake .. -DWITH_AVX=OFF \ -DWITH_MKL=OFF \ -DWITH_GPU=ON \ -DWITH_TESTING=ON \ - -DWITH_TIMER=ON \ -DWITH_PROFILER=ON \ - -DWITH_FLUID_ONLY=ON make -j `nproc` pip install -U "$WHEEL_PATH/$(ls $WHEEL_PATH)" diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index df961be9115375..f24cf96cce30bf 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -261,7 +261,7 @@ paddle.fluid.layers.increment ArgSpec(args=['x', 'value', 'in_place'], varargs=N paddle.fluid.layers.array_write ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.create_array ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.less_than ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords='ignored', defaults=(None, None)) -paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords='ignored', defaults=(None,)) +paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.array_read ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.array_length ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.IfElse.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,)) @@ -427,7 +427,7 @@ paddle.fluid.optimizer.MomentumOptimizer.__init__ ArgSpec(args=['self', 'learnin paddle.fluid.optimizer.MomentumOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None) paddle.fluid.optimizer.MomentumOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.optimizer.MomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) -paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1e-06, None, None)) +paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'epsilon', 'regularization', 'name', 'initial_accumulator_value'], varargs=None, keywords=None, defaults=(1e-06, None, None, 0.0)) paddle.fluid.optimizer.AdagradOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None) paddle.fluid.optimizer.AdagradOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.optimizer.AdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) @@ -473,11 +473,11 @@ paddle.fluid.LoDTensor.has_valid_recursive_sequence_lengths has_valid_recursive_ paddle.fluid.LoDTensor.lod lod(self: paddle.fluid.core.LoDTensor) -> List[List[int]] paddle.fluid.LoDTensor.recursive_sequence_lengths recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor) -> List[List[int]] paddle.fluid.LoDTensor.set 1. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CPUPlace) -> None 2. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CPUPlace) -> None 3. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CPUPlace) -> None 4. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CPUPlace) -> None 5. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CPUPlace) -> None 6. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CPUPlace) -> None 7. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CPUPlace) -> None 8. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CPUPlace) -> None 9. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPlace) -> None 10. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPlace) -> None 11. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPlace) -> None 12. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPlace) -> None 13. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPlace) -> None 14. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPlace) -> None 15. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPlace) -> None 16. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPlace) -> None 17. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float32], arg1: paddle::platform::CUDAPinnedPlace) -> None 18. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int32], arg1: paddle::platform::CUDAPinnedPlace) -> None 19. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[float64], arg1: paddle::platform::CUDAPinnedPlace) -> None 20. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int64], arg1: paddle::platform::CUDAPinnedPlace) -> None 21. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[bool], arg1: paddle::platform::CUDAPinnedPlace) -> None 22. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint16], arg1: paddle::platform::CUDAPinnedPlace) -> None 23. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[uint8], arg1: paddle::platform::CUDAPinnedPlace) -> None 24. set(self: paddle.fluid.core.Tensor, arg0: numpy.ndarray[int8], arg1: paddle::platform::CUDAPinnedPlace) -> None -paddle.fluid.LoDTensor.set_lod set_lod(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None -paddle.fluid.LoDTensor.set_recursive_sequence_lengths set_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor, arg0: List[List[int]]) -> None +paddle.fluid.LoDTensor.set_lod set_lod(self: paddle.fluid.core.LoDTensor, lod: List[List[int]]) -> None +paddle.fluid.LoDTensor.set_recursive_sequence_lengths set_recursive_sequence_lengths(self: paddle.fluid.core.LoDTensor, recursive_sequence_lengths: List[List[int]]) -> None paddle.fluid.LoDTensor.shape shape(self: paddle.fluid.core.Tensor) -> List[int] paddle.fluid.LoDTensorArray.__init__ __init__(self: paddle.fluid.core.LoDTensorArray) -> None -paddle.fluid.LoDTensorArray.append append(self: paddle.fluid.core.LoDTensorArray, arg0: paddle.fluid.core.LoDTensor) -> None +paddle.fluid.LoDTensorArray.append append(self: paddle.fluid.core.LoDTensorArray, tensor: paddle.fluid.core.LoDTensor) -> None paddle.fluid.CPUPlace.__init__ __init__(self: paddle.fluid.core.CPUPlace) -> None paddle.fluid.CUDAPlace.__init__ __init__(self: paddle.fluid.core.CUDAPlace, arg0: int) -> None paddle.fluid.CUDAPinnedPlace.__init__ __init__(self: paddle.fluid.core.CUDAPinnedPlace) -> None diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index e88084424baf7e..dc308fd2592bb1 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -50,7 +50,12 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_ cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope) -cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper) +if(WITH_GPU) +cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info) +else() +cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info) +endif() + cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass) cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info) cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper) diff --git a/paddle/fluid/framework/details/all_reduce_deps_pass.cc b/paddle/fluid/framework/details/all_reduce_deps_pass.cc index b7d6edd389d8e4..2e20c436dfdb61 100644 --- a/paddle/fluid/framework/details/all_reduce_deps_pass.cc +++ b/paddle/fluid/framework/details/all_reduce_deps_pass.cc @@ -30,8 +30,6 @@ namespace paddle { namespace framework { namespace details { -static constexpr char kAllOpDescs[] = "all_op_descs"; - VarHandle* GetValidInput(const OpHandleBase* a) { for (auto p : a->Inputs()) { VarHandle* b = dynamic_cast(p); diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.cc b/paddle/fluid/framework/details/all_reduce_op_handle.cc index dd77f7099f581a..c1f9c2b60c9153 100644 --- a/paddle/fluid/framework/details/all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc @@ -53,7 +53,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, #endif void AllReduceOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); + platform::RecordEvent record_event(Name()); WaitInputVarGenerated(); auto in_var_handles = DynamicCast(this->Inputs()); diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index 89d626edddfee3..fdff83b92819b3 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -22,7 +22,7 @@ namespace framework { namespace details { void BroadcastOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + platform::RecordEvent record_event(Name()); if (places_.size() == 1) return; @@ -30,7 +30,7 @@ void BroadcastOpHandle::RunImpl() { VarHandle *in_var_handle; { auto in_var_handles = DynamicCast(inputs_); - PADDLE_ENFORCE_EQ(in_var_handles.size(), 1, + PADDLE_ENFORCE_EQ(in_var_handles.size(), 1UL, "The number of input should be one."); in_var_handle = in_var_handles[0]; } diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index f8030c53f72bc8..8c6c9f35e84f4f 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -34,9 +34,11 @@ namespace details { static inline bool SeqOnlyAllReduceOps(const BuildStrategy &strategy) { // Should fix the allreduce op order if scheduling // them in multiple threads or processes to avoid hang. + // NOTE: ParallelGraph would execute this pass on each graph, so + // don't need to append it here. return (!strategy.enable_sequential_execution_ && - strategy.num_trainers_ > 1) || - strategy.enable_parallel_graph_; + strategy.num_trainers_ > 1) && + !strategy.enable_parallel_graph_; } class ParallelExecutorPassBuilder : public ir::PassBuilder { @@ -240,7 +242,9 @@ std::unique_ptr BuildStrategy::Apply( continue; } } + VLOG(3) << "Start Apply Pass " << pass->Type(); graph = pass->Apply(std::move(graph)); + VLOG(3) << "Finish Apply Pass " << pass->Type(); } return graph; } diff --git a/paddle/fluid/framework/details/data_balance_op_handle.cc b/paddle/fluid/framework/details/data_balance_op_handle.cc index 48dcc52623369f..c9b52b68205ade 100644 --- a/paddle/fluid/framework/details/data_balance_op_handle.cc +++ b/paddle/fluid/framework/details/data_balance_op_handle.cc @@ -86,7 +86,7 @@ std::vector> DataBalanceOpHandle::GetBalancePlan( } void DataBalanceOpHandle::RunImpl() { - PADDLE_ENFORCE_GT(places_.size(), 1, + PADDLE_ENFORCE_GT(places_.size(), 1UL, "Data balance can only be enabled when the number of " "places to run larger than 1."); auto in_var_handles = DynamicCast(this->Inputs()); diff --git a/paddle/fluid/framework/details/fuse_vars_op_handle.cc b/paddle/fluid/framework/details/fuse_vars_op_handle.cc index d65b0920698748..14292c0a5d06aa 100644 --- a/paddle/fluid/framework/details/fuse_vars_op_handle.cc +++ b/paddle/fluid/framework/details/fuse_vars_op_handle.cc @@ -23,7 +23,7 @@ void FuseVarsOpHandle::RunImpl() { auto in_var_handles = DynamicCast(this->Inputs()); auto out_var_handles = DynamicCast(this->Outputs()); - PADDLE_ENFORCE_EQ(in_var_handles.size(), 0); + PADDLE_ENFORCE_EQ(in_var_handles.size(), 0UL); PADDLE_ENFORCE_EQ(out_var_handles.size() - 1, inputs_numel_.size(), ""); auto scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle.cc b/paddle/fluid/framework/details/fused_broadcast_op_handle.cc index 51dfa2d0711f49..f48561ea32e6a3 100644 --- a/paddle/fluid/framework/details/fused_broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle.cc @@ -22,7 +22,7 @@ namespace framework { namespace details { void FusedBroadcastOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + platform::RecordEvent record_event(Name()); if (places_.size() == 1UL) return; diff --git a/paddle/fluid/framework/details/inplace_op_pass.cc b/paddle/fluid/framework/details/inplace_op_pass.cc index b0c5968499be3a..c91fc81b2defc9 100644 --- a/paddle/fluid/framework/details/inplace_op_pass.cc +++ b/paddle/fluid/framework/details/inplace_op_pass.cc @@ -49,7 +49,7 @@ DEFINE_bool( "If this option turns on, only these op in whitelist can be inplaced." "If it turns off, all of the running op can be candidate of inplaced op." "Such as scale, elementwise_add" - "By default, it's turned on"); + "By default, it's turned off"); DECLARE_string(memory_optimize_debug); diff --git a/paddle/fluid/framework/details/memory_optimize_helper.cc b/paddle/fluid/framework/details/memory_optimize_helper.cc index 6345ba335997ec..db4e805bb692ee 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper.cc +++ b/paddle/fluid/framework/details/memory_optimize_helper.cc @@ -13,13 +13,19 @@ // limitations under the License. #include "paddle/fluid/framework/details/memory_optimize_helper.h" +#include #include #include -#include +#include #include #include #include #include "paddle/fluid/framework/var_desc.h" +#include "paddle/fluid/platform/cpu_info.h" + +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/gpu_info.h" +#endif // PADDLE_WITH_CUDA namespace paddle { namespace framework { @@ -123,7 +129,13 @@ size_t NodeSize(const VarDesc& node) { } size_t NodeSize(ir::Node* n) { - auto* desc = FindVarDescInBlock(n); + VarDesc* desc = nullptr; + // some op do not have block pointer + if (n->inputs[0]->Op() != nullptr) { + desc = FindVarDescInBlock(n); + } else { + desc = n->Var(); + } return NodeSize(*desc); } @@ -166,6 +178,11 @@ struct NodeComparator { bool operator()(ir::Node* lhs, ir::Node* rhs) const { auto* lhs_desc = FindVarDescInBlock(lhs); auto* rhs_desc = FindVarDescInBlock(rhs); + // match data type + if (lhs_desc->GetDataType() != rhs_desc->GetDataType()) { + return false; + } + // match shape auto lhs_shape = lhs_desc->GetShape(); auto rhs_shape = rhs_desc->GetShape(); if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) || @@ -230,6 +247,27 @@ ir::Node* OrderedSet::FindBestFitNode(ir::Node* var) const { return found_node; } +ir::Node* OrderedSet::FindNextBestFitNode(ir::Node* var, ir::Node* prev) const { + ir::Node* found_node = nullptr; + NodeComparator functor; + auto it = + std::find_if(nodes_.begin(), nodes_.end(), [&](const NodeVector& v) { + if (v.front() == prev) + return true; + else + return false; + }); + PADDLE_ENFORCE(it != nodes_.end(), "Not found previous in node list!"); + for (it = std::next(it); it != nodes_.end(); ++it) { + auto& candidate = it->front(); + if (functor(var, candidate)) { + found_node = candidate; + break; + } + } + return found_node; +} + bool OrderedSet::Has(ir::Node* var) const { if (mark_table_.count(var->Name())) { auto& node_in_samename = mark_table_.at(var->Name()); @@ -241,10 +279,15 @@ bool OrderedSet::Has(ir::Node* var) const { return false; } +void OrderedSet::Erase(const std::string& var) { + PADDLE_ENFORCE(mark_table_.count(var)); + nodes_.erase(mark_table_[var]); + mark_table_.erase(var); +} + void OrderedSet::Erase(ir::Node* var) { - PADDLE_ENFORCE(mark_table_.count(var->Name())); - nodes_.erase(mark_table_[var->Name()]); - mark_table_.erase(var->Name()); + PADDLE_ENFORCE(var != nullptr); + Erase(var->Name()); } std::string OrderedSet::ToString() const { @@ -274,14 +317,35 @@ bool NodeCanReused(ir::Node* node) { return flag; } +int MinChunkSize() { + int size{0}; +#ifdef PADDLE_WITH_CUDA + size = platform::GpuMinChunkSize(); +#else + size = platform::CpuMinChunkSize(); +#endif // PADDLE_WITH_CUDA + return size; +} + bool NodeCanReused(const VarDesc& node) { auto type = node.GetType(); + // only these types holds bulk of gpu memory if (!(type == proto::VarType::LOD_TENSOR || type == proto::VarType::SELECTED_ROWS || type == proto::VarType::LOD_TENSOR_ARRAY)) { return false; } - if (node.Persistable() || node.GetShape().empty()) { + // persistable variable is parameter + if (node.Persistable()) { + return false; + } + // shape < min_chunk_size is meaningless. + // further more, fetched loss always has size = 1 + // which should not be reused. + auto shape = node.GetShape(); + int size = std::abs( + std::accumulate(shape.begin(), shape.end(), 1, std::multiplies())); + if (shape.empty() || size < MinChunkSize()) { return false; } // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad @@ -461,7 +525,9 @@ ir::Node* ControlFlowGraph::GetNodeByName(const std::string& name, for (auto* node : ops_) { if (node == op) break; for (auto& output : node->outputs) { - if (output->Name() == name) { + PADDLE_ENFORCE((output != nullptr && output->IsVar()), + "Output is empty!"); + if (output->Var() && output->Name() == name) { found_node = output; } } diff --git a/paddle/fluid/framework/details/memory_optimize_helper.h b/paddle/fluid/framework/details/memory_optimize_helper.h index 0bfaf827fea840..377367faf3c529 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper.h +++ b/paddle/fluid/framework/details/memory_optimize_helper.h @@ -29,8 +29,6 @@ namespace paddle { namespace framework { namespace details { -constexpr char kAllOpDescs[] = "all_op_descs"; - std::vector SortOpLikeDescOrder(const ir::Graph& graph); // NOTE(dzh): A ordered set for node reuse in memory optimize. @@ -55,6 +53,7 @@ class OrderedSet { void Insert(ir::Node* var); void Erase(ir::Node* var); + void Erase(const std::string& var); bool Has(ir::Node* var) const; void Clear() { mark_table_.clear(); @@ -62,6 +61,7 @@ class OrderedSet { } // find the bestfit shape node block with var. ir::Node* FindBestFitNode(ir::Node* var) const; + ir::Node* FindNextBestFitNode(ir::Node* var, ir::Node* prev) const; // map store non-const iterator, can not promise const int GetNodeIndexInPool(ir::Node* var); // pool all node to string diff --git a/paddle/fluid/framework/details/memory_optimize_helper_test.cc b/paddle/fluid/framework/details/memory_optimize_helper_test.cc index 5c13dda9e54910..3cfe297a73cf41 100644 --- a/paddle/fluid/framework/details/memory_optimize_helper_test.cc +++ b/paddle/fluid/framework/details/memory_optimize_helper_test.cc @@ -107,6 +107,52 @@ TEST(OrderedSet, Normal) { ASSERT_EQ(pool.GetNodeIndexInPool(cache), 5); // match 4:[5,2] } } + +TEST(OrderedSet, FindBestFitNode) { + OrderedSet pool; + std::vector> nodes; + ProgramDesc prog; + BlockDesc* block_desc = prog.MutableBlock(0); + auto* op_desc = block_desc->AppendOp(); + op_desc->SetType("dummy"); + std::unique_ptr op = ir::CreateNodeForTest(op_desc); + + { + auto desc = block_desc->Var("a"); + desc->SetShape({128, 128}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + { + auto desc = block_desc->Var("b"); + desc->SetShape({128, 129}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + { + auto desc = block_desc->Var("c"); + desc->SetShape({128, 128}); + std::unique_ptr node = ir::CreateNodeForTest(desc); + node->inputs.emplace_back(op.get()); + nodes.emplace_back(std::move(node)); + } + + for (auto& node : nodes) { + pool.Insert(node.get()); + } + + // FindNextBestFitNode + auto* n = nodes[0].get(); + auto* cache = pool.FindBestFitNode(n); + PADDLE_ENFORCE(cache->Name() == "a"); + cache = pool.FindNextBestFitNode(n, cache); + PADDLE_ENFORCE(cache->Name() == "c"); + cache = pool.FindNextBestFitNode(n, cache); + PADDLE_ENFORCE(cache->Name() == "b"); +} + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/memory_optimize_pass.cc b/paddle/fluid/framework/details/memory_optimize_pass.cc index 41e4a834df0aba..fd02bc4697e72c 100644 --- a/paddle/fluid/framework/details/memory_optimize_pass.cc +++ b/paddle/fluid/framework/details/memory_optimize_pass.cc @@ -69,55 +69,59 @@ std::unique_ptr MemoryOptimizePass::ApplyImpl( } for (auto& var : op->outputs) { - if (!NodeCanReused(var) || cfg_->Use(op).count(var->Name()) == 0 || - skip_set_.count(var->Name())) + if (var->IsVar() && !var->IsCtrlVar() && skip_set_.count(var->Name())) { + VLOG(3) << "Skip set contains variable of " << var->Name() + << "disable reuse on it. skipped"; continue; - ir::Node* cache = pool_.FindBestFitNode(var); - - if (var->Name() == FLAGS_memory_optimize_debug) { - VLOG(3) << "start match var " << DebugString(var) << " of op " - << op->Name(); - VLOG(3) << pool_.ToString(); - VLOG(3) << "matched in pool : " - << ((cache == nullptr) ? "False" : "True"); } + if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) { + ir::Node* cache = pool_.FindBestFitNode(var); + while (cache != nullptr && var->Name() == cache->Name()) { + VLOG(3) << "The same cache variable is cascade reused. " + << cache->Name() << " is re-filled to the pool after " + << "the reused op is finished. Current op can not " + << "replace it again. Skip this candidate."; + cache = pool_.FindNextBestFitNode(var, cache); + } + if (var->Name() == FLAGS_memory_optimize_debug) { + VLOG(3) << "start match var " << DebugString(var) << " of op " + << op->Name(); + VLOG(3) << pool_.ToString(); + VLOG(3) << "matched in pool : " + << ((cache == nullptr) ? "False" : "True"); + } - if (cache == nullptr) continue; - if (var->Name() == cache->Name()) { - VLOG(3) << "The same cache variable is cascade reused." << var->Name() - << " is re-filled to the pool after" - << "the reused op is finished. Current op can not " - << "replace it again. Skip this candidate."; - continue; - - int node_idx_in_pool = pool_.GetNodeIndexInPool(cache); - VLOG(3) << string::Sprintf( - "!!! %s, %s => %s, cache idx %d, pool size %d", - std::to_string(reuse_id++), DebugString(var), DebugString(cache), - node_idx_in_pool, static_cast(pool_.size())); - - // update CFG Graph on the fly. - // reused var maybe re-fill into the pool - cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx); - // NOTE(dzhwinter): we need to both update the ProgramDesc - // and IR Graph. because op_desc/var_desc is used in CreateOp, - // CreateVar when running happens. But IR Graph - // define the dependence relationship between nodes. - RenameVarInGraphDesc(var->Name(), cache->Name(), idx); - RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get()); - - pool_.Erase(cache); - } + if (cache != nullptr) { + int node_idx_in_pool = pool_.GetNodeIndexInPool(cache); + VLOG(3) << string::Sprintf( + "!!! %s, %s => %s, cache idx %d, pool size %d", + std::to_string(reuse_id++), DebugString(var), DebugString(cache), + node_idx_in_pool, static_cast(pool_.size())); + // NOTE(dzhwinter): update the ProgramDesc/IR Graph + // and the CFG Graph on the fly. + // + // IR Graph define the dependence relationship between nodes. + // + // ProgramDesc defines the input/output vars. Its used in + // CreateOp, CreateVar when running happens. + // + // CFG Graph store the liveness information, when reuse happens + // we also need to update the variable liveness. + const std::string var_name = var->Name(); + const std::string cache_name = cache->Name(); - // fill the pool - std::unordered_set unlived_vars; - for (auto var : cfg_->LiveIn(op)) { - if (cfg_->LiveOut(op).count(var) == 0) { - unlived_vars.emplace(var); + cfg_->RenameVarInCFGGraph(var_name, cache_name, idx); + RenameVarInGraphDesc(var_name, cache_name, idx); + RenameVarInGraphNode(var_name, cache_name, idx, graph.get()); + pool_.Erase(cache_name); } } - for (auto var : unlived_vars) { + } + // fill the pool + for (auto var : cfg_->LiveIn(op)) { + if (cfg_->LiveOut(op).count(var) == 0) { ir::Node* var_node = cfg_->GetNodeByName(var, op); + if (var_node == nullptr || var_node->IsCtrlVar()) continue; if (NodeCanReused(var_node) && !pool_.Has(var_node)) { pool_.Insert(var_node); } @@ -190,7 +194,8 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const { // effect. Because it is a single op in graph. No need to // update the ir nodes. sub_op_desc->Rename(var->Name(), cache->Name()); - if (sub_op_desc->Block()->HasVar(var->Name())) { + if (sub_op_desc->Block() != nullptr && + sub_op_desc->Block()->HasVar(var->Name())) { sub_op_desc->Block()->RemoveVar(var->Name()); } } @@ -231,7 +236,13 @@ void MemoryOptimizePass::RenameVarInGraphDesc(const std::string& var, auto* op_desc = op->Op(); op_desc->RenameInput(var, cache_var); op_desc->RenameOutput(var, cache_var); - if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var); + if (op_desc->Block() != nullptr) { + op_desc->Block()->RemoveVar(var); + } else { + LOG(WARNING) << "op " << op->Name() << " not know its block." + << "Is the op_desc created without block pointer? " + << "Can not find " << var << " in Block(0)"; + } op_desc->Flush(); } } @@ -273,8 +284,7 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, // redirect the input to the latest version of cache_var for (auto* node : op->inputs) { if (node->Name() == var) { - ir::Node* cache_node = graph->CreateVarNode(var_desc.get()); - var_nodes_[cache_var].emplace_back(cache_node); + ir::Node* cache_node = var_nodes_[cache_var].back(); // swap node to cache_node cache_node->outputs.insert(cache_node->outputs.end(), @@ -283,11 +293,15 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, auto* prev_op = node->inputs[0]; std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node, cache_node); - cache_node->inputs.emplace_back(prev_op); for (auto* next_op : node->outputs) { std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, cache_node); } + + // erase unused node + auto& nodes = var_nodes_.at(var); + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + graph->RemoveNode(node); } } @@ -307,15 +321,14 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var, std::replace(next_op->inputs.begin(), next_op->inputs.end(), node, cache_node); } + + // erase unused node + auto& nodes = var_nodes_.at(var); + nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end()); + graph->RemoveNode(node); } } } - - // release node of unused var in graph - for (auto* node : var_nodes_[var]) { - graph->RemoveNode(node); - } - var_nodes_.at(var).clear(); } } // namespace details diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc index 75f922d2cca685..7d1e63f3682bca 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc @@ -392,20 +392,32 @@ void MultiDevSSAGraphBuilderBase::CreateComputationalOp(ir::Graph *result, void MultiDevSSAGraphBuilderBase::CreateAllReduceOp( ir::Graph *result, const std::string &og) const { + OpHandleBase *op_handle = nullptr; + + auto append_allreduce_op = [&]( + const std::vector &scopes, + const std::vector &places) -> OpHandleBase * { #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) - result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( - result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), - local_scopes_, places_, nccl_ctxs_)); + result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( + result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), + scopes, places, nccl_ctxs_)); #else - result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( - result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), - local_scopes_, places_)); + result->Get(kGraphOps).emplace_back(new AllReduceOpHandle( + result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), + scopes, places)); #endif - auto *op_handle = result->Get(kGraphOps).back(); + return result->Get(kGraphOps).back(); + }; + + if (!strategy_.enable_parallel_graph_) + op_handle = append_allreduce_op(local_scopes_, places_); for (size_t i = 0; i < places_.size(); ++i) { - auto &p = places_[i]; - SetCommunicationContext(op_handle, p); + if (strategy_.enable_parallel_graph_) { + op_handle = append_allreduce_op({local_scopes_[i]}, {places_[i]}); + } + + SetCommunicationContext(op_handle, places_[i]); auto &vars = result->Get(kGraphVars)[i][og]; PADDLE_ENFORCE(!vars.empty()); auto &prev_grad = vars.back(); @@ -413,7 +425,7 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp( auto var = new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable), - vars.size(), i, og, p); + vars.size(), i, og, places_[i]); vars.emplace_back(var); op_handle->AddOutput(var); } diff --git a/paddle/fluid/framework/details/multi_devices_helper.h b/paddle/fluid/framework/details/multi_devices_helper.h index 1a2b75fbc0c289..9afbb91005c9c3 100644 --- a/paddle/fluid/framework/details/multi_devices_helper.h +++ b/paddle/fluid/framework/details/multi_devices_helper.h @@ -36,13 +36,14 @@ namespace details { // map from variable name to variables. The variables, who have the same name, // will have a differsent version. The offset in the // `std::vector` is the version of varaibles. -typedef std::vector>> +typedef std::vector>> GraphVars; const char kGraphVars[] = "vars"; // aux variables to represent dependency. Useful to resolve data hazard. -typedef std::unordered_set GraphDepVars; +typedef std::unordered_set GraphDepVars; const char kGraphDepVars[] = "dep_vars"; + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index b1a82e8771b92f..e0aa352e95bc36 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -70,6 +70,9 @@ class OpHandleBase { auto it = dev_ctxes_.find(place); return it != dev_ctxes_.end() ? it->second : nullptr; } + const std::map &DeviceContext() { + return dev_ctxes_; + } void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) { dev_ctxes_[place] = ctx_; diff --git a/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc b/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc index e8deb5bfc6c013..4c8f69c68ce17d 100644 --- a/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/parallel_ssa_graph_executor.cc @@ -13,22 +13,92 @@ // limitations under the License. #include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h" +#include "paddle/fluid/framework/ir/graph_helper.h" namespace paddle { namespace framework { namespace details { +std::vector> +ParallelSSAGraphExecutor::SeparateMultiDevicesGraph( + std::unique_ptr &&graph) { + std::vector> graphs; + graphs.reserve(places_.size()); + for (size_t i = 0; i < places_.size(); ++i) { + ProgramDesc empty; + graphs.emplace_back(std::unique_ptr(new ir::Graph(empty))); + auto &g = graphs.back(); + g->Set(kGraphVars, new GraphVars(1UL)); + g->Set(kGraphDepVars, new GraphDepVars); + } + auto op_handles = ir::FilterByNodeWrapper(*graph); + + for (auto &op : op_handles) { + auto &dev_ctx = op->DeviceContext(); + auto &p = dev_ctx.begin()->first; + int dev_id = boost::get(p).device; + auto &dev_dummys = graphs[dev_id]->Get(kGraphDepVars); + graphs[dev_id]->AddNode(graph->RemoveNode(op->Node()).release()); + + for (auto &var : op->Inputs()) { + auto dummy_ptr = dynamic_cast(var); + if (dummy_ptr) { + dev_dummys.insert(var); + if (graph->Nodes().count(var->Node())) + graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release()); + } + } + for (auto &var : op->Outputs()) { + auto dummy_ptr = dynamic_cast(var); + if (dummy_ptr) { + dev_dummys.insert(var); + if (graph->Nodes().count(var->Node())) + graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release()); + } + } + } + + for (size_t dev_id = 0; dev_id < places_.size(); ++dev_id) { + auto &dev_vars = graphs[dev_id]->Get(kGraphVars)[0]; + auto &origin_vars = graph->Get(kGraphVars)[dev_id]; + for (auto &name_pair : origin_vars) { + dev_vars.emplace(name_pair.first, name_pair.second); + for (auto &version_pair : name_pair.second) { + if (graph->Nodes().count(version_pair->Node())) { + graphs[dev_id]->AddNode( + graph->RemoveNode(version_pair->Node()).release()); + } + } + } + } + + return graphs; +} + ParallelSSAGraphExecutor::ParallelSSAGraphExecutor( const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::vector> &&graphs) + const framework::ProgramDesc &main_prog, std::unique_ptr &&graph) : strategy_(std::move(strategy)), local_scopes_(std::move(local_scopes)), pool_(places.size() >= 2 ? new ::ThreadPool(places.size()) : nullptr), places_(std::move(places)), - graphs_(std::move(graphs)) { + main_prog_(main_prog), + // TODO(Yancey1989): Copying graphs is not safely since it deleted the + // attrs. + graphs_(SeparateMultiDevicesGraph(std::move(graph))) { PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size()); + auto seq_allreduce_pass = + ir::PassRegistry::Instance().Get("all_reduce_deps_pass"); + seq_allreduce_pass->Erase(details::kAllOpDescs); + seq_allreduce_pass->Set>( + details::kAllOpDescs, + new std::vector(main_prog_.Block(0).AllOps())); + for (size_t i = 0; i < graphs_.size(); ++i) { + graphs_[i] = seq_allreduce_pass->Apply(std::move(graphs_[i])); + } + // set the correct size of thread pool to each device. strategy_.num_threads_ = strategy_.num_threads_ < places_.size() ? 1UL @@ -37,7 +107,7 @@ ParallelSSAGraphExecutor::ParallelSSAGraphExecutor( << " to run the operators of the graph on each device."; for (size_t i = 0; i < places.size(); ++i) { executors_.emplace_back(new details::ThreadedSSAGraphExecutor( - strategy_, {local_scopes_[i]}, {places_[i]}, std::move(graphs_[i]))); + strategy_, local_scopes_, {places_[i]}, std::move(graphs_.at(i)))); } } diff --git a/paddle/fluid/framework/details/parallel_ssa_graph_executor.h b/paddle/fluid/framework/details/parallel_ssa_graph_executor.h index c00c5bc2d1b4b7..1c35d45fdd356a 100644 --- a/paddle/fluid/framework/details/parallel_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/parallel_ssa_graph_executor.h @@ -18,7 +18,9 @@ #include #include "ThreadPool.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" +#include "paddle/fluid/framework/ir/graph.h" namespace paddle { namespace framework { @@ -29,17 +31,23 @@ class ParallelSSAGraphExecutor : public SSAGraphExecutor { ParallelSSAGraphExecutor(const ExecutionStrategy &strategy, const std::vector &local_scopes, const std::vector &places, - std::vector> &&graphs); + const framework::ProgramDesc &main_prog, + std::unique_ptr &&graph); ~ParallelSSAGraphExecutor() final = default; + const ir::Graph &Graph() const override { return *graphs_[0]; } FeedFetchList Run(const std::vector &fetch_tensors) override; private: + std::vector> SeparateMultiDevicesGraph( + std::unique_ptr &&graph); + ExecutionStrategy strategy_; std::vector local_scopes_; std::unique_ptr<::ThreadPool> pool_{nullptr}; std::vector places_; + framework::ProgramDesc main_prog_; std::vector> graphs_; std::vector> executors_; diff --git a/paddle/fluid/framework/details/reduce_op_handle.cc b/paddle/fluid/framework/details/reduce_op_handle.cc index ee4c8a6ecf77e5..4e2477c205db59 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.cc +++ b/paddle/fluid/framework/details/reduce_op_handle.cc @@ -139,7 +139,7 @@ void ReduceOpHandle::GatherSelectedRows( #endif void ReduceOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); + platform::RecordEvent record_event(Name()); if (places_.size() == 1) return; // the input and output may have dummy var. @@ -153,7 +153,7 @@ void ReduceOpHandle::RunImpl() { { auto out_var_handles = DynamicCast(outputs_); - PADDLE_ENFORCE_EQ(out_var_handles.size(), 1, + PADDLE_ENFORCE_EQ(out_var_handles.size(), 1UL, "The number of output should be one."); out_var_handle = out_var_handles.front(); } diff --git a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc index 91e4f9adb41897..7b13112986f9ad 100644 --- a/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc @@ -63,7 +63,7 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run( eptr = std::current_exception(); } - platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr); + platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun"); ++drop_scope_counter_; bool stream_end = false; diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 677a2937945b03..72acc337b7cc48 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -37,7 +37,7 @@ ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor( FeedFetchList ThreadedSSAGraphExecutor::Run( const std::vector &fetch_tensors) { std::unique_ptr event( - new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr)); + new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare")); std::unordered_map pending_ops; std::unordered_set pending_vars; auto ready_vars = std::make_shared>(); @@ -219,7 +219,7 @@ void ThreadedSSAGraphExecutor::RunOp( VLOG(10) << op << " " << op->Name() << " Done "; running_ops_--; ready_var_q->Extend(op->Outputs()); - VLOG(10) << op << " " << op->Name() << "Signal posted"; + VLOG(10) << op << " " << op->Name() << " Signal posted"; } catch (...) { exception_holder_.Catch(std::current_exception()); } diff --git a/paddle/fluid/framework/inplace_op_inference_test.cc b/paddle/fluid/framework/inplace_op_inference_test.cc index 3e4d715c6f0894..bf9d1dcd380cdf 100644 --- a/paddle/fluid/framework/inplace_op_inference_test.cc +++ b/paddle/fluid/framework/inplace_op_inference_test.cc @@ -179,11 +179,11 @@ TEST(InferInplace, SingleOpInplaceInToOut) { op->SetOutput("Out", {"test2_out"}); prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64}); + prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64, 128, 128}); prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_out"); - prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 128, 128}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -201,11 +201,11 @@ TEST(InferInplace, SingleGradOpInplaceInToOut) { op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"}); prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR); - prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16, 1024, 1024}); prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR); prog.MutableBlock(0)->Var("test2_out"); - prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -233,12 +233,12 @@ TEST(InferInplace, MultiOutInplaceInToOut) { prog.MutableBlock(0)->Var("o0"); prog.MutableBlock(0)->Var("y0"); prog.MutableBlock(0)->Var("z0"); - prog.MutableBlock(0)->Var("a0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("b0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("c0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("o0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("y0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("z0")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); @@ -267,12 +267,12 @@ TEST(InferInplace, MultiGradInplaceInToOut) { prog.MutableBlock(0)->Var("o0"); prog.MutableBlock(0)->Var("y0"); prog.MutableBlock(0)->Var("z0"); - prog.MutableBlock(0)->Var("a0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("b0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("c0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("o0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("y0")->SetShape({32, 16}); - prog.MutableBlock(0)->Var("z0")->SetShape({32, 16}); + prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024}); + prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024}); auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_; auto in_to_outs = infer_inplace(*op, op->Block()); diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc index 846a14e365e6bd..04765dd1440331 100644 --- a/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc @@ -169,7 +169,7 @@ std::unique_ptr ConvBNFusePass::ApplyImpl( if (has_bias && conv->Op()->Input("Bias").size() > 0) { // reuse existing conv bias node auto conv_bias_names = conv->Op()->Input("Bias"); - PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1); + PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1UL); auto* conv_bias_var = scope->FindVar(conv_bias_names[0]); auto* conv_bias_tensor = conv_bias_var->GetMutable(); PADDLE_ENFORCE_EQ(conv_bias_tensor->dims(), diff --git a/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc b/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc index 0d94008ea82d0e..fe844caed2e757 100644 --- a/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc +++ b/paddle/fluid/framework/ir/fuse_relu_depthwise_conv_pass.cc @@ -111,7 +111,7 @@ std::unique_ptr FuseReluDepthwiseConvPass::FuseReluDepthwiseConv( xg_var = subgraph.at(xg)->Var(); } - PADDLE_ENFORCE_EQ(layer_op->Input("Input").size(), 1); + PADDLE_ENFORCE_EQ(layer_op->Input("Input").size(), 1UL); PADDLE_ENFORCE_EQ(layer_op->Input("Input")[0], y_var->Name()); layer_op->SetInput("Input", {x_var->Name()}); subgraph.at(layer)->inputs.push_back(subgraph.at(x)); @@ -119,13 +119,13 @@ std::unique_ptr FuseReluDepthwiseConvPass::FuseReluDepthwiseConv( VLOG(4) << "replace " << y_var->Name() << " -> " << x_var->Name(); if (!only_forward) { - PADDLE_ENFORCE_EQ(layer_g_op->Input("Input").size(), 1); + PADDLE_ENFORCE_EQ(layer_g_op->Input("Input").size(), 1UL); PADDLE_ENFORCE_EQ(layer_g_op->Input("Input")[0], y_var->Name()); layer_g_op->SetInput("Input", {x_var->Name()}); subgraph.at(layer_g)->inputs.push_back(subgraph.at(x)); subgraph.at(x)->outputs.push_back(subgraph.at(layer_g)); - PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input")).size(), 1); + PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input")).size(), 1UL); PADDLE_ENFORCE_EQ(layer_g_op->Output(GradVarName("Input"))[0], yg_var->Name()); layer_g_op->SetOutput(GradVarName("Input"), {xg_var->Name()}); diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h index feb3330176490e..296f3b83961c13 100644 --- a/paddle/fluid/framework/ir/graph.h +++ b/paddle/fluid/framework/ir/graph.h @@ -26,6 +26,14 @@ limitations under the License. */ namespace paddle { namespace framework { + +namespace details { + +// This attr is not recommended, because the graph should not dependence +// the program once it is built. +constexpr char kAllOpDescs[] = "all_op_descs"; +} // namespace details + namespace ir { /* @@ -168,10 +176,13 @@ class Graph { return ret; } - void RemoveNode(ir::Node *node) { + std::unique_ptr RemoveNode(ir::Node *node) { PADDLE_ENFORCE(node_set_.find(node) != node_set_.end()); - node_set_.erase(node); + std::unique_ptr ret; + ret.reset(nodes_.at(node).release()); nodes_.erase(node); + node_set_.erase(node); + return ret; } // NOTE low performance, but simple and secure. @@ -184,13 +195,6 @@ class Graph { return nullptr; } - void ResolveHazard( - const std::map> &var_nodes); - - private: - std::map> InitFromProgram( - const ProgramDesc &program); - // This method takes ownership of `node`. ir::Node *AddNode(ir::Node *node) { PADDLE_ENFORCE(node_set_.find(node) == node_set_.end()); @@ -199,6 +203,13 @@ class Graph { return node; } + void ResolveHazard( + const std::map> &var_nodes); + + private: + std::map> InitFromProgram( + const ProgramDesc &program); + // NOTE: program_ shouldn't be exposed to user. const ProgramDesc program_; std::map attrs_; diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index 9ea0729e1f3339..c0c34d186b0081 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -38,7 +38,7 @@ size_t PDPattern::id_ = 0UL; PDNode *PDPattern::NewNode(const std::string &name) { if (!name.empty()) { - PADDLE_ENFORCE_EQ(node_map_.count(name), 0, + PADDLE_ENFORCE_EQ(node_map_.count(name), 0UL, "PDNode's name should be unique, get duplicate [%s]", name); } @@ -51,7 +51,7 @@ PDNode *PDPattern::NewNode(const std::string &name) { PDNode *PDPattern::NewNode(PDNode::teller_t &&teller, const std::string &name) { if (!name.empty()) { - PADDLE_ENFORCE_EQ(node_map_.count(name), 0, + PADDLE_ENFORCE_EQ(node_map_.count(name), 0UL, "PDNode's name should be unique, get duplicate [%s]", name); } diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index e15c838f4fbe44..9a0348871b0502 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -177,9 +177,7 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { // in concurrency scenerio. Here use an `if` to fix this issue. // Please not remove the `if`, ask @Superjomn if there are any concern. if (platform::IsProfileEnabled()) { - platform::DeviceContextPool& pool = - platform::DeviceContextPool::Instance(); - platform::RecordEvent record_event(Type(), pool.Get(place)); + platform::RecordEvent record_event(Type()); RunImpl(scope, place); } else { RunImpl(scope, place); diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index ff7ef0cce2f12f..56da5660095aff 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -21,6 +21,7 @@ limitations under the License. */ #include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/details/all_reduce_deps_pass.h" #include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h" #include "paddle/fluid/framework/details/multi_devices_helper.h" #include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h" @@ -193,7 +194,6 @@ ParallelExecutor::ParallelExecutor( member_->use_all_reduce_ = build_strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce; member_->nranks_ = build_strategy.num_trainers_ * places.size(); - if (!member_->use_all_reduce_) { PADDLE_ENFORCE(places.size() > 1, "If you set build_strategy.reduce with 'Reduce'," @@ -221,9 +221,10 @@ ParallelExecutor::ParallelExecutor( // choice the execution strategy. build_strategy.enable_parallel_graph_ = EnableParallelGraphExecution(main_program, exec_strategy, build_strategy); - - VLOG(1) << "Enable ParallelGraph Execution: " - << build_strategy.enable_parallel_graph_; + if (build_strategy.enable_parallel_graph_) + VLOG(0) << "The Executor would execute the graph by ParallelGraph " + "Execution which can get better performance," + << "you can force it off by env FLAGS_enable_parallel_graph=0"; if (member_->use_cuda_) { // Bcast Parameters to all GPUs @@ -257,60 +258,44 @@ ParallelExecutor::ParallelExecutor( // Step 2. Convert main_program to SSA form and dependency graph. Also, insert // ncclOp - std::vector> graphs; + std::unique_ptr graph; #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) - if (build_strategy.enable_parallel_graph_) { - for (size_t i = 0; i < member_->places_.size(); ++i) { - std::unique_ptr graph = build_strategy.Apply( - main_program, {member_->places_[i]}, loss_var_name, - {member_->local_scopes_[i]}, member_->nranks_, member_->use_cuda_, - member_->nccl_ctxs_.get()); - graphs.push_back(std::move(graph)); - } - } else { - std::unique_ptr graph = build_strategy.Apply( - main_program, member_->places_, loss_var_name, member_->local_scopes_, - member_->nranks_, member_->use_cuda_, member_->nccl_ctxs_.get()); - graphs.push_back(std::move(graph)); - } + graph = build_strategy.Apply(main_program, member_->places_, loss_var_name, + member_->local_scopes_, member_->nranks_, + member_->use_cuda_, member_->nccl_ctxs_.get()); #else - std::unique_ptr graph = build_strategy.Apply( - main_program, member_->places_, loss_var_name, member_->local_scopes_, - member_->nranks_, member_->use_cuda_); - graphs.push_back(std::move(graph)); + graph = build_strategy.Apply(main_program, member_->places_, loss_var_name, + member_->local_scopes_, member_->nranks_, + member_->use_cuda_); #endif auto max_memory_size = GetEagerDeletionThreshold(); VLOG(10) << "Eager Deletion Threshold " << static_cast(max_memory_size) / (1 << 30); if (max_memory_size >= 0) { - for (size_t i = 0; i < graphs.size(); ++i) { - graphs[i] = member_->PrepareGCAndRefCnts( - std::move(graphs[i]), static_cast(max_memory_size)); - } + graph = member_->PrepareGCAndRefCnts(std::move(graph), + static_cast(max_memory_size)); } // Step 3. Create vars in each scope. Passes may also create new vars. // skip control vars and empty vars std::vector var_infos; - for (auto &graph : graphs) { - for (auto &node : graph->Nodes()) { - if (node->IsVar() && !node->IsCtrlVar() && node->Var()) { - var_infos.emplace_back(); - var_infos.back().name_ = node->Var()->Name(); - var_infos.back().type_ = node->Var()->GetType(); - var_infos.back().persistable_ = node->Var()->Persistable(); - } + for (auto &node : graph->Nodes()) { + if (node->IsVar() && !node->IsCtrlVar() && node->Var()) { + var_infos.emplace_back(); + var_infos.back().name_ = node->Var()->Name(); + var_infos.back().type_ = node->Var()->GetType(); + var_infos.back().persistable_ = node->Var()->Persistable(); } } // If the loss_var_name is given, the number of graph should be only one. if (loss_var_name.size()) { - size_t graph_num = ir::GraphNum(*graphs[0]); + size_t graph_num = ir::GraphNum(*graph); if (graph_num > 1) { LOG(WARNING) << "The number of graph should be only one, " "but the current graph has " - << ir::GraphNum(*graphs[0]) + << ir::GraphNum(*graph) << " sub_graphs. If you want to see the nodes of the " "sub_graphs, you should use 'FLAGS_print_sub_graph_dir' " "to specify the output dir. NOTES: if you not do training, " @@ -319,18 +304,25 @@ ParallelExecutor::ParallelExecutor( } if (build_strategy.enable_parallel_graph_) { +#ifdef PADDLE_WITH_CUDA + // TODO(Yancey1989): Remove passing in the main_program when + // allreduce_seq_pass doesn't need it as the attr. member_->executor_.reset(new details::ParallelSSAGraphExecutor( - exec_strategy, member_->local_scopes_, member_->places_, - std::move(graphs))); + exec_strategy, member_->local_scopes_, member_->places_, main_program, + std::move(graph))); +#else + PADDLE_THROW( + "Paddle should be compiled with CUDA for ParallelGraph Execution."); +#endif } else { if (exec_strategy.type_ == ExecutionStrategy::kDefault) { member_->executor_.reset(new details::ThreadedSSAGraphExecutor( exec_strategy, member_->local_scopes_, member_->places_, - std::move(graphs[0]))); + std::move(graph))); } else { member_->executor_.reset(new details::FastThreadedSSAGraphExecutor( exec_strategy, member_->local_scopes_, member_->places_, - std::move(graphs[0]))); + std::move(graph))); } } @@ -482,11 +474,10 @@ bool ParallelExecutor::EnableParallelGraphExecution( } if (!member_->use_all_reduce_ || !member_->use_cuda_) - enable_parallel_graph = false; - if (build_strategy.enable_sequential_execution_ || - exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental) - enable_parallel_graph = false; + if (build_strategy.enable_sequential_execution_ || + exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental) + enable_parallel_graph = false; return enable_parallel_graph; } diff --git a/paddle/fluid/inference/api/api.cc b/paddle/fluid/inference/api/api.cc index 6cd18277d63200..f83537f064187e 100644 --- a/paddle/fluid/inference/api/api.cc +++ b/paddle/fluid/inference/api/api.cc @@ -92,7 +92,7 @@ void PaddleBuf::Reset(void *data, size_t length) { void PaddleBuf::Free() { if (memory_owned_ && data_) { - PADDLE_ENFORCE_GT(length_, 0); + PADDLE_ENFORCE_GT(length_, 0UL); free(static_cast(data_)); data_ = nullptr; length_ = 0; diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc index dd953e0dccbb37..bd0059e18485c0 100644 --- a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc @@ -56,14 +56,14 @@ struct DataRecord { std::vector slot_data; split_to_float(data[1], ' ', &slot_data); std::string name = data[0]; - PADDLE_ENFORCE_EQ(slot_data.size() % 11, 0, + PADDLE_ENFORCE_EQ(slot_data.size() % 11, 0UL, "line %d, %s should be divisible", num_lines, name); datasets[name].emplace_back(std::move(slot_data)); } num_samples = num_lines / num_slots; PADDLE_ENFORCE_EQ(num_samples * num_slots, static_cast(num_lines), "num samples should be divisible"); - PADDLE_ENFORCE_GT(num_samples, 0); + PADDLE_ENFORCE_GT(num_samples, 0UL); } void Prepare(int bs) { diff --git a/paddle/fluid/inference/tests/test.cmake b/paddle/fluid/inference/tests/test.cmake index 29f0f034a2aab5..6c5fe043ffa3f3 100644 --- a/paddle/fluid/inference/tests/test.cmake +++ b/paddle/fluid/inference/tests/test.cmake @@ -1,18 +1,43 @@ +include(ExternalProject) set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url") set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING "A path setting inference demo download directories.") -function (inference_download install_dir url filename) - message(STATUS "Download inference test stuff from ${url}/${filename}") - file(DOWNLOAD "${url}/${filename}" "${install_dir}/${filename}") - message(STATUS "finish downloading ${filename}") + +function(inference_download INSTALL_DIR URL FILENAME) + message(STATUS "Download inference test stuff from ${URL}/${FILENAME}") + string(REGEX REPLACE "[-%.]" "_" FILENAME_EX ${FILENAME}) + ExternalProject_Add( + extern_inference_download_${FILENAME_EX} + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${INSTALL_DIR} + URL ${URL}/${FILENAME} + DOWNLOAD_COMMAND wget -q -O ${INSTALL_DIR}/${FILENAME} ${URL}/${FILENAME} + DOWNLOAD_DIR ${INSTALL_DIR} + DOWNLOAD_NO_PROGRESS 1 + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + UPDATE_COMMAND "" + INSTALL_COMMAND "" + ) endfunction() -function (inference_download_and_uncompress install_dir url filename) - inference_download(${install_dir} ${url} ${filename}) - execute_process( - COMMAND ${CMAKE_COMMAND} -E tar xzf ${install_dir}/${filename} - WORKING_DIRECTORY ${install_dir} - ) +function(inference_download_and_uncompress INSTALL_DIR URL FILENAME) + message(STATUS "Download inference test stuff from ${URL}/${FILENAME}") + string(REGEX REPLACE "[-%.]" "_" FILENAME_EX ${FILENAME}) + set(EXTERNAL_PROJECT_NAME "extern_inference_download_${FILENAME_EX}") + set(UNPACK_DIR "${INSTALL_DIR}/src/${EXTERNAL_PROJECT_NAME}") + ExternalProject_Add( + ${EXTERNAL_PROJECT_NAME} + ${EXTERNAL_PROJECT_LOG_ARGS} + PREFIX ${INSTALL_DIR} + URL ${URL}/${FILENAME} + DOWNLOAD_DIR ${INSTALL_DIR} + DOWNLOAD_NO_PROGRESS 1 + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + UPDATE_COMMAND "" + INSTALL_COMMAND ${CMAKE_COMMAND} -E copy_directory ${UNPACK_DIR} ${INSTALL_DIR} + ) endfunction() set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec") diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 75fa611c0d701d..861f69f4d2143b 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -171,9 +171,7 @@ void TestInference(const std::string& dirname, // Enable the profiler paddle::platform::EnableProfiler(state); { - paddle::platform::RecordEvent record_event( - "init_program", - paddle::platform::DeviceContextPool::Instance().Get(place)); + paddle::platform::RecordEvent record_event("init_program"); inference_program = InitProgram(&executor, scope, dirname, is_combined); } @@ -230,9 +228,7 @@ void TestInference(const std::string& dirname, // Run repeat times to profile the performance for (int i = 0; i < repeat; ++i) { - paddle::platform::RecordEvent record_event( - "run_inference", - paddle::platform::DeviceContextPool::Instance().Get(place)); + paddle::platform::RecordEvent record_event("run_inference"); if (PrepareContext) { // Note: if you change the inference_program, you need to call diff --git a/paddle/fluid/memory/allocation/legacy_allocator.cc b/paddle/fluid/memory/allocation/legacy_allocator.cc index e983ae327d6938..1936f9d4cd83c5 100644 --- a/paddle/fluid/memory/allocation/legacy_allocator.cc +++ b/paddle/fluid/memory/allocation/legacy_allocator.cc @@ -356,7 +356,7 @@ void MemInfo::Minus(const size_t &size) { usage_ -= size; } -uint64_t MemInfo::GetPeakUsage() { return peak_usage_; } +uint64_t MemInfo::GetPeakUsage() const { return peak_usage_; } LegacyMemMonitor::~LegacyMemMonitor() { for (auto &item : gpu_mem_info_) delete item.second; @@ -380,10 +380,10 @@ void LegacyMemMonitor::Minus(const int &device, const size_t &size) { gpu_mem_info_[device]->Minus(size); } -uint64_t LegacyMemMonitor::GetMemUsage(const int &device) { +uint64_t LegacyMemMonitor::GetMemUsage(const int &device) const { return gpu_mem_info_.find(device) == gpu_mem_info_.end() ? 0 - : gpu_mem_info_[device]->GetPeakUsage(); + : gpu_mem_info_.at(device)->GetPeakUsage(); } void LegacyMemMonitor::PrintMemUsage() { diff --git a/paddle/fluid/memory/allocation/legacy_allocator.h b/paddle/fluid/memory/allocation/legacy_allocator.h index ccbc8c70d8e9a1..d9bdae153da643 100644 --- a/paddle/fluid/memory/allocation/legacy_allocator.h +++ b/paddle/fluid/memory/allocation/legacy_allocator.h @@ -27,20 +27,20 @@ namespace allocation { class MemInfo { public: MemInfo() : usage_(0), peak_usage_(0) {} - MemInfo(const MemInfo &) = delete; - MemInfo &operator=(const MemInfo &) = delete; // return a flag to indicate current operation will create a peak point or not bool Add(const size_t &); void Minus(const size_t &); - uint64_t GetPeakUsage(); + uint64_t GetPeakUsage() const; private: /* current memory usage*/ uint64_t usage_; uint64_t peak_usage_; std::mutex mutex_; + + DISABLE_COPY_AND_ASSIGN(MemInfo); }; class LegacyMemMonitor { @@ -56,11 +56,11 @@ class LegacyMemMonitor { void Add(const int &, const size_t &); void Minus(const int &, const size_t &); - uint64_t GetMemUsage(const int &); + uint64_t GetMemUsage(const int &) const; void PrintMemUsage(); - protected: + private: MemUsage gpu_mem_info_; }; diff --git a/paddle/fluid/operators/attention_lstm_op.cc b/paddle/fluid/operators/attention_lstm_op.cc index b6996be4b0984b..912ec79910301b 100644 --- a/paddle/fluid/operators/attention_lstm_op.cc +++ b/paddle/fluid/operators/attention_lstm_op.cc @@ -293,7 +293,7 @@ class AttentionLSTMKernel : public framework::OpKernel { int len = x_lod[0][i + 1] - x_lod[0][i]; max_seq_len = max_seq_len < len ? len : max_seq_len; } - PADDLE_ENFORCE_EQ(x_lod.size(), 1, "Input(X)'s lod size must be 1."); + PADDLE_ENFORCE_EQ(x_lod.size(), 1UL, "Input(X)'s lod size must be 1."); PADDLE_ENFORCE_EQ(c0->dims()[0], N, "C0 dims should be %d x %d.", N, D); fc_out->Resize({max_seq_len, 1}); diff --git a/paddle/fluid/operators/controlflow/compare_op.cc b/paddle/fluid/operators/controlflow/compare_op.cc index 688457d4a75168..5d3f9b43f8c08d 100644 --- a/paddle/fluid/operators/controlflow/compare_op.cc +++ b/paddle/fluid/operators/controlflow/compare_op.cc @@ -51,6 +51,11 @@ class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { comment.type)); AddInput("Y", string::Sprintf("the right hand operand of %s operator", comment.type)); + AddAttr( + "axis", + "The start dimension index for broadcasting Y onto X. [default -1]") + .SetDefault(-1) + .EqualGreaterThan(-1); AddAttr("force_cpu", "Force fill output variable to cpu " "memory. Otherwise, fill output variable to the running " @@ -64,11 +69,6 @@ N-dim tensor. X and Y could be any type. The each element of the Out tensor is calculated by $%s$ )DOC", comment.equation)); - AddAttr( - "axis", - "The start dimension index for broadcasting Y onto X. [default -1]") - .SetDefault(-1) - .EqualGreaterThan(-1); } }; diff --git a/paddle/fluid/operators/controlflow/get_places_op.cc b/paddle/fluid/operators/controlflow/get_places_op.cc index db6ff782569017..1a157688f3d021 100644 --- a/paddle/fluid/operators/controlflow/get_places_op.cc +++ b/paddle/fluid/operators/controlflow/get_places_op.cc @@ -52,7 +52,7 @@ class GetPlacesOp : public framework::OperatorBase { device_count = is_gpu ? CUDADevCount() : std::thread::hardware_concurrency(); } - PADDLE_ENFORCE_NE(device_count, 0, "Cannot indicate %s device count", + PADDLE_ENFORCE_NE(device_count, 0UL, "Cannot indicate %s device count", is_gpu ? "GPU" : "CPU"); auto out_var_name = Output("Out"); diff --git a/paddle/fluid/operators/crf_decoding_op.cc b/paddle/fluid/operators/crf_decoding_op.cc index 81c9e9e543191d..e053ae57739d3d 100644 --- a/paddle/fluid/operators/crf_decoding_op.cc +++ b/paddle/fluid/operators/crf_decoding_op.cc @@ -84,12 +84,12 @@ class CRFDecodingOp : public framework::OperatorWithKernel { "Output(ViterbiPath) should be not null."); auto emission_dims = ctx->GetInputDim("Emission"); - PADDLE_ENFORCE_EQ(emission_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(emission_dims.size(), 2, "The Input(Emission) should be a 2-D tensor."); PADDLE_ENFORCE(emission_dims[0], "An empty mini-batch is not allowed."); auto transition_dims = ctx->GetInputDim("Transition"); - PADDLE_ENFORCE_EQ(transition_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(transition_dims.size(), 2, "The Input(Transition) should be a 2-D tensor."); PADDLE_ENFORCE_EQ( transition_dims[0] - 2, transition_dims[1], diff --git a/paddle/fluid/operators/detection/anchor_generator_op.cc b/paddle/fluid/operators/detection/anchor_generator_op.cc index f2984d1af2f26d..4a333b559f82e6 100644 --- a/paddle/fluid/operators/detection/anchor_generator_op.cc +++ b/paddle/fluid/operators/detection/anchor_generator_op.cc @@ -85,7 +85,7 @@ class AnchorGeneratorOpMaker : public framework::OpProtoAndCheckerMaker { " For instance, the anchor size of 64 means the area of this anchor " "equals to 64**2.") .AddCustomChecker([](const std::vector& anchor_sizes) { - PADDLE_ENFORCE_GT(anchor_sizes.size(), 0, + PADDLE_ENFORCE_GT(anchor_sizes.size(), 0UL, "Size of anchor_sizes must be at least 1."); for (size_t i = 0; i < anchor_sizes.size(); ++i) { PADDLE_ENFORCE_GT(anchor_sizes[i], 0.0, @@ -103,7 +103,7 @@ class AnchorGeneratorOpMaker : public framework::OpProtoAndCheckerMaker { "(vector) List of variances to be used " "in box regression deltas") .AddCustomChecker([](const std::vector& variances) { - PADDLE_ENFORCE_EQ(variances.size(), 4, + PADDLE_ENFORCE_EQ(variances.size(), 4UL, "Must and only provide 4 variance."); for (size_t i = 0; i < variances.size(); ++i) { PADDLE_ENFORCE_GT(variances[i], 0.0, @@ -117,7 +117,7 @@ class AnchorGeneratorOpMaker : public framework::OpProtoAndCheckerMaker { .SetDefault(std::vector(2, 16.0)) .AddCustomChecker([](const std::vector& stride) { PADDLE_ENFORCE_EQ( - stride.size(), 2, + stride.size(), 2UL, "Must and only provide 2 stride for width and height."); for (size_t i = 0; i < stride.size(); ++i) { PADDLE_ENFORCE_GT(stride[i], 0.0, diff --git a/paddle/fluid/operators/distributed/brpc/brpc_client.cc b/paddle/fluid/operators/distributed/brpc/brpc_client.cc index b8e63f42e20407..a1a3443348129b 100644 --- a/paddle/fluid/operators/distributed/brpc/brpc_client.cc +++ b/paddle/fluid/operators/distributed/brpc/brpc_client.cc @@ -80,7 +80,7 @@ VarHandlePtr BRPCClient::AsyncSendVar(const std::string& ep, google::protobuf::Closure* done = brpc::NewCallback( &HandleSendResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); ch_ctx->stub->SendVariable(cntl, &request, response, done); @@ -184,7 +184,7 @@ VarHandlePtr BRPCClient::_AsyncGetVar(const std::string& ep, google::protobuf::Closure* done = brpc::NewCallback( &HandleGetResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); if (method_name == kGetMonomerRPC) { ch_ctx->stub->GetMonomerVariable(cntl, &req, response, done); @@ -272,7 +272,7 @@ VarHandlePtr BRPCClient::AsyncPrefetchVar(const std::string& ep, &cntl->request_attachment(), out_var_name_val, false, 0, table_name_val); - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); google::protobuf::Closure* done = brpc::NewCallback( &HandleGetResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); @@ -311,7 +311,7 @@ VarHandlePtr BRPCClient::AsyncSendFetchBarrier(const std::string& ep, VarHandlePtr var_h( new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr)); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); google::protobuf::Closure* done = brpc::NewCallback( &HandleFetchBarrierResponse, cntl, response, var_h, ch_ptr, ch_ctx, this); @@ -406,7 +406,7 @@ VarHandlePtr BRPCClient::AsyncSendVarMessage( sendrecv::VoidMessage* response = new sendrecv::VoidMessage(); cntl->set_timeout_ms(time_out); - platform::RecordRPCEvent record_event(method_name, nullptr); + platform::RecordRPCEvent record_event(method_name); VarHandlePtr var_h( new VarHandle(ep, method_name, req.varname(), nullptr, nullptr)); diff --git a/paddle/fluid/operators/distributed/grpc/grpc_client.cc b/paddle/fluid/operators/distributed/grpc/grpc_client.cc index 52310f8d04db6a..61e94dae3c7a10 100644 --- a/paddle/fluid/operators/distributed/grpc/grpc_client.cc +++ b/paddle/fluid/operators/distributed/grpc/grpc_client.cc @@ -89,7 +89,7 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, // stub context s->response_call_back_ = nullptr; - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); @@ -184,7 +184,7 @@ VarHandlePtr GRPCClient::_AsyncGetVar( // stub context s->response_call_back_ = ProcGetResponse; - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); auto call = s->stub_g_.PrepareUnaryCall(s->context_.get(), rpc_path, buf, &cq_); @@ -235,7 +235,7 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep, // stub context s->response_call_back_ = ProcGetResponse; - platform::RecordRPCEvent record_event(method, p_ctx); + platform::RecordRPCEvent record_event(method); auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req, @@ -265,7 +265,7 @@ VarHandlePtr GRPCClient::AsyncSendBatchBarrier(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(BATCH_BARRIER_MESSAGE); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -290,7 +290,7 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(FETCH_BARRIER_MESSAGE); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -317,7 +317,7 @@ VarHandlePtr GRPCClient::AsyncGetMonomerBarrier(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(var_name); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncGetMonomerBarrier(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -342,7 +342,7 @@ VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep, sendrecv::VariableMessage req; req.set_varname(COMPLETE_MESSAGE); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); @@ -372,7 +372,7 @@ VarHandlePtr GRPCClient::AsyncCheckpointNotify(const std::string& ep, req.set_varname(CHECKPOINT_SAVE_MESSAGE); req.set_out_varname(dir); - platform::RecordRPCEvent record_event(method, nullptr); + platform::RecordRPCEvent record_event(method); auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); diff --git a/paddle/fluid/operators/distributed/grpc/grpc_serde.cc b/paddle/fluid/operators/distributed/grpc/grpc_serde.cc index 6df4fd36f95b12..6e65aa5fae8353 100644 --- a/paddle/fluid/operators/distributed/grpc/grpc_serde.cc +++ b/paddle/fluid/operators/distributed/grpc/grpc_serde.cc @@ -38,7 +38,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, ::grpc::ByteBuffer* msg, const std::string& out_name, const int trainer_id, const std::string& table_name) { - platform::RecordRPCEvent record_event("serial", &ctx); + platform::RecordRPCEvent record_event("serial"); VarMsg request; TensorPayload* payload = nullptr; @@ -147,7 +147,7 @@ void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, const framework::Scope* scope, framework::Variable** var, int* trainer_id) { - platform::RecordRPCEvent record_event("deserial", &ctx); + platform::RecordRPCEvent record_event("deserial"); operators::distributed::GRPCVariableResponse resp(scope, &ctx); PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!"); *var = resp.GetVar(); diff --git a/paddle/fluid/operators/fc_op.cc b/paddle/fluid/operators/fc_op.cc index 38e57a41ed253e..eb4617a9359353 100644 --- a/paddle/fluid/operators/fc_op.cc +++ b/paddle/fluid/operators/fc_op.cc @@ -47,7 +47,7 @@ void FCOp::InferShape(framework::InferShapeContext* ctx) const { PADDLE_ENFORCE(in_dims.size() == 2 || in_dims.size() == 4, "Fully Connected input should be 2-D or 4-D tensor."); } - PADDLE_ENFORCE_EQ(w_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(w_dims.size(), 2, "Fully Connected input should be 2-D tensor."); int in_num_col_dims = ctx->Attrs().Get("in_num_col_dims"); PADDLE_ENFORCE_GT( diff --git a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h index 758432fd9e4197..33a1b47d150f65 100644 --- a/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h +++ b/paddle/fluid/operators/fused/fused_embedding_seq_pool_op.h @@ -21,6 +21,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/jit/kernels.h" #include "paddle/fluid/operators/math/blas.h" namespace paddle { @@ -37,32 +38,25 @@ struct EmbeddingVSumFunctor { const LoDTensor *table_t, const LoDTensor *ids_t, LoDTensor *output_t) { auto *table = table_t->data(); - int64_t row_number = table_t->dims()[0]; - int64_t row_width = table_t->dims()[1]; - int64_t last_dim = output_t->dims()[1]; + int64_t table_height = table_t->dims()[0]; + int64_t table_width = table_t->dims()[1]; + int64_t out_width = output_t->dims()[1]; const int64_t *ids = ids_t->data(); auto ids_lod = ids_t->lod()[0]; - int64_t ids_count = ids_t->numel() / ids_lod.back(); - + int64_t idx_width = ids_t->numel() / ids_lod.back(); auto *output = output_t->mutable_data(context.GetPlace()); - auto blas = math::GetBlas(context); - for (int64_t i = 0; i != ids_lod.size() - 1; ++i) { - size_t begin = ids_lod[i] * ids_count; - for (int64_t j = 0; j != ids_count; ++j) { - PADDLE_ENFORCE_LT(ids[begin], row_number); - PADDLE_ENFORCE_GE(ids[begin], 0, "ids %d", i); - blas.VCOPY(row_width, table + ids[begin + j] * row_width, - output + i * last_dim + j * row_width); - } - - for (int64_t r = (ids_lod[i] + 1) * ids_count; - r < ids_lod[i + 1] * ids_count; ++r) { - PADDLE_ENFORCE_LT(ids[r], row_number); - PADDLE_ENFORCE_GE(ids[r], 0, "ids %d", i); - blas.AXPY(row_width, 1., table + ids[r] * row_width, - output + i * last_dim + (r % ids_count) * row_width); - } + PADDLE_ENFORCE_LE(table_width * idx_width, out_width); + PADDLE_ENFORCE_GT(ids_lod.size(), 1UL); + + jit::emb_seq_pool_attr_t attr(table_height, table_width, 0, idx_width, + out_width, jit::SeqPoolType::kSum); + for (size_t i = 0; i != ids_lod.size() - 1; ++i) { + attr.index_height = ids_lod[i + 1] - ids_lod[i]; + auto emb_seqpool = jit::Get, + platform::CPUPlace>(attr); + emb_seqpool(table, ids + ids_lod[i] * idx_width, output + i * out_width, + &attr); } } }; diff --git a/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc b/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc index e9e2a3b1f5c1c0..8ecdf2ed9d40e7 100644 --- a/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc +++ b/paddle/fluid/operators/fused/fusion_repeated_fc_relu_op.cc @@ -37,7 +37,7 @@ void FusionRepeatedFCReluOp::InferShape( "Output(Out) of FusionRepeatedFCReluOp should not be null."); auto i_dims = ctx->GetInputDim("X"); - PADDLE_ENFORCE_EQ(i_dims.size(), 2UL, "Input shape size should be 2"); + PADDLE_ENFORCE_EQ(i_dims.size(), 2, "Input shape size should be 2"); auto w_dims = ctx->GetInputsDim("W"); auto b_dims = ctx->GetInputsDim("Bias"); @@ -49,7 +49,7 @@ void FusionRepeatedFCReluOp::InferShape( "inpute width should be equal with weight height"); for (size_t i = 1; i < sz; ++i) { - PADDLE_ENFORCE_EQ(w_dims[i].size(), 2UL, + PADDLE_ENFORCE_EQ(w_dims[i].size(), 2, "Every weight shape size should be 2."); PADDLE_ENFORCE_EQ(framework::product(b_dims[i]), w_dims[i][1], "The length of Bias must be equal with w_dims[1]."); diff --git a/paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc b/paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc index aaef46de0d3b88..d091da5aa8a7e7 100644 --- a/paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc +++ b/paddle/fluid/operators/fused/fusion_seqexpand_concat_fc_op.cc @@ -39,7 +39,7 @@ void FusionSeqExpandConcatFCOp::InferShape( auto ins_dims = ctx->GetInputsDim("X"); auto w_dims = ctx->GetInputDim("FCWeight"); // (M0+M1+M2+..) x D - PADDLE_ENFORCE_EQ(w_dims.size(), 2UL, "Input(FCWeight)'s rank must be 2."); + PADDLE_ENFORCE_EQ(w_dims.size(), 2, "Input(FCWeight)'s rank must be 2."); const int D = w_dims[1]; int sum = ins_dims[0][1]; for (size_t i = 1; i < ins_dims.size(); ++i) { diff --git a/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc b/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc index b181140db750a8..d48bdafe0aa38c 100644 --- a/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc +++ b/paddle/fluid/operators/fused/fusion_seqpool_concat_op.cc @@ -39,7 +39,7 @@ void FusionSeqPoolConcatOp::InferShape( // The output height should be confirmed in Compute, // since input lod is not accessible here. - PADDLE_ENFORCE_EQ(ins_dims[0].size(), 2UL, + PADDLE_ENFORCE_EQ(ins_dims[0].size(), 2, "The dims size of first input should be 2."); ctx->SetOutputDim("Out", {-1, ins_dims[0][axis] * static_cast(n)}); } diff --git a/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc b/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc index 8c8b079633aacb..8493f4468fc994 100644 --- a/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc +++ b/paddle/fluid/operators/fused/fusion_squared_mat_sub_op.cc @@ -42,7 +42,7 @@ void FusionSquaredMatSubOp::InferShape( auto y_dims = ctx->GetInputDim("Y"); PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(), "Input tensors dims size should be equal."); - PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "Input tensors should be a Matrix."); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input tensors should be a Matrix."); PADDLE_ENFORCE_EQ(x_dims[1], y_dims[0], "Inputs Matrix should be multiply."); ctx->SetOutputDim("SquaredX", x_dims); diff --git a/paddle/fluid/operators/jit/benchmark.cc b/paddle/fluid/operators/jit/benchmark.cc index 77a2d04ebf176c..3348778ee782ef 100644 --- a/paddle/fluid/operators/jit/benchmark.cc +++ b/paddle/fluid/operators/jit/benchmark.cc @@ -301,6 +301,37 @@ void BenchSeqPoolKernel() { } } +template +void BenchEmbSeqPoolKernel() { + std::vector pool_types = {jit::SeqPoolType::kSum}; + int64_t tbl_h = 1e4; + for (int tbl_w : {10, 16, 256}) { + Tensor table; + table.Resize({tbl_h, tbl_w}); + RandomVec(tbl_h * tbl_w, table.mutable_data(PlaceType()), -2.f, 2.f); + const T* table_data = table.data(); + for (auto type : pool_types) { + for (int idx_w : {1, 2, 10, 16}) { + for (int idx_h : {1, 2, 9, 13, 16}) { + int64_t out_w = tbl_w * idx_w; + jit::emb_seq_pool_attr_t attr(tbl_h, tbl_w, idx_h, idx_w, out_w, + type); + Tensor idx, out; + idx.Resize({idx_h, idx_w}); + out.Resize({out_w}); + RandomVec(idx_h * idx_w, + idx.mutable_data(PlaceType()), 0, + tbl_h - 1); + const int64_t* idx_data = idx.data(); + T* o_data = out.mutable_data(PlaceType()); + BenchAllImpls, PlaceType>( + attr, table_data, idx_data, o_data, &attr); + } + } + } + } +} + template void BenchMatMulKernel() { for (int m : {1, 2, 3, 4}) { @@ -441,6 +472,11 @@ BENCH_FP32_CPU(kGRUHtPart2) { BenchGRUKernel(); } // seq pool function BENCH_FP32_CPU(kSeqPool) { BenchSeqPoolKernel(); } +// embedding seq pool function +BENCH_FP32_CPU(kEmbSeqPool) { + BenchEmbSeqPoolKernel(); +} + // matmul BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel(); } diff --git a/paddle/fluid/operators/jit/gen/CMakeLists.txt b/paddle/fluid/operators/jit/gen/CMakeLists.txt index efc7eb79d36c5c..294f73d9646c93 100644 --- a/paddle/fluid/operators/jit/gen/CMakeLists.txt +++ b/paddle/fluid/operators/jit/gen/CMakeLists.txt @@ -31,3 +31,4 @@ USE_JITKERNEL_GEN(kNCHW16CMulNC) USE_JITKERNEL_GEN(kSeqPool) USE_JITKERNEL_GEN(kHMax) USE_JITKERNEL_GEN(kHSum) +USE_JITKERNEL_GEN(kEmbSeqPool) diff --git a/paddle/fluid/operators/jit/gen/embseqpool.cc b/paddle/fluid/operators/jit/gen/embseqpool.cc new file mode 100644 index 00000000000000..23837a3fb9886a --- /dev/null +++ b/paddle/fluid/operators/jit/gen/embseqpool.cc @@ -0,0 +1,149 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#include "paddle/fluid/operators/jit/gen/embseqpool.h" +#include // offsetof +#include +#include "paddle/fluid/operators/jit/gen/act.h" // for exp_float_consts ones +#include "paddle/fluid/operators/jit/registry.h" +#include "paddle/fluid/platform/cpu_info.h" + +namespace paddle { +namespace operators { +namespace jit { +namespace gen { + +void EmbSeqPoolJitCode::genCode() { + preCode(); + constexpr int block = YMM_FLOAT_BLOCK; + constexpr int max_num_regs = 8; + const int num_block = tbl_w_ / block; + const int num_groups = num_block / max_num_regs; + const size_t block_size = sizeof(float) * block; + std::vector groups(num_groups, max_num_regs); + int rest_num_regs = num_block % max_num_regs; + if (rest_num_regs > 0) { + groups.push_back(rest_num_regs); + } + + // protect param_dst + mov(reg_ptr_param_dst, param_dst); + mov(reg_idx_width_in_byte, + qword[param_attr + offsetof(emb_seq_pool_attr_t, index_width)]); + mov(reg_idx_height, + qword[param_attr + offsetof(emb_seq_pool_attr_t, index_height)]); + mov(rax, sizeof(int64_t)); + mul(reg_idx_width_in_byte); + mov(reg_idx_width_in_byte, rax); + const size_t tbl_width_in_byte = sizeof(float) * tbl_w_; + int acc_num_regs = 0; + for (int num_regs : groups) { + Label l_next_idx_w, l_next_idx_h, l_save_now; + xor_(reg_idx_w_i_in_byte, reg_idx_w_i_in_byte); + mov(reg_ptr_dst_i, reg_ptr_param_dst); + add(reg_ptr_dst_i, acc_num_regs * block_size); + + L(l_next_idx_w); + { + // h == 0 + mov(reg_ptr_idx_i, param_idx); + add(reg_ptr_idx_i, reg_idx_w_i_in_byte); + mov(reg_idx, qword[reg_ptr_idx_i]); + mov(rax, tbl_width_in_byte); + mul(reg_idx); + mov(reg_ptr_tbl_i, rax); // reg is offset now + add(reg_ptr_tbl_i, param_tbl); // reg is ptr_i now + size_t w_offset = 0; + for (int reg_i = 0; reg_i < num_regs; ++reg_i) { + vmovups(ymm_t(reg_i + num_regs), ptr[reg_ptr_tbl_i + w_offset]); + w_offset += block_size; + } + add(reg_ptr_idx_i, reg_idx_width_in_byte); + + // end condition of idx h + mov(reg_idx_h_end, reg_idx_height); + mov(rax, reg_idx_width_in_byte); + mul(reg_idx_h_end); + mov(reg_idx_h_end, rax); + add(reg_idx_h_end, reg_idx_w_i_in_byte); + add(reg_idx_h_end, param_idx); + + cmp(reg_ptr_idx_i, reg_idx_h_end); + jge(l_save_now, T_NEAR); + L(l_next_idx_h); + { + mov(reg_idx, qword[reg_ptr_idx_i]); + mov(reg_ptr_tbl_i, reg_idx); + mov(rax, tbl_width_in_byte); + mul(reg_idx); + mov(reg_ptr_tbl_i, rax); + add(reg_ptr_tbl_i, param_tbl); + size_t w_offset = 0; + for (int reg_i = 0; reg_i < num_regs; ++reg_i) { + vmovups(ymm_t(reg_i), ptr[reg_ptr_tbl_i + w_offset]); + vaddps(ymm_t(reg_i + num_regs), ymm_t(reg_i + num_regs), + ymm_t(reg_i)); + w_offset += block_size; + } + add(reg_ptr_idx_i, reg_idx_width_in_byte); + cmp(reg_ptr_idx_i, reg_idx_h_end); + jl(l_next_idx_h, T_NEAR); + } // end of idx h + L(l_save_now); + // avg or sqrt here, if needed + w_offset = 0; + for (int reg_i = 0; reg_i < num_regs; ++reg_i) { + vmovups(ptr[reg_ptr_dst_i + w_offset], ymm_t(reg_i + num_regs)); + w_offset += block_size; + } + add(reg_ptr_dst_i, tbl_width_in_byte); + add(reg_idx_w_i_in_byte, sizeof(int64_t)); + cmp(reg_idx_w_i_in_byte, reg_idx_width_in_byte); + jl(l_next_idx_w, T_NEAR); + } // end of idx w + + acc_num_regs += num_regs; + add(param_tbl, num_regs * block_size); // do not use acc_num_regs + } // end of groups + postCode(); +} + +class EmbSeqPoolCreator : public JitCodeCreator { + public: + bool UseMe(const emb_seq_pool_attr_t& attr) const override { + return platform::MayIUse(platform::avx) && + attr.table_width % YMM_FLOAT_BLOCK == 0; + } + size_t CodeSize(const emb_seq_pool_attr_t& attr) const override { + return 96 + (attr.table_width / YMM_FLOAT_BLOCK) * 96 * 8; + } + std::unique_ptr CreateJitCode( + const emb_seq_pool_attr_t& attr) const override { + PADDLE_ENFORCE_GT(attr.table_height, 0); + PADDLE_ENFORCE_GT(attr.table_width, 0); + PADDLE_ENFORCE_GT(attr.index_height, 0); + PADDLE_ENFORCE_GT(attr.index_width, 0); + PADDLE_ENFORCE_GT(attr.out_width, 0); + return make_unique(attr, CodeSize(attr)); + } +}; + +} // namespace gen +} // namespace jit +} // namespace operators +} // namespace paddle + +namespace gen = paddle::operators::jit::gen; + +REGISTER_JITKERNEL_GEN(kEmbSeqPool, gen::EmbSeqPoolCreator); diff --git a/paddle/fluid/operators/jit/gen/embseqpool.h b/paddle/fluid/operators/jit/gen/embseqpool.h new file mode 100644 index 00000000000000..5afcfbdc1786be --- /dev/null +++ b/paddle/fluid/operators/jit/gen/embseqpool.h @@ -0,0 +1,81 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. */ + +#pragma once + +#include +#include "glog/logging.h" +#include "paddle/fluid/operators/jit/gen/jitcode.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace operators { +namespace jit { +namespace gen { + +class EmbSeqPoolJitCode : public JitCode { + public: + explicit EmbSeqPoolJitCode(const emb_seq_pool_attr_t& attr, + size_t code_size = 256 * 1024, + void* code_ptr = nullptr) + : JitCode(code_size, code_ptr), + tbl_w_(attr.table_width), + type_(attr.pool_type) { + if (type_ != SeqPoolType::kSum) { + LOG(FATAL) << "Only support sum pool yet "; + } + this->genCode(); + } + + std::string name() const override { + std::string base = "EmbSeqPoolJitCode"; + if (type_ == SeqPoolType::kSum) { + base += "_Sum"; + } else if (type_ == SeqPoolType::kAvg) { + base += "_Avg"; + } else if (type_ == SeqPoolType::kSqrt) { + base += "_Sqrt"; + } + base += ("_W" + std::to_string(tbl_w_)); + return base; + } + void genCode() override; + + private: + int tbl_w_; + SeqPoolType type_; + reg64_t param_tbl{abi_param1}; + reg64_t param_idx{abi_param2}; + reg64_t param_dst{abi_param3}; + reg64_t param_attr{abi_param4}; + + reg64_t reg_tmp{rax}; + + reg64_t reg_idx_width_in_byte{r8}; + reg64_t reg_idx_height{r9}; + + reg64_t reg_ptr_tbl_i{r10}; + reg64_t reg_idx{r10}; // could use same of reg_ptr_tbl_i + reg64_t reg_ptr_idx_i{r11}; + reg64_t reg_ptr_dst_i{r12}; + reg64_t reg_ptr_param_dst{r13}; // rdx is used in mul so protect param_dst + + reg64_t reg_idx_w_i_in_byte{r14}; + reg64_t reg_idx_h_end{r15}; +}; + +} // namespace gen +} // namespace jit +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/jit/gen/seqpool.h b/paddle/fluid/operators/jit/gen/seqpool.h index 4108ee2f46433f..e909bc7c7939ee 100644 --- a/paddle/fluid/operators/jit/gen/seqpool.h +++ b/paddle/fluid/operators/jit/gen/seqpool.h @@ -32,7 +32,7 @@ class SeqPoolJitCode : public JitCode { : JitCode(code_size, code_ptr), w_(attr.w), type_(attr.type) { if (!(type_ == SeqPoolType::kSum || type_ == SeqPoolType::kAvg || type_ == SeqPoolType::kSqrt)) { - LOG(FATAL) << "Only support sum pool yet "; + LOG(FATAL) << "Only supported pool type: sum, avg and sqrt."; } fp_h_[0] = 1.f; this->genCode(); diff --git a/paddle/fluid/operators/jit/helper.cc b/paddle/fluid/operators/jit/helper.cc index e7292fe2bd8031..a7665361328989 100644 --- a/paddle/fluid/operators/jit/helper.cc +++ b/paddle/fluid/operators/jit/helper.cc @@ -54,6 +54,7 @@ const char* to_string(KernelType kt) { ONE_CASE(kHMax); ONE_CASE(kHSum); ONE_CASE(kSoftmax); + ONE_CASE(kEmbSeqPool); default: PADDLE_THROW("Not support type: %d, or forget to add it.", kt); return "NOT JITKernel"; diff --git a/paddle/fluid/operators/jit/helper.h b/paddle/fluid/operators/jit/helper.h index d5773d65940127..07998588a5a560 100644 --- a/paddle/fluid/operators/jit/helper.h +++ b/paddle/fluid/operators/jit/helper.h @@ -172,6 +172,15 @@ inline std::ostream& operator<<(std::ostream& os, const seq_pool_attr_t& attr) { return os; } +inline std::ostream& operator<<(std::ostream& os, + const emb_seq_pool_attr_t& attr) { + os << "table_height[" << attr.table_height << "],table_width[" + << attr.table_width << "],index_height[" << attr.index_height + << "],index_width[" << attr.index_width << "],output_width[" + << attr.out_width << "],pool_type[" << to_string(attr.pool_type) << "]"; + return os; +} + inline std::ostream& operator<<(std::ostream& os, const matmul_attr_t& attr) { os << "M[" << attr.m << "],N[" << attr.n << "],K[" << attr.k << "]"; return os; diff --git a/paddle/fluid/operators/jit/kernel_base.h b/paddle/fluid/operators/jit/kernel_base.h index 4a8f61146a1921..20b6a32bef9860 100644 --- a/paddle/fluid/operators/jit/kernel_base.h +++ b/paddle/fluid/operators/jit/kernel_base.h @@ -13,6 +13,7 @@ * limitations under the License. */ #pragma once +#include #include "paddle/fluid/operators/jit/macro.h" #include "paddle/fluid/platform/macros.h" @@ -20,34 +21,35 @@ namespace paddle { namespace operators { namespace jit { -// TODO(TJ): reorder by alphabet typedef enum { kNone = 0, - kVMul = 1, - kVAdd = 2, - kVAddRelu, - kVSub, - kVScal, - kVAddBias, - kVRelu, - kVIdentity, - kVSquare, - kVExp, - kVSigmoid, - kVTanh, - kLSTMCtHt, - kLSTMC1H1, + // sort by alphabet + kCRFDecoding = 1, + kEmbSeqPool = 2, kGRUH1, kGRUHtPart1, kGRUHtPart2, - kCRFDecoding, + kHSum, // horizontal max + kHMax, // horizontal sum + kLSTMCtHt, + kLSTMC1H1, kLayerNorm, + kMatMul, kNCHW16CMulNC, kSeqPool, - kMatMul, - kHSum, // horizontal max - kHMax, // horizontal sum kSoftmax, + kVAdd, + kVAddBias, + kVAddRelu, + kVExp, + kVIdentity, + kVMul, + kVRelu, + kVScal, + kVSigmoid, + kVSquare, + kVSub, + kVTanh, } KernelType; typedef enum { @@ -145,6 +147,32 @@ struct SeqPoolTuples { typedef void (*func_type)(const T*, T*, const seq_pool_attr_t*); }; +typedef struct emb_seq_pool_attr_s { + int64_t table_height, table_width; + int64_t index_height, index_width; + int64_t out_width; + SeqPoolType pool_type; + emb_seq_pool_attr_s() = default; + explicit emb_seq_pool_attr_s(int64_t tbl_height, int64_t tbl_width, + int64_t idx_height, int64_t idx_width, + int64_t output_width, + SeqPoolType seqpool_type = SeqPoolType::kSum) + : table_height(tbl_height), + table_width(tbl_width), + index_height(idx_height), + index_width(idx_width), + out_width(output_width), + pool_type(seqpool_type) {} +} emb_seq_pool_attr_t; + +template +struct EmbSeqPoolTuples { + typedef T data_type; + typedef emb_seq_pool_attr_t attr_type; + typedef void (*func_type)(const T*, const int64_t*, T*, + const emb_seq_pool_attr_t*); +}; + typedef struct matmul_attr_s { int m, n, k; void* packed_weight{nullptr}; diff --git a/paddle/fluid/operators/jit/kernel_key.cc b/paddle/fluid/operators/jit/kernel_key.cc index 1e4a8884e78c5d..e659c6d254391f 100644 --- a/paddle/fluid/operators/jit/kernel_key.cc +++ b/paddle/fluid/operators/jit/kernel_key.cc @@ -56,6 +56,11 @@ size_t JitCodeKey(const matmul_attr_t& attr) { return (key << shift * 2) + ((static_cast(attr.n)) << shift) + attr.k; } +template <> +size_t JitCodeKey(const emb_seq_pool_attr_t& attr) { + return attr.table_width; +} + } // namespace jit } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/jit/more/mkl/CMakeLists.txt b/paddle/fluid/operators/jit/more/mkl/CMakeLists.txt index f9e5aea32e7cd4..d209f31007255b 100644 --- a/paddle/fluid/operators/jit/more/mkl/CMakeLists.txt +++ b/paddle/fluid/operators/jit/more/mkl/CMakeLists.txt @@ -13,3 +13,4 @@ USE_JITKERNEL_MORE(kVSigmoid, mkl) USE_JITKERNEL_MORE(kVTanh, mkl) USE_JITKERNEL_MORE(kSeqPool, mkl) USE_JITKERNEL_MORE(kSoftmax, mkl) +USE_JITKERNEL_MORE(kEmbSeqPool, mkl) diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.cc b/paddle/fluid/operators/jit/more/mkl/mkl.cc index 4c999131ab116e..29a451f832fa74 100644 --- a/paddle/fluid/operators/jit/more/mkl/mkl.cc +++ b/paddle/fluid/operators/jit/more/mkl/mkl.cc @@ -174,6 +174,16 @@ bool SeqPoolKernel::UseMe(const seq_pool_attr_t& attr) const { return true; } +template <> +bool EmbSeqPoolKernel::UseMe(const emb_seq_pool_attr_t& attr) const { + return true; +} + +template <> +bool EmbSeqPoolKernel::UseMe(const emb_seq_pool_attr_t& attr) const { + return true; +} + template <> bool MatMulKernel::UseMe(const matmul_attr_t& attr) const { return platform::MayIUse(platform::avx); @@ -227,6 +237,7 @@ REGISTER_MKL_KERNEL(kVSquare, VSquare); REGISTER_MKL_KERNEL(kVSigmoid, VSigmoid); REGISTER_MKL_KERNEL(kVTanh, VTanh); REGISTER_MKL_KERNEL(kSeqPool, SeqPool); +REGISTER_MKL_KERNEL(kEmbSeqPool, EmbSeqPool); REGISTER_MKL_KERNEL(kSoftmax, Softmax); #undef REGISTER_MKL_KERNEL diff --git a/paddle/fluid/operators/jit/more/mkl/mkl.h b/paddle/fluid/operators/jit/more/mkl/mkl.h index 8130b87326f188..9a72ba83022de2 100644 --- a/paddle/fluid/operators/jit/more/mkl/mkl.h +++ b/paddle/fluid/operators/jit/more/mkl/mkl.h @@ -18,6 +18,7 @@ #include #include #include "paddle/fluid/operators/jit/kernel_base.h" +#include "paddle/fluid/platform/enforce.h" namespace paddle { namespace operators { @@ -91,6 +92,32 @@ void SeqPool(const T* x, T* y, const seq_pool_attr_t* attr) { } } +template +void EmbSeqPool(const T* table, const int64_t* idx, T* out, + const emb_seq_pool_attr_t* attr) { + PADDLE_ENFORCE_EQ(attr->table_width * attr->index_width, attr->out_width); + auto check_idx_value_valid = [&](int64_t i) { + PADDLE_ENFORCE_LT(idx[i], attr->table_height, "idx value: %d, i: %d", + idx[i], i); + PADDLE_ENFORCE_GE(idx[i], 0, "idx value: %d, i: %d", idx[i], i); + }; + + for (int64_t w = 0; w != attr->index_width; ++w) { + check_idx_value_valid(w); + VCopy(table + idx[w] * attr->table_width, out + w * attr->table_width, + attr->table_width); + } + + for (int64_t h = 1; h < attr->index_height; ++h) { + for (int64_t w = 0; w < attr->index_width; ++w) { + int64_t i = h * attr->index_width + w; + check_idx_value_valid(i); + VAXPY(static_cast(1), table + idx[i] * attr->table_width, + out + w * attr->table_width, attr->table_width); + } + } +} + template void ASum(const T* x, T* res, int n); @@ -142,6 +169,8 @@ DECLARE_MKL_KERNEL(VSquare, XYNTuples); DECLARE_MKL_KERNEL(SeqPool, SeqPoolTuples); +DECLARE_MKL_KERNEL(EmbSeqPool, EmbSeqPoolTuples); + DECLARE_MKL_KERNEL(Softmax, SoftmaxTuples); #undef DECLARE_MKL_KERNEL diff --git a/paddle/fluid/operators/jit/refer/CMakeLists.txt b/paddle/fluid/operators/jit/refer/CMakeLists.txt index 9f2935828ca300..218d801c084be4 100644 --- a/paddle/fluid/operators/jit/refer/CMakeLists.txt +++ b/paddle/fluid/operators/jit/refer/CMakeLists.txt @@ -32,3 +32,4 @@ USE_JITKERNEL_REFER(kVSquare) USE_JITKERNEL_REFER(kHSum) USE_JITKERNEL_REFER(kHMax) USE_JITKERNEL_REFER(kSoftmax) +USE_JITKERNEL_REFER(kEmbSeqPool) diff --git a/paddle/fluid/operators/jit/refer/refer.cc b/paddle/fluid/operators/jit/refer/refer.cc index b8adb40ec7e1b6..7e7dd6960b66e4 100644 --- a/paddle/fluid/operators/jit/refer/refer.cc +++ b/paddle/fluid/operators/jit/refer/refer.cc @@ -57,4 +57,6 @@ REGISTER_REFER_KERNEL(kHSum, HSum); REGISTER_REFER_KERNEL(kSoftmax, Softmax); +REGISTER_REFER_KERNEL(kEmbSeqPool, EmbSeqPool); + #undef REGISTER_REFER_KERNEL diff --git a/paddle/fluid/operators/jit/refer/refer.h b/paddle/fluid/operators/jit/refer/refer.h index 0c4a985f8e8ece..fd1193aa41e50e 100644 --- a/paddle/fluid/operators/jit/refer/refer.h +++ b/paddle/fluid/operators/jit/refer/refer.h @@ -16,6 +16,7 @@ #include #include +#include #include "paddle/fluid/operators/jit/helper.h" #include "paddle/fluid/operators/jit/kernel_base.h" #include "paddle/fluid/platform/enforce.h" @@ -414,6 +415,37 @@ void Softmax(const T* x, T* y, int n, int bs = 1) { } } +// embedding seq pool +// table is a matrix with (tbl_h, tbl_w) +// idx is a matrix with (idx_h, idx_w) +// output is a vector with length tbl_w * idx_w +template +void EmbSeqPool(const T* table, const int64_t* idx, T* out, + const emb_seq_pool_attr_t* attr) { + PADDLE_ENFORCE_EQ(attr->table_width * attr->index_width, attr->out_width); + + auto check_idx_value_valid = [&](int64_t i) { + PADDLE_ENFORCE_LT(idx[i], attr->table_height, "idx value: %d, i: %d", + idx[i], i); + PADDLE_ENFORCE_GE(idx[i], 0, "idx value: %d, i: %d", idx[i], i); + }; + + for (int64_t w = 0; w != attr->index_width; ++w) { + check_idx_value_valid(w); + std::memcpy(out + w * attr->table_width, table + idx[w] * attr->table_width, + attr->table_width * sizeof(T)); + } + + for (int64_t h = 1; h < attr->index_height; ++h) { + for (int64_t w = 0; w < attr->index_width; ++w) { + int64_t i = h * attr->index_width + w; + check_idx_value_valid(i); + VAdd(table + idx[i] * attr->table_width, out + w * attr->table_width, + out + w * attr->table_width, attr->table_width); + } + } +} + #define DECLARE_REFER_KERNEL(name, tuples) \ template \ class name##Kernel : public ReferKernel> { \ @@ -462,6 +494,8 @@ DECLARE_REFER_KERNEL(HSum, XRNTuples); DECLARE_REFER_KERNEL(Softmax, SoftmaxTuples); +DECLARE_REFER_KERNEL(EmbSeqPool, EmbSeqPoolTuples); + #undef DECLARE_REFER_KERNEL } // namespace refer diff --git a/paddle/fluid/operators/jit/test.cc b/paddle/fluid/operators/jit/test.cc index 85b50b79d95070..2632bfb6de1751 100644 --- a/paddle/fluid/operators/jit/test.cc +++ b/paddle/fluid/operators/jit/test.cc @@ -270,6 +270,32 @@ struct TestFuncWithRefer, std::vector, std::vector, } }; +template +struct TestFuncWithRefer, std::vector, + std::vector, std::vector, + typename jit::EmbSeqPoolTuples::attr_type> { + void operator()(const typename jit::EmbSeqPoolTuples::func_type tgt, + const std::vector& table, const std::vector& idx, + const std::vector& oref, + const typename jit::EmbSeqPoolTuples::attr_type& attr) { + EXPECT_TRUE(tgt != nullptr); + EXPECT_EQ(table.size(), + static_cast(attr.table_height * attr.table_width)); + EXPECT_EQ(idx.size(), + static_cast(attr.index_height * attr.index_width)); + EXPECT_EQ(oref.size(), + static_cast(attr.table_width * attr.index_width)); + const T* table_data = table.data(); + const int64_t* idx_data = idx.data(); + const T* oref_data = oref.data(); + int o_w = oref.size(); + std::vector out(o_w); + T* o_data = out.data(); + tgt(table_data, idx_data, o_data, &attr); + ExpectEQ(o_data, oref_data, o_w); + } +}; + template struct TestFuncWithRefer, std::vector, std::vector, std::vector, @@ -644,6 +670,40 @@ void TestSoftmaxKernel() { } } +template +void TestEmbSeqPoolKernel() { + VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); + int64_t tbl_h = 1e4; + std::vector pool_types = { + jit::SeqPoolType::kSum}; // only support sum yet + for (int tbl_w : TestSizes()) { + std::vector table(tbl_h * tbl_w); + RandomVec(tbl_h * tbl_w, table.data(), -2.f, 2.f); + const T* table_data = table.data(); + for (auto type : pool_types) { + for (int idx_w : {1, 2, 10, 16}) { + for (int idx_h : {1, 2, 9, 13, 16}) { + auto ref = jit::GetRefer>(); + EXPECT_TRUE(ref != nullptr); + std::vector idx(idx_h * idx_w); + RandomVec(idx_h * idx_w, idx.data(), 0, tbl_h - 1); + int64_t out_w = tbl_w * idx_w; + std::vector oref(out_w); + const int64_t* idx_data = idx.data(); + T* o_data = oref.data(); + jit::emb_seq_pool_attr_t attr(tbl_h, tbl_w, idx_h, idx_w, out_w, + type); + ref(table_data, idx_data, o_data, &attr); + + TestAllImpls, PlaceType, std::vector, + std::vector, std::vector>(attr, table, idx, + oref, attr); + } + } + } + } +} + template void TestNCHW16CMulNCKernel() { VLOG(10) << "===== Test JITKernel " << jit::to_string(KT); @@ -878,6 +938,11 @@ TEST(JITKernel, kSoftmax) { TestSoftmaxKernel(); } +TEST(JITKernel, kEmbSeqPool) { + TestEmbSeqPoolKernel(); + TestEmbSeqPoolKernel(); +} + TEST(JITKernel, kNCHW16CMulNC) { TestNCHW16CMulNCKernel(); TestNCHW16CMulNCKernel(); diff --git a/paddle/fluid/operators/layer_norm_op.cc b/paddle/fluid/operators/layer_norm_op.cc index f83fe355b85566..b9db6daf0825b5 100644 --- a/paddle/fluid/operators/layer_norm_op.cc +++ b/paddle/fluid/operators/layer_norm_op.cc @@ -44,11 +44,11 @@ class LayerNormOp : public framework::OperatorWithKernel { int left = static_cast(matrix_dim[0]); int right = static_cast(matrix_dim[1]); if (ctx->HasInput("Scale")) { - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], right); } if (ctx->HasInput("Bias")) { - PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1); PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias")[0], right); } diff --git a/paddle/fluid/operators/linear_chain_crf_op.cc b/paddle/fluid/operators/linear_chain_crf_op.cc index 1da14631e35608..e17b6cb5989852 100644 --- a/paddle/fluid/operators/linear_chain_crf_op.cc +++ b/paddle/fluid/operators/linear_chain_crf_op.cc @@ -144,12 +144,12 @@ class LinearChainCRFOp : public framework::OperatorWithKernel { "Output(LogLikelihood) should be not null."); auto emission_dims = ctx->GetInputDim("Emission"); - PADDLE_ENFORCE_EQ(emission_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(emission_dims.size(), 2, "The Input(Emission) should be a 2-D tensor."); PADDLE_ENFORCE(emission_dims[0], "An empty mini-batch is not allowed."); auto transition_dims = ctx->GetInputDim("Transition"); - PADDLE_ENFORCE_EQ(transition_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(transition_dims.size(), 2, "The Input(Transition) should be a 2-D tensor."); PADDLE_ENFORCE_EQ( transition_dims[0] - 2, transition_dims[1], @@ -202,13 +202,13 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel { "Input(LogLikelihood@GRAD) shoudl be not null."); auto emission_exps_dims = ctx->GetInputDim("EmissionExps"); - PADDLE_ENFORCE_EQ(emission_exps_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(emission_exps_dims.size(), 2, "The Input(EmissionExps) should be a 2-D tensor."); PADDLE_ENFORCE(emission_exps_dims[0], "An empty mini-batch is not allowed."); auto transition_exps_dims = ctx->GetInputDim("TransitionExps"); - PADDLE_ENFORCE_EQ(transition_exps_dims.size(), 2UL, + PADDLE_ENFORCE_EQ(transition_exps_dims.size(), 2, "The Input(TransitionExps) should be a 2-D tensor."); PADDLE_ENFORCE_EQ( transition_exps_dims[0] - 2, transition_exps_dims[1], diff --git a/paddle/fluid/operators/ngraph/ngraph_bridge.cc b/paddle/fluid/operators/ngraph/ngraph_bridge.cc index 08d72a5b397809..4bfcba6c3ce312 100644 --- a/paddle/fluid/operators/ngraph/ngraph_bridge.cc +++ b/paddle/fluid/operators/ngraph/ngraph_bridge.cc @@ -36,11 +36,14 @@ std::map("epsilon"); const float momentum = op_attrs.Get("momentum"); + PADDLE_ENFORCE( + data_layout == "NHWC" || data_layout == "NCHW" || data_layout == "NC", + "The BatchNorm operator only supports NHWC/NCHW/NC data format"); + if (data_layout == "NHWC") { x = paddle::platform::Nhwc2Nchw(x); } @@ -110,6 +114,9 @@ void BuildBatchNormGradNode( "BN grap input size needs to be 2 or 4"); PADDLE_ENFORCE_EQ(x_shape.size(), dy_shape.size(), "BN grap input and delta size needs to be equal"); + PADDLE_ENFORCE( + data_layout == "NHWC" || data_layout == "NCHW" || data_layout == "NC", + "The BatchNorm operator only supports NHWC/NCHW/NC data format"); if (x_shape.size() == 2) { x = std::make_shared( diff --git a/paddle/fluid/operators/ngraph/ops/cross_entropy_op.h b/paddle/fluid/operators/ngraph/ops/cross_entropy_op.h new file mode 100644 index 00000000000000..f88a2cb94103b2 --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/cross_entropy_op.h @@ -0,0 +1,145 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include + +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildCrossEntropyNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map); + auto label = paddle::platform::GetInputNode(op, "Label", ngb_node_map); + auto label_shape = label->get_shape(); + auto x_shape = x->get_shape(); + auto label_rank = label_shape.size(); + auto x_rank = x_shape.size(); + std::shared_ptr x_2d = x, label_2d = label; + auto label_2d_shape = label_shape, x_2d_shape = x_shape; + + if (label_rank > 2) { + label_2d_shape = paddle::platform::FlattenTo2d(label_shape, label_rank - 1); + label_2d = paddle::platform::NgReshaper(label, label_2d_shape); + } + if (x_rank > 2) { + x_2d_shape = paddle::platform::FlattenTo2d(x_shape, x_rank - 1); + x_2d = paddle::platform::NgReshaper(x, x_2d_shape); + } + + auto batch_size = x_2d_shape.at(0); + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + const bool is_soft_label = op_attrs.Get("soft_label"); + + std::shared_ptr node_1_hot = label_2d; + if (!is_soft_label) { + auto label_1d = paddle::platform::NgReshaper( + label_2d, ngraph::Shape{label_2d_shape.at(0)}); + node_1_hot = std::make_shared(label_1d, x_2d_shape, 1); + } + if (x->get_element_type() != node_1_hot->get_element_type()) { + node_1_hot = std::make_shared(node_1_hot, + x->get_element_type()); + } + + auto node_log = std::make_shared(x_2d); + auto high_clip = ngraph::op::Constant::create(node_log->get_element_type(), + node_log->get_shape(), {1e20}); + auto low_clip = ngraph::op::Constant::create(node_log->get_element_type(), + node_log->get_shape(), {-1e20}); + auto node_min = std::make_shared(node_log, high_clip); + auto node_max = std::make_shared(node_min, low_clip); + auto node_mul = node_1_hot * node_log; + auto node_sum = + std::make_shared(node_mul, ngraph::AxisSet{1}); + auto node_neg = std::make_shared(node_sum); + auto xe = + paddle::platform::NgReshaper(node_neg, ngraph::Shape{batch_size, 1}); + + if (!is_soft_label) { + auto ignore_index = op_attrs.Get("ignore_index"); + auto ignore_node = ngraph::op::Constant::create( + label->get_element_type(), label_2d_shape, {ignore_index}); + auto not_equal_node = + std::make_shared(label_2d, ignore_node); + auto mask = std::make_shared(not_equal_node, + xe->get_element_type()); + xe = xe * mask; + } + + paddle::platform::SetOutputNode(op, "Y", xe, ngb_node_map); +} + +void BuildCrossEntropyGradNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + const bool is_soft_label = op_attrs.Get("soft_label"); + + auto x = paddle::platform::GetInputNode(op, "X", ngb_node_map); + auto label = paddle::platform::GetInputNode(op, "Label", ngb_node_map); + auto dy = paddle::platform::GetInputNode(op, "Y@GRAD", ngb_node_map); + auto x_shape = x->get_shape(); + auto rank = x_shape.size(); + + std::shared_ptr mask; + if (!is_soft_label) { + auto label_shape = label->get_shape(); + label_shape.pop_back(); + label = paddle::platform::NgReshaper(label, label_shape); + + auto ignore_index = op_attrs.Get("ignore_index"); + auto ignore_node = ngraph::op::Constant::create( + label->get_element_type(), label_shape, {ignore_index}); + auto not_equal_node = + std::make_shared(label, ignore_node); + mask = std::make_shared(not_equal_node, + x->get_element_type()); + mask = std::make_shared(mask, x_shape, + ngraph::AxisSet{rank - 1}); + + label = std::make_shared(label, x_shape, rank - 1); + } + + auto dy_shape = dy->get_shape(); + dy_shape.pop_back(); + auto dy_reshape = paddle::platform::NgReshaper(dy, dy_shape); + auto dy_bcast = std::make_shared( + dy_reshape, x_shape, ngraph::AxisSet{rank - 1}); + if (x->get_element_type() != label->get_element_type()) { + label = std::make_shared(label, x->get_element_type()); + } + + auto xe_grad = -label * dy_bcast / x; + + if (!is_soft_label) { + xe_grad = xe_grad * mask; + } + + paddle::platform::SetOutputNode(op, "X@GRAD", xe_grad, ngb_node_map); +} +} // namespace ngraphs +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/ngraph/ops/fill_constant_op.h b/paddle/fluid/operators/ngraph/ops/fill_constant_op.h index 406a4314f89810..58783bc220fa02 100644 --- a/paddle/fluid/operators/ngraph/ops/fill_constant_op.h +++ b/paddle/fluid/operators/ngraph/ops/fill_constant_op.h @@ -46,8 +46,6 @@ void BuildFillConstantNode( ng_dtype = ngraph::element::i64; } else if (data_type == paddle::framework::proto::VarType::INT32) { ng_dtype = ngraph::element::i32; - } else if (data_type == paddle::framework::proto::VarType::BOOL) { - ng_dtype = ngraph::element::boolean; } else { PADDLE_THROW("unsupported data type: %s", data_type); } diff --git a/paddle/fluid/operators/ngraph/ops/momentum_op.h b/paddle/fluid/operators/ngraph/ops/momentum_op.h new file mode 100644 index 00000000000000..f1b365c488d31c --- /dev/null +++ b/paddle/fluid/operators/ngraph/ops/momentum_op.h @@ -0,0 +1,101 @@ +/*Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include "ngraph/ngraph.hpp" +#include "paddle/fluid/platform/ngraph_helper.h" + +namespace paddle { +namespace operators { +namespace ngraphs { + +void BuildMomentumNode( + const std::shared_ptr& op, + std::shared_ptr< + std::unordered_map>> + ngb_node_map) { + auto op_attrs = paddle::framework::AttrReader(op->Attrs()); + auto param = paddle::platform::GetInputNode(op, "Param", ngb_node_map); + auto grad = paddle::platform::GetInputNode(op, "Grad", ngb_node_map); + auto velocity = paddle::platform::GetInputNode(op, "Velocity", ngb_node_map); + auto learning_rate = + paddle::platform::GetInputNode(op, "LearningRate", ngb_node_map); + + auto mu = op_attrs.Get("mu"); + bool use_nesterov = op_attrs.Get("use_nesterov"); + + auto param_shape = param->get_shape(); + auto velocity_shape = velocity->get_shape(); + auto grad_shape = grad->get_shape(); + auto lr_shape = learning_rate->get_shape(); + + auto shape_velocity = ngraph::Shape{velocity_shape}; + auto mu_create = + ngraph::op::Constant::create(ngraph::element::f32, shape_velocity, {mu}); + + auto vel_mul = std::make_shared(velocity, mu_create); + auto vel_out = std::make_shared(vel_mul, grad); + + ngraph::NodeVector result; + if (use_nesterov) { + auto mul_res = std::make_shared(vel_out, mu_create); + auto add_res = std::make_shared(grad, mul_res); + + auto add_2d = paddle::platform::FlattenTo2d(add_res->get_shape(), 0); + auto vel_reshape = paddle::platform::NgReshaper(vel_out, add_2d); + + auto lr_bcast = std::make_shared( + learning_rate, vel_reshape->get_shape(), + ngraph::AxisSet{vel_reshape->get_shape().size() - 1}); + + auto lr_1d = paddle::platform::FlattenTo1d(lr_bcast->get_shape(), 0); + auto lr_reshape = std::make_shared( + lr_bcast, ngraph::AxisVector{0, 1}, lr_1d); + + lr_reshape = std::make_shared( + lr_reshape, ngraph::AxisVector{0}, param->get_shape()); + + auto mul_res1 = std::make_shared(add_res, lr_reshape); + auto res = std::make_shared(param, mul_res1); + paddle::platform::SetOutputNode(op, "ParamOut", res, ngb_node_map); + } else { + auto vel_2d = paddle::platform::FlattenTo2d(vel_out->get_shape(), 0); + auto vel_reshape = paddle::platform::NgReshaper(vel_out, vel_2d); + + auto lr_bcast = std::make_shared( + learning_rate, vel_reshape->get_shape(), + ngraph::AxisSet{vel_reshape->get_shape().size() - 1}); + + auto lr_1d = paddle::platform::FlattenTo1d(lr_bcast->get_shape(), 0); + auto lr_reshape = std::make_shared( + lr_bcast, ngraph::AxisVector{0, 1}, lr_1d); + + lr_reshape = std::make_shared( + lr_reshape, ngraph::AxisVector{0}, param->get_shape()); + + auto mul_result = + std::make_shared(lr_reshape, vel_out); + + auto res = std::make_shared(param, mul_result); + paddle::platform::SetOutputNode(op, "ParamOut", res, ngb_node_map); + } + paddle::platform::SetOutputNode(op, "VelocityOut", vel_out, ngb_node_map); +} + +} // namespace ngraphs +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reader/read_op.cc b/paddle/fluid/operators/reader/read_op.cc index 8fe638ac2fdc6e..846b2ed77e46d8 100644 --- a/paddle/fluid/operators/reader/read_op.cc +++ b/paddle/fluid/operators/reader/read_op.cc @@ -85,9 +85,7 @@ class ReadOp : public framework::OperatorBase { std::vector ins; // For profiling - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(dev_place); - platform::RecordEvent record_event(Type(), &ctx); + platform::RecordEvent record_event(Type()); reader->ReadNext(&ins); if (ins.empty()) { diff --git a/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc b/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc index 1eebadc2c980dd..0932211cadf30d 100644 --- a/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc +++ b/paddle/fluid/operators/sequence_ops/sequence_enumerate_op.cc @@ -31,10 +31,10 @@ class SequenceEnumerateOp : public framework::OperatorWithKernel { const auto x_dims = ctx->GetInputDim("X"); PADDLE_ENFORCE_EQ( - x_dims.size(), 2UL, + x_dims.size(), 2, "Input(X) of SequenceEnumerate operator's rank should be 2."); PADDLE_ENFORCE_EQ( - x_dims[1], 1UL, + x_dims[1], 1, "Input(X) of SequenceEnumerate operator's 2nd dimension should be 1."); const auto win_size = ctx->Attrs().Get("win_size"); diff --git a/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc b/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc index 27e0201bd70df5..f6c42415301bc8 100644 --- a/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc +++ b/paddle/fluid/operators/sequence_ops/sequence_expand_op.cc @@ -48,10 +48,10 @@ class SequenceExpandOp : public framework::OperatorWithKernel { auto& x_lod = x_var->Get().lod(); auto& y_lod = y_var->Get().lod(); - PADDLE_ENFORCE_LE(x_lod.size(), 1, + PADDLE_ENFORCE_LE(x_lod.size(), 1UL, "Level number of Input(X)'s lod should not be " "greater than 1."); - PADDLE_ENFORCE_GT(y_lod.size(), 0, + PADDLE_ENFORCE_GT(y_lod.size(), 0UL, "Level number of Input(Y)'s lod should be " "greater than 0."); PADDLE_ENFORCE( @@ -69,7 +69,8 @@ class SequenceExpandOp : public framework::OperatorWithKernel { "size of Input(X)'s first level lod should be equal to " "size of Input(Y)'s referred level lod."); } else { - PADDLE_ENFORCE_EQ(x_dims[0], y_lod[ref_level].size() - 1, + PADDLE_ENFORCE_EQ(x_dims[0], + static_cast(y_lod[ref_level].size()) - 1, "When Input(X)'s lod is null, the dims[0] of " "Input(X) should match the " "size of Input(Y)'s referred level lod."); diff --git a/paddle/fluid/operators/shape_op.cc b/paddle/fluid/operators/shape_op.cc index 1be9fe47af71d3..efc497fa47d1d9 100644 --- a/paddle/fluid/operators/shape_op.cc +++ b/paddle/fluid/operators/shape_op.cc @@ -35,14 +35,15 @@ class ShapeOp : public framework::OperatorWithKernel { class ShapeOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Input", "(Tensor), The input tensor."); - AddOutput("Out", - "(Tensor), The shape of input tensor, the data type of the shape" - " is int32_t, will be on the same device with the input Tensor."); + AddInput("Input", "(LoDTensor), The input tensor."); + AddOutput( + "Out", + "(LoDTensor), The shape of input tensor, the data type of the shape" + " is int32_t, will be on the same device with the input Tensor."); AddComment(R"DOC( -Shape Operator +Shape Operator. -Get the shape of input tensor. Only support CPU input Tensor now. +Return the shape of the input. )DOC"); } }; diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt index fbb2ac3fe8c5de..5833fee35b14d6 100644 --- a/paddle/fluid/platform/CMakeLists.txt +++ b/paddle/fluid/platform/CMakeLists.txt @@ -36,7 +36,7 @@ cc_test(cpu_info_test SRCS cpu_info_test.cc DEPS cpu_info) nv_library(gpu_info SRCS gpu_info.cc DEPS gflags glog enforce) -cc_library(place SRCS place.cc DEPS enforce boost lib_any) +cc_library(place SRCS place.cc DEPS enforce boost) cc_test(place_test SRCS place_test.cc DEPS place glog gflags) add_subdirectory(dynload) @@ -88,7 +88,11 @@ cc_library(timer SRCS timer.cc) cc_test(timer_test SRCS timer_test.cc DEPS timer) cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto framework_proto ${GPU_CTX_DEPS}) -cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) +if(WITH_GPU) + nv_library(profiler SRCS profiler.cc profiler.cu DEPS device_context device_tracer) +else() + cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) +endif() cc_test(profiler_test SRCS profiler_test.cc DEPS profiler) nv_test(float16_gpu_test SRCS float16_test.cu DEPS lod_tensor) diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index 2493fb71c019f9..ed0dbdeb13ce93 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -291,7 +291,7 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace place) if (dynload::HasCUDNN()) { auto local_cudnn_version = cudnn_dso_ver / 100; auto compile_cudnn_version = CUDNN_VERSION / 100; - if (local_cudnn_version < compile_cudnn_version) { + if (local_cudnn_version < static_cast(compile_cudnn_version)) { LOG_FIRST_N(WARNING, 1) << "WARNING: device: " << place_.device << ". The installed Paddle is compiled with CUDNN " diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc index 0a4563ead65b1e..f42212d09508ce 100644 --- a/paddle/fluid/platform/device_tracer.cc +++ b/paddle/fluid/platform/device_tracer.cc @@ -14,17 +14,23 @@ limitations under the License. */ #include "paddle/fluid/platform/device_tracer.h" #include +#include #include +#include #include #include // NOLINT #include +#include #include #include // NOLINT +#include +#include #include #include "glog/logging.h" #include "google/protobuf/text_format.h" #include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/string/printf.h" namespace paddle { @@ -33,17 +39,31 @@ namespace { // Tracking the nested block stacks of each thread. thread_local std::deque block_id_stack; // Tracking the nested event stacks. -thread_local std::deque annotation_stack; +thread_local std::deque annotation_stack; + +std::map system_thread_id_map; std::once_flag tracer_once_flag; DeviceTracer *tracer = nullptr; + +void PrintCuptiHint() { + static bool showed = false; + if (showed) return; + showed = true; + LOG(WARNING) << "Invalid timestamp occured. Please try increasing the " + "FLAGS_multiple_of_cupti_buffer_size."; +} + } // namespace #ifdef PADDLE_WITH_CUPTI namespace { -// TODO(panyx0718): Revisit the buffer size here. -uint64_t kBufSize = 32 * 1024; +// The experimental best performance is +// the same size with CUPTI device buffer size(8M) +uint64_t kBufSize = 1024 * 1024 * 8; uint64_t kAlignSize = 8; +std::unordered_map runtime_cbid_str, + driver_cbid_str; #define ALIGN_BUFFER(buffer, align) \ (((uintptr_t)(buffer) & ((align)-1)) \ @@ -92,15 +112,33 @@ std::string MemcpyKind(CUpti_ActivityMemcpyKind kind) { return "MEMCPY"; } +std::string DriverKind(CUpti_CallbackId cbid) { + auto iter = driver_cbid_str.find(cbid); + if (iter == driver_cbid_str.end()) + return "Driver API " + std::to_string(cbid); + return iter->second; +} + +std::string RuntimeKind(CUpti_CallbackId cbid) { + auto iter = runtime_cbid_str.find(cbid); + if (iter == runtime_cbid_str.end()) + return "Runtime API " + std::to_string(cbid); + return iter->second; +} + void EnableActivity() { // Device activity record is created when CUDA initializes, so we // want to enable it before cuInit() or any CUDA runtime call. CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMCPY)); - CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL)); - CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DEVICE)); - CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMSET)); - CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_OVERHEAD)); + CUPTI_CALL( + dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL)); + // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL)); + CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DRIVER)); + CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_RUNTIME)); // We don't track these activities for now. + // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMSET)); + // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_OVERHEAD)); + // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DEVICE)); // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONTEXT)); // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_DRIVER)); // CUPTI_CALL(dynload::cuptiActivityEnable(CUPTI_ACTIVITY_KIND_RUNTIME)); @@ -110,16 +148,17 @@ void EnableActivity() { void DisableActivity() { CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMCPY)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_KERNEL)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DEVICE)); + CUPTI_CALL( + dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DEVICE)); // Disable all other activity record kinds. - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_CONTEXT)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_CONTEXT)); CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_DRIVER)); CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_RUNTIME)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMSET)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_NAME)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MARKER)); - CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_OVERHEAD)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MEMSET)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_NAME)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_MARKER)); + // CUPTI_CALL(dynload::cuptiActivityDisable(CUPTI_ACTIVITY_KIND_OVERHEAD)); } void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size, @@ -132,6 +171,11 @@ void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size, void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, size_t size, size_t validSize) { + static std::thread::id cupti_thread_id(0); + if (cupti_thread_id == std::thread::id(0)) + cupti_thread_id = std::this_thread::get_id(); + PADDLE_ENFORCE_EQ(std::this_thread::get_id(), cupti_thread_id, + "Only one thread is allowed to call bufferCompleted()"); CUptiResult status; CUpti_Activity *record = NULL; if (validSize > 0) { @@ -168,6 +212,23 @@ void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, memcpy->correlationId, memcpy->bytes); break; } + case CUPTI_ACTIVITY_KIND_DRIVER: { + auto *api = reinterpret_cast(record); + if (api->start != 0 && api->end != 0) + // -1 device id represents CUDA api call + tracer->AddCPURecords( + DriverKind(api->cbid), api->start, api->end, -1, + GetThreadIdFromSystemThreadId(api->threadId)); + break; + } + case CUPTI_ACTIVITY_KIND_RUNTIME: { + auto *api = reinterpret_cast(record); + if (api->start != 0 && api->end != 0) + tracer->AddCPURecords( + RuntimeKind(api->cbid), api->start, api->end, -1, + GetThreadIdFromSystemThreadId(api->threadId)); + break; + } default: { break; } } } else if (status == CUPTI_ERROR_MAX_LIMIT_REACHED) { @@ -183,21 +244,35 @@ void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId, uint8_t *buffer, dynload::cuptiActivityGetNumDroppedRecords(ctx, streamId, &dropped)); if (dropped != 0) { fprintf(stderr, "Dropped %u activity records\n", (unsigned int)dropped); + PrintCuptiHint(); } } free(buffer); } + +void initCuptiCbidStr(); + } // namespace #endif // PADDLE_WITH_CUPTI class DeviceTracerImpl : public DeviceTracer { public: - DeviceTracerImpl() : enabled_(false) {} + DeviceTracerImpl() : enabled_(false) { +#ifdef PADDLE_WITH_CUPTI + initCuptiCbidStr(); +#endif + } - void AddAnnotation(uint64_t id, const std::string &anno) { - std::lock_guard l(trace_mu_); - correlations_[id] = anno; + void AddAnnotation(uint32_t id, Event *event) { + thread_local std::forward_list> + *local_correlations_pairs = nullptr; + if (local_correlations_pairs == nullptr) { + std::lock_guard l(trace_mu_); + correlations_pairs.emplace_front(); + local_correlations_pairs = &correlations_pairs.front(); + } + local_correlations_pairs->push_front(std::make_pair(id, event)); } void AddCPURecords(const std::string &anno, uint64_t start_ns, @@ -206,8 +281,13 @@ class DeviceTracerImpl : public DeviceTracer { VLOG(1) << "Empty timeline annotation."; return; } - std::lock_guard l(trace_mu_); - cpu_records_.push_back( + thread_local std::forward_list *local_cpu_records_ = nullptr; + if (local_cpu_records_ == nullptr) { + std::lock_guard l(trace_mu_); + cpu_records_.emplace_front(); + local_cpu_records_ = &cpu_records_.front(); + } + local_cpu_records_->push_front( CPURecord{anno, start_ns, end_ns, device_id, thread_id}); } @@ -215,25 +295,27 @@ class DeviceTracerImpl : public DeviceTracer { uint64_t end_ns, int64_t device_id, int64_t stream_id, uint32_t correlation_id, uint64_t bytes) { // 0 means timestamp information could not be collected for the kernel. - if (start_ns == 0 || end_ns == 0) { + if (start_ns == 0 || end_ns == 0 || start_ns == end_ns) { VLOG(3) << name << " cannot be traced"; + PrintCuptiHint(); return; } - std::lock_guard l(trace_mu_); - mem_records_.push_back(MemRecord{name, start_ns, end_ns, device_id, - stream_id, correlation_id, bytes}); + // NOTE(liangdun): lock is not needed, only one thread call this function. + mem_records_.push_front(MemRecord{name, start_ns, end_ns, device_id, + stream_id, correlation_id, bytes}); } void AddKernelRecords(std::string name, uint64_t start, uint64_t end, int64_t device_id, int64_t stream_id, uint32_t correlation_id) { // 0 means timestamp information could not be collected for the kernel. - if (start == 0 || end == 0) { + if (start == 0 || end == 0 || start == end) { VLOG(3) << correlation_id << " cannot be traced"; + PrintCuptiHint(); return; } - std::lock_guard l(trace_mu_); - kernel_records_.push_back( + // NOTE(liangdun): lock is not needed, only one thread call this function. + kernel_records_.push_front( KernelRecord{name, start, end, device_id, stream_id, correlation_id}); } @@ -263,25 +345,80 @@ class DeviceTracerImpl : public DeviceTracer { } else if (ret != CUPTI_SUCCESS) { fprintf(stderr, "Failed to create CUPTI subscriber.\n"); } - CUPTI_CALL( - dynload::cuptiEnableCallback(1, subscriber_, CUPTI_CB_DOMAIN_DRIVER_API, - CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel)); + const std::vector cbids { + CUPTI_RUNTIME_TRACE_CBID_cudaMemcpy_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaMemcpyAsync_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaLaunch_v3020, + CUPTI_RUNTIME_TRACE_CBID_cudaLaunchKernel_v7000 +#if CUDA_VERSION >= 9000 + , + CUPTI_RUNTIME_TRACE_CBID_cudaLaunchCooperativeKernel_v9000, + CUPTI_RUNTIME_TRACE_CBID_cudaLaunchCooperativeKernelMultiDevice_v9000 +#endif + }; + for (auto cbid : cbids) + CUPTI_CALL(dynload::cuptiEnableCallback( + 1, subscriber_, CUPTI_CB_DOMAIN_RUNTIME_API, cbid)); CUPTI_CALL(dynload::cuptiGetTimestamp(&start_ns_)); #endif // PADDLE_WITH_CUPTI enabled_ = true; } + void Reset() { +#ifdef PADDLE_WITH_CUPTI + CUPTI_CALL( + dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED)); +#endif + std::lock_guard l(trace_mu_); + kernel_records_.clear(); + mem_records_.clear(); + correlations_.clear(); + for (auto &tmp : correlations_pairs) tmp.clear(); + for (auto &tmp : cpu_records_) tmp.clear(); + } + + void GenEventKernelCudaElapsedTime() { +#ifdef PADDLE_WITH_CUPTI + if (correlations_.empty()) + for (auto &tmp : correlations_pairs) + for (auto &pair : tmp) correlations_[pair.first] = pair.second; + for (const KernelRecord &r : kernel_records_) { + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + Event *e = c->second; + e->AddCudaElapsedTime(r.start_ns, r.end_ns); + } + } + for (const auto &r : mem_records_) { + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + Event *e = c->second; + e->AddCudaElapsedTime(r.start_ns, r.end_ns); + } + } +#endif + } + proto::Profile GenProfile(const std::string &profile_path) { + int miss = 0, find = 0; std::lock_guard l(trace_mu_); proto::Profile profile_pb; profile_pb.set_start_ns(start_ns_); profile_pb.set_end_ns(end_ns_); + if (correlations_.empty()) + for (auto &tmp : correlations_pairs) + for (auto &pair : tmp) correlations_[pair.first] = pair.second; for (const KernelRecord &r : kernel_records_) { auto *event = profile_pb.add_events(); event->set_type(proto::Event::GPUKernel); - if (correlations_.find(r.correlation_id) != correlations_.end()) { - event->set_name(correlations_.at(r.correlation_id)); + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + event->set_name(c->second->name()); + event->set_detail_info(r.name); + find++; } else { + VLOG(10) << "Missing Kernel Event: " + r.name; + miss++; event->set_name(r.name); } event->set_start_ns(r.start_ns); @@ -289,31 +426,41 @@ class DeviceTracerImpl : public DeviceTracer { event->set_sub_device_id(r.stream_id); event->set_device_id(r.device_id); } - - for (const CPURecord &r : cpu_records_) { - auto *event = profile_pb.add_events(); - event->set_type(proto::Event::CPU); - event->set_name(r.name); - event->set_start_ns(r.start_ns); - event->set_end_ns(r.end_ns); - event->set_sub_device_id(r.thread_id); - event->set_device_id(r.device_id); - } + VLOG(1) << "KernelRecord event miss: " << miss << " find: " << find; + for (auto &tmp : cpu_records_) + for (const CPURecord &r : tmp) { + auto *event = profile_pb.add_events(); + event->set_type(proto::Event::CPU); + event->set_name(r.name); + event->set_start_ns(r.start_ns); + event->set_end_ns(r.end_ns); + event->set_sub_device_id(r.thread_id); + event->set_device_id(r.device_id); + } + miss = find = 0; for (const MemRecord &r : mem_records_) { auto *event = profile_pb.add_events(); event->set_type(proto::Event::GPUKernel); - event->set_name(r.name); + auto c = correlations_.find(r.correlation_id); + if (c != correlations_.end() && c->second != nullptr) { + event->set_name(c->second->name()); + event->set_detail_info(r.name); + find++; + } else { + miss++; + event->set_name(r.name); + } event->set_start_ns(r.start_ns); event->set_end_ns(r.end_ns); event->set_sub_device_id(r.stream_id); event->set_device_id(r.device_id); event->mutable_memcopy()->set_bytes(r.bytes); } + VLOG(1) << "MemRecord event miss: " << miss << " find: " << find; std::ofstream profile_f; - profile_f.open(profile_path, std::ios::out | std::ios::trunc); - std::string profile_str; - profile_pb.SerializeToString(&profile_str); - profile_f << profile_str; + profile_f.open(profile_path, + std::ios::out | std::ios::trunc | std::ios::binary); + profile_pb.SerializeToOstream(&profile_f); profile_f.close(); return profile_pb; } @@ -321,12 +468,13 @@ class DeviceTracerImpl : public DeviceTracer { void Disable() { #ifdef PADDLE_WITH_CUPTI // flush might cause additional calls to DeviceTracker. - dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED); + CUPTI_CALL( + dynload::cuptiActivityFlushAll(CUPTI_ACTIVITY_FLAG_FLUSH_FORCED)); #endif // PADDLE_WITH_CUPTI std::lock_guard l(trace_mu_); #ifdef PADDLE_WITH_CUPTI DisableActivity(); - dynload::cuptiUnsubscribe(subscriber_); + CUPTI_CALL(dynload::cuptiUnsubscribe(subscriber_)); CUPTI_CALL(dynload::cuptiGetTimestamp(&end_ns_)); #endif // PADDLE_WITH_CUPTI enabled_ = false; @@ -337,18 +485,10 @@ class DeviceTracerImpl : public DeviceTracer { static void CUPTIAPI ApiCallback(void *userdata, CUpti_CallbackDomain domain, CUpti_CallbackId cbid, const void *cbdata) { auto *cbInfo = reinterpret_cast(cbdata); - DeviceTracer *tracer = reinterpret_cast(userdata); - - if ((domain == CUPTI_CB_DOMAIN_DRIVER_API) && - (cbid == CUPTI_DRIVER_TRACE_CBID_cuLaunchKernel)) { - if (cbInfo->callbackSite == CUPTI_API_ENTER) { - const std::string anno = !annotation_stack.empty() - ? annotation_stack.back() - : cbInfo->symbolName; - tracer->AddAnnotation(cbInfo->correlationId, anno); - } - } else { - VLOG(1) << "Unhandled API Callback for " << domain << " " << cbid; + DeviceTracerImpl *tracer = reinterpret_cast(userdata); + if (cbInfo->callbackSite == CUPTI_API_ENTER) { + Event *event = CurAnnotation(); + tracer->AddAnnotation(cbInfo->correlationId, event); } } CUpti_SubscriberHandle subscriber_; @@ -357,10 +497,12 @@ class DeviceTracerImpl : public DeviceTracer { bool enabled_; uint64_t start_ns_; uint64_t end_ns_; - std::vector kernel_records_; - std::vector mem_records_; - std::vector cpu_records_; - std::unordered_map correlations_; + std::forward_list kernel_records_; + std::forward_list mem_records_; + std::forward_list> cpu_records_; + std::forward_list>> + correlations_pairs; + std::unordered_map correlations_; }; void CreateTracer(DeviceTracer **t) { *t = new DeviceTracerImpl(); } @@ -370,21 +512,104 @@ DeviceTracer *GetDeviceTracer() { return tracer; } -void SetCurAnnotation(const std::string &anno) { - annotation_stack.push_back(anno); -} +void SetCurAnnotation(Event *event) { annotation_stack.push_back(event); } void ClearCurAnnotation() { annotation_stack.pop_back(); } -std::string CurAnnotation() { - if (annotation_stack.empty()) return ""; +Event *CurAnnotation() { + if (annotation_stack.empty()) return nullptr; return annotation_stack.back(); } +std::string CurAnnotationName() { + if (annotation_stack.empty()) return ""; + return annotation_stack.back()->name(); +} void SetCurBlock(int block_id) { block_id_stack.push_back(block_id); } void ClearCurBlock() { block_id_stack.pop_back(); } int BlockDepth() { return block_id_stack.size(); } + +uint32_t GetCurSystemThreadId() { + std::stringstream ss; + ss << std::this_thread::get_id(); + uint32_t id = static_cast(std::stoull(ss.str())); + return id; +} + +void RecoreCurThreadId(int32_t id) { + auto gid = GetCurSystemThreadId(); + VLOG(1) << "RecoreCurThreadId: " << gid << " -> " << id; + system_thread_id_map[gid] = id; +} + +int32_t GetThreadIdFromSystemThreadId(uint32_t id) { + auto it = system_thread_id_map.find(id); + if (it != system_thread_id_map.end()) return it->second; + // return origin id if no event is recorded in this thread. + return static_cast(id); +} + +#ifdef PADDLE_WITH_CUPTI +namespace { + +void initCuptiCbidStr() { + static bool called = false; + if (called) return; + called = true; +#define REGISTER_RUNTIME_CBID_STR(cbid) \ + runtime_cbid_str[CUPTI_RUNTIME_TRACE_CBID_##cbid] = #cbid + + REGISTER_RUNTIME_CBID_STR(cudaBindTexture_v3020); + REGISTER_RUNTIME_CBID_STR(cudaConfigureCall_v3020); + REGISTER_RUNTIME_CBID_STR(cudaDeviceGetAttribute_v5000); + REGISTER_RUNTIME_CBID_STR(cudaDeviceGetStreamPriorityRange_v5050); + REGISTER_RUNTIME_CBID_STR(cudaDeviceSynchronize_v3020); + REGISTER_RUNTIME_CBID_STR(cudaDriverGetVersion_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventCreateWithFlags_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventDestroy_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventDestroy_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventQuery_v3020); + REGISTER_RUNTIME_CBID_STR(cudaEventRecord_v3020); + REGISTER_RUNTIME_CBID_STR(cudaFreeHost_v3020); + REGISTER_RUNTIME_CBID_STR(cudaFree_v3020); + REGISTER_RUNTIME_CBID_STR(cudaFuncGetAttributes_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetDeviceCount_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetDeviceProperties_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetDevice_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetErrorString_v3020); + REGISTER_RUNTIME_CBID_STR(cudaGetLastError_v3020); + REGISTER_RUNTIME_CBID_STR(cudaHostAlloc_v3020); + REGISTER_RUNTIME_CBID_STR(cudaHostGetDevicePointer_v3020); + REGISTER_RUNTIME_CBID_STR(cudaLaunchKernel_v7000); + REGISTER_RUNTIME_CBID_STR(cudaMallocHost_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMalloc_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMemcpyAsync_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMemcpy_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMemsetAsync_v3020); + REGISTER_RUNTIME_CBID_STR(cudaMemset_v3020); + REGISTER_RUNTIME_CBID_STR( + cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags_v7000); + REGISTER_RUNTIME_CBID_STR(cudaPeekAtLastError_v3020); + REGISTER_RUNTIME_CBID_STR(cudaRuntimeGetVersion_v3020); + REGISTER_RUNTIME_CBID_STR(cudaSetDevice_v3020); + REGISTER_RUNTIME_CBID_STR(cudaStreamCreate_v3020); + REGISTER_RUNTIME_CBID_STR(cudaStreamCreateWithFlags_v5000); + REGISTER_RUNTIME_CBID_STR(cudaStreamCreateWithPriority_v5050); + REGISTER_RUNTIME_CBID_STR(cudaStreamDestroy_v5050); + REGISTER_RUNTIME_CBID_STR(cudaStreamSynchronize_v3020); + REGISTER_RUNTIME_CBID_STR(cudaStreamWaitEvent_v3020); + REGISTER_RUNTIME_CBID_STR(cudaUnbindTexture_v3020); +#if CUDA_VERSION >= 9000 + REGISTER_RUNTIME_CBID_STR(cudaLaunchCooperativeKernel_v9000); + REGISTER_RUNTIME_CBID_STR(cudaLaunchCooperativeKernelMultiDevice_v9000); +#endif + +#undef REGISTER_RUNTIME_CBID_STR +} +} // namespace +#endif // PADDLE_WITH_CUPTI + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/device_tracer.h b/paddle/fluid/platform/device_tracer.h index bf0786be2d0faf..6ee2c36146215e 100644 --- a/paddle/fluid/platform/device_tracer.h +++ b/paddle/fluid/platform/device_tracer.h @@ -32,6 +32,8 @@ inline uint64_t PosixInNsec() { return 1000 * (static_cast(tv.tv_sec) * 1000000 + tv.tv_usec); } +class Event; + // DeviceTracer performs the following tasks: // 1. Register cuda callbacks for various events: kernel, memcpy, etc. // 2. Collect cuda statistics: start/end ts, memory, etc. @@ -68,11 +70,13 @@ class DeviceTracer { virtual void Enable() = 0; // Needs to be called once after use. virtual void Disable() = 0; + // Needs to be called once before reuse. + virtual void Reset() = 0; // Add a pair to correlate internal cuda id with high level - // annotation (string). So cuda statistics can be represented by + // annotation event(with string). So cuda statistics can be represented by // human-readable annotations. - virtual void AddAnnotation(uint64_t id, const std::string& anno) = 0; + virtual void AddAnnotation(uint32_t id, Event* event) = 0; virtual void AddMemRecords(const std::string& name, uint64_t start_ns, uint64_t end_ns, int64_t device_id, @@ -92,6 +96,9 @@ class DeviceTracer { // Generate a proto after done (Disabled). virtual proto::Profile GenProfile(const std::string& profile_path) = 0; + // generate kernel elapsed time into Event + virtual void GenEventKernelCudaElapsedTime() = 0; + virtual bool IsEnabled() = 0; }; @@ -99,14 +106,19 @@ class DeviceTracer { DeviceTracer* GetDeviceTracer(); // Set a name for the cuda kernel operation being launched by the thread. -void SetCurAnnotation(const std::string& anno); +void SetCurAnnotation(Event* event); // Clear the name after the operation is done. void ClearCurAnnotation(); // Current name of the operation being run in the thread. -std::string CurAnnotation(); +std::string CurAnnotationName(); +Event* CurAnnotation(); void SetCurBlock(int block_id); void ClearCurBlock(); int BlockDepth(); + +// Set current thread id, so we can map the system thread id to thread id. +void RecoreCurThreadId(int32_t id); +int32_t GetThreadIdFromSystemThreadId(uint32_t id); } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h index d32f9c8667d342..54ad18a8e4abbd 100644 --- a/paddle/fluid/platform/enforce.h +++ b/paddle/fluid/platform/enforce.h @@ -31,6 +31,8 @@ limitations under the License. */ #include #include #include +#include +#include #include "glog/logging.h" #include "paddle/fluid/platform/macros.h" @@ -280,16 +282,62 @@ inline void throw_on_error(ncclResult_t stat, const std::string& msg) { } \ } while (0) -#define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...) \ +namespace details { +template +inline constexpr bool IsArithmetic() { + return std::is_arithmetic::value; +} + +template +struct TypeConverterImpl { + using Type1 = typename std::common_type::type; + using Type2 = Type1; +}; + +template +struct TypeConverterImpl { + using Type1 = T1; + using Type2 = T2; +}; + +template +struct TypeConverter { + private: + static constexpr bool kIsArithmetic = + IsArithmetic() && IsArithmetic(); + + public: + using Type1 = typename TypeConverterImpl::Type1; + using Type2 = typename TypeConverterImpl::Type2; +}; + +template +using CommonType1 = typename std::add_lvalue_reference< + typename std::add_const::Type1>::type>::type; + +template +using CommonType2 = typename std::add_lvalue_reference< + typename std::add_const::Type2>::type>::type; +} // namespace details + +#define __PADDLE_BINARY_COMPARE(__VAL1, __VAL2, __CMP, __INV_CMP, ...) \ do { \ - auto __cond1__ = (__VAL0); \ - auto __cond2__ = (__VAL1); \ - if (UNLIKELY(!((__cond1__)__CMP(__cond2__)))) { \ + auto __val1 = (__VAL1); \ + auto __val2 = (__VAL2); \ + using __TYPE1__ = decltype(__val1); \ + using __TYPE2__ = decltype(__val2); \ + using __COMMON_TYPE1__ = \ + ::paddle::platform::details::CommonType1<__TYPE1__, __TYPE2__>; \ + using __COMMON_TYPE2__ = \ + ::paddle::platform::details::CommonType2<__TYPE1__, __TYPE2__>; \ + bool __is_not_error = (static_cast<__COMMON_TYPE1__>(__val1))__CMP( \ + static_cast<__COMMON_TYPE2__>(__val2)); \ + if (UNLIKELY(!__is_not_error)) { \ PADDLE_THROW("Enforce failed. Expected %s " #__CMP \ " %s, but received %s:%s " #__INV_CMP " %s:%s.\n%s", \ - #__VAL0, #__VAL1, #__VAL0, \ - ::paddle::string::to_string(__cond1__), #__VAL1, \ - ::paddle::string::to_string(__cond2__), \ + #__VAL1, #__VAL2, #__VAL1, \ + ::paddle::string::to_string(__val1), #__VAL2, \ + ::paddle::string::to_string(__val2), \ ::paddle::string::Sprintf(__VA_ARGS__)); \ } \ } while (0) diff --git a/paddle/fluid/platform/enforce_test.cc b/paddle/fluid/platform/enforce_test.cc index 1091badae54a80..adcc95367f11df 100644 --- a/paddle/fluid/platform/enforce_test.cc +++ b/paddle/fluid/platform/enforce_test.cc @@ -118,59 +118,58 @@ TEST(ENFORCE_GT, OK) { PADDLE_ENFORCE_GT(2, 1); } TEST(ENFORCE_GT, FAIL) { bool caught_exception = false; try { - PADDLE_ENFORCE_GT(1, 2UL); + PADDLE_ENFORCE_GT(1, 2); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE(HasPrefix( - StringPiece(error.what()), - "Enforce failed. Expected 1 > 2UL, but received 1:1 <= 2UL:2.")); + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), + "Enforce failed. Expected 1 > 2, but received 1:1 <= 2:2.")); } EXPECT_TRUE(caught_exception); } TEST(ENFORCE_GE, OK) { - PADDLE_ENFORCE_GE(2, 2UL); - PADDLE_ENFORCE_GE(3, 2UL); + PADDLE_ENFORCE_GE(2, 2); PADDLE_ENFORCE_GE(3, 2); - PADDLE_ENFORCE_GE(3.21, 2UL); + PADDLE_ENFORCE_GE(3.21, 2.0); } TEST(ENFORCE_GE, FAIL) { bool caught_exception = false; try { - PADDLE_ENFORCE_GE(1, 2UL); + PADDLE_ENFORCE_GE(1, 2); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE(HasPrefix( - StringPiece(error.what()), - "Enforce failed. Expected 1 >= 2UL, but received 1:1 < 2UL:2.")); + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), + "Enforce failed. Expected 1 >= 2, but received 1:1 < 2:2.")); } EXPECT_TRUE(caught_exception); } TEST(ENFORCE_LE, OK) { PADDLE_ENFORCE_LE(1, 1); - PADDLE_ENFORCE_LE(1, 1UL); - PADDLE_ENFORCE_LE(2, 3UL); - PADDLE_ENFORCE_LE(2UL, 3); - PADDLE_ENFORCE_LE(2UL, 3.2); + PADDLE_ENFORCE_LE(1UL, 1UL); + PADDLE_ENFORCE_LE(2, 3); + PADDLE_ENFORCE_LE(2UL, 3UL); + PADDLE_ENFORCE_LE(2.0, 3.2); } TEST(ENFORCE_LE, FAIL) { bool caught_exception = false; try { - PADDLE_ENFORCE_GT(1, 2UL); + PADDLE_ENFORCE_GT(1, 2); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE(HasPrefix( - StringPiece(error.what()), - "Enforce failed. Expected 1 > 2UL, but received 1:1 <= 2UL:2.")); + EXPECT_TRUE( + HasPrefix(StringPiece(error.what()), + "Enforce failed. Expected 1 > 2, but received 1:1 <= 2:2.")); } EXPECT_TRUE(caught_exception); } TEST(ENFORCE_LT, OK) { PADDLE_ENFORCE_LT(3, 10); - PADDLE_ENFORCE_LT(2, 3UL); - PADDLE_ENFORCE_LT(2UL, 3); + PADDLE_ENFORCE_LT(2UL, 3UL); + PADDLE_ENFORCE_LT(2, 3); } TEST(ENFORCE_LT, FAIL) { bool caught_exception = false; @@ -235,7 +234,13 @@ TEST(ENFORCE_USER_DEFINED_CLASS, EQ) { TEST(ENFORCE_USER_DEFINED_CLASS, NE) { Dims a{{1, 2, 3, 4}}, b{{5, 6, 7, 8}}; - ASSERT_THROW(PADDLE_ENFORCE_EQ(a, b), paddle::platform::EnforceNotMet); + bool caught_exception = false; + try { + PADDLE_ENFORCE_EQ(a, b); + } catch (paddle::platform::EnforceNotMet&) { + caught_exception = true; + } + EXPECT_TRUE(caught_exception); } TEST(EOF_EXCEPTION, THROW_EOF) { diff --git a/paddle/fluid/platform/init.cc b/paddle/fluid/platform/init.cc index ac86b38a61c9d8..4dcf7e79043af0 100644 --- a/paddle/fluid/platform/init.cc +++ b/paddle/fluid/platform/init.cc @@ -22,6 +22,7 @@ limitations under the License. */ #include "paddle/fluid/string/split.h" #ifdef PADDLE_WITH_CUDA #include "paddle/fluid/platform/cuda_device_guard.h" +#include "paddle/fluid/platform/dynload/cupti.h" #endif #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/init.h" @@ -30,6 +31,9 @@ limitations under the License. */ DEFINE_int32(paddle_num_threads, 1, "Number of threads for each paddle instance."); +DEFINE_int32(multiple_of_cupti_buffer_size, 1, + "Multiple of the CUPTI device buffer size. If the timestamps have " + "been dropped when you are profiling, try increasing this value."); namespace paddle { namespace framework { @@ -78,7 +82,32 @@ void InitP2P(std::vector devices) { #endif } +void InitCupti() { +#ifdef PADDLE_WITH_CUPTI + if (FLAGS_multiple_of_cupti_buffer_size == 1) return; + size_t attrValue = 0, attrValueSize = sizeof(size_t); +#define MULTIPLY_ATTR_VALUE(attr) \ + { \ + PADDLE_ENFORCE(!platform::dynload::cuptiActivityGetAttribute( \ + attr, &attrValueSize, &attrValue)); \ + attrValue *= FLAGS_multiple_of_cupti_buffer_size; \ + LOG(WARNING) << "Set " #attr " " << attrValue << " byte"; \ + PADDLE_ENFORCE(!platform::dynload::cuptiActivitySetAttribute( \ + attr, &attrValueSize, &attrValue)); \ + } + MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE); + MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_DEVICE_BUFFER_SIZE_CDP); +#if CUDA_VERSION >= 9000 + MULTIPLY_ATTR_VALUE(CUPTI_ACTIVITY_ATTR_PROFILING_SEMAPHORE_POOL_SIZE); +#endif +#undef MULTIPLY_ATTR_VALUE +#endif +} + void InitDevices(bool init_p2p) { + // CUPTI attribute should be set before any CUDA context is created (see CUPTI + // documentation about CUpti_ActivityAttribute). + InitCupti(); /*Init all available devices by default */ std::vector devices; #ifdef PADDLE_WITH_CUDA diff --git a/paddle/fluid/platform/ngraph_helper.h b/paddle/fluid/platform/ngraph_helper.h index 5ee985ea719f8c..e74f57a79a66ea 100644 --- a/paddle/fluid/platform/ngraph_helper.h +++ b/paddle/fluid/platform/ngraph_helper.h @@ -43,6 +43,13 @@ std::shared_ptr Nchw2Nhwc(std::shared_ptr in) { return std::make_shared(in, axis_vec, in_shape); } +ngraph::Shape FlattenTo1d(ngraph::Shape sh, int num) { + auto x1 = std::accumulate(std::begin(sh), std::end(sh) + num, 1, + std::multiplies()); + size_t x1_l = (size_t)x1; + return ngraph::Shape{x1_l}; +} + ngraph::Shape FlattenTo2d(ngraph::Shape sh, int num) { auto x1 = std::accumulate(std::begin(sh), std::begin(sh) + num, 1, std::multiplies()); diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 85977366e61c67..28f93b4b1259e9 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -12,6 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "paddle/fluid/platform/profiler.h" + #include #include #include @@ -27,7 +29,6 @@ limitations under the License. */ #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/platform/device_tracer.h" #include "paddle/fluid/platform/port.h" -#include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/string/printf.h" DEFINE_bool(enable_rpc_profiler, false, "Enable rpc profiler or not."); @@ -66,12 +67,13 @@ struct EventList { ((kEventSize + kEventAlign - 1) / kEventAlign * kEventAlign); template - void Record(Args&&... args) { + Event* Record(Args&&... args) { if (event_blocks.empty() || event_blocks.front().size() == kNumBlock) { event_blocks.emplace_front(); event_blocks.front().reserve(kNumBlock); } event_blocks.front().emplace_back(std::forward(args)...); + return &event_blocks.front().back(); } std::vector Reduce() { @@ -98,21 +100,8 @@ inline uint64_t GetTimeInNsec() { .count(); } -Event::Event(EventType type, std::string name, uint32_t thread_id, - const DeviceContext* dev_ctx) - : type_(type), name_(name), thread_id_(thread_id), has_cuda_(false) { -#ifdef PADDLE_WITH_CUDA - has_cuda_ = dev_ctx ? platform::is_gpu_place(dev_ctx->GetPlace()) : false; - if (has_cuda_) { - auto* cuda_dev_ctx = static_cast(dev_ctx); - PADDLE_ENFORCE(cudaSetDevice( - boost::get(cuda_dev_ctx->GetPlace()).device)); - PADDLE_ENFORCE(cudaGetDevice(&device_)); - PADDLE_ENFORCE(cudaEventCreate(&event_)); - auto stream = cuda_dev_ctx->stream(); - PADDLE_ENFORCE(cudaEventRecord(event_, stream)); - } -#endif +Event::Event(EventType type, std::string name, uint32_t thread_id) + : type_(type), name_(name), thread_id_(thread_id) { cpu_ns_ = GetTimeInNsec(); } @@ -123,89 +112,70 @@ double Event::CpuElapsedMs(const Event& e) const { } double Event::CudaElapsedMs(const Event& e) const { -#ifdef PADDLE_WITH_CUDA - if (!has_cuda_) return 0.0; - PADDLE_ENFORCE(e.has_cuda() && has_cuda()); - PADDLE_ENFORCE(e.device() == device()); - PADDLE_ENFORCE(cudaEventSynchronize(event_)); - PADDLE_ENFORCE(cudaEventSynchronize(e.event())); - float ms; - PADDLE_ENFORCE(cudaEventElapsedTime(&ms, event_, e.event())); - return ms; +#ifdef PADDLE_WITH_CUPTI + return gpu_ns_ / 1000000.0; #else - PADDLE_THROW("CUDA is not enabled"); + LOG_FIRST_N(WARNING, 1) << "CUDA CUPTI is not enabled"; + return 0; #endif } -#ifdef PADDLE_WITH_CUDA -static void ForEachDevice(std::function func) { - auto original_device = GetCurrentDeviceId(); - int count = GetCUDADeviceCount(); - for (int i = 0; i < count; i++) { - SetDeviceId(i); - func(i); - } - SetDeviceId(original_device); -} -#endif - inline EventList& GetEventList() { if (!g_event_list) { std::lock_guard guard(g_all_event_lists_mutex); g_event_list = std::make_shared(); g_thread_id = g_next_thread_id++; g_all_event_lists.emplace_front(g_event_list); + RecoreCurThreadId(g_thread_id); } return *g_event_list; } -void Mark(const std::string& name, const DeviceContext* dev_ctx) { - GetEventList().Record(EventType::kMark, name, g_thread_id, dev_ctx); +void Mark(const std::string& name) { + GetEventList().Record(EventType::kMark, name, g_thread_id); } -void PushEvent(const std::string& name, const DeviceContext* dev_ctx) { - GetEventList().Record(EventType::kPushRange, name, g_thread_id, dev_ctx); +Event* PushEvent(const std::string& name) { + return GetEventList().Record(EventType::kPushRange, name, g_thread_id); } -void PopEvent(const std::string& name, const DeviceContext* dev_ctx) { - GetEventList().Record(EventType::kPopRange, name, g_thread_id, dev_ctx); +void PopEvent(const std::string& name) { + GetEventList().Record(EventType::kPopRange, name, g_thread_id); } -RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx) +RecordEvent::RecordEvent(const std::string& name) : is_enabled_(false), start_ns_(PosixInNsec()) { if (g_state == ProfilerState::kDisabled) return; - std::lock_guard l(profiler_mu); + // lock is not needed, the code below is thread-safe is_enabled_ = true; - dev_ctx_ = dev_ctx; name_ = name; - PushEvent(name_, dev_ctx_); + Event* e = PushEvent(name_); // Maybe need the same push/pop behavior. - SetCurAnnotation(name_); + SetCurAnnotation(e); } RecordEvent::~RecordEvent() { if (g_state == ProfilerState::kDisabled || !is_enabled_) return; - std::lock_guard l(profiler_mu); + // lock is not needed, the code below is thread-safe DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { - tracer->AddCPURecords(CurAnnotation(), start_ns_, PosixInNsec(), + tracer->AddCPURecords(CurAnnotationName(), start_ns_, PosixInNsec(), BlockDepth(), g_thread_id); } ClearCurAnnotation(); - PopEvent(name_, dev_ctx_); + PopEvent(name_); } -RecordRPCEvent::RecordRPCEvent(const std::string& name, - const DeviceContext* dev_ctx) { +RecordRPCEvent::RecordRPCEvent(const std::string& name) { if (FLAGS_enable_rpc_profiler) { - event_.reset(new platform::RecordEvent(name, dev_ctx)); + event_.reset(new platform::RecordEvent(name)); } } RecordBlock::RecordBlock(int block_id) : is_enabled_(false), start_ns_(PosixInNsec()) { - std::lock_guard l(profiler_mu); + // lock is not needed, the code below is thread-safe if (g_state == ProfilerState::kDisabled) return; is_enabled_ = true; SetCurBlock(block_id); @@ -213,7 +183,7 @@ RecordBlock::RecordBlock(int block_id) } RecordBlock::~RecordBlock() { - std::lock_guard l(profiler_mu); + // lock is not needed, the code below is thread-safe if (g_state == ProfilerState::kDisabled || !is_enabled_) return; DeviceTracer* tracer = GetDeviceTracer(); if (tracer) { @@ -225,11 +195,21 @@ RecordBlock::~RecordBlock() { ClearCurBlock(); } +void SynchronizeAllDevice() { +#ifdef PADDLE_WITH_CUDA + int count = GetCUDADeviceCount(); + for (int i = 0; i < count; i++) { + SetDeviceId(i); + PADDLE_ENFORCE(cudaDeviceSynchronize()); + } +#endif +} + void EnableProfiler(ProfilerState state) { PADDLE_ENFORCE(state != ProfilerState::kDisabled, "Can't enable profiling, since the input state is ", "ProfilerState::kDisabled"); - + SynchronizeAllDevice(); std::lock_guard l(profiler_mu); if (state == g_state) { return; @@ -238,23 +218,20 @@ void EnableProfiler(ProfilerState state) { should_send_profile_state = true; GetDeviceTracer()->Enable(); #ifdef PADDLE_WITH_CUDA - if (g_state == ProfilerState::kCUDA) { + if (g_state == ProfilerState::kCUDA || g_state == ProfilerState::kAll || + g_state == ProfilerState::kCPU) { // Generate some dummy events first to reduce the startup overhead. - for (int i = 0; i < 5; i++) { - ForEachDevice([](int d) { - DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(d)); - Mark("_cuda_startup_", dev_ctx); - dev_ctx->Wait(); - delete dev_ctx; - }); - } + DummyKernelAndEvent(); + GetDeviceTracer()->Reset(); } #endif // Mark the profiling start. - Mark("_start_profiler_", nullptr); + Mark("_start_profiler_"); } void ResetProfiler() { + SynchronizeAllDevice(); + GetDeviceTracer()->Reset(); std::lock_guard guard(g_all_event_lists_mutex); for (auto it = g_all_event_lists.begin(); it != g_all_event_lists.end(); ++it) { @@ -481,20 +458,23 @@ void ParseEvents(const std::vector>& events, void DisableProfiler(EventSortingKey sorted_key, const std::string& profile_path) { + SynchronizeAllDevice(); std::lock_guard l(profiler_mu); if (g_state == ProfilerState::kDisabled) return; // Mark the profiling stop. - Mark("_stop_profiler_", nullptr); + Mark("_stop_profiler_"); - std::vector> all_events = GetAllEvents(); - ParseEvents(all_events, true, sorted_key); - ParseEvents(all_events, false, sorted_key); - ResetProfiler(); DeviceTracer* tracer = GetDeviceTracer(); if (tracer->IsEnabled()) { tracer->Disable(); tracer->GenProfile(profile_path); + tracer->GenEventKernelCudaElapsedTime(); } + + std::vector> all_events = GetAllEvents(); + ParseEvents(all_events, true, sorted_key); + ParseEvents(all_events, false, sorted_key); + ResetProfiler(); g_state = ProfilerState::kDisabled; should_send_profile_state = true; } diff --git a/paddle/fluid/platform/profiler.cu b/paddle/fluid/platform/profiler.cu new file mode 100644 index 00000000000000..e115c554caf383 --- /dev/null +++ b/paddle/fluid/platform/profiler.cu @@ -0,0 +1,50 @@ +/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. + +licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/platform/profiler.h" + +#include + +namespace paddle { +namespace platform { + +__global__ void DummyKernel(int *a) { a[0] = 0; } + +static void ForEachDevice(std::function func) { + auto original_device = GetCurrentDeviceId(); + int count = GetCUDADeviceCount(); + for (int i = 0; i < count; i++) { + SetDeviceId(i); + func(i); + } + SetDeviceId(original_device); +} + +void DummyKernelAndEvent() { + for (int i = 0; i < 5; i++) { + ForEachDevice([](int d) { + CUDADeviceContext *dev_ctx = new CUDADeviceContext(CUDAPlace(d)); + Mark("_cuda_startup_"); + int *ptr; + PADDLE_ENFORCE(cudaMalloc(&ptr, sizeof(int))); + DummyKernel<<<1, 1, 0, dev_ctx->stream()>>>(ptr); + dev_ctx->Wait(); + PADDLE_ENFORCE(cudaFree(ptr)); + delete dev_ctx; + }); + } +} + +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index f5d3490634f319..55d94f0fd84a5e 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -28,17 +28,17 @@ class Event { public: // The DeviceContext is used to get the cuda stream. // If CPU profiling mode, can pass nullptr. - Event(EventType type, std::string name, uint32_t thread_id, - const DeviceContext* dev_ctx); + Event(EventType type, std::string name, uint32_t thread_id); const EventType& type() const; std::string name() const { return name_; } uint32_t thread_id() const { return thread_id_; } - bool has_cuda() const { return has_cuda_; } #ifdef PADDLE_WITH_CUDA +#ifndef PADDLE_WITH_CUPTI cudaEvent_t event() const { return event_; } int device() const { return device_; } +#endif #endif double CpuElapsedMs(const Event& e) const; @@ -49,11 +49,21 @@ class Event { std::string name_; uint32_t thread_id_; int64_t cpu_ns_; - bool has_cuda_; #ifdef PADDLE_WITH_CUDA +#ifdef PADDLE_WITH_CUPTI + int64_t gpu_ns_ = 0; + + public: + void AddCudaElapsedTime(int64_t start_ns, int64_t end_ns) { + gpu_ns_ += end_ns - start_ns; + } + + private: +#else cudaEvent_t event_ = nullptr; int device_ = -1; #endif +#endif }; enum ProfilerState { @@ -63,22 +73,19 @@ enum ProfilerState { kAll, // Profile both CPU and GPU. (Currently experimental). }; -void Mark(const std::string& name, const DeviceContext* dev_ctx); +void Mark(const std::string& name); -void PushEvent(const std::string& name, const DeviceContext* dev_ctx); +Event* PushEvent(const std::string& name); -void PopEvent(const std::string& name, const DeviceContext* dev_ctx); +void PopEvent(const std::string& name); struct RecordEvent { - // dev_ctx can be set to nullptr if device is cpu. - RecordEvent(const std::string& name, const DeviceContext* dev_ctx); + explicit RecordEvent(const std::string& name); ~RecordEvent(); bool is_enabled_; uint64_t start_ns_; - // The device context is used by Event to get the current cuda stream. - const DeviceContext* dev_ctx_; // Event name std::string name_; // Need to distinguish name by op type, block_id, program_id and perhaps @@ -88,8 +95,7 @@ struct RecordEvent { class RecordRPCEvent { public: - // dev_ctx can be set to nullptr if device is cpu. - RecordRPCEvent(const std::string& name, const DeviceContext* dev_ctx); + explicit RecordRPCEvent(const std::string& name); ~RecordRPCEvent() {} private: @@ -132,5 +138,9 @@ bool ShouldSendProfileState(); void SetProfileListener(); int64_t ListenerId(); +#ifdef PADDLE_WITH_CUDA +void DummyKernelAndEvent(); +#endif + } // namespace platform } // namespace paddle diff --git a/paddle/fluid/platform/profiler.proto b/paddle/fluid/platform/profiler.proto index 7b42aa785ec6ad..e761d7b266e92f 100644 --- a/paddle/fluid/platform/profiler.proto +++ b/paddle/fluid/platform/profiler.proto @@ -31,6 +31,7 @@ message Event { optional int64 sub_device_id = 6; optional MemCopy memcopy = 7; + optional string detail_info = 9; } message Profile { diff --git a/paddle/fluid/platform/profiler_test.cc b/paddle/fluid/platform/profiler_test.cc index 61f467814ba4a2..528fe03c67a282 100644 --- a/paddle/fluid/platform/profiler_test.cc +++ b/paddle/fluid/platform/profiler_test.cc @@ -23,76 +23,49 @@ TEST(Event, CpuElapsedTime) { using paddle::platform::Event; using paddle::platform::EventType; - Event start_event(EventType::kPushRange, "test", 0, nullptr); - EXPECT_TRUE(start_event.has_cuda() == false); + Event start_event(EventType::kPushRange, "test", 0); int counter = 0; while (counter != 1000) { counter++; } - Event stop_event(EventType::kPopRange, "test", 0, nullptr); + Event stop_event(EventType::kPopRange, "test", 0); EXPECT_GT(start_event.CpuElapsedMs(stop_event), 0); } -#ifdef PADDLE_WITH_CUDA -TEST(Event, CudaElapsedTime) { - using paddle::platform::DeviceContext; - using paddle::platform::CUDADeviceContext; - using paddle::platform::CUDAPlace; - using paddle::platform::Event; - using paddle::platform::EventType; - - DeviceContext* dev_ctx = new CUDADeviceContext(CUDAPlace(0)); - Event start_event(EventType::kPushRange, "test", 0, dev_ctx); - EXPECT_TRUE(start_event.has_cuda() == true); - int counter = 0; - while (counter != 1000) { - counter++; - } - Event stop_event(EventType::kPopRange, "test", 0, dev_ctx); - EXPECT_GT(start_event.CudaElapsedMs(stop_event), 0); -} -#endif - TEST(RecordEvent, RecordEvent) { using paddle::platform::DeviceContext; using paddle::platform::Event; using paddle::platform::EventType; using paddle::platform::RecordEvent; + using paddle::platform::PushEvent; + using paddle::platform::PopEvent; using paddle::platform::ProfilerState; using paddle::platform::EventSortingKey; ProfilerState state = ProfilerState::kCPU; - DeviceContext* dev_ctx = nullptr; -#ifdef PADDLE_WITH_CUDA - using paddle::platform::CUDADeviceContext; - using paddle::platform::CUDAPlace; - state = ProfilerState::kCUDA; - dev_ctx = - new paddle::platform::CUDADeviceContext(paddle::platform::CUDAPlace(0)); -#endif EnableProfiler(state); /* Usage 1: - * PushEvent(evt_name, dev_ctx); + * PushEvent(evt_name); * ... * code to be analyzed * ... - * PopEvent(evt_name, dev_ctx); + * PopEvent(evt_name); */ LOG(INFO) << "Usage 1: PushEvent & PopEvent"; for (int loop = 0; loop < 3; ++loop) { for (int i = 1; i < 5; ++i) { std::string name = "op_" + std::to_string(i); - PushEvent(name, dev_ctx); + PushEvent(name); int counter = 1; while (counter != i * 1000) counter++; - PopEvent(name, dev_ctx); + PopEvent(name); } } /* Usage 2: * { - * RecordEvent record_event(name, dev_ctx); + * RecordEvent record_event(name); * ... * code to be analyzed * ... @@ -101,7 +74,7 @@ TEST(RecordEvent, RecordEvent) { LOG(INFO) << "Usage 2: RecordEvent"; for (int i = 1; i < 5; ++i) { std::string name = "evs_op_" + std::to_string(i); - RecordEvent record_event(name, dev_ctx); + RecordEvent record_event(name); int counter = 1; while (counter != i * 1000) counter++; } @@ -123,20 +96,20 @@ TEST(RecordEvent, RecordEvent) { LOG(INFO) << "Usage 3: nested RecordEvent"; for (int i = 1; i < 5; ++i) { std::string name = "ano_evs_op_" + std::to_string(i); - RecordEvent record_event(name, dev_ctx); + RecordEvent record_event(name); int counter = 1; while (counter != i * 100) counter++; { std::string nested_name = "nested_ano_evs_op_" + std::to_string(i); - RecordEvent nested_record_event(nested_name, dev_ctx); + RecordEvent nested_record_event(nested_name); int nested_counter = 1; while (nested_counter != i * 100) nested_counter++; } } // Bad Usage: - PushEvent("event_without_pop", dev_ctx); - PopEvent("event_without_push", dev_ctx); + PushEvent("event_without_pop"); + PopEvent("event_without_push"); std::vector> events = paddle::platform::GetAllEvents(); int cuda_startup_count = 0; diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index a4a01ad647b038..d8e57a1ac6ccfc 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -106,6 +106,11 @@ bool IsCompiledWithDIST() { #endif } +template +static inline bool IsSamePlace(const PlaceType1 &p1, const PlaceType2 &p2) { + return paddle::platform::Place(p1) == paddle::platform::Place(p2); +} + PYBIND11_MODULE(core, m) { // Not used, just make sure cpu_info.cc is linked. paddle::platform::CpuTotalPhysicalMemory(); @@ -373,7 +378,13 @@ PYBIND11_MODULE(core, m) { PADDLE_ENFORCE(CheckLoD(new_lod, vectorize(self.dims()).front()), "the provided lod info is invalid"); self.set_lod(new_lod); - }) + }, + py::arg("lod"), R"DOC( + Set LoD of the LoDTensor. + + Args: + lod (List[List[int]]): the lod to be set. + )DOC") .def("set_recursive_sequence_lengths", [](LoDTensor &self, const std::vector> &recursive_sequence_lengths) { @@ -389,7 +400,17 @@ PYBIND11_MODULE(core, m) { CheckLoD(new_offset_lod, vectorize(self.dims()).front()), "the provided recursive_sequence_lengths info is invalid"); self.set_lod(new_offset_lod); - }) + }, + py::arg("recursive_sequence_lengths"), R"DOC( + Set LoD of the LoDTensor according to recursive sequence length. + + For example, if recursive_sequence_lengths=[[2, 3]], meaning that + there are two sequences with length 2 and 3 respectively, the + corresponding lod would be [[0, 2, 2+3]], i.e, [[0, 2, 5]]. + + Args: + recursive_sequence_lengths (List[List[int]]): sequence lengths. + )DOC") .def("lod", [](LoDTensor &self) -> std::vector> { // output the offset-based lod info @@ -398,7 +419,13 @@ PYBIND11_MODULE(core, m) { new_lod.reserve(lod.size()); std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); return new_lod; - }) + }, + R"DOC( + Return the LoD of the LoDTensor. + + Returns: + out (List[List[int]]): the lod of the LoDTensor. + )DOC") // Set above comments of set_lod. .def("recursive_sequence_lengths", [](LoDTensor &self) -> std::vector> { @@ -408,12 +435,25 @@ PYBIND11_MODULE(core, m) { new_lod.reserve(lod.size()); std::copy(lod.begin(), lod.end(), std::back_inserter(new_lod)); return new_lod; - }) - .def("has_valid_recursive_sequence_lengths", [](LoDTensor &self) -> bool { - // Check that the lod info is valid and match the outermost - // dimension of the LoDTensor data - return CheckLoD(self.lod(), vectorize(self.dims()).front()); - }); + }, + R"DOC( + Return the sequence length of the LoDTensor corresponding to LoD. + + Returns: + out (List[List[int]): the sequence lengths. + )DOC") + .def("has_valid_recursive_sequence_lengths", + [](LoDTensor &self) -> bool { + // Check that the lod info is valid and match the outermost + // dimension of the LoDTensor data + return CheckLoD(self.lod(), vectorize(self.dims()).front()); + }, + R"DOC( + Check whether the lod of the LoDTensor is valid. + + Returns: + out (bool): whether the lod is valid. + )DOC"); py::class_(m, "SelectedRows") .def("__init__", @@ -549,11 +589,45 @@ All parameter, weight, gradient are variables in Paddle. [](Scope &self, const std::string &name) -> Variable * { return self.Var(name); }, + py::arg("name"), + R"DOC( + Find or create variable named :code:`name` in the current scope. + + If the variable named :code:`name` does not exist in the + current scope, the variable would be created. Otherwise, + return the existing variable. + + Args: + name (str): the variable name. + + Returns: + out (core.Variable): the found or created variable. + )DOC", + py::return_value_policy::reference) + .def("find_var", &Scope::FindVar, py::arg("name"), + R"DOC( + Find variable named :code:`name` in the current scope or + its parent scope. Return None if not found. + + Args: + name (str): the variable name. + + Returns: + out (core.Variable|None): the found variable or None. + )DOC", py::return_value_policy::reference) - .def("find_var", &Scope::FindVar, py::return_value_policy::reference) .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); }, + R"DOC( + Create a new sub-scope of the current scope. + + Returns: + out (core._Scope): the created sub-scope. + )DOC", py::return_value_policy::reference) - .def("drop_kids", &Scope::DropKids); + .def("drop_kids", &Scope::DropKids, + R"DOC( + Delete all sub-scopes of the current scope. + )DOC"); m.def("Scope", []() -> Scope * { @@ -561,6 +635,12 @@ All parameter, weight, gradient are variables in Paddle. ScopePool::Instance().Insert(std::unique_ptr(s)); return s; }, + R"DOC( + Create a new scope. + + Returns: + out (core._Scope): the created scope. + )DOC", py::return_value_policy::reference); //! @note: Be careful! PyBind will return std::string as an unicode, not @@ -657,23 +737,45 @@ All parameter, weight, gradient are variables in Paddle. PADDLE_THROW("Cannot use CUDAPlace in CPU only version"); #endif }) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", + &IsSamePlace) .def("__str__", string::to_string); py::class_(m, "CPUPlace") .def(py::init<>()) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", + &IsSamePlace) .def("__str__", string::to_string); py::class_(m, "CUDAPinnedPlace") .def("__init__", - [](platform::CUDAPinnedPlace &) { + [](platform::CUDAPinnedPlace &self) { #ifndef PADDLE_WITH_CUDA PADDLE_THROW("Cannot use CUDAPinnedPlace in CPU only version"); #endif + new (&self) platform::CUDAPinnedPlace(); }) + .def("_equals", &IsSamePlace) + .def("_equals", + &IsSamePlace) + .def("_equals", + &IsSamePlace) + .def("_equals", + &IsSamePlace) .def("__str__", string::to_string); py::class_(m, "Place") .def(py::init<>()) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) + .def("_equals", &IsSamePlace) .def("is_gpu_place", [](platform::Place &self) { return platform::is_gpu_place(self); }) .def("gpu_device_id", @@ -789,11 +891,13 @@ All parameter, weight, gradient are variables in Paddle. self[i].ShareDataWith(t); self[i].set_lod(t.lod()); }) - .def("append", [](LoDTensorArray &self, const LoDTensor &t) { - self.emplace_back(); - self.back().ShareDataWith(t); - self.back().set_lod(t.lod()); - }); + .def("append", + [](LoDTensorArray &self, const LoDTensor &t) { + self.emplace_back(); + self.back().ShareDataWith(t); + self.back().set_lod(t.lod()); + }, + py::arg("tensor"), "Append a LoDensor to LoDTensorArray."); m.def("IsInplace", [](std::string op) -> bool { return operators::IsInplace(op); }); diff --git a/paddle/fluid/train/demo/README.md b/paddle/fluid/train/demo/README.md index 191da20669e185..bd53ab4b0c023b 100644 --- a/paddle/fluid/train/demo/README.md +++ b/paddle/fluid/train/demo/README.md @@ -9,7 +9,6 @@ PADDLE_LIB=/paddle/lib/dir cmake .. -DFLUID_INSTALL_DIR=$PADDLE_LIB \ -DCMAKE_BUILD_TYPE=Release \ - -DWITH_FLUID_ONLY=ON \ -DWITH_GPU=OFF \ -DWITH_STYLE_CHECK=OFF \ -DWITH_MKL=OFF \ diff --git a/paddle/scripts/README.md b/paddle/scripts/README.md index 6c608fce3cdad3..1db262f06d9766 100644 --- a/paddle/scripts/README.md +++ b/paddle/scripts/README.md @@ -66,12 +66,10 @@ Users can specify the following Docker build arguments with either "ON" or "OFF" | `WITH_AVX` | OFF | Set to "ON" to enable AVX support. | | `WITH_TESTING` | OFF | Build unit tests binaries. | | `WITH_MKL` | ON | Build with [Intel® MKL](https://software.intel.com/en-us/mkl) and [Intel® MKL-DNN](https://github.com/01org/mkl-dnn) support. | -| `WITH_GOLANG` | OFF | Build fault-tolerant parameter server written in go. | | `WITH_PYTHON` | ON | Build with python support. Turn this off if build is only for capi. | | `WITH_STYLE_CHECK` | ON | Check the code style when building. | | `PYTHON_ABI` | "" | Build for different python ABI support, can be cp27-cp27m or cp27-cp27mu | | `RUN_TEST` | OFF | Run unit test immediently after the build. | -| `WITH_DOC` | OFF | Build docs after build binaries. | | `WOBOQ` | OFF | Generate WOBOQ code viewer under `build/woboq_out` | ## Docker Images diff --git a/paddle/scripts/fast_install.sh b/paddle/scripts/fast_install.sh index b960d0f00a2619..0461944ca8c6c5 100644 --- a/paddle/scripts/fast_install.sh +++ b/paddle/scripts/fast_install.sh @@ -1,5 +1,37 @@ #!/bin/bash +## purple to echo +function purple(){ + echo -e "\033[35m$1\033[0m" +} + + +## green to echo +function green(){ + echo -e "\033[32m$1\033[0m" +} + +## Error to warning with blink +function bred(){ + echo -e "\033[31m\033[01m\033[05m$1\033[0m" +} + +## Error to warning with blink +function byellow(){ + echo -e "\033[33m\033[01m\033[05m$1\033[0m" +} + + +## Error +function red(){ + echo -e "\033[31m\033[01m$1\033[0m" +} + +## warning +function yellow(){ + echo -e "\033[33m\033[01m$1\033[0m" +} + path='http://paddlepaddle.org/download?url=' #release_version=`curl -s https://pypi.org/project/paddlepaddle/|grep -E "/project/paddlepaddle/"|grep "release"|awk -F '/' '{print $(NF-1)}'|head -1` release_version=1.2.0 @@ -228,36 +260,128 @@ function checkLinuxPaddleVersion(){ done } -function checkLinuxPip(){ +function checkPythonVirtualenv(){ while true do - echo "请输入您要使用的pip目录（您可以另起终端，并使用which pip来查看）：" - read -p "" pip_path - if [ "$pip_path" == "" -o ! -f "$pip_path" ];then - echo "检测结果：pip不存在,请重新输入" - continue - fi - python_version=`$pip_path --version|awk -F "[ |)]" '{print $6}'|sed 's#\.##g'` - if [ "$python_version" == "27" ];then - uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27mu"` - if [[ "$uncode" == "" ]];then - uncode= - else - uncode=u - fi - fi - if [ "$python_version" == "" ];then - echo "检测结果：pip不存在,请重新输入" - else - version_list=`echo "${python_list[@]}" | grep "$python_version" ` - if [ "$version_list" != "" ];then - echo "检测结果：找到python${python_version}版本" - break - else - echo "检测结果：找不到可用的 pip, 我们只支持Python27/35/36/37及其对应的pip, 请重新输入， 或使用ctrl + c退出 " - fi - fi + read -p " + 是否使用python virtualenv虚环境安装(y/n)": check_virtualenv + case $check_virtualenv in + y) + echo "为您使用python虚环境安装" + ;; + n) + break + ;; + *) + continue + ;; + esac + + virtualenv_path=`which virtualenv 2>&1` + if [ "$virtualenv_path" == "" ];then + $python_path -m pip install virtualenv + if [ "$?" != '0' ];then + echo "安装虚拟环境失败,请检查本地环境" + fi + fi + + while true + do + read -p "请输入虚拟环境名字：" virtualenv_name + if [ "$virtualenv_name" == "" ];then + echo "不能为空" + continue + fi + break + done + + virtualenv -p $python_path ${virtualenv_name} + if [ "$?" != 0 ];then + echo "创建虚环境失败,请检查环境" + exit 2 + fi + cd ${virtualenv_name} + source ./bin/activate + + if [ "$?" == 0 ];then + use_virtualenv= + python_path=`which python` + break + else + echo "创建虚环境失败,请检查环境" + exit 2 + fi + done +} + +function checkLinuxPython(){ + python_path=`which python 2>/dev/null` + while true + do + if [ "$python_path" == '' ];then + while true + do + read -p "没有找到默认的python版本,请输入要安装的python路径:" python_path + python_path=`$python_path -V` + if [ "$python_path" != "" ];then + break + else + echo "输入路径有误,未找到pyrhon" + fi done + fi + + python_version=`$python_path -V 2>&1|awk -F '[ .]' '{print $2$3}'` + pip_version=`$python_path -m pip -V|awk -F '[ .]' '{print $2}'` + while true + do + read -p " + 找到python版本$python_version,使用请输入y,选择其他版本请输n(y/n):" check_python + case $check_python in + n) + read -p "请指定您的python路径:" new_python_path + python_V=`$new_python_path -V 2>/dev/null` + if [ "$python_V" != "" ];then + python_path=$new_python_path + python_version=`$python_path -V 2>&1|awk -F '[ .]' '{print $2$3}'` + pip_version=`python -m pip -V|awk -F '[ .]' '{print $2}'` + echo "您的python版本为${python_version}" + break + else + echo 输入有误,未找到python路径 + fi + ;; + y) + break + ;; + *) + echo "输入有误，请重新输入." + continue + ;; + esac + done + + if [ "$pip_version" -lt 9 ];then + echo "您的pip版本小于9.0.1 请升级pip (pip install --upgrade pip)" + exit 0 + fi + + if [ "$python_version" == "27" ];then + uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27mu"` + if [[ "$uncode" == "" ]];then + uncode= + else + uncode=u + fi + fi + + version_list=`echo "${python_list[@]}" | grep "$python_version" ` + if [ "$version_list" == "" ];then + echo "找不到可用的 pip, 我们只支持Python27/35/36/37及其对应的pip, 请重新输入， 或使用ctrl + c退出 " + else + break + fi + done } function checkLinuxAVX(){ @@ -287,25 +411,36 @@ function PipLinuxInstall(){ wheel_cpu_develop="http://paddle-wheel.bj.bcebos.com/latest-cpu-${AVX}-${math}/paddlepaddle-latest-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" wheel_gpu_develop="http://paddle-wheel.bj.bcebos.com/latest-gpu-cuda${CUDA}-cudnn${CUDNN}-${AVX}-${math}/paddlepaddle_gpu-latest-cp${python_version}-cp${python_version}m${uncode}-linux_x86_64.whl" - if [[ "$paddle_version" == "2" ]];then if [[ "$GPU" == "gpu" ]];then if [[ ${AVX} == "avx" ]];then rm -rf `echo $wheel_gpu_release|awk -F '/' '{print $NF}'` wget -q $wheel_gpu_release if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi else rm -rf `echo $wheel_gpu_release_novax|awk -F '/' '{print $NF}'` wget -q $wheel_gpu_release_novax if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release_noavx + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_release_noavx + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi fi @@ -313,9 +448,15 @@ function PipLinuxInstall(){ rm -rf `echo $wheel_cpu_release|awk -F '/' '{print $NF}'` wget -q $wheel_cpu_release if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_release + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_release + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi fi @@ -324,18 +465,30 @@ function PipLinuxInstall(){ rm -rf `echo $wheel_gpu_develop|awk -F '/' '{print $NF}'` wget -q $wheel_gpu_develop if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_develop + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_gpu_develop + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi else rm -rf `echo $wheel_cpu_develop|awk -F '/' '{print $NF}'` wget -q $wheel_cpu_develop if [ "$?" == "0" ];then - $pip_path install --user -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_develop + $python_path -m pip install ${use_virtualenv} -i https://mirrors.aliyun.com/pypi/simple --trusted-host=mirrors.aliyun.com $wheel_cpu_develop + if [ "$?" == 0 ];then + echo 安装成功 + else + echo 安装失败 + exit 1 + fi else - echo "paddlepaddle whl包下载失败" + echo paddlepaddle whl包下载失败 exit 1 fi fi @@ -575,95 +728,122 @@ gpu_list=( echo echo "Step 5. 检测pip版本" echo - checkLinuxPip + checkLinuxPython echo checkLinuxAVX + echo + echo "Step 6.是否使用Python的虚拟环境" + use_virtualenv="--user" + checkPythonVirtualenv echo "*********************2. 开始安装*****************************" PipLinuxInstall + if [ "$check_virtualenv" == 'y' ];then + echo "虚环境创建成功，请cd 进入${virtualenv_name}, 执行 source bin/activate　进入虚环境。退出虚环境执行 deactivate命令。 + 更多虚环境使用方法请参考virtualenv官网:https://virtualenv.pypa.io/en/latest/" + fi +} + +function clearMacPythonEnv(){ + python_version="" + python_brief_version="" + python_root="" } function checkMacPython2(){ while true do - read -p " - => 未能在常规路径下找到Python2，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python2（注意Python版本不能低于2.7.15） - 如希望自定义Python路径，请输入路径：" python_root - echo python_version=`$python_root --version 2>&1 1>&1` - if [ $? == "0" ];then - : + if [[ $? == "0" ]];then + if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ];then + clearMacPythonEnv + else + check_python=`echo $python_version | grep "Python 2"` + if [[ -n "$check_python" ]];then + while true + do + echo -e " => 在您的环境中找到 \033[32m[ $python_version ]\033[0m, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " + read -p "" use_python + echo + use_python=`echo $use_python | tr 'A-Z' 'a-z'` + if [[ "$use_python" == "y" ]]||[[ "$use_python" == "" ]];then + use_python="y" + break + elif [[ "$use_python" == "n" ]];then + clearMacPythonEnv + break + else + red " 输入错误，请重新输入(y/n)" + fi + done + if [[ "$use_python" == "y" ]];then + return 0 + fi + else + red " 您输入Python的不是Python2" + clearMacPythonEnv + fi + fi else - python_version="" + clearMacPythonEnv + red " => 未能在常规路径下找到可用的Python2，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python2（注意Python版本不能低于2.7.15）" + read -p " 如希望自定义Python路径，请输入路径 + 如果希望重新选择Python版本，请回车：" python_root + echo + if [[ "$python_root" == "" ]];then + python_V="" + clearMacPythonEnv + return 1 + fi fi - check_python=`echo $python_version | grep "Python 2"` - if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then - python_version="" - elif [ -n "$check_python" ];then - while true - do - read -p " - => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " use_python - echo - use_python=`echo $use_python | tr 'A-Z' 'a-z'` - if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then - use_python="y" - break - elif [ "$use_python" == "n" ];then - python_root="" - break - else - echo "输入错误，请重新输入(y/n)" - fi - done - if [ "$use_python" == "y" ];then - break - fi - else - echo "您输入Python的不是Python2" - python_version="" - fi done } function checkMacPython3(){ while true do - read -p " - => 未能在常规路径下找到Python3，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载Python3 - 如希望自定义Python路径，请输入路径：" python_root - python_version=`$python_root --version 2>&1 1>&1` - if [ $? == "0" ];then - : + python_version=`$python_root --version 2>&1 1>&1` + if [[ $? == "0" ]];then + if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then + clearMacPythonEnv + else + check_python=`echo $python_version | grep "Python 3"` + if [[ -n "$check_python" ]];then + while true + do + echo -e " => 在您的环境中找到 \033[32m[ $python_version ]\033[0m, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " + read -p "" use_python + echo + use_python=`echo $use_python | tr 'A-Z' 'a-z'` + if [[ "$use_python" == "y" ]]||[[ "$use_python" == "" ]];then + use_python="y" + break + elif [[ "$use_python" == "n" ]];then + clearMacPythonEnv + break + else + red " 输入错误，请重新输入(y/n)" + fi + done + if [[ "$use_python" == "y" ]];then + return 0 + fi + else + red " 您输入Python的不是Python3" + clearMacPythonEnv + fi + fi else - python_version="" + clearMacPythonEnv + red " => 未能在常规路径下找到可用的Python3，请使用ctrl+c命令退出安装程序，并使用brew或pypi.org下载安装Python3（注意Python版本不能低于3.5.x)" + read -p " 如希望自定义Python路径，请输入路径 + 如果希望重新选择Python版本，请回车：" python_root + echo + if [[ "$python_root" == "" ]];then + python_V="" + clearMacPythonEnv + return 1 + fi fi - check_python=`echo $python_version | grep "Python 3"` - if [ "$python_version" == "" ] || [ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ] ;then - python_version="" - elif [ -n "$check_python" ] ;then - while true - do - read -p " - => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车: " use_python - echo - use_python=`echo $use_python | tr 'A-Z' 'a-z'` - if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then - use_python="y" - break - elif [ "$use_python" == "n" ];then - python_root="" - break - else - echo "输入错误，请重新输入(y/n)" - fi - done - if [ "$use_python" == "y" ];then - break - fi - else - echo "您输入Python的不是Python3" - python_version="" - fi done } @@ -672,145 +852,160 @@ function checkMacPaddleVersion(){ do read -n1 -p "Step 2. 选择PaddlePaddle的版本，请按回车键继续..." echo - read -p " - 1. 开发版：对应Github上develop分支，如您需要开发、或希望使用PaddlePaddle最新功能，请选用此版本 - 2. 稳定版（推荐）：如您无特殊开发需求，建议使用此版本，目前最新的版本号为 ${release_version} - - => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. 稳定版 】 。请在这里输入并回车：" paddle_version - if [ "$paddle_version" == "1" ]||[ "$paddle_version" == "2" ];then + yellow " 1. 开发版：对应Github上develop分支，如您需要开发、或希望使用PaddlePaddle最新功能，请选用此版本" + yellow " 2. 稳定版（推荐）：如您无特殊开发需求，建议使用此版本，目前最新的版本号为 ${release_version}" + read -p " => 请输入数字1或2。如输入其他字符或直接回车，将会默认选择【 2. 稳定版 】 。请在这里输入并回车：" paddle_version + if [[ "$paddle_version" == "1" ]]||[[ "$paddle_version" == "2" ]];then echo - echo "您选择了数字【"$paddle_version" 】" + yellow " 您选择了数字【"$paddle_version" 】" echo break else paddle_version="2" echo - echo "您选择了数字【2】" + yellow " 您选择了数字【2】" echo break fi done } +function initCheckMacPython2(){ + echo + yellow " 您选择了Python "$python_V"，正在寻找符合要求的Python 2版本" + echo + python_root=`which python2.7` + if [[ "$python_root" == "" ]];then + python_root=`which python` + fi + checkMacPython2 + if [[ "$?" == "1" ]];then + return 1 + else + return 0 + fi +} -function checkMacPythonVersion(){ - while true - do - read -n1 -p "Step 3. 选择Python版本，请按回车键继续..." - read -p " - 2. 使用python 2.x - 3. 使用python 3.x +function initCheckMacPython3(){ + echo + yellow " 您选择了Python "$python_V"，正在寻找符合您要求的Python 2版本" + echo + python_root=`which python3` + checkMacPython3 + if [[ "$?" == "1" ]];then + return 1 + else + return 0 + fi +} - => 请输入数字2或3。如输入其他字符或直接回车，将会默认使用【Python 2 】。请在这里输入并回车：" python_V - echo - if [ "$python_V" == "" ];then - python_V="2" +function checkMacPip(){ + if [[ "$python_V" == "2" ]]||[[ "$python_V" == "3" ]];then + + python_brief_version=`$python_root -m pip -V |awk -F "[ |)]" '{print $6}'|sed 's#\.##g'` + if [[ ${python_brief_version} == "" ]];then + red "您输入的python：${python_root} 对应的pip不可用，请检查此pip或重新选择其他python" + echo + return 1 fi - echo "您选择了数字【"$python_V"】，正在寻找符合您要求的Python版本，请按回车键继续..." - echo - if [ "$python_V" == "2" ];then - python_root=`which python2.7` - if [ "$python_root" == "" ];then - python_root=`which python` - fi - python_version=`$python_root --version 2>&1 1>&1` - if [ $? == "0" ];then - : - else - python_version="" - fi - if [ "$python_root" == "" ]||[ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ]||[ "$python_root" == "/usr/bin/python2.7" -a "$python_version" == "Python 2.7.10" ];then - checkMacPython2 - fi - while true - do - read -p " - => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车：" use_python - echo - use_python=`echo $use_python | tr 'A-Z' 'a-z'` - if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then - break - elif [ "$use_python" == "n" ];then - python_root="" - checkMacPython2 - break + pip_version=`$python_root -m pip -V |awk -F '[ .]' '{print $2}'` + if [[ 9 -le ${pip_version} ]];then + : + else + red "您的pip版本过低，请安装pip 9.0.1及以上的版本" + echo + return 1 + fi + if [[ "$python_brief_version" == "" ]];then + clearMacPythonEnv + red "您的 $python_root 对应的pip存在问题，请按ctrl + c退出后重新安装pip，或切换其他python版本" + echo + return 1 + else + if [[ $python_brief_version == "27" ]];then + uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27"` + if [[ $uncode == "" ]];then + uncode="mu" else - echo "输入错误，请重新输入(y/n)" + uncode="m" fi - done - - elif [ "$python_V" == "3" ];then - python_root=`which python3` - python_version=`$python_root --version 2>&1 1>&1` - if [ $? == "0" ];then - : - else - python_version="" - fi - if [ "$python_root" == "" ]||[ "$python_root" == "/usr/bin/python" -a "$python_version" == "Python 2.7.10" ];then - checkMacPython3 - fi - while true - do - read -p " - => 在您的环境中找到 $python_version, 确认使用此版本请输入y；如您希望自定义Python路径请输入n。请在这里输入（y/n）并回车：" use_python + fi + version_list=`echo "${python_list[@]}" | grep "$python_brief_version" ` + if [[ "$version_list" != "" ]];then + return 0 + else + red "未找到可用的pip或pip3。PaddlePaddle目前支持：Python2.7/3.5/3.6/3.7及其对应的pip, 请重新输入，或使用ctrl + c退出" echo - use_python=`echo $use_python | tr 'A-Z' 'a-z'` - if [ "$use_python" == "y" ]||[ "$use_python" == "" ];then - break - elif [ "$use_python" == "n" ];then - checkMacPython3 - break - else - echo "输入错误，请重新输入(y/n)" - fi - done - else - : - fi + clearMacPythonEnv + return 1 + fi + fi + fi +} - if [ "$python_V" == "2" ]||[ "$python_V" == "3" ];then - python_brief_version=`$python_root -m pip -V |awk -F "[ |)]" '{print $6}'|sed 's#\.##g'` - if [[ $python_brief_version == "27" ]];then - uncode=`python -c "import pip._internal;print(pip._internal.pep425tags.get_supported())"|grep "cp27"` - if [[ $uncode == "" ]];then - uncode="mu" - else - uncode="m" - fi - fi - version_list=`echo "${python_list[@]}" | grep "$python_brief_version" ` - if [ "$version_list" != "" ];then - break +function checkMacPythonVersion(){ + while true + do + read -n1 -p "Step 3. 选择Python版本，请按回车键继续..." + echo + yellow " 2. 使用python 2.x" + yellow " 3. 使用python 3.x" + read -p " => 请输入数字2或3。如输入其他字符或直接回车，将会默认使用【Python 2 】。请在这里输入并回车：" python_V + if [[ "$python_V" == "" ]];then + python_V="2" + fi + if [[ "$python_V" == "2" ]];then + initCheckMacPython2 + if [[ "$?" == "0" ]];then + checkMacPip + if [[ "$?" == "0" ]];then + return 0 + else + : + fi else - echo "未找到可用的pip或pip3。PaddlePaddle目前支持：Python2.7/3.5/3.6/3.7及其对应的pip, 请重新输入，或使用ctrl + c退出" - fi - else - echo "输入错误，请重新输入" - fi + : + fi + elif [[ "$python_V" == "3" ]];then + initCheckMacPython3 + if [[ "$?" == "0" ]];then + checkMacPip + if [[ "$?" == "0" ]];then + return 0 + else + : + fi + else + : + fi + else + red "输入错误，请重新输入" + fi done } function checkMacAVX(){ read -n1 -p "Step 4. 检测您的Mac是否支持AVX指令集，请按回车键继续..." - echo if [[ $AVX != "" ]];then AVX="avx" - echo "检测结果：支持" + echo "" + green " 检测结果：支持" + echo "" + return 0 else - read -n1 -p "检测结果：不支持。非常抱歉，PaddlePaddle在Mac系统暂不提供no_avx类型的安装包，您可以选择在Linux系统中安装no_avx版的PaddlePaddle, 请按回车键退出..." - exit + red " 检测结果：不支持。非常抱歉，PaddlePaddle在Mac系统暂不提供no_avx类型的安装包，您可以选择在Linux系统中安装no_avx版的PaddlePaddle, 请按回车键退出..." + echo + return 1 fi - echo } function checkMacGPU(){ read -n1 -p "Step 5. 选择CPU/GPU版本，请按回车键继续..." echo if [[ $GPU != "" ]];then - echo "MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" + yellow " MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" else - echo "MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" + yellow " MacOS环境下，暂未提供GPU版本的PaddlePaddle安装包，将为您安装CPU版本的PaddlePaddle" GPU=cpu fi echo @@ -822,38 +1017,44 @@ function macos() { while true do + checkMacPaddleVersion + checkMacPythonVersion + checkMacAVX + checkMacGPU - echo "*********************2. 开始安装*****************************" + green "*********************2. 开始安装*****************************" echo - read -n1 -p "即将为您下载并安装PaddlePaddle，请按回车键继续..." + yellow "即将为您下载并安装PaddlePaddle，请按回车键继续..." + read -n1 -p "" echo if [[ $paddle_version == "2" ]];then $python_root -m pip install paddlepaddle - if [ $? == "0" ];then - echo "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + if [[ $? == "0" ]];then + green "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" break else rm $whl_cpu_release - echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" echo"" echo "==========================================================================================" echo"" exit 1 fi else - if [ -f $whl_cpu_develop ];then + if [[ -f $whl_cpu_develop ]];then $python_root -m pip install $whl_cpu_develop - if [ $? == "0" ];then + if [[ $? == "0" ]];then rm -rf $whl_cpu_develop - echo "安装成功！小提示：可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + # TODO add install success check here + green "安装成功！小提示：可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" break else - echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" echo"" echo "==========================================================================================" echo"" @@ -861,15 +1062,15 @@ function macos() { fi else wget ${path}$whl_cpu_develop -O $whl_cpu_develop - if [ $? == "0" ];then + if [[ $? == "0" ]];then $python_root -m pip install $whl_cpu_develop - if [ $? == "0" ];then + if [[ $? == "0" ]];then rm $wheel_cpu_develop - echo "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" + green "安装成功，可以使用: ${python_root} 来启动安装了PaddlePaddle的Python解释器" break else rm $whl_cpu_release - echo "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" + red "未能正常安装PaddlePaddle，请尝试更换您输入的python路径，或者ctrl + c退出后请检查您使用的python对应的pip或pip源是否可用" echo"" echo "==========================================================================================" echo"" @@ -877,7 +1078,7 @@ function macos() { fi else rm $whl_cpu_develop - echo "未能正常安装PaddlePaddle，请检查您的网络 或者确认您是否安装有 wget，或者ctrl + c退出后反馈至https://github.com/PaddlePaddle/Paddle/issues" + red "未能正常安装PaddlePaddle，请检查您的网络 或者确认您是否安装有 wget，或者ctrl + c退出后反馈至https://github.com/PaddlePaddle/Paddle/issues" echo"" echo "==========================================================================================" echo"" @@ -890,33 +1091,35 @@ function macos() { function main() { echo "*********************************" - echo "欢迎使用PaddlePaddle快速安装脚本" + green "欢迎使用PaddlePaddle快速安装脚本" echo "*********************************" echo - echo "如果您在安装过程中遇到任何问题，请在https://github.com/PaddlePaddle/Paddle/issues反馈，我们的工作人员将会帮您答疑解惑" + yellow "如果您在安装过程中遇到任何问题，请在https://github.com/PaddlePaddle/Paddle/issues反馈，我们的工作人员将会帮您答疑解惑" echo - echo "本安装包将帮助您在Linux或Mac系统下安装PaddlePaddle，包括 1）安装前的准备和 2）开始安装 两部分" + echo "本安装包将帮助您在Linux或Mac系统下安装PaddlePaddle,包括" + yellow "1）安装前的准备" + yellow "2）开始安装" echo read -n1 -p "请按回车键进行下一步..." echo echo - echo "*********************1. 安装前的准备*****************************" + green "*********************1. 安装前的准备*****************************" echo echo "Step 1. 正在检测您的操作系统信息..." echo SYSTEM=`uname -s` - if [ "$SYSTEM" == "Darwin" ];then - echo "您的系统为：MAC OSX" + if [[ "$SYSTEM" == "Darwin" ]];then + yellow " 您的系统为：MAC OSX" echo macos else - echo "您的系统为：Linux" + yellow " 您的系统为：Linux" echo OS=`cat /etc/issue|awk 'NR==1 {print $1}'` - if [ $OS == "\S" ] || [ "$OS" == "CentOS" ] || [ $OS == "Ubuntu" ];then + if [[ $OS == "\S" ]] || [[ "$OS" == "CentOS" ]] || [[ $OS == "Ubuntu" ]];then linux else - echo "您的系统不在本安装包的支持范围，如您需要在windows环境下安装PaddlePaddle，请您参考PaddlePaddle官网的windows安装文档" + red "您的系统不在本安装包的支持范围，如您需要在windows环境下安装PaddlePaddle，请您参考PaddlePaddle官网的windows安装文档" fi fi } diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index 1135caf4f8c329..26b26c9b1faf7b 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -87,7 +87,7 @@ function cmake_gen() { PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.5/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/Library/Frameworks/Python.framework/Versions/3.5/include/python3.5m/ -DPYTHON_LIBRARY:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.5/lib/libpython3.5m.dylib" - WITH_FLUID_ONLY=${WITH_FLUID_ONLY:-ON} + pip3.5 uninstall -y protobuf pip3.5 install --user -r ${PADDLE_ROOT}/python/requirements.txt else exit 1 @@ -100,7 +100,7 @@ function cmake_gen() { PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.6/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/Library/Frameworks/Python.framework/Versions/3.6/include/python3.6m/ -DPYTHON_LIBRARY:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.6/lib/libpython3.6m.dylib" - WITH_FLUID_ONLY=${WITH_FLUID_ONLY:-ON} + pip3.6 uninstall -y protobuf pip3.6 install --user -r ${PADDLE_ROOT}/python/requirements.txt else exit 1 @@ -113,7 +113,7 @@ function cmake_gen() { PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.7/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/Library/Frameworks/Python.framework/Versions/3.7/include/python3.7m/ -DPYTHON_LIBRARY:FILEPATH=/Library/Frameworks/Python.framework/Versions/3.7/lib/libpython3.7m.dylib" - WITH_FLUID_ONLY=${WITH_FLUID_ONLY:-ON} + pip3.7 uninstall -y protobuf pip3.7 install --user -r ${PADDLE_ROOT}/python/requirements.txt else exit 1 @@ -128,31 +128,44 @@ function cmake_gen() { PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7 -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so" + pip uninstall -y protobuf + pip install -r ${PADDLE_ROOT}/python/requirements.txt elif [ "$1" == "cp27-cp27mu" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:} export PATH=/opt/python/cp27-cp27mu/bin/:${PATH} PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7 -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so" + pip uninstall -y protobuf + pip install -r ${PADDLE_ROOT}/python/requirements.txt elif [ "$1" == "cp35-cp35m" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-3.5.1/lib/:${LD_LIBRARY_PATH} export PATH=/opt/_internal/cpython-3.5.1/bin/:${PATH} export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.5.1/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/opt/_internal/cpython-3.5.1/include/python3.5m -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-3.5.1/lib/libpython3.so" + pip3.5 uninstall -y protobuf + pip3.5 install -r ${PADDLE_ROOT}/python/requirements.txt elif [ "$1" == "cp36-cp36m" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-3.6.0/lib/:${LD_LIBRARY_PATH} export PATH=/opt/_internal/cpython-3.6.0/bin/:${PATH} export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.6.0/bin/python3 -DPYTHON_INCLUDE_DIR:PATH=/opt/_internal/cpython-3.6.0/include/python3.6m -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-3.6.0/lib/libpython3.so" + pip3.6 uninstall -y protobuf + pip3.6 install -r ${PADDLE_ROOT}/python/requirements.txt elif [ "$1" == "cp37-cp37m" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-3.7.0/lib/:${LD_LIBRARY_PATH} export PATH=/opt/_internal/cpython-3.7.0/bin/:${PATH} export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/_internal/cpython-3.7.0/bin/python3.7 -DPYTHON_INCLUDE_DIR:PATH=/opt/_internal/cpython-3.7.0/include/python3.7m -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-3.7.0/lib/libpython3.so" + pip3.7 uninstall -y protobuf + pip3.7 install -r ${PADDLE_ROOT}/python/requirements.txt fi + else + pip uninstall -y protobuf + pip install -r ${PADDLE_ROOT}/python/requirements.txt fi fi @@ -186,7 +199,6 @@ function cmake_gen() { -DWITH_TESTING=${WITH_TESTING:-ON} -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON - -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DWITH_CONTRIB=${WITH_CONTRIB:-ON} -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} @@ -219,7 +231,6 @@ EOF -DCUDNN_ROOT=/usr/ \ -DWITH_TESTING=${WITH_TESTING:-ON} \ -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \ - -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \ -DWITH_CONTRIB=${WITH_CONTRIB:-ON} \ -DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} \ @@ -382,9 +393,7 @@ EOF pip3.7 install --user ${INSTALL_PREFIX:-/paddle/build}/opt/paddle/share/wheels/*.whl fi - if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]] ; then - paddle version - fi + paddle version if [ "$1" == "cp27-cp27m" ]; then pip uninstall -y paddlepaddle @@ -539,7 +548,6 @@ EOF -DCMAKE_BUILD_TYPE=Release \ -DWITH_GPU=OFF \ -DWITH_MKL=OFF \ - -DWITH_FLUID_ONLY=ON local LIB_TYPE=$1 case $LIB_TYPE in @@ -615,13 +623,8 @@ EOF NCCL_DEPS="true" fi - if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]]; then - PADDLE_VERSION="paddle version" - CMD='"paddle", "version"' - else - PADDLE_VERSION="true" - CMD='"true"' - fi + PADDLE_VERSION="paddle version" + CMD='"paddle", "version"' if [ "$1" == "cp35-cp35m" ]; then cat >> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile < 1 and trainers_endpoints: assert self._build_strategy.num_trainers == len( @@ -217,7 +220,7 @@ def _compile(self, scope, place): if self._compiled: if scope and self._scope != scope: raise ValueError("Cannot compile with different scope") - if place and self._place != place: + if place and not self._place._equals(place): raise ValueError("Cannot compile with different place") return self self._compiled = True diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index ef304b11106628..15367c724e5304 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -557,7 +557,8 @@ def generated_op_attr_names(): return { core.op_proto_and_checker_maker.kOpRoleAttrName(), core.op_proto_and_checker_maker.kOpRoleVarAttrName(), - core.op_proto_and_checker_maker.kOpNameScopeAttrName() + core.op_proto_and_checker_maker.kOpNameScopeAttrName(), + core.op_proto_and_checker_maker.kOpCreationCallstackAttrName() } diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index a2abbf36c0267d..24e102b6c2612b 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -766,7 +766,10 @@ def __load_persistable_vars(executor, dirname, need_load_vars): dtype=slice_var.dtype, persistable=True) - dim1_flatten = reduce(lambda x, y: x * y, slice.shape[1:]) + dim1_flatten = 1 + if len(slice.shape) >= 2: + dim1_flatten = reduce(lambda x, y: x * y, slice.shape[1:]) + start = int(offset / dim1_flatten) end = int(offset / dim1_flatten + slice.shape[0]) diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 3a6753b01f152f..539c9675b2d69b 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -506,9 +506,9 @@ class While(object): while loop control flow. Args: - cond (Variable): condition used to compare. + cond(Variable): condition used to compare. is_test(bool): A flag indicating whether execution is in test phase. - name (str): The name of this layer. + name(str): The name of this layer. Examples: .. code-block:: python @@ -589,7 +589,8 @@ def _complete(self): def lod_rank_table(x, level=0): - """LoD Rank Table Operator. Given an input variable **x** and a level number + """ + LoD Rank Table Operator. Given an input variable **x** and a level number of LoD, this layer creates a LodRankTable object. A LoDRankTable object contains a list of bi-element tuples. Each tuple consists of an index and a length, both of which are int type. Refering to specified level of LoD, @@ -883,10 +884,8 @@ def less_than(x, y, force_cpu=None, cond=None, **ignored): return cond -def equal(x, y, cond=None, **ignored): +def equal(x, y, cond=None): """ - **equal** - This layer returns the truth value of :math:`x == y` elementwise. Args: @@ -1458,7 +1457,6 @@ def step_input(self, x): Returns: The current timestep in the input sequence. - """ self._assert_in_rnn_block_("step_input") if not isinstance(x, Variable): @@ -1535,8 +1533,7 @@ def static_input(self, x): @signature_safe_contextmanager def block(self): """ - The block for user to define operators in RNN. See the class docstring - for more details. + The block for user to define operators in RNN. """ if self.status != DynamicRNN.BEFORE_RNN: raise ValueError("rnn.block() can only be invoke once") @@ -1640,8 +1637,7 @@ def memory(self, dtype(str|numpy.dtype): The data type of the initialized memory. Returns: - the memory variable. - + The memory variable. """ self._assert_in_rnn_block_('memory') self._init_zero_idx_() @@ -1740,7 +1736,7 @@ def update_memory(self, ex_mem, new_mem): def output(self, *outputs): """ - mark the RNN output variables. + Mark the RNN output variables. Args: outputs: The output variables. diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index b88be66906e806..a9b391fd53a98d 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -56,7 +56,10 @@ def data(name, Args: name(str): The name/alias of the function - shape(list): Tuple declaring the shape. + shape(list): Tuple declaring the shape. If :code:`append_batch_size` is + True and there is no -1 inside :code:`shape`, it should be + considered as the shape of the each sample. Otherwise, it + should be considered as the shape of the batched data. append_batch_size(bool): 1. If true, it prepends -1 to the shape. For example if shape=[1], the resulting shape is [-1, 1]. diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 09b1b30216b03e..da6c24100452ba 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -24,7 +24,7 @@ from ..layer_helper import LayerHelper __all__ = [ - 'deprecated', 'generate_layer_fn', 'generate_layer_fn_noattr', 'autodoc', + 'deprecated', 'generate_layer_fn', 'generate_activation_fn', 'autodoc', 'templatedoc' ] @@ -89,6 +89,9 @@ def _generate_doc_string_(op_proto, additional_args_lines=None): buf.write('\n') skip_attrs = OpProtoHolder.generated_op_attr_names() + # attr use_mkldnn and is_test also should not be visible to users. + skip_attrs.add("use_mkldnn") + skip_attrs.add("is_test") for each_attr in op_proto.attrs: if each_attr.name in skip_attrs: @@ -226,7 +229,7 @@ def func(*args, **kwargs): return func -def generate_layer_fn_noattr(op_type): +def generate_activation_fn(op_type): """Register the Python layer for an Operator without Attribute. Args: @@ -246,6 +249,7 @@ def func(x, name=None): func.__name__ = op_type func.__doc__ = _generate_doc_string_(op_proto) + return func diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index fe257f9a37fd52..1a7d076835841e 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -8744,16 +8744,17 @@ def slice(input, axes, starts, ends): return out -@templatedoc() def shape(input): """ - ${comment} + **Shape Layer** + + Get the shape of the input. Args: - input (Variable): ${input_comment} + input (Variable): The input variable. Returns: - out (Variable): ${out_comment} + Variable: The shape of the input variable. Examples: .. code-block:: python diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 3dcf9dc06998be..6b4dc4ac89af43 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -14,7 +14,7 @@ from __future__ import print_function import os -from .layer_function_generator import generate_layer_fn, generate_layer_fn_noattr +from .layer_function_generator import generate_layer_fn, generate_activation_fn from .. import core from ..framework import convert_np_dtype_to_dtype_ @@ -53,7 +53,7 @@ __all__ += __activations_noattr__ for _OP in set(__activations_noattr__): - globals()[_OP] = generate_layer_fn_noattr(_OP) + globals()[_OP] = generate_activation_fn(_OP) __all__ += ["uniform_random"] diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index fbd04f1eb46126..cb799b639648fc 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -649,6 +649,7 @@ class AdagradOptimizer(Optimizer): regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer. name: A optional name prefix. + initial_accumulator_value (float): Initial value for moment accumulator. Examples: .. code-block:: python @@ -662,7 +663,8 @@ def __init__(self, learning_rate, epsilon=1.0e-6, regularization=None, - name=None): + name=None, + initial_accumulator_value=0.0): assert learning_rate is not None assert epsilon is not None super(AdagradOptimizer, self).__init__( @@ -671,6 +673,7 @@ def __init__(self, name=name) self.type = "adagrad" self._epsilon = epsilon + self.initial_accumulator_value = initial_accumulator_value def _create_accumulators(self, block, parameters): assert isinstance(block, framework.Block) @@ -683,6 +686,16 @@ def _append_optimize_op(self, block, param_and_grad): moment_acc = self._get_accumulator(self._moment_acc_str, param_and_grad[0]) + startup_block = framework.default_startup_program().global_block() + startup_block.append_op( + type='fill_constant', + inputs={}, + outputs={'Out': [moment_acc]}, + attrs={ + 'dtype': moment_acc.dtype, + 'value': self.initial_accumulator_value, + 'shape': moment_acc.shape, + }) # Create the adagrad optimizer op adagrad_op = block.append_op( @@ -1368,9 +1381,9 @@ class FtrlOptimizer(Optimizer): Args: learning_rate (float|Variable): global learning rate. - l1 (float): - l2 (float): - lr_power (float): + l1 (float): L1 regularization strength. + l2 (float): L2 regularization strength. + lr_power (float): Learning Rate Power. regularization: A Regularizer, such as fluid.regularizer.L2DecayRegularizer. name: A optional name prefix. diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 22212ae9a216ac..8586670c2481a0 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -148,6 +148,8 @@ def __init__(self, else framework.default_main_program() # FIXME(dzhwinter): enable_inplace should be after memory_optimize # if turn on python memory optimize, turn off the inplace_pass. + if build_strategy.memory_optimize is None: + build_strategy.memory_optimize = False if main._is_mem_optimized else True if build_strategy.enable_inplace is None: build_strategy.enable_inplace = False if main._is_mem_optimized else True scope = scope if scope is not None else executor.global_scope() diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index 534411219b5007..a1cf5fad138f06 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -77,6 +77,7 @@ list(REMOVE_ITEM TEST_OPS test_bilinear_interp_op) list(REMOVE_ITEM TEST_OPS test_nearest_interp_op) list(REMOVE_ITEM TEST_OPS test_imperative_resnet) list(REMOVE_ITEM TEST_OPS test_imperative_optimizer) +list(REMOVE_ITEM TEST_OPS test_ir_memory_optimize_transformer) foreach(TEST_OP ${TEST_OPS}) py_test_modules(${TEST_OP} MODULES ${TEST_OP}) endforeach(TEST_OP) @@ -107,14 +108,16 @@ py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SE py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL) set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 450) py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL) +if(NOT WIN32) +py_test_modules(test_ir_memory_optimize_transformer MODULES test_ir_memory_optimize_transformer SERIAL) +endif() if(NOT APPLE) py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL) - if(CMAKE_BUILD_TYPE STREQUAL "Debug") - # change the timeout from 600 to 1200, because in debug mode, this test need more time. - set_tests_properties(test_image_classification_resnet PROPERTIES TIMEOUT 1200) - endif() endif() - +if(CMAKE_BUILD_TYPE STREQUAL "Debug") + # change the timeout from 600 to 1200, because in debug mode, this test need more time. + set_tests_properties(test_parallel_executor_seresnext PROPERTIES TIMEOUT 1200) +endif() if (WITH_NGRAPH) add_subdirectory(ngraph) diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_accuracy_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_accuracy_ngraph_op.py index 84b9198dbf6569..5298c3c2f6f011 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_accuracy_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_accuracy_ngraph_op.py @@ -15,39 +15,7 @@ from __future__ import print_function import unittest -import numpy as np -from paddle.fluid.tests.unittests.op_test import OpTest - - -class TestNGRAPHAccuracyOp(OpTest): - def setUp(self): - self.op_type = "accuracy" - self.dtype = np.float32 - self.init_dtype() - n = 128 - infer = np.random.random((n, 1)).astype(self.dtype) - indices = np.random.randint(0, 2, (n, 1)) - label = np.random.randint(0, 2, (n, 1)) - self.inputs = {'Out': infer, 'Indices': indices, "Label": label} - num_correct = 0 - for rowid in range(n): - for ele in indices[rowid]: - if ele == label[rowid]: - num_correct += 1 - break - self.outputs = { - 'Accuracy': np.array([num_correct / float(n)]).astype(self.dtype), - 'Correct': np.array([num_correct]).astype("int64"), - 'Total': np.array([n]).astype("int64") - } - self._cpu_only = True - - def init_dtype(self): - pass - - def test_check_output(self): - self.check_output() - +from paddle.fluid.tests.unittests.test_accuracy_op import TestAccuracyOp if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_batch_norm_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_batch_norm_ngraph_op.py index 511173af5e5b2a..34fb73f3cf7e8b 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_batch_norm_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_batch_norm_ngraph_op.py @@ -17,21 +17,5 @@ import unittest from paddle.fluid.tests.unittests.test_batch_norm_op import TestBatchNormOpTraining, TestBatchNormOpInference - -class TestNGRAPHBatchNormOpTraining(TestBatchNormOpTraining): - def init_kernel_type(self): - super(TestNGRAPHBatchNormOpTraining, self).init_kernel_type() - - -class TestNGRAPHBatchNormOpInference(TestBatchNormOpInference): - def init_kernel_type(self): - super(TestNGRAPHBatchNormOpInference, self).init_kernel_type() - - -class TestNGRAPHBatchNormOpWithReluInference(TestBatchNormOpInference): - def init_kernel_type(self): - super(TestNGRAPHBatchNormOpWithReluInference, self).init_kernel_type() - - if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py index dbc8557b4e1c96..ff2e865b66a5f1 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_conv2d_ngraph_op.py @@ -17,60 +17,5 @@ import unittest from paddle.fluid.tests.unittests.test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride, TestWithGroup, TestWith1x1, TestWithInput1x1Filter1x1 - -class TestNGRAPH(TestConv2dOp): - def setUp(self): - super(TestNGRAPH, self).setUp() - self._cpu_only = True - - def init_kernel_type(self): - super(TestNGRAPH, self).init_kernel_type() - - -class TestNGRAPHWithPad(TestWithPad): - def setUp(self): - super(TestNGRAPHWithPad, self).setUp() - self._cpu_only = True - - def init_kernel_type(self): - super(TestNGRAPHWithPad, self).init_kernel_type() - - -class TestNGRAPHWithStride(TestWithStride): - def setUp(self): - super(TestNGRAPHWithStride, self).setUp() - self._cpu_only = True - - def init_kernel_type(self): - super(TestNGRAPHWithStride, self).init_kernel_type() - - -class TestNGRAPHWithGroup(TestWithGroup): - def setUp(self): - super(TestNGRAPHWithGroup, self).setUp() - self._cpu_only = True - - def init_kernel_type(self): - super(TestNGRAPHWithGroup, self).init_kernel_type() - - -class TestNGRAPHWith1x1(TestWith1x1): - def setUp(self): - super(TestNGRAPHWith1x1, self).setUp() - self._cpu_only = True - - def init_kernel_type(self): - super(TestNGRAPHWith1x1, self).init_kernel_type() - - -class TestNGRAPHWithInput1x1Filter1x1(TestWithInput1x1Filter1x1): - def setUp(self): - super(TestNGRAPHWithInput1x1Filter1x1, self).setUp() - self._cpu_only = True - - def init_kernel_type(self): - super(TestNGRAPHWithInput1x1Filter1x1, self).init_kernel_type() - - if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_cross_entropy_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_cross_entropy_ngraph_op.py new file mode 100644 index 00000000000000..3057218a1d80de --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_cross_entropy_ngraph_op.py @@ -0,0 +1,21 @@ +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +from paddle.fluid.tests.unittests.test_cross_entropy_op import TestCrossEntropyOp, TestCrossEntropyOp2, TestCrossEntropyOp3, TestCrossEntropyOp4, TestCrossEntropyOp5, TestCrossEntropyOp6, TestCrossEntropyOp7 + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_elementwise_add_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_elementwise_add_ngraph_op.py index 67f749bfeeb1bb..3fb9af3a542d5e 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_elementwise_add_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_elementwise_add_ngraph_op.py @@ -13,18 +13,9 @@ # limitations under the License. from __future__ import print_function -import unittest -from paddle.fluid.tests.unittests.test_elementwise_add_op import TestElementwiseAddOp - - -class TestNGRAPHElementwiseAddOp(TestElementwiseAddOp): - def setUp(self): - super(TestNGRAPHElementwiseAddOp, self).setUp() - self._cpu_only = True - - def init_input_output(self): - super(TestNGRAPHElementwiseAddOp, self).init_input_output() +import unittest +from paddle.fluid.tests.unittests.test_elementwise_add_op import TestElementwiseAddOp, TestElementwiseAddOp_broadcast_0 if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_fill_constant_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_fill_constant_ngraph_op.py index 835376ffe78f91..2b10b8f7a3ac0f 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_fill_constant_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_fill_constant_ngraph_op.py @@ -13,24 +13,34 @@ # limitations under the License. from __future__ import print_function + import unittest +import numpy as np from paddle.fluid.tests.unittests.test_fill_constant_op import TestFillConstantOp1, TestFillConstantOp2, TestFillConstantOpWithSelectedRows -class TestNGRAPHFillConstantOp1(TestFillConstantOp1): +class TestNGRAPHFillConstantFP64(TestFillConstantOp1): def setUp(self): - super(TestNGRAPHFillConstantOp1, self).setUp() + super(TestNGRAPHFillConstantFP64, self).setUp() + + self.attrs = {'shape': [123, 92], 'value': 3.8, 'dtype': 6} + self.outputs = {'Out': np.full((123, 92), 3.8)} -class TestNGRAPHFillConstantOp2(TestFillConstantOp2): +class TestNGRAPHFillConstantINT32(TestFillConstantOp2): def setUp(self): - super(TestNGRAPHFillConstantOp2, self).setUp() + super(TestNGRAPHFillConstantINT32, self).setUp() + self.attrs = {'shape': [123, 92], 'dtype': 2} + self.outputs = {'Out': np.full((123, 92), 0)} -class TestNGRAPHFillConstantOpWithSelectedRows( - TestFillConstantOpWithSelectedRows): + +class TestNGRAPHFillConstantINT64(TestFillConstantOp2): def setUp(self): - super(TestFillConstantOpWithSelectedRows, self).setUp() + super(TestNGRAPHFillConstantINT64, self).setUp() + + self.attrs = {'shape': [123, 92], 'dtype': 3} + self.outputs = {'Out': np.full((123, 92), 0)} if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_mean_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_mean_ngraph_op.py index 11881ac6e5292c..b4894734cbcc11 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_mean_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_mean_ngraph_op.py @@ -16,12 +16,5 @@ import unittest from paddle.fluid.tests.unittests.test_mean_op import TestMeanOp - -class TestNGRAPHMeanOp(TestMeanOp): - def setUp(self): - super(TestNGRAPHMeanOp, self).setUp() - self._cpu_only = True - - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_momentum_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_momentum_ngraph_op.py new file mode 100644 index 00000000000000..2c3549d907f5f6 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/ngraph/test_momentum_ngraph_op.py @@ -0,0 +1,21 @@ +# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +from paddle.fluid.tests.unittests.test_momentum_op import TestMomentumOp1, TestMomentumOp2, TestLarsMomentumOp, TestSparseMomentumOp, TestSparseMomentumOp2 + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_mul_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_mul_ngraph_op.py index a916c8d450f4a2..549d03f6e92dc7 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_mul_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_mul_ngraph_op.py @@ -15,39 +15,7 @@ from __future__ import print_function import unittest -import numpy as np -from paddle.fluid.tests.unittests.op_test import OpTest - - -class TestNGRAPHMulOp(OpTest): - def setUp(self): - self.op_type = "mul" - self.dtype = np.float32 - self.init_dtype_type() - self.inputs = { - 'X': np.random.random((2, 4)).astype(self.dtype), - 'Y': np.random.random((4, 4)).astype(self.dtype) - } - self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} - self._cpu_only = True - - def init_dtype_type(self): - pass - - def test_check_output(self): - self.check_output() - - def test_check_grad_normal(self): - self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5) - - def test_check_grad_ingore_x(self): - self.check_grad( - ['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X")) - - def test_check_grad_ingore_y(self): - self.check_grad( - ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) - +from paddle.fluid.tests.unittests.test_mul_op import TestMulOp, TestMulOp2 if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py index 96a2b72d8add9c..ff82e9fa1d3d34 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_pool2d_ngraph_op.py @@ -14,61 +14,25 @@ from __future__ import print_function -from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5 - - -class TestNGRAPHPool2D_Op(TestPool2D_Op): - def setUp(self): - super(TestNGRAPHPool2D_Op, self).setUp() - self._cpu_only = True - - def init_test_case(self): - super(TestNGRAPHPool2D_Op, self).init_test_case() - - -class TestNGRAPHCase1(TestCase1): - def setUp(self): - super(TestNGRAPHCase1, self).setUp() - self._cpu_only = True - - def init_test_case(self): - super(TestNGRAPHCase1, self).init_test_case() +import unittest - -class TestNGRAPHCase2(TestCase2): - def setUp(self): - super(TestNGRAPHCase2, self).setUp() - self._cpu_only = True - - def init_test_case(self): - super(TestNGRAPHCase2, self).init_test_case() - - -class TestNGRAPHCase3(TestCase3): - def setUp(self): - super(TestNGRAPHCase3, self).setUp() - self._cpu_only = True - - def init_pool_type(self): - super(TestNGRAPHCase3, self).init_pool_type() +from paddle.fluid.tests.unittests.test_pool2d_op import TestPool2D_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5 -class TestNGRAPHCase4(TestCase4): +class TestNGRAPHCeilMode(TestCase1): def setUp(self): - super(TestNGRAPHCase4, self).setUp() - self._cpu_only = True + super(TestNGRAPHCeilMode, self).setUp() - def init_pool_type(self): - super(TestNGRAPHCase4, self).init_pool_type() + def init_ceil_mode(self): + self.ceil_mode = True -class TestNGRAPHCase5(TestCase5): +class TestNGRAPHAdaptive(TestCase1): def setUp(self): - super(TestNGRAPHCase5, self).setUp() - self._cpu_only = True + super(TestNGRAPHAdaptive, self).setUp() - def init_pool_type(self): - super(TestNGRAPHCase5, self).init_pool_type() + def init_adaptive(self): + self.adaptive = True if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_scale_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_scale_ngraph_op.py index 4da5ca4583c65d..8beb44f55e487e 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_scale_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_scale_ngraph_op.py @@ -15,24 +15,5 @@ import unittest from paddle.fluid.tests.unittests.test_scale_op import TestScaleOp, TestScaleOpSelectedRows - -class TestNGRAPHScaleOp(TestScaleOp): - def setUp(self): - super(TestNGRAPHScaleOp, self).setUp() - self._cpu_only = True - - def init_dtype_type(self): - pass - - -class TestNGRAPHScaleOpSelectedRows(TestScaleOpSelectedRows): - def setUp(self): - super(TestNGRAPHScaleOpSelectedRows, self).setUp() - self._cpu_only = True - - def init_dtype_type(self): - pass - - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_softmax_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_softmax_ngraph_op.py index 81894c6e3872e4..0cb08842df0797 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_softmax_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_softmax_ngraph_op.py @@ -16,11 +16,5 @@ import unittest from paddle.fluid.tests.unittests.test_softmax_op import TestSoftmaxOp - -class TestSoftmaxNGRAPHOp(TestSoftmaxOp): - def setUp(self): - super(TestSoftmaxNGRAPHOp, self).setUp() - - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/ngraph/test_top_k_ngraph_op.py b/python/paddle/fluid/tests/unittests/ngraph/test_top_k_ngraph_op.py index fa68df1adf2cfb..d2319c4d921fcc 100644 --- a/python/paddle/fluid/tests/unittests/ngraph/test_top_k_ngraph_op.py +++ b/python/paddle/fluid/tests/unittests/ngraph/test_top_k_ngraph_op.py @@ -16,30 +16,5 @@ import unittest from paddle.fluid.tests.unittests.test_top_k_op import TestTopkOp, TestTopkOp3d, TestTopkOp2, TestTopkOp3, TestTopkOp4 - -class TestNGRAPHTopkOp(TestTopkOp): - def setUp(self): - super(TestNGRAPHTopkOp, self).setUp() - self._cpu_only = True - - -class TestNGRAPHTopkOp2(TestTopkOp2): - def setUp(self): - super(TestNGRAPHTopkOp2, self).setUp() - self._cpu_only = True - - -class TestNGRAPHTopkOp3(TestTopkOp3): - def setUp(self): - super(TestNGRAPHTopkOp3, self).setUp() - self._cpu_only = True - - -class TestNGRAPHTopkOp4(TestTopkOp4): - def setUp(self): - super(TestNGRAPHTopkOp4, self).setUp() - self._cpu_only = True - - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index 0fe836683b0296..823445724302db 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -14,6 +14,7 @@ from __future__ import print_function +import os import unittest import numpy as np import random @@ -374,6 +375,9 @@ def _get_places(self): return [] places = [fluid.CPUPlace()] cpu_only = self._cpu_only if hasattr(self, '_cpu_only') else False + use_ngraph = bool(os.getenv("FLAGS_use_ngraph", False)) + if use_ngraph: + cpu_only = True if core.is_compiled_with_cuda() and core.op_support_gpu(self.op_type)\ and not cpu_only: places.append(core.CUDAPlace(0)) diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py index c429c8af7d37cb..a94487e67dc90d 100644 --- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -79,7 +79,7 @@ def run_executor(exe, binary, feed, fetch_list): if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops build_strategy.fuse_relu_depthwise_conv = fuse_relu_depthwise_conv - build_strategy.memory_optimize = use_ir_memory_optimize + build_strategy.memory_optimize = False if memory_opt else use_ir_memory_optimize # python memory optimization is conflict with inplace pass. # Use ir graph memory optimization after inplace pass is the correct way. build_strategy.enable_inplace = False if memory_opt else enable_inplace diff --git a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py index 03471a4432f2b6..c1fb53ecf52d95 100644 --- a/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py +++ b/python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py @@ -121,6 +121,8 @@ def _optimizer(learning_rate=1e-6): regularization=fluid.regularizer.L2Decay(1e-6)) return optimizer + # NOTE(dzh): + # need to make it compatible with elewise fuse act not_fuse_op_first_loss, not_fuse_op_last_loss = self.check_network_convergence( model, feed_dict={"image": img, @@ -128,6 +130,7 @@ def _optimizer(learning_rate=1e-6): use_cuda=use_cuda, fuse_elewise_add_act_ops=False, memory_opt=False, + use_ir_memory_optimize=False, optimizer=_optimizer) fuse_op_first_loss, fuse_op_last_loss = self.check_network_convergence( model, @@ -136,6 +139,7 @@ def _optimizer(learning_rate=1e-6): use_cuda=use_cuda, fuse_elewise_add_act_ops=True, memory_opt=False, + use_ir_memory_optimize=False, optimizer=_optimizer) for loss in zip(not_fuse_op_first_loss, fuse_op_first_loss): diff --git a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py index 82aff18b728f45..7cf3bf13d2072b 100644 --- a/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py +++ b/python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py @@ -40,6 +40,8 @@ def __init__(self, self._dropout = dropout self._input = None self._num_steps = num_steps + from paddle.fluid.layer_helper import LayerHelper + self._helper = LayerHelper('SimpleLSTMRNN', act="tanh") def _build_once(self, input_embedding, init_hidden=None, init_cell=None): self.weight_1_arr = [] @@ -50,17 +52,21 @@ def _build_once(self, input_embedding, init_hidden=None, init_cell=None): self.mask_array = [] for i in range(self._num_layers): - weight_1 = fluid.layers.create_parameter( + weight_1 = self._helper.create_parameter( + attr=fluid.ParamAttr( + initializer=fluid.initializer.UniformInitializer( + low=-self._init_scale, high=self._init_scale)), shape=[self._hidden_size * 2, self._hidden_size * 4], dtype="float32", - name="fc_weight1_" + str(i), default_initializer=fluid.initializer.UniformInitializer( low=-self._init_scale, high=self._init_scale)) self.weight_1_arr.append(weight_1) - bias_1 = fluid.layers.create_parameter( - [self._hidden_size * 4], + bias_1 = self._helper.create_parameter( + attr=fluid.ParamAttr( + initializer=fluid.initializer.UniformInitializer( + low=-self._init_scale, high=self._init_scale)), + shape=[self._hidden_size * 4], dtype="float32", - name="fc_bias1_" + str(i), default_initializer=fluid.initializer.Constant(0.0)) self.bias_arr.append(bias_1) @@ -137,6 +143,8 @@ def __init__(self, self.num_layers = num_layers self.num_steps = num_steps self.dropout = dropout + from paddle.fluid.layer_helper import LayerHelper + self._helper = LayerHelper('PtbModel', act="tanh") self.simple_lstm_rnn = SimpleLSTMRNN( hidden_size, num_steps, @@ -151,16 +159,16 @@ def __init__(self, name='embedding_para', initializer=fluid.initializer.UniformInitializer( low=-init_scale, high=init_scale))) - self.softmax_weight = fluid.layers.create_parameter( - [self.hidden_size, self.vocab_size], + self.softmax_weight = self._helper.create_parameter( + attr=fluid.ParamAttr(), + shape=[self.hidden_size, self.vocab_size], dtype="float32", - name="softmax_weight", default_initializer=fluid.initializer.UniformInitializer( low=-self.init_scale, high=self.init_scale)) - self.softmax_bias = fluid.layers.create_parameter( - [self.vocab_size], + self.softmax_bias = self._helper.create_parameter( + attr=fluid.ParamAttr(), + shape=[self.vocab_size], dtype="float32", - name='softmax_bias', default_initializer=fluid.initializer.UniformInitializer( low=-self.init_scale, high=self.init_scale)) @@ -256,7 +264,6 @@ def test_ptb_rnn_cpu_float32(self): with new_program_scope(): fluid.default_startup_program().random_seed = seed fluid.default_main_program().random_seed = seed - # TODO: marsyang1993 Change seed to ptb_model = PtbModel( hidden_size=hidden_size, vocab_size=vocab_size, diff --git a/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py b/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py new file mode 100644 index 00000000000000..c0f480e34dcac3 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py @@ -0,0 +1,48 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +import unittest +import paddle.fluid as fluid +import paddle.fluid.core as core + +os.environ['FLAGS_eager_delete_tensor_gb'] = "0.0" +os.environ[ + 'RECORDIO_FILENAME'] = '/tmp/ir_memory_optimize_transformer.wmt16.recordio' + +from test_parallel_executor_transformer import TestTransformer +from test_parallel_executor_transformer import transformer + + +# NOTE(dzhwinter): test diferent strategy colisions. +# open the eager delete tensor strategy by default. +class TestTransformerWithIR(TestTransformer): + def test_main(self): + if core.is_compiled_with_cuda(): + # check python transpiler + self.check_network_convergence( + transformer, + use_cuda=True, + memory_opt=True, + use_ir_memory_optimize=False) + # check IR memory optimize + self.check_network_convergence( + transformer, + use_cuda=True, + memory_opt=False, + use_ir_memory_optimize=True) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py index 8fc391a1ff2529..69e060341ed9db 100644 --- a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py +++ b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py @@ -173,13 +173,16 @@ def lod_multiclass_nms(boxes, scores, background, score_threshold, normalized, shared=False) if nmsed_num == 0: - #lod.append(1) continue lod.append(nmsed_num) + tmp_det_out = [] for c, indices in nmsed_outs.items(): for idx in indices: xmin, ymin, xmax, ymax = box[idx, c, :] - det_outs.append([c, score[idx][c], xmin, ymin, xmax, ymax]) + tmp_det_out.append([c, score[idx][c], xmin, ymin, xmax, ymax]) + sorted_det_out = sorted( + tmp_det_out, key=lambda tup: tup[0], reverse=False) + det_outs.extend(sorted_det_out) if len(lod) == 0: lod.append(1) diff --git a/python/paddle/fluid/tests/unittests/test_optimizer.py b/python/paddle/fluid/tests/unittests/test_optimizer.py index 34c9b7e006950f..95ddc135b3da5b 100644 --- a/python/paddle/fluid/tests/unittests/test_optimizer.py +++ b/python/paddle/fluid/tests/unittests/test_optimizer.py @@ -274,7 +274,7 @@ def test_adagrad_optimizer(self): # Check init_program init_ops = init_program.global_block().ops - self.assertEqual(len(init_ops), 2) + self.assertEqual(len(init_ops), 3) self.assertEqual(init_ops[0].type, "fill_constant") self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate) self.assertEqual(init_ops[1].type, "fill_constant") diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_pg.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_pg.py new file mode 100644 index 00000000000000..041c56fce11e6f --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_pg.py @@ -0,0 +1,107 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest + +import numpy as np +import os +os.environ['FLAGS_enable_parallel_graph'] = str(1) +import paddle.fluid.core as core +import os +import paddle.fluid as fluid +from parallel_executor_test_base import TestParallelExecutorBase + + +def simple_fc_net(use_feed): + img = fluid.layers.data(name='image', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = img + for _ in range(4): + hidden = fluid.layers.fc( + hidden, + size=200, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + prediction = fluid.layers.fc(hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +class TestMNIST(TestParallelExecutorBase): + @classmethod + def setUpClass(cls): + os.environ['CPU_NUM'] = str(4) + + def _init_data(self): + np.random.seed(5) + img = np.random.random(size=[32, 784]).astype(np.float32) + label = np.ones(shape=[32, 1], dtype='int64') + return img, label + + # simple_fc + def check_simple_fc_convergence(self, use_cuda, use_reduce=False): + if use_cuda and not core.is_compiled_with_cuda(): + return + + img, label = self._init_data() + + self.check_network_convergence( + simple_fc_net, + feed_dict={"image": img, + "label": label}, + use_cuda=use_cuda, + use_reduce=use_reduce) + + def test_simple_fc(self): + # use_cuda + self.check_simple_fc_convergence(True) + + def check_simple_fc_parallel_accuracy(self, use_cuda): + if use_cuda and not core.is_compiled_with_cuda(): + return + + img, label = self._init_data() + + single_first_loss, single_last_loss = self.check_network_convergence( + method=simple_fc_net, + seed=1, + feed_dict={"image": img, + "label": label}, + use_cuda=use_cuda, + use_parallel_executor=False) + parallel_first_loss, parallel_last_loss = self.check_network_convergence( + method=simple_fc_net, + seed=1, + feed_dict={"image": img, + "label": label}, + use_cuda=use_cuda, + use_parallel_executor=True) + + self.assertAlmostEquals( + np.mean(parallel_first_loss), + single_first_loss, + delta=1e-6, ) + self.assertAlmostEquals( + np.mean(parallel_last_loss), single_last_loss, delta=1e-6) + + def test_simple_fc_parallel_accuracy(self): + self.check_simple_fc_parallel_accuracy(True) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_profiler.py b/python/paddle/fluid/tests/unittests/test_profiler.py index 7934164b84931f..39d778b82a04f4 100644 --- a/python/paddle/fluid/tests/unittests/test_profiler.py +++ b/python/paddle/fluid/tests/unittests/test_profiler.py @@ -16,15 +16,19 @@ import unittest import os +import tempfile import numpy as np import paddle.fluid as fluid import paddle.fluid.profiler as profiler import paddle.fluid.layers as layers import paddle.fluid.core as core +import paddle.fluid.proto.profiler.profiler_pb2 as profiler_pb2 class TestProfiler(unittest.TestCase): - def net_profiler(self, state, profile_path='/tmp/profile'): + def net_profiler(self, state, use_parallel_executor=False): + profile_path = os.path.join(tempfile.gettempdir(), "profile") + open(profile_path, "w").write("") startup_program = fluid.Program() main_program = fluid.Program() @@ -60,6 +64,11 @@ def net_profiler(self, state, profile_path='/tmp/profile'): place = fluid.CPUPlace() if state == 'CPU' else fluid.CUDAPlace(0) exe = fluid.Executor(place) exe.run(startup_program) + if use_parallel_executor: + pe = fluid.ParallelExecutor( + state != 'CPU', + loss_name=avg_cost.name, + main_program=main_program) pass_acc_calculator = fluid.average.WeightedAverage() with profiler.profiler(state, 'total', profile_path) as prof: @@ -69,6 +78,9 @@ def net_profiler(self, state, profile_path='/tmp/profile'): x = np.random.random((32, 784)).astype("float32") y = np.random.randint(0, 10, (32, 1)).astype("int64") + if use_parallel_executor: + pe.run(feed={'x': x, 'y': y}, fetch_list=[avg_cost.name]) + continue outs = exe.run(main_program, feed={'x': x, 'y': y}, @@ -77,21 +89,37 @@ def net_profiler(self, state, profile_path='/tmp/profile'): b_size = np.array(outs[2]) pass_acc_calculator.add(value=acc, weight=b_size) pass_acc = pass_acc_calculator.eval() + data = open(profile_path, 'rb').read() + self.assertGreater(len(data), 0) + profile_pb = profiler_pb2.Profile() + profile_pb.ParseFromString(data) + self.assertGreater(len(profile_pb.events), 0) + for event in profile_pb.events: + if event.type == profiler_pb2.Event.GPUKernel: + if not event.detail_info and not event.name.startswith("MEM"): + raise Exception( + "Kernel %s missing event. Has this kernel been recorded by RecordEvent?" + % event.name) + elif event.type == profiler_pb2.Event.CPU and ( + event.name.startswith("Driver API") or + event.name.startswith("Runtime API")): + print("Warning: unregister", event.name) def test_cpu_profiler(self): self.net_profiler('CPU') + self.net_profiler('CPU', use_parallel_executor=True) @unittest.skipIf(not core.is_compiled_with_cuda(), "profiler is enabled only with GPU") def test_cuda_profiler(self): self.net_profiler('GPU') + self.net_profiler('GPU', use_parallel_executor=True) @unittest.skipIf(not core.is_compiled_with_cuda(), "profiler is enabled only with GPU") def test_all_profiler(self): - self.net_profiler('All', '/tmp/profile_out') - with open('/tmp/profile_out', 'rb') as f: - self.assertGreater(len(f.read()), 0) + self.net_profiler('All') + self.net_profiler('All', use_parallel_executor=True) if __name__ == '__main__': diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index a3293afbbd7cef..eb54068650e8b3 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -1020,7 +1020,11 @@ def _get_slice_var_info(self, slice_var): skip_dim0 = 0 slice_vars = self.param_var_mapping[orig_var_name] - orig_dim1_flatten = reduce(lambda x, y: x * y, slice_vars[0].shape[1:]) + orig_dim1_flatten = 1 + + if len(slice_vars[0].shape) >= 2: + orig_dim1_flatten = reduce(lambda x, y: x * y, + slice_vars[0].shape[1:]) for slice_var in slice_vars[:block_idx]: skip_dim0 += slice_var.shape[0] diff --git a/python/requirements.txt b/python/requirements.txt index 5a70f1aa3ffc0a..36bd5d4261cc7a 100644 --- a/python/requirements.txt +++ b/python/requirements.txt @@ -1,6 +1,6 @@ requests==2.9.2 numpy>=1.12 -protobuf==3.1 +protobuf>=3.1.0 recordio>=0.1.0 matplotlib==2.2.3 # TODO: let python3 paddlepaddle package use latest matplotlib rarfile diff --git a/tools/manylinux1/Dockerfile.x64 b/tools/manylinux1/Dockerfile.x64 index 48fd145e5fe673..c2fd743f62f536 100644 --- a/tools/manylinux1/Dockerfile.x64 +++ b/tools/manylinux1/Dockerfile.x64 @@ -31,10 +31,10 @@ RUN wget --no-check-certificate -qO- https://storage.googleapis.com/golang/go1.8 ENV GOROOT=/usr/local/go GOPATH=/root/gopath ENV PATH=${GOROOT}/bin:${GOPATH}/bin:${PATH} -# protobuf 3.1.0 -RUN cd /opt && wget -q --no-check-certificate https://github.com/google/protobuf/releases/download/v3.1.0/protobuf-cpp-3.1.0.tar.gz && \ - tar xzf protobuf-cpp-3.1.0.tar.gz && \ - cd protobuf-3.1.0 && ./configure && make -j4 && make install && cd .. && rm -f protobuf-cpp-3.1.0.tar.gz +# protobuf 3.6.1 +RUN cd /opt && wget -q --no-check-certificate https://github.com/google/protobuf/releases/download/v3.6.1/protobuf-cpp-3.6.1.tar.gz && \ + tar xzf protobuf-cpp-3.6.1.tar.gz && \ + cd protobuf-3.6.1 && ./configure && make -j4 && make install && cd .. && rm -f protobuf-cpp-3.6.1.tar.gz RUN wget https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/python/requirements.txt -O /root/requirements.txt diff --git a/tools/manylinux1/build_all.sh b/tools/manylinux1/build_all.sh index 097bedb5265d00..caf21722158b74 100755 --- a/tools/manylinux1/build_all.sh +++ b/tools/manylinux1/build_all.sh @@ -24,3 +24,8 @@ sed 's//9.0-cudnn7-devel-centos6/g' Dockerfile.x64 | \ sed 's//NVCC_GENCODE="-gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_60,code=compute_60 -gencode=arch=compute_61,code=sm_61 -gencode=arch=compute_62,code=sm_62 -gencode=arch=compute_70,code=sm_70"/g'> Dockerfile.tmp docker build -t ${REPO}/paddle_manylinux_devel:cuda9.0_cudnn7 -f Dockerfile.tmp . docker push ${REPO}/paddle_manylinux_devel:cuda9.0_cudnn7 + +sed 's//10.0-devel-centos6/g' Dockerfile.x64 | \ +sed 's//NVCC_GENCODE="-gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_60,code=sm_60 -gencode=arch=compute_60,code=compute_60 -gencode=arch=compute_61,code=sm_61 -gencode=arch=compute_62,code=sm_62 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_75,code=sm_75"/g'> Dockerfile.tmp +docker build -t ${REPO}/paddle_manylinux_devel:cuda10.0_cudnn7 -f Dockerfile.tmp . +docker push ${REPO}/paddle_manylinux_devel:cuda10.0_cudnn7 diff --git a/tools/manylinux1/build_scripts/build.sh b/tools/manylinux1/build_scripts/build.sh index 6c551eceb4543b..1b0059a8c69fca 100644 --- a/tools/manylinux1/build_scripts/build.sh +++ b/tools/manylinux1/build_scripts/build.sh @@ -17,7 +17,7 @@ OPENSSL_ROOT=openssl-1.1.0i OPENSSL_HASH=ebbfc844a8c8cc0ea5dc10b86c9ce97f401837f3fa08c17b2cdadc118253cf99 EPEL_RPM_HASH=e5ed9ecf22d0c4279e92075a64c757ad2b38049bcf5c16c4f2b75d5f6860dc0d DEVTOOLS_HASH=a8ebeb4bed624700f727179e6ef771dafe47651131a00a78b342251415646acc -PATCHELF_HASH=d9afdff4baeacfbc64861454f368b7f2c15c44d245293f7587bbf726bfe722fb +PATCHELF_HASH=f2aa40a6148cb3b0ca807a1bf836b081793e55ec9e5540a5356d800132be7e0a CURL_ROOT=curl-7.49.1 CURL_HASH=eb63cec4bef692eab9db459033f409533e6d10e20942f4b060b32819e81885f1 AUTOCONF_ROOT=autoconf-2.69 @@ -107,11 +107,13 @@ curl-config --features rm -rf /usr/local/ssl # Install patchelf (latest with unreleased bug fixes) -curl -sLO http://nipy.bic.berkeley.edu/manylinux/patchelf-0.9njs2.tar.gz -check_sha256sum patchelf-0.9njs2.tar.gz $PATCHELF_HASH -tar -xzf patchelf-0.9njs2.tar.gz -(cd patchelf-0.9njs2 && ./configure && make && make install) -rm -rf patchelf-0.9njs2.tar.gz patchelf-0.9njs2 +# FIXME(typhoonzero): restore this when the link is fixed. +# curl -sLO http://nipy.bic.berkeley.edu/manylinux/patchelf-0.9njs2.tar.gz +# check_sha256sum patchelf-0.9njs2.tar.gz $PATCHELF_HASH +# tar -xzf patchelf-0.9njs2.tar.gz +# (cd patchelf-0.9njs2 && ./configure && make && make install) +# rm -rf patchelf-0.9njs2.tar.gz patchelf-0.9njs2 +yum install -y patchelf # Install latest pypi release of auditwheel LD_LIBRARY_PATH="${ORIGINAL_LD_LIBRARY_PATH}:$(dirname ${PY35_BIN})/lib" $PY35_BIN/pip install auditwheel diff --git a/tools/manylinux1/build_scripts/build_utils.sh b/tools/manylinux1/build_scripts/build_utils.sh index 48cce15a145138..083101249cd856 100755 --- a/tools/manylinux1/build_scripts/build_utils.sh +++ b/tools/manylinux1/build_scripts/build_utils.sh @@ -87,6 +87,8 @@ function do_cpython_build { # NOTE Make libpython shared library visible to python calls below LD_LIBRARY_PATH="${prefix}/lib" ${prefix}/bin/python get-pip.py LD_LIBRARY_PATH="${prefix}/lib" ${prefix}/bin/pip install wheel + cd / + ls ${MY_DIR} local abi_tag=$(LD_LIBRARY_PATH="${prefix}/lib" ${prefix}/bin/python ${MY_DIR}/python-tag-abi-tag.py) ln -s ${prefix} /opt/python/${abi_tag} } diff --git a/tools/timeline.py b/tools/timeline.py index f850476831d847..ebadb29bdbe00c 100644 --- a/tools/timeline.py +++ b/tools/timeline.py @@ -131,8 +131,12 @@ def _allocate_pids(self): if (k, event.device_id, "CPU") not in self._devices: pid = self._allocate_pid() self._devices[(k, event.device_id, "CPU")] = pid - self._chrome_trace.emit_pid("%s:cpu:block:%d" % - (k, event.device_id), pid) + # -1 device id represents CUDA api call + if event.device_id == -1: + self._chrome_trace.emit_pid("%s:cuda_api" % k, pid) + else: + self._chrome_trace.emit_pid( + "%s:cpu:block:%d" % (k, event.device_id), pid) elif event.type == profiler_pb2.Event.GPUKernel: if (k, event.device_id, "GPUKernel") not in self._devices: pid = self._allocate_pid() @@ -150,7 +154,9 @@ def _allocate_events(self): pid = self._devices[(k, event.device_id, type)] args = {'name': event.name} if event.memcopy.bytes > 0: - args = {'mem_bytes': event.memcopy.bytes} + args['mem_bytes'] = event.memcopy.bytes + if event.detail_info: + args['detail_info'] = event.detail_info # TODO(panyx0718): Chrome tracing only handles ms. However, some # ops takes micro-seconds. Hence, we keep the ns here. self._chrome_trace.emit_region( @@ -173,7 +179,7 @@ def generate_chrome_trace(self): profile_paths = profile_path.split(',') profile_dict = dict() if len(profile_paths) == 1: - with open(profile_path, 'r') as f: + with open(profile_path, 'rb') as f: profile_s = f.read() profile_pb = profiler_pb2.Profile() profile_pb.ParseFromString(profile_s) @@ -181,7 +187,7 @@ def generate_chrome_trace(self): else: for profile_path in profile_paths: k, v = profile_path.split('=') - with open(v, 'r') as f: + with open(v, 'rb') as f: profile_s = f.read() profile_pb = profiler_pb2.Profile() profile_pb.ParseFromString(profile_s)