diff --git a/.gitignore b/.gitignore index 90138f996cf9cacc3c1cbff0cf2600eefca3f305..fa0c8882606b76ac71b43dcde7e1df6770c46c31 100644 --- a/.gitignore +++ b/.gitignore @@ -28,3 +28,4 @@ third_party/ build_* # clion workspace. cmake-build-* +model_test diff --git a/CMakeLists.txt b/CMakeLists.txt index df00e977ebb547980e69ee421779c57717d771a9..291a960b1471b22a6cb53c4ca49b45609afb4dc6 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -41,6 +41,7 @@ option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_F option(WITH_AMD_GPU "Compile PaddlePaddle with AMD GPU" OFF) option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND}) option(WITH_MKL "Compile PaddlePaddle with MKL support." ${AVX_FOUND}) +option(WITH_NGRAPH "Compile PaddlePaddle with nGraph support." OFF) option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON) option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF) option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON) @@ -62,13 +63,12 @@ option(WITH_DISTRIBUTE "Compile with distributed 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_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" 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) option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE}) option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF) -option(WITH_INFERENCE "Compile fluid inference library" ON) +option(ON_INFER "Turn on inference optimization." OFF) option(WITH_INFERENCE_API_TEST "Test fluid inference high-level api interface" OFF) option(WITH_SYSTEM_BLAS "Use system blas library" OFF) option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION}) @@ -104,6 +104,8 @@ if(ANDROID OR IOS) "Disable RDMA when cross-compiling for Android and iOS" FORCE) set(WITH_MKL OFF CACHE STRING "Disable MKL when cross-compiling for Android and iOS" FORCE) + set(WITH_NGRAPH OFF CACHE STRING + "Disable nGraph when cross-compiling for Android and iOS" FORCE) set(WITH_GOLANG OFF CACHE STRING "Disable golang when cross-compiling for Android and iOS" FORCE) @@ -127,6 +129,9 @@ set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING "A path setting fluid shared and static libraries") +set(FLUID_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_inference_install_dir" CACHE STRING + "A path setting fluid inference shared and static libraries") + if (WITH_C_API AND WITH_PYTHON) message(WARNING "It is suggest not embedded a python interpreter in Paddle " "when using C-API. It will give an unpredictable behavior when using a " @@ -169,6 +174,7 @@ include(external/protobuf) # download, build, install protobuf include(external/python) # download, build, install python include(external/openblas) # download, build, install openblas include(external/mkldnn) # download, build, install mkldnn +include(external/ngraph) # download, build, install nGraph include(external/swig) # download, build, install swig include(external/boost) # download boost include(external/any) # download libn::any @@ -176,6 +182,7 @@ include(external/eigen) # download eigen3 include(external/pybind11) # download pybind11 include(external/cares) include(external/cub) +include(external/xxhash) # download xxhash if (NOT WIN32) # there is no official support of snappystream, warpctc, nccl, cupti in windows @@ -298,3 +305,11 @@ if(WITH_DOC) find_python_module(recommonmark REQUIRED) add_subdirectory(doc) endif() + +if (ON_INFER) + message(STATUS "On inference mode, will take place some specific optimization.") + add_definitions(-DPADDLE_ON_INFERENCE) +else() + #TODO(luotao), combine this warning with `make inference_lib_dist` command. + message(WARNING "On inference mode, will take place some specific optimization. Turn on the ON_INFER flag when building inference_lib only.") +endif() diff --git a/Dockerfile b/Dockerfile index 738bba9bc2e1ab19709722fe04f1490b1b13bd4f..c8b9eed6d60e5d3b32fc14c0c7af80a785145d1b 100644 --- a/Dockerfile +++ b/Dockerfile @@ -75,14 +75,14 @@ RUN pip3 install -U wheel && \ pip3 install -U docopt PyYAML sphinx==1.5.6 && \ pip3 install sphinx-rtd-theme==0.1.9 recommonmark && \ easy_install -U pip && \ - pip install -U wheel && \ + pip install -U pip setuptools wheel && \ pip install -U docopt PyYAML sphinx==1.5.6 && \ pip install sphinx-rtd-theme==0.1.9 recommonmark -RUN pip3 install pre-commit 'ipython==5.3.0' && \ +RUN pip3 install 'pre-commit==1.10.4' 'ipython==5.3.0' && \ pip3 install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ pip3 install opencv-python && \ - pip install pre-commit 'ipython==5.3.0' && \ + pip install 'pre-commit==1.10.4' 'ipython==5.3.0' && \ pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \ pip install opencv-python diff --git a/README.md b/README.md index 46fdef5e376d3f5bf49ef10c62f5b3a6637913c1..56d6c10c642787836abb55cb2974bda0b8d22da4 100644 --- a/README.md +++ b/README.md @@ -2,8 +2,8 @@ [![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle) -[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html) -[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html) +[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) +[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) [![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases) [![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE) @@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle. Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle. -### Latest PaddlePaddle Release: [Fluid 0.15.0](https://github.com/PaddlePaddle/Paddle/tree/v0.15.0) +### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1) ### Install Latest Stable Release: ``` # Linux CPU @@ -27,9 +27,9 @@ pip install paddlepaddle # Linux GPU cuda9cudnn7 pip install paddlepaddle-gpu # Linux GPU cuda8cudnn7 -pip install paddlepaddle-gpu==0.15.0.post87 +pip install paddlepaddle-gpu==1.1.0.post87 # Linux GPU cuda8cudnn5 -pip install paddlepaddle-gpu==0.15.0.post85 +pip install paddlepaddle-gpu==1.1.0.post85 # For installation on other platform, refer to http://paddlepaddle.org/ ``` @@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==0.15.0.post85 ## Installation -It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/install/install_doc.html) on our website. +It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website. ## Documentation -We provide [English](http://paddlepaddle.org/documentation/docs/en/0.15.0/getstarted/index_en.html) and -[Chinese](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/beginners_guide/index.html) documentation. +We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and +[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation. - [Deep Learning 101](https://github.com/PaddlePaddle/book) You might want to start from this online interactive book that can run in a Jupyter Notebook. -- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/user_guides/howto/training/cluster_howto.html) +- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html) You can run distributed training jobs on MPI clusters. -- [Python API](http://paddlepaddle.org/documentation/api/zh/0.15.0/fluid.html) +- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html) Our new API enables much shorter programs. -- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/0.15.0/new_docs/advanced_usage/development/contribute_to_paddle.html) +- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html) We appreciate your contributions! diff --git a/benchmark/fluid/args.py b/benchmark/fluid/args.py index 9540900b112f54594bbfdbc8d7cd3b6e1f5269dd..ff616ddbb2cb1cb7f348d6d164815823b08b7629 100644 --- a/benchmark/fluid/args.py +++ b/benchmark/fluid/args.py @@ -142,5 +142,10 @@ def parse_args(): choices=['reduce', 'all_reduce'], default='all_reduce', help='Specify the reduce strategy, can be reduce, all_reduce') + parser.add_argument( + '--fuse_broadcast_op', + action='store_true', + help='If set, would fuse multiple broadcast operators into one fused_broadcast operator.' + ) args = parser.parse_args() return args diff --git a/benchmark/fluid/fluid_benchmark.py b/benchmark/fluid/fluid_benchmark.py index ddd9fe809853a830ca676cc98f1819f683866def..5f3ce300acc44ad8d2898c27296b866c403f3cc8 100644 --- a/benchmark/fluid/fluid_benchmark.py +++ b/benchmark/fluid/fluid_benchmark.py @@ -177,6 +177,7 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog, else: build_strategy.reduce_strategy = fluid.BuildStrategy( ).ReduceStrategy.AllReduce + build_strategy.fuse_broadcast_op = args.fuse_broadcast_op avg_loss = train_args[0] @@ -240,7 +241,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog, if args.use_fake_data or args.use_reader_op: try: - fetch_ret = exe.run(fetch_list) except fluid.core.EOFException as eof: break diff --git a/benchmark/fluid/run.sh b/benchmark/fluid/run.sh old mode 100644 new mode 100755 diff --git a/cmake/configure.cmake b/cmake/configure.cmake index e9852f00b1835adec31373f58ac538f9685251e0..7f5771e561f6cc419fc9b3094174645ac432546e 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -50,11 +50,7 @@ if(NOT WITH_PROFILER) endif(NOT WITH_PROFILER) if(NOT CMAKE_CROSSCOMPILING) - if(WITH_AVX AND AVX512F_FOUND) - set(SIMD_FLAG ${AVX512F_FLAG}) - elseif(WITH_AVX AND AVX2_FOUND) - set(SIMD_FLAG ${AVX2_FLAG}) - elseif(WITH_AVX AND AVX_FOUND) + if(WITH_AVX AND AVX_FOUND) set(SIMD_FLAG ${AVX_FLAG}) elseif(SSE3_FOUND) set(SIMD_FLAG ${SSE3_FLAG}) diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake index baf253df2755657b01b67c410f63b7d8422d4df3..785148d4f9f44032e2ce5bf93f0dc80fc865808b 100644 --- a/cmake/external/mkldnn.cmake +++ b/cmake/external/mkldnn.cmake @@ -37,7 +37,6 @@ SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE) SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib") INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers. -INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include mkldnn.h IF(${CBLAS_PROVIDER} STREQUAL "MKLML") SET(MKLDNN_DEPENDS ${MKLML_PROJECT}) @@ -45,7 +44,7 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML") ELSE() MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN") ENDIF() -SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result") +SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-error=array-bounds") SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value") SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}") SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}") @@ -54,7 +53,7 @@ ExternalProject_Add( ${EXTERNAL_PROJECT_LOG_ARGS} DEPENDS ${MKLDNN_DEPENDS} GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git" - GIT_TAG "64e03a1939e0d526aa8e9f2e3f7dc0ad8d372944" + GIT_TAG "21fb5f2af1dd14e132af4f1b79160977ee487818" PREFIX ${MKLDNN_SOURCES_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER} diff --git a/cmake/external/ngraph.cmake b/cmake/external/ngraph.cmake new file mode 100644 index 0000000000000000000000000000000000000000..2e335579f32df4f146c8d88e05e684a9a8105e20 --- /dev/null +++ b/cmake/external/ngraph.cmake @@ -0,0 +1,92 @@ +# 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. + +add_library(ngraph INTERFACE) + +IF(WIN32 OR APPLE) + MESSAGE(WARNING + "Windows or Mac is not supported with nGraph in Paddle yet." + "Force WITH_NGRAPH=OFF") + SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph in Windows and MacOS" FORCE) +ENDIF() + +IF(${WITH_NGRAPH} AND NOT ${WITH_MKLDNN}) + MESSAGE(WARNING + "nGraph needs mkl-dnn to be enabled." + "Force WITH_NGRAPH=OFF") + SET(WITH_NGRAPH OFF CACHE STRING "Disable nGraph if mkl-dnn is disabled" FORCE) +ENDIF() + +IF(NOT ${WITH_NGRAPH}) + return() +ENDIF() + +INCLUDE(ExternalProject) + +SET(NGRAPH_PROJECT "extern_ngraph") +SET(NGRAPH_VERSION "0.9") +SET(NGRAPH_GIT_TAG "f9fd9d4cc318dc59dd4b68448e7fbb5f67a28bd0") +SET(NGRAPH_SOURCES_DIR ${THIRD_PARTY_PATH}/ngraph) +SET(NGRAPH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/ngraph) +SET(NGRAPH_INC_DIR ${NGRAPH_INSTALL_DIR}/include) +SET(NGRAPH_SHARED_LIB_NAME libngraph.so.${NGRAPH_VERSION}) +SET(NGRAPH_CPU_LIB_NAME libcpu_backend.so) +SET(NGRAPH_TBB_LIB_NAME libtbb.so.2) +SET(NGRAPH_GIT_REPO "https://github.com/NervanaSystems/ngraph.git") + +ExternalProject_Add( + ${NGRAPH_PROJECT} + ${EXTERNAL_PROJECT_LOG_ARGS} + DEPENDS ${MKLDNN_PROJECT} ${MKLML_PROJECT} + GIT_REPOSITORY ${NGRAPH_GIT_REPO} + GIT_TAG ${NGRAPH_GIT_TAG} + PREFIX ${NGRAPH_SOURCES_DIR} + UPDATE_COMMAND "" + CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR} + CMAKE_ARGS -DNGRAPH_UNIT_TEST_ENABLE=FALSE + CMAKE_ARGS -DNGRAPH_TOOLS_ENABLE=FALSE + CMAKE_ARGS -DNGRAPH_INTERPRETER_ENABLE=FALSE + CMAKE_ARGS -DNGRAPH_DEX_ONLY=TRUE + CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} + CMAKE_ARGS -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR} + CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib +) + +if(UNIX AND NOT APPLE) + include(GNUInstallDirs) + SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}) +else() + SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/lib) +endif() +MESSAGE(STATUS "nGraph lib will be installed at: ${NGRAPH_LIB_DIR}") + +SET(NGRAPH_SHARED_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME}) +SET(NGRAPH_CPU_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME}) +SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME}) + +# Workaround for nGraph expecting mklml to be in mkldnn install directory. +ExternalProject_Add_Step( + ${NGRAPH_PROJECT} + PrepareMKL + COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_LIB} ${MKLDNN_INSTALL_DIR}/lib/libmklml_intel.so + COMMAND ${CMAKE_COMMAND} -E create_symlink ${MKLML_IOMP_LIB} ${MKLDNN_INSTALL_DIR}/lib/libiomp5.so + DEPENDEES download + DEPENDERS configure +) + +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/protobuf.cmake b/cmake/external/protobuf.cmake index 550b0dada8e90c1e2b33705fd53c065672113b45..45ef9b4550291cadaa9571f05dbaefdf4a0c223a 100644 --- a/cmake/external/protobuf.cmake +++ b/cmake/external/protobuf.cmake @@ -30,66 +30,61 @@ UNSET_VAR(PROTOBUF_LITE_LIBRARY) UNSET_VAR(PROTOBUF_LIBRARY) UNSET_VAR(PROTOBUF_INCLUDE_DIR) UNSET_VAR(Protobuf_PROTOC_EXECUTABLE) +function(protobuf_generate_python SRCS) + # shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake + if(NOT ARGN) + message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files") + return() + endif() -if(NOT COMMAND protobuf_generate_python) # before cmake 3.4, protobuf_genrerate_python is not defined. - function(protobuf_generate_python SRCS) - # shameless copy from https://github.com/Kitware/CMake/blob/master/Modules/FindProtobuf.cmake - if(NOT ARGN) - message(SEND_ERROR "Error: PROTOBUF_GENERATE_PYTHON() called without any proto files") - return() - endif() - - if(PROTOBUF_GENERATE_CPP_APPEND_PATH) - # Create an include path for each file specified - foreach(FIL ${ARGN}) - get_filename_component(ABS_FIL ${FIL} ABSOLUTE) - get_filename_component(ABS_PATH ${ABS_FIL} PATH) - list(FIND _protobuf_include_path ${ABS_PATH} _contains_already) - if(${_contains_already} EQUAL -1) - list(APPEND _protobuf_include_path -I ${ABS_PATH}) - endif() - endforeach() - else() - set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR}) - endif() - - if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS) - set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}") - endif() - - if(DEFINED Protobuf_IMPORT_DIRS) - foreach(DIR ${Protobuf_IMPORT_DIRS}) - get_filename_component(ABS_PATH ${DIR} ABSOLUTE) - list(FIND _protobuf_include_path ${ABS_PATH} _contains_already) - if(${_contains_already} EQUAL -1) - list(APPEND _protobuf_include_path -I ${ABS_PATH}) - endif() - endforeach() - endif() - - set(${SRCS}) + if(PROTOBUF_GENERATE_CPP_APPEND_PATH) + # Create an include path for each file specified foreach(FIL ${ARGN}) get_filename_component(ABS_FIL ${FIL} ABSOLUTE) - get_filename_component(FIL_WE ${FIL} NAME_WE) - if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH) - get_filename_component(FIL_DIR ${FIL} DIRECTORY) - if(FIL_DIR) - set(FIL_WE "${FIL_DIR}/${FIL_WE}") - endif() + get_filename_component(ABS_PATH ${ABS_FIL} PATH) + list(FIND _protobuf_include_path ${ABS_PATH} _contains_already) + if(${_contains_already} EQUAL -1) + list(APPEND _protobuf_include_path -I ${ABS_PATH}) endif() + endforeach() + else() + set(_protobuf_include_path -I ${CMAKE_CURRENT_SOURCE_DIR}) + endif() + if(DEFINED PROTOBUF_IMPORT_DIRS AND NOT DEFINED Protobuf_IMPORT_DIRS) + set(Protobuf_IMPORT_DIRS "${PROTOBUF_IMPORT_DIRS}") + endif() - list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py") - add_custom_command( - OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py" - COMMAND ${Protobuf_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL} - DEPENDS ${ABS_FIL} ${Protobuf_PROTOC_EXECUTABLE} - COMMENT "Running Python protocol buffer compiler on ${FIL}" - VERBATIM ) + if(DEFINED Protobuf_IMPORT_DIRS) + foreach(DIR ${Protobuf_IMPORT_DIRS}) + get_filename_component(ABS_PATH ${DIR} ABSOLUTE) + list(FIND _protobuf_include_path ${ABS_PATH} _contains_already) + if(${_contains_already} EQUAL -1) + list(APPEND _protobuf_include_path -I ${ABS_PATH}) + endif() endforeach() + endif() - set(${SRCS} ${${SRCS}} PARENT_SCOPE) - endfunction() -endif() + set(${SRCS}) + foreach(FIL ${ARGN}) + get_filename_component(ABS_FIL ${FIL} ABSOLUTE) + get_filename_component(FIL_WE ${FIL} NAME_WE) + if(NOT PROTOBUF_GENERATE_CPP_APPEND_PATH) + get_filename_component(FIL_DIR ${FIL} DIRECTORY) + if(FIL_DIR) + set(FIL_WE "${FIL_DIR}/${FIL_WE}") + endif() + endif() + list(APPEND ${SRCS} "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py") + add_custom_command( + OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/${FIL_WE}_pb2.py" + COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} --python_out ${CMAKE_CURRENT_BINARY_DIR} ${_protobuf_include_path} ${ABS_FIL} + DEPENDS ${ABS_FIL} ${PROTOBUF_PROTOC_EXECUTABLE} + COMMENT "Running Python protocol buffer compiler on ${FIL}" + VERBATIM ) + endforeach() + + set(${SRCS} ${${SRCS}} PARENT_SCOPE) +endfunction() # Print and set the protobuf library information, # finish this cmake process and exit from this file. @@ -126,6 +121,7 @@ macro(PROMPT_PROTOBUF_LIB) # FIND_Protobuf.cmake uses `Protobuf_PROTOC_EXECUTABLE`. # make `protobuf_generate_cpp` happy. SET(Protobuf_PROTOC_EXECUTABLE ${PROTOBUF_PROTOC_EXECUTABLE}) + FOREACH(dep ${protobuf_DEPS}) ADD_DEPENDENCIES(protobuf ${dep}) ADD_DEPENDENCIES(protobuf_lite ${dep}) diff --git a/cmake/external/xxhash.cmake b/cmake/external/xxhash.cmake new file mode 100644 index 0000000000000000000000000000000000000000..c227e09719bd5f0e825f81fb96f78105aa10c79b --- /dev/null +++ b/cmake/external/xxhash.cmake @@ -0,0 +1,50 @@ +INCLUDE(ExternalProject) + +set(XXHASH_SOURCE_DIR ${THIRD_PARTY_PATH}/xxhash) +set(XXHASH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/xxhash) +set(XXHASH_INCLUDE_DIR "${XXHASH_INSTALL_DIR}/include") + +IF(WITH_STATIC_LIB) + SET(BUILD_CMD make lib) +ELSE() + IF(APPLE) + SET(BUILD_CMD sed -i \"\" "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib) + ELSE(APPLE) + SET(BUILD_CMD sed -i "s/-Wstrict-prototypes -Wundef/-Wstrict-prototypes -Wundef -fPIC/g" ${XXHASH_SOURCE_DIR}/src/extern_xxhash/Makefile && make lib) + ENDIF(APPLE) +ENDIF() + +ExternalProject_Add( + extern_xxhash + ${EXTERNAL_PROJECT_LOG_ARGS} + GIT_REPOSITORY "https://github.com/Cyan4973/xxHash" + GIT_TAG "v0.6.5" + PREFIX ${XXHASH_SOURCE_DIR} + DOWNLOAD_NAME "xxhash" + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_IN_SOURCE 1 + PATCH_COMMAND + BUILD_COMMAND ${BUILD_CMD} + INSTALL_COMMAND export PREFIX=${XXHASH_INSTALL_DIR}/ && make install + TEST_COMMAND "" +) + +set(XXHASH_LIBRARIES "${XXHASH_INSTALL_DIR}/lib/libxxhash.a") +INCLUDE_DIRECTORIES(${XXHASH_INCLUDE_DIR}) + +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) + +IF(WITH_C_API) + INSTALL(DIRECTORY ${XXHASH_INCLUDE_DIR} DESTINATION third_party/xxhash) + IF(ANDROID) + INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib/${ANDROID_ABI}) + ELSE() + INSTALL(FILES ${XXHASH_LIBRARIES} DESTINATION third_party/xxhash/lib) + ENDIF() +ENDIF() diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 5bf82b4ddf10a646ca540ac4ee2cfd3d3bc6cf58..62227c67849dbb476339a176e0c98e295cbf529c 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -261,6 +261,13 @@ function(cc_library TARGET_NAME) add_dependencies(${TARGET_NAME} mklml) target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed") endif() + # remove link to python, see notes at: + # https://github.com/pybind/pybind11/blob/master/docs/compiling.rst#building-manually + if("${cc_library_DEPS};" MATCHES "python;") + list(REMOVE_ITEM cc_library_DEPS python) + add_dependencies(${TARGET_NAME} python) + target_link_libraries(${TARGET_NAME} "-Wl,-undefined,dynamic_lookup") + endif() target_link_libraries(${TARGET_NAME} ${cc_library_DEPS}) add_dependencies(${TARGET_NAME} ${cc_library_DEPS}) endif() @@ -311,6 +318,8 @@ function(cc_test TARGET_NAME) set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true) set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true) set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true) + # No unit test should exceed 10 minutes. + set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600) endif() endfunction(cc_test) @@ -629,6 +638,8 @@ function(py_test TARGET_NAME) PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS} ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}) + # No unit test should exceed 10 minutes. + set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600) endif() endfunction() diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index 077072f6eadb0c48f4ae32f94828613d89ed01c9..efdb093a7b28e19f3b2a774dd54f2e7f042e9ca7 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -18,7 +18,7 @@ function(copy TARGET) set(oneValueArgs "") set(multiValueArgs SRCS DSTS DEPS) cmake_parse_arguments(copy_lib "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - set(inference_lib_dist_dep ${TARGET} ${inference_lib_dist_dep} PARENT_SCOPE) + set(fluid_lib_dist_dep ${TARGET} ${fluid_lib_dist_dep} PARENT_SCOPE) list(LENGTH copy_lib_SRCS copy_lib_SRCS_len) list(LENGTH copy_lib_DSTS copy_lib_DSTS_len) @@ -31,7 +31,7 @@ function(copy TARGET) foreach(index RANGE ${len}) list(GET copy_lib_SRCS ${index} src) list(GET copy_lib_DSTS ${index} dst) - add_custom_command(TARGET ${TARGET} PRE_BUILD + add_custom_command(TARGET ${TARGET} PRE_BUILD COMMAND mkdir -p "${dst}" COMMAND cp -r "${src}" "${dst}" COMMENT "copying ${src} -> ${dst}") @@ -67,6 +67,13 @@ copy(boost_lib DEPS boost ) +set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/xxhash") +copy(xxhash_lib + SRCS ${XXHASH_INCLUDE_DIR} ${XXHASH_LIBRARIES} + DSTS ${dst_dir} ${dst_dir}/lib + DEPS xxhash +) + if(NOT PROTOBUF_FOUND) set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/protobuf") copy(protobuf_lib @@ -150,16 +157,16 @@ if (WITH_ANAKIN AND WITH_MKL) SRCS ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libinference_anakin_api* # compiled anakin api ${ANAKIN_INSTALL_DIR} # anakin release - DSTS ${dst_dir}/inference/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin) + DSTS ${FLUID_INSTALL_DIR}/third_party/install/anakin ${FLUID_INSTALL_DIR}/third_party/install/anakin) list(APPEND inference_deps anakin_inference_lib) endif() set(module "inference") copy(inference_lib DEPS ${inference_deps} SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.* - ${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci + ${src_dir}/${module}/api/paddle_inference_api.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/paddle_inference_pass.h - DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} + DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ) set(module "platform") @@ -185,20 +192,41 @@ copy(cmake_cache SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt DSTS ${FLUID_INSTALL_DIR}) -add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) +# This command generates a complete fluid library for both train and inference +add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep}) + +# Following commands generate a inference-only fluid library +# third_party, version.txt and CMakeCache.txt are the same position with ${FLUID_INSTALL_DIR} +copy(third_party DEPS fluid_lib_dist + SRCS ${FLUID_INSTALL_DIR}/third_party ${FLUID_INSTALL_DIR}/CMakeCache.txt + DSTS ${FLUID_INFERENCE_INSTALL_DIR} ${FLUID_INFERENCE_INSTALL_DIR} +) + +# only need libpaddle_fluid.so/a and paddle_inference_api.h for inference-only library +copy(inference_api_lib DEPS fluid_lib_dist + SRCS ${FLUID_INSTALL_DIR}/paddle/fluid/inference/libpaddle_fluid.* + ${FLUID_INSTALL_DIR}/paddle/fluid/inference/paddle_inference_api.h + DSTS ${FLUID_INFERENCE_INSTALL_DIR}/paddle/lib ${FLUID_INFERENCE_INSTALL_DIR}/paddle/include +) + +add_custom_target(inference_lib_dist DEPENDS third_party inference_api_lib) # paddle fluid version -execute_process( - COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1 - WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} - OUTPUT_VARIABLE PADDLE_GIT_COMMIT) -set(version_file ${FLUID_INSTALL_DIR}/version.txt) -file(WRITE ${version_file} - "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n" - "WITH_MKL: ${WITH_MKL}\n" - "WITH_GPU: ${WITH_GPU}\n") -if(WITH_GPU) - file(APPEND ${version_file} - "CUDA version: ${CUDA_VERSION}\n" - "CUDNN version: v${CUDNN_MAJOR_VERSION}\n") -endif() +function(version version_file) + execute_process( + COMMAND ${GIT_EXECUTABLE} log --pretty=format:%H -1 + WORKING_DIRECTORY ${PADDLE_SOURCE_DIR} + OUTPUT_VARIABLE PADDLE_GIT_COMMIT) + file(WRITE ${version_file} + "GIT COMMIT ID: ${PADDLE_GIT_COMMIT}\n" + "WITH_MKL: ${WITH_MKL}\n" + "WITH_MKLDNN: ${WITH_MKLDNN}\n" + "WITH_GPU: ${WITH_GPU}\n") + if(WITH_GPU) + file(APPEND ${version_file} + "CUDA version: ${CUDA_VERSION}\n" + "CUDNN version: v${CUDNN_MAJOR_VERSION}\n") + endif() +endfunction() +version(${FLUID_INSTALL_DIR}/version.txt) +version(${FLUID_INFERENCE_INSTALL_DIR}/version.txt) diff --git a/cmake/simd.cmake b/cmake/simd.cmake index 3eacf4d86aa0385eddb690d72e85e3384929bb99..566dc75fda019eb66759eb403f60e16f18cffef1 100644 --- a/cmake/simd.cmake +++ b/cmake/simd.cmake @@ -89,7 +89,9 @@ CHECK_CXX_SOURCE_RUNS(" #include int main() { - __m512i a = _mm512_undefined_epi32(); + __m512i a = _mm512_set_epi32 (-1, 2, -3, 4, -1, 2, -3, 4, + 13, -5, 6, -7, 9, 2, -6, 3); + __m512i result = _mm512_abs_epi32 (a); return 0; }" AVX512F_FOUND) diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index 6653244507742b33d9524a7a0e4a5b2b575d358a..6b665a9effba4bef083d007c0c74f2f4c79e647e 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -24,6 +24,7 @@ if(NOT WITH_FLUID_ONLY) endif() add_subdirectory(testing) +set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests CACHE INTERNAL "python tests directory") if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API) add_subdirectory(fluid) endif() diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index c6dd919a93d119723b389d3a695f0af82d711a06..3378d210cdf6a625e11b1dd5fe348aa04cdb9361 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -61,21 +61,22 @@ paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100)) paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None)) +paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None)) paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None)) paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None)) -paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'], varargs=None, keywords=None, defaults=None) -paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn'], varargs=None, keywords=None, defaults=(None, None, False)) -paddle.fluid.layers.softmax ArgSpec(args=['input', 'param_attr', 'bias_attr', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None)) -paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None)) -paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None)) +paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type', 'is_test'], varargs=None, keywords=None, defaults=(False,)) +paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None)) +paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None)) +paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)) +paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True)) paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False)) paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None)) paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None)) paddle.fluid.layers.sequence_expand_as ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,)) -paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen', 'name'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.sequence_unpad ArgSpec(args=['x', 'length', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None)) paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) @@ -84,7 +85,8 @@ paddle.fluid.layers.reduce_min ArgSpec(args=['input', 'dim', 'keep_dim', 'name'] paddle.fluid.layers.reduce_prod ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None)) paddle.fluid.layers.sequence_first_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.sequence_last_step ArgSpec(args=['input'], varargs=None, keywords=None, defaults=None) -paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name'], varargs=None, keywords=None, defaults=(False, None, None)) +paddle.fluid.layers.sequence_slice ArgSpec(args=['input', 'offset', 'length', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.dropout ArgSpec(args=['x', 'dropout_prob', 'is_test', 'seed', 'name', 'dropout_implementation'], varargs=None, keywords=None, defaults=(False, None, None, 'downgrade_in_infer')) paddle.fluid.layers.split ArgSpec(args=['input', 'num_or_sections', 'dim', 'name'], varargs=None, keywords=None, defaults=(-1, None)) paddle.fluid.layers.ctc_greedy_decoder ArgSpec(args=['input', 'blank', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.edit_distance ArgSpec(args=['input', 'label', 'normalized', 'ignored_tokens'], varargs=None, keywords=None, defaults=(True, None)) @@ -95,17 +97,17 @@ paddle.fluid.layers.warpctc ArgSpec(args=['input', 'label', 'blank', 'norm_by_ti paddle.fluid.layers.sequence_reshape ArgSpec(args=['input', 'new_dim'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.transpose ArgSpec(args=['x', 'perm', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.im2sequence ArgSpec(args=['input', 'filter_size', 'stride', 'padding', 'input_image_size', 'out_stride', 'name'], varargs=None, keywords=None, defaults=(1, 1, 0, None, 1, None)) -paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples'], varargs=None, keywords=None, defaults=(None, None, None, None)) -paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.layers.nce ArgSpec(args=['input', 'label', 'num_total_classes', 'sample_weight', 'param_attr', 'bias_attr', 'num_neg_samples', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None)) +paddle.fluid.layers.hsigmoid ArgSpec(args=['input', 'label', 'num_classes', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.beam_search ArgSpec(args=['pre_ids', 'pre_scores', 'ids', 'scores', 'beam_size', 'end_id', 'level', 'name'], varargs=None, keywords=None, defaults=(0, None)) paddle.fluid.layers.row_conv ArgSpec(args=['input', 'future_context_size', 'param_attr', 'act'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.layers.multiplex ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.layer_norm ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None)) -paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100)) +paddle.fluid.layers.softmax_with_cross_entropy ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode', 'return_softmax'], varargs=None, keywords=None, defaults=(False, -100, False, False)) paddle.fluid.layers.smooth_l1 ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.one_hot ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.autoincreased_step_counter ArgSpec(args=['counter_name', 'begin', 'step'], varargs=None, keywords=None, defaults=(None, 1, 1)) -paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, True, None)) +paddle.fluid.layers.reshape ArgSpec(args=['x', 'shape', 'actual_shape', 'act', 'inplace', 'name'], varargs=None, keywords=None, defaults=(None, None, False, None)) paddle.fluid.layers.squeeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.unsqueeze ArgSpec(args=['input', 'axes', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.lod_reset ArgSpec(args=['x', 'y', 'target_lod'], varargs=None, keywords=None, defaults=(None, None)) @@ -114,10 +116,12 @@ paddle.fluid.layers.pad ArgSpec(args=['x', 'paddings', 'pad_value', 'name'], var paddle.fluid.layers.pad_constant_like ArgSpec(args=['x', 'y', 'pad_value', 'name'], varargs=None, keywords=None, defaults=(0.0, None)) paddle.fluid.layers.label_smooth ArgSpec(args=['label', 'prior_dist', 'epsilon', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 0.1, 'float32', None)) paddle.fluid.layers.roi_pool ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1, 1, 1.0)) +paddle.fluid.layers.roi_align ArgSpec(args=['input', 'rois', 'pooled_height', 'pooled_width', 'spatial_scale', 'sampling_ratio', 'name'], varargs=None, keywords=None, defaults=(1, 1, 1.0, -1, None)) paddle.fluid.layers.dice_loss ArgSpec(args=['input', 'label', 'epsilon'], varargs=None, keywords=None, defaults=(1e-05,)) -paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR')) +paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'resample', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, 'BILINEAR', None)) paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)) -paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) +paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None)) +paddle.fluid.layers.resize_nearest ArgSpec(args=['input', 'out_shape', 'scale', 'name', 'actual_shape'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.sequence_scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)) @@ -127,6 +131,7 @@ paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None paddle.fluid.layers.log ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.margin_rank_loss ArgSpec(args=['label', 'left', 'right', 'margin', 'name'], varargs=None, keywords=None, defaults=(0.1, None)) paddle.fluid.layers.elu ArgSpec(args=['x', 'alpha', 'name'], varargs=None, keywords=None, defaults=(1.0, None)) paddle.fluid.layers.relu6 ArgSpec(args=['x', 'threshold', 'name'], varargs=None, keywords=None, defaults=(6.0, None)) paddle.fluid.layers.pow ArgSpec(args=['x', 'factor', 'name'], varargs=None, keywords=None, defaults=(1.0, None)) @@ -170,6 +175,16 @@ paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None)) paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None)) +paddle.fluid.layers.similarity_focus ArgSpec(args=['input', 'axis', 'indexes', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None)) +paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None)) +paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,)) +paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None)) paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)) paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None)) paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None) @@ -178,6 +193,7 @@ paddle.fluid.layers.batch ArgSpec(args=['reader', 'batch_size'], varargs=None, k paddle.fluid.layers.double_buffer ArgSpec(args=['reader', 'place', 'name'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.layers.random_data_generator ArgSpec(args=['low', 'high', 'shapes', 'lod_levels', 'for_parallel'], varargs=None, keywords=None, defaults=(True,)) paddle.fluid.layers.py_reader ArgSpec(args=['capacity', 'shapes', 'dtypes', 'lod_levels', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, None, True)) +paddle.fluid.layers.create_py_reader_by_data ArgSpec(args=['capacity', 'feed_list', 'name', 'use_double_buffer'], varargs=None, keywords=None, defaults=(None, True)) paddle.fluid.layers.Preprocessor.__init__ ArgSpec(args=['self', 'reader', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.Preprocessor.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.layers.Preprocessor.inputs ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None) @@ -187,6 +203,7 @@ paddle.fluid.layers.create_tensor ArgSpec(args=['dtype', 'name', 'persistable'], paddle.fluid.layers.create_parameter ArgSpec(args=['shape', 'dtype', 'name', 'attr', 'is_bias', 'default_initializer'], varargs=None, keywords=None, defaults=(None, None, False, None)) paddle.fluid.layers.create_global_var ArgSpec(args=['shape', 'value', 'dtype', 'persistable', 'force_cpu', 'name'], varargs=None, keywords=None, defaults=(False, False, None)) paddle.fluid.layers.cast ArgSpec(args=['x', 'dtype'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.tensor_array_to_tensor ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None)) paddle.fluid.layers.concat ArgSpec(args=['input', 'axis', 'name'], varargs=None, keywords=None, defaults=(0, None)) paddle.fluid.layers.sums ArgSpec(args=['input', 'out'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.assign ArgSpec(args=['input', 'output'], varargs=None, keywords=None, defaults=(None,)) @@ -257,6 +274,7 @@ paddle.fluid.layers.hard_shrink ArgSpec(args=['x', 'threshold'], varargs=None, k paddle.fluid.layers.cumsum ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.thresholded_relu ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.prior_box ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False)) +paddle.fluid.layers.density_prior_box ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, None)) paddle.fluid.layers.multi_box_head ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False)) paddle.fluid.layers.bipartite_match ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.target_assign ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) @@ -348,6 +366,8 @@ paddle.fluid.optimizer.ModelAverage.__init__ ArgSpec(args=['self', 'average_wind paddle.fluid.optimizer.ModelAverage.apply ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None) paddle.fluid.optimizer.ModelAverage.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.optimizer.ModelAverage.restore ArgSpec(args=['self', 'executor'], varargs=None, keywords=None, defaults=None) +paddle.fluid.optimizer.LarsMomentumOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'momentum', 'lars_coeff', 'lars_weight_decay', 'regularization', 'name'], varargs=None, keywords=None, defaults=(0.001, 0.0005, None, None)) +paddle.fluid.optimizer.LarsMomentumOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.backward.append_backward ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.regularizer.L1DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)) paddle.fluid.regularizer.L2DecayRegularizer.__init__ ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)) diff --git a/paddle/fluid/CMakeLists.txt b/paddle/fluid/CMakeLists.txt index 519a00fb073b08f6c88de8186de187476b548fd3..7d48f0057140cf021a21ea7e304b7e38cc8b9ec2 100644 --- a/paddle/fluid/CMakeLists.txt +++ b/paddle/fluid/CMakeLists.txt @@ -9,9 +9,6 @@ add_subdirectory(pybind) add_subdirectory(recordio) endif(NOT WIN32) -if(WITH_INFERENCE) - # NOTE: please add subdirectory inference at last. - add_subdirectory(inference) -endif() - +# NOTE: please add subdirectory inference at last. +add_subdirectory(inference) add_subdirectory(train) diff --git a/paddle/fluid/framework/attribute.cc b/paddle/fluid/framework/attribute.cc index 0dcecb62dba971b48c4f11c0ef47494be40eeea0..fabf2abfc803b8838edb48aa01ab8896799c97ac 100644 --- a/paddle/fluid/framework/attribute.cc +++ b/paddle/fluid/framework/attribute.cc @@ -64,6 +64,13 @@ Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) { case proto::AttrType::LONG: { return attr_desc.l(); } + case proto::AttrType::LONGS: { + std::vector val(attr_desc.longs_size()); + for (int i = 0; i < attr_desc.longs_size(); ++i) { + val[i] = attr_desc.longs(i); + } + return val; + } default: PADDLE_THROW("Unsupport attr type %d", attr_desc.type()); } diff --git a/paddle/fluid/framework/attribute.h b/paddle/fluid/framework/attribute.h index 14ca3e96209ed17f12e87fda8506806514698977..d9c76881b7e98d0b7cd29024b98c8f7720398c66 100644 --- a/paddle/fluid/framework/attribute.h +++ b/paddle/fluid/framework/attribute.h @@ -26,6 +26,113 @@ limitations under the License. */ namespace paddle { namespace framework { + +template +struct ExtractAttribute { + explicit ExtractAttribute(const std::string& attr_name) + : attr_name_(attr_name) {} + + T* operator()(Attribute& attr) const { + T* attr_value = nullptr; + try { + attr_value = &boost::get(attr); + } catch (boost::bad_get& bad_get) { + PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s", + attr_name_, paddle::platform::demangle(typeid(T).name()), + paddle::platform::demangle(attr.type().name())); + } + return attr_value; + } + + const std::string& attr_name_; +}; + +// special handle bool +// FIXME(yuyang18): Currently we cast bool into int in python binding. It is +// hard to change the logic there. In another way, we should correct handle +// if the user set `some_flag=1`. +// +// FIX ME anytime if there is a better solution. +template <> +struct ExtractAttribute { + explicit ExtractAttribute(const std::string& attr_name) + : attr_name_(attr_name) {} + + bool* operator()(Attribute& attr) const { + if (attr.type() == typeid(int)) { // NOLINT + int val = boost::get(attr); + attr = static_cast(val); + } else if (attr.type() == typeid(float)) { // NOLINT + float val = boost::get(attr); + attr = static_cast(val); + } + bool* attr_value = nullptr; + try { + attr_value = &boost::get(attr); + } catch (boost::bad_get& bad_get) { + PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s", + attr_name_, paddle::platform::demangle(attr.type().name())); + } + return attr_value; + } + + const std::string& attr_name_; +}; + +template <> +struct ExtractAttribute { + explicit ExtractAttribute(const std::string& attr_name) + : attr_name_(attr_name) {} + + int64_t* operator()(Attribute& attr) const { + if (attr.type() == typeid(int)) { // NOLINT + int val = boost::get(attr); + attr = static_cast(val); + } else if (attr.type() == typeid(float)) { // NOLINT + int val = boost::get(attr); + attr = static_cast(val); + } + int64_t* attr_value = nullptr; + try { + attr_value = &boost::get(attr); + } catch (boost::bad_get& bad_get) { + PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s", + attr_name_, paddle::platform::demangle(attr.type().name())); + } + return attr_value; + } + + const std::string& attr_name_; +}; + +template <> +struct ExtractAttribute> { + explicit ExtractAttribute(const std::string& attr_name) + : attr_name_(attr_name) {} + + std::vector* operator()(Attribute& attr) const { + if (attr.type() == typeid(std::vector)) { // NOLINT + std::vector val = boost::get>(attr); + std::vector vec(val.begin(), val.end()); + attr = vec; + } else if (attr.type() == typeid(std::vector)) { // NOLINT + std::vector val = boost::get>(attr); + std::vector vec(val.begin(), val.end()); + attr = vec; + } + std::vector* attr_value = nullptr; + try { + attr_value = &boost::get>(attr); + } catch (boost::bad_get& bad_get) { + PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s", + attr_name_, paddle::platform::demangle(attr.type().name())); + } + return attr_value; + } + + const std::string& attr_name_; +}; + template inline proto::AttrType AttrTypeID() { Attribute tmp = T(); @@ -42,7 +149,11 @@ class AttrReader { inline const T& Get(const std::string& name) const { PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap", name); - return boost::get(attrs_.at(name)); + + Attribute& attr = const_cast(attrs_.at(name)); + ExtractAttribute extract_attr(name); + T* attr_value = extract_attr(attr); + return *attr_value; } private: @@ -82,7 +193,7 @@ class DefaultValueSetter { public: explicit DefaultValueSetter(T default_value) : default_value_(default_value) {} - void operator()(T& value) const { value = default_value_; } + void operator()(T& value) const { value = default_value_; } // NOLINT private: T default_value_; @@ -117,84 +228,6 @@ class EnumInContainer { std::unordered_set container_; }; -template -struct ExtractAttribute { - explicit ExtractAttribute(const std::string& attr_name) - : attr_name_(attr_name) {} - - T* operator()(Attribute& attr) const { - T* attr_value = nullptr; - try { - attr_value = &boost::get(attr); - } catch (boost::bad_get& bad_get) { - PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s", - attr_name_, paddle::platform::demangle(typeid(T).name()), - paddle::platform::demangle(attr.type().name())); - } - return attr_value; - } - - const std::string& attr_name_; -}; - -// special handle bool -// FIXME(yuyang18): Currently we cast bool into int in python binding. It is -// hard to change the logic there. In another way, we should correct handle -// if the user set `some_flag=1`. -// -// FIX ME anytime if there is a better solution. -template <> -struct ExtractAttribute { - explicit ExtractAttribute(const std::string& attr_name) - : attr_name_(attr_name) {} - - bool* operator()(Attribute& attr) const { - if (attr.type() == typeid(int)) { // NOLINT - int val = boost::get(attr); - attr = static_cast(val); - } else if (attr.type() == typeid(float)) { // NOLINT - float val = boost::get(attr); - attr = static_cast(val); - } - bool* attr_value = nullptr; - try { - attr_value = &boost::get(attr); - } catch (boost::bad_get& bad_get) { - PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s", - attr_name_, paddle::platform::demangle(attr.type().name())); - } - return attr_value; - } - - const std::string& attr_name_; -}; - -template <> -struct ExtractAttribute { - explicit ExtractAttribute(const std::string& attr_name) - : attr_name_(attr_name) {} - - int64_t* operator()(Attribute& attr) const { - if (attr.type() == typeid(int)) { // NOLINT - int val = boost::get(attr); - attr = static_cast(val); - } else if (attr.type() == typeid(float)) { // NOLINT - int val = boost::get(attr); - attr = static_cast(val); - } - int64_t* attr_value = nullptr; - try { - attr_value = &boost::get(attr); - } catch (boost::bad_get& bad_get) { - PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s", - attr_name_, paddle::platform::demangle(attr.type().name())); - } - return attr_value; - } - - const std::string& attr_name_; -}; - // check whether a certain attribute fit its limits // an attribute can have more than one limits template @@ -235,7 +268,7 @@ class TypedAttrChecker { return *this; } - void operator()(AttributeMap& attr_map) const { + void operator()(AttributeMap& attr_map) const { // NOLINT if (!attr_map.count(attr_name_)) { // user do not set this attr PADDLE_ENFORCE(!default_value_setter_.empty(), @@ -271,7 +304,7 @@ class OpAttrChecker { return *(checker.target>()); } - void Check(AttributeMap& attr_map) const { + void Check(AttributeMap& attr_map) const { // NOLINT for (const auto& checker : attr_checkers_) { checker(attr_map); } diff --git a/paddle/fluid/framework/data_device_transform.cc b/paddle/fluid/framework/data_device_transform.cc index fee6ba40047053ed5662fe044eceb0c687bd4db9..57ff061fe5e612495add86df8f82fe7d9f9107dc 100644 --- a/paddle/fluid/framework/data_device_transform.cc +++ b/paddle/fluid/framework/data_device_transform.cc @@ -18,8 +18,8 @@ namespace framework { void TransDataDevice(const Tensor &in, const platform::Place &dst_place, Tensor *out) { - VLOG(3) << "DeviceTransform in, src_place " << in.place() - << " dst_place: " << dst_place; + VLOG(30) << "DeviceTransform in, src_place " << in.place() + << " dst_place: " << dst_place; PADDLE_ENFORCE_NE( in.place().which(), dst_place.which(), diff --git a/paddle/fluid/framework/data_device_transform_test.cu b/paddle/fluid/framework/data_device_transform_test.cu index f2c55e533a2747325b1b16fdada37945a8ed3c42..21e0cb3f91cc0ae05513c3bbd470650ca71194d7 100644 --- a/paddle/fluid/framework/data_device_transform_test.cu +++ b/paddle/fluid/framework/data_device_transform_test.cu @@ -49,10 +49,10 @@ class TestOpWithKernel : public OperatorWithKernel { OpKernelType GetExpectedKernelType( const ExecutionContext& ctx) const override { if (Attr("use_gpu")) { - VLOG(3) << "force use gpu kernel"; + VLOG(30) << "force use gpu kernel"; return OpKernelType(proto::VarType::FP32, platform::CUDAPlace(0)); } else { - VLOG(3) << "use default kernel"; + VLOG(30) << "use default kernel"; return OpKernelType(proto::VarType::FP32, ctx.Input("input")->place()); } @@ -148,7 +148,7 @@ TEST(Operator, CPUtoGPU) { // get output auto* output2 = scope.Var("OUT2"); gpu_op->Run(scope, cuda_place); - VLOG(3) << "after gpu_op run"; + VLOG(30) << "after gpu_op run"; // auto* output2_ptr = output2->Get().data(); paddle::platform::DeviceContextPool& pool = diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index e0a3ef5a9c6c53c42ebea1a41cac0d18a77781b2..d6b5ad4570c1d8402dedb8596cc75d9eae5a91c7 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -1,5 +1,6 @@ cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto node) cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context lod_tensor) +cc_library(op_graph_view SRCS op_graph_view.cc DEPS op_handle_base) cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry) @@ -16,32 +17,39 @@ if(WITH_GPU) dynload_cuda variable_visitor) nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda) nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda) + nv_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle) else() cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory variable_visitor) cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim) cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor) + cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle) endif() cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor) 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) -if(WITH_GPU) +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) + +if (WITH_GPU) cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass) endif() +cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass) + cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle - scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle) + scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle) -if(WITH_GPU) - cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto reference_count_pass) -else() - cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto) +set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass) +if (WITH_GPU) + list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass) endif() +cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS}) + cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context) @@ -54,8 +62,9 @@ cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_execu # device_context reduce_op_handle ) cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context) +cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fused_broadcast_op_handle) cc_library(build_strategy SRCS build_strategy.cc DEPS graph_viz_pass multi_devices_graph_pass multi_devices_graph_print_pass multi_devices_graph_check_pass - fuse_elewise_add_act_pass) + fuse_elewise_add_act_pass multi_batch_merge_pass) diff --git a/paddle/fluid/framework/details/all_reduce_op_handle.cc b/paddle/fluid/framework/details/all_reduce_op_handle.cc index 7c5f5bd80a937bf1a1c891155764833d7b21c5c2..b8690156763e4037811245b8016982710445e6a2 100644 --- a/paddle/fluid/framework/details/all_reduce_op_handle.cc +++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc @@ -34,7 +34,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, nccl_ctxs_(ctxs) { if (nccl_ctxs_) { for (auto &p : places_) { - this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p); + this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p)); } } } @@ -46,7 +46,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, #endif void AllReduceOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); if (NoDummyInputSize() == 1) { return; // No need to all reduce when GPU count = 1; @@ -127,7 +127,7 @@ void AllReduceOpHandle::RunImpl() { *local_scopes_[i]->FindVar(kLocalExecScopeName)->Get(); auto &p = places_[i]; auto *var = scope.FindVar(out_var_handles[i]->name_); - auto *dev_ctx = dev_ctxes_[p]; + auto *dev_ctx = dev_ctxes_.at(p); RunAndRecordEvent(p, [&trg, var, dev_ctx, p] { auto &tensor_gpu = *var->GetMutable(); diff --git a/paddle/fluid/framework/details/broadcast_op_handle.cc b/paddle/fluid/framework/details/broadcast_op_handle.cc index 4fdab5cd94358d08eac7f8b041bf16d09042f0bd..8e5e5427659387d63eac21a200c1a20da493e539 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle.cc @@ -48,16 +48,27 @@ void BroadcastOpHandle::RunImpl() { var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get()); } + BroadcastOneVar(*in_var_handle, out_var_handles, var_scopes); +} + +void BroadcastOpHandle::BroadcastOneVar( + const VarHandle &in_var_handle, + const std::vector &out_var_handles, + const std::vector &var_scopes) { auto *in_var = - var_scopes.at(in_var_handle->scope_idx_)->FindVar(in_var_handle->name_); + var_scopes.at(in_var_handle.scope_idx_)->FindVar(in_var_handle.name_); PADDLE_ENFORCE_NOT_NULL(in_var); Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var); + if (UNLIKELY(!in_tensor.IsInitialized())) { + VLOG(30) << "in var " << in_var_handle.name_ << "not inited, return!"; + return; + } - InitOutputValue(*in_var_handle, out_var_handles); + InitOutputValue(in_var_handle, out_var_handles); if (platform::is_cpu_place(in_tensor.place())) { for (auto *out_var_handle : out_var_handles) { - if (out_var_handle->IsTheSameVar(*in_var_handle)) { + if (out_var_handle->IsTheSameVar(in_var_handle)) { continue; } auto &out_p = out_var_handle->place_; @@ -114,12 +125,12 @@ void BroadcastOpHandle::RunImpl() { } } - if (!out_handle->IsTheSameVar(*in_var_handle)) { - auto out_var = var_scopes.at(in_var_handle->scope_idx_) + if (!out_handle->IsTheSameVar(in_var_handle)) { + auto out_var = var_scopes.at(in_var_handle.scope_idx_) ->FindVar(out_var_handles[0]->name_); paddle::framework::TensorCopy( - in_tensor, in_var_handle->place_, - *(dev_ctxes_.at(in_var_handle->place_)), + in_tensor, in_var_handle.place_, + *(dev_ctxes_.at(in_var_handle.place_)), &VariableVisitor::GetMutableTensor(out_var)); } }); diff --git a/paddle/fluid/framework/details/broadcast_op_handle.h b/paddle/fluid/framework/details/broadcast_op_handle.h index fe4e733e43417977df324fde808f52b228a27d19..72180fac864256ddda076c57e50ab1083c113d32 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle.h +++ b/paddle/fluid/framework/details/broadcast_op_handle.h @@ -44,7 +44,8 @@ struct BroadcastOpHandle : public OpHandleBase { nccl_ctxs_(nccl_ctxs) { if (nccl_ctxs_) { for (auto &p_ctx : nccl_ctxs_->contexts_) { - dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get(); + this->SetDeviceContext(platform::CUDAPlace(p_ctx.first), + p_ctx.second.ctx_.get()); } } } @@ -61,7 +62,10 @@ struct BroadcastOpHandle : public OpHandleBase { protected: void RunImpl() override; - private: + void BroadcastOneVar(const VarHandle &in_var_handle, + const std::vector &out_var_handles, + const std::vector &var_scopes); + std::vector local_scopes_; std::vector places_; #ifdef PADDLE_WITH_CUDA diff --git a/paddle/fluid/framework/details/broadcast_op_handle_test.cc b/paddle/fluid/framework/details/broadcast_op_handle_test.cc index ab7412a19fbd13fa39dbae9af528d158cc9ddbd0..650de5a48de6b1fdab120cdeda563a169fd1a1c1 100644 --- a/paddle/fluid/framework/details/broadcast_op_handle_test.cc +++ b/paddle/fluid/framework/details/broadcast_op_handle_test.cc @@ -12,232 +12,12 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/details/broadcast_op_handle.h" -#include "gtest/gtest.h" - -#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/framework/details/broadcast_op_handle_test.h" namespace paddle { namespace framework { namespace details { -namespace f = paddle::framework; -namespace p = paddle::platform; - -// test data amount -const f::DDim kDims = {20, 20}; - -struct TestBroadcastOpHandle { - std::vector> ctxs_; - std::vector local_scopes_; - std::vector param_scopes_; - Scope g_scope_; - std::unique_ptr op_handle_; - std::vector> vars_; - std::vector gpu_list_; - bool use_gpu_; -#ifdef PADDLE_WITH_CUDA - std::unique_ptr nccl_ctxs_; -#endif - - void WaitAll() { - for (size_t j = 0; j < ctxs_.size(); ++j) { - ctxs_[j]->Wait(); - } -#ifdef PADDLE_WITH_CUDA - if (nccl_ctxs_) { - nccl_ctxs_->WaitAll(); - } -#endif - } - - void InitCtxOnGpu(bool use_gpu) { - use_gpu_ = use_gpu; - if (use_gpu_) { -#ifdef PADDLE_WITH_CUDA - int count = p::GetCUDADeviceCount(); - if (count <= 1) { - LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA " - "device count is " - << count; - exit(0); - } - for (int i = 0; i < count; ++i) { - auto p = p::CUDAPlace(i); - gpu_list_.push_back(p); - ctxs_.emplace_back(new p::CUDADeviceContext(p)); - } - nccl_ctxs_.reset(new platform::NCCLContextMap(gpu_list_)); -#else - PADDLE_THROW("CUDA is not support."); -#endif - } else { - int count = 8; - for (int i = 0; i < count; ++i) { - auto p = p::CPUPlace(); - gpu_list_.push_back(p); - ctxs_.emplace_back(new p::CPUDeviceContext(p)); - } -#ifdef PADDLE_WITH_CUDA - nccl_ctxs_.reset(nullptr); -#endif - } - } - - void InitBroadcastOp(size_t input_scope_idx) { - for (size_t j = 0; j < gpu_list_.size(); ++j) { - local_scopes_.push_back(&(g_scope_.NewScope())); - Scope& local_scope = local_scopes_.back()->NewScope(); - *local_scopes_.back() - ->Var(details::kLocalExecScopeName) - ->GetMutable() = &local_scope; - local_scope.Var("out"); - param_scopes_.emplace_back(&local_scope); - } - param_scopes_[input_scope_idx]->Var("input"); - - std::unique_ptr n = - ir::CreateNodeForTest("node0", ir::Node::Type::kOperation); - if (use_gpu_) { -#ifdef PADDLE_WITH_CUDA - op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_, - nccl_ctxs_.get())); -#else - PADDLE_THROW("CUDA is not support."); -#endif - } else { -#ifdef PADDLE_WITH_CUDA - op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_, - nccl_ctxs_.get())); -#else - op_handle_.reset( - new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_)); -#endif - } - - std::unique_ptr v = - ir::CreateNodeForTest("node1", ir::Node::Type::kVariable); - auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input", - gpu_list_[input_scope_idx]); - vars_.emplace_back(in_var_handle); - op_handle_->AddInput(in_var_handle); - - // add dummy var - - std::unique_ptr v2 = - ir::CreateNodeForTest("node2", ir::Node::Type::kVariable); - vars_.emplace_back(new DummyVarHandle(v2.get())); - DummyVarHandle* dummy_var_handle = - static_cast(vars_.back().get()); - dummy_var_handle->ClearGeneratedOp(); - op_handle_->AddInput(dummy_var_handle); - - for (size_t j = 0; j < gpu_list_.size(); ++j) { - if (!use_gpu_) { - op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); - } - std::unique_ptr v3 = - ir::CreateNodeForTest("node3", ir::Node::Type::kVariable); - VarHandle* out_var_handle = - new VarHandle(v3.get(), 2, j, "out", gpu_list_[j]); - vars_.emplace_back(out_var_handle); - op_handle_->AddOutput(out_var_handle); - } - - // add dummy var - std::unique_ptr v4 = - ir::CreateNodeForTest("node4", ir::Node::Type::kVariable); - vars_.emplace_back(new DummyVarHandle(v4.get())); - DummyVarHandle* out_dummy_var_handle = - static_cast(vars_.back().get()); - out_dummy_var_handle->ClearGeneratedOp(); - op_handle_->AddOutput(out_dummy_var_handle); - } - - void TestBroadcastLodTensor(size_t input_scope_idx) { - auto in_var = param_scopes_[input_scope_idx]->FindVar("input"); - PADDLE_ENFORCE_NOT_NULL(in_var); - auto in_lod_tensor = in_var->GetMutable(); - in_lod_tensor->mutable_data(kDims, gpu_list_[input_scope_idx]); - - std::vector send_vector(static_cast(f::product(kDims))); - for (size_t k = 0; k < send_vector.size(); ++k) { - send_vector[k] = k; - } - f::LoD lod{{0, 10, 20}}; - paddle::framework::TensorFromVector( - send_vector, *(ctxs_[input_scope_idx]), in_lod_tensor); - in_lod_tensor->set_lod(lod); - in_lod_tensor->Resize(kDims); - - op_handle_->Run(false); - - WaitAll(); - - p::CPUPlace cpu_place; - for (size_t j = 0; j < gpu_list_.size(); ++j) { - auto out_var = param_scopes_[j]->FindVar("out"); - PADDLE_ENFORCE_NOT_NULL(out_var); - auto out_tensor = out_var->Get(); - PADDLE_ENFORCE_EQ(out_tensor.lod(), lod, "lod is not equal."); - - f::Tensor result_tensor; - f::TensorCopySync(out_tensor, cpu_place, &result_tensor); - float* ct = result_tensor.mutable_data(cpu_place); - - for (int64_t i = 0; i < f::product(kDims); ++i) { - ASSERT_NEAR(ct[i], send_vector[i], 1e-5); - } - } - } - - void TestBroadcastSelectedRows(size_t input_scope_idx) { - auto in_var = param_scopes_[input_scope_idx]->FindVar("input"); - PADDLE_ENFORCE_NOT_NULL(in_var); - auto in_selected_rows = in_var->GetMutable(); - auto value = in_selected_rows->mutable_value(); - value->mutable_data(kDims, gpu_list_[input_scope_idx]); - int height = static_cast(kDims[0]) * 2; - std::vector rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1, - 2, 4, 6, 3, 1, 1, 1, 1, 3, 7}; - in_selected_rows->set_height(height); - in_selected_rows->set_rows(rows); - - std::vector send_vector(static_cast(f::product(kDims))); - for (size_t k = 0; k < send_vector.size(); ++k) { - send_vector[k] = k; - } - paddle::framework::TensorFromVector( - send_vector, *(ctxs_[input_scope_idx]), value); - - op_handle_->Run(false); - - WaitAll(); - - p::CPUPlace cpu_place; - for (size_t j = 0; j < gpu_list_.size(); ++j) { - auto out_var = param_scopes_[j]->FindVar("out"); - PADDLE_ENFORCE_NOT_NULL(out_var); - auto& out_select_rows = out_var->Get(); - auto rt = out_select_rows.value(); - - PADDLE_ENFORCE_EQ(out_select_rows.height(), height, - "height is not equal."); - for (size_t k = 0; k < out_select_rows.rows().size(); ++k) { - PADDLE_ENFORCE_EQ(out_select_rows.rows()[k], rows[k]); - } - - f::Tensor result_tensor; - f::TensorCopySync(rt, cpu_place, &result_tensor); - float* ct = result_tensor.data(); - - for (int64_t i = 0; i < f::product(kDims); ++i) { - ASSERT_NEAR(ct[i], send_vector[i], 1e-5); - } - } - } -}; - TEST(BroadcastTester, TestCPUBroadcastTestLodTensor) { TestBroadcastOpHandle test_op; size_t input_scope_idx = 0; diff --git a/paddle/fluid/framework/details/broadcast_op_handle_test.h b/paddle/fluid/framework/details/broadcast_op_handle_test.h new file mode 100644 index 0000000000000000000000000000000000000000..4305eb65733a7c871450949ce2c48cab013bac81 --- /dev/null +++ b/paddle/fluid/framework/details/broadcast_op_handle_test.h @@ -0,0 +1,273 @@ +// 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 "gtest/gtest.h" +#include "paddle/fluid/framework/details/broadcast_op_handle.h" + +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace framework { +namespace details { + +namespace f = paddle::framework; +namespace p = paddle::platform; + +// test data amount +const f::DDim kDims = {20, 20}; + +struct TestBroadcastOpHandle { + std::vector> ctxs_; + std::vector local_scopes_; + std::vector param_scopes_; + Scope g_scope_; + OpHandleBase* op_handle_; + std::vector vars_; + std::vector> nodes_; + std::vector place_list_; + bool use_gpu_; +#ifdef PADDLE_WITH_CUDA + std::unique_ptr nccl_ctxs_; +#endif + + void WaitAll() { + for (size_t j = 0; j < ctxs_.size(); ++j) { + ctxs_[j]->Wait(); + } +#ifdef PADDLE_WITH_CUDA + if (nccl_ctxs_) { + nccl_ctxs_->WaitAll(); + } +#endif + } + + void InitCtxOnGpu(bool use_gpu) { + use_gpu_ = use_gpu; + if (use_gpu_) { +#ifdef PADDLE_WITH_CUDA + int count = p::GetCUDADeviceCount(); + if (count <= 1) { + LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA " + "device count is " + << count; + exit(0); + } + for (int i = 0; i < count; ++i) { + auto p = p::CUDAPlace(i); + place_list_.push_back(p); + ctxs_.emplace_back(new p::CUDADeviceContext(p)); + } + nccl_ctxs_.reset(new platform::NCCLContextMap(place_list_)); +#else + PADDLE_THROW("CUDA is not support."); +#endif + } else { + int count = 8; + for (int i = 0; i < count; ++i) { + auto p = p::CPUPlace(); + place_list_.push_back(p); + ctxs_.emplace_back(new p::CPUDeviceContext(p)); + } +#ifdef PADDLE_WITH_CUDA + nccl_ctxs_.reset(nullptr); +#endif + } + } + + void InitBroadcastOp(size_t input_scope_idx) { + nodes_.clear(); + for (size_t j = 0; j < place_list_.size(); ++j) { + local_scopes_.push_back(&(g_scope_.NewScope())); + Scope& local_scope = local_scopes_.back()->NewScope(); + *local_scopes_.back() + ->Var(details::kLocalExecScopeName) + ->GetMutable() = &local_scope; + local_scope.Var("out"); + param_scopes_.emplace_back(&local_scope); + } + param_scopes_[input_scope_idx]->Var("input"); + + nodes_.emplace_back( + ir::CreateNodeForTest("node0", ir::Node::Type::kOperation)); + if (use_gpu_) { +#ifdef PADDLE_WITH_CUDA + op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_, + place_list_, nccl_ctxs_.get()); +#else + PADDLE_THROW("CUDA is not support."); +#endif + } else { +#ifdef PADDLE_WITH_CUDA + op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_, + place_list_, nccl_ctxs_.get()); +#else + op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_, + place_list_); +#endif + } + + nodes_.emplace_back( + ir::CreateNodeForTest("node1", ir::Node::Type::kVariable)); + auto* in_var_handle = new VarHandle(nodes_.back().get(), 1, input_scope_idx, + "input", place_list_[input_scope_idx]); + vars_.emplace_back(in_var_handle); + op_handle_->AddInput(in_var_handle); + + // add dummy var + + nodes_.emplace_back( + ir::CreateNodeForTest("node2", ir::Node::Type::kVariable)); + vars_.emplace_back(new DummyVarHandle(nodes_.back().get())); + DummyVarHandle* dummy_var_handle = + static_cast(vars_.back()); + dummy_var_handle->ClearGeneratedOp(); + op_handle_->AddInput(dummy_var_handle); + + for (size_t j = 0; j < place_list_.size(); ++j) { + if (!use_gpu_) { + op_handle_->SetDeviceContext(place_list_[j], ctxs_[j].get()); + } + nodes_.emplace_back( + ir::CreateNodeForTest("node3", ir::Node::Type::kVariable)); + VarHandle* out_var_handle = + new VarHandle(nodes_.back().get(), 2, j, "out", place_list_[j]); + vars_.emplace_back(out_var_handle); + op_handle_->AddOutput(out_var_handle); + } + + // add dummy var + nodes_.emplace_back( + ir::CreateNodeForTest("node4", ir::Node::Type::kVariable)); + vars_.emplace_back(new DummyVarHandle(nodes_.back().get())); + DummyVarHandle* out_dummy_var_handle = + static_cast(vars_.back()); + out_dummy_var_handle->ClearGeneratedOp(); + op_handle_->AddOutput(out_dummy_var_handle); + } + + std::vector InitLoDTensor(const std::string& varname, + size_t input_scope_idx, const f::LoD& lod, + float val_scalar = 0.0) { + auto var = param_scopes_[input_scope_idx]->FindVar(varname); + + PADDLE_ENFORCE_NOT_NULL(var); + auto lod_tensor = var->GetMutable(); + std::vector send_vector(static_cast(f::product(kDims))); + for (size_t k = 0; k < send_vector.size(); ++k) { + send_vector[k] = k + val_scalar; + } + paddle::framework::TensorFromVector( + send_vector, *(ctxs_[input_scope_idx]), lod_tensor); + lod_tensor->set_lod(lod); + lod_tensor->Resize(kDims); + return send_vector; + } + + std::vector InitSelectedRows(const std::string& varname, + size_t input_scope_idx, + const std::vector& rows, + int height, float value_scalar = 0.0) { + std::vector send_vector(static_cast(f::product(kDims))); + for (size_t k = 0; k < send_vector.size(); ++k) { + send_vector[k] = k + value_scalar; + } + + auto var = param_scopes_[input_scope_idx]->FindVar(varname); + PADDLE_ENFORCE_NOT_NULL(var); + auto selected_rows = var->GetMutable(); + auto value = selected_rows->mutable_value(); + value->mutable_data(kDims, place_list_[input_scope_idx]); + selected_rows->set_height(height); + selected_rows->set_rows(rows); + + paddle::framework::TensorFromVector( + send_vector, *(ctxs_[input_scope_idx]), value); + + return send_vector; + } + + void SelectedRowsEqual(const std::string& varname, int input_scope_idx, + const std::vector& send_vector, + const std::vector& rows, int height) { + auto var = param_scopes_[input_scope_idx]->FindVar(varname); + PADDLE_ENFORCE_NOT_NULL(var); + auto& selected_rows = var->Get(); + auto rt = selected_rows.value(); + PADDLE_ENFORCE_EQ(selected_rows.height(), height, "height is not equal."); + + for (size_t k = 0; k < selected_rows.rows().size(); ++k) { + PADDLE_ENFORCE_EQ(selected_rows.rows()[k], rows[k]); + } + + p::CPUPlace cpu_place; + f::Tensor result_tensor; + f::TensorCopySync(rt, cpu_place, &result_tensor); + float* ct = result_tensor.data(); + + for (int64_t i = 0; i < f::product(kDims); ++i) { + ASSERT_NEAR(ct[i], send_vector[i], 1e-5); + } + } + + void LoDTensorEqual(const std::string& varname, + const std::vector& send_vec, const f::LoD& lod, + framework::Scope* scope) { + p::CPUPlace cpu_place; + auto var = scope->FindVar(varname); + PADDLE_ENFORCE_NOT_NULL(var); + auto tensor = var->Get(); + PADDLE_ENFORCE_EQ(tensor.lod(), lod, "lod is not equal."); + f::Tensor result_tensor; + f::TensorCopySync(tensor, cpu_place, &result_tensor); + float* ct = result_tensor.mutable_data(cpu_place); + for (int64_t k = 0; k < f::product(kDims); ++k) { + ASSERT_NEAR(ct[k], send_vec[k], 1e-5); + } + } + + void TestBroadcastLodTensor(size_t input_scope_idx) { + f::LoD lod{{0, 10, 20}}; + auto send_vector = InitLoDTensor("input", input_scope_idx, lod); + + op_handle_->Run(false); + + WaitAll(); + for (size_t j = 0; j < place_list_.size(); ++j) { + LoDTensorEqual("out", send_vector, lod, param_scopes_[j]); + } + } + + void TestBroadcastSelectedRows(size_t input_scope_idx) { + std::vector rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1, + 2, 4, 6, 3, 1, 1, 1, 1, 3, 7}; + int height = static_cast(kDims[0] * 2); + auto send_vector = InitSelectedRows("input", input_scope_idx, rows, height); + + op_handle_->Run(false); + + WaitAll(); + for (size_t j = 0; j < place_list_.size(); ++j) { + SelectedRowsEqual("out", input_scope_idx, send_vector, rows, height); + } + } +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/build_strategy.cc b/paddle/fluid/framework/details/build_strategy.cc index 6a6b497fa897e3882995688bf36704b1d77ea962..48f94a1f05614d4b797562ac67cdb9828fd0456e 100644 --- a/paddle/fluid/framework/details/build_strategy.cc +++ b/paddle/fluid/framework/details/build_strategy.cc @@ -16,6 +16,7 @@ limitations under the License. */ #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h" #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h" +#include "paddle/fluid/framework/details/sequential_execution_pass.h" #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/graph_viz_pass.h" @@ -27,6 +28,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { public: explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy) : ir::PassBuilder(), strategy_(strategy) { + if (strategy_.enable_sequential_execution_) { + AppendPass("sequential_execution_pass"); + } + // Add a graph viz pass to record a graph. if (!strategy_.debug_graphviz_path_.empty()) { auto viz_pass = AppendPass("graph_viz_pass"); @@ -64,6 +69,10 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder { // Verify that the graph is correct for multi-device executor. AppendPass("multi_devices_check_pass"); + + if (strategy_.remove_unnecessary_lock_) { + AppendPass("modify_op_lock_and_record_event_pass"); + } } private: @@ -110,6 +119,11 @@ std::unique_ptr BuildStrategy::Apply( pass->Erase("nccl_ctxs"); pass->SetNotOwned("nccl_ctxs", nctx); #endif + } else if (pass->Type() == "sequential_execution_pass") { + pass->Erase(kAllOpDescs); + pass->Set>( + kAllOpDescs, + new std::vector(main_program.Block(0).AllOps())); } graph = pass->Apply(std::move(graph)); } @@ -121,6 +135,9 @@ std::unique_ptr BuildStrategy::Apply( USE_PASS(fuse_elewise_add_act_pass); USE_PASS(graph_viz_pass); +USE_PASS(multi_batch_merge_pass); USE_PASS(multi_devices_pass); USE_PASS(multi_devices_check_pass); USE_PASS(multi_devices_print_pass); +USE_PASS(sequential_execution_pass); +USE_PASS(modify_op_lock_and_record_event_pass); diff --git a/paddle/fluid/framework/details/build_strategy.h b/paddle/fluid/framework/details/build_strategy.h index 02c4bea16916d58a6d0fce8918f8fceb9ff9356e..6c7b54db8f610aa34cd51dcbc13063290cae3ac0 100644 --- a/paddle/fluid/framework/details/build_strategy.h +++ b/paddle/fluid/framework/details/build_strategy.h @@ -69,6 +69,12 @@ struct BuildStrategy { bool enable_data_balance_{false}; + bool enable_sequential_execution_{false}; + + bool fuse_broadcast_op_{false}; + + bool remove_unnecessary_lock_{false}; + // User normally doesn't need to call this API. // The PassBuilder allows for more customized insert, remove of passes // from python side. diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc index b6282debdb4eb6b1f29c39e54ac4f3e2296838da..7ad1e40c600c6e70cea822fac777ff20163078e6 100644 --- a/paddle/fluid/framework/details/computation_op_handle.cc +++ b/paddle/fluid/framework/details/computation_op_handle.cc @@ -29,15 +29,21 @@ ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope, void ComputationOpHandle::RunImpl() { WaitInputVarGenerated(place_); - this->RunAndRecordEvent([this] { + auto run_func = [this]() { op_->Run(*scope_->FindVar(kLocalExecScopeName)->Get(), place_); - }); + }; + + if (is_lock_and_record_event_free_) { + run_func(); + } else { + this->RunAndRecordEvent(run_func); + } } bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) { bool need_wait = in_var && in_var->GeneratedOp() && - in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_[place_]; + in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_); return need_wait; } diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h index e98f1ab148db083ac63a1afd43e334fbfae62539..662a91d6b4dfcfed563fdf2e46c22f83f90b40af 100644 --- a/paddle/fluid/framework/details/computation_op_handle.h +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -36,6 +36,8 @@ struct ComputationOpHandle : public OpHandleBase { const platform::Place &GetPlace() const { return place_; } + void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; } + protected: void RunImpl() override; @@ -45,6 +47,7 @@ struct ComputationOpHandle : public OpHandleBase { std::unique_ptr op_; Scope *scope_; platform::Place place_; + bool is_lock_and_record_event_free_{false}; }; } // namespace details } // namespace framework diff --git a/paddle/fluid/framework/details/data_balance_op_handle.cc b/paddle/fluid/framework/details/data_balance_op_handle.cc index 525d24322442ef4dd6e8c24212af61c908959b87..0b772f9b63e2cfb78175f5e0d7011db8e6a5ec20 100644 --- a/paddle/fluid/framework/details/data_balance_op_handle.cc +++ b/paddle/fluid/framework/details/data_balance_op_handle.cc @@ -28,7 +28,7 @@ DataBalanceOpHandle::DataBalanceOpHandle( : OpHandleBase(node), local_scopes_(local_scopes), places_(places) { if (ctxs) { for (auto &p : places_) { - this->dev_ctxes_[p] = ctxs->DevCtx(p); + this->SetDeviceContext(p, ctxs->DevCtx(p)); } } } @@ -89,8 +89,8 @@ void DataBalanceOpHandle::RunImpl() { PADDLE_ENFORCE_GT(places_.size(), 1, "Data balance can only be enabled when the number of " "places to run larger than 1."); - auto in_var_handles = DynamicCast(inputs_); - auto out_var_handles = DynamicCast(outputs_); + auto in_var_handles = DynamicCast(this->Inputs()); + auto out_var_handles = DynamicCast(this->Outputs()); PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0); PADDLE_ENFORCE_EQ( in_var_handles.size(), out_var_handles.size(), diff --git a/paddle/fluid/framework/details/execution_strategy.h b/paddle/fluid/framework/details/execution_strategy.h index 5183be878eb49cccc68603c3fdd8023be5578036..15c496130c2b6c7643ff96661be09e5ac4870344 100644 --- a/paddle/fluid/framework/details/execution_strategy.h +++ b/paddle/fluid/framework/details/execution_strategy.h @@ -13,6 +13,7 @@ // limitations under the License. #pragma once +#include // for size_t namespace paddle { namespace framework { @@ -26,6 +27,7 @@ struct ExecutionStrategy { bool allow_op_delay_{false}; size_t num_iteration_per_drop_scope_{100}; ExecutorType type_{kDefault}; + bool dry_run_{false}; }; } // namespace details diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc index 4ec1accd2e6a9c5a1ca45bc7e66445837af2f46e..949510e03705a4a0900f1c7b8758a8f7308aa44b 100644 --- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc @@ -16,6 +16,7 @@ #include #include "paddle/fluid/framework/details/fetch_op_handle.h" #include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/ir/graph_helper.h" namespace paddle { namespace framework { @@ -32,13 +33,11 @@ FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor( pool_(strategy.num_threads_), prepare_pool_(1), // add one more thread for generate op_deps fetch_ctxs_(places) { - auto &ops = graph_->Get("ops"); - - for (auto &op : ops) { + for (auto &op : ir::FilterByNodeWrapper(*graph_)) { int dep = static_cast(op->NotReadyInputSize()); - op_deps_.emplace(op.get(), dep); + op_deps_.emplace(op, dep); if (dep == 0) { - bootstrap_ops_.emplace_back(op.get()); + bootstrap_ops_.emplace_back(op); } } @@ -54,13 +53,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( paddle::framework::FeedFetchList fetches; fetches.resize(fetch_tensors.size()); std::unordered_map> fetched_vars; - std::vector> fetch_ops; + std::vector fetch_ops; for (auto &fetch_var_name : fetch_tensors) { for (auto &var_map : graph_->Get("vars")) { auto it = var_map.find(fetch_var_name); if (it != var_map.end()) { - fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); + fetched_vars[fetch_var_name].push_back(*it->second.rbegin()); } } } @@ -92,13 +91,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( size_t num_complete = 0; remaining_ = 0; - BlockingQueue complete_q; + auto complete_q = std::make_shared>(); for (auto op : bootstrap_ops_) { - RunOpAsync(op_deps.get(), op, &complete_q); + RunOpAsync(op_deps.get(), op, complete_q); } while (num_complete != op_deps->size()) { - size_t num_comp = complete_q.Pop(); + size_t num_comp = complete_q->Pop(); if (num_comp == -1UL) { int remaining = 0; while (true) { @@ -107,10 +106,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( break; } for (int i = 0; i < remaining; ++i) { - complete_q.Pop(); + complete_q->Pop(); } } - exception_.ReThrow(); + if (exception_.IsCaught()) { + ClearFetchOp(graph_.get(), &fetch_ops); + exception_.ReThrow(); + } } num_complete += num_comp; } @@ -120,14 +122,17 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run( } void FastThreadedSSAGraphExecutor::RunOpAsync( std::unordered_map> *op_deps, - OpHandleBase *op, BlockingQueue *complete_q) { + OpHandleBase *op, + const std::shared_ptr> &complete_q) { ++remaining_; this->pool_.enqueue([=] { OpHandleBase *op_to_run = op; size_t complete = 0; while (op_to_run != nullptr) { try { - op_to_run->Run(strategy_.use_cuda_); + if (LIKELY(!strategy_.dry_run_)) { + op_to_run->Run(strategy_.use_cuda_); + } ++complete; } catch (...) { exception_.Catch(std::current_exception()); @@ -144,7 +149,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync( if (op_to_run == nullptr) { op_to_run = pending_op; } else { - this->RunOpAsync(op_deps, pending_op, complete_q); + RunOpAsync(op_deps, pending_op, complete_q); } } } diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h index 043f9d3fb7849591696c78708d3090be87920e96..949616f02d5168e6abab932d608e4b20ee64304a 100644 --- a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h @@ -51,7 +51,8 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor { std::atomic remaining_; void RunOpAsync(std::unordered_map> *op_deps, - OpHandleBase *op, BlockingQueue *complete_q); + OpHandleBase *op, + const std::shared_ptr> &complete_q); void PrepareAtomicOpDeps(); diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc index fe18b2060c5cd7e157374da53c5a985f70545ab7..648adae06facb504042d8286f6eab5d98e99c015 100644 --- a/paddle/fluid/framework/details/fetch_op_handle.cc +++ b/paddle/fluid/framework/details/fetch_op_handle.cc @@ -28,11 +28,7 @@ FetchOpHandle::FetchOpHandle(ir::Node *node, FeedFetchList *data, size_t offset, offset_(offset), local_scopes_(local_scopes) {} -FetchOpHandle::~FetchOpHandle() { - for (auto *input_var : inputs_) { - input_var->RemoveOutput(this, this->Node()); - } -} +FetchOpHandle::~FetchOpHandle() {} void FetchOpHandle::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) { PADDLE_THROW("Nobody should wait FetchOp. Unexpceted Error"); diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle.cc b/paddle/fluid/framework/details/fused_broadcast_op_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..51dfa2d0711f49aaefab0af3549283dbf77eee4a --- /dev/null +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle.cc @@ -0,0 +1,55 @@ +// 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/framework/details/fused_broadcast_op_handle.h" +#include "paddle/fluid/framework/details/container_cast.h" +#include "paddle/fluid/framework/details/variable_visitor.h" +#include "paddle/fluid/platform/profiler.h" + +namespace paddle { +namespace framework { +namespace details { + +void FusedBroadcastOpHandle::RunImpl() { + platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + + if (places_.size() == 1UL) return; + + auto in_var_handles = DynamicCast(inputs_); + auto out_var_handles = DynamicCast(outputs_); + + WaitInputVarGenerated(); + + std::vector var_scopes; + for (auto *s : local_scopes_) { + var_scopes.emplace_back(s->FindVar(kLocalExecScopeName)->Get()); + } + + size_t place_num = places_.size(); + PADDLE_ENFORCE_EQ(in_var_handles.size() * place_num, out_var_handles.size()); + + for (size_t i = 0; i < in_var_handles.size(); ++i) { + BroadcastOneVar( + *in_var_handles[i], + std::vector(out_var_handles.begin() + i * place_num, + out_var_handles.begin() + (i + 1) * place_num), + var_scopes); + } +} + +std::string FusedBroadcastOpHandle::Name() const { return "fused_broadcast"; } + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle.h b/paddle/fluid/framework/details/fused_broadcast_op_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..e37259526a5f6f57d51a0ca8bca96a18211a4790 --- /dev/null +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle.h @@ -0,0 +1,57 @@ +// 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 + +#include "paddle/fluid/framework/details/broadcast_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/platform/device_context.h" + +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/nccl_helper.h" +#endif + +namespace paddle { +namespace framework { +namespace details { + +struct FusedBroadcastOpHandle : public BroadcastOpHandle { + public: +#ifdef PADDLE_WITH_CUDA + FusedBroadcastOpHandle(ir::Node *node, + const std::vector local_scopes, + const std::vector &places, + const platform::NCCLContextMap *nccl_ctx) + : BroadcastOpHandle(node, local_scopes, places, nccl_ctx) {} +#else + FusedBroadcastOpHandle(ir::Node* node, const std::vector local_scopes, + const std::vector& places) + : BroadcastOpHandle(node, local_scopes, places) {} +#endif + std::string Name() const override; + + protected: + void RunImpl() override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc b/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..541993c74332cc483a8b854a6b8f227c7c9a19a9 --- /dev/null +++ b/paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc @@ -0,0 +1,167 @@ +// 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/framework/details/fused_broadcast_op_handle.h" +#include "gtest/gtest.h" +#include "paddle/fluid/framework/details/broadcast_op_handle_test.h" + +namespace paddle { +namespace framework { +namespace details { + +struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle { + std::vector out_varnames_; + std::vector> nodes_; + + void InitFusedBroadcastOp(std::vector input_scope_idxes) { + nodes_.clear(); + // initialize scope and var + for (size_t i = 0; i < place_list_.size(); ++i) { + local_scopes_.push_back(&(g_scope_.NewScope())); + Scope& local_scope = local_scopes_.back()->NewScope(); + *local_scopes_.back() + ->Var(details::kLocalExecScopeName) + ->GetMutable() = &local_scope; + for (size_t j = 0; j < input_scope_idxes.size(); ++j) { + local_scope.Var("out_var" + j); + if (i == j) local_scope.Var("in_var" + j); + } + param_scopes_.emplace_back(&local_scope); + } + + // create op handle node + nodes_.emplace_back( + ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation)); + if (use_gpu_) { +#ifdef PADDLE_WITH_CUDA + op_handle_ = new FusedBroadcastOpHandle( + nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get()); +#else + PADDLE_THROW("CUDA is not supported."); +#endif + } else { +#ifdef PADDLE_WITH_CUDA + op_handle_ = new FusedBroadcastOpHandle( + nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get()); +#else + op_handle_ = new FusedBroadcastOpHandle(nodes_.back().get(), + local_scopes_, place_list_); +#endif + } + + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + // add input var handle + nodes_.emplace_back( + ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable)); + VarHandle* in_var_handle = + new VarHandle(nodes_.back().get(), 1, input_scope_idxes[i], + "in_var" + i, place_list_[input_scope_idxes[i]]); + vars_.emplace_back(in_var_handle); + op_handle_->AddInput(in_var_handle); + + // add output var handle + for (size_t j = 0; j < place_list_.size(); ++j) { + nodes_.emplace_back( + ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable)); + VarHandle* out_var_handle = new VarHandle( + nodes_.back().get(), 2, j, "out_var" + i, place_list_[j]); + vars_.emplace_back(out_var_handle); + op_handle_->AddOutput(out_var_handle); + } + } + } + + void TestFusedBroadcastLoDTensor(std::vector input_scope_idxes) { + std::vector> send_vec; + f::LoD lod{{0, 10, 20}}; + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + const std::string varname("in_var" + i); + float val_scalar = static_cast(i); + send_vec.push_back( + InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar)); + } + + op_handle_->Run(false); + + WaitAll(); + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + const std::string& varname("out_var" + i); + for (size_t j = 0; j < place_list_.size(); ++j) { + LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]); + } + } + } + + void TestFusedBroadcastSelectedRows(std::vector input_scope_idxes) { + std::vector> send_vector; + std::vector rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1, + 2, 4, 6, 3, 1, 1, 1, 1, 3, 7}; + int height = static_cast(kDims[0] * 2); + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + const std::string varname("in_var" + i); + float val_scalar = static_cast(i); + send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i], + rows, height, val_scalar)); + } + + op_handle_->Run(false); + + WaitAll(); + for (size_t i = 0; i < input_scope_idxes.size(); ++i) { + const std::string& varname("out_var" + i); + for (size_t j = 0; j < place_list_.size(); ++j) { + SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows, + height); + } + } + } +}; + +TEST(FusedBroadcastTester, CPULodTensor) { + TestFusedBroadcastOpHandle test_op; + std::vector input_scope_idxes = {0, 1}; + test_op.InitCtxOnGpu(false); + test_op.InitFusedBroadcastOp(input_scope_idxes); + test_op.TestFusedBroadcastLoDTensor(input_scope_idxes); +} + +TEST(FusedBroadcastTester, CPUSelectedRows) { + TestFusedBroadcastOpHandle test_op; + std::vector input_scope_idxes = {0, 1}; + test_op.InitCtxOnGpu(false); + test_op.InitFusedBroadcastOp(input_scope_idxes); + test_op.TestFusedBroadcastSelectedRows(input_scope_idxes); +} + +#ifdef PADDLE_WITH_CUDA +TEST(FusedBroadcastTester, GPULodTensor) { + TestFusedBroadcastOpHandle test_op; + std::vector input_scope_idxes = {0, 1}; + test_op.InitCtxOnGpu(true); + test_op.InitFusedBroadcastOp(input_scope_idxes); + test_op.TestFusedBroadcastLoDTensor(input_scope_idxes); +} + +TEST(FusedBroadcastTester, GPUSelectedRows) { + TestFusedBroadcastOpHandle test_op; + std::vector input_scope_idxes = {0, 1}; + test_op.InitCtxOnGpu(true); + test_op.InitFusedBroadcastOp(input_scope_idxes); + test_op.TestFusedBroadcastSelectedRows(input_scope_idxes); +} +#endif + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/gather_op_handle.cc b/paddle/fluid/framework/details/gather_op_handle.cc index 9aae19fc73de4387186da47c55710c94d53f1b88..ca4633c5a8f22fc9f7319b06aa766f9fe37dc68c 100644 --- a/paddle/fluid/framework/details/gather_op_handle.cc +++ b/paddle/fluid/framework/details/gather_op_handle.cc @@ -36,7 +36,7 @@ void GatherOpHandle::RunImpl() { VarHandle *out_var_handle; { - auto out_var_handles = DynamicCast(outputs_); + auto out_var_handles = DynamicCast(this->Outputs()); PADDLE_ENFORCE_EQ(out_var_handles.size(), 1, "The number of output should be one."); out_var_handle = out_var_handles.front(); @@ -99,7 +99,7 @@ void GatherOpHandle::RunImpl() { Tensor *out_tensor = out_value->mutable_value(); // copy - auto dev_ctx = dev_ctxes_[out_var_handle->place_]; + auto dev_ctx = dev_ctxes_.at(out_var_handle->place_); RunAndRecordEvent(out_var_handle->place_, [in_tensors, out_tensor, &dev_ctx, t_out_p] { int s = 0, e = 0; diff --git a/paddle/fluid/framework/details/gather_op_handle_test.cc b/paddle/fluid/framework/details/gather_op_handle_test.cc index ed67e88ff6a7fe9efd93e5dfd4d7bdf4c43aac2e..e8cb7feb8bea92a7486b8a9d84ba4b9e2b93dbfb 100644 --- a/paddle/fluid/framework/details/gather_op_handle_test.cc +++ b/paddle/fluid/framework/details/gather_op_handle_test.cc @@ -31,9 +31,10 @@ struct TestGatherOpHandle { std::vector local_scopes_; std::vector param_scopes_; Scope g_scope_; - std::unique_ptr op_handle_; - std::vector> vars_; + OpHandleBase* op_handle_; + std::vector vars_; std::vector gpu_list_; + std::vector> nodes_; void WaitAll() { for (size_t j = 0; j < ctxs_.size(); ++j) { @@ -70,7 +71,7 @@ struct TestGatherOpHandle { } void InitGatherOp(size_t input_scope_idx) { - std::vector> nodes; + nodes_.clear(); for (size_t j = 0; j < gpu_list_.size(); ++j) { local_scopes_.push_back(&(g_scope_.NewScope())); Scope& local_scope = local_scopes_.back()->NewScope(); @@ -82,44 +83,45 @@ struct TestGatherOpHandle { } param_scopes_[input_scope_idx]->Var("out"); - nodes.emplace_back( + nodes_.emplace_back( ir::CreateNodeForTest("node", ir::Node::Type::kOperation).release()); - op_handle_.reset( - new GatherOpHandle(nodes.back().get(), local_scopes_, gpu_list_)); + op_handle_ = + new GatherOpHandle(nodes_.back().get(), local_scopes_, gpu_list_); // add input for (size_t j = 0; j < gpu_list_.size(); ++j) { op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get()); - nodes.emplace_back( + nodes_.emplace_back( ir::CreateNodeForTest("node1", ir::Node::Type::kVariable).release()); auto* in_var_handle = - new VarHandle(nodes.back().get(), 1, j, "input", gpu_list_[j]); + new VarHandle(nodes_.back().get(), 1, j, "input", gpu_list_[j]); vars_.emplace_back(in_var_handle); op_handle_->AddInput(in_var_handle); } // add dummy var - nodes.emplace_back( + nodes_.emplace_back( ir::CreateNodeForTest("node2", ir::Node::Type::kVariable).release()); - vars_.emplace_back(new DummyVarHandle(nodes.back().get())); + vars_.emplace_back(new DummyVarHandle(nodes_.back().get())); DummyVarHandle* in_dummy_var_handle = - static_cast(vars_.back().get()); + static_cast(vars_.back()); in_dummy_var_handle->ClearGeneratedOp(); op_handle_->AddInput(in_dummy_var_handle); // add output - nodes.emplace_back( + nodes_.emplace_back( ir::CreateNodeForTest("node3", ir::Node::Type::kVariable).release()); - auto* out_var_handle = new VarHandle(nodes.back().get(), 2, input_scope_idx, - "out", gpu_list_[input_scope_idx]); + auto* out_var_handle = + new VarHandle(nodes_.back().get(), 2, input_scope_idx, "out", + gpu_list_[input_scope_idx]); vars_.emplace_back(out_var_handle); op_handle_->AddOutput(out_var_handle); // add dummy var - nodes.emplace_back( + nodes_.emplace_back( ir::CreateNodeForTest("node4", ir::Node::Type::kVariable).release()); - vars_.emplace_back(new DummyVarHandle(nodes.back().get())); + vars_.emplace_back(new DummyVarHandle(nodes_.back().get())); DummyVarHandle* dummy_var_handle = - static_cast(vars_.back().get()); + static_cast(vars_.back()); op_handle_->AddOutput(dummy_var_handle); } diff --git a/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..bf3f3637b551a8a8084e6e4f1ca6a94b65361f17 --- /dev/null +++ b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.cc @@ -0,0 +1,60 @@ +// 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/framework/details/modify_op_lock_and_record_event_pass.h" +#include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/details/op_graph_view.h" +#include "paddle/fluid/framework/ir/graph_helper.h" + +namespace paddle { +namespace framework { +namespace details { + +static bool IsLockAndRecordEventFreeComputationOpHandle( + ComputationOpHandle *op, const OpGraphView &graph_view) { + if (!platform::is_gpu_place(op->GetPlace())) return false; + for (auto &pending_op : graph_view.PendingOps(op)) { + auto *tmp = dynamic_cast(pending_op); + if (tmp == nullptr || !(tmp->GetPlace() == op->GetPlace())) { + return false; + } + } + return true; +} + +std::unique_ptr ModifyOpLockAndRecordEventPass::ApplyImpl( + std::unique_ptr ir_graph) const { + auto all_ops = ir::FilterByNodeWrapper(*ir_graph); + OpGraphView graph_view(all_ops); + for (auto &op : all_ops) { + auto *compute_op = dynamic_cast(op); + if (compute_op == nullptr) continue; + bool is_lock_and_record_event_free = + IsLockAndRecordEventFreeComputationOpHandle(compute_op, graph_view); + compute_op->SetLockAndRecordEventFree(is_lock_and_record_event_free); + if (is_lock_and_record_event_free) { + VLOG(100) << "Set is_lock_and_record_event_free be true in op " + << compute_op->DebugString(); + } + } + return ir_graph; +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(modify_op_lock_and_record_event_pass, + paddle::framework::details::ModifyOpLockAndRecordEventPass); diff --git a/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..b54e1b318be95e1e0abf6830f8c918895df02718 --- /dev/null +++ b/paddle/fluid/framework/details/modify_op_lock_and_record_event_pass.h @@ -0,0 +1,32 @@ +// 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 "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +class ModifyOpLockAndRecordEventPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc index c9c255864a2477ed29873f8521acce37fa928c06..c8ea18804630fea4ada98062256730dbf4c24860 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_check_pass.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h" #include #include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" namespace paddle { namespace framework { @@ -36,20 +37,20 @@ bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const { for (auto &var_map : graph->Get(kGraphVars)) { for (auto &name_pair : var_map) { for (auto &version_pair : name_pair.second) { - insert_pending_var(version_pair.get()); + insert_pending_var(version_pair); } } } for (auto &var : graph->Get(kGraphDepVars)) { - insert_pending_var(var.get()); + insert_pending_var(var); } - for (auto &op : graph->Get(kGraphOps)) { + for (OpHandleBase *op : ir::FilterByNodeWrapper(*graph)) { if (op->Inputs().empty()) { - ready_ops.insert(op.get()); + ready_ops.insert(op); } else { - pending_ops.insert({op.get(), op.get()->NoDupInputSize()}); + pending_ops.insert({op, op->NoDupInputSize()}); } } @@ -89,6 +90,4 @@ bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const { REGISTER_PASS(multi_devices_check_pass, paddle::framework::details::SSAGraghBuilderWithChecker) .RequireGraphAttr(paddle::framework::details::kGraphVars) - .RequireGraphAttr(paddle::framework::details::kGraphDepVars) - .RequireGraphAttr(paddle::framework::details::kGraphOps) - .RequireGraphAttr(paddle::framework::details::kShardedVarDevice); + .RequireGraphAttr(paddle::framework::details::kGraphDepVars); diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc index 134fcee826715672a6e021e9bf694bb771ebb830..8c98b781301e884d5d5c7d141f3d901d74d51285 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc @@ -21,6 +21,7 @@ #include "paddle/fluid/framework/details/broadcast_op_handle.h" #include "paddle/fluid/framework/details/computation_op_handle.h" #include "paddle/fluid/framework/details/data_balance_op_handle.h" +#include "paddle/fluid/framework/details/fused_broadcast_op_handle.h" #include "paddle/fluid/framework/details/multi_devices_graph_pass.h" #include "paddle/fluid/framework/details/reduce_op_handle.h" #include "paddle/fluid/framework/details/rpc_op_handle.h" @@ -33,7 +34,14 @@ namespace paddle { namespace framework { namespace details { + namespace { +// TODO(panyx0718): Clean this up as well. +// all operators. NOTE that even we use a vector here, the operators is +// unordered. +typedef std::vector GraphOps; +const char kGraphOps[] = "ops"; + void PolishGraphToSupportDataHazards(ir::Graph *graph) { for (auto &var_map : graph->Get(kGraphVars)) { for (auto &name_pair : var_map) { @@ -91,7 +99,7 @@ VarHandle *CreateOrGetLatestVarHandle(ir::Graph *graph, ir::Node *node, } var_holder.emplace_back(var); } else { - var = var_holder.rbegin()->get(); + var = *var_holder.rbegin(); } return var; } @@ -153,7 +161,7 @@ void MultiDevSSAGraphBuilder::CreateOpHandleIOs(ir::Graph *result, ir::Node *node, size_t place_id) const { auto p = places_[place_id]; - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); op_handle->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p)); @@ -252,9 +260,9 @@ std::vector SortOpsAndDelayOptimizeOp(const ir::Graph &graph) { std::vector sorted_ret; for (size_t i = 0; i < ret.size(); ++i) { if (i < last_backward) { - if (boost::get(ret[i]->Op()->GetAttr( - OpProtoAndCheckerMaker::OpRoleAttrName())) == - static_cast(OpRole::kOptimize)) { + if (static_cast(boost::get(ret[i]->Op()->GetAttr( + OpProtoAndCheckerMaker::OpRoleAttrName())) & + static_cast(OpRole::kOptimize))) { optimize_ops.push_back(ret[i]); } else { sorted_ret.push_back(ret[i]); @@ -302,7 +310,6 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( result.Set(kGraphVars, new GraphVars(places_.size())); result.Set(kGraphDepVars, new GraphDepVars); result.Set(kGraphOps, new GraphOps); - result.Set(kShardedVarDevice, new ShardedVarDevice); // find send/recv vars so that we can place the distributed training // related op in the place 0 @@ -316,11 +323,13 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( bool is_forwarding = true; bool is_dist_train = false; + std::unordered_map sharded_var_device; + for (ir::Node *node : sorted_ops) { if (boost::get( node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) == static_cast(OpRole::kRPC)) { - int op_dev_id = CreateRPCOp(&result, node); + int op_dev_id = CreateRPCOp(&result, node, &sharded_var_device); PADDLE_ENFORCE(op_dev_id != -1, "Can not schedule the RPC operator to the right place."); if (node->Op()->Type() == "recv") { @@ -336,7 +345,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( } else if (boost::get(node->Op()->GetAttr( OpProtoAndCheckerMaker::OpRoleAttrName())) == static_cast(OpRole::kDist)) { - int op_dev_id = CreateDistTrainOp(&result, node); + int op_dev_id = CreateDistTrainOp(&result, node, &sharded_var_device); if (node->Op()->Type() == "concat") { auto origin_param_name = node->Op()->OutputArgumentNames()[0]; bcast_var_name_set[op_dev_id].emplace(origin_param_name); @@ -347,7 +356,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( BuildStrategy::GradientScaleStrategy::kCustomized) { // TODO(paddle-dev): Why is there no input for this op_handle? auto loss_grad_name = node->Op()->OutputArgumentNames()[0]; - CreateScaleLossGradOp(&result, loss_grad_name); + CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0]); } // This assumes the backward generating code will ensure IsScaleLossOp // is true only for the op that scale the final scalar loss. @@ -355,12 +364,11 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( // the block. is_forwarding = false; } else { - int op_dev_id = GetOpDeviceID(result, node); + int op_dev_id = GetOpDeviceID(result, node, sharded_var_device); if (op_dev_id != -1) { // This op only runs on one specific device. CreateComputationalOp(&result, node, op_dev_id); for (ir::Node *n : node->outputs) { - graph->Get(kShardedVarDevice) - .emplace(n->Name(), op_dev_id); + sharded_var_device.emplace(n->Name(), op_dev_id); } } else { // This op runs on all devices, and its output may have parameter's @@ -391,14 +399,13 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( for (size_t i = 0; i < backward_vars.size(); i += 2) { auto &p_name = backward_vars[i]; auto &g_name = backward_vars[i + 1]; - VLOG(10) << "Bcast " << g_name << " for parameter " << p_name; + VLOG(100) << "Bcast " << g_name << " for parameter " << p_name; switch (strategy_.reduce_) { case BuildStrategy::ReduceStrategy::kReduce: cur_device_id = GetAppropriateDeviceID({g_name}); CreateReduceOp(&result, g_name, cur_device_id); - graph->Get(kShardedVarDevice) - .emplace(g_name, cur_device_id); + sharded_var_device.emplace(g_name, cur_device_id); if (!is_dist_train) { bcast_var_name_set[cur_device_id].emplace(p_name); } @@ -436,10 +443,14 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( if ((use_gpu && strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) || is_dist_train) { - for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) { - auto &to_bcast_set = bcast_var_name_set[dev_id]; - for (auto &bcast_name : to_bcast_set) { - CreateBroadcastOp(&result, bcast_name, dev_id); + if (strategy_.fuse_broadcast_op_) { + CreateFusedBroadcastOp(&result, bcast_var_name_set); + } else { + for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) { + auto &to_bcast_set = bcast_var_name_set[dev_id]; + for (auto &bcast_name : to_bcast_set) { + CreateBroadcastOp(&result, bcast_name, dev_id); + } } } } @@ -453,7 +464,7 @@ std::unique_ptr MultiDevSSAGraphBuilder::ApplyImpl( * Only variables should be the leaves of graph. */ AddOutputToLeafOps(&result); - PADDLE_ENFORCE(!ir::HasCircle(result)); + result.Erase(kGraphOps); return graph; } @@ -493,7 +504,7 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result, result->Get(kGraphOps).emplace_back(op_handle); auto *in = - result->Get(kGraphVars).at(src_dev_id).at(p_name).back().get(); + result->Get(kGraphVars).at(src_dev_id).at(p_name).back(); op_handle->AddInput(in); for (size_t i = 0; i < places_.size(); ++i) { @@ -508,6 +519,44 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result, } } +void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp( + ir::Graph *result, + const std::vector> &bcast_varnames) const { +#ifdef PADDLE_WITH_CUDA + auto *op_handle = new FusedBroadcastOpHandle( + result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation), + local_scopes_, places_, nccl_ctxs_); +#else + auto *op_handle = new FusedBroadcastOpHandle( + result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation), + local_scopes_, places_); +#endif + result->Get(kGraphOps).emplace_back(op_handle); + + for (size_t i = 0; i < places_.size(); ++i) { + auto &p = places_[i]; + SetCommunicationContext(op_handle, p); + } + + for (size_t dev_id = 0; dev_id < bcast_varnames.size(); ++dev_id) { + for (auto &p_name : bcast_varnames[dev_id]) { + auto *in = + result->Get(kGraphVars).at(dev_id).at(p_name).back(); + op_handle->AddInput(in); + for (size_t out_dev_id = 0; out_dev_id < places_.size(); ++out_dev_id) { + auto &p = places_[out_dev_id]; + auto &vars = + result->Get(kGraphVars).at(out_dev_id).at(p_name); + auto *out_var = new VarHandle( + result->CreateEmptyNode(p_name, ir::Node::Type::kVariable), + vars.size(), out_dev_id, p_name, p); + vars.emplace_back(out_var); + op_handle->AddOutput(out_var); + } + } + } +} + void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result, ir::Node *node, int dev_id) const { @@ -528,7 +577,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result, result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation), local_scopes_, places_)); #endif - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); for (size_t i = 0; i < places_.size(); ++i) { auto &p = places_[i]; @@ -536,7 +585,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result, auto &vars = result->Get(kGraphVars)[i][og]; PADDLE_ENFORCE(!vars.empty()); auto &prev_grad = vars.back(); - op_handle->AddInput(prev_grad.get()); + op_handle->AddInput(prev_grad); auto var = new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable), @@ -557,14 +606,14 @@ void MultiDevSSAGraphBuilder::InsertDataBalanceOp( result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation), local_scopes_, places_)); #endif - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); for (size_t i = 0; i < places_.size(); ++i) { auto &p = places_[i]; SetCommunicationContext(op_handle, p); for (const std::string &d_name : datas) { auto &vars = result->Get(kGraphVars)[i][d_name]; PADDLE_ENFORCE(!vars.empty()); - op_handle->AddInput(vars.back().get()); + op_handle->AddInput(vars.back()); auto var = new VarHandle( result->CreateEmptyNode(d_name, ir::Node::Type::kVariable), vars.size(), i, d_name, p); @@ -574,8 +623,9 @@ void MultiDevSSAGraphBuilder::InsertDataBalanceOp( } } -int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph, - ir::Node *node) const { +int MultiDevSSAGraphBuilder::GetOpDeviceID( + const ir::Graph &graph, ir::Node *node, + const std::unordered_map &sharded_var_device) const { if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) { return -1; } @@ -588,21 +638,28 @@ int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph, node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName())); PADDLE_ENFORCE_EQ(param_grad.size(), 2U); - int dev_id = GetVarDeviceID(graph, param_grad[1]); + int dev_id = GetVarDeviceID(graph, param_grad[1], sharded_var_device); PADDLE_ENFORCE_NE(dev_id, -1, "dev_id should not be -1.[%s, %s, %s]", node->Op()->Type(), param_grad[0], param_grad[1]); return dev_id; } -int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph, - const std::string &varname) const { - auto &sharded_var_device = graph.Get(kShardedVarDevice); +int MultiDevSSAGraphBuilder::GetVarDeviceID( + const ir::Graph &graph, const std::string &varname, + const std::unordered_map &sharded_var_device) const { auto got = sharded_var_device.find(varname); + if (got == sharded_var_device.end()) { + auto pos = varname.find(framework::kNewGradSuffix); + if (pos != std::string::npos) { + got = sharded_var_device.find(varname.substr(0, pos)); + } + } return got == sharded_var_device.end() ? -1 : got->second; } void MultiDevSSAGraphBuilder::CreateScaleLossGradOp( - ir::Graph *result, const std::string &loss_grad_name) const { + ir::Graph *result, const std::string &loss_grad_name, + ir::Node *out_var_node) const { for (size_t i = 0; i < places_.size(); ++i) { // Insert ScaleCost OpHandle auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]); @@ -617,10 +674,8 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp( // loss->pending_ops_.emplace_back(op_handle); // op_handle->inputs_.emplace_back(loss); - CreateOpOutput( - result, op_handle, - result->CreateEmptyNode(loss_grad_name, ir::Node::Type::kVariable), - places_[i], i); + CreateOpOutput(result, op_handle, + result->CreateVarNode(out_var_node->Var()), places_[i], i); } } @@ -648,7 +703,7 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result, result->CreateEmptyNode("reduce", ir::Node::Type::kOperation), local_scopes_, places_)); #endif - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); for (size_t i = 0; i < places_.size(); ++i) { auto &p = places_[i]; @@ -656,7 +711,7 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result, auto &vars = result->Get(kGraphVars)[i][og]; PADDLE_ENFORCE(!vars.empty()); auto &prev_grad = vars.back(); - op_handle->AddInput(prev_grad.get()); + op_handle->AddInput(prev_grad); } auto &vars = result->Get(kGraphVars)[dst_dev_id][og]; auto var = @@ -667,8 +722,9 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result, return var; } -int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result, - ir::Node *node) const { +int MultiDevSSAGraphBuilder::CreateDistTrainOp( + ir::Graph *result, ir::Node *node, + std::unordered_map *sharded_var_device) const { int op_dev_id = -1; std::vector input_var_names; std::vector output_var_names; @@ -680,25 +736,25 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result, } if (node->Op()->Type() == "split_byref" || - node->Op()->Type() == "split_selected_rows") { + node->Op()->Type() == "split_selected_rows" || + node->Op()->Type() == "split_ids") { // TODO(paddle-dev): getting the first var is not safe. - op_dev_id = GetVarDeviceID(*result, input_var_names[0]); + op_dev_id = + GetVarDeviceID(*result, input_var_names[0], *sharded_var_device); if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) { op_dev_id = GetAppropriateDeviceID(input_var_names); for (auto &varname : input_var_names) { - result->Get(kShardedVarDevice) - .emplace(varname, op_dev_id); + sharded_var_device->emplace(varname, op_dev_id); } } for (auto &varname : output_var_names) { - result->Get(kShardedVarDevice) - .emplace(varname, op_dev_id); + sharded_var_device->emplace(varname, op_dev_id); } } else if (node->Op()->Type() == "concat") { - op_dev_id = GetVarDeviceID(*result, input_var_names[0]); + op_dev_id = + GetVarDeviceID(*result, input_var_names[0], *sharded_var_device); for (auto &varname : output_var_names) { - result->Get(kShardedVarDevice) - .emplace(varname, op_dev_id); + sharded_var_device->emplace(varname, op_dev_id); } } else { LOG(ERROR) << "got unexpected dist op: " << node->Op()->Type(); @@ -716,14 +772,14 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result, } void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) { - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); for (ir::Node *input : node->inputs) { VarHandle *var = nullptr; for (int place_offset = 0; place_offset < num_places; ++place_offset) { auto &var_holders = result->Get(kGraphVars)[place_offset]; auto &var_holder = var_holders[input->Name()]; if (!var_holder.empty()) { - var = var_holder.rbegin()->get(); + var = *var_holder.rbegin(); op_handle->AddInput(var); } } @@ -731,12 +787,14 @@ void SetOpInputsAllPlaces(ir::Graph *result, ir::Node *node, int num_places) { } // Create RPC related op handles that connects its in ops and out ops. -int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, - ir::Node *node) const { +int MultiDevSSAGraphBuilder::CreateRPCOp( + ir::Graph *result, ir::Node *node, + std::unordered_map *sharded_var_device) const { int op_dev_id = -1; if (node->Op()->Type() == "send") { // TODO(paddle-dev): getting the first var is not safe. - op_dev_id = GetVarDeviceID(*result, node->inputs[0]->Name()); + op_dev_id = + GetVarDeviceID(*result, node->inputs[0]->Name(), *sharded_var_device); PADDLE_ENFORCE(!ir::IsControlDepVar(*node->inputs[0]), "This hack no longer holds, please fix."); // the variable name which contains .block means it was splited by @@ -751,14 +809,12 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName())); PADDLE_ENFORCE_EQ(send_param_grad.size(), 2U); op_dev_id = GetAppropriateDeviceID({send_param_grad[1]}); - VLOG(10) << "send grad " << input_var_names[0] << " origin " - << send_param_grad[1] << " place: " << op_dev_id; + VLOG(100) << "send grad " << input_var_names[0] << " origin " + << send_param_grad[1] << " place: " << op_dev_id; for (auto &varname : input_var_names) { - result->Get(kShardedVarDevice) - .emplace(varname, op_dev_id); + sharded_var_device->emplace(varname, op_dev_id); } - result->Get(kShardedVarDevice) - .emplace(send_param_grad[1], op_dev_id); + sharded_var_device->emplace(send_param_grad[1], op_dev_id); } } else if (node->Op()->Type() == "recv") { std::vector output_var_names; @@ -768,16 +824,16 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, auto recv_param_grad = boost::get>( node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName())); if (recv_param_grad.size() == 2U) { - op_dev_id = GetVarDeviceID(*result, recv_param_grad[1]); - VLOG(10) << "recv param " << recv_param_grad[0] - << " get grad place: " << recv_param_grad[1] - << " place: " << op_dev_id; + op_dev_id = + GetVarDeviceID(*result, recv_param_grad[1], *sharded_var_device); + VLOG(100) << "recv param " << recv_param_grad[0] + << " get grad place: " << recv_param_grad[1] + << " place: " << op_dev_id; } else { op_dev_id = GetAppropriateDeviceID(output_var_names); } for (auto &varname : output_var_names) { - result->Get(kShardedVarDevice) - .emplace(varname, op_dev_id); + sharded_var_device->emplace(varname, op_dev_id); } } else { // send_barrier, fetch_barrier will run on place 0; @@ -796,7 +852,7 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, // send_barrier, recv, fetch_barrier's inputs are deps var, get them from // all places auto p = places_[op_dev_id]; - auto *op_handle = result->Get(kGraphOps).back().get(); + auto *op_handle = result->Get(kGraphOps).back(); op_handle->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p)); @@ -804,7 +860,8 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result, for (ir::Node *output : node->outputs) { int outvar_dev_id = op_dev_id; if (node->Op()->Type() == "fetch_barrier") { - outvar_dev_id = GetVarDeviceID(*result, output->Name()); + outvar_dev_id = + GetVarDeviceID(*result, output->Name(), *sharded_var_device); PADDLE_ENFORCE_NE(outvar_dev_id, -1); } p = places_[outvar_dev_id]; diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.h b/paddle/fluid/framework/details/multi_devices_graph_pass.h index cdf9f13cde608b546d17a1e53e0f6acea9e12566..f3ec2d29415240b7012f458070223469d0947166 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_pass.h +++ b/paddle/fluid/framework/details/multi_devices_graph_pass.h @@ -44,12 +44,18 @@ class MultiDevSSAGraphBuilder : public ir::Pass { mutable platform::NCCLContextMap *nccl_ctxs_; #endif - int GetVarDeviceID(const ir::Graph &graph, const std::string &varname) const; + int GetVarDeviceID( + const ir::Graph &graph, const std::string &varname, + const std::unordered_map &sharded_var_device) const; bool IsScaleLossOp(ir::Node *node) const; - int CreateRPCOp(ir::Graph *result, ir::Node *node) const; - int CreateDistTrainOp(ir::Graph *result, ir::Node *node) const; + int CreateRPCOp( + ir::Graph *result, ir::Node *node, + std::unordered_map *sharded_var_device) const; + int CreateDistTrainOp( + ir::Graph *result, ir::Node *node, + std::unordered_map *sharded_var_device) const; std::vector FindDistTrainSendVars( const std::vector &nodes) const; @@ -61,14 +67,17 @@ class MultiDevSSAGraphBuilder : public ir::Pass { size_t num_places) const; void CreateScaleLossGradOp(ir::Graph *result, - const std::string &loss_grad_name) const; + const std::string &loss_grad_name, + ir::Node *out_var_node) const; VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og, int dst_dev_id) const; void CreateComputationalOp(ir::Graph *result, ir::Node *node, int dev_id) const; - int GetOpDeviceID(const ir::Graph &graph, ir::Node *node) const; + int GetOpDeviceID( + const ir::Graph &graph, ir::Node *node, + const std::unordered_map &sharded_var_device) const; void InsertAllReduceOp(ir::Graph *result, const std::string &og) const; @@ -78,6 +87,10 @@ class MultiDevSSAGraphBuilder : public ir::Pass { void CreateBroadcastOp(ir::Graph *result, const std::string &p_name, size_t src_dev_id) const; + void CreateFusedBroadcastOp( + ir::Graph *result, + const std::vector> &bcast_varnames) const; + bool IsSparseGradient(const std::string &og) const; size_t GetAppropriateDeviceID( diff --git a/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc index 361c91dc78c08a2cbf84ee88211d389c1e2312e5..8f92f0948d7d397ab0f20c01eae9e313f739adec 100644 --- a/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc +++ b/paddle/fluid/framework/details/multi_devices_graph_print_pass.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h" #include #include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_helper.h" namespace paddle { namespace framework { @@ -62,7 +63,7 @@ void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph, }); size_t op_id = 0; - for (auto &op : graph.Get(kGraphOps)) { + for (auto &op : ir::FilterByNodeWrapper(graph)) { std::string op_name = "op_" + std::to_string(op_id++); sout << op_name << " [label=\"" << op->Name() << "\", shape=rect]" << std::endl; diff --git a/paddle/fluid/framework/details/multi_devices_helper.h b/paddle/fluid/framework/details/multi_devices_helper.h index 175c5a9950be69d7bf6ae9e386af762007a18a51..1a2b75fbc0c28984ce5cf00e0a2ce0f804349bb1 100644 --- a/paddle/fluid/framework/details/multi_devices_helper.h +++ b/paddle/fluid/framework/details/multi_devices_helper.h @@ -35,23 +35,14 @@ namespace details { // The outside vector is the device vector. Each element of this vector is a // 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< - std::unordered_map>>> +// `std::vector` is the version of varaibles. +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"; - -// all operators. NOTE that even we use a vector here, the operators is -// unordered. -typedef std::vector> GraphOps; -const char kGraphOps[] = "ops"; - -typedef std::unordered_map ShardedVarDevice; -const char kShardedVarDevice[] = "sharded_var_device"; } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/op_graph_view.cc b/paddle/fluid/framework/details/op_graph_view.cc new file mode 100644 index 0000000000000000000000000000000000000000..4838c4198ff35ba3fb562f3a7c0563ee60179e3b --- /dev/null +++ b/paddle/fluid/framework/details/op_graph_view.cc @@ -0,0 +1,66 @@ +// 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/framework/details/op_graph_view.h" +#include +#include + +namespace paddle { +namespace framework { +namespace details { + +OpGraphView::OpGraphView(const std::vector &ops) { Build(ops); } + +void OpGraphView::Build(const std::vector &ops) { + for (auto &op : ops) { + preceding_ops_[op]; + pending_ops_[op]; + for (auto &var : op->Outputs()) { + for (auto &pending_op : var->PendingOps()) { + preceding_ops_[pending_op].insert(op); + pending_ops_[op].insert(pending_op); + } + } + } + PADDLE_ENFORCE( + preceding_ops_.size() == ops.size() && pending_ops_.size() == ops.size(), + "There are duplicate ops in graph."); +} + +std::unordered_set OpGraphView::AllOps() const { + std::unordered_set ret; + for (auto &pair : preceding_ops_) { + ret.insert(pair.first); + } + return ret; +} + +bool OpGraphView::HasOp(OpHandleBase *op) const { + return preceding_ops_.count(op) != 0; +} + +void OpGraphView::EnforceHasOp(OpHandleBase *op) const { + PADDLE_ENFORCE(HasOp(op), "Cannot find op %s in OpGraphView", + op == nullptr ? "nullptr" : op->DebugString()); +} + +const std::unordered_set &OpGraphView::PendingOps( + OpHandleBase *op) const { + EnforceHasOp(op); + return pending_ops_.at(op); +} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/op_graph_view.h b/paddle/fluid/framework/details/op_graph_view.h new file mode 100644 index 0000000000000000000000000000000000000000..afb3e8e59461eeba10d7027fc70b89cc170c1805 --- /dev/null +++ b/paddle/fluid/framework/details/op_graph_view.h @@ -0,0 +1,49 @@ +// 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 +#include +#include "paddle/fluid/framework/details/op_handle_base.h" + +namespace paddle { +namespace framework { +namespace details { + +class OpGraphView { + public: + explicit OpGraphView(const std::vector &ops); + + std::unordered_set AllOps() const; + + const std::unordered_set &PendingOps(OpHandleBase *op) const; + + bool HasOp(OpHandleBase *op) const; + + private: + void Build(const std::vector &ops); + void EnforceHasOp(OpHandleBase *op) const; + + std::unordered_map> + preceding_ops_; + std::unordered_map> + pending_ops_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc index 3812f0abf1b7069525c4420054c61c01c908acfe..4822627ac3b65972f41d9a23d9fe3dba3de3f97d 100644 --- a/paddle/fluid/framework/details/op_handle_base.cc +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -103,7 +103,7 @@ void OpHandleBase::WaitInputVarGenerated() { void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) { for (auto *in : inputs_) { if (NeedWait(in)) { - in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[place]); + in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(place)); } } } diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index 9fbefabc841e3f6940860f60d959fee97495e4c9..ba12ca3c61c05b3e856fffa8353d4ec5bf79bc39 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -31,7 +31,10 @@ constexpr char kLocalExecScopeName[] = "@LCOAL_SCOPE@"; // It's responsible for populating necessary fields of ir::Node. class OpHandleBase { public: - explicit OpHandleBase(ir::Node *node) : node_(node) {} + // Owned by `node`. No need to be deleted explicitly. + explicit OpHandleBase(ir::Node *node) : node_(node) { + node_->WrappedBy(this); + } virtual ~OpHandleBase(); @@ -64,7 +67,8 @@ class OpHandleBase { virtual bool IsMultiDeviceTransfer() { return false; } const platform::DeviceContext *DeviceContext(platform::Place place) { - return dev_ctxes_[place]; + auto it = dev_ctxes_.find(place); + return it != dev_ctxes_.end() ? it->second : nullptr; } void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) { diff --git a/paddle/fluid/framework/details/reduce_op_handle.cc b/paddle/fluid/framework/details/reduce_op_handle.cc index 7fc06f234d42a992328c0b6164f17945d8075c28..4503123eac810917cabcf1e62cff98552ed2f742 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.cc +++ b/paddle/fluid/framework/details/reduce_op_handle.cc @@ -27,7 +27,7 @@ namespace framework { namespace details { void ReduceOpHandle::RunImpl() { - platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second); + platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second); if (places_.size() == 1) return; // the input and output may have dummy var. diff --git a/paddle/fluid/framework/details/reduce_op_handle.h b/paddle/fluid/framework/details/reduce_op_handle.h index a6289b055f97b7b0e57928358d84117b33cf2df8..999828ae457ba43541da06088ce7c25331fd05ec 100644 --- a/paddle/fluid/framework/details/reduce_op_handle.h +++ b/paddle/fluid/framework/details/reduce_op_handle.h @@ -46,7 +46,8 @@ struct ReduceOpHandle : public OpHandleBase { nccl_ctxs_(nccl_ctxs) { if (nccl_ctxs_) { for (auto &p_ctx : nccl_ctxs_->contexts_) { - dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get(); + this->SetDeviceContext(platform::CUDAPlace(p_ctx.first), + p_ctx.second.ctx_.get()); } } } diff --git a/paddle/fluid/framework/details/reduce_op_handle_test.cc b/paddle/fluid/framework/details/reduce_op_handle_test.cc index 3a9a58412391b188c5e804b41fa47b3607a36bd1..72299c0bfa916d3b92e1c5020ddd69dadad3701d 100644 --- a/paddle/fluid/framework/details/reduce_op_handle_test.cc +++ b/paddle/fluid/framework/details/reduce_op_handle_test.cc @@ -30,8 +30,8 @@ struct TestReduceOpHandle { Scope g_scope_; std::vector local_scopes_; std::vector param_scopes_; - std::unique_ptr op_handle_; - std::vector> vars_; + OpHandleBase *op_handle_; + std::vector vars_; std::vector gpu_list_; std::vector> ctxs_; diff --git a/paddle/fluid/framework/details/reference_count_op_handle.h b/paddle/fluid/framework/details/reference_count_op_handle.h index fc479a4c4a1e7d5c824d3c202e0cccf743dd52c9..cc4ccfbdfc720284e683a8f3f59a4aa57a3a9eb1 100644 --- a/paddle/fluid/framework/details/reference_count_op_handle.h +++ b/paddle/fluid/framework/details/reference_count_op_handle.h @@ -51,7 +51,7 @@ class ReferenceCountOpHandle : public OpHandleBase { dev_ctx_ = static_cast( platform::DeviceContextPool::Instance().Get(place)); if (IsStreamGarabageCollector()) { - PADDLE_ENFORCE(cudaSetDevice(place.device)); + platform::SetDeviceId(place.device); PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming)); } @@ -61,7 +61,7 @@ class ReferenceCountOpHandle : public OpHandleBase { ~ReferenceCountOpHandle() { if (IsStreamGarabageCollector()) { auto gpu_place = boost::get(dev_ctx_->GetPlace()); - PADDLE_ENFORCE(cudaSetDevice(gpu_place.device)); + platform::SetDeviceId(gpu_place.device); PADDLE_ENFORCE(cudaEventDestroy(event_)); } } diff --git a/paddle/fluid/framework/details/reference_count_pass.cc b/paddle/fluid/framework/details/reference_count_pass.cc index 2d1f688d64ece3322e253b0c070264b9eb73d678..28443cc886e4c3f5db707d6d8fe9971618d8c2f7 100644 --- a/paddle/fluid/framework/details/reference_count_pass.cc +++ b/paddle/fluid/framework/details/reference_count_pass.cc @@ -19,6 +19,7 @@ #include "paddle/fluid/framework/details/computation_op_handle.h" #include "paddle/fluid/framework/details/multi_devices_helper.h" #include "paddle/fluid/framework/details/reference_count_pass.h" +#include "paddle/fluid/framework/ir/graph_helper.h" namespace paddle { namespace framework { @@ -43,6 +44,23 @@ static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) { return nullptr; } +static void AddDependencyBetween(OpHandleBase *in, OpHandleBase *out, + ir::Graph *graph) { + auto it = std::find_if( + in->Outputs().begin(), in->Outputs().end(), [](VarHandleBase *var) { + return dynamic_cast(var) != nullptr; + }); + + if (it != in->Outputs().end()) { + out->AddInput(*it); + } else { + auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); + graph->Get(kGraphDepVars).emplace(dep_var); + in->AddOutput(dep_var); + out->AddInput(dep_var); + } +} + std::unique_ptr ReferenceCountPass::ApplyImpl( std::unique_ptr graph) const { auto &ref_cnts = Get(kGlobalReferenceCount); @@ -54,14 +72,13 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( // Step 2: Find all variables in non-computation ops which refers to variables // in computation ops std::unordered_set names; - std::unordered_map> + std::unordered_map compute_ref_cnt_map; auto get_ref_cnts_from_compute_op = [&]( - const std::unique_ptr &op, - const std::vector &vars) { + OpHandleBase *op, const std::vector &vars) { std::vector var_names_in_op; - auto *compute_op = dynamic_cast(op.get()); + auto *compute_op = dynamic_cast(op); if (compute_op == nullptr || !platform::is_gpu_place(compute_op->GetPlace())) return var_names_in_op; @@ -104,9 +121,8 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( }; auto update_ref_cnts_from_non_compute_op = [&]( - const std::unique_ptr &op, - const std::vector &vars) { - if (dynamic_cast(op.get()) != nullptr) return; + OpHandleBase *op, const std::vector &vars) { + if (dynamic_cast(op) != nullptr) return; for (VarHandleBase *var_handle_base : vars) { auto *var_handle = dynamic_cast(var_handle_base); if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue; @@ -124,8 +140,8 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( if (next_compute_op != nullptr) { if (compute_ref_cnt_map.count(next_compute_op)) { compute_ref_cnt_map[next_compute_op]->AddVar(var_name); - VLOG(5) << "Add reference count of " << var_name << " to Operator " - << next_compute_op->Name(); + VLOG(50) << "Add reference count of " << var_name << " to Operator " + << next_compute_op->Name(); } else { // Create new reference_count_op_handle ir::Node *ref_cnt_node = graph->CreateEmptyNode( @@ -133,40 +149,30 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( auto *ref_cnt_handle = new ReferenceCountOpHandle( ref_cnt_node, next_compute_op->GetScope(), place, {var_name}, gcs[place.device].get(), cur_ref_cnts[place.device].get()); - if (next_compute_op->Outputs().empty()) { - auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); - next_compute_op->AddOutput(dep_var); - graph->Get(kGraphDepVars).emplace(dep_var); - } - ref_cnt_handle->AddInput(next_compute_op->Outputs().front()); - compute_ref_cnt_map[next_compute_op].reset(ref_cnt_handle); + AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get()); + compute_ref_cnt_map[next_compute_op] = ref_cnt_handle; } } } } }; - auto &all_ops = graph->Get(kGraphOps); + auto all_ops = ir::FilterByNodeWrapper(*graph); for (auto &op : all_ops) { auto in_var_names = get_ref_cnts_from_compute_op(op, op->Inputs()); auto out_var_names = get_ref_cnts_from_compute_op(op, op->Outputs()); if (in_var_names.empty() && out_var_names.empty()) continue; in_var_names.insert(in_var_names.end(), out_var_names.begin(), out_var_names.end()); - auto *compute_op = dynamic_cast(op.get()); + auto *compute_op = dynamic_cast(op); auto place = boost::get(compute_op->GetPlace()); ir::Node *ref_cnt_node = graph->CreateEmptyNode("reference_count", ir::Node::Type::kOperation); auto *ref_cnt_handle = new ReferenceCountOpHandle( ref_cnt_node, compute_op->GetScope(), place, in_var_names, gcs[place.device].get(), cur_ref_cnts[place.device].get()); - if (compute_op->Outputs().empty()) { - auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar()); - compute_op->AddOutput(dep_var); - graph->Get(kGraphDepVars).emplace(dep_var); - } - ref_cnt_handle->AddInput(compute_op->Outputs().front()); - compute_ref_cnt_map[compute_op].reset(ref_cnt_handle); + AddDependencyBetween(compute_op, ref_cnt_handle, graph.get()); + compute_ref_cnt_map[compute_op] = ref_cnt_handle; } for (auto &op : all_ops) { @@ -174,11 +180,11 @@ std::unique_ptr ReferenceCountPass::ApplyImpl( update_ref_cnts_from_non_compute_op(op, op->Outputs()); } - std::vector> new_all_ops; + std::vector new_all_ops; new_all_ops.reserve(compute_ref_cnt_map.size() + all_ops.size()); for (auto &op : all_ops) { new_all_ops.emplace_back(std::move(op)); - auto it = compute_ref_cnt_map.find(new_all_ops.back().get()); + auto it = compute_ref_cnt_map.find(new_all_ops.back()); if (it != compute_ref_cnt_map.end()) { // Add LeafNode to ReferenceCountOpHandle auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar()); diff --git a/paddle/fluid/framework/details/rpc_op_handle.cc b/paddle/fluid/framework/details/rpc_op_handle.cc index f44b374edb29228dff5a8bf003d945291f166d49..dfa6c1ade1a024bb9087144d0e96fa5b0417f06a 100644 --- a/paddle/fluid/framework/details/rpc_op_handle.cc +++ b/paddle/fluid/framework/details/rpc_op_handle.cc @@ -29,22 +29,19 @@ RPCOpHandle::RPCOpHandle(ir::Node *node, const framework::OpDesc &op_desc, place_(place) {} void RPCOpHandle::RunImpl() { - // TODO(wuyi): need further analysis whether wait VarDummyHandle. - // Wait input done for (auto *in : inputs_) { auto &p = static_cast(in)->place_; - // FIXME(Yancey1989): need a better solution instead of use DebugString() - if (ir::IsControlDepVar(*in->Node())) { // HACK + if (ir::IsControlDepVar(*in->Node())) { continue; } if (in->GeneratedOp()) { - in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[p]); + in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(p)); } } - auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get(); - // FIXME(wuyi): can not use RunAndRecordEvent here, for it will cause dead - // lock. - op_->Run(*tmp_scope, place_); + this->RunAndRecordEvent([this] { + op_->Run(*local_scope_->FindVar(kLocalExecScopeName)->Get(), + place_); + }); } std::string RPCOpHandle::Name() const { return name_; } diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc index ba243979b34aa1f683de707525403becaf0a1c00..6ab6cb2332b0af3fa16b986f115513ee098fae4f 100644 --- a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc @@ -27,7 +27,7 @@ ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev, coeff_(static_cast(1.0 / num_dev)), scope_(scope), place_(place) { - dev_ctxes_[place_] = dev_ctx; + this->SetDeviceContext(place_, dev_ctx); } ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {} @@ -46,12 +46,12 @@ void ScaleLossGradOpHandle::RunImpl() { } else { #ifdef PADDLE_WITH_CUDA this->RunAndRecordEvent([&] { - auto stream = - static_cast(this->dev_ctxes_[place_]) - ->stream(); + auto stream = static_cast( + this->dev_ctxes_.at(place_)) + ->stream(); memory::Copy(boost::get(place_), tmp, platform::CPUPlace(), &coeff_, sizeof(float), stream); - VLOG(10) << place_ << "RUN Scale loss grad op"; + VLOG(100) << place_ << "RUN Scale loss grad op"; }); #endif } diff --git a/paddle/fluid/framework/details/sequential_execution_pass.cc b/paddle/fluid/framework/details/sequential_execution_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..f78a47bb78e6f1d81db6abed11a7762f21dd2226 --- /dev/null +++ b/paddle/fluid/framework/details/sequential_execution_pass.cc @@ -0,0 +1,109 @@ +// 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/framework/details/sequential_execution_pass.h" +#include +#include +#include +#include +#include "paddle/fluid/framework/op_proto_maker.h" + +namespace paddle { +namespace framework { +namespace details { + +static bool IsSameOpDesc(OpDesc *op1, OpDesc *op2) { + return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() && + op1->Outputs() == op2->Outputs(); +} + +std::unique_ptr SequentialExecutionPass::ApplyImpl( + std::unique_ptr graph) const { + // FIXME(zjl): Insert dependencies between some distributed ops may cause + // the multi_devices_graph_pass fails. So we skip these ops here. + // Indeed, maybe we should not insert dependencies between these ops + // casually, which may cause deadlock easily. + // We should add more skipped distributed ops when found errors in + // multi_devices_graph_pass + static std::unordered_set skip_dist_ops{ + "send", "recv", "send_barrier", "fetch_barrier"}; + + auto &ops = Get>(kAllOpDescs); + std::vector op_node_list; + op_node_list.reserve(ops.size()); + + std::unordered_map op_deps; + std::unordered_map> pending_ops; + std::unordered_set ready_ops; + + for (ir::Node *node : graph->Nodes()) { + if (!node->IsOp()) continue; + std::unordered_set preceding_ops; + for (auto *in : node->inputs) { + PADDLE_ENFORCE(in->IsVar(), + "Preceding Node of Op Nodes must be Var Node"); + if (in->inputs.empty()) continue; + PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp(), + "Preceding Op Node of Var Node must be unique"); + preceding_ops.insert(in->inputs[0]); + pending_ops[in->inputs[0]].insert(node); + } + op_deps[node] = preceding_ops.size(); + if (preceding_ops.empty()) { + ready_ops.insert(node); + } + } + + for (auto *op_desc : ops) { + ir::Node *found_node = nullptr; + for (auto *node : ready_ops) { + if (IsSameOpDesc(op_desc, node->Op())) { + PADDLE_ENFORCE(found_node == nullptr, + "Found multiple op_desc in graph: %s", op_desc->Type()); + found_node = node; + } + } + + PADDLE_ENFORCE_NOT_NULL(found_node, "Cannot find op_desc in graph: %s", + op_desc->Type()); + for (auto *pending_op : pending_ops[found_node]) { + if (--op_deps.at(pending_op) == 0) { + ready_ops.insert(pending_op); + } + } + ready_ops.erase(found_node); + if (skip_dist_ops.count(op_desc->Type()) == 0) { + op_node_list.push_back(found_node); + } + } + + for (size_t i = 1; i < op_node_list.size(); ++i) { + auto *dep_var = graph->CreateControlDepVar(); + op_node_list[i]->inputs.push_back(dep_var); + op_node_list[i - 1]->outputs.push_back(dep_var); + dep_var->outputs.push_back(op_node_list[i]); + dep_var->inputs.push_back(op_node_list[i - 1]); + VLOG(100) << "Add dependencies between " << op_node_list[i - 1]->Name() + << " and " << op_node_list[i]->Name(); + } + return graph; +} + +} // namespace details +} // namespace framework +} // namespace paddle + +REGISTER_PASS(sequential_execution_pass, + paddle::framework::details::SequentialExecutionPass) + .RequirePassAttr(paddle::framework::details::kAllOpDescs); diff --git a/paddle/fluid/framework/details/sequential_execution_pass.h b/paddle/fluid/framework/details/sequential_execution_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..a04c08bc2eb3bae797d648b30a22a5fee7ba0eaa --- /dev/null +++ b/paddle/fluid/framework/details/sequential_execution_pass.h @@ -0,0 +1,34 @@ +// 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 "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace details { + +constexpr char kAllOpDescs[] = "all_op_descs"; + +class SequentialExecutionPass : public ir::Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_executor.cc b/paddle/fluid/framework/details/ssa_graph_executor.cc index 780da5478ff34ecd7096d0ef62b72bf1088dd221..af2cbd5c876fdd7c27cd679f7e9412d1b0604ecc 100644 --- a/paddle/fluid/framework/details/ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/ssa_graph_executor.cc @@ -19,14 +19,16 @@ namespace framework { namespace details { SSAGraphExecutor::~SSAGraphExecutor() {} -void ClearFetchOp(ir::Graph* graph, - std::vector>* fetch_ops) { +void ClearFetchOp(ir::Graph* graph, std::vector* fetch_ops) { if (fetch_ops->empty()) return; for (auto& op : *fetch_ops) { for (auto& out_var : op->Node()->outputs) { graph->RemoveNode(out_var); } + for (auto& in_var : op->Inputs()) { + in_var->RemoveOutput(op, op->Node()); + } graph->RemoveNode(op->Node()); } fetch_ops->clear(); diff --git a/paddle/fluid/framework/details/ssa_graph_executor.h b/paddle/fluid/framework/details/ssa_graph_executor.h index d5cf7737d565c523995e6685b73c57e5a6f0197b..860eaa25b58e4579ad792ff18618de3b90707e8d 100644 --- a/paddle/fluid/framework/details/ssa_graph_executor.h +++ b/paddle/fluid/framework/details/ssa_graph_executor.h @@ -38,8 +38,7 @@ class SSAGraphExecutor { virtual FeedFetchList Run(const std::vector& fetch_tensors) = 0; }; -void ClearFetchOp(ir::Graph* graph, - std::vector>* fetch_ops); +void ClearFetchOp(ir::Graph* graph, std::vector* fetch_ops); } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc index 31beef3ae829d72570ee7c879dac71ed600cd216..f781f02a076594b5a70fd4863ebf273e88607dfd 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" #include "paddle/fluid/framework/details/multi_devices_helper.h" +#include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/platform/profiler.h" namespace paddle { @@ -39,7 +40,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr)); std::unordered_map pending_ops; std::unordered_set pending_vars; - BlockingQueue ready_vars; + auto ready_vars = std::make_shared>(); std::unordered_set ready_ops; // For ops (e.g. nccl_all_reduce) that need to coordinate multiple // streams from multiple GPUs, it's faster to buffer them and schedule @@ -51,34 +52,34 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( for (auto &var_map : graph_->Get(details::kGraphVars)) { for (auto &name_pair : var_map) { for (auto &version_pair : name_pair.second) { - InsertPendingVar(&pending_vars, &ready_vars, version_pair.get()); + InsertPendingVar(&pending_vars, ready_vars.get(), version_pair); } } } for (auto &var : graph_->Get(details::kGraphDepVars)) { - InsertPendingVar(&pending_vars, &ready_vars, var.get()); + InsertPendingVar(&pending_vars, ready_vars.get(), var); } - for (auto &op : graph_->Get(details::kGraphOps)) { + for (auto &op : ir::FilterByNodeWrapper(*graph_)) { if (op->Inputs().empty()) { // Special case, Op has no input. - ready_ops.insert(op.get()); + ready_ops.insert(op); } else { - InsertPendingOp(&pending_ops, op.get()); + InsertPendingOp(&pending_ops, op); } } // Step 2. Insert FetchOps - std::vector> fetch_ops; - std::unordered_set> fetch_dependencies; + std::vector fetch_ops; + std::unordered_set fetch_dependencies; FeedFetchList fetch_data(fetch_tensors.size()); InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops, - &pending_vars, &ready_vars, &fetch_data); + &pending_vars, ready_vars.get(), &fetch_data); auto run_all_ops = [&](std::unordered_set &set) { for (auto *op : set) { running_ops_++; - RunOp(&ready_vars, op); + RunOp(ready_vars, op); } set.clear(); }; @@ -87,7 +88,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( run_op_futures_.clear(); exception_holder_.Clear(); event.reset(nullptr); - // Step 3. Execution while (!pending_vars.empty()) { // 1. Run All Ready ops @@ -103,13 +103,14 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( // 2. Find ready variable bool timeout; - auto cur_ready_vars = ready_vars.PopAll(1, &timeout); + auto cur_ready_vars = ready_vars->PopAll(1, &timeout); if (timeout) { if (exception_holder_.IsCaught()) { for (auto &run_op_future : run_op_futures_) { run_op_future.wait(); } + ClearFetchOp(graph_.get(), &fetch_ops); exception_holder_.ReThrow(); } else { continue; @@ -133,7 +134,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( } } PADDLE_ENFORCE(ready_ops.empty()); - // Wait FetchOps. ClearFetchOp(graph_.get(), &fetch_ops); @@ -142,8 +142,8 @@ FeedFetchList ThreadedSSAGraphExecutor::Run( void ThreadedSSAGraphExecutor::InsertFetchOps( const std::vector &fetch_tensors, - std::vector> *fetch_ops, - std::unordered_set> *fetch_dependencies, + std::vector *fetch_ops, + std::unordered_set *fetch_dependencies, std::unordered_map *pending_ops, std::unordered_set *pending_vars, BlockingQueue *ready_vars, FeedFetchList *fetch_data) { @@ -153,7 +153,7 @@ void ThreadedSSAGraphExecutor::InsertFetchOps( for (auto &var_map : graph_->Get(details::kGraphVars)) { auto it = var_map.find(fetch_var_name); if (it != var_map.end()) { - fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); + fetched_vars[fetch_var_name].push_back(*it->second.rbegin()); } } } @@ -206,17 +206,20 @@ void ThreadedSSAGraphExecutor::InsertPendingVar( } void ThreadedSSAGraphExecutor::RunOp( - BlockingQueue *ready_var_q, details::OpHandleBase *op) { + const std::shared_ptr> &ready_var_q, + details::OpHandleBase *op) { auto op_run = [ready_var_q, op, this] { try { - if (VLOG_IS_ON(10)) { - VLOG(10) << op << " " << op->Name() << " : " << op->DebugString(); + if (VLOG_IS_ON(100)) { + VLOG(100) << op << " " << op->Name() << " : " << op->DebugString(); + } + if (LIKELY(!strategy_.dry_run_)) { + op->Run(strategy_.use_cuda_); } - op->Run(strategy_.use_cuda_); - VLOG(10) << op << " " << op->Name() << " Done "; + VLOG(100) << op << " " << op->Name() << " Done "; running_ops_--; ready_var_q->Extend(op->Outputs()); - VLOG(10) << op << " " << op->Name() << "Signal posted"; + VLOG(100) << op << " " << op->Name() << "Signal posted"; } catch (...) { exception_holder_.Catch(std::current_exception()); } diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h index 512f8a4ca5a9b82a395dde11722b8db44ea5ec27..24da56c09e3e0f3894d58e5af8838c98e3e1e67c 100644 --- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -48,10 +48,10 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { // Use topological sort algorithm FeedFetchList Run(const std::vector &fetch_tensors) override; - ~ThreadedSSAGraphExecutor() {} + ~ThreadedSSAGraphExecutor() final = default; private: - void RunOp(BlockingQueue *ready_var_q, + void RunOp(const std::shared_ptr> &ready_var_q, details::OpHandleBase *op); private: @@ -70,13 +70,13 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor { BlockingQueue *ready_vars, VarHandleBase *var) const; - void InsertFetchOps( - const std::vector &fetch_tensors, - std::vector> *fetch_ops, - std::unordered_set> *fetch_dependencies, - std::unordered_map *pending_ops, - std::unordered_set *pending_vars, - BlockingQueue *ready_vars, FeedFetchList *fetch_data); + void InsertFetchOps(const std::vector &fetch_tensors, + std::vector *fetch_ops, + std::unordered_set *fetch_dependencies, + std::unordered_map *pending_ops, + std::unordered_set *pending_vars, + BlockingQueue *ready_vars, + FeedFetchList *fetch_data); private: ExecutionStrategy strategy_; diff --git a/paddle/fluid/framework/details/var_handle.cc b/paddle/fluid/framework/details/var_handle.cc index 5457870e9ff5d7cf67c9c7076b9aae94eeada779..30da029ca2a90e7faa6288557ff2f1aeb21cc1c6 100644 --- a/paddle/fluid/framework/details/var_handle.cc +++ b/paddle/fluid/framework/details/var_handle.cc @@ -20,6 +20,8 @@ namespace details { VarHandleBase::~VarHandleBase() {} +VarHandle::~VarHandle() { VLOG(4) << "deleting var handle " << DebugString(); } + std::string VarHandle::DebugString() const { std::stringstream ss; ss << name_ << ":" << place_; @@ -27,6 +29,10 @@ std::string VarHandle::DebugString() const { } std::string DummyVarHandle::DebugString() const { return node_->Name(); } + +DummyVarHandle::~DummyVarHandle() { + VLOG(4) << "deleting dummy var handle " << DebugString(); +} } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/var_handle.h b/paddle/fluid/framework/details/var_handle.h index d8c2bc40b9458a1d5a7dd8a32277d04f69295f09..3b007d7b1a52df765a2dbd41939f8f865123cb43 100644 --- a/paddle/fluid/framework/details/var_handle.h +++ b/paddle/fluid/framework/details/var_handle.h @@ -35,7 +35,10 @@ class OpHandleBase; // A variable can only be generated by a single operator. i.e. // This is a single assignment graph. struct VarHandleBase { - explicit VarHandleBase(ir::Node* node) : node_(node) {} + // Owned by `node`. No need to be deleted explicitly. + explicit VarHandleBase(ir::Node* node) : node_(node) { + node_->WrappedBy(this); + } virtual ~VarHandleBase(); @@ -49,6 +52,8 @@ struct VarHandleBase { void AddOutput(OpHandleBase* out, ir::Node* node) { if (pending_ops_.find(out) == pending_ops_.end()) { + PADDLE_ENFORCE(out != nullptr, "The output of %s should not be nullptr", + this->Node()->Name()); pending_ops_.insert(out); node_->outputs.push_back(node); } @@ -92,6 +97,8 @@ struct VarHandleBase { struct VarHandle : public VarHandleBase { explicit VarHandle(ir::Node* node) : VarHandleBase(node) {} + virtual ~VarHandle(); + std::string DebugString() const override; VarHandle(ir::Node* node, size_t version, size_t scope_index, @@ -119,6 +126,8 @@ struct VarHandle : public VarHandleBase { struct DummyVarHandle : public VarHandleBase { explicit DummyVarHandle(ir::Node* node) : VarHandleBase(node) {} + virtual ~DummyVarHandle(); + std::string DebugString() const override; }; diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 70ec6e90a4d0106b7f838e51b8357798daa4b10d..fc6b32528661fb56b39d007465046ac6fb893046 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -43,15 +43,52 @@ ExecutorPrepareContext::ExecutorPrepareContext( } ExecutorPrepareContext::~ExecutorPrepareContext() { - VLOG(5) << "destroy ExecutorPrepareContext"; + VLOG(50) << "destroy ExecutorPrepareContext"; +} + +template +static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op, + GarbageCollector* gc, + RefCntMap* ref_cnts) { + std::unordered_set erase_tensors; + + auto handler = [&](const VariableNameMap& name_map) { + for (auto& name_pair : name_map) { + for (auto& name : name_pair.second) { + auto it = ref_cnts->find(name); + if (it == ref_cnts->end()) continue; + if ((it->second)-- == 1) { + auto* var = scope.FindVar(name); + if (var != nullptr) { + VLOG(100) << "Erase tensor \'" << name << "\'"; + if (var->IsType()) { + erase_tensors.insert(var->GetMutable()); + } else if (var->IsType()) { + erase_tensors.insert( + var->GetMutable()->mutable_value()); + } + } + } + } + } + }; + + handler(op->Inputs()); + handler(op->Outputs()); + + if (!erase_tensors.empty()) { + gc->Add(erase_tensors); + } } Executor::Executor(const platform::Place& place) : place_(place) {} void Executor::Close() { #ifdef PADDLE_WITH_DISTRIBUTE + // TODO(typhoonzero): complete message will need to use real trainer_id, + // except 0. ::paddle::operators::distributed::RPCClient::GetInstance< - ::paddle::operators::distributed::GRPCClient>() + ::paddle::operators::distributed::GRPCClient>(0) ->SendComplete(); #endif } @@ -66,7 +103,7 @@ void InitializeVariable(Variable* var, proto::VarType::Type var_type) { } else if (var_type == proto::VarType::FETCH_LIST) { var->GetMutable(); } else if (var_type == proto::VarType::STEP_SCOPES) { - var->GetMutable>(); + var->GetMutable>(); } else if (var_type == proto::VarType::LOD_RANK_TABLE) { var->GetMutable(); } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) { @@ -104,21 +141,21 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, if (var->Persistable()) { auto* ptr = const_cast(ancestor_scope)->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(3) << "Create Variable " << var->Name() - << " global, which pointer is " << ptr; + VLOG(30) << "Create Variable " << var->Name() + << " global, which pointer is " << ptr; } else { auto* ptr = scope->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(3) << "Create Variable " << var->Name() - << " locally, which pointer is " << ptr; + VLOG(30) << "Create Variable " << var->Name() + << " locally, which pointer is " << ptr; } } } else { for (auto& var : global_block.AllVars()) { auto* ptr = scope->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(3) << "Create variable " << var->Name() << ", which pointer is " - << ptr; + VLOG(30) << "Create variable " << var->Name() << ", which pointer is " + << ptr; } } } @@ -249,7 +286,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, int i = 0; for (auto& feed_target : (*feed_targets)) { std::string var_name = feed_target.first; - VLOG(3) << "feed target's name: " << var_name; + VLOG(30) << "feed target's name: " << var_name; // prepend feed op auto* op = global_block->PrependOp(); @@ -272,7 +309,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, int i = 0; for (auto& fetch_target : (*fetch_targets)) { std::string var_name = fetch_target.first; - VLOG(3) << "fetch target's name: " << var_name; + VLOG(30) << "fetch target's name: " << var_name; // append fetch op auto* op = global_block->AppendOp(); @@ -331,9 +368,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, } int64_t max_memory_size = GetEagerDeletionThreshold(); - std::unique_ptr> gc; - if (max_memory_size >= 0) { + // WhileOp would set keep_kids to false + // WhileGradOp would need the scopes created in WhileOp + // Perhaps, we should not perform eager deletion in WhileOp + // The scopes and variables created by WhileOp would be deleted + // in WhileGradOp. + if (max_memory_size >= 0 && !keep_kids) { ctx->ResetReferenceCount(); #ifdef PADDLE_WITH_CUDA if (platform::is_gpu_place(place_)) { @@ -352,50 +393,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, op->Run(*local_scope, place_); if (gc != nullptr) { - std::vector erase_vars; - for (auto& input : op->Inputs()) { - for (auto& input_name : input.second) { - auto it = ctx->cur_ref_cnts_.find(input_name); - if (it == ctx->cur_ref_cnts_.end()) continue; - if (it->second == 1) { // should delete it - erase_vars.emplace_back(input_name); - ctx->cur_ref_cnts_.erase(input_name); - } else { - --(it->second); - } - } - } - - for (auto& output : op->Outputs()) { - for (auto& output_name : output.second) { - auto it = ctx->cur_ref_cnts_.find(output_name); - if (it == ctx->cur_ref_cnts_.end()) continue; - if (it->second == 1) { - erase_vars.emplace_back(output_name); - ctx->cur_ref_cnts_.erase(output_name); - } else { - --(it->second); - } - } - } - - if (!erase_vars.empty()) { - std::vector erase_tensors; - for (auto& name : erase_vars) { - auto* var = local_scope->FindVar(name); - if (var == nullptr) continue; - if (var->IsType()) { - auto* tensor = var->GetMutable(); - erase_tensors.push_back(tensor); - } - } - if (!erase_tensors.empty()) gc->Add(erase_tensors); - } + DeleteUnusedTensors(*local_scope, op.get(), gc.get(), + &(ctx->cur_ref_cnts_)); } if (FLAGS_benchmark) { - VLOG(2) << "Memory used after operator " + op->Type() + " running: " - << memory::memory_usage(place_); + VLOG(20) << "Memory used after operator " + op->Type() + " running: " + << memory::memory_usage(place_); } } @@ -420,10 +424,10 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, } if (FLAGS_benchmark) { - VLOG(2) << "-------------------------------------------------------"; - VLOG(2) << "Memory used after deleting local scope: " - << memory::memory_usage(place_); - VLOG(2) << "-------------------------------------------------------"; + VLOG(20) << "-------------------------------------------------------"; + VLOG(20) << "Memory used after deleting local scope: " + << memory::memory_usage(place_); + VLOG(20) << "-------------------------------------------------------"; } } @@ -467,7 +471,7 @@ void Executor::RunPreparedContext( void Executor::EnableMKLDNN(const ProgramDesc& program) { #ifdef PADDLE_WITH_MKLDNN - VLOG(3) << "use_mkldnn=True"; + VLOG(30) << "use_mkldnn=True"; for (size_t bid = 0; bid < program.Size(); ++bid) { auto* block = const_cast(program).MutableBlock(bid); for (auto* op : block->AllOps()) { diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index f0cc1338a8af50030a70a9797cbcd1b0567272b5..36b36d49c2728dbef93042158dffa26d8f56d529 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -32,38 +32,32 @@ template std::unordered_map GetNonPersistableReferenceCount( const ProgramDesc& prog, size_t block_id) { auto& block = prog.Block(block_id); - std::unordered_set ignored_vars; std::unordered_map ref_cnts; - for (auto var_desc : block.AllVars()) { - auto type = var_desc->Proto()->type().type(); - if (type != proto::VarType::LOD_TENSOR || var_desc->Persistable()) { - ignored_vars.insert(var_desc->Name()); // ignore persistable vars - } - } - - for (auto op_desc : block.AllOps()) { - for (auto& input : op_desc->Inputs()) { - for (auto& input_name : input.second) { - if (!ignored_vars.count(input_name)) { - if (ref_cnts.count(input_name)) - ++ref_cnts[input_name]; - else - ref_cnts[input_name] = 1; + auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) { + for (auto& name_pair : name_map) { + for (auto& name : name_pair.second) { + auto* var_desc = block.FindVar(name); + if (var_desc == nullptr || var_desc->Persistable()) continue; + auto type = var_desc->Proto()->type().type(); + if (type != proto::VarType::LOD_TENSOR && + type != proto::VarType::SELECTED_ROWS) { + continue; } - } - } - for (auto& output : op_desc->Outputs()) { - for (auto output_name : output.second) { - if (!ignored_vars.count(output_name)) { - if (ref_cnts.count(output_name)) - ++ref_cnts[output_name]; - else - ref_cnts[output_name] = 1; + auto it = ref_cnts.find(name); + if (it != ref_cnts.end()) { + ++it->second; + } else { + ref_cnts[name] = 1; } } } + }; + + for (auto op_desc : block.AllOps()) { + update_ref_cnts(op_desc, op_desc->Inputs()); + update_ref_cnts(op_desc, op_desc->Outputs()); } return ref_cnts; } diff --git a/paddle/fluid/framework/feed_fetch_method.cc b/paddle/fluid/framework/feed_fetch_method.cc index 8e1f93c5ebd448903d70f9668539e077875836e4..1f3c19c0d5901cec9acc4ac9c5dab538d620c956 100644 --- a/paddle/fluid/framework/feed_fetch_method.cc +++ b/paddle/fluid/framework/feed_fetch_method.cc @@ -25,10 +25,9 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input, const std::string& var_name, size_t index) { // If var_name Variable is not found in GlobalScope, a new variable will // be created. - VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index; + VLOG(30) << "SetFeedVariable name=" << var_name << " index=" << index; Variable* g_feed_value = scope->Var(var_name); - auto& feed_inputs = - *(g_feed_value->GetMutable>()); + auto& feed_inputs = *(g_feed_value->GetMutable()); if (index >= feed_inputs.size()) { feed_inputs.resize(index + 1); } @@ -48,8 +47,8 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name, typeid(FeedFetchList).name()); auto& fetch_outputs = *g_fetch_value->GetMutable(); auto& tensor = fetch_outputs[index]; - VLOG(3) << "Fetch " << var_name << " with index " << index - << " shape= " << tensor.dims(); + VLOG(30) << "Fetch " << var_name << " with index " << index + << " shape= " << tensor.dims(); PADDLE_ENFORCE_LT(index, fetch_outputs.size()); return tensor; } diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto index 25f0ba418433571343c5b2bbfdbf9fb940eaec52..efdabffb9b33ddf007c13008d0f3afb7a3961eda 100644 --- a/paddle/fluid/framework/framework.proto +++ b/paddle/fluid/framework/framework.proto @@ -35,6 +35,7 @@ enum AttrType { BLOCK = 8; LONG = 9; BLOCKS = 10; + LONGS = 11; } // OpDesc describes an instance of a C++ framework::OperatorBase @@ -55,6 +56,7 @@ message OpDesc { optional int32 block_idx = 12; optional int64 l = 13; repeated int32 blocks_idx = 14; + repeated int64 longs = 15; }; message Var { @@ -80,7 +82,6 @@ message OpProto { optional bool duplicable = 3 [ default = false ]; optional bool intermediate = 4 [ default = false ]; optional bool dispensable = 5 [ default = false ]; - optional string reuse = 6; } // AttrProto describes the C++ type Attribute. diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index 0076a8bece31f9a977b375717c25688fc0c95819..4cf973253cc4f1f22d2fc578a1ac3a8c95e479c9 100644 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -10,7 +10,7 @@ function(pass_library TARGET DEST) set(oneValueArgs "") set(multiValueArgs SRCS DEPS) cmake_parse_arguments(op_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass ${op_library_DEPS}) + cc_library(${TARGET} SRCS ${TARGET}.cc DEPS graph_pattern_detector pass fuse_pass_base ${op_library_DEPS}) # add more DEST here, such as train, dist and collect USE_PASS into a file automatically. if (${DEST} STREQUAL "base" OR ${DEST} STREQUAL "inference") message(STATUS "add pass ${TARGET} ${DEST}") @@ -25,19 +25,27 @@ cc_library(graph_helper SRCS graph_helper.cc DEPS graph) cc_library(pass SRCS pass.cc DEPS graph node graph_helper) cc_library(graph_traits SRCS graph_traits.cc DEPS graph) cc_library(graph_pattern_detector SRCS graph_pattern_detector.cc DEPS graph graph_helper graph_traits) +cc_library(fuse_pass_base SRCS fuse_pass_base.cc DEPS pass) pass_library(graph_to_program_pass base) pass_library(graph_viz_pass base) pass_library(fc_fuse_pass inference) -if (WITH_MKLDNN) - pass_library(conv_relu_mkldnn_fuse_pass inference) -endif () pass_library(attention_lstm_fuse_pass inference) pass_library(infer_clean_graph_pass inference) pass_library(fc_lstm_fuse_pass inference) pass_library(embedding_fc_lstm_fuse_pass inference) pass_library(fc_gru_fuse_pass inference) pass_library(seq_concat_fc_fuse_pass inference) +pass_library(multi_batch_merge_pass base) +pass_library(conv_bn_fuse_pass inference) +pass_library(seqconv_eltadd_relu_fuse_pass inference) +if(WITH_MKLDNN) + pass_library(mkldnn_placement_pass base) + pass_library(depthwise_conv_mkldnn_pass base) + pass_library(conv_bias_mkldnn_fuse_pass inference) + pass_library(conv_relu_mkldnn_fuse_pass inference) + pass_library(conv_elementwise_add_mkldnn_fuse_pass inference) +endif() cc_library(fuse_elewise_add_act_pass SRCS fuse_elewise_add_act_pass.cc DEPS pass graph_pattern_detector ) @@ -45,6 +53,7 @@ set(GLOB_PASS_LIB ${PASS_LIBRARY} CACHE INTERNAL "Global PASS library") cc_library(pass_builder SRCS pass_builder.cc DEPS pass) +cc_test(node_test SRCS node_test.cc DEPS node) cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper) cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry) cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry) @@ -52,5 +61,7 @@ cc_test(graph_to_program_pass_test SRCS graph_to_program_pass_test.cc DEPS graph cc_test(test_graph_pattern_detector SRCS graph_pattern_detector_tester.cc DEPS graph_pattern_detector) cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass framework_proto) if (WITH_MKLDNN) + cc_test(test_depthwise_conv_mkldnn_pass SRCS depthwise_conv_mkldnn_pass_tester.cc DEPS depthwise_conv_mkldnn_pass) cc_test(test_conv_relu_mkldnn_fuse_pass SRCS conv_relu_mkldnn_fuse_pass_tester.cc DEPS conv_relu_mkldnn_fuse_pass) + cc_test(test_conv_elementwise_add_mkldnn_fuse_pass SRCS conv_elementwise_add_mkldnn_fuse_pass_tester.cc DEPS conv_elementwise_add_mkldnn_fuse_pass) endif () diff --git a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc index 1c75cb5a82029b6a542a3a2f031a353f5e40f4ea..6b284b1c1a4a37803229f4d55b100ca1da3a741d 100644 --- a/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc +++ b/paddle/fluid/framework/ir/attention_lstm_fuse_pass.cc @@ -147,19 +147,19 @@ void PrepareParameters(Graph* graph, const Param& param) { scope->Var(param.LSTMX)->GetMutable(); scope->Var(param.LSTMOUT)->GetMutable(); -#define GATE_W(name__) \ - auto* W_##name__##_w0 = scope->FindVar(#name__ ".w_0"); \ - auto* W_##name__##_w1 = scope->FindVar(#name__ ".w_1"); \ - auto* W_##name__##_b0 = scope->FindVar(#name__ ".b_0"); \ - CHECK_P3(W_##name__##_w0, W_##name__##_w1, W_##name__##_b0); \ - VLOG(4) << #name__ "_w0" \ - << " shape: " << W_##name__##_w0->Get().dims(); \ - VLOG(4) << #name__ "_w1" \ - << " shape: " << W_##name__##_w1->Get().dims(); \ - VLOG(4) << #name__ "_b0" \ - << " shape: " << W_##name__##_b0->Get().dims(); \ - auto& W_##name__##_w0_t = W_##name__##_w0->Get(); \ - auto& W_##name__##_w1_t = W_##name__##_w1->Get(); \ +#define GATE_W(name__) \ + auto* W_##name__##_w0 = scope->FindVar(#name__ ".w_0"); \ + auto* W_##name__##_w1 = scope->FindVar(#name__ ".w_1"); \ + auto* W_##name__##_b0 = scope->FindVar(#name__ ".b_0"); \ + CHECK_P3(W_##name__##_w0, W_##name__##_w1, W_##name__##_b0); \ + VLOG(40) << #name__ "_w0" \ + << " shape: " << W_##name__##_w0->Get().dims(); \ + VLOG(40) << #name__ "_w1" \ + << " shape: " << W_##name__##_w1->Get().dims(); \ + VLOG(40) << #name__ "_b0" \ + << " shape: " << W_##name__##_b0->Get().dims(); \ + auto& W_##name__##_w0_t = W_##name__##_w0->Get(); \ + auto& W_##name__##_w1_t = W_##name__##_w1->Get(); \ auto& W_##name__##_b0_t = W_##name__##_b0->Get(); GATE_W(forget); @@ -208,7 +208,7 @@ void PrepareLSTMWeight(const LoDTensor& W_forget_w0, int D = W_forget_w0.dims()[0]; int M = W_forget_w1.dims()[0]; out->Resize(make_ddim({D + M, 4 * D})); - VLOG(3) << "LSTMWeight resized to " << out->dims(); + VLOG(30) << "LSTMWeight resized to " << out->dims(); float* out_data = out->mutable_data(platform::CPUPlace()); std::array tensors( @@ -262,7 +262,7 @@ std::unique_ptr AttentionLSTMFusePass::ApplyImpl( std::unordered_set specified_vars({"data_lod_attention", "cell_init", "hidden_init", "data", "week", "minute"}); - int count = 0; + size_t count = 0; for (auto* node : graph->Nodes()) { if (node->IsVar() && specified_vars.count(node->Name())) { ++count; diff --git a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..c9c4d5afe5a0cd67ea14ae7abcf2b2bad1407e39 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc @@ -0,0 +1,137 @@ +// 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/framework/ir/conv_bias_mkldnn_fuse_pass.h" +#include +#include +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace framework { +namespace ir { + +template +LoDTensor tensor_apply_eltwise(const LoDTensor& vec_a, const LoDTensor& vec_b, + BinaryOperation f) { + PADDLE_ENFORCE_EQ(vec_a.dims(), vec_b.dims()); + LoDTensor vec_y; + vec_y.Resize(vec_a.dims()); + const float* a = vec_a.data(); + const float* b = vec_b.data(); + float* y = vec_y.mutable_data(platform::CPUPlace()); + for (int i = 0; i < vec_a.numel(); i++) { + y[i] = f(a[i], b[i]); + } + return vec_y; +} + +std::unique_ptr ConvBiasFusePass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init(name_scope_, graph.get()); + + auto* scope = param_scope(); + PADDLE_ENFORCE(scope); + + GraphPatternDetector gpd; + auto* conv_input = + gpd.mutable_pattern() + ->NewNode(patterns::PDNodeName(name_scope_, "conv_input")) + ->AsInput() + ->assert_is_op_input("conv2d", "Input"); + patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_); + conv_bias_pattern(conv_input); + int found_conv_bias_count = 0; + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(40) << "handle ConvBias fuse"; + GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight, + conv_bias_pattern); // Filter + GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_bias_pattern); // tmp + GET_IR_NODE_FROM_SUBGRAPH(conv, conv, conv_bias_pattern); // CONV op + // bias + GET_IR_NODE_FROM_SUBGRAPH(eltwise_bias, eltwise_bias, conv_bias_pattern); + // output + GET_IR_NODE_FROM_SUBGRAPH(eltwise_out, eltwise_out, conv_bias_pattern); + // elementwise_add op + GET_IR_NODE_FROM_SUBGRAPH(eltwise, eltwise, conv_bias_pattern); + + PADDLE_ENFORCE(subgraph.count(conv_input)); + + // check if fuse can be done and if MKL-DNN should be used + FuseOptions fuse_option = FindFuseOption(*conv, *eltwise); + if (fuse_option == DO_NOT_FUSE || fuse_option == FUSE_NATIVE) { + VLOG(30) << "do not perform conv+bias fuse"; + return; + } + + auto* eltwise_bias_tensor = + scope->FindVar(eltwise_bias->Name())->GetMutable(); + + auto input_names = conv->Op()->InputNames(); + bool has_bias = std::find(input_names.begin(), input_names.end(), "Bias") != + input_names.end(); + if (has_bias && conv->Op()->Input("Bias").size() > 0) { + auto conv_bias_names = conv->Op()->Input("Bias"); + // add eltwise bias to existing conv bias + PADDLE_ENFORCE_EQ(conv_bias_names.size(), 1); + 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(), eltwise_bias_tensor->dims()); + *conv_bias_tensor = tensor_apply_eltwise( + *conv_bias_tensor, *eltwise_bias_tensor, std::plus()); + + conv->Op()->SetOutput("Output", + std::vector({eltwise_out->Name()})); + + GraphSafeRemoveNodes(graph.get(), {eltwise, conv_out}); + + IR_NODE_LINK_TO(conv, eltwise_out); + } else { + // take eltwise bias as conv bias + OpDesc desc; + + desc.SetInput( + "Input", std::vector({subgraph.at(conv_input)->Name()})); + desc.SetInput("Filter", std::vector({conv_weight->Name()})); + desc.SetInput("Bias", std::vector({eltwise_bias->Name()})); + desc.SetOutput("Output", std::vector({eltwise_out->Name()})); + desc.SetType("conv2d"); + + for (auto& attr : conv->Op()->GetAttrMap()) { + desc.SetAttr(attr.first, attr.second); + } + auto conv_bias_node = g->CreateOpNode(&desc); + + IR_NODE_LINK_TO(subgraph.at(conv_input), conv_bias_node); + IR_NODE_LINK_TO(conv_weight, conv_bias_node); + IR_NODE_LINK_TO(eltwise_bias, conv_bias_node); + IR_NODE_LINK_TO(conv_bias_node, eltwise_out); + + GraphSafeRemoveNodes(graph.get(), {conv, eltwise, conv_out}); + } + + found_conv_bias_count++; + }; + gpd(graph.get(), handler); + AddStatis(found_conv_bias_count); + return graph; +} +} // namespace ir +} // namespace framework +} // namespace paddle +REGISTER_PASS(conv_bias_mkldnn_fuse_pass, + paddle::framework::ir::ConvBiasFusePass); diff --git a/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..5775b83b88730ec298c421a15f5c0b83c27b0750 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h @@ -0,0 +1,36 @@ +// 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 "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" +#include "paddle/fluid/framework/ir/pass.h" +namespace paddle { +namespace framework { +namespace ir { +/* +* Fuse the Conv and Elementwise_add to a ConvBiasOp. +*/ +class ConvBiasFusePass : public FusePassBase { + public: + virtual ~ConvBiasFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + const std::string name_scope_{"conv_bias_mkldnn_fuse"}; +}; +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..34b4c26ae3a8c281cd2729f67e49c78a8f440cc5 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_bn_fuse_pass.cc @@ -0,0 +1,298 @@ +// 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/framework/ir/conv_bn_fuse_pass.h" +#include +#include +#include +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/operators/math/cpu_vec.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace framework { +namespace ir { + +#define GET_CONV_BN_NODES(pattern_name) \ + /* OPERATORS */ \ + GET_IR_NODE_FROM_SUBGRAPH(conv, conv, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(batch_norm, batch_norm, pattern_name); \ + /* CONV inputs */ \ + GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight, pattern_name); \ + /* CONV outputs */ \ + GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, pattern_name); \ + /* BN inputs */ \ + GET_IR_NODE_FROM_SUBGRAPH(bn_scale, bn_scale, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_bias, bn_bias, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_mean, bn_mean, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_variance, bn_variance, pattern_name); \ + /* BN outputs */ \ + GET_IR_NODE_FROM_SUBGRAPH(bn_out, bn_out, pattern_name); /* Out */ \ + GET_IR_NODE_FROM_SUBGRAPH(bn_mean_out, bn_mean_out, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_variance_out, bn_variance_out, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_saved_mean, bn_saved_mean, pattern_name); \ + GET_IR_NODE_FROM_SUBGRAPH(bn_saved_variance, bn_saved_variance, pattern_name) + +void recompute_bias_and_weights(const Scope* scope, + ir::Node* conv_weight, // + const ir::Node& bn_scale, // + const LoDTensor& bn_bias_tensor, // + const ir::Node& bn_mean, // + const ir::Node& bn_variance, // + LoDTensor* eltwise_y_in_tensor, // + float epsilon) { + using EigenVectorArrayMap = + Eigen::Map>; + using ConstEigenVectorArrayMap = + Eigen::Map>; + using EigenMatrixArrayMap = Eigen::Map< + Eigen::Array>; + + // Re-compute bias of conv2d from BN + PADDLE_ENFORCE_EQ(eltwise_y_in_tensor->dims(), bn_bias_tensor.dims()); + + auto* scale_tensor = scope->FindVar(bn_scale.Name())->GetMutable(); + auto* variance_tensor = + scope->FindVar(bn_variance.Name())->GetMutable(); + auto* mean_tensor = scope->FindVar(bn_mean.Name())->GetMutable(); + + ConstEigenVectorArrayMap scale_array(scale_tensor->data(), + scale_tensor->numel(), 1); + EigenVectorArrayMap variance_array( + variance_tensor->mutable_data(platform::CPUPlace()), + variance_tensor->numel(), 1); + ConstEigenVectorArrayMap mean_array(mean_tensor->data(), + mean_tensor->numel(), 1); + ConstEigenVectorArrayMap bn_bias_array(bn_bias_tensor.data(), + bn_bias_tensor.numel(), 1); + + // variance will not be used anymore, so make it std_array and then tmp_array + variance_array += epsilon; + variance_array = variance_array.sqrt(); + variance_array = scale_array / variance_array; + + EigenVectorArrayMap eltwise_y_in_array( + eltwise_y_in_tensor->mutable_data(platform::CPUPlace()), + eltwise_y_in_tensor->numel(), 1); + + eltwise_y_in_array = + ((eltwise_y_in_array - mean_array) * variance_array) + bn_bias_array; + + // Re-compute weight of conv2d from BN + auto* weights = scope->FindVar(conv_weight->Name())->GetMutable(); + auto weights_shape = weights->dims(); + auto weights_shape_2d = flatten_to_2d(weights_shape, 1); + + EigenMatrixArrayMap weights_array_2d( + weights->mutable_data(platform::CPUPlace()), weights_shape_2d[0], + weights_shape_2d[1]); + + weights_array_2d.colwise() *= variance_array; +} + +std::unique_ptr ConvBNFusePass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init(name_scope_, graph.get()); + + auto* scope = param_scope(); + PADDLE_ENFORCE(scope); + + GraphPatternDetector gpd; + auto* conv_input = + gpd.mutable_pattern() + ->NewNode(patterns::PDNodeName(name_scope_, "conv_input")) + ->AsInput() + ->assert_is_op_input("conv2d", "Input"); + patterns::ConvBN conv_bn_pattern(gpd.mutable_pattern(), name_scope_); + conv_bn_pattern(conv_input, false /*with_eltwise_add*/); + + int found_conv_bn_count = 0; + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(40) << "handle ConvBN fuse"; + + // conv, batch_norm, + // conv_weight, conv_out, + // bn_scale, bn_bias, bn_mean, bn_variance, + // bn_out, bn_mean_out, bn_variance_out, bn_saved_mean, + // bn_saved_variance + GET_CONV_BN_NODES(conv_bn_pattern); + + // check if fuse can be done and if MKL-DNN should be used + FuseOptions fuse_option = FindFuseOption(*conv, *batch_norm); + if (fuse_option == DO_NOT_FUSE) { + VLOG(30) << "do not perform conv+bn fuse"; + return; + } + + // Create eltwise_y (conv bias) variable + VarDesc eltwise_y_in_desc( + patterns::PDNodeName(name_scope_, "eltwise_y_in")); + eltwise_y_in_desc.SetPersistable(true); + auto* eltwise_y_in_node = g->CreateVarNode(&eltwise_y_in_desc); + auto* eltwise_y_in_tensor = + scope->Var(eltwise_y_in_node->Name())->GetMutable(); + + // Get batch norm bias + auto* bn_bias_tensor = + scope->FindVar(bn_bias->Name())->GetMutable(); + + // Initialize eltwise_y + eltwise_y_in_tensor->Resize(bn_bias_tensor->dims()); + std::fill_n(eltwise_y_in_tensor->mutable_data(platform::CPUPlace()), + eltwise_y_in_tensor->numel(), 0.0f); + + // update weights and biases + float epsilon = boost::get(batch_norm->Op()->GetAttr("epsilon")); + recompute_bias_and_weights(scope, conv_weight, *bn_scale, *bn_bias_tensor, + *bn_mean, *bn_variance, eltwise_y_in_tensor, + epsilon); + + // with MKL-DNN fuse conv+bn into conv with bias + // without MKL-DNN fuse conv+bn into conv+elementwise_add + if (fuse_option == FUSE_MKLDNN) { + auto input_names = conv->Op()->InputNames(); + bool has_bias = std::find(input_names.begin(), input_names.end(), + "Bias") != input_names.end(); + 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); + 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(), + eltwise_y_in_tensor->dims()); + + auto eigen_conv_bias = EigenVector::From(*conv_bias_tensor); + eigen_conv_bias += EigenVector::From(*eltwise_y_in_tensor); + } else { + // add new conv_bias node + conv->Op()->SetInput( + "Bias", std::vector({eltwise_y_in_node->Name()})); + IR_NODE_LINK_TO(eltwise_y_in_node, conv); + } + conv->Op()->SetOutput("Output", + std::vector({bn_out->Name()})); + + GraphSafeRemoveNodes( + graph.get(), + {conv_out, bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, + bn_mean_out, bn_variance_out, bn_saved_mean, bn_saved_variance}); + + IR_NODE_LINK_TO(conv, bn_out); + found_conv_bn_count++; + } else { // fuse_option == FUSE_NATIVE + // create an elementwise add node. + OpDesc desc; + desc.SetInput("X", std::vector({conv_out->Name()})); + desc.SetInput("Y", std::vector({eltwise_y_in_node->Name()})); + desc.SetOutput("Out", std::vector({bn_out->Name()})); + desc.SetType("elementwise_add"); + desc.SetAttr("axis", 1); + auto eltwise_op = g->CreateOpNode(&desc); // OpDesc will be copied. + + GraphSafeRemoveNodes( + graph.get(), + {bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out, + bn_variance_out, bn_saved_mean, bn_saved_variance}); + + IR_NODE_LINK_TO(conv_out, eltwise_op); + IR_NODE_LINK_TO(eltwise_y_in_node, eltwise_op); + IR_NODE_LINK_TO(eltwise_op, bn_out); + found_conv_bn_count++; + } + }; + + gpd(graph.get(), handler); + + AddStatis(found_conv_bn_count); + return graph; +} + +std::unique_ptr ConvEltwiseAddBNFusePass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init(name_scope_, graph.get()); + + auto* scope = param_scope(); + PADDLE_ENFORCE(scope); + + GraphPatternDetector gpd; + auto* conv_input = + gpd.mutable_pattern() + ->NewNode(patterns::PDNodeName(name_scope_, "conv_input")) + ->AsInput() + ->assert_is_op_input("conv2d", "Input"); + patterns::ConvBN conv_bn_pattern(gpd.mutable_pattern(), name_scope_); + conv_bn_pattern(conv_input, true /*with_eltwise_add*/); + + int found_conv_bn_count = 0; + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(40) << "handle ConvBN fuse"; + + // conv, batch_norm, + // conv_weight, conv_out, + // bn_scale, bn_bias, bn_mean, bn_variance, + // bn_out, bn_mean_out, bn_variance_out, bn_saved_mean,bn_saved_variance + GET_CONV_BN_NODES(conv_bn_pattern); + // OPERATORS + GET_IR_NODE_FROM_SUBGRAPH(eltwise, eltwise, conv_bn_pattern); + // BIAS inputs + GET_IR_NODE_FROM_SUBGRAPH(eltwise_y_in, eltwise_y_in, conv_bn_pattern); + // BIAS outputs + GET_IR_NODE_FROM_SUBGRAPH(eltwise_out, eltwise_out, conv_bn_pattern); + + // Get eltwise_y (conv bias) variable + auto* eltwise_y_in_tensor = + scope->FindVar(eltwise_y_in->Name())->GetMutable(); + + // Get batch norm bias + auto* bn_bias_tensor = + scope->FindVar(bn_bias->Name())->GetMutable(); + + // update weights and biases + float epsilon = boost::get(batch_norm->Op()->GetAttr("epsilon")); + recompute_bias_and_weights(scope, conv_weight, *bn_scale, *bn_bias_tensor, + *bn_mean, *bn_variance, eltwise_y_in_tensor, + epsilon); + + // Update the elementwise_add node + eltwise->Op()->SetAttr("axis", 1); + eltwise->Op()->SetOutput("Out", std::vector({bn_out->Name()})); + + GraphSafeRemoveNodes( + graph.get(), + {bn_scale, bn_bias, bn_mean, bn_variance, batch_norm, bn_mean_out, + bn_variance_out, bn_saved_mean, bn_saved_variance, eltwise_out}); + + IR_NODE_LINK_TO(eltwise, bn_out); + + found_conv_bn_count++; + }; + + gpd(graph.get(), handler); + + AddStatis(found_conv_bn_count); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(conv_bn_fuse_pass, paddle::framework::ir::ConvBNFusePass); +REGISTER_PASS(conv_eltwiseadd_bn_fuse_pass, + paddle::framework::ir::ConvEltwiseAddBNFusePass); diff --git a/paddle/fluid/framework/ir/conv_bn_fuse_pass.h b/paddle/fluid/framework/ir/conv_bn_fuse_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..2c9eb574fe8e054e0ae221f08f664b91f05d95c9 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_bn_fuse_pass.h @@ -0,0 +1,49 @@ +// 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 "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +/* + * Fuse the Conv and BatchNorm to a ConvBNMKLDNNOp. + */ +class ConvBNFusePass : public FusePassBase { + public: + virtual ~ConvBNFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + const std::string name_scope_{"conv_bn_fuse"}; +}; + +class ConvEltwiseAddBNFusePass : public FusePassBase { + public: + virtual ~ConvEltwiseAddBNFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + const std::string name_scope_{"conv_eltwiseadd_bn_fuse"}; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..8d0035ae98b093979eb8bbcc0a8d6ae5356d951f --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.cc @@ -0,0 +1,154 @@ +// 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/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h" +#include +#include + +#include "paddle/fluid/framework/ir/graph_traits.h" + +namespace paddle { +namespace framework { +namespace ir { +namespace { + +// The function keeps the graph consistent by replacing +// a node 'from' in the set of inputs nodes +// of the visited node by a node 'to'. +void CorrectGraphEdges(Graph* graph, Node* from, Node* to) { + for (auto& node : GraphTraits::DFS(*graph)) { + auto from_in_inputs = + std::find(std::begin(node.inputs), std::end(node.inputs), from); + + if (from_in_inputs != std::end(node.inputs)) { + IR_NODE_LINK_TO(to, (&node)); + + auto inputs = node.Op()->Inputs(); + + using input_type = VariableNameMap::value_type; + + std::for_each(std::begin(inputs), std::end(inputs), + [from, to, &node](const input_type& i) -> void { + auto param_names = i.second; + auto pi = std::find(std::begin(param_names), + std::end(param_names), from->Name()); + + if (pi != std::end(param_names)) { + node.Op()->SetInput(i.first, {to->Name()}); + } + }); + } + } +} +} // namespace +using graph_ptr = std::unique_ptr; + +graph_ptr ConvElementwiseAddMKLDNNFusePass::ApplyImpl(graph_ptr graph) const { + FusePassBase::Init(name_scope_, graph.get()); + + GraphPatternDetector gpd; + auto pattern = gpd.mutable_pattern(); + + patterns::Conv conv_pattern{pattern, name_scope_}; + auto conv_output = conv_pattern(); + + patterns::ElementwiseAdd elementwise_add_pattern{pattern, name_scope_}; + elementwise_add_pattern(conv_output); + + conv_output->AsIntermediate(); + + auto conv_op_has_bias = [](const Node& conv_op) -> std::pair { + auto bias_input_names = conv_op.Op()->Inputs(); + auto bias_it = bias_input_names.find("Bias"); + + if (bias_it != std::end(bias_input_names)) { + bool has_bias = !bias_it->second.empty(); + + if (has_bias) { + auto conv_bias_names = bias_it->second; + auto conv_bias_names_it = + std::find_if(std::begin(conv_op.inputs), std::end(conv_op.inputs), + [&conv_bias_names](Node* n) -> bool { + return n->Name() == conv_bias_names[0]; + }); + return std::make_pair(has_bias, *conv_bias_names_it); + } + } + + return std::make_pair(false, nullptr); + }; + + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + GET_IR_NODE_FROM_SUBGRAPH(conv_op, conv_op, conv_pattern); + GET_IR_NODE_FROM_SUBGRAPH(conv_input, conv_input, conv_pattern); + GET_IR_NODE_FROM_SUBGRAPH(conv_filter, conv_filter, conv_pattern); + GET_IR_NODE_FROM_SUBGRAPH(conv_output, conv_output, conv_pattern); + GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_op, elementwise_add_op, + elementwise_add_pattern); + GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_x, elementwise_add_x, + elementwise_add_pattern); + GET_IR_NODE_FROM_SUBGRAPH(elementwise_add_out, elementwise_add_out, + elementwise_add_pattern); + + if (FindFuseOption(*conv_op, *elementwise_add_op) != FUSE_MKLDNN) return; + + OpDesc op_desc; + op_desc.SetType("conv2d"); + + op_desc.SetInput("Input", {conv_input->Name()}); + op_desc.SetInput("Filter", {conv_filter->Name()}); + op_desc.SetInput("ResidualData", {elementwise_add_x->Name()}); + op_desc.SetOutput("Output", {conv_output->Name()}); + + bool has_bias; + Node* conv_bias; + + std::tie(has_bias, conv_bias) = conv_op_has_bias(*conv_op); + + if (has_bias) { + op_desc.SetInput("Bias", {conv_bias->Name()}); + } + + for (const auto& attr : conv_op->Op()->GetAttrMap()) { + op_desc.SetAttr(attr.first, attr.second); + } + + op_desc.SetAttr("fuse_residual_connection", true); + + auto fused_conv_op = g->CreateOpNode(&op_desc); + + IR_NODE_LINK_TO(conv_input, fused_conv_op); + IR_NODE_LINK_TO(conv_filter, fused_conv_op); + IR_NODE_LINK_TO(elementwise_add_x, fused_conv_op); + IR_NODE_LINK_TO(fused_conv_op, conv_output); + + if (has_bias) { + IR_NODE_LINK_TO(conv_bias, fused_conv_op); + } + + CorrectGraphEdges(g, elementwise_add_out, conv_output); + GraphSafeRemoveNodes(g, {elementwise_add_out, conv_op, elementwise_add_op}); + }; + + gpd(graph.get(), handler); + + return graph; +} +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(conv_elementwise_add_mkldnn_fuse_pass, + paddle::framework::ir::ConvElementwiseAddMKLDNNFusePass); diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..f4a899f1adb5e993895a40a8cfb846a67b41bb22 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h @@ -0,0 +1,38 @@ +// 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 "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +class ConvElementwiseAddMKLDNNFusePass : public FusePassBase { + public: + virtual ~ConvElementwiseAddMKLDNNFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + + const std::string name_scope_{"residual_connections_fuse_pass"}; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..348a3dfc5da78e860742595a60a0b7a8b2d92243 --- /dev/null +++ b/paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass_tester.cc @@ -0,0 +1,247 @@ +// 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 +#include + +#include "paddle/fluid/framework/ir/conv_elementwise_add_mkldnn_fuse_pass.h" +#include "paddle/fluid/framework/ir/graph_traits.h" + +namespace paddle { +namespace framework { +namespace ir { + +namespace { +constexpr int nodes_removed = 3; +constexpr int nodes_added = 1; + +void SetOp(ProgramDesc* prog, const std::string& type, + const std::vector>& inputs, + const std::pair& output) { + auto op = prog->MutableBlock(0)->AppendOp(); + op->SetType(type); + op->SetAttr("use_mkldnn", true); + + for (const auto& input : inputs) { + op->SetInput(input.first, {input.second}); + } + + op->SetOutput(output.first, {output.second}); +} + +struct IsReachable { + using func = std::function; + + auto operator()(const std::unique_ptr& graph) -> func { + auto find_node = [](const std::unique_ptr& graph, + const std::string& name) -> Node* { + for (auto& node : GraphTraits::DFS(*graph)) { + if (name == node.Name()) { + return &node; + } + } + + return nullptr; + }; + + return [&](std::string from, const std::string to) -> bool { + if (from == to) return true; + + std::map visited; + + for (auto& node : GraphTraits::DFS(*graph)) { + visited[node.Name()] = false; + } + + visited[from] = true; + + std::list queue; + queue.push_back(from); + + while (!queue.empty()) { + auto cur = find_node(graph, queue.front()); + queue.pop_front(); + + if (cur == nullptr) return false; + + for (auto n : cur->outputs) { + if (n->Name() == to) return true; + + if (!visited[n->Name()]) { + visited[n->Name()] = true; + queue.push_back(n->Name()); + } + } + } + return false; + }; + } +}; + +void AssertOpsCount(const std::unique_ptr& graph) { + int conv_count = 0; + int elementwise_add_count = 0; + + for (auto* node : graph->Nodes()) { + if (node->IsOp() && node->Op()->Type() == "conv2d") { + ++conv_count; + } + if (node->IsOp() && node->Op()->Type() == "elementwise_add") { + ++elementwise_add_count; + } + } + EXPECT_EQ(conv_count, 1); + EXPECT_EQ(elementwise_add_count, 0); +} + +ProgramDesc BuildProgramDesc(const std::vector& transient_vars, + const std::vector& persistent_vars) { + ProgramDesc prog; + + auto add_var_to_prog = [&prog](const std::string& var_name) -> VarDesc* { + auto var = prog.MutableBlock(0)->Var(var_name); + var->SetType(proto::VarType::LOD_TENSOR); + + return var; + }; + + for (const auto& v : transient_vars) { + add_var_to_prog(v); + } + + for (const auto& v : persistent_vars) { + auto var = add_var_to_prog(v); + var->SetPersistable(true); + } + + return prog; +} +} // namespace + +TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionWithElementwiseAddRelu) { + auto prog = + BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"}); + + SetOp(&prog, "conv2d", + {{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}}, + {"Output", "b"}); + SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"}); + SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"}); + + std::unique_ptr graph(new ir::Graph(prog)); + + IsReachable is_reachable; + EXPECT_TRUE(is_reachable(graph)("a", "relu")); + + auto pass = + PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass"); + int original_nodes_num = graph->Nodes().size(); + graph = pass->Apply(std::move(graph)); + int current_nodes_num = graph->Nodes().size(); + + EXPECT_TRUE(is_reachable(graph)("a", "relu")); + + EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added, + current_nodes_num); + + AssertOpsCount(graph); +} + +TEST(ConvElementwiseAddMKLDNNFusePass, + ConvolutionWithElementwiseAddReluNoBias) { + auto prog = BuildProgramDesc({"a", "b", "c", "d", "e"}, {"weights"}); + SetOp(&prog, "conv2d", {{"Input", "a"}, {"Filter", "weights"}}, + {"Output", "b"}); + SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"}); + SetOp(&prog, "relu", {{"X", "d"}}, {"Out", "e"}); + + std::unique_ptr graph(new ir::Graph(prog)); + + IsReachable is_reachable; + + EXPECT_TRUE(is_reachable(graph)("a", "relu")); + + auto pass = + PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass"); + int original_nodes_num = graph->Nodes().size(); + graph = pass->Apply(std::move(graph)); + int current_nodes_num = graph->Nodes().size(); + + EXPECT_TRUE(is_reachable(graph)("a", "relu")); + + EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added, + current_nodes_num); + + AssertOpsCount(graph); +} + +TEST(ConvElementwiseAddMKLDNNFusePass, ConvolutionElementwiseAdd) { + auto prog = BuildProgramDesc({"a", "b", "c", "d"}, {"bias", "weights"}); + SetOp(&prog, "conv2d", + {{"Input", "a"}, {"Bias", "bias"}, {"Filter", "weights"}}, + {"Output", "b"}); + SetOp(&prog, "elementwise_add", {{"X", "b"}, {"Y", "c"}}, {"Out", "d"}); + + std::unique_ptr graph(new ir::Graph(prog)); + + IsReachable is_reachable; + EXPECT_TRUE(is_reachable(graph)("a", "d")); + + auto pass = + PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass"); + int original_nodes_num = graph->Nodes().size(); + graph = pass->Apply(std::move(graph)); + int current_nodes_num = graph->Nodes().size(); + + EXPECT_FALSE(is_reachable(graph)("a", "d")); + + EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added, + current_nodes_num); + AssertOpsCount(graph); +} + +TEST(ConvElementwiseAddMKLDNNFusePass, SigmoidConvolutionAddElementwiseRelu) { + auto prog = + BuildProgramDesc({"a", "b", "c", "d", "e", "f"}, {"bias", "weights"}); + SetOp(&prog, "sigmoid", {{"X", "a"}}, {"Out", "b"}); + SetOp(&prog, "conv2d", + {{"Input", "b"}, {"Bias", "bias"}, {"Filter", "weights"}}, + {"Output", "c"}); + SetOp(&prog, "elementwise_add", {{"X", "c"}, {"Y", "d"}}, {"Out", "e"}); + SetOp(&prog, "relu", {{"X", "e"}}, {"Out", "f"}); + + std::unique_ptr graph(new ir::Graph(prog)); + + IsReachable is_reachable; + + EXPECT_TRUE(is_reachable(graph)("a", "f")); + + auto pass = + PassRegistry::Instance().Get("conv_elementwise_add_mkldnn_fuse_pass"); + int original_nodes_num = graph->Nodes().size(); + graph = pass->Apply(std::move(graph)); + int current_nodes_num = graph->Nodes().size(); + + EXPECT_TRUE(is_reachable(graph)("a", "f")); + + EXPECT_EQ(original_nodes_num - nodes_removed + nodes_added, + current_nodes_num); + AssertOpsCount(graph); +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +USE_PASS(conv_elementwise_add_mkldnn_fuse_pass); diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc index d7df6389cfd595324e284e0da10f65213ccee80f..048868e1f913e9df3d985b9e66c075a02a7f0bcb 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc +++ b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.cc @@ -38,7 +38,7 @@ std::unique_ptr ConvReLUFusePass::ApplyImpl( int found_conv_relu_count = 0; auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(4) << "handle ConvReLU fuse"; + VLOG(40) << "handle ConvReLU fuse"; GET_IR_NODE_FROM_SUBGRAPH(conv_weight, conv_weight, conv_relu_pattern); // Filter GET_IR_NODE_FROM_SUBGRAPH(conv_out, conv_out, conv_relu_pattern); // tmp @@ -46,6 +46,12 @@ std::unique_ptr ConvReLUFusePass::ApplyImpl( GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, conv_relu_pattern); // Out GET_IR_NODE_FROM_SUBGRAPH(relu, relu, conv_relu_pattern); // ReLU op + FuseOptions fuse_option = FindFuseOption(*conv, *relu); + if (fuse_option == DO_NOT_FUSE) { + VLOG(30) << "do not perform conv+relu fuse"; + return; + } + // Transform Conv node into ConvReLU node. OpDesc* desc = conv->Op(); desc->SetOutput("Output", std::vector({relu_out->Name()})); diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h index b5de0d548713772e7ad41cfb6d8b3e9460683efb..fe585bd7c41bb32ae00462e989ab4c0051fc89a8 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h +++ b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h @@ -31,7 +31,8 @@ class ConvReLUFusePass : public FusePassBase { virtual ~ConvReLUFusePass() {} protected: - std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; }; } // namespace ir diff --git a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc index 9dd780ec89ab991d6d99cb66fa2a9b683be2b9ca..19248b4dfee1da81d18cd2effac08ba68dde80fb 100644 --- a/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc +++ b/paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass_tester.cc @@ -15,25 +15,30 @@ #include "paddle/fluid/framework/ir/conv_relu_mkldnn_fuse_pass.h" #include +#include "paddle/fluid/framework/op_proto_maker.h" namespace paddle { namespace framework { namespace ir { -void SetOp(ProgramDesc* prog, const std::string& type, +void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name, const std::vector& inputs, - const std::vector& outputs) { + const std::vector& outputs, bool use_mkldnn = false) { auto* op = prog->MutableBlock(0)->AppendOp(); op->SetType(type); if (type == "conv2d") { - op->SetAttr("use_mkldnn", true); + op->SetAttr("use_mkldnn", use_mkldnn); + op->SetAttr("name", name); op->SetInput("Input", {inputs[0]}); op->SetInput("Filter", {inputs[1]}); op->SetInput("Bias", {inputs[2]}); } else if (type == "relu") { + op->SetAttr("use_mkldnn", use_mkldnn); op->SetInput("X", inputs); } op->SetOutput("Out", outputs); + op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(OpRole::kForward)); } // a->OP0->b @@ -43,7 +48,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, ProgramDesc BuildProgramDesc() { ProgramDesc prog; for (auto& v : - std::vector({"a", "b", "c", "weights", "bias", "f", "g"})) { + std::vector({"a", "b", "c", "weights", "bias", "f", "g", + "h", "weights2", "bias2", "k", "l"})) { auto* var = prog.MutableBlock(0)->Var(v); var->SetType(proto::VarType::SELECTED_ROWS); if (v == "weights" || v == "bias") { @@ -51,14 +57,24 @@ ProgramDesc BuildProgramDesc() { } } - SetOp(&prog, "OP0", std::vector({"a"}), + SetOp(&prog, "OP0", "op0", std::vector({"a"}), std::vector({"b"})); - SetOp(&prog, "OP1", std::vector({"b"}), + SetOp(&prog, "OP1", "op1", std::vector({"b"}), std::vector({"c"})); - SetOp(&prog, "conv2d", std::vector({"c", "weights", "bias"}), - std::vector({"f"})); - SetOp(&prog, "relu", std::vector({"f"}), - std::vector({"g"})); + // conv+relu, both with MKL-DNN + SetOp(&prog, "conv2d", "conv1", + std::vector({"c", "weights", "bias"}), + std::vector({"f"}), true); + SetOp(&prog, "relu", "relu1", std::vector({"f"}), + std::vector({"g"}), true); + SetOp(&prog, "OP3", "op3", std::vector({"g"}), + std::vector({"h"})); + // conv+relu, only one with MKL-DNN + SetOp(&prog, "conv2d", "conv2", + std::vector({"h", "weights2", "bias2"}), + std::vector({"k"}), true); + SetOp(&prog, "relu", "relu2", std::vector({"k"}), + std::vector({"l"})); return prog; } @@ -88,10 +104,16 @@ TEST(ConvReLUFusePass, basic) { auto* op = node->Op(); ASSERT_TRUE(op->HasAttr("use_mkldnn")); EXPECT_TRUE(boost::get(op->GetAttr("use_mkldnn"))); - ASSERT_TRUE(op->HasAttr("fuse_relu")); - bool fuse_relu = boost::get(op->GetAttr("fuse_relu")); - if (fuse_relu) { - ++conv_relu_count; + // check if only "conv1" convolution is fused + auto op_name = boost::get(op->GetAttr("name")); + if (op_name == "conv1") { + ASSERT_TRUE(op->HasAttr("fuse_relu")); + bool fuse_relu = boost::get(op->GetAttr("fuse_relu")); + if (fuse_relu) { + ++conv_relu_count; + } + } else if (op_name == "conv2") { + ASSERT_FALSE(op->HasAttr("fuse_relu")); } } } diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..5f3334578d10f64b197215bfc11d08e30747cb90 --- /dev/null +++ b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.cc @@ -0,0 +1,58 @@ +/* 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/framework/ir/depthwise_conv_mkldnn_pass.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +#define GET_NODE(id, pattern) \ + PADDLE_ENFORCE(subgraph.count(pattern.RetrieveNode(#id)), \ + "pattern has no Node called %s", #id); \ + auto* id = subgraph.at(pattern.RetrieveNode(#id)); \ + PADDLE_ENFORCE_NOT_NULL(id, "subgraph has no node %s", #id); + +std::unique_ptr DepthwiseConvMKLDNNPass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + FusePassBase::Init("depthwise_conv_mkldnn_pass", graph.get()); + GraphPatternDetector gpd; + + auto* pattern = gpd.mutable_pattern(); + pattern->NewNode("depthwise_conv") + ->assert_is_op("depthwise_conv2d") + ->assert_op_attr("use_mkldnn", true); + + int found_depthwise_conv_mkldnn_count = 0; + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(30) << "handle DepthwiseConvMKLDNN fuse"; + GET_NODE(depthwise_conv, (*pattern)); + depthwise_conv->Op()->SetType("conv2d"); + found_depthwise_conv_mkldnn_count++; + }; + + gpd(graph.get(), handler); + AddStatis(found_depthwise_conv_mkldnn_count); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(depthwise_conv_mkldnn_pass, + paddle::framework::ir::DepthwiseConvMKLDNNPass); diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..8ca6a7325186401c26eb7f9375cf83b7b97cc1c9 --- /dev/null +++ b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass.h @@ -0,0 +1,34 @@ +/* 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 "paddle/fluid/framework/ir/fuse_pass_base.h" + +namespace paddle { +namespace framework { +namespace ir { + +class DepthwiseConvMKLDNNPass : public FusePassBase { + public: + virtual ~DepthwiseConvMKLDNNPass() {} + + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..09d0b15f46a7e50afb6aea46383013ce6a6c6118 --- /dev/null +++ b/paddle/fluid/framework/ir/depthwise_conv_mkldnn_pass_tester.cc @@ -0,0 +1,123 @@ +// 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/framework/ir/depthwise_conv_mkldnn_pass.h" + +#include + +namespace paddle { +namespace framework { +namespace ir { + +void SetOp(ProgramDesc* prog, const std::string& type, const std::string& name, + const std::vector& inputs, + const std::vector& outputs, bool use_mkldnn = false) { + auto* op = prog->MutableBlock(0)->AppendOp(); + op->SetType(type); + op->SetAttr("use_mkldnn", use_mkldnn); + op->SetAttr("name", name); + op->SetInput("Input", {inputs[0]}); + op->SetInput("Filter", {inputs[1]}); + op->SetInput("Bias", {inputs[2]}); + op->SetOutput("Out", outputs); +} + +// (a, weights, bias)->depthwise conv mkldnn->b +// (b, weights2, bias2)->depthwise conv no mkldnn->c +// (c, weights3, bias3)->conv mkldnn->d +// (d, weights3, bias3)->conv no mkldnn->e +ProgramDesc BuildProgramDesc() { + ProgramDesc prog; + for (auto& v : std::vector( + {"a", "b", "c", "d", "e", "weights", "bias", "weights2", "bias2", + "weights3", "bias3", "weights4", "bias4"})) { + auto* var = prog.MutableBlock(0)->Var(v); + var->SetType(proto::VarType::SELECTED_ROWS); + if (v == "weights" || v == "bias" || v == "weights2" || v == "bias2" || + v == "weights3" || v == "bias3" || v == "weights4" || v == "bias4") { + var->SetPersistable(true); + } + } + + // depthwise conv with MKL-DNN + SetOp(&prog, "depthwise_conv2d", "conv1", + std::vector({"a", "weights", "bias"}), + std::vector({"b"}), true); + // depthwise conv without MKL-DNN + SetOp(&prog, "depthwise_conv2d", "conv2", + std::vector({"b", "weights2", "bias2"}), + std::vector({"c"}), false); + // conv with MKL-DNN + SetOp(&prog, "conv2d", "conv3", + std::vector({"c", "weights3", "bias3"}), + std::vector({"d"}), true); + // conv without MKL-dNN + SetOp(&prog, "conv2d", "conv4", + std::vector({"d", "weights4", "bias4"}), + std::vector({"e"}), false); + + return prog; +} + +TEST(DepthwiseConvMKLDNNPass, basic) { + auto prog = BuildProgramDesc(); + + std::unique_ptr graph(new ir::Graph(prog)); + + auto pass = PassRegistry::Instance().Get("depthwise_conv_mkldnn_pass"); + + struct counters { + int mkldnn_depthwise_conv_nodes; + int other_depthwise_conv_nodes; + int mkldnn_conv_nodes; + int other_conv_nodes; + }; + + counters before{1, 1, 1, 1}; + + graph = pass->Apply(std::move(graph)); + + // initialize counters before loop + counters after{0, 0, 0, 0}; + + for (auto* node : graph->Nodes()) { + if (node->IsOp()) { + auto* op = node->Op(); + if (op->Type() == "conv2d") { + if (boost::get(op->GetAttr("use_mkldnn"))) + after.mkldnn_conv_nodes++; + else + after.other_conv_nodes++; + } else if (op->Type() == "depthwise_conv2d") { + if (boost::get(op->GetAttr("use_mkldnn"))) + after.mkldnn_depthwise_conv_nodes++; + else + after.other_depthwise_conv_nodes++; + } + } + } + + EXPECT_EQ(after.other_depthwise_conv_nodes, + before.other_depthwise_conv_nodes); + EXPECT_EQ(after.other_conv_nodes, before.other_conv_nodes); + EXPECT_EQ(after.mkldnn_depthwise_conv_nodes, + before.mkldnn_depthwise_conv_nodes - 1); + EXPECT_EQ(after.mkldnn_conv_nodes, before.mkldnn_conv_nodes + 1); +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +USE_PASS(depthwise_conv_mkldnn_pass); diff --git a/paddle/fluid/framework/ir/fc_fuse_pass.cc b/paddle/fluid/framework/ir/fc_fuse_pass.cc index ca704c7f5631bbaa88f1bc2caaa22fd021de11c4..3348abb19b3339b2b3e8b50485133b15a1973a32 100644 --- a/paddle/fluid/framework/ir/fc_fuse_pass.cc +++ b/paddle/fluid/framework/ir/fc_fuse_pass.cc @@ -39,7 +39,7 @@ std::unique_ptr FCFusePass::ApplyImpl( int found_fc_count = 0; auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* g) { - VLOG(4) << "handle FC fuse"; + VLOG(40) << "handle FC fuse"; GET_IR_NODE_FROM_SUBGRAPH(w, w, fc_pattern); GET_IR_NODE_FROM_SUBGRAPH(fc_bias, bias, fc_pattern); GET_IR_NODE_FROM_SUBGRAPH(fc_out, Out, fc_pattern); diff --git a/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc b/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc index 06286a109d01af638e74e06ccc83e2a5500663ea..2db7d95cae1c8c59691fd642e2462e92ed58814f 100644 --- a/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc +++ b/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/framework/ir/fc_fuse_pass.h" #include +#include "paddle/fluid/framework/op_proto_maker.h" namespace paddle { namespace framework { @@ -32,6 +33,8 @@ void SetOp(ProgramDesc* prog, const std::string& type, op->SetInput("X", inputs); } op->SetOutput("Out", outputs); + op->SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(OpRole::kForward)); } // a->OP0->b diff --git a/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc b/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc index 648acc4a759417240d9a39749b059289182ebb1e..8ed68905beed2faedc34f194070cc76e8ff3c32d 100644 --- a/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc +++ b/paddle/fluid/framework/ir/fuse_elewise_add_act_pass.cc @@ -61,7 +61,7 @@ std::unique_ptr FuseElewiseAddActPass::FuseElewiseAddAct( auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph, Graph *g) { - VLOG(4) << "handle FuseElewiseAddAct fuse"; + VLOG(40) << "handle FuseElewiseAddAct fuse"; GET_IR_NODE_FROM_SUBGRAPH(ele_y, ele_y, elewise_add_act_pattern); GET_IR_NODE_FROM_SUBGRAPH(ele_out, elewise_add_out, elewise_add_act_pattern); @@ -77,10 +77,10 @@ std::unique_ptr FuseElewiseAddActPass::FuseElewiseAddAct( Node *elewise_add_act_node = CreateFuseElewiseAddActNode( g, act, ele_add, ele_x_n, ele_y_n, ele_out_n, act_out_n); - VLOG(4) << "\n\t " << ele_x_n << " and " << ele_y_n << " -> " - << ele_add->Name() << " -> " << ele_out_n << "\n" - << "\t " << ele_out_n << " -> " << act->Name() << " -> " - << act_out_n; + VLOG(40) << "\n\t " << ele_x_n << " and " << ele_y_n << " -> " + << ele_add->Name() << " -> " << ele_out_n << "\n" + << "\t " << ele_out_n << " -> " << act->Name() << " -> " + << act_out_n; ReLinkNodes(g, ele_out, ele_add, act, elewise_add_act_node); found_elewise_add_act_count++; @@ -113,7 +113,7 @@ std::unique_ptr FuseElewiseAddActPass::FuseActElewiseAdd( auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph, Graph *g) { - VLOG(4) << "handle FuseElewiseAddAct fuse"; + VLOG(40) << "handle FuseElewiseAddAct fuse"; GET_IR_NODE_FROM_SUBGRAPH(act_out, act_out, act_elewise_add_pattern); GET_IR_NODE_FROM_SUBGRAPH(ele_x, ele_x, act_elewise_add_pattern); GET_IR_NODE_FROM_SUBGRAPH(ele_out, elewise_add_out, @@ -129,9 +129,9 @@ std::unique_ptr FuseElewiseAddActPass::FuseActElewiseAdd( Node *elewise_add_act_node = CreateFuseElewiseAddActNode( g, ele_add, act, elewise_add_x_n, act_i_n, act_o_n, elewise_add_out_n); - VLOG(4) << "\n\t " << act_i_n << " -> " << act->Name() << " -> " << act_o_n - << "\n\t " << act_o_n << " and " << elewise_add_x_n << " -> " - << ele_add->Name() << " -> " << elewise_add_out_n; + VLOG(40) << "\n\t " << act_i_n << " -> " << act->Name() << " -> " << act_o_n + << "\n\t " << act_o_n << " and " << elewise_add_x_n << " -> " + << ele_add->Name() << " -> " << elewise_add_out_n; ReLinkNodes(g, act_out, act, ele_add, elewise_add_act_node); found_elewise_add_act_count++; @@ -165,7 +165,7 @@ std::unique_ptr FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad( auto handler = [&](const GraphPatternDetector::subgraph_t &subgraph, Graph *g) { - VLOG(4) << "handle FuseElewiseAddActGrad1 fuse"; + VLOG(40) << "handle FuseElewiseAddActGrad1 fuse"; GET_IR_NODE_FROM_SUBGRAPH(act_out, act_out, elewise_add_act_grad_pattern); GET_IR_NODE_FROM_SUBGRAPH(act_grad, act_grad, elewise_add_act_grad_pattern); GET_IR_NODE_FROM_SUBGRAPH(d_itermediate_out, d_itermediate_out, @@ -208,10 +208,10 @@ std::unique_ptr FuseElewiseAddActPass::FuseElewiseAddActInplaceGrad( auto fused_node = g->CreateOpNode(&desc); - VLOG(4) << "\n\t " << d_act_out_n << " and " << act_out_n << " -> " - << act_grad->Name() << " -> " << d_itermediate_out_n << "\n\t " - << d_itermediate_out_n << " and " << act_out_n << " -> " - << ele_add_grad->Name() << " -> " << d_itermediate_out_n; + VLOG(40) << "\n\t " << d_act_out_n << " and " << act_out_n << " -> " + << act_grad->Name() << " -> " << d_itermediate_out_n << "\n\t " + << d_itermediate_out_n << " and " << act_out_n << " -> " + << ele_add_grad->Name() << " -> " << d_itermediate_out_n; ReLinkNodes(g, d_itermediate_out, act_grad, ele_add_grad, fused_node); found_elewise_add_act_count++; diff --git a/paddle/fluid/framework/ir/fuse_pass_base.cc b/paddle/fluid/framework/ir/fuse_pass_base.cc new file mode 100644 index 0000000000000000000000000000000000000000..d70010089e4b4fbb4542ef7748b8e9ece48d3942 --- /dev/null +++ b/paddle/fluid/framework/ir/fuse_pass_base.cc @@ -0,0 +1,62 @@ +// 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/framework/ir/fuse_pass_base.h" + +namespace paddle { +namespace framework { +namespace ir { + +void FusePassBase::Init(const std::string& repr, Graph* graph) const { + repr_ = repr; + graph_ = graph; +} + +Scope* FusePassBase::param_scope() const { + PADDLE_ENFORCE(graph_->Has(kParamScopeAttr)); + return graph_->Get(kParamScopeAttr); +} + +void FusePassBase::AddStatis(int count_of_fused) const { + PADDLE_ENFORCE(graph_); + PADDLE_ENFORCE(!repr_.empty()); + if (!graph_->Has(kFuseStatisAttr)) { + graph_->Set(kFuseStatisAttr, new std::unordered_map); + } + auto& info = + graph_->Get>(kFuseStatisAttr); + info[repr_] = count_of_fused; +} + +FuseOptions FusePassBase::FindFuseOption(const Node& node1, + const Node& node2) const { +#ifdef PADDLE_WITH_MKLDNN + bool node1_mkldnn = node1.Op()->HasAttr("use_mkldnn") && + boost::get(node1.Op()->GetAttr("use_mkldnn")); + bool node2_mkldnn = node2.Op()->HasAttr("use_mkldnn") && + boost::get(node2.Op()->GetAttr("use_mkldnn")); + if (node1_mkldnn && node2_mkldnn) + return FUSE_MKLDNN; + else if (!node1_mkldnn && !node2_mkldnn) + return FUSE_NATIVE; + else + return DO_NOT_FUSE; +#else + return FUSE_NATIVE; +#endif +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/fuse_pass_base.h b/paddle/fluid/framework/ir/fuse_pass_base.h index 877bbeb502252cac77095981641d7ce283ca1eb7..c53b2a6186741d86f14faf1d21fa19aa09cec036 100644 --- a/paddle/fluid/framework/ir/fuse_pass_base.h +++ b/paddle/fluid/framework/ir/fuse_pass_base.h @@ -25,32 +25,24 @@ namespace ir { static const char kParamScopeAttr[] = "__param_scope__"; static const char kFuseStatisAttr[] = "__fuse_statis__"; +enum FuseOptions { + DO_NOT_FUSE, // fusing will not be done + FUSE_NATIVE, // fusing will be done without MKL-DNN + FUSE_MKLDNN // fusing will be done with MKL-DNN +}; + class FusePassBase : public Pass { public: - void Init(const std::string& repr, Graph* graph) const { - repr_ = repr; - graph_ = graph; - } - - Scope* param_scope() const { - PADDLE_ENFORCE(graph_->Has(kParamScopeAttr)); - return graph_->Get(kParamScopeAttr); - } - - void AddStatis(int count_of_fused) const { - PADDLE_ENFORCE(graph_); - PADDLE_ENFORCE(!repr_.empty()); - if (!graph_->Has(kFuseStatisAttr)) { - graph_->Set(kFuseStatisAttr, new std::unordered_map); - } - auto& info = - graph_->Get>(kFuseStatisAttr); - info[repr_] = count_of_fused; - } + void Init(const std::string& repr, Graph* graph) const; + Scope* param_scope() const; + void AddStatis(int count_of_fused) const; virtual ~FusePassBase() {} protected: + virtual FuseOptions FindFuseOption(const Node& node1, + const Node& node2) const; + mutable Graph* graph_; mutable std::string repr_; }; diff --git a/paddle/fluid/framework/ir/graph.cc b/paddle/fluid/framework/ir/graph.cc index 398f7095968e62f92d610f560d7574b27706d13e..a2a8baa5e45d1791120e32c62dd0dbc533668290 100644 --- a/paddle/fluid/framework/ir/graph.cc +++ b/paddle/fluid/framework/ir/graph.cc @@ -23,80 +23,83 @@ limitations under the License. */ namespace paddle { namespace framework { namespace ir { - -std::vector FindDistTrainSendVars( - const std::vector &nodes) { - std::vector send_vars; - // since parameters are all in block 0, - // it's enough to only scan send ops in block 0 - for (auto &node : nodes) { - auto op_vars = node->Op()->InputArgumentNames(); - send_vars.reserve(send_vars.size() + - std::distance(op_vars.begin(), op_vars.end())); - send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end()); - } - return send_vars; -} - -std::vector FindDistTrainRecvVars( - const std::vector &nodes) { - std::vector recv_vars; - for (auto &node : nodes) { - auto op_vars = node->Op()->OutputArgumentNames(); - recv_vars.reserve(recv_vars.size() + - std::distance(op_vars.begin(), op_vars.end())); - recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end()); - } - return recv_vars; -} - -bool IsDistTrainOp(ir::Node *node, const std::vector &send_vars, - const std::vector &recv_vars) { - if (send_vars.size() == 0 || recv_vars.size() == 0) { - return false; - } - - /** - * Check any of opvars contains `.block` and in sendvars - */ - auto checker = [](const std::vector &opvars, - const std::vector &rpc_vars) -> bool { - for (auto &var : opvars) { - // a variable name with the suffix `.block` means it's a splited - // variable by (DistributeTranspiler) - // [python/paddle/fluid/transpiler/distribute_transpiler.py] - if (var.find(".block") != std::string::npos && - std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) { - return true; - } +namespace { + +void CheckProgram(const ProgramDesc &program) { +#define _INT(role) static_cast(role) + + std::map visit; + for (OpDesc *op : program.Block(0).AllOps()) { + // For backward compatibility, some program doesn't have role added. + if (!op->HasAttr(OpProtoAndCheckerMaker::OpRoleAttrName())) continue; + int role_id = + boost::get(op->GetAttr(OpProtoAndCheckerMaker::OpRoleAttrName())); + visit[role_id] = true; + switch (role_id) { + case _INT(OpRole::kForward): + if (visit.find(_INT(OpRole::kBackward)) != visit.end()) { + LOG(ERROR) + << "Cannot add backward operator before forward operator %s." + << op->Type(); + } + break; + case _INT(OpRole::kBackward): + case _INT(OpRole::kBackward) | _INT(OpRole::kLoss): + PADDLE_ENFORCE( + visit.find(_INT(OpRole::kOptimize)) == visit.end(), + "Cannot add backward operator %s after optimize operator.", + op->Type()); + break; + case _INT(OpRole::kForward) | _INT(OpRole::kLoss): + PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward) | + _INT(OpRole::kLoss)) == visit.end(), + "Cannot add backward|loss operator before " + "forward|loss operator %s.", + op->Type()); + PADDLE_ENFORCE( + visit.find(_INT(OpRole::kOptimize)) == visit.end(), + "Cannot add forward|loss operator %s after optimize operator.", + op->Type()); + break; + case _INT(OpRole::kOptimize): + case _INT(OpRole::kOptimize) | _INT(OpRole::kLRSched): + PADDLE_ENFORCE(visit.find(_INT(OpRole::kBackward)) != visit.end(), + "Optimize operators %s must follow backward operator.", + op->Type()); + break; + case _INT(OpRole::kLRSched): + case _INT(OpRole::kDist): + case _INT(OpRole::kRPC): + case _INT(OpRole::kNotSpecified): + break; + default: + LOG(FATAL) << "Unknown operator role. Don't add new role because " + "you don't know what you are doing."; } - return false; - }; - - std::vector input_var_names; - std::vector output_var_names; - for (ir::Node *input : node->inputs) { - input_var_names.push_back(input->Name()); - } - for (ir::Node *output : node->outputs) { - output_var_names.push_back(output->Name()); } - return checker(output_var_names, send_vars) || - checker(input_var_names, recv_vars); +#undef _INT } +} // namespace Graph::Graph(const ProgramDesc &program) : program_(program) { + CheckProgram(program_); // Make the nodes id start from 0. Node::ResetId(); + auto var_nodes = InitFromProgram(program_); + ResolveHazard(var_nodes); +} - VLOG(3) << "block in program:" << program_.Size(); +std::map> Graph::InitFromProgram( + const ProgramDesc &program) { + VLOG(30) << "block in program:" << program_.Size(); std::unordered_map all_vars; + // var nodes for each var name, will have multiple versions in SSA + std::map> var_nodes; for (auto *var : program.Block(0).AllVars()) { all_vars.emplace(var->Name(), var); } - std::map> var_nodes; for (auto *op : program.Block(0).AllOps()) { ir::Node *node = CreateOpNode(op); // For input args, reuse the same var name if it was created before. @@ -134,7 +137,11 @@ Graph::Graph(const ProgramDesc &program) : program_(program) { var->inputs.push_back(node); } } + return std::move(var_nodes); +} +void Graph::ResolveHazard( + const std::map> &var_nodes) { /** * We should handle write after read(WAR) and write after write(WAW) here. * Because some of the operators of the program can be executed parallelly. @@ -153,6 +160,7 @@ Graph::Graph(const ProgramDesc &program) : program_(program) { auto it_old = versions.rbegin(); ++it_old; for (; it_old != versions.rend(); it_new = it_old, ++it_old) { + VLOG(30) << "deal with var: " << (*it_new)->Name(); ir::Node *write_op = (*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0]; const auto &read_ops = (*it_old)->outputs; diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h index ab687e760a761d4e445726bd5149966adc2403d0..6384d89d2f2af4ab1d733af5eb1561cab2d09728 100644 --- a/paddle/fluid/framework/ir/graph.h +++ b/paddle/fluid/framework/ir/graph.h @@ -89,7 +89,7 @@ class Graph { attr_name); attrs_[attr_name] = attr; attr_dels_[attr_name] = [attr, attr_name]() { - VLOG(3) << "deleting " << attr_name; + VLOG(30) << "deleting " << attr_name; delete attr; }; } @@ -102,6 +102,15 @@ class Graph { attr_dels_[attr_name] = []() {}; } + template + void Erase(const std::string &attr_name) { + PADDLE_ENFORCE(attrs_.count(attr_name) != 0, "%s not set in the graph", + attr_name); + attr_dels_[attr_name](); + attrs_.erase(attr_name); + attr_dels_.erase(attr_name); + } + const std::unordered_set &Nodes() const { return node_set_; } // Create a normal variable with non-null VarDesc. @@ -160,6 +169,12 @@ class Graph { return nullptr; } + std::map> InitFromProgram( + const ProgramDesc &program); + + void ResolveHazard( + const std::map> &var_nodes); + private: // This method takes ownership of `node`. ir::Node *AddNode(ir::Node *node) { diff --git a/paddle/fluid/framework/ir/graph_helper.cc b/paddle/fluid/framework/ir/graph_helper.cc index c54766d95a61ac1a4b61566c6de62cbc86685a1d..98112c1ed317c230cb5150e7cbc6d0d173256601 100644 --- a/paddle/fluid/framework/ir/graph_helper.cc +++ b/paddle/fluid/framework/ir/graph_helper.cc @@ -33,8 +33,9 @@ void SortHelper( } } - VLOG(3) << "topology sort insert: " << node->Name() - << reinterpret_cast(node) << " input " << node->inputs.size(); + VLOG(30) << "topology sort insert: " << node->Name() + << reinterpret_cast(node) << " input " + << node->inputs.size(); ret->push_back(node); } @@ -103,9 +104,9 @@ std::map> BuildOperationAdjList( for (auto &var : n->inputs) { for (auto &adj_n : var->inputs) { PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation); - VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast(adj_n) - << " -> " << n->Name() << reinterpret_cast(n) - << " via " << var->Name() << reinterpret_cast(var); + VLOG(40) << "adj " << adj_n->Name() << reinterpret_cast(adj_n) + << " -> " << n->Name() << reinterpret_cast(n) + << " via " << var->Name() << reinterpret_cast(var); adj_list[n].insert(adj_n); } } @@ -120,19 +121,25 @@ size_t GraphNum(const Graph &graph) { std::deque q_nodes; std::vector> graph_nodes; std::unordered_set g_nodes; + // q_set used to record records in the queue. + std::unordered_set q_set; size_t graph_count = 0; - auto traverse_nodes = [&visited_nodes, - &q_nodes](const std::vector &nodes) { - std::copy_if( - nodes.begin(), nodes.end(), std::back_inserter(q_nodes), - [&visited_nodes](Node *node) { return !visited_nodes.count(node); }); + auto traverse_nodes = [&visited_nodes, &q_nodes, + &q_set](const std::vector &nodes) { + for (auto n : nodes) { + if (visited_nodes.count(n) == 0 && q_set.count(n) == 0) { + q_nodes.push_back(n); + q_set.insert(n); + } + } }; while (visited_nodes.size() != nodes.size()) { if (!q_nodes.empty()) { auto cur_node = q_nodes.front(); q_nodes.pop_front(); + q_set.erase(cur_node); visited_nodes.insert(cur_node); g_nodes.insert(cur_node); traverse_nodes(cur_node->inputs); @@ -146,6 +153,7 @@ size_t GraphNum(const Graph &graph) { for (auto &n : nodes) { if (visited_nodes.count(n) == 0) { q_nodes.push_back(n); + q_set.insert(n); break; } } @@ -156,10 +164,10 @@ size_t GraphNum(const Graph &graph) { graph_nodes.emplace_back(g_nodes); } - if (VLOG_IS_ON(10)) { - VLOG(10) << "graph_num: " << graph_nodes.size(); + if (VLOG_IS_ON(100)) { + VLOG(100) << "graph_num: " << graph_nodes.size(); for (auto &g_n : graph_nodes) { - VLOG(10) << "graph_nodes: " << g_n.size(); + VLOG(100) << "graph_nodes: " << g_n.size(); if (g_n.size() < 10) { std::stringstream out; for (auto &node : g_n) { @@ -173,7 +181,7 @@ size_t GraphNum(const Graph &graph) { } out << "]"; } - VLOG(10) << out.str(); + VLOG(100) << out.str(); } } } diff --git a/paddle/fluid/framework/ir/graph_helper.h b/paddle/fluid/framework/ir/graph_helper.h index ec46b38c01b8c369ab37b4fbd5497ec120d8db91..8d92c406689ab3a97596a8666ceb452aec4be170 100644 --- a/paddle/fluid/framework/ir/graph_helper.h +++ b/paddle/fluid/framework/ir/graph_helper.h @@ -37,6 +37,15 @@ std::vector TopologySortOperations(const Graph &graph); std::map> BuildOperationAdjList( const Graph &graph); +template +std::vector FilterByNodeWrapper(const Graph &graph) { + std::vector ret; + for (ir::Node *n : graph.Nodes()) { + if (n->IsWrappedBy()) ret.push_back(&n->Wrapper()); + } + return ret; +} + } // namespace ir } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_helper_test.cc b/paddle/fluid/framework/ir/graph_helper_test.cc index cea902809339f9d45b0e2525163f08a3c1c44c95..260a73ae763bd2cdea9948e4d928377a7c718dda 100644 --- a/paddle/fluid/framework/ir/graph_helper_test.cc +++ b/paddle/fluid/framework/ir/graph_helper_test.cc @@ -200,15 +200,15 @@ TEST(GraphHelperTest, GraphNum) { Graph g(prog); BuildZeroGraph(&g); - ASSERT_EQ(GraphNum(g), 0); + ASSERT_EQ(GraphNum(g), 0UL); Graph g2(prog); BuildOneGraph(&g2); - ASSERT_EQ(GraphNum(g2), 1); + ASSERT_EQ(GraphNum(g2), 1UL); Graph g3(prog); BuildTwoGraphs(&g3); - ASSERT_EQ(GraphNum(g3), 2); + ASSERT_EQ(GraphNum(g3), 2UL); } } // namespace ir diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index 46c6a52c09e896596aa6d8e1e901955a68a4957d..b534a5509279ef7bfc5fc92ec726224e6c5ed16f 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -12,6 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include #include #include #include @@ -91,19 +92,19 @@ void GraphPatternDetector::operator()(Graph *graph, PrettyLogEndl(Style::detail(), "--- detect %d subgraphs", subgraphs.size()); int id = 0; for (auto &g : subgraphs) { - VLOG(3) << "optimizing #" << id++ << " subgraph"; + VLOG(30) << "optimizing #" << id++ << " subgraph"; handler(g, graph); } } bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) { - VLOG(3) << "mark pdnodes in graph"; + VLOG(30) << "mark pdnodes in graph"; if (graph.Nodes().empty()) return false; for (auto &node : GraphTraits::DFS(graph)) { for (const auto &pdnode : pattern_.nodes()) { if (pdnode->Tell(&node)) { - VLOG(4) << "pdnode " << pdnode->name() << " marked"; + VLOG(40) << "pdnode " << pdnode->name() << " marked"; pdnodes2nodes_[pdnode.get()].insert(&node); } } @@ -111,7 +112,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) { // Check to early stop if some PDNode can't find matched Node. for (auto &pdnode : pattern_.nodes()) { if (!pdnodes2nodes_.count(pdnode.get())) { - VLOG(4) << pdnode->name() << " can't find matched Node, early stop"; + VLOG(40) << pdnode->name() << " can't find matched Node, early stop"; // return false; } } @@ -120,7 +121,7 @@ bool GraphPatternDetector::MarkPDNodesInGraph(const ir::Graph &graph) { GetMarkedNodes(const_cast(&graph)).insert(n); } } - VLOG(3) << pdnodes2nodes_.size() << " nodes marked"; + VLOG(30) << pdnodes2nodes_.size() << " nodes marked"; return !pdnodes2nodes_.empty(); } @@ -166,10 +167,12 @@ struct HitGroup { bool Match(Node *node, PDNode *pat) { if (nodes_.count(node)) { - if (!roles.count(pat)) return false; - return roles[pat] == node; + if (roles.count(pat) && roles[pat] == node) return true; + return false; + } else { + if (roles.count(pat) && roles[pat] != node) return false; + return true; } - return !roles.count(pat) || roles.at(pat) == node; } void Register(Node *node, PDNode *pat) { @@ -197,7 +200,6 @@ GraphPatternDetector::DetectPatterns() { std::vector result; std::vector init_groups; std::array, 2> bi_records; - // PADDLE_ENFORCE(!pattern_.edges().empty(), "At least one edge is needed"); auto *first_pnode = pattern_.edges().empty() ? pattern().nodes().front().get() : pattern_.edges().front().first; if (!pdnodes2nodes_.count(first_pnode)) return result; @@ -213,7 +215,7 @@ GraphPatternDetector::DetectPatterns() { // Extend a PDNode to subgraphs by deducing the connection relations defined // in edges of PDNodes. for (const auto &edge : pattern_.edges()) { - VLOG(4) << "check " << edge.first->name() << " -> " << edge.second->name(); + VLOG(40) << "check " << edge.first->name() << " -> " << edge.second->name(); // TODO(Superjomn) Fix bug here, the groups might be duplicate here. // Each role has two PDNodes, which indicates two roles. // Detect two Nodes that can match these two roles and they are connected. @@ -224,14 +226,15 @@ GraphPatternDetector::DetectPatterns() { // source -> target for (Node *source : pdnodes2nodes_[edge.first]) { for (Node *target : pdnodes2nodes_[edge.second]) { - VLOG(8) << "check " << source->id() << " -- " << target->id(); + VLOG(80) << "check " << source->id() << " -- " << target->id(); // TODO(Superjomn) add some prune strategies. for (const auto &group : pre_groups) { - HitGroup new_group = group; - if (IsNodesLink(source, target) && - new_group.Match(source, edge.first)) { - new_group.Register(source, edge.first); - if (new_group.Match(target, edge.second)) { + if (IsNodesLink(source, target)) { + HitGroup new_group = group; + bool flag = new_group.Match(source, edge.first) && + new_group.Match(target, edge.second); + if (flag) { + new_group.Register(source, edge.first); new_group.Register(target, edge.second); cur_groups.push_back(new_group); // TODO(Superjomn) need to unique @@ -240,12 +243,13 @@ GraphPatternDetector::DetectPatterns() { } } } - VLOG(3) << "step " << step << " get records: " << cur_groups.size(); + VLOG(30) << "step " << step << " get records: " << cur_groups.size(); for (auto &group : cur_groups) { for (auto &item : group.roles) { - VLOG(4) << "node " << item.second->id() << " as " << item.first->name(); + VLOG(40) << "node " << item.second->id() << " as " + << item.first->name(); } - VLOG(4) << "========================================================="; + VLOG(40) << "========================================================="; } } @@ -259,18 +263,35 @@ GraphPatternDetector::DetectPatterns() { return result; } +struct GraphItemLessThan { + bool operator()(const std::pair &a, + const std::pair &b) { + if (a.first != b.first) { + return a.first < b.first; + } else { + return a.second < b.second; + } + } +}; + +// TODO(Superjomn) enhance the function as it marks unique unique as duplicates +// see https://github.com/PaddlePaddle/Paddle/issues/13550 void GraphPatternDetector::UniquePatterns( std::vector *subgraphs) { if (subgraphs->empty()) return; std::vector result; std::unordered_set set; + std::hash hasher; for (auto &g : *subgraphs) { - size_t key = 0; - for (auto &item : g) { - key ^= std::hash{}(item.first); - key ^= std::hash{}(item.second); + // Sort the items in the sub-graph, and transform to a string key. + std::vector> sorted_keys(g.begin(), g.end()); + std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemLessThan()); + std::stringstream ss; + for (auto &item : sorted_keys) { + ss << item.first << ":" << item.second; } + auto key = hasher(ss.str()); if (!set.count(key)) { result.emplace_back(g); set.insert(key); @@ -626,6 +647,112 @@ bool VarLinksFromOp(Node *node, const std::string &op_type) { return false; } +PDNode *patterns::ConvBN::operator()(paddle::framework::ir::PDNode *conv_input, + bool with_eltwise_add) { + // Create Operators + conv_input->assert_is_op_input("conv2d", "Input"); + auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d"); + + PDNode *eltwise_op = nullptr; + if (with_eltwise_add) { + eltwise_op = + pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add"); + } + auto *batch_norm_op = + pattern->NewNode(batch_norm_repr())->assert_is_op("batch_norm"); + // Create variables + // Conv Filter + auto *conv_weight_var = pattern->NewNode(conv_weight_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("conv2d", "Filter"); + + auto *conv_out_var = pattern->NewNode(conv_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("conv2d"); + + PDNode *eltwise_y_in_var = nullptr; + PDNode *eltwise_out_var = nullptr; + if (with_eltwise_add) { + // Conv output as Bias input + conv_out_var->assert_is_op_input("elementwise_add", "X"); + // Bias + eltwise_y_in_var = pattern->NewNode(eltwise_y_in_repr()) + ->assert_is_op_input("elementwise_add", "Y") + ->AsInput(); + eltwise_out_var = pattern->NewNode(eltwise_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("elementwise_add"); + } else { + // Conv output as BN input + conv_out_var->assert_is_op_input("batch_norm", "X"); + } + + // BN Scale + auto *bn_scale_var = pattern->NewNode(bn_scale_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("batch_norm", "Scale"); + // BN Bias + auto *bn_bias_var = pattern->NewNode(bn_bias_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("batch_norm", "Bias"); + // BN Mean + auto *bn_mean_var = pattern->NewNode(bn_mean_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("batch_norm", "Mean"); + // BN Variance + auto *bn_variance_var = pattern->NewNode(bn_variance_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("batch_norm", "Variance"); + + // BN output + auto *bn_out_var = pattern->NewNode(bn_out_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm"); + + auto *bn_mean_out_var = pattern->NewNode(bn_mean_out_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm", "MeanOut"); + + auto *bn_variance_out_var = + pattern->NewNode(bn_variance_out_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm", "VarianceOut"); + + auto *bn_saved_mean_var = + pattern->NewNode(bn_saved_mean_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm", "SavedMean"); + + auto *bn_saved_variance_var = + pattern->NewNode(bn_saved_variance_repr()) + ->AsOutput() + ->assert_is_op_output("batch_norm", "SavedVariance"); + + conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var}); + + if (with_eltwise_add) { + eltwise_op->LinksFrom({conv_out_var, eltwise_y_in_var}) + .LinksTo({eltwise_out_var}); + batch_norm_op + ->LinksFrom({eltwise_out_var, bn_scale_var, bn_bias_var, bn_mean_var, + bn_variance_var}) + .LinksTo({bn_out_var, bn_mean_out_var, bn_variance_out_var, + bn_saved_mean_var, bn_saved_variance_var}); + } else { + batch_norm_op + ->LinksFrom({conv_out_var, bn_scale_var, bn_bias_var, bn_mean_var, + bn_variance_var}) + .LinksTo({bn_out_var, bn_mean_out_var, bn_variance_out_var, + bn_saved_mean_var, bn_saved_variance_var}); + } + return bn_out_var; +} + PDNode *patterns::ConvReLU::operator()( paddle::framework::ir::PDNode *conv_input) { // Create Operators @@ -653,6 +780,51 @@ PDNode *patterns::ConvReLU::operator()( return relu_out_var; } +PDNode *patterns::SeqConvEltAddRelu::operator()( + paddle::framework::ir::PDNode *seqconv_input) { + // Create Operators + seqconv_input->assert_is_op_input("sequence_conv", "X"); + auto *seqconv_op = pattern->NewNode(seqconv_repr()) + ->assert_is_op("sequence_conv") + ->assert_op_attr("paddingTrainable", false) + ->assert_op_attr("contextStride", 1); + + auto *eltadd_op = + pattern->NewNode(eltadd_repr())->assert_is_op("elementwise_add"); + auto *relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu"); + // Create variables + // Filter + auto *seqconv_weight_var = + pattern->NewNode(seqconv_weight_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("sequence_conv", "Filter"); + // Bias + auto *eltadd_bias_var = pattern->NewNode(eltadd_bias_repr()) + ->AsInput() + ->assert_is_op_input("elementwise_add"); + // intermediate variable, will be removed in the IR after fuse. + auto *seqconv_out_var = pattern->NewNode(seqconv_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("sequence_conv") + ->assert_is_op_input("elementwise_add"); + auto *eltadd_out_var = pattern->NewNode(eltadd_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("elementwise_add") + ->assert_is_only_input_of_op("relu"); + // output + auto *relu_out_var = pattern->NewNode(relu_out_repr()) + ->AsOutput() + ->assert_is_op_output("relu"); + + seqconv_op->LinksFrom({seqconv_input, seqconv_weight_var}) + .LinksTo({seqconv_out_var}); + eltadd_op->LinksFrom({seqconv_out_var, eltadd_bias_var}) + .LinksTo({eltadd_out_var}); + relu_op->LinksFrom({eltadd_out_var}).LinksTo({relu_out_var}); + return relu_out_var; +} + PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x, bool with_bias) { // Create shared nodes. @@ -858,6 +1030,79 @@ PDNode *patterns::ElewiseAddActInplaceGrad::operator()( return ele_add_grad; } +PDNode *patterns::ConvBias::operator()( + paddle::framework::ir::PDNode *conv_input) { + // Create Operators + conv_input->assert_is_op_input("conv2d", "Input"); + auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d"); + auto *eltiwse_op = + pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add"); + // Create variables + // Filter + auto *conv_weight_var = pattern->NewNode(conv_weight_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("conv2d", "Filter"); + // intermediate variable, will be removed in the IR after fuse. + auto *conv_out_var = pattern->NewNode(conv_out_repr()) + ->AsIntermediate() + ->assert_is_only_output_of_op("conv2d") + ->assert_is_op_input("elementwise_add"); + // Bias stored in elementwise_add + auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr()) + ->AsInput() + ->assert_is_persistable_var() + ->assert_is_op_input("elementwise_add", "Y"); + // output + auto *eltwise_out_var = pattern->NewNode(eltwise_out_repr()) + ->AsOutput() + ->assert_is_op_output("elementwise_add"); + conv_op->LinksFrom({conv_input, conv_weight_var}).LinksTo({conv_out_var}); + eltiwse_op->LinksFrom({conv_out_var, eltwise_bias_var}) + .LinksTo({eltwise_out_var}); + return eltwise_out_var; +} + +PDNode *patterns::Conv::operator()() { + auto conv_op = pattern->NewNode(conv_op_repr())->assert_is_op("conv2d"); + + auto input_var = pattern->NewNode(conv_input_repr()) + ->AsInput() + ->assert_is_op_input("conv2d", "Input"); + + auto filter_var = pattern->NewNode(conv_filter_repr()) + ->AsInput() + ->assert_is_op_input("conv2d", "Filter"); + + auto output_var = pattern->NewNode(conv_output_repr()) + ->AsOutput() + ->assert_is_op_output("conv2d", "Output"); + + conv_op->LinksFrom({input_var, filter_var}); + conv_op->LinksTo({output_var}); + + return output_var; +} + +PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var) { + auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr()) + ->assert_is_op("elementwise_add"); + + x_var->assert_is_op_input("elementwise_add", "X"); + + auto y_var = pattern->NewNode(elementwise_add_x_repr()) + ->AsInput() + ->assert_is_op_input("elementwise_add", "Y"); + + auto out_var = pattern->NewNode(elementwise_add_out_repr()) + ->AsOutput() + ->assert_is_op_output("elementwise_add", "Out"); + + elementwise_add_op->LinksFrom({x_var, y_var}); + elementwise_add_op->LinksTo({out_var}); + + return out_var; +} } // namespace ir } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h index 508113bf4fcab274394f2705c36eddbf4ba3c77a..9e462ac671ee931fc17a31f32a76049a0990341f 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.h +++ b/paddle/fluid/framework/ir/graph_pattern_detector.h @@ -128,6 +128,15 @@ struct PDNode { const std::unordered_set& op_types, const std::string& argument, int nth); + template + PDNode* assert_op_attr(const std::string& attr_name, const T& attr) { + asserts_.emplace_back([=](Node* x) { + return x && x->IsOp() && x->Op()->HasAttr(attr_name) && + boost::get(x->Op()->GetAttr(attr_name)) == attr; + }); + return this; + } + private: PDNode(PDPattern* pattern, const std::string& name = "", Type type = Type::kVar) @@ -375,6 +384,44 @@ struct PatternBase { size_t id_; }; +// Conv with batch norm +// op: conv + (elementwise_add +) batch_norm +// named nodes: +// conv_weight, conv_out, conv, +// bn_x, bn_scale, bn_bias, bn_mean, bn_variance, +// bn_batch_norm, bn_y, bn_mean_out, bn_variance_out, +// bn_saved_mean, bn_saved_variance +struct ConvBN : public PatternBase { + ConvBN(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "conv_bn") {} + + PDNode* operator()(PDNode* conv_input, bool with_eltwise_add); + + // declare operator node's name + PATTERN_DECL_NODE(conv); + PATTERN_DECL_NODE(batch_norm); + PATTERN_DECL_NODE(eltwise); // ELEMENTWISE_ADD + // CONV inputs + PATTERN_DECL_NODE(conv_weight); // Filter + // CONV outputs + PATTERN_DECL_NODE(conv_out); // tmp + // ELTWISE inputs + PATTERN_DECL_NODE(eltwise_y_in); + // ELTWISE outputs + PATTERN_DECL_NODE(eltwise_out); // tmp + // BN inputs + PATTERN_DECL_NODE(bn_scale); + PATTERN_DECL_NODE(bn_bias); + PATTERN_DECL_NODE(bn_mean); + PATTERN_DECL_NODE(bn_variance); + // BN outputs + PATTERN_DECL_NODE(bn_out); // Out + PATTERN_DECL_NODE(bn_mean_out); + PATTERN_DECL_NODE(bn_variance_out); + PATTERN_DECL_NODE(bn_saved_mean); + PATTERN_DECL_NODE(bn_saved_variance); +}; + // CONV with ReLU // op: conv + relu // named nodes: @@ -396,6 +443,31 @@ struct ConvReLU : public PatternBase { PATTERN_DECL_NODE(relu_out); }; +// SEQCONV with Elementwise_Add ReLU +// op: seqconv + elementwise_add + relu +// named nodes: +// seqconv_input, seqconv_weight, +// seqconv_out, seqconv, +// elementwise_add_bias, elementwise_add_out, elementwise_add +// relu_out, relu +struct SeqConvEltAddRelu : public PatternBase { + SeqConvEltAddRelu(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "seqconv_eltadd_relu") {} + + PDNode* operator()(PDNode* seqconv_input); + + // declare operator node's name + PATTERN_DECL_NODE(seqconv); + PATTERN_DECL_NODE(eltadd); + PATTERN_DECL_NODE(relu); + // declare variable node's name + PATTERN_DECL_NODE(seqconv_weight); + PATTERN_DECL_NODE(seqconv_out); + PATTERN_DECL_NODE(eltadd_bias); + PATTERN_DECL_NODE(eltadd_out); + PATTERN_DECL_NODE(relu_out); +}; + // FC with bias // op: mul + elementwise_add // named nodes: @@ -540,6 +612,65 @@ struct ElewiseAddActInplaceGrad : public PatternBase { PATTERN_DECL_NODE(d_ele_y); PATTERN_DECL_NODE(ele_y); }; + +// Conv with Elementwise_add as bias +// op: conv + elementwise_add +// named nodes: +// conv_input, conv_weight, +// conv_out, conv, +// eltwise_bias, eltwise_out, +// elementwise_add +struct ConvBias : public PatternBase { + ConvBias(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "conv_bias") {} + PDNode* operator()(PDNode* conv_input); + // declare operator node's name + PATTERN_DECL_NODE(conv); + PATTERN_DECL_NODE(eltwise); + // declare variable node's name + PATTERN_DECL_NODE(conv_weight); + PATTERN_DECL_NODE(conv_out); + PATTERN_DECL_NODE(eltwise_bias); + PATTERN_DECL_NODE(eltwise_out); +}; + +// Convolution op +// Forward pass for convolution. +// conv_input, conv_bias and conv_filter are inputs. +// conv_output is a result of the operator. +// residual_data is data used by skip connection. +// If residual connection fusion is on, the formula is: +// conv_output = conv_op(conv_filter, conv_input, conv_bias) +// + conv_residual_data +// If the fusion is off, conv_residual_data is not added. +struct Conv : public PatternBase { + Conv(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "convolution") {} + + PDNode* operator()(); + + PATTERN_DECL_NODE(conv_op); + PATTERN_DECL_NODE(conv_input); + PATTERN_DECL_NODE(conv_filter); + PATTERN_DECL_NODE(conv_residual_data); + PATTERN_DECL_NODE(conv_output); +}; + +// ElementwiseAdd used in residual connections. +// y_var is used and convolution output. +// The operator is removed, when residual +// connection fusion is on. +struct ElementwiseAdd : public PatternBase { + ElementwiseAdd(PDPattern* pattern, const std::string& name_scope) + : PatternBase(pattern, name_scope, "elementwise_add") {} + + PDNode* operator()(PDNode* x_var); + + PATTERN_DECL_NODE(elementwise_add_op); + PATTERN_DECL_NODE(elementwise_add_x); + PATTERN_DECL_NODE(elementwise_add_y); + PATTERN_DECL_NODE(elementwise_add_out); +}; } // namespace patterns // Link two ir::Nodes from each other. diff --git a/paddle/fluid/framework/ir/graph_test.cc b/paddle/fluid/framework/ir/graph_test.cc index cadda49c399a6d65079cacedfea61f4fd580a69a..7ed2f96eb24239d87965192d73f4ba200ff5dbeb 100644 --- a/paddle/fluid/framework/ir/graph_test.cc +++ b/paddle/fluid/framework/ir/graph_test.cc @@ -124,7 +124,7 @@ TEST(GraphTest, Basic) { ASSERT_EQ(n->outputs.size(), 0UL); } } - ASSERT_EQ(nodes.size(), 5); + ASSERT_EQ(nodes.size(), 5UL); } TEST(GraphTest, WriteAfterRead) { diff --git a/paddle/fluid/framework/ir/graph_viz_pass.cc b/paddle/fluid/framework/ir/graph_viz_pass.cc index 31ed98db72c8fd4af8c970861d386687962001ce..13dd354dc59b2bf00a741c565a4c97719eac76c3 100644 --- a/paddle/fluid/framework/ir/graph_viz_pass.cc +++ b/paddle/fluid/framework/ir/graph_viz_pass.cc @@ -41,7 +41,7 @@ std::string FormatName(const Node* node) { std::unique_ptr GraphVizPass::ApplyImpl( std::unique_ptr graph) const { const std::string graph_viz_path = Get(kGraphVizPath); - VLOG(3) << "draw IR graph viz to " << graph_viz_path; + VLOG(30) << "draw IR graph viz to " << graph_viz_path; std::unique_ptr fout(new std::ofstream(graph_viz_path)); PADDLE_ENFORCE(fout->good()); std::ostream& sout = *fout; diff --git a/paddle/fluid/framework/ir/mkldnn_placement_pass.cc b/paddle/fluid/framework/ir/mkldnn_placement_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..145a3a455c8ae2c1e6a5bc4fefa3491f420af5ba --- /dev/null +++ b/paddle/fluid/framework/ir/mkldnn_placement_pass.cc @@ -0,0 +1,37 @@ +/* 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/framework/ir/mkldnn_placement_pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +std::unique_ptr MKLDNNPlacementPass::ApplyImpl( + std::unique_ptr graph) const { + VLOG(30) << "Aplies MKL-DNN placement strategy."; + for (const Node* n : graph->Nodes()) { + if (n->IsOp() && n->Op()->HasAttr("use_mkldnn")) { + n->Op()->SetAttr("use_mkldnn", true); + } + } + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(mkldnn_placement_pass, + paddle::framework::ir::MKLDNNPlacementPass); diff --git a/paddle/fluid/framework/ir/mkldnn_placement_pass.h b/paddle/fluid/framework/ir/mkldnn_placement_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..3d4dc9e2b6ecccddea4d63e45710c80d55ef2772 --- /dev/null +++ b/paddle/fluid/framework/ir/mkldnn_placement_pass.h @@ -0,0 +1,31 @@ +/* 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 "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +class MKLDNNPlacementPass : public Pass { + protected: + std::unique_ptr ApplyImpl( + std::unique_ptr graph) const override; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/multi_batch_merge_pass.cc b/paddle/fluid/framework/ir/multi_batch_merge_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..532961e4d59ad3611dc93b20738080d1755290e8 --- /dev/null +++ b/paddle/fluid/framework/ir/multi_batch_merge_pass.cc @@ -0,0 +1,315 @@ +// 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/framework/ir/multi_batch_merge_pass.h" + +#include +#include +#include + +#include "paddle/fluid/framework/ir/graph_helper.h" +#include "paddle/fluid/framework/op_proto_maker.h" + +namespace paddle { +namespace framework { +namespace ir { + +static const char kNumRepeats[] = "num_repeats"; +typedef std::unordered_map> SSAVarList; + +ir::Node* SameNameVar(std::unordered_set all, ir::Node* target) { + for (auto n : all) { + if (target->IsVar() && target->Name() == n->Name()) { + return n; + } + } + return nullptr; +} + +VarDesc CopyVarDesc(VarDesc* var_desc) { + VarDesc repeated_var(var_desc->Name()); + // copy other variable attributes + if (var_desc->GetType() != proto::VarType::READER) { + repeated_var.SetType(var_desc->GetType()); + repeated_var.SetShape(var_desc->GetShape()); + repeated_var.SetDataType(var_desc->GetDataType()); + repeated_var.SetLoDLevel(var_desc->GetLoDLevel()); + repeated_var.SetPersistable(var_desc->Persistable()); + } else { + // TODO(typhoonzero): copy reader var + } + return repeated_var; +} + +VarDesc UpdateGradVarDesc( + VarDesc* var_desc, int repeat, + const std::unordered_set& grad_names, + const std::unordered_set& bn_vars_need_rename) { + if (grad_names.find(var_desc->Name()) != grad_names.end() || + bn_vars_need_rename.find(var_desc->Name()) != bn_vars_need_rename.end()) { + std::string new_gname = + string::Sprintf("%s.repeat.%d", var_desc->Name(), repeat); + VarDesc repeated_var = CopyVarDesc(var_desc); + repeated_var.SetName(new_gname); + VLOG(30) << "update " << var_desc->Name() << " to repeat " << repeat; + return repeated_var; + } + return *var_desc; +} + +std::unique_ptr BatchMergePass::ApplyImpl( + std::unique_ptr graph) const { + int num_repeats = Get(kNumRepeats); + std::vector forward_backward_ops; + std::vector optimize_ops; + std::vector lr_ops; // ops other than forward/backward/optimize + std::unordered_set grad_names; + + std::vector nodes = TopologySortOperations(*graph); + auto origin_nodes = graph->ReleaseNodes(); + VLOG(30) << "origin nodes count: " << origin_nodes.size(); + ir::Graph& result = *graph; + + // 1. record op nodes of different roles + for (auto node : nodes) { + if (node->IsVar()) continue; + int op_role = boost::get(node->Op()->GetAttr( + framework::OpProtoAndCheckerMaker::OpRoleAttrName())); + if ((op_role == static_cast(framework::OpRole::kForward)) || + (op_role & static_cast(framework::OpRole::kBackward)) || + (op_role & static_cast(framework::OpRole::kLoss))) { + forward_backward_ops.push_back(node); + } else if ((op_role & static_cast(framework::OpRole::kOptimize)) || + (op_role & static_cast(framework::OpRole::kDist)) || + (op_role & static_cast(framework::OpRole::kRPC))) { + optimize_ops.push_back(node); + auto op_role_var = node->Op()->GetNullableAttr( + OpProtoAndCheckerMaker::OpRoleVarAttrName()); + auto op_role_vars = boost::get>(op_role_var); + for (size_t i = 0; i < op_role_vars.size(); i += 2) { + grad_names.insert(op_role_vars[i + 1]); + } + } else if (op_role & static_cast(framework::OpRole::kLRSched)) { + lr_ops.push_back(node); + } else { // NOLINT + PADDLE_THROW("Invalid op_role: %d", static_cast(op_role)); + } + } + + // 2. copy forward backward + ir::Node* prev_repeat_last_op_node = nullptr; + // record origin_grad -> repeated grad list map. + std::map> grad_repeated_map; + std::map> created; + std::unordered_set bn_vars_need_rename; + for (int i = 0; i < num_repeats; ++i) { + std::unordered_set copied; + for (size_t node_idx = 0; node_idx < forward_backward_ops.size(); + ++node_idx) { + auto node = forward_backward_ops[node_idx]; + OpDesc repeated_op(*(node->Op()), node->Op()->Block()); + // 3. rename grad outputs to current repeat. + for (auto outname : repeated_op.OutputArgumentNames()) { + if (grad_names.find(outname) != grad_names.end()) { + std::string new_gname = string::Sprintf("%s.repeat.%d", outname, i); + repeated_op.RenameOutput(outname, new_gname); + } + } + // 3.5 let batch_norm ops use independent vars, note batch_norm_grad do + // not need this update + if (node->Name() == "batch_norm") { + // NOTE: assume bn op created by layers use save var as output mean and + // variance + std::string new_mean_name = + string::Sprintf("%s.repeat.%d", repeated_op.Input("Mean")[0], i); + std::string new_var_name = string::Sprintf( + "%s.repeat.%d", repeated_op.Input("Variance")[0], i); + bn_vars_need_rename.insert(repeated_op.Input("Mean")[0]); + bn_vars_need_rename.insert(repeated_op.Input("Variance")[0]); + VLOG(30) << "renaming " << repeated_op.Input("Mean")[0] << " to " + << new_mean_name; + repeated_op.RenameInput(repeated_op.Input("Mean")[0], new_mean_name); + repeated_op.RenameInput(repeated_op.Input("Variance")[0], new_var_name); + repeated_op.RenameOutput(repeated_op.Output("MeanOut")[0], + new_mean_name); + repeated_op.RenameOutput(repeated_op.Output("VarianceOut")[0], + new_var_name); + } + + // 3.9 do copy + auto repeated_node = result.CreateOpNode(&repeated_op); + copied.insert(node); + + // 4. add deps between repeats + if (node_idx == forward_backward_ops.size() - 1) { + prev_repeat_last_op_node = repeated_node; + } + if (node_idx == 0 && prev_repeat_last_op_node) { + auto* depvar = result.CreateControlDepVar(); + prev_repeat_last_op_node->outputs.push_back(depvar); + depvar->inputs.push_back(prev_repeat_last_op_node); + repeated_node->inputs.push_back(depvar); + depvar->outputs.push_back(repeated_node); + } + + for (auto in_node : node->inputs) { + if (in_node->IsCtrlVar()) { + continue; + } + ir::Node* var = nullptr; + auto updated_var = UpdateGradVarDesc(in_node->Var(), i, grad_names, + bn_vars_need_rename); + // should be initialized by startup, how to initilize tensor in the + // scope? + if (node->Name() == "batch_norm" && + bn_vars_need_rename.find(in_node->Name()) != + bn_vars_need_rename.end()) { + // Create bn mean/variance for each repeat + var = result.CreateVarNode(&updated_var); + created[updated_var.Name()].push_back(var); + copied.insert(in_node); + repeated_node->inputs.push_back(var); + var->outputs.push_back(repeated_node); + continue; + } + + // for other ops + if (in_node->inputs.empty() && i > 0) { + // do not copy head vars (inputs, params) in repeats > 0 + var = created.at(in_node->Name()).back(); + } else { + if (copied.find(in_node) == copied.end()) { + var = result.CreateVarNode(&updated_var); + if (grad_names.find(in_node->Var()->Name()) != grad_names.end()) { + grad_repeated_map[in_node].push_back(var); + } + copied.insert(in_node); + created[updated_var.Name()].push_back(var); + } else { + var = created.at(updated_var.Name()).back(); + } + } + repeated_node->inputs.push_back(var); + var->outputs.push_back(repeated_node); + } + for (auto out_node : node->outputs) { + if (out_node->IsCtrlVar()) { + continue; + } + ir::Node* var = nullptr; + auto updated_var = UpdateGradVarDesc(out_node->Var(), i, grad_names, + bn_vars_need_rename); + if (copied.find(out_node) == copied.end()) { + var = result.CreateVarNode(&updated_var); + if (grad_names.find(out_node->Var()->Name()) != grad_names.end()) { + grad_repeated_map[out_node].push_back(var); + } + copied.insert(out_node); + created[updated_var.Name()].push_back(var); + } else { + var = created.at(updated_var.Name()).back(); + } + repeated_node->outputs.push_back(var); + var->inputs.push_back(repeated_node); + } + } + } + + // 5. create GRAD merge op node + for (auto kv : grad_repeated_map) { + OpDesc sum_op; + sum_op.SetType("sum"); + std::vector repeated_grad_names; + for (auto r : kv.second) { + repeated_grad_names.push_back(r->Var()->Name()); + } + sum_op.SetInput("X", repeated_grad_names); + sum_op.SetOutput("Out", {kv.first->Var()->Name()}); + sum_op.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(OpRole::kBackward)); + auto sum_op_node = result.CreateOpNode(&sum_op); + for (auto r : kv.second) { + sum_op_node->inputs.push_back(r); + r->outputs.push_back(sum_op_node); + } + auto sum_out_var_node = result.CreateVarNode(kv.first->Var()); + sum_op_node->outputs.push_back(sum_out_var_node); + sum_out_var_node->inputs.push_back(sum_op_node); + created[sum_out_var_node->Name()].push_back(sum_out_var_node); + + OpDesc scale_op; + scale_op.SetType("scale"); + scale_op.SetInput("X", {sum_out_var_node->Var()->Name()}); + // NOTE: inplace scale. + scale_op.SetOutput("Out", {sum_out_var_node->Var()->Name()}); + scale_op.SetAttr("scale", static_cast(1.0f / num_repeats)); + scale_op.SetAttr(OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(OpRole::kBackward)); + auto scale_op_node = result.CreateOpNode(&scale_op); + scale_op_node->inputs.push_back(sum_out_var_node); + sum_out_var_node->outputs.push_back(scale_op_node); + auto scale_out_var_node = result.CreateVarNode(sum_out_var_node->Var()); + scale_op_node->outputs.push_back(scale_out_var_node); + scale_out_var_node->inputs.push_back(scale_op_node); + created[scale_out_var_node->Name()].push_back(scale_out_var_node); + } + // 6. add optimize ops + { + auto copy_node = [&result, &created](ir::Node* node) { + auto op_node = result.CreateOpNode(node->Op()); + // copy op ins/outs + // NOTE: for send/recv ops, the OpDesc uses ctrldepvar to describe + // dependencies, so create those depvars if OpDesc have in/outs. + for (auto in_node : node->inputs) { + if (in_node->IsCtrlVar() && !in_node->Var()) { + continue; + } + ir::Node* var = nullptr; + if (created.find(in_node->Name()) == created.end()) { + var = result.CreateVarNode(in_node->Var()); + created[in_node->Name()].push_back(var); + } else { + var = created.at(in_node->Name()).back(); + } + op_node->inputs.push_back(var); + var->outputs.push_back(op_node); + } + for (auto out_node : node->outputs) { + if (out_node->IsCtrlVar() && !out_node->Var()) { + continue; + } + auto var = result.CreateVarNode(out_node->Var()); + created[out_node->Name()].push_back(var); + op_node->outputs.push_back(var); + var->inputs.push_back(op_node); + } + }; + for (auto node : lr_ops) { + copy_node(node); + } + for (auto node : optimize_ops) { + copy_node(node); + } + } + + result.ResolveHazard(created); + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(multi_batch_merge_pass, paddle::framework::ir::BatchMergePass) + .RequirePassAttr(paddle::framework::ir::kNumRepeats); diff --git a/paddle/fluid/framework/ir/multi_batch_merge_pass.h b/paddle/fluid/framework/ir/multi_batch_merge_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..c1e5aef20dbc60c18ed03038818bfd8ab217bf28 --- /dev/null +++ b/paddle/fluid/framework/ir/multi_batch_merge_pass.h @@ -0,0 +1,44 @@ +// 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 "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +// BatchMergePass is used to copy forward and backward ops for several +// times to run several batches to simulate large batch size training +// as if we have more than 1 GPUs. +// User can define how many batches to run, gradients will be merged +// through those repeats, and then do optimization using merged gradients. +// This pass is extremely useful when doing large batch-size distributed +// sync training, we can simulate even large batch size as if we have more +// GPUs. + +class BatchMergePass : public Pass { + public: + virtual ~BatchMergePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const override; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/node.h b/paddle/fluid/framework/ir/node.h index 5d6da9f1d76a3c0fc64b7ff35264e385cf19a14b..eedb375cf46165ebb09af56e5ab052a0327f1d0c 100644 --- a/paddle/fluid/framework/ir/node.h +++ b/paddle/fluid/framework/ir/node.h @@ -15,7 +15,10 @@ limitations under the License. */ #pragma once #include +#include +#include #include + #include "paddle/fluid/framework/op_desc.h" #include "paddle/fluid/framework/var_desc.h" #include "paddle/fluid/platform/macros.h" @@ -24,9 +27,33 @@ namespace paddle { namespace framework { namespace ir { -// Node should normally created by Graph::CreateXXXNode(). +// Node should only created by Graph::CreateXXXNode(). +// 1. Every Node should be part of a graph. No dangling Node exists. +// 2. Node only contains members necessary for building graph structure. +// It doesn't contain other unrelated members, such as device, etc. +// +// Sometimes, for specific usages, Node needs to have additional members, +// such as device_placement, version in order to be executed. It is suggested +// to use composition pattern. +// +// class RunnableOp { +// RunnableOp(ir::Node* n) : n_(n) { n_.WrappedBy(this); } +// +// int any_thing_; +// } +// +// RunnableOp is owned by the ir::Node that composes it. In other words. +// ir::Node will be responsible for deleting RunnableOp, say, when ir::Node +// is deleted from the graph. class Node { public: + virtual ~Node() { + if (!wrapper_.empty()) { + VLOG(4) << "ir::Node deleting a wrapper node " << Name(); + wrapper_deleter_(); + } + } + enum class Type { kOperation, kVariable }; static constexpr char kControlDepVarName[] = "__control_var"; @@ -44,6 +71,30 @@ class Node { return op_desc_.get(); } + // Set the `wrapper` that wraps the Node. `wrapper` is owned by Node. + template + void WrappedBy(T* wrapper) { + if (!wrapper_.empty()) { + wrapper_deleter_(); + } + wrapper_ = wrapper; + wrapper_deleter_ = [wrapper]() { delete wrapper; }; + wrapper_type_ = std::type_index(typeid(T)); + } + + // Return a reference to the `wrapper`. + template + T& Wrapper() { + return *boost::any_cast(wrapper_); + } + + // Test if the Node is wrapped by type T. + template + bool IsWrappedBy() { + return std::type_index(typeid(T)) == wrapper_type_; + } + + // Please don't use this API! int id() const { return id_; } bool IsOp() const { return type_ == Type::kOperation; } @@ -92,7 +143,13 @@ class Node { Node() = delete; static int count_; + // Please don't use this API or make this public. static void ResetId() { count_ = 0; } + + boost::any wrapper_; + std::function wrapper_deleter_; + std::type_index wrapper_type_ = std::type_index(typeid(void)); + DISABLE_COPY_AND_ASSIGN(Node); }; diff --git a/paddle/fluid/framework/ir/node_test.cc b/paddle/fluid/framework/ir/node_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..694efadda078169c993457181c00f7b357a09e87 --- /dev/null +++ b/paddle/fluid/framework/ir/node_test.cc @@ -0,0 +1,80 @@ +/* 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 +#include "gtest/gtest.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +class RunnableOp { + public: + RunnableOp(Node* node, bool* alive) : node_(node), alive_(alive) { + node_->WrappedBy(this); + } + + virtual ~RunnableOp() { *alive_ = false; } + + private: + Node* node_; + bool* alive_; +}; + +class RunnableOp2 { + public: + RunnableOp2(Node* node, bool* alive) : node_(node), alive_(alive) { + node_->WrappedBy(this); + } + + virtual ~RunnableOp2() { *alive_ = false; } + + private: + Node* node_; + bool* alive_; +}; + +TEST(NodeTest, Basic) { + bool alive1 = true; + bool alive2 = true; + std::unique_ptr n1(CreateNodeForTest("n1", Node::Type::kVariable)); + std::unique_ptr n2(CreateNodeForTest("n2", Node::Type::kVariable)); + + EXPECT_FALSE(n1->IsWrappedBy()); + EXPECT_FALSE(n1->IsWrappedBy()); + EXPECT_FALSE(n2->IsWrappedBy()); + EXPECT_FALSE(n2->IsWrappedBy()); + + new RunnableOp(n1.get(), &alive1); + new RunnableOp2(n2.get(), &alive2); + + EXPECT_TRUE(n1->IsWrappedBy()); + EXPECT_FALSE(n1->IsWrappedBy()); + EXPECT_FALSE(n2->IsWrappedBy()); + EXPECT_TRUE(n2->IsWrappedBy()); + + EXPECT_TRUE(alive1); + EXPECT_TRUE(alive2); + + n1.reset(nullptr); + n2.reset(nullptr); + EXPECT_FALSE(alive1); + EXPECT_FALSE(alive2); +} + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/pass.h b/paddle/fluid/framework/ir/pass.h index 9570c59cff2a6afeb1c607f7219b7b455974d6ce..8ac8d7677e1fe339d7802ea262e61a02a678aab5 100644 --- a/paddle/fluid/framework/ir/pass.h +++ b/paddle/fluid/framework/ir/pass.h @@ -76,7 +76,7 @@ class Pass { attr_name); attrs_[attr_name] = attr; attr_dels_[attr_name] = [attr, attr_name]() { - VLOG(3) << "deleting " << attr_name; + VLOG(30) << "deleting " << attr_name; delete attr; }; } diff --git a/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc b/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc index a7d5161c35db804703415066990f34da8109fbd9..b7687d61de3eacd47ff1208ba14c3f482215c1d4 100644 --- a/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc +++ b/paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.cc @@ -12,10 +12,13 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h" +#include +#include + #include "paddle/fluid/framework/ir/fuse_pass_base.h" #include "paddle/fluid/framework/ir/graph_pattern_detector.h" #include "paddle/fluid/framework/ir/graph_viz_pass.h" +#include "paddle/fluid/framework/ir/seq_concat_fc_fuse_pass.h" #include "paddle/fluid/framework/lod_tensor.h" namespace paddle { @@ -159,10 +162,7 @@ PDNode* BuildFCPattern(PDPattern* pattern, PDNode* fc_x) { std::set acts({"sigmoid", "tanh", "relu", "identity"}); PDNode* act = pattern->NewNode( - [=](Node* x) { - return x && x->IsOp() && acts.count(x->Op()->Type()); - - }, + [=](Node* x) { return x && x->IsOp() && acts.count(x->Op()->Type()); }, "act"); PDNode* fc_out = pattern->NewNode( @@ -196,7 +196,7 @@ std::unique_ptr SeqConcatFcFusePass::ApplyImpl( detector(graph.get(), [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) { - VLOG(4) << "get one concat pattern"; + VLOG(40) << "get one concat pattern"; // fc GET_NODE(fc_w, detector.pattern()); GET_NODE(fc_bias, detector.pattern()); diff --git a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc new file mode 100644 index 0000000000000000000000000000000000000000..015b5e3c6363cc96e31e21095fbbb007543c99af --- /dev/null +++ b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc @@ -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. + +#include "paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h" +#include +#include "paddle/fluid/framework/lod_tensor.h" + +namespace paddle { +namespace framework { +namespace ir { + +int BuildFusion(Graph* graph, const std::string& name_scope, Scope* scope) { + GraphPatternDetector gpd; + auto* pattern = gpd.mutable_pattern(); + + PDNode* x = pattern->NewNode(patterns::PDNodeName(name_scope, "X")) + ->assert_is_op_input("sequence_conv") + ->assert_var_not_persistable(); + patterns::SeqConvEltAddRelu fuse_pattern(pattern, name_scope); + fuse_pattern(x); + + // Create New OpDesc + auto fuse_creator = [&](Node* seqconv, Node* input, Node* seqconv_weight, + Node* eltadd_bias, Node* relu_out) { + OpDesc op_desc; + op_desc.SetType("fusion_seqconv_eltadd_relu"); + op_desc.SetInput("X", {input->Name()}); + op_desc.SetInput("Filter", {seqconv_weight->Name()}); + op_desc.SetInput("Bias", {eltadd_bias->Name()}); + op_desc.SetAttr("contextLength", seqconv->Op()->GetAttr("contextLength")); + op_desc.SetAttr("contextStart", seqconv->Op()->GetAttr("contextStart")); + op_desc.SetAttr("contextStride", seqconv->Op()->GetAttr("contextStride")); + PADDLE_ENFORCE(graph->Has(kParamScopeAttr)); + auto* scope = graph->Get(kParamScopeAttr); + const std::string ColMat = patterns::UniqueKey("SeqConvColMat"); + op_desc.SetOutput("ColMat", {ColMat}); + op_desc.SetOutput("Out", {relu_out->Name()}); + scope->Var(ColMat)->GetMutable(); + + auto* op = graph->CreateOpNode(&op_desc); + IR_NODE_LINK_TO(input, op); + IR_NODE_LINK_TO(seqconv_weight, op); + IR_NODE_LINK_TO(eltadd_bias, op); + IR_NODE_LINK_TO(op, relu_out); + return op; + }; + + int fusion_count{0}; + + auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, + Graph* g) { + VLOG(40) << "handle SeqConv EltAdd Relu fuse"; + GET_IR_NODE_FROM_SUBGRAPH(seqconv, seqconv, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(seqconv_weight, seqconv_weight, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(seqconv_out, seqconv_out, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(eltadd, eltadd, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(eltadd_bias, eltadd_bias, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(eltadd_out, eltadd_out, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(relu, relu, fuse_pattern); + GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, fuse_pattern); + + fuse_creator(seqconv, subgraph.at(x), seqconv_weight, eltadd_bias, + relu_out); + std::unordered_set marked_nodes( + {seqconv, seqconv_out, eltadd, eltadd_out, relu}); + GraphSafeRemoveNodes(graph, marked_nodes); + ++fusion_count; + }; + + gpd(graph, handler); + + return fusion_count; +} + +std::unique_ptr SeqConvEltAddReluFusePass::ApplyImpl( + std::unique_ptr graph) const { + FusePassBase::Init(name_scope_, graph.get()); + + int fusion_count = BuildFusion(graph.get(), name_scope_, param_scope()); + AddStatis(fusion_count); + + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(seqconv_eltadd_relu_fuse_pass, + paddle::framework::ir::SeqConvEltAddReluFusePass); diff --git a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h new file mode 100644 index 0000000000000000000000000000000000000000..dac9de71930c1768bdf416520caae6468449cd3d --- /dev/null +++ b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h @@ -0,0 +1,38 @@ +// 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 "paddle/fluid/framework/ir/fuse_pass_base.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detector.h" + +namespace paddle { +namespace framework { +namespace ir { + +class SeqConvEltAddReluFusePass : public FusePassBase { + public: + virtual ~SeqConvEltAddReluFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; + + const std::string name_scope_{"seqconv_eltadd_relu_fuse"}; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/lod_rank_table.cc b/paddle/fluid/framework/lod_rank_table.cc index 6bc795b642bf79b7556869c5ebe9b0323d3cc5fc..660ce2ec85131bafae27e8b7800fbfa3c238b59a 100644 --- a/paddle/fluid/framework/lod_rank_table.cc +++ b/paddle/fluid/framework/lod_rank_table.cc @@ -31,7 +31,7 @@ void LoDRankTable::Reset(const LoD& lod, size_t level) { TableItem item; item.index = i; item.length = vec[i + 1] - vec[i]; - VLOG(10) << "Add item to rank table " << item.index << " " << item.length; + VLOG(100) << "Add item to rank table " << item.index << " " << item.length; items_.emplace_back(item); } // NOTE(yuyang18): diff --git a/paddle/fluid/framework/lod_tensor.cc b/paddle/fluid/framework/lod_tensor.cc index 1e7da9a69c7cbf8c13306656599a759515802b76..669d08c70c9b7453264806b346a6c9eb211cfd4a 100644 --- a/paddle/fluid/framework/lod_tensor.cc +++ b/paddle/fluid/framework/lod_tensor.cc @@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor( PADDLE_ENFORCE_EQ(new_lod.size(), lod.size()); for (size_t j = 0; j < lod.size(); ++j) { auto &sub_lod = new_lod[j]; - auto &offset = sub_lod.back(); + size_t offset = sub_lod.back(); for (size_t k = 1; k < lod[j].size(); ++k) { sub_lod.push_back(lod[j][k] + offset); } diff --git a/paddle/fluid/framework/lod_tensor_array.h b/paddle/fluid/framework/lod_tensor_array.h index 6d7b6a4ada8729e3698dab5d2b1861aac632be79..36a5c3c5d601390beedaf37ceb98ee2c63ecf5a6 100644 --- a/paddle/fluid/framework/lod_tensor_array.h +++ b/paddle/fluid/framework/lod_tensor_array.h @@ -18,6 +18,8 @@ limitations under the License. */ namespace paddle { namespace framework { + using LoDTensorArray = std::vector; -} + +} // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/mixed_vector.h b/paddle/fluid/framework/mixed_vector.h index 77386f4f069489b6ff7b927a281bdc286ff816e0..e1aac6dc5a92fb616f00de5806f044b83c2f503f 100644 --- a/paddle/fluid/framework/mixed_vector.h +++ b/paddle/fluid/framework/mixed_vector.h @@ -542,6 +542,33 @@ class CPUVector : public std::vector> { this->reserve(this->size() + size_t(end - begin)); this->insert(this->end(), begin, end); } + + const T *CUDAData(platform::Place place) const { + PADDLE_THROW( + "Vector::CUDAData() method is not supported in CPU-only version"); + } + + T *CUDAMutableData(platform::Place place) { + PADDLE_THROW( + "Vector::CUDAMutableData() method is not supported in CPU-only " + "version"); + } + + const T *Data(platform::Place place) const { + PADDLE_ENFORCE( + platform::is_cpu_place(place), + "Vector::Data() method is not supported when not in CPUPlace"); + return this->data(); + } + + T *MutableData(platform::Place place) { + PADDLE_ENFORCE( + platform::is_cpu_place(place), + "Vector::MutableData() method is not supported when not in CPUPlace"); + return this->data(); + } + + const void *Handle() const { return static_cast(this); } }; template diff --git a/paddle/fluid/framework/mixed_vector_test.cc b/paddle/fluid/framework/mixed_vector_test.cc index 0599c8d384641606b0a5ebb5ba1781b56f539e63..0330cae377c32b2d49d409eff42b968d81356d49 100644 --- a/paddle/fluid/framework/mixed_vector_test.cc +++ b/paddle/fluid/framework/mixed_vector_test.cc @@ -51,7 +51,7 @@ TEST(mixed_vector, InitWithCount) { TEST(mixed_vector, ForEach) { vec tmp; for (auto& v : tmp) { - VLOG(3) << v; + VLOG(30) << v; } } diff --git a/paddle/fluid/framework/naive_executor.cc b/paddle/fluid/framework/naive_executor.cc index ba10687d65cfbbac89cfc76879c8b202ebd03229..8e660f97f051b194a0305dc82371fbe64da7e061 100644 --- a/paddle/fluid/framework/naive_executor.cc +++ b/paddle/fluid/framework/naive_executor.cc @@ -37,7 +37,7 @@ static void InitializeVariable(Variable *var, proto::VarType::Type var_type) { } else if (var_type == proto::VarType::FETCH_LIST) { var->GetMutable(); } else if (var_type == proto::VarType::STEP_SCOPES) { - var->GetMutable>(); + var->GetMutable>(); } else if (var_type == proto::VarType::LOD_RANK_TABLE) { var->GetMutable(); } else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) { @@ -71,7 +71,7 @@ void NaiveExecutor::Prepare(Scope *parent_scope, void NaiveExecutor::Run() { for (auto &op : ops_) { - VLOG(4) << "run " << op->Type(); + VLOG(40) << "run " << op->Type(); op->Run(*scope_, place_); } } @@ -95,21 +95,21 @@ void NaiveExecutor::CreateVariables(const ProgramDesc &desc, Scope *scope, if (var->Persistable()) { auto *ptr = const_cast(ancestor_scope)->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(3) << "Create Variable " << var->Name() - << " global, which pointer is " << ptr; + VLOG(30) << "Create Variable " << var->Name() + << " global, which pointer is " << ptr; } else { // Create temporary variables in local scope. auto *ptr = scope->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(3) << "Create Variable " << var->Name() - << " locally, which pointer is " << ptr; + VLOG(30) << "Create Variable " << var->Name() + << " locally, which pointer is " << ptr; } } } else { for (auto &var : global_block.AllVars()) { auto *ptr = scope->Var(var->Name()); InitializeVariable(ptr, var->GetType()); - VLOG(3) << "Create variable " << var->Name() << ", which pointer is " - << ptr; + VLOG(30) << "Create variable " << var->Name() << ", which pointer is " + << ptr; } } } @@ -146,22 +146,5 @@ void NaiveExecutor::CleanFeedFetchOps() { ops_.swap(ops); } -void NaiveExecutor::EnableMKLDNN(const ProgramDesc &program) { -#ifdef PADDLE_WITH_MKLDNN - VLOG(3) << "use_mkldnn=True"; - for (size_t block_id = 0; block_id < program.Size(); ++block_id) { - auto *block = const_cast(program).MutableBlock(block_id); - for (auto *op : block->AllOps()) { - if (op->HasAttr("use_mkldnn")) { - op->SetAttr("use_mkldnn", true); - } - } - } -#else - LOG(WARNING) - << "'MKLDNN' is not supported, Please re-compile with WITH_MKLDNN option"; -#endif -} - } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/naive_executor.h b/paddle/fluid/framework/naive_executor.h index 9374f3f4a35cc0f90e5b2d6e8b397784b8eae123..ddfa6e1f4d8b73f594fc381ab505797491cdd378 100644 --- a/paddle/fluid/framework/naive_executor.h +++ b/paddle/fluid/framework/naive_executor.h @@ -48,8 +48,6 @@ class NaiveExecutor { void CleanFeedFetchOps(); - void EnableMKLDNN(const ProgramDesc& program); - protected: void CreateVariables(const ProgramDesc& desc, Scope* scope, int block_id); diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index b29ac44699463312a1fdcea55e003daa75997302..fbaa169df6324761ef9136aa173dce4e2182ed38 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -82,13 +82,9 @@ class CompileTimeInferShapeContext : public InferShapeContext { auto *in_var = block_.FindVarRecursive(Inputs(in)[i]); auto *out_var = block_.FindVarRecursive(Outputs(out)[j]); if (in_var->GetType() != proto::VarType::LOD_TENSOR) { - VLOG(3) << "input " << in << " is not LodTensor"; + VLOG(30) << "input " << in << " is not LodTensor"; return; } - PADDLE_ENFORCE_EQ(in_var->GetType(), proto::VarType::LOD_TENSOR, - "The %d-th output of Output(%s) must be LoDTensor.", j, - out); - out_var->SetLoDLevel(in_var->GetLoDLevel()); } @@ -245,32 +241,32 @@ void OpDesc::SetAttr(const std::string &name, const Attribute &v) { const proto::OpProto::Attr &attr = GetProtoAttr(name); switch (attr.type()) { case proto::AttrType::BOOLEANS: { - VLOG(11) << "SetAttr: " << Type() << ", " << name - << " from INTS to BOOLEANS"; + VLOG(110) << "SetAttr: " << Type() << ", " << name + << " from INTS to BOOLEANS"; this->attrs_[name] = std::vector(); break; } case proto::AttrType::INTS: { - VLOG(11) << "SetAttr: " << Type() << ", " << name - << " from INTS to INTS"; + VLOG(110) << "SetAttr: " << Type() << ", " << name + << " from INTS to INTS"; this->attrs_[name] = std::vector(); break; } case proto::AttrType::FLOATS: { - VLOG(11) << "SetAttr: " << Type() << ", " << name - << " from INTS to FLOATS"; + VLOG(110) << "SetAttr: " << Type() << ", " << name + << " from INTS to FLOATS"; this->attrs_[name] = std::vector(); break; } case proto::AttrType::STRINGS: { - VLOG(11) << "SetAttr: " << Type() << ", " << name - << " from INTS to STRINGS"; + VLOG(110) << "SetAttr: " << Type() << ", " << name + << " from INTS to STRINGS"; this->attrs_[name] = std::vector(); break; } case proto::AttrType::BLOCKS: { - VLOG(11) << "SetAttr: " << Type() << ", " << name - << " from INTS to BLOCKS"; + VLOG(110) << "SetAttr: " << Type() << ", " << name + << " from INTS to BLOCKS"; this->SetBlocksAttr(name, std::vector()); return; } @@ -423,8 +419,15 @@ struct SetAttrDescVisitor : public boost::static_visitor { } VectorToRepeated(blocks_idx, attr_->mutable_blocks_idx()); } + void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); } + void operator()(int64_t v) const { attr_->set_l(v); } + + void operator()(const std::vector &v) const { + VectorToRepeated(v, attr_->mutable_longs()); + } + void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); } }; @@ -496,13 +499,13 @@ void OpDesc::CheckAttrs() { } void OpDesc::InferShape(const BlockDesc &block) const { - VLOG(3) << "CompileTime infer shape on " << Type(); + VLOG(30) << "CompileTime infer shape on " << Type(); InitInferShapeFuncs(); auto &infer_shape = OpInfoMap::Instance().Get(this->Type()).infer_shape_; PADDLE_ENFORCE(static_cast(infer_shape), "%s's infer_shape has not been registered", this->Type()); CompileTimeInferShapeContext ctx(*this, block); - if (VLOG_IS_ON(10)) { + if (VLOG_IS_ON(100)) { std::ostringstream sout; auto inames = this->InputArgumentNames(); sout << " From ["; @@ -513,26 +516,20 @@ void OpDesc::InferShape(const BlockDesc &block) const { std::copy(onames.begin(), onames.end(), std::ostream_iterator(sout, ", ")); sout << "]"; - VLOG(10) << sout.str(); + VLOG(100) << sout.str(); } infer_shape(&ctx); } void OpDesc::InferVarType(BlockDesc *block) const { + // There are a few places that var type can be set. + // When VarDesc is created, default set to LOD_TENSOR. + // When output variable is created, default is defaut set to LOD_TENSOR. + // We limit here to be the only place that operator defines its customized + // var type inference. Hence, we don't do any "default" setting here. auto &info = OpInfoMap::Instance().Get(this->Type()); if (info.infer_var_type_) { info.infer_var_type_(*this, block); - } else { - // all output type is LoDTensor by default - VLOG(10) << this->Type() - << " has not registered InferVarType. Set output variables to " - "LOD_TENSOR"; - for (auto &out_pair : this->outputs_) { - for (auto &out_var_name : out_pair.second) { - block->FindRecursiveOrCreateVar(out_var_name) - .SetType(proto::VarType::LOD_TENSOR); - } - } } } @@ -610,7 +607,7 @@ DDim CompileTimeInferShapeContext::GetDim(const std::string &name) const { auto shape = var->GetShape(); res = shape.empty() ? make_ddim({0UL}) : make_ddim(shape); } catch (...) { - VLOG(5) << "GetDim of variable " << name << " error"; + VLOG(50) << "GetDim of variable " << name << " error"; std::rethrow_exception(std::current_exception()); } return res; @@ -627,7 +624,7 @@ std::vector CompileTimeInferShapeContext::GetRepeatedDims( res.push_back(s.empty() ? make_ddim({0UL}) : make_ddim(s)); } } catch (...) { - VLOG(5) << "GetRepeatedDim of variable " << name << " error."; + VLOG(50) << "GetRepeatedDim of variable " << name << " error."; std::rethrow_exception(std::current_exception()); } return res; diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index b4205aba83e774fb9c08193124adb93935c00157..30c8a26c3d2f0068674aa70b4ff875a2f73c1dca 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -100,16 +100,6 @@ class OpDesc { std::vector InputNames() const { return MapKeys(inputs_); } std::vector OutputNames() const { return MapKeys(outputs_); } - void SetInputMap(const VariableNameMap &input) { - this->inputs_ = input; - this->need_update_ = true; - } - - void SetOutputMap(const VariableNameMap &output) { - this->outputs_ = output; - this->need_update_ = true; - } - const VariableNameMap &Inputs() const { return inputs_; } const VariableNameMap &Outputs() const { return outputs_; } @@ -131,10 +121,6 @@ class OpDesc { BlockDesc *Block() { return this->block_; } - const BlockDesc &BlockRef() const { return *this->block_; } - - void SetBlock(BlockDesc *block) { this->block_ = block; } - private: template static std::vector MapKeys(const MapType &map) { diff --git a/paddle/fluid/framework/op_proto_maker.cc b/paddle/fluid/framework/op_proto_maker.cc index df2a7a27ca4a6011b214202ac9bf4f30dc482ece..ca31303f77c4a30eb64c43404e214779ea78aeaf 100644 --- a/paddle/fluid/framework/op_proto_maker.cc +++ b/paddle/fluid/framework/op_proto_maker.cc @@ -21,7 +21,6 @@ namespace framework { void OpProtoAndCheckerMaker::Validate() { validated_ = true; CheckNoDuplicatedInOutAttrs(); - CheckReuseVars(); } OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput( @@ -40,40 +39,6 @@ OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput( return OpProtoAndCheckerMaker::VariableBuilder{output}; } -void OpProtoAndCheckerMaker::Reuse(const std::string& name, - const std::string& reused_name) { - bool found = false; - proto::OpProto::Var* var; - - for (auto& var : proto_->inputs()) { - if (var.name() == reused_name) { - found = true; - break; - } - } - PADDLE_ENFORCE(found == true, - "Input/Output name: %s reused_name: %s, one of them is not " - "exists or not matched.", - name, reused_name); - - found = false; - for (int i = 0; i < proto_->outputs().size(); ++i) { - var = proto_->mutable_outputs()->Mutable(i); - if (var->name() == name) { - PADDLE_ENFORCE(!var->has_reuse(), - "Output(%s) has been set reused var of %s", name, - var->reuse()); - found = true; - var->set_reuse(reused_name); - break; - } - } - PADDLE_ENFORCE(found == true, - "Input/Output name: %s reused_name: %s, one of them is not " - "exists or not matched.", - name, reused_name); -} - void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { std::unordered_set names; auto checker = [&](const std::string& name) { @@ -91,24 +56,6 @@ void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() { } } -void OpProtoAndCheckerMaker::CheckReuseVars() { - std::unordered_set names; - for (auto& input : proto_->inputs()) { - names.insert(input.name()); - } - auto checker = [&](const std::string& name, const std::string& reused) { - PADDLE_ENFORCE( - names.count(reused), - "Output [%s] reuse Input [%s], but the input is not registered.", name, - reused); - }; - for (auto& output : proto_->outputs()) { - if (output.has_reuse()) { - checker(output.name(), output.reuse()); - } - } -} - void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto, OpAttrChecker* attr_checker) { proto_ = proto; @@ -124,6 +71,8 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto, static_cast(OpRole::kLoss) | static_cast(OpRole::kForward), static_cast(OpRole::kLoss) | static_cast(OpRole::kBackward), + static_cast(OpRole::kOptimize) | + static_cast(OpRole::kLRSched), static_cast(OpRole::kNotSpecified)}) .SetDefault(static_cast(OpRole::kNotSpecified)); AddAttr>(OpRoleVarAttrName(), diff --git a/paddle/fluid/framework/op_proto_maker.h b/paddle/fluid/framework/op_proto_maker.h index 4ed3cc45d66849267ef4945a03da1db76b53e4ea..4c59c73d8779eceb267ad532aabccabbd54b0df2 100644 --- a/paddle/fluid/framework/op_proto_maker.h +++ b/paddle/fluid/framework/op_proto_maker.h @@ -14,25 +14,26 @@ limitations under the License. */ #pragma once #include -#include - #include "glog/logging.h" #include "paddle/fluid/framework/attribute.h" #include "paddle/fluid/framework/framework.pb.h" namespace paddle { namespace framework { +////////////////////////// +// Don't add more roles to make this too complicated! +////////////////////////// enum class OpRole { kForward = 0x0000, kBackward = 0x0001, kOptimize = 0x0002, // RPC role is for send/recv releated op - kRPC = 0x0003, + kRPC = 0x0004, // Dist role is for split_byref/split_selected_rows/concat // used for distributed training. - kDist = 0x0004, + kDist = 0x0008, // Tag all learning rate scheduler operators. - kLRSched = 0x0005, + kLRSched = 0x0010, kLoss = 0x0100, // The default value of op's role. This should be only used for unittests and @@ -73,11 +74,6 @@ class OpProtoAndCheckerMaker { var_->set_dispensable(true); return *this; } - - VariableBuilder &Reuse(const std::string &name) { - var_->set_reuse(name); - return *this; - } }; VariableBuilder AddInput(const std::string &name, const std::string &comment); @@ -85,8 +81,6 @@ class OpProtoAndCheckerMaker { VariableBuilder AddOutput(const std::string &name, const std::string &comment); - void Reuse(const std::string &name, const std::string &reused_name); - template TypedAttrChecker &AddAttr(const std::string &name, const std::string &comment, @@ -105,8 +99,6 @@ class OpProtoAndCheckerMaker { void CheckNoDuplicatedInOutAttrs(); void Validate(); - void CheckReuseVars(); - proto::OpProto *proto_; OpAttrChecker *op_checker_; bool validated_{false}; diff --git a/paddle/fluid/framework/op_proto_maker_test.cc b/paddle/fluid/framework/op_proto_maker_test.cc index b71c7b646857e11f291748c4c7c2af92b6d53231..a8030d377fdb4d4aef74b315e21792dad10fac96 100644 --- a/paddle/fluid/framework/op_proto_maker_test.cc +++ b/paddle/fluid/framework/op_proto_maker_test.cc @@ -47,120 +47,3 @@ TEST(ProtoMaker, DuplicatedInOut) { ASSERT_THROW(proto_maker(&op_proto, &op_checker), paddle::platform::EnforceNotMet); } - -class TestInplaceProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { - public: - void Make() { - AddInput("X", "input of test op"); - AddOutput("XOut", "output of test op").Reuse("X"); - } -}; - -class TestInplaceProtoMaker2 - : public paddle::framework::OpProtoAndCheckerMaker { - public: - void Make() { - AddInput("X", "input of test op"); - AddOutput("XOut", "output of test op").Reuse("X"); - AddOutput("NoOut", "output of test op").Reuse("NotExists"); - } -}; - -TEST(ProtoMaker, InplaceOutput) { - paddle::framework::proto::OpProto op_proto, op_proto2; - paddle::framework::OpAttrChecker op_checker; - TestInplaceProtoMaker proto_maker; - TestInplaceProtoMaker2 proto_maker2; - - proto_maker(&op_proto, &op_checker); - - ASSERT_THROW(proto_maker2(&op_proto2, &op_checker), - paddle::platform::EnforceNotMet); -} - -// normal reuse -class TestReuseProtoMaker : public paddle::framework::OpProtoAndCheckerMaker { - public: - void Make() { - AddInput("X", "input of test op"); - AddInput("Y", "input of test op"); - AddOutput("Out", "output of test op"); - AddOutput("XOut", "output of test op"); - // avoid destructor exception. - // Validate(); - TestReuse(); - } - - virtual void TestReuse() {} -}; - -// test duplicate reuse error -class TestReuseProtoMaker2 : public TestReuseProtoMaker { - public: - void TestReuse() { - Reuse("Out", "X"); - Reuse("Out", "Y"); - } -}; - -// NotExists Input -class TestReuseProtoMaker3 : public TestReuseProtoMaker { - public: - void TestReuse() { - Reuse("Out", "NotExists"); - Reuse("XOut", "X"); - } -}; - -// NotExists Output -class TestReuseProtoMaker4 : public TestReuseProtoMaker { - public: - void TestReuse() { Reuse("NotExists", "X"); } -}; - -TEST(ProtoMaker, Reuse) { - paddle::framework::proto::OpProto op_proto; - paddle::framework::OpAttrChecker op_checker; - TestReuseProtoMaker proto_maker; - proto_maker(&op_proto, &op_checker); -} - -// NOTE(dzhwinter): -// There is a Fatal CHECK on base class destructor, which will call abort inside -// instead of -// throw an exception. If we throw an exception in Make(), we will trigger the -// CHECK and terminate the tests. -// -// I had tried to replace the default CHECK with a exception, however, it's -// still not supported by glog. -// the details: -// https://github.com/google/glog/issues/249 -// https://github.com/facebookresearch/TensorComprehensions/issues/351 -/* -TEST(ProtoMaker, ReuseWithException) { - paddle::framework::proto::OpProto op_proto2, op_proto3, op_proto4; - paddle::framework::OpAttrChecker op_checker; - TestReuseProtoMaker2 proto_maker2; - TestReuseProtoMaker3 proto_maker3; - TestReuseProtoMaker4 proto_maker4; - EXPECT_THROW(proto_maker2(&op_proto2, &op_checker), - paddle::platform::EnforceNotMet); - - EXPECT_THROW(proto_maker3(&op_proto3, &op_checker), - paddle::platform::EnforceNotMet); - - EXPECT_THROW(proto_maker4(&op_proto4, &op_checker), - paddle::platform::EnforceNotMet); -} - -void FailureFunction() { - throw std::runtime_error("Check failed in destructor."); - // return 0; -} - -int main(int argc, char** argv) { - testing::InitGoogleTest(&argc, argv); - google::InstallFailureFunction(&FailureFunction); - return RUN_ALL_TESTS(); -} -*/ diff --git a/paddle/fluid/framework/op_registry.cc b/paddle/fluid/framework/op_registry.cc index bfc411ca2c4a483e344b368da089392d8e4a87c1..4a841bae8323f5733ba413a2c623a8147ec32f67 100644 --- a/paddle/fluid/framework/op_registry.cc +++ b/paddle/fluid/framework/op_registry.cc @@ -46,9 +46,9 @@ static VariableNameMap ConvertOpDescVarsToVarNameMap( std::unique_ptr OpRegistry::CreateOp( const proto::OpDesc& op_desc) { - VLOG(1) << "CreateOp directly from OpDesc is deprecated. It should only be" - "used in unit tests. Use CreateOp(const OpDesc& op_desc) " - "instead."; + VLOG(10) << "CreateOp directly from OpDesc is deprecated. It should only be" + "used in unit tests. Use CreateOp(const OpDesc& op_desc) " + "instead."; VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs()); VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs()); AttributeMap attrs; diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index 9f930065324f13f5aa79c214e820fb6fc2f3a166..5624878d439873e5f6aee6ec9234e31d5c77ff97 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -140,7 +140,7 @@ static LoD GetLoD(const Scope& scope, const std::string& name) { } void OperatorBase::Run(const Scope& scope, const platform::Place& place) { - VLOG(4) << place << " " << DebugStringEx(&scope); + VLOG(40) << place << " " << DebugStringEx(&scope); if (platform::is_gpu_place(place)) { #ifndef PADDLE_WITH_CUDA PADDLE_THROW("Cannot run operator on place %s", place); @@ -149,10 +149,18 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) { platform::SetDeviceId(dev_id); #endif } - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - platform::RecordEvent record_event(Type(), pool.Get(place)); - RunImpl(scope, place); - VLOG(3) << place << " " << DebugStringEx(&scope); + + // The profile has a process-wide mutex, results in serious performance issue + // 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)); + RunImpl(scope, place); + } else { + RunImpl(scope, place); + } + VLOG(30) << place << " " << DebugStringEx(&scope); } bool OperatorBase::HasInputs(const std::string& name) const { @@ -251,6 +259,8 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const { if (row_size >= 0) { ss << "[row_size=" << row_size << "]"; } + std::string dtype = GetDtype(*scope, output.second[i]); + ss << ":" << dtype; ss << "[" << GetDims(*scope, var_name, true) << "]"; ss << "(" << GetLoD(*scope, var_name) << ")"; } @@ -346,22 +356,22 @@ void OperatorBase::GenerateTemporaryNames() { } } -static bool VarIsTensor(const Variable* var) { - return var->IsType() || var->IsType(); +static bool VarIsTensor(const Variable& var) { + return var.IsType() || var.IsType(); } -static const Tensor* GetTensorFromVar(Variable* var) { - if (var->IsType()) { - return var->GetMutable(); - } else if (var->IsType()) { - return var->GetMutable()->mutable_value(); +const Tensor* GetLoDTensorOrSelectedRowsValueFromVar(const Variable& var) { + if (var.IsType()) { + return static_cast(&(var.Get())); + } else if (var.IsType()) { + return &(var.Get().value()); } else { PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.", - var->Type().name()); + var.Type().name()); } } -static Tensor* GetMutableTensorFromVar(Variable* var) { +Tensor* GetMutableLoDTensorOrSelectedRowsValueFromVar(Variable* var) { if (var->IsType()) { return var->GetMutable(); } else if (var->IsType()) { @@ -406,9 +416,7 @@ bool ExecutionContext::HasOutput(const std::string& name) const { template <> const Tensor* ExecutionContext::Input(const std::string& name) const { - auto* var = InputVar(name); - return var == nullptr ? nullptr - : GetTensorFromVar(const_cast(var)); + return Input(name); } template <> @@ -418,17 +426,21 @@ const std::vector ExecutionContext::MultiInput( std::vector res; res.reserve(names.size()); std::transform(names.begin(), names.end(), std::back_inserter(res), - [&](const std::string& sub_name) { + [&](const std::string& sub_name) -> const Tensor* { auto var = scope_.FindVar(sub_name); - return var == nullptr ? nullptr : GetTensorFromVar(var); + if (var == nullptr) return nullptr; + PADDLE_ENFORCE( + var->IsType(), + "%s should be LoDTensor, but the received type is %s", + sub_name, var->Type().name()); + return &(var->Get()); }); return res; } template <> Tensor* ExecutionContext::Output(const std::string& name) const { - auto var = OutputVar(name); - return var == nullptr ? nullptr : GetMutableTensorFromVar(var); + return Output(name); } template <> @@ -438,10 +450,14 @@ std::vector ExecutionContext::MultiOutput( std::vector res; res.reserve(names.size()); std::transform(names.begin(), names.end(), std::back_inserter(res), - [&](const std::string& sub_name) { + [&](const std::string& sub_name) -> Tensor* { auto var = scope_.FindVar(sub_name); - return var == nullptr ? nullptr - : GetMutableTensorFromVar(var); + if (var == nullptr) return nullptr; + PADDLE_ENFORCE( + var->IsType(), + "%s should be LoDTensor, but the received type is %s", + sub_name, var->Type().name()); + return var->GetMutable(); }); return res; } @@ -701,14 +717,14 @@ void OperatorWithKernel::RunImpl(const Scope& scope, auto expected_kernel_key = this->GetExpectedKernelType(ExecutionContext(*this, scope, *dev_ctx)); - VLOG(3) << "expected_kernel_key:" << expected_kernel_key; + VLOG(30) << "expected_kernel_key:" << expected_kernel_key; auto kernel_iter = kernels.find(expected_kernel_key); #ifdef PADDLE_WITH_MKLDNN // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set if (kernel_iter == kernels.end() && expected_kernel_key.library_type_ == LibraryType::kMKLDNN) { - VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one"; + VLOG(30) << "missing MKLDNN kernel: fallbacking to PLAIN one"; expected_kernel_key.library_type_ = LibraryType::kPlain; expected_kernel_key.data_layout_ = DataLayout::kAnyLayout; kernel_iter = kernels.find(expected_kernel_key); @@ -760,10 +776,14 @@ void OperatorWithKernel::TransferInplaceVarsBack( const Scope& scope, const std::vector& inplace_vars, const Scope& transfer_scope) const { for (auto& var_name : inplace_vars) { - VLOG(3) << "share inplace var " + var_name + " back to it's original scope"; - auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name)); - auto* transformed_tensor = - GetTensorFromVar(transfer_scope.FindVar(var_name)); + VLOG(30) << "share inplace var " + var_name + + " back to it's original scope"; + auto* original_tensor = + GetMutableLoDTensorOrSelectedRowsValueFromVar(scope.FindVar(var_name)); + auto* var = transfer_scope.FindVar(var_name); + PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr", + var_name); + auto* transformed_tensor = GetLoDTensorOrSelectedRowsValueFromVar(*var); original_tensor->ShareDataWith(*transformed_tensor); } } @@ -776,11 +796,11 @@ Scope* OperatorWithKernel::TryTransferData( for (auto& var_name : var_name_item.second) { auto* var = scope.FindVar(var_name); // Only tensor can be tranfer to another device. - if (var == nullptr || !VarIsTensor(var)) { + if (var == nullptr || !VarIsTensor(*var)) { continue; } - auto* tensor_in = GetTensorFromVar(var); + auto* tensor_in = GetLoDTensorOrSelectedRowsValueFromVar(*var); if (!tensor_in->IsInitialized()) { continue; } @@ -798,8 +818,8 @@ Scope* OperatorWithKernel::TryTransferData( transfered_inplace_vars->emplace_back(var_name); } - VLOG(3) << "Transform Variable " << var_name << " from " - << kernel_type_for_var << " to " << expected_kernel_key; + VLOG(30) << "Transform Variable " << var_name << " from " + << kernel_type_for_var << " to " << expected_kernel_key; if (new_scope == nullptr) { new_scope = &scope.NewScope(); diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h index 626b50edfd39424473be33e9f8baec5970471477..40b0130b265471a1288d966c4cbcd4f0e1bdb9f1 100644 --- a/paddle/fluid/framework/operator.h +++ b/paddle/fluid/framework/operator.h @@ -54,6 +54,9 @@ constexpr char kGradVarSuffix[] = "@GRAD"; /// Variables with this suffix are supposed to be filled up with zeros. constexpr char kZeroVarSuffix[] = "@ZERO"; +/// Variables with this suffix are the new Gradient. +constexpr char kNewGradSuffix[] = "@NEWGRAD@"; + // define some kernel priority /* Define multiple kernel type fallback order*/ extern std::vector> kKernelPriority; @@ -63,6 +66,8 @@ inline std::string GradVarName(const std::string& var_name) { } proto::VarType::Type GetDataTypeOfVar(const Variable* var); +const Tensor* GetLoDTensorOrSelectedRowsValueFromVar(const Variable& var); +Tensor* GetMutableLoDTensorOrSelectedRowsValueFromVar(Variable* var); class OperatorBase; class ExecutionContext; @@ -223,7 +228,7 @@ class ExecutionContext { std::vector res; res.reserve(names.size()); std::transform(names.begin(), names.end(), std::back_inserter(res), - [&](const std::string& sub_name) { + [&](const std::string& sub_name) -> const T* { auto var = scope_.FindVar(sub_name); return var == nullptr ? nullptr : &var->Get(); }); @@ -236,7 +241,7 @@ class ExecutionContext { std::vector res; res.reserve(names.size()); std::transform(names.begin(), names.end(), std::back_inserter(res), - [&](const std::string& sub_name) { + [&](const std::string& sub_name) -> T* { auto var = scope_.FindVar(sub_name); return var == nullptr ? nullptr : var->GetMutable(); }); diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index f06bad6c78c05804e583f859906b88fb7b500372..39b47415ff7e378cabc79e668fe2be63eb71d87f 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -38,9 +38,20 @@ class ParallelExecutorPrivate { explicit ParallelExecutorPrivate(const std::vector &places) : places_(places) {} + ~ParallelExecutorPrivate() { + if (own_local_scope_) { + for (size_t i = 1; i < local_scopes_.size(); ++i) { + // Skip the first scope, since it is the global scope. + Scope *local_scope = local_scopes_[i]; + if (global_scope_->HasKid(local_scope)) { + global_scope_->DeleteScope(local_scope); + } + } + } + } std::vector places_; std::vector local_scopes_; - Scope *global_scope_; + Scope *global_scope_; // not owned std::unique_ptr executor_; #ifdef PADDLE_WITH_CUDA @@ -109,18 +120,9 @@ ParallelExecutor::ParallelExecutor( if (member_->local_scopes_.size() != 1 && local_scopes.empty()) { BCastParamsToDevices(bcast_vars); } - // Startup Program has been run. All local scopes has correct parameters. - - // Step 2. Create vars in each scope; - std::vector var_infos; - for (auto *var : main_program.Block(0).AllVars()) { - var_infos.emplace_back(); - var_infos.back().name_ = var->Name(); - var_infos.back().type_ = var->GetType(); - var_infos.back().persistable_ = var->Persistable(); - } +// Startup Program has been run. All local scopes has correct parameters. -// Step 3. Convert main_program to SSA form and dependency graph. Also, insert +// Step 2. Convert main_program to SSA form and dependency graph. Also, insert // ncclOp #ifdef PADDLE_WITH_CUDA std::unique_ptr graph = build_strategy.Apply( @@ -156,13 +158,22 @@ ParallelExecutor::ParallelExecutor( params, member_->local_scopes_, member_->use_cuda_); #endif - if (VLOG_IS_ON(5)) { - // If the loss_var_name is given, the number of graph should be only one. - if (loss_var_name.size()) { - PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1, - "The number of graph should be only one"); + // 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 &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()) { + PADDLE_ENFORCE_EQ(ir::GraphNum(*graph), 1, + "The number of graph should be only one"); + } if (exec_strategy.type_ == ExecutionStrategy::kDefault) { member_->executor_.reset(new details::ThreadedSSAGraphExecutor( @@ -187,6 +198,10 @@ void ParallelExecutor::BCastParamsToDevices( } auto &main_tensor = main_var->Get(); + if (!main_tensor.IsInitialized()) { + VLOG(30) << "one in var not inited, return!"; + continue; + } auto &dims = main_tensor.dims(); if (paddle::platform::is_gpu_place(main_tensor.place())) { #ifdef PADDLE_WITH_CUDA @@ -299,14 +314,12 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes( } ParallelExecutor::~ParallelExecutor() { - if (member_->own_local_scope_) { - for (size_t i = 1; i < member_->local_scopes_.size(); ++i) { - Scope *local_scope = member_->local_scopes_[i]; - if (member_->global_scope_->HasKid(local_scope)) { - member_->global_scope_->DeleteScope(local_scope); - } - } + for (auto &p : member_->places_) { + platform::DeviceContextPool::Instance().Get(p)->Wait(); } + // member_ must be destructed before gcs_ since the destructor of + // ReferenceCountOpHandle use raw pointers of gcs_ inside. + member_.reset(); } } // namespace framework diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h index fd386a5987f11ff64964e95eb7e9b83572dc790c..ef09b98b2aa91a9d729b94d15dbb676dde4092b6 100644 --- a/paddle/fluid/framework/parallel_executor.h +++ b/paddle/fluid/framework/parallel_executor.h @@ -75,7 +75,7 @@ class ParallelExecutor { private: void BCastParamsToDevices(const std::unordered_set &vars) const; - ParallelExecutorPrivate *member_; + std::unique_ptr member_; #ifdef PADDLE_WITH_CUDA // ref_cnts_ is only initialized when ParallelExecutor constructs, and then diff --git a/paddle/fluid/framework/program_desc.cc b/paddle/fluid/framework/program_desc.cc index 589905828f7793c614c0fe12259e9ba5ab11ceac..4b9667113bc7918c1323f0213213a6ffdb7eed8e 100644 --- a/paddle/fluid/framework/program_desc.cc +++ b/paddle/fluid/framework/program_desc.cc @@ -126,7 +126,7 @@ const std::vector ProgramDesc::GetFeedTargetNames() { std::vector feed_target_names; for (auto *op : global_block.AllOps()) { if (op->Type() == kFeedOpType) { - int col = boost::get(op->GetAttr("col")); + size_t col = boost::get(op->GetAttr("col")); if (col >= feed_target_names.size()) { feed_target_names.resize(col + 1); } @@ -143,7 +143,7 @@ const std::vector ProgramDesc::GetFetchTargetNames() { std::vector fetch_target_names; for (auto *op : global_block.AllOps()) { if (op->Type() == kFetchOpType) { - int col = boost::get(op->GetAttr("col")); + size_t col = boost::get(op->GetAttr("col")); if (col >= fetch_target_names.size()) { fetch_target_names.resize(col + 1); } diff --git a/paddle/fluid/framework/program_desc_test.cc b/paddle/fluid/framework/program_desc_test.cc index 7e689a37da8a16bd9b1ac6650b9322d2eb5a2c85..48bde2785e6a51afc0d2905ac31fe20a3c3019b6 100644 --- a/paddle/fluid/framework/program_desc_test.cc +++ b/paddle/fluid/framework/program_desc_test.cc @@ -103,7 +103,7 @@ TEST(ProgramDesc, copy_ctor) { ASSERT_EQ(1, op->GetBlockAttrId("sub_block")); found_sub_block = true; - ASSERT_EQ(2, op->GetBlocksAttrIds("sub_blocks").size()); + ASSERT_EQ(2UL, op->GetBlocksAttrIds("sub_blocks").size()); found_sub_blocks = true; } } diff --git a/paddle/fluid/framework/reader_test.cc b/paddle/fluid/framework/reader_test.cc index f0d07cb7c1367576084b9494e7758103bb45d1e5..d812417a38200bcfdbdeac78800190647510a144 100644 --- a/paddle/fluid/framework/reader_test.cc +++ b/paddle/fluid/framework/reader_test.cc @@ -39,8 +39,8 @@ TEST(READER, decorate_chain) { { auto endpoints = root->GetEndPoints(); ASSERT_EQ(endpoints.size(), 2U); - ASSERT_NE(endpoints.count(end_point1.get()), 0); - ASSERT_NE(endpoints.count(end_point2.get()), 0); + ASSERT_NE(endpoints.count(end_point1.get()), 0UL); + ASSERT_NE(endpoints.count(end_point2.get()), 0UL); } { diff --git a/paddle/fluid/framework/scope.cc b/paddle/fluid/framework/scope.cc index 1a727a2c8c759d010606d5b605823b7252b35c69..0c407f8c1d11a8a0f99551fc51d2ef2be5262c63 100644 --- a/paddle/fluid/framework/scope.cc +++ b/paddle/fluid/framework/scope.cc @@ -49,18 +49,18 @@ int64_t GetEagerDeletionThreshold() { Scope::~Scope() { DropKids(); } Scope& Scope::NewScope() const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); kids_.push_back(new Scope(this)); return *kids_.back(); } Variable* Scope::Var(const std::string& name) { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); return VarInternal(name); } Variable* Scope::Var(std::string* name) { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); auto new_name = string::Sprintf("%p.%d", this, vars_.size()); if (name != nullptr) { *name = new_name; @@ -69,29 +69,34 @@ Variable* Scope::Var(std::string* name) { } Variable* Scope::FindVar(const std::string& name) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); return FindVarInternal(name); } +Variable* Scope::FindLocalVar(const std::string& name) const { + std::lock_guard lock(mutex_); + return FindVarLocally(name); +} + const Scope* Scope::FindScope(const Variable* var) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); return FindScopeInternal(var); } void Scope::DropKids() { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); for (Scope* s : kids_) delete s; kids_.clear(); } bool Scope::HasKid(const Scope* scope) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); auto it = std::find(this->kids_.begin(), this->kids_.end(), scope); return it != this->kids_.end(); } std::vector Scope::LocalVarNames() const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); std::vector known_vars; known_vars.reserve(this->vars_.size()); for (auto& p : vars_) { @@ -101,7 +106,7 @@ std::vector Scope::LocalVarNames() const { } void Scope::DeleteScope(Scope* scope) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); auto it = std::find(this->kids_.begin(), this->kids_.end(), scope); PADDLE_ENFORCE(it != this->kids_.end(), "Cannot find %p as kid scope", scope); this->kids_.erase(it); @@ -114,7 +119,7 @@ void Scope::DeleteScope(Scope* scope) const { } void Scope::EraseVars(const std::vector& var_names) { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); std::set var_set(var_names.begin(), var_names.end()); for (auto it = vars_.begin(); it != vars_.end();) { if (var_set.find(it->first) != var_set.end()) { @@ -127,12 +132,12 @@ void Scope::EraseVars(const std::vector& var_names) { void Scope::Rename(const std::string& origin_name, const std::string& new_name) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); RenameInternal(origin_name, new_name); } std::string Scope::Rename(const std::string& origin_name) const { - std::unique_lock lock(mutex_); + std::lock_guard lock(mutex_); auto new_name = string::Sprintf("%p.%d", this, vars_.size()); RenameInternal(origin_name, new_name); return new_name; @@ -144,7 +149,7 @@ Variable* Scope::VarInternal(const std::string& name) { v = new Variable(); vars_[name].reset(v); - VLOG(3) << "Create variable " << name; + VLOG(30) << "Create variable " << name; v->name_ = &(vars_.find(name)->first); return v; } diff --git a/paddle/fluid/framework/scope.h b/paddle/fluid/framework/scope.h index e42fff1d79d92fb7ed61768a614d8cd98f6775a0..9462620e829ec815e1553f6378a67463ea3b8aa3 100644 --- a/paddle/fluid/framework/scope.h +++ b/paddle/fluid/framework/scope.h @@ -63,6 +63,11 @@ class Scope { /// Caller doesn't own the returned Variable. Variable* FindVar(const std::string& name) const; + /// Find a variable in the current scope. + /// Return nullptr if cannot find. + /// Caller doesn't own the returned Variable. + Variable* FindLocalVar(const std::string& name) const; + const Scope* parent() const { return parent_; } /// Find the scope or an ancestor scope that contains the given variable. @@ -73,6 +78,8 @@ class Scope { /// Drop all kids scopes belonged to this scope. void DropKids(); + std::list& kids() const { return kids_; } + /// Find if a scope exists in the kid scopes bool HasKid(const Scope* scope) const; diff --git a/paddle/fluid/framework/selected_rows.cc b/paddle/fluid/framework/selected_rows.cc index 8c290bb095d554a973e66a3a19606a06759fd668..3319c772ec789bc5b28307906adfb2a9417d9182 100644 --- a/paddle/fluid/framework/selected_rows.cc +++ b/paddle/fluid/framework/selected_rows.cc @@ -176,7 +176,7 @@ void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value, PADDLE_ENFORCE(value->IsInitialized(), "The value tensor should be initialized."); if (ids.numel() == 0) { - VLOG(3) << "keys is empty, please check data!"; + VLOG(30) << "keys is empty, please check data!"; } else { int64_t value_width = value_->numel() / value_->dims()[0]; PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0], diff --git a/paddle/fluid/framework/selected_rows_test.cc b/paddle/fluid/framework/selected_rows_test.cc index 928e1ad8b9168e61ddc5782066a4aa29a4296a94..9c427a4ae4c9660b107ca891a60db306cb09301f 100644 --- a/paddle/fluid/framework/selected_rows_test.cc +++ b/paddle/fluid/framework/selected_rows_test.cc @@ -91,7 +91,7 @@ TEST(SelectedRows, SparseTable) { ASSERT_TRUE(table.HasKey(10)); ASSERT_TRUE(table.HasKey(8)); ASSERT_TRUE(table.HasKey(6)); - ASSERT_EQ(table.rows().size(), 3); + ASSERT_EQ(table.rows().size(), 3UL); framework::Tensor ids; ids.Resize(framework::make_ddim({4})); diff --git a/paddle/fluid/framework/tensor_test.cc b/paddle/fluid/framework/tensor_test.cc index cb2061c06a429d8e8116001a4aa4e8c46ea13428..a0a9a573603ceb6b577529101cb331adbc81337a 100644 --- a/paddle/fluid/framework/tensor_test.cc +++ b/paddle/fluid/framework/tensor_test.cc @@ -75,6 +75,19 @@ TEST(Tensor, MutableData) { platform::CPUPlace()); EXPECT_EQ(p1, p2); } + // Not sure if it's desired, but currently, Tensor type can be changed. + { + framework::Tensor src_tensor; + int8_t* p1 = src_tensor.mutable_data(framework::make_ddim({1}), + platform::CPUPlace()); + EXPECT_NE(p1, nullptr); + *p1 = 1; + + uint8_t* p2 = src_tensor.mutable_data(framework::make_ddim({1}), + platform::CPUPlace()); + EXPECT_NE(p2, nullptr); + EXPECT_EQ(static_cast(p2[0]), 1); + } #ifdef PADDLE_WITH_CUDA { diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc index 1d7a2eb5b38255531880fe3d2e5321024caf0c6b..8d8f07a1f52b3062498b59a4dbc20219d42e4735 100644 --- a/paddle/fluid/framework/tensor_util.cc +++ b/paddle/fluid/framework/tensor_util.cc @@ -22,8 +22,8 @@ namespace framework { void TensorCopy(const Tensor& src, const platform::Place& dst_place, const platform::DeviceContext& ctx, Tensor* dst) { - VLOG(3) << "TensorCopy " << src.dims() << " from " << src.place() << " to " - << dst_place; + VLOG(30) << "TensorCopy " << src.dims() << " from " << src.place() << " to " + << dst_place; src.check_memory_size(); dst->Resize(src.dims()); @@ -36,6 +36,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto size = src.numel() * SizeOfType(src.type()); if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) { + if (src_ptr == dst_ptr) { + VLOG(30) << "Skip copy the same data async from " << src_place << " to " + << dst_place; + return; + } memory::Copy(boost::get(dst_place), dst_ptr, boost::get(src_place), src_ptr, size); } @@ -71,6 +76,11 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, auto stream = reinterpret_cast(ctx).stream(); if (platform::is_same_place(src_place, dst_place)) { + if (src_ptr == dst_ptr) { + VLOG(30) << "Skip copy the same data async from " << src_place << " to " + << dst_place; + return; + } memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream); } else { @@ -104,8 +114,8 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place, void TensorCopySync(const Tensor& src, const platform::Place& dst_place, Tensor* dst) { - VLOG(3) << "TensorCopySync " << src.dims() << " from " << src.place() - << " to " << dst_place; + VLOG(30) << "TensorCopySync " << src.dims() << " from " << src.place() + << " to " << dst_place; src.check_memory_size(); dst->Resize(src.dims()); dst->set_layout(src.layout()); @@ -114,6 +124,11 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, auto dst_ptr = dst->mutable_data(dst_place, src.type()); auto size = src.numel() * SizeOfType(src.type()); if (platform::is_cpu_place(src_place) && platform::is_cpu_place(dst_place)) { + if (src_ptr == dst_ptr) { + VLOG(30) << "Skip copy the same data from " << src_place << " to " + << dst_place; + return; + } memory::Copy(boost::get(dst_place), dst_ptr, boost::get(src_place), src_ptr, size); } @@ -130,9 +145,20 @@ void TensorCopySync(const Tensor& src, const platform::Place& dst_place, memory::Copy(dst_gpu_place, dst_ptr, src_cpu_place, src_ptr, size, nullptr); } else if (platform::is_gpu_place(src_place) && platform::is_gpu_place(dst_place)) { + if (src_ptr == dst_ptr && platform::is_same_place(src_place, dst_place)) { + VLOG(30) << "Skip copy the same data from " << src_place << " to " + << dst_place; + return; + } auto src_gpu_place = boost::get(src_place); auto dst_gpu_place = boost::get(dst_place); memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, nullptr); + } else if (platform::is_cuda_pinned_place(src_place) && + platform::is_gpu_place(dst_place)) { + auto src_pinned_place = boost::get(src_place); + auto dst_gpu_place = boost::get(dst_place); + memory::Copy(dst_gpu_place, dst_ptr, src_pinned_place, src_ptr, size, + nullptr); } #endif } diff --git a/paddle/fluid/framework/tensor_util_test.cc b/paddle/fluid/framework/tensor_util_test.cc index a1e5b967a86d10f3439db662af54bb82888027b9..793ccfc79fe56707f226477b9d50b1d972ab6a59 100644 --- a/paddle/fluid/framework/tensor_util_test.cc +++ b/paddle/fluid/framework/tensor_util_test.cc @@ -41,6 +41,11 @@ TEST(TensorCopy, Tensor) { EXPECT_EQ(src_ptr[i], dst_ptr[i]); } + TensorCopy(dst_tensor, *cpu_place, &dst_tensor); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], dst_ptr[i]); + } + EXPECT_TRUE(dst_tensor.layout() == src_tensor.layout()); Tensor slice_tensor = src_tensor.Slice(1, 2); @@ -82,6 +87,15 @@ TEST(TensorCopy, Tensor) { EXPECT_EQ(src_ptr[i], dst_ptr[i]); } + // Copy the same tensor + TensorCopy(gpu_tensor, *gpu_place, gpu_ctx, &gpu_tensor); + gpu_ctx.Wait(); + const int* dst_ptr_tmp = dst_tensor.data(); + EXPECT_NE(src_ptr, dst_ptr_tmp); + for (size_t i = 0; i < 9; ++i) { + EXPECT_EQ(src_ptr[i], dst_ptr_tmp[i]); + } + Tensor slice_tensor = src_tensor.Slice(1, 2); // CPU Slice Tensor to GPU Tensor diff --git a/paddle/fluid/framework/threadpool.cc b/paddle/fluid/framework/threadpool.cc index 18cdca3a658a6a89e6ab637a7f38825756acfea8..2dab4e793eeacd65239786976948b8043aeeb215 100644 --- a/paddle/fluid/framework/threadpool.cc +++ b/paddle/fluid/framework/threadpool.cc @@ -25,7 +25,6 @@ DEFINE_int32(dist_threadpool_size, 0, namespace paddle { namespace framework { - std::unique_ptr ThreadPool::threadpool_(nullptr); std::once_flag ThreadPool::init_flag_; @@ -40,15 +39,14 @@ void ThreadPool::Init() { int num_threads = std::thread::hardware_concurrency(); if (FLAGS_dist_threadpool_size > 0) { num_threads = FLAGS_dist_threadpool_size; - VLOG(1) << "set dist_threadpool_size to " << num_threads; + VLOG(10) << "set dist_threadpool_size to " << num_threads; } PADDLE_ENFORCE_GT(num_threads, 0); threadpool_.reset(new ThreadPool(num_threads)); } } -ThreadPool::ThreadPool(int num_threads) - : total_threads_(num_threads), idle_threads_(num_threads), running_(true) { +ThreadPool::ThreadPool(int num_threads) : running_(true) { threads_.resize(num_threads); for (auto& thread : threads_) { // TODO(Yancey1989): binding the thread on the specify CPU number @@ -59,9 +57,10 @@ ThreadPool::ThreadPool(int num_threads) ThreadPool::~ThreadPool() { { // notify all threads to stop running + std::unique_lock l(mutex_); running_ = false; - scheduled_.notify_all(); } + scheduled_.notify_all(); for (auto& t : threads_) { t->join(); @@ -69,36 +68,30 @@ ThreadPool::~ThreadPool() { } } -void ThreadPool::Wait() { - std::unique_lock lock(mutex_); - completed_.wait(lock, [=] { return Done() == true; }); -} - void ThreadPool::TaskLoop() { - while (running_) { - std::unique_lock lock(mutex_); - scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; }); - - if (!running_) { - break; - } - // pop a task from the task queue - auto task = std::move(tasks_.front()); - tasks_.pop(); - - --idle_threads_; - lock.unlock(); - - // run the task - task(); + while (true) { + Task task; { std::unique_lock lock(mutex_); - ++idle_threads_; - if (Done()) { - completed_.notify_all(); + scheduled_.wait( + lock, [this] { return !this->tasks_.empty() || !this->running_; }); + + if (!running_ && tasks_.empty()) { + return; + } + + if (tasks_.empty()) { + PADDLE_THROW("This thread has no task to Run"); } + + // pop a task from the task queue + task = std::move(tasks_.front()); + tasks_.pop(); } + + // run the task + task(); } } diff --git a/paddle/fluid/framework/threadpool.h b/paddle/fluid/framework/threadpool.h index 94111ee335b1a5df327b3e46d62069b4735c54f6..7a51d18fbbf65f68725aa86a6a0ce4d15dff5673 100644 --- a/paddle/fluid/framework/threadpool.h +++ b/paddle/fluid/framework/threadpool.h @@ -57,17 +57,8 @@ class ThreadPool { ~ThreadPool(); - // Returns the number of threads created by the constructor. - size_t Threads() const { return total_threads_; } - - // Returns the number of currently idle threads. - size_t IdleThreads() { - std::unique_lock lock(mutex_); - return idle_threads_; - } - // Run pushes a function to the task queue and returns a std::future - // object. To wait for the completion of the task, call + // object. To wait for the completion of the task, call // std::future::wait(). template std::future Run(Callback fn) { @@ -78,7 +69,6 @@ class ThreadPool { template std::future> RunAndGetException( Callback fn) { - std::unique_lock lock(mutex_); Task task([fn]() -> std::unique_ptr { try { fn(); @@ -93,26 +83,20 @@ class ThreadPool { return nullptr; }); std::future> f = task.get_future(); - tasks_.push(std::move(task)); - lock.unlock(); + { + std::unique_lock lock(mutex_); + if (!running_) { + PADDLE_THROW("enqueue on stopped ThreadPool"); + } + tasks_.push(std::move(task)); + } scheduled_.notify_one(); return f; } - // Wait until all the tasks are completed. - void Wait(); - private: DISABLE_COPY_AND_ASSIGN(ThreadPool); - // If the task queue is empty and avaialbe is equal to the number of - // threads, means that all tasks are completed. Note: this function - // is not thread-safe. Returns true if all tasks are completed. - // Note: don't delete the data member total_threads_ and use - // threads_.size() instead; because you'd need to lock the mutex - // before accessing threads_. - bool Done() { return tasks_.empty() && idle_threads_ == total_threads_; } - // The constructor starts threads to run TaskLoop, which retrieves // and runs tasks from the queue. void TaskLoop(); @@ -125,14 +109,11 @@ class ThreadPool { static std::once_flag init_flag_; std::vector> threads_; - const size_t total_threads_; - size_t idle_threads_; std::queue tasks_; std::mutex mutex_; bool running_; std::condition_variable scheduled_; - std::condition_variable completed_; }; class ThreadPoolIO : ThreadPool { diff --git a/paddle/fluid/framework/threadpool_test.cc b/paddle/fluid/framework/threadpool_test.cc index 27a4ffd4fcbf293a3dea1744b29384d0bee0c137..884d61e23428a0ad758946295ca9c470767e93ef 100644 --- a/paddle/fluid/framework/threadpool_test.cc +++ b/paddle/fluid/framework/threadpool_test.cc @@ -19,10 +19,11 @@ limitations under the License. */ namespace framework = paddle::framework; -void do_sum(framework::ThreadPool* pool, std::atomic* sum, int cnt) { - std::vector> fs; +void do_sum(std::vector>* fs, std::mutex* mu, + std::atomic* sum, int cnt) { for (int i = 0; i < cnt; ++i) { - fs.push_back(framework::Async([sum]() { sum->fetch_add(1); })); + std::lock_guard l(*mu); + fs->push_back(framework::Async([sum]() { sum->fetch_add(1); })); } } @@ -40,18 +41,21 @@ TEST(ThreadPool, ConcurrentInit) { } TEST(ThreadPool, ConcurrentRun) { - framework::ThreadPool* pool = framework::ThreadPool::GetInstance(); std::atomic sum(0); std::vector threads; + std::vector> fs; + std::mutex fs_mu; int n = 50; // sum = (n * (n + 1)) / 2 for (int i = 1; i <= n; ++i) { - std::thread t(do_sum, pool, &sum, i); + std::thread t(do_sum, &fs, &fs_mu, &sum, i); threads.push_back(std::move(t)); } for (auto& t : threads) { t.join(); } - pool->Wait(); + for (auto& t : fs) { + t.wait(); + } EXPECT_EQ(sum, ((n + 1) * n) / 2); } diff --git a/paddle/fluid/framework/type_defs.h b/paddle/fluid/framework/type_defs.h index e099e40f121ff13657e563eb608feecbca0551be..2de6233a9e0d320ec9a06d547db3575eb61925c0 100644 --- a/paddle/fluid/framework/type_defs.h +++ b/paddle/fluid/framework/type_defs.h @@ -36,7 +36,7 @@ using Attribute = boost::variant, std::vector, std::vector, bool, std::vector, BlockDesc*, int64_t, - std::vector>; + std::vector, std::vector>; using AttributeMap = std::unordered_map; diff --git a/paddle/fluid/framework/var_desc.cc b/paddle/fluid/framework/var_desc.cc index 7e3f002b53351ba5892aaa50482b21a83db94069..29ef459b454075a30c3a4d0ff0f9ef1212292b4b 100644 --- a/paddle/fluid/framework/var_desc.cc +++ b/paddle/fluid/framework/var_desc.cc @@ -61,10 +61,10 @@ size_t VarDesc::GetTensorDescNum() const { void VarDesc::SetShapes( const std::vector> &multiple_dims) { if (multiple_dims.size() != GetTensorDescNum()) { - VLOG(3) << "WARNING: The number of given shapes(" << multiple_dims.size() - << ") doesn't match the existing tensor number(" - << GetTensorDescNum() - << "). The Reader is going to be reinitialized."; + VLOG(30) << "WARNING: The number of given shapes(" << multiple_dims.size() + << ") doesn't match the existing tensor number(" + << GetTensorDescNum() + << "). The Reader is going to be reinitialized."; SetTensorDescNum(multiple_dims.size()); } std::vector tensors = mutable_tensor_descs(); @@ -94,11 +94,11 @@ void VarDesc::SetDataType(proto::VarType::Type data_type) { void VarDesc::SetDataTypes( const std::vector &multiple_data_type) { if (multiple_data_type.size() != GetTensorDescNum()) { - VLOG(3) << "WARNING: The number of given data types(" - << multiple_data_type.size() - << ") doesn't match the existing tensor number(" - << GetTensorDescNum() - << "). The Reader is going to be reinitialized."; + VLOG(30) << "WARNING: The number of given data types(" + << multiple_data_type.size() + << ") doesn't match the existing tensor number(" + << GetTensorDescNum() + << "). The Reader is going to be reinitialized."; SetTensorDescNum(multiple_data_type.size()); } std::vector tensor_descs = @@ -139,11 +139,11 @@ void VarDesc::SetLoDLevel(int32_t lod_level) { void VarDesc::SetLoDLevels(const std::vector &multiple_lod_level) { if (multiple_lod_level.size() != GetTensorDescNum()) { - VLOG(3) << "WARNING: The number of given lod_levels(" - << multiple_lod_level.size() - << ") doesn't match the existing tensor number(" - << GetTensorDescNum() - << "). The Reader is going to be reinitialized."; + VLOG(30) << "WARNING: The number of given lod_levels(" + << multiple_lod_level.size() + << ") doesn't match the existing tensor number(" + << GetTensorDescNum() + << "). The Reader is going to be reinitialized."; SetTensorDescNum(multiple_lod_level.size()); } switch (desc_.type().type()) { diff --git a/paddle/fluid/framework/var_desc.h b/paddle/fluid/framework/var_desc.h index e33849ef502fb10b913e7e28cbd0abdb8b8ff9bb..9d3fb811191c207c75845ef8f8486e8beac7525a 100644 --- a/paddle/fluid/framework/var_desc.h +++ b/paddle/fluid/framework/var_desc.h @@ -59,6 +59,7 @@ class VarDesc { public: explicit VarDesc(const std::string &name) { desc_.set_name(name); + // TODO(paddle-dev): Why default to lodtensor. desc_.mutable_type()->set_type(proto::VarType::LOD_TENSOR); } diff --git a/paddle/fluid/framework/var_type_inference.h b/paddle/fluid/framework/var_type_inference.h index f3035cd712bdea517068b4c172bb2794d5fccddb..64236b78d2e390ea5f6c43c76a4b33b62c67629f 100644 --- a/paddle/fluid/framework/var_type_inference.h +++ b/paddle/fluid/framework/var_type_inference.h @@ -13,6 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include +#include "paddle/fluid/framework/block_desc.h" +#include "paddle/fluid/framework/op_desc.h" #include "paddle/fluid/framework/type_defs.h" namespace paddle { @@ -24,5 +27,27 @@ class VarTypeInference { virtual void operator()(const OpDesc& op_desc, BlockDesc* block) const = 0; }; +class PassInDtypeAndVarTypeToOutput : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const final { + auto in_out_var_names = this->GetInputOutputWithSameType(); + + for (auto& i_o_n : in_out_var_names) { + auto& x_name = op_desc.Input(i_o_n.first).at(0); + auto& out_name = op_desc.Output(i_o_n.second).at(0); + + auto& x = block->FindRecursiveOrCreateVar(x_name); + auto& out = block->FindRecursiveOrCreateVar(out_name); + out.SetType(x.GetType()); + out.SetDataType(x.GetDataType()); + } + } + + protected: + virtual std::unordered_map + GetInputOutputWithSameType() const = 0; +}; + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/variable.h b/paddle/fluid/framework/variable.h index 067e0c2b8389f88639fd9b95bd680702517efee1..873e1b20a584df3ba90cf5c1a62a3879bf98ce5c 100644 --- a/paddle/fluid/framework/variable.h +++ b/paddle/fluid/framework/variable.h @@ -38,8 +38,12 @@ class Variable { template T* GetMutable() { - if (!IsType()) { + if (!holder_) { holder_.reset(new PlaceholderImpl(new T())); + } else { + PADDLE_ENFORCE(IsType(), + "Variable must be type %s, the holding type is %s", + typeid(T).name(), holder_->Type().name()); } return static_cast(holder_->Ptr()); } diff --git a/paddle/fluid/framework/variable_test.cc b/paddle/fluid/framework/variable_test.cc index c5c1d215f4a6affae0a3bdafacec40a2aee2ca19..003dcfd3dfe5ecfd563a686bb72b061aff602f73 100644 --- a/paddle/fluid/framework/variable_test.cc +++ b/paddle/fluid/framework/variable_test.cc @@ -33,9 +33,10 @@ TEST(Variable, GetMutable) { const Tensor& tt = v->Get(); EXPECT_EQ(1234, tt.content_); - std::string* s = v->GetMutable(); - *s = "hello"; - - const std::string& ss = v->Get(); - EXPECT_EQ("hello", ss); + try { + v->GetMutable(); + } catch (std::exception& e) { + return; + } + EXPECT_TRUE(false); } diff --git a/paddle/fluid/inference/CMakeLists.txt b/paddle/fluid/inference/CMakeLists.txt index ec1bc7825dd21628f5c37ea44a154abe7b7e8c73..e5678cf607a8ff3763e79c1f321a81c021846fb1 100644 --- a/paddle/fluid/inference/CMakeLists.txt +++ b/paddle/fluid/inference/CMakeLists.txt @@ -1,3 +1,6 @@ +if(WITH_TESTING) + include(tests/test.cmake) # some generic cmake funtion for inference +endif() # analysis and tensorrt must be added before creating static library, # otherwise, there would be undefined reference to them in static library. add_subdirectory(analysis) @@ -19,9 +22,19 @@ cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api) add_subdirectory(api) +set(STATIC_INFERENCE_APIS paddle_fluid_api paddle_inference_api analysis_predictor) +set(SHARED_INFERENCE_SRCS + io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc + ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc + ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc) +if (WITH_GPU AND TENSORRT_FOUND) + set(STATIC_INFERENCE_APIS ${STATIC_INFERENCE_APIS} paddle_inference_tensorrt_subgraph_engine) + set(SHARED_INFERENCE_SRCS ${SHARED_INFERENCE_SRCS} ${CMAKE_CURRENT_SOURCE_DIR}/api/api_tensorrt_subgraph_engine.cc) +endif() + # Create static library -cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api - analysis_predictor zero_copy_tensor) +cc_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array) + if(NOT APPLE) # TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac. set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym") @@ -29,11 +42,8 @@ if(NOT APPLE) endif() # Create shared library -cc_library(paddle_fluid_shared SHARED - SRCS io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc - ${CMAKE_CURRENT_SOURCE_DIR}/api/analysis_predictor.cc - ${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc - DEPS ${fluid_modules} paddle_fluid_api) +cc_library(paddle_fluid_shared SHARED SRCS ${SHARED_INFERENCE_SRCS} + DEPS ${fluid_modules} paddle_fluid_api reset_tensor_array) set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid) if(NOT APPLE) diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt index d4d2fd4634f9e11f3f002e11e177c332ced49885..0354f9e6e9588af601210b8a71ae98c1f90d62f0 100644 --- a/paddle/fluid/inference/analysis/CMakeLists.txt +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -20,22 +20,17 @@ cc_test(test_node SRCS node_tester.cc DEPS analysis) cc_test(test_dot SRCS dot_tester.cc DEPS analysis) cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid) -function (inference_analysis_test TARGET) - if(WITH_TESTING) - set(options "") - set(oneValueArgs "") - set(multiValueArgs SRCS ARGS EXTRA_DEPS) - cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - set(mem_opt "") - if(WITH_GPU) - set(mem_opt "--fraction_of_gpu_memory_to_use=0.5") - endif() - cc_test(${TARGET} - SRCS "${analysis_test_SRCS}" - DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS} - ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt} ${analysis_test_ARGS}) - set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec) - endif(WITH_TESTING) +function(inference_analysis_test TARGET) + if(WITH_TESTING) + set(options "") + set(oneValueArgs "") + set(multiValueArgs SRCS ARGS EXTRA_DEPS) + cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + inference_base_test(${TARGET} + SRCS ${analysis_test_SRCS} + DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS} + ARGS --inference_model_dir=${WORD2VEC_MODEL_DIR} ${analysis_test_ARGS}) + endif() endfunction(inference_analysis_test) inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api) diff --git a/paddle/fluid/inference/analysis/analyzer.cc b/paddle/fluid/inference/analysis/analyzer.cc index 8a8aeb5e09a0d9a6746f6d6d61c547363e0e2d30..d55303a51e9fee3057455470b4b3139dc5f85e89 100644 --- a/paddle/fluid/inference/analysis/analyzer.cc +++ b/paddle/fluid/inference/analysis/analyzer.cc @@ -60,7 +60,7 @@ class DfgPassManagerImpl final : public DfgPassManager { private: void AddPass(const std::string& name, AnalysisPass* pass) { - VLOG(3) << "Adding pass " << name; + VLOG(30) << "Adding pass " << name; Register(name, pass); AddGraphvizDebugerPass(pass); } @@ -70,7 +70,7 @@ class DfgPassManagerImpl final : public DfgPassManager { auto trt_teller = [&](const Node* node) { std::unordered_set teller_set( {"mul", "conv2d", "pool2d", "relu", "softmax", "sigmoid", - "depthwise_conv2d", "batch_norm", "concat", "tanh", + "depthwise_conv2d", "batch_norm", "concat", "tanh", "pad", "elementwise_add", "dropout"}); if (!node->IsFunction()) return false; @@ -101,13 +101,25 @@ Analyzer::Analyzer() { Register("manager1", new DfgPassManagerImpl); } void Analyzer::Run(Argument* argument) { std::vector passes; - for (auto& pass : all_ir_passes_) { - if (!disabled_ir_passes_.count(pass)) { + passes.push_back("graph_viz_pass"); // add graphviz for debug. +#ifdef PADDLE_WITH_MKLDNN + if (use_mkldnn_) { + VLOG(30) << "Adding MKL-DNN placement pass"; + passes.push_back("mkldnn_placement_pass"); + } +#endif + // infer_clean_graph_pass should be the first default pass + // after mkldnn_placement_pass. + passes.push_back("infer_clean_graph_pass"); + passes.push_back("graph_viz_pass"); // add graphviz for debug. + for (auto& pass : ir_passes_) { + // skip mkldnn pass when use_mkldnn_ = false; + bool skip_pass = (!use_mkldnn_) && pass.find("mkldnn") != std::string::npos; + if (!disabled_ir_passes_.count(pass) && !skip_pass) { passes.push_back(pass); passes.push_back("graph_viz_pass"); // add graphviz for debug. } } - passes.push_back("graph_viz_pass"); argument->Set(kFluidToIrPassesAttr, new std::vector(passes)); for (auto& x : data_) { @@ -117,11 +129,26 @@ void Analyzer::Run(Argument* argument) { } } +Analyzer& Analyzer::IncludeAllIrPasses() { + ir_passes_ = all_ir_passes_; + return *this; +} + Analyzer& Analyzer::DisableIrPasses(const std::vector& passes) { disabled_ir_passes_.insert(passes.begin(), passes.end()); return *this; } +Analyzer& Analyzer::IncludeIrPasses(const std::vector& passes) { + ir_passes_ = passes; + return *this; +} + +Analyzer& Analyzer::SetUseMkldnn(bool use_mkldnn) { + use_mkldnn_ = use_mkldnn; + return *this; +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/analyzer.h b/paddle/fluid/inference/analysis/analyzer.h index 0aa9367bf5692e53e2a1f1247523cf9a4f0b3a1d..3af1d572dfd81197797dd7e57d87ba12c2f3548e 100644 --- a/paddle/fluid/inference/analysis/analyzer.h +++ b/paddle/fluid/inference/analysis/analyzer.h @@ -54,6 +54,9 @@ class Analyzer : public OrderedRegistry { void Run(Argument* argument); Analyzer& DisableIrPasses(const std::vector& passes); + Analyzer& IncludeIrPasses(const std::vector& passes); + Analyzer& IncludeAllIrPasses(); + Analyzer& SetUseMkldnn(bool use_mkldnn); DISABLE_COPY_AND_ASSIGN(Analyzer); @@ -64,21 +67,29 @@ class Analyzer : public OrderedRegistry { // larger fusion. const std::vector all_ir_passes_{{ // Manual update the passes here. - "infer_clean_graph_pass", // - "attention_lstm_fuse_pass", // - "embedding_fc_lstm_fuse_pass", // - "fc_lstm_fuse_pass", // - "mul_lstm_fuse_pass", // - "fc_gru_fuse_pass", // - "mul_gru_fuse_pass", // - "seq_concat_fc_fuse_pass", // - "fc_fuse_pass", // + "attention_lstm_fuse_pass", // + "seqconv_eltadd_relu_fuse_pass", // + "embedding_fc_lstm_fuse_pass", // + "fc_lstm_fuse_pass", // + "mul_lstm_fuse_pass", // + "fc_gru_fuse_pass", // + "mul_gru_fuse_pass", // + "seq_concat_fc_fuse_pass", // + "fc_fuse_pass", // + "conv_bn_fuse_pass", // + "conv_eltwiseadd_bn_fuse_pass", // #ifdef PADDLE_WITH_MKLDNN - "conv_relu_mkldnn_fuse_pass", // + "depthwise_conv_mkldnn_pass", // + "conv_bias_mkldnn_fuse_pass", // + "conv_relu_mkldnn_fuse_pass", // + "conv_elementwise_add_mkldnn_fuse_pass", // #endif }}; std::unordered_set disabled_ir_passes_; + // Ir passes to run + std::vector ir_passes_; + bool use_mkldnn_; }; } // namespace analysis diff --git a/paddle/fluid/inference/analysis/analyzer_tester.cc b/paddle/fluid/inference/analysis/analyzer_tester.cc index f90910ac0d0a897ef01d4ca2bd0bca575baf4c40..5430e5c1ef1c70d27295ebc1a9bd427cd95f006a 100644 --- a/paddle/fluid/inference/analysis/analyzer_tester.cc +++ b/paddle/fluid/inference/analysis/analyzer_tester.cc @@ -51,9 +51,7 @@ void TestWord2vecPrediction(const std::string& model_path) { config.model_dir = model_path; config.use_gpu = false; config.device = 0; - auto predictor = - ::paddle::CreatePaddlePredictor( - config); + auto predictor = ::paddle::CreatePaddlePredictor(config); // One single batch diff --git a/paddle/fluid/inference/analysis/argument.h b/paddle/fluid/inference/analysis/argument.h index e8fb0775b45761f64fd6fd28306c35b76d1e40c4..9495e2435c79ff660c64322d2acd8e058e09e563 100644 --- a/paddle/fluid/inference/analysis/argument.h +++ b/paddle/fluid/inference/analysis/argument.h @@ -68,8 +68,8 @@ struct Argument { key); attrs_[key] = data; attr_deleters_[key] = [data, key]() { - VLOG(3) << "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"; - VLOG(3) << "argument delete attr: " << key; + VLOG(30) << "xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"; + VLOG(30) << "argument delete attr: " << key; delete data; }; } diff --git a/paddle/fluid/inference/analysis/data_flow_graph.cc b/paddle/fluid/inference/analysis/data_flow_graph.cc index 8c7d58678fd29cb25d13d64a08e6c6f26f242d8b..545017da07f8e414f21a02c35fca96aba6de41aa 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph.cc @@ -112,8 +112,8 @@ void DataFlowGraph::Build(const framework::proto::ProgramDesc &prog) { out_alias->SetPbMsg(out->pb_msg()); var2id[out_alias->name()] = out_alias->id(); // update variable's alias Node - LOG(INFO) << "loop found in graph, create SSA alias node [" - << out_alias->repr() << "] for [" << out->repr() << "]"; + VLOG(40) << "loop found in graph, create SSA alias node [" + << out_alias->repr() << "] for [" << out->repr() << "]"; out = out_alias; } out->inlinks.push_back(o); @@ -132,7 +132,7 @@ void DataFlowGraph::Build(const framework::ir::Graph &graph) { Node *x{nullptr}; if (ir_node->IsOp()) { PADDLE_ENFORCE(ir_node->Op()); - VLOG(4) << "get op " << ir_node << " " << ir_node->Name(); + VLOG(40) << "get op " << ir_node << " " << ir_node->Name(); x = nodes.Create(Node::Type::kFunction); x->attr("ir_node").Pointer() = ir_node; PADDLE_ENFORCE(ir_node->Op()->Proto()); @@ -141,7 +141,7 @@ void DataFlowGraph::Build(const framework::ir::Graph &graph) { } else if (ir_node->IsVar()) { // Not create a Node for IR ControlDepVar, considering Inference currently // just used in single thread scenerio. - VLOG(4) << "get var " << ir_node->Name(); + VLOG(40) << "get var " << ir_node->Name(); x = nodes.Create(Node::Type::kValue); x->attr("ir_node").Pointer() = ir_node; x->SetName(ir_node->Name()); @@ -151,9 +151,9 @@ void DataFlowGraph::Build(const framework::ir::Graph &graph) { } ir_node_map.emplace(ir_node, x); } - VLOG(4) << "finish creating Nodes"; + VLOG(40) << "finish creating Nodes"; - VLOG(4) << "to create edge"; + VLOG(40) << "to create edge"; // Create links for (auto *ir_node : graph.Nodes()) { auto it = ir_node_map.find(ir_node); @@ -175,7 +175,7 @@ void DataFlowGraph::Build(const framework::ir::Graph &graph) { "Can't deduce any inputs from the graph, Is the graph empty?"); ir_graph = &graph; - VLOG(3) << "finished build from IR"; + VLOG(30) << "finished build from IR"; } void DataFlowGraph::Clean() { diff --git a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc index 1682011c3d8cc9927a4b026b370671798cace625..50ce20621fb289023ecccf7bb39d98169765d5ee 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/framework/op_proto_maker.h" #include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/inference/analysis/ut_helper.h" @@ -130,6 +131,8 @@ void SetOp(framework::ProgramDesc* prog, const std::string& type, op->SetType(type); op->SetInput("Xs", inputs); op->SetOutput("Xs", outputs); + op->SetAttr(framework::OpProtoAndCheckerMaker::OpRoleAttrName(), + static_cast(framework::OpRole::kForward)); } TEST(DataFlowGraph, Build_IR_Graph) { diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc index cb549f4b50cf56154a951d16b58b022dbad3e990..dbe138514b20a4be20e7cca800f8e12b230e7824 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc @@ -239,9 +239,10 @@ void DataFlowGraphToFluidPass::AddEngineOp(Node *node) { framework::BlockDesc block_desc(nullptr, &proto); block_desc.Proto()->set_parent_idx(-1); block_desc.Proto()->set_idx(0); - VLOG(4) << "origin variable size: " - << argument_->origin_program_desc->blocks(0).vars().size(); - VLOG(4) << "transformed variable size: " << block_desc.Proto()->vars().size(); + VLOG(40) << "origin variable size: " + << argument_->origin_program_desc->blocks(0).vars().size(); + VLOG(40) << "transformed variable size: " + << block_desc.Proto()->vars().size(); // copy ops. for (auto *node : block_node->subgraph) { diff --git a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc index 648b8f7d6a6ec4bafbad2838c5631e776c8699b1..8888529a57a29c4349095c2ff4c527346716e026 100644 --- a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc +++ b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc @@ -29,7 +29,7 @@ void DFG_GraphvizDrawPass::Run(DataFlowGraph *graph) { auto png_path = dot_path.substr(0, dot_path.size() - 4) + ".png"; std::string message; - VLOG(3) << "draw to " << png_path; + VLOG(30) << "draw to " << png_path; ExecShellCommand("dot -Tpng " + dot_path + " -o " + png_path, &message); } diff --git a/paddle/fluid/inference/analysis/fluid_to_ir_pass.cc b/paddle/fluid/inference/analysis/fluid_to_ir_pass.cc index fc60ca3bd0bf706407defb2655a093d999aef7c2..9f52af670b8e26c31230bc003af4c9ddc5b67802 100644 --- a/paddle/fluid/inference/analysis/fluid_to_ir_pass.cc +++ b/paddle/fluid/inference/analysis/fluid_to_ir_pass.cc @@ -29,7 +29,7 @@ void FluidToIrPass::EnableParamModify(const std::string &model_dir, PADDLE_ENFORCE(argument_); argument_->Set(framework::ir::kParamScopeAttr, new framework::Scope); // Load parameters. - VLOG(3) << "Loading parameters from " << model_dir; + VLOG(30) << "Loading parameters from " << model_dir; LoadParams(&argument_->Get(framework::ir::kParamScopeAttr), model_dir, prog_file, param_file); } diff --git a/paddle/fluid/inference/analysis/model_store_pass.cc b/paddle/fluid/inference/analysis/model_store_pass.cc index c313db08875669010ddcca13aa66b383ee6d26f8..4f40a7a1adc324b824af8e3831901abcbffaeca6 100644 --- a/paddle/fluid/inference/analysis/model_store_pass.cc +++ b/paddle/fluid/inference/analysis/model_store_pass.cc @@ -35,21 +35,21 @@ void ModelStorePass::Run(DataFlowGraph *x) { std::stringstream ss; // NOTE these commands only works on linux. ss << "mkdir -p " << *argument_->model_output_store_path; - VLOG(3) << "run command: " << ss.str(); + VLOG(30) << "run command: " << ss.str(); PADDLE_ENFORCE_EQ(system(ss.str().c_str()), 0); ss.str(""); ss << "cp " << *argument_->fluid_model_dir << "/*" << " " << *argument_->model_output_store_path; - VLOG(3) << "run command: " << ss.str(); + VLOG(30) << "run command: " << ss.str(); PADDLE_ENFORCE_EQ(system(ss.str().c_str()), 0); // Store program PADDLE_ENFORCE_NOT_NULL(argument_->transformed_program_desc, "program desc is not transformed, should call " "DataFlowGraphToFluidPass first."); - VLOG(3) << "store analyzed program to " - << *argument_->model_output_store_path; + VLOG(30) << "store analyzed program to " + << *argument_->model_output_store_path; const std::string program_output_path = *argument_->model_output_store_path + "/__model__"; std::ofstream file(program_output_path, std::ios::binary); diff --git a/paddle/fluid/inference/analysis/pass_manager.cc b/paddle/fluid/inference/analysis/pass_manager.cc index a6ac0ee49f8f408faa7a17bf5ef5d2799a9a6238..ce390ee8313d6e3e2f0d79fb59d2225e2779180b 100644 --- a/paddle/fluid/inference/analysis/pass_manager.cc +++ b/paddle/fluid/inference/analysis/pass_manager.cc @@ -23,7 +23,7 @@ namespace analysis { bool PassManager::Initialize(Argument* argument) { argument_ = argument; for (auto& pass : data_) { - VLOG(3) << "Initializing pass [" << pass->repr() << "]"; + VLOG(30) << "Initializing pass [" << pass->repr() << "]"; if (!pass->Initialize(argument)) { LOG(ERROR) << "Failed to initialize pass [" << pass->repr() << "]"; return false; @@ -34,7 +34,7 @@ bool PassManager::Initialize(Argument* argument) { void DfgPassManager::RunAll() { PADDLE_ENFORCE(argument_); - VLOG(3) << "Total " << data_.size() << " Analysys passes"; + VLOG(30) << "Total " << data_.size() << " Analysys passes"; for (auto& pass : data_) { string::PrettyLogEndl(string::Style::H1(), "* Running Analysis pass [%s]", pass->repr()); diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.cc b/paddle/fluid/inference/analysis/subgraph_splitter.cc index 526bbbadfe90c3064d7c620cc22e30f7fef99088..3688ea15d959309d33901c360cb1055e2ac489a5 100644 --- a/paddle/fluid/inference/analysis/subgraph_splitter.cc +++ b/paddle/fluid/inference/analysis/subgraph_splitter.cc @@ -232,7 +232,7 @@ std::vector> SubGraphSplitter::ExtractSubGraphs() { BriefNode *brief_node = itr.second; if (!brief_node->node->attr(kMarkerAttrName).Bool()) { - VLOG(4) << brief_node->node->id() << " node not a trt candicate."; + VLOG(40) << brief_node->node->id() << " node not a trt candicate."; continue; } diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc index cc1746ecb34c983d219693bcec17c8789c38fa9f..3aa65f223a9e70b8ba7e387d1766ec6a97aee385 100644 --- a/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc @@ -25,9 +25,9 @@ TensorRTSubGraphPass::TensorRTSubGraphPass( void TensorRTSubGraphPass::Run(DataFlowGraph *graph) { SubGraphFuse(graph, node_inside_subgraph_teller_, argument_)(); - VLOG(4) << "debug info " - << graph->HumanReadableInfo(false /*show_values*/, - true /*show_functions*/); + VLOG(40) << "debug info " + << graph->HumanReadableInfo(false /*show_values*/, + true /*show_functions*/); } } // namespace analysis diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt index 0ddd5d53f836131fe37d412fc867cb38f11ee2b5..fd05c967774b336275f4c7bd98313bc1d750502f 100644 --- a/paddle/fluid/inference/api/CMakeLists.txt +++ b/paddle/fluid/inference/api/CMakeLists.txt @@ -17,32 +17,14 @@ if(APPLE) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") endif(APPLE) - set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB}) if(WITH_GPU AND TENSORRT_FOUND) set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor) endif() -function(inference_api_test TARGET_NAME) - if (WITH_TESTING) - set(options "") - set(oneValueArgs SRC) - set(multiValueArgs ARGS) - cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) - - cc_test(${TARGET_NAME} - SRCS ${inference_test_SRC} - DEPS "${inference_deps}" - ARGS --dirname=${PYTHON_TESTS_DIR}/book/) - if(inference_test_ARGS) - set_tests_properties(${TARGET_NAME} - PROPERTIES DEPENDS "${inference_test_ARGS}") - endif() - endif(WITH_TESTING) -endfunction(inference_api_test) - -cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS lod_tensor scope) +cc_library(reset_tensor_array SRCS details/reset_tensor_array.cc DEPS lod_tensor scope) +cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS reset_tensor_array lod_tensor scope) cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor) cc_library(zero_copy_tensor SRCS details/zero_copy_tensor.cc DEPS paddle_inference_api) cc_library(zero_copy_tensor_dummy SRCS details/zero_copy_tensor_dummy.cc DEPS paddle_inference_api) @@ -50,19 +32,22 @@ cc_test(test_paddle_inference_api SRCS api_tester.cc DEPS paddle_inference_api) -inference_api_test(test_api_impl SRC api_impl_tester.cc - ARGS test_word2vec test_image_classification) - -set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests) -cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} paddle_inference_api - ARGS --dirname=${PYTHON_TESTS_DIR}/book) +if(WITH_TESTING) + inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps} + ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book) + set_tests_properties(test_api_impl PROPERTIES DEPENDS test_image_classification) +endif() +cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} + ARGS --dirname=${WORD2VEC_MODEL_DIR}) if(WITH_GPU AND TENSORRT_FOUND) cc_library(paddle_inference_tensorrt_subgraph_engine SRCS api_tensorrt_subgraph_engine.cc DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api tensorrt_converter zero_copy_tensor_dummy) - -inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec) + if(WITH_TESTING) + inference_base_test(test_api_tensorrt_subgraph_engine SRCS api_tensorrt_subgraph_engine_tester.cc DEPS ${inference_deps} + ARGS --dirname=${WORD2VEC_MODEL_DIR}) + endif() endif() if (WITH_ANAKIN AND WITH_MKL) # only needed in CI diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc index 3bc6af5241c41bd805699121d614d431d46d863f..dd295854a87c9707aaa85e1e2c6111089e3fe885 100644 --- a/paddle/fluid/inference/api/analysis_predictor.cc +++ b/paddle/fluid/inference/api/analysis_predictor.cc @@ -25,9 +25,11 @@ #include "paddle/fluid/inference/api/paddle_inference_api.h" #include "paddle/fluid/inference/api/paddle_inference_pass.h" #include "paddle/fluid/inference/utils/singleton.h" +#include "paddle/fluid/platform/cpu_helper.h" #include "paddle/fluid/platform/profiler.h" DECLARE_bool(profile); +DECLARE_int32(paddle_num_threads); namespace paddle { @@ -36,7 +38,7 @@ using contrib::AnalysisConfig; bool AnalysisPredictor::Init( const std::shared_ptr &parent_scope, const std::shared_ptr &program) { - VLOG(3) << "Predictor::init()"; + VLOG(30) << "Predictor::init()"; #if !defined(_WIN32) if (FLAGS_profile) { LOG(WARNING) << "Profiler is actived, might affect the performance"; @@ -47,6 +49,9 @@ bool AnalysisPredictor::Init( } #endif + // no matter with or without MKLDNN + paddle::platform::SetNumThreads(FLAGS_paddle_num_threads); + if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); LOG(WARNING) << "ir optimize only supports CPU currently, enable_ir_optim " @@ -72,22 +77,19 @@ bool AnalysisPredictor::Init( inference_program_ = program; } - if (config_._use_mkldnn) { - executor_->EnableMKLDNN(*inference_program_); - } - executor_->Prepare(scope_.get(), *inference_program_, 0, config_.use_feed_fetch_ops); // Get the feed_target_names and fetch_target_names PrepareFeedFetch(); + return true; } bool AnalysisPredictor::Run(const std::vector &inputs, std::vector *output_data, int batch_size) { - VLOG(3) << "Predictor::predict"; + VLOG(30) << "Predictor::predict"; inference::Timer timer; timer.tic(); // set feed variable @@ -107,13 +109,17 @@ bool AnalysisPredictor::Run(const std::vector &inputs, LOG(ERROR) << "fail to get fetches"; return false; } - VLOG(3) << "predict cost: " << timer.toc() << "ms"; + VLOG(30) << "predict cost: " << timer.toc() << "ms"; + + // Fix TensorArray reuse not cleaned bug. + tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get()); + tensor_array_batch_cleaner_.ResetTensorArray(); return true; } bool AnalysisPredictor::SetFeed(const std::vector &inputs, framework::Scope *scope) { - VLOG(3) << "Predictor::set_feed"; + VLOG(30) << "Predictor::set_feed"; if (inputs.size() != feeds_.size()) { LOG(ERROR) << "wrong feed input size, need " << feeds_.size() << " but get " << inputs.size(); @@ -178,7 +184,7 @@ void AnalysisPredictor::GetFetchOne(const framework::LoDTensor &fetch, bool AnalysisPredictor::GetFetch(std::vector *outputs, framework::Scope *scope) { - VLOG(3) << "Predictor::get_fetch"; + VLOG(30) << "Predictor::get_fetch"; outputs->resize(fetchs_.size()); for (size_t i = 0; i < fetchs_.size(); ++i) { int idx = boost::get(fetchs_[i]->GetAttr("col")); @@ -220,13 +226,27 @@ void AnalysisPredictor::OptimizeInferenceProgram() { argument_.origin_program_desc.reset( new ProgramDesc(*inference_program_->Proto())); - PADDLE_ENFORCE( - config_.ir_mode == contrib::AnalysisConfig::IrPassMode::kExclude, - "Only kExclude is supported yet."); - Analyzer().DisableIrPasses(config_.ir_passes).Run(&argument_); + + switch (config_.ir_mode) { + case contrib::AnalysisConfig::IrPassMode::kExclude: + Analyzer() + .IncludeAllIrPasses() + .SetUseMkldnn(config_._use_mkldnn) + .DisableIrPasses(config_.ir_passes) + .Run(&argument_); + break; + case contrib::AnalysisConfig::IrPassMode::kInclude: + Analyzer() + .SetUseMkldnn(config_._use_mkldnn) + .IncludeIrPasses(config_.ir_passes) + .Run(&argument_); + break; + default: + LOG(ERROR) << "Only kExclude and kInclude modes are supoorted yet."; + } CHECK(argument_.transformed_program_desc); - VLOG(5) << "to prepare executor"; + VLOG(50) << "to prepare executor"; inference_program_.reset( new framework::ProgramDesc(*argument_.transformed_program_desc)); if (argument_.Has(framework::ir::kParamScopeAttr)) { @@ -240,7 +260,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() { template <> std::unique_ptr CreatePaddlePredictor< AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig &config) { - VLOG(3) << "create AnalysisConfig"; + VLOG(30) << "create AnalysisConfig"; if (config.use_gpu) { // 1. GPU memeroy PADDLE_ENFORCE_GT( @@ -254,7 +274,7 @@ std::unique_ptr CreatePaddlePredictor< std::string flag = "--fraction_of_gpu_memory_to_use=" + std::to_string(config.fraction_of_gpu_memory); flags.push_back(flag); - VLOG(3) << "set flag: " << flag; + VLOG(30) << "set flag: " << flag; framework::InitGflags(flags); } } @@ -307,6 +327,9 @@ std::unique_ptr AnalysisPredictor::GetOutputTensor( bool AnalysisPredictor::ZeroCopyRun() { executor_->Run(); + // Fix TensorArray reuse not cleaned bug. + tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get()); + tensor_array_batch_cleaner_.ResetTensorArray(); return true; } @@ -335,6 +358,19 @@ bool AnalysisPredictor::LoadProgramDesc() { } return true; } + +AnalysisPredictor::~AnalysisPredictor() { +#if !defined(_WIN32) + if (FLAGS_profile) { + platform::DisableProfiler(platform::EventSortingKey::kTotal, + "./profile.log"); + } +#endif + if (sub_scope_) { + scope_->DeleteScope(sub_scope_); + } +} + std::unique_ptr AnalysisPredictor::Clone() { auto *x = new AnalysisPredictor(config_); x->Init(scope_, inference_program_); diff --git a/paddle/fluid/inference/api/analysis_predictor.h b/paddle/fluid/inference/api/analysis_predictor.h index 0d01d7ac2b29ea6364b07af9bb3bdeb5ced6bd00..a9f4cce6dfa1c92301f57a7b1dd024a61f99d5ab 100644 --- a/paddle/fluid/inference/api/analysis_predictor.h +++ b/paddle/fluid/inference/api/analysis_predictor.h @@ -13,11 +13,14 @@ // limitations under the License. #pragma once +#include +#include #include #include #include "paddle/fluid/framework/naive_executor.h" #include "paddle/fluid/inference/analysis/analyzer.h" #include "paddle/fluid/inference/api/api_impl.h" +#include "paddle/fluid/inference/api/details/reset_tensor_array.h" #include "paddle/fluid/inference/api/paddle_inference_api.h" #include "paddle/fluid/string/printf.h" @@ -72,6 +75,7 @@ class AnalysisPredictor : public PaddlePredictor { template void GetFetchOne(const framework::LoDTensor &fetchs, PaddleTensor *output_data); + ~AnalysisPredictor(); private: contrib::AnalysisConfig config_; @@ -87,6 +91,7 @@ class AnalysisPredictor : public PaddlePredictor { // Memory buffer for feed inputs. The temporary LoDTensor will cause serious // concurrency problems, so cache them. std::vector feed_tensors_; + details::TensorArrayBatchCleaner tensor_array_batch_cleaner_; }; } // namespace paddle diff --git a/paddle/fluid/inference/api/analysis_predictor_tester.cc b/paddle/fluid/inference/api/analysis_predictor_tester.cc index 1d25f55b3188a684fe38df1417d114348cfa2e8a..f75c45f3a0438bc437e716160af8c5eab5b10fce 100644 --- a/paddle/fluid/inference/api/analysis_predictor_tester.cc +++ b/paddle/fluid/inference/api/analysis_predictor_tester.cc @@ -24,12 +24,10 @@ using contrib::AnalysisConfig; TEST(AnalysisPredictor, ZeroCopy) { AnalysisConfig config; - config.model_dir = FLAGS_dirname + "/word2vec.inference.model"; + config.model_dir = FLAGS_dirname; config.use_feed_fetch_ops = false; - auto predictor = - CreatePaddlePredictor( - config); + auto predictor = CreatePaddlePredictor(config); auto w0 = predictor->GetInputTensor("firstw"); auto w1 = predictor->GetInputTensor("secondw"); diff --git a/paddle/fluid/inference/api/api.cc b/paddle/fluid/inference/api/api.cc index 01ea942d3c8d20180cfc9664b8601ba87a898e86..20fab8078fedf837564496aa296648bf5970a348 100644 --- a/paddle/fluid/inference/api/api.cc +++ b/paddle/fluid/inference/api/api.cc @@ -16,7 +16,6 @@ #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/inference/api/paddle_inference_api.h" #include "paddle/fluid/platform/enforce.h" -#include "paddle_inference_api.h" namespace paddle { diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc index 6682e0a81b20c82aa668a249d37986386d769c83..fcbc3803d04def9a9855f2fee489e7e2c561b454 100644 --- a/paddle/fluid/inference/api/api_impl.cc +++ b/paddle/fluid/inference/api/api_impl.cc @@ -22,10 +22,13 @@ limitations under the License. */ #include "paddle/fluid/framework/feed_fetch_method.h" #include "paddle/fluid/inference/api/api_impl.h" +#include "paddle/fluid/inference/api/details/reset_tensor_array.h" #include "paddle/fluid/inference/api/helper.h" +#include "paddle/fluid/platform/cpu_helper.h" #include "paddle/fluid/platform/profiler.h" DEFINE_bool(profile, false, "Turn on profiler for fluid"); +DECLARE_int32(paddle_num_threads); namespace paddle { namespace { @@ -72,6 +75,9 @@ bool NativePaddlePredictor::Init( } #endif + // no matter with or without MKLDNN + paddle::platform::SetNumThreads(FLAGS_paddle_num_threads); + if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); } else { @@ -151,7 +157,11 @@ bool NativePaddlePredictor::Run(const std::vector &inputs, LOG(ERROR) << "fail to get fetches"; return false; } - VLOG(3) << "predict cost: " << timer.toc() << "ms"; + VLOG(30) << "predict cost: " << timer.toc() << "ms"; + + // Fix TensorArray reuse not cleaned bug. + tensor_array_batch_cleaner_.CollectTensorArrays(scope_.get()); + tensor_array_batch_cleaner_.ResetTensorArray(); return true; } diff --git a/paddle/fluid/inference/api/api_impl.h b/paddle/fluid/inference/api/api_impl.h index 7882f6a53c7ce9a2486158ea9b50c018d1814091..4e4ab47ca9c5e37f2714ebd48d250c23c7e9b117 100644 --- a/paddle/fluid/inference/api/api_impl.h +++ b/paddle/fluid/inference/api/api_impl.h @@ -26,11 +26,11 @@ limitations under the License. */ #include #include -#include "paddle/fluid/inference/api/paddle_inference_api.h" - #include "paddle/fluid/framework/ddim.h" #include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/naive_executor.h" +#include "paddle/fluid/inference/api/details/reset_tensor_array.h" #include "paddle/fluid/inference/api/paddle_inference_api.h" #include "paddle/fluid/inference/io.h" #include "paddle/fluid/platform/init.h" @@ -77,6 +77,7 @@ class NativePaddlePredictor : public PaddlePredictor { std::vector fetchs_; // Do not use unique_ptr, use parent scope to delete framework::Scope *sub_scope_{nullptr}; + details::TensorArrayBatchCleaner tensor_array_batch_cleaner_; }; } // namespace paddle diff --git a/paddle/fluid/inference/api/api_impl_tester.cc b/paddle/fluid/inference/api/api_impl_tester.cc index bed7c871311e476b4ed113d982c690383ad732de..5152b8670ddb206f0927c03149684af4a096df42 100644 --- a/paddle/fluid/inference/api/api_impl_tester.cc +++ b/paddle/fluid/inference/api/api_impl_tester.cc @@ -27,7 +27,9 @@ limitations under the License. */ #define ACC_DIFF 1e-3 #endif -DEFINE_string(dirname, "", "Directory of the inference model."); +DEFINE_string(word2vec_dirname, "", + "Directory of the word2vec inference model."); +DEFINE_string(book_dirname, "", "Directory of the book inference model."); namespace paddle { @@ -49,7 +51,7 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { NativeConfig GetConfig() { NativeConfig config; - config.model_dir = FLAGS_dirname + "/word2vec.inference.model"; + config.model_dir = FLAGS_word2vec_dirname; LOG(INFO) << "dirname " << config.model_dir; config.fraction_of_gpu_memory = 0.15; #ifdef PADDLE_WITH_CUDA @@ -116,7 +118,7 @@ void MainImageClassification(bool use_gpu) { NativeConfig config = GetConfig(); config.use_gpu = use_gpu; config.model_dir = - FLAGS_dirname + "/image_classification_resnet.inference.model"; + FLAGS_book_dirname + "/image_classification_resnet.inference.model"; const bool is_combined = false; std::vector> feed_target_shapes = @@ -187,7 +189,7 @@ void MainThreadsWord2Vec(bool use_gpu) { std::vector threads; for (int tid = 0; tid < num_jobs; ++tid) { threads.emplace_back([&, tid]() { - auto predictor = main_predictor->Clone(); + auto predictor = CreatePaddlePredictor(config); auto& local_inputs = paddle_tensor_feeds[tid]; std::vector local_outputs; ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs)); @@ -205,7 +207,7 @@ void MainThreadsWord2Vec(bool use_gpu) { float* ref_data = refs[tid].data(); EXPECT_EQ(refs[tid].numel(), static_cast(len / sizeof(float))); for (int i = 0; i < refs[tid].numel(); ++i) { - EXPECT_NEAR(ref_data[i], data[i], ACC_DIFF); + EXPECT_NEAR(ref_data[i], data[i], 2e-3); } }); } @@ -220,7 +222,7 @@ void MainThreadsImageClassification(bool use_gpu) { NativeConfig config = GetConfig(); config.use_gpu = use_gpu; config.model_dir = - FLAGS_dirname + "/image_classification_resnet.inference.model"; + FLAGS_book_dirname + "/image_classification_resnet.inference.model"; auto main_predictor = CreatePaddlePredictor(config); std::vector jobs(num_jobs); @@ -245,7 +247,7 @@ void MainThreadsImageClassification(bool use_gpu) { std::vector threads; for (int tid = 0; tid < num_jobs; ++tid) { threads.emplace_back([&, tid]() { - auto predictor = main_predictor->Clone(); + auto predictor = CreatePaddlePredictor(config); auto& local_inputs = paddle_tensor_feeds[tid]; std::vector local_outputs; ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs)); @@ -271,7 +273,7 @@ TEST(inference_api_native, word2vec_cpu_threads) { MainThreadsWord2Vec(false /*use_gpu*/); } TEST(inference_api_native, image_classification_cpu) { - MainThreadsImageClassification(false /*use_gpu*/); + MainImageClassification(false /*use_gpu*/); } TEST(inference_api_native, image_classification_cpu_threads) { MainThreadsImageClassification(false /*use_gpu*/); @@ -279,15 +281,17 @@ TEST(inference_api_native, image_classification_cpu_threads) { #ifdef PADDLE_WITH_CUDA TEST(inference_api_native, word2vec_gpu) { MainWord2Vec(true /*use_gpu*/); } -TEST(inference_api_native, word2vec_gpu_threads) { - MainThreadsWord2Vec(true /*use_gpu*/); -} +// Turn off temporarily for the unstable result. +// TEST(inference_api_native, word2vec_gpu_threads) { +// MainThreadsWord2Vec(true /*use_gpu*/); +// } TEST(inference_api_native, image_classification_gpu) { - MainThreadsImageClassification(true /*use_gpu*/); -} -TEST(inference_api_native, image_classification_gpu_threads) { - MainThreadsImageClassification(true /*use_gpu*/); + MainImageClassification(true /*use_gpu*/); } +// Turn off temporarily for the unstable result. +// TEST(inference_api_native, image_classification_gpu_threads) { +// MainThreadsImageClassification(true /*use_gpu*/); +// } #endif diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc index 5ee6a5a93168f58770067f76ca7f6bb6f67b2965..94b3933497daac1a4db1787994ea1bc33ec4e74f 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc @@ -34,7 +34,7 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor { bool Init(const std::shared_ptr& parent_scope) { FLAGS_IA_enable_tensorrt_subgraph_engine = true; - VLOG(3) << "Predictor::init()"; + VLOG(30) << "Predictor::init()"; if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); } else { @@ -70,7 +70,7 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor { OptimizeInferenceProgram(); ctx_ = executor_->Prepare(*inference_program_, 0); - VLOG(5) << "to create variables"; + VLOG(50) << "to create variables"; executor_->CreateVariables(*inference_program_, sub_scope_ ? sub_scope_ : scope_.get(), 0); // Get the feed_target_names and fetch_target_names @@ -114,9 +114,9 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor { new ProgramDesc(*inference_program_->Proto())); Singleton::Global().Run(&argument); CHECK(argument.transformed_program_desc); - VLOG(5) << "transformed program:\n" - << argument.transformed_program_desc->SerializeAsString(); - VLOG(5) << "to prepare executor"; + VLOG(50) << "transformed program:\n" + << argument.transformed_program_desc->SerializeAsString(); + VLOG(50) << "to prepare executor"; inference_program_.reset( new framework::ProgramDesc(*argument.transformed_program_desc)); } @@ -129,7 +129,7 @@ template <> std::unique_ptr CreatePaddlePredictor( const MixedRTConfig& config) { - VLOG(3) << "create TensorRTSubgraphPredictor"; + VLOG(30) << "create TensorRTSubgraphPredictor"; if (config.use_gpu) { // 1. GPU memeroy PADDLE_ENFORCE_GT( @@ -143,7 +143,7 @@ CreatePaddlePredictor( std::string flag = "--fraction_of_gpu_memory_to_use=" + std::to_string(config.fraction_of_gpu_memory); flags.push_back(flag); - VLOG(3) << "set flag: " << flag; + VLOG(30) << "set flag: " << flag; framework::InitGflags(flags); } } @@ -185,3 +185,4 @@ USE_TRT_CONVERTER(softmax); USE_TRT_CONVERTER(batch_norm); USE_TRT_CONVERTER(concat); USE_TRT_CONVERTER(dropout); +USE_TRT_CONVERTER(pad); diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc index fc6310e90b0257bc84742fb617a00f5778bb1866..89c9a65cb06ba565f0e0cbdb9b6031c6adbcb64e 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc @@ -29,23 +29,20 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) { //# 1. Create PaddlePredictor with a config. NativeConfig config0; - config0.model_dir = FLAGS_dirname + "word2vec.inference.model"; + config0.model_dir = FLAGS_dirname; config0.use_gpu = true; config0.fraction_of_gpu_memory = 0.3; config0.device = 0; MixedRTConfig config1; - config1.model_dir = FLAGS_dirname + "word2vec.inference.model"; + config1.model_dir = FLAGS_dirname; config1.use_gpu = true; config1.fraction_of_gpu_memory = 0.3; config1.device = 0; config1.max_batch_size = 10; - auto predictor0 = - CreatePaddlePredictor(config0); - auto predictor1 = - CreatePaddlePredictor(config1); + auto predictor0 = CreatePaddlePredictor(config0); + auto predictor1 = CreatePaddlePredictor(config1); for (int batch_id = 0; batch_id < 1; batch_id++) { //# 2. Prepare input. diff --git a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt index d4e6bb3e4a4ceb361ccd35121d0ecf84a764243e..49683eab07a2f5bc008272038a27bdb277396284 100644 --- a/paddle/fluid/inference/api/demo_ci/CMakeLists.txt +++ b/paddle/fluid/inference/api/demo_ci/CMakeLists.txt @@ -3,6 +3,7 @@ project(cpp_inference_demo CXX C) option(WITH_MKL "Compile demo with MKL/OpenBlas support, default use MKL." ON) option(WITH_GPU "Compile demo with GPU/CPU, default use CPU." OFF) option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static." ON) +option(USE_TENSORRT "Compile demo with TensorRT." OFF) macro(safe_set_static_flag) foreach(flag_var @@ -51,6 +52,7 @@ include_directories("${PADDLE_LIB}") include_directories("${PADDLE_LIB}/third_party/install/protobuf/include") include_directories("${PADDLE_LIB}/third_party/install/glog/include") include_directories("${PADDLE_LIB}/third_party/install/gflags/include") +include_directories("${PADDLE_LIB}/third_party/install/xxhash/include") if (NOT WIN32) include_directories("${PADDLE_LIB}/third_party/install/snappy/include") include_directories("${PADDLE_LIB}/third_party/install/snappystream/include") @@ -60,6 +62,13 @@ endif(NOT WIN32) include_directories("${PADDLE_LIB}/third_party/boost") include_directories("${PADDLE_LIB}/third_party/eigen3") +if (NOT WIN32) + if (USE_TENSORRT AND WITH_GPU) + include_directories("${TENSORRT_INCLUDE_DIR}") + link_directories("${TENSORRT_LIB_DIR}") + endif() +endif(NOT WIN32) + if (NOT WIN32) link_directories("${PADDLE_LIB}/third_party/install/snappy/lib") link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib") @@ -69,13 +78,14 @@ endif(NOT WIN32) link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib") link_directories("${PADDLE_LIB}/third_party/install/glog/lib") link_directories("${PADDLE_LIB}/third_party/install/gflags/lib") -link_directories("${PADDLE_LIB}/paddle/fluid/inference") +link_directories("${PADDLE_LIB}/third_party/install/xxhash/lib") +link_directories("${PADDLE_LIB}/paddle/lib") add_executable(${DEMO_NAME} ${DEMO_NAME}.cc) if(WITH_MKL) include_directories("${PADDLE_LIB}/third_party/install/mklml/include") - set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} + set(MATH_LIB ${PADDLE_LIB}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} ${PADDLE_LIB}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX}) set(MKLDNN_PATH "${PADDLE_LIB}/third_party/install/mkldnn") if(EXISTS ${MKLDNN_PATH}) @@ -89,17 +99,17 @@ endif() # Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a if(WITH_STATIC_LIB) set(DEPS - ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) + ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}) else() set(DEPS - ${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX}) + ${PADDLE_LIB}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX}) endif() if (NOT WIN32) set(EXTERNAL_LIB "-lrt -ldl -lpthread") set(DEPS ${DEPS} ${MATH_LIB} ${MKLDNN_LIB} - glog gflags protobuf snappystream snappy z + glog gflags protobuf snappystream snappy z xxhash ${EXTERNAL_LIB}) else() set(DEPS ${DEPS} @@ -112,6 +122,10 @@ endif(NOT WIN32) if(WITH_GPU) if(NOT WIN32) + if (USE_TENSORRT) + set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer${CMAKE_STATIC_LIBRARY_SUFFIX}) + set(DEPS ${DEPS} ${TENSORRT_LIB_DIR}/libnvinfer_plugin${CMAKE_STATIC_LIBRARY_SUFFIX}) + endif() set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX}) else() set(DEPS ${DEPS} ${CUDA_LIB}/cudart${CMAKE_STATIC_LIBRARY_SUFFIX} ) diff --git a/paddle/fluid/inference/api/demo_ci/run.sh b/paddle/fluid/inference/api/demo_ci/run.sh index 44335a872f2e00b34e29a9e7601cb390a460362c..ff718077c1ba6b10fe87aac10d84f96a23ad6bba 100755 --- a/paddle/fluid/inference/api/demo_ci/run.sh +++ b/paddle/fluid/inference/api/demo_ci/run.sh @@ -3,19 +3,28 @@ PADDLE_ROOT=$1 TURN_ON_MKL=$2 # use MKL or Openblas TEST_GPU_CPU=$3 # test both GPU/CPU mode or only CPU mode DATA_DIR=$4 # dataset +TENSORRT_INCLUDE_DIR=$5 # TensorRT header file dir, defalut to /usr/local/TensorRT/include +TENSORRT_LIB_DIR=$6 # TensorRT lib file dir, default to /usr/local/TensorRT/lib +inference_install_dir=${PADDLE_ROOT}/build/fluid_inference_install_dir + cd `dirname $0` current_dir=`pwd` if [ $2 == ON ]; then # You can export yourself if move the install path - MKL_LIB=${PADDLE_ROOT}/build/fluid_install_dir/third_party/install/mklml/lib + MKL_LIB=${inference_install_dir}/third_party/install/mklml/lib export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:${MKL_LIB} fi if [ $3 == ON ]; then use_gpu_list='true false' -else +else use_gpu_list='false' fi +USE_TENSORRT=OFF +if [ -d "$TENSORRT_INCLUDE_DIR" -a -d "$TENSORRT_LIB_DIR" ]; then + USE_TENSORRT=ON +fi + PREFIX=inference-vis-demos%2F URL_ROOT=http://paddlemodels.cdn.bcebos.com/${PREFIX} @@ -47,13 +56,13 @@ cd build for WITH_STATIC_LIB in ON OFF; do # -----simple_on_word2vec----- rm -rf * - cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \ + cmake .. -DPADDLE_LIB=${inference_install_dir} \ -DWITH_MKL=$TURN_ON_MKL \ -DDEMO_NAME=simple_on_word2vec \ -DWITH_GPU=$TEST_GPU_CPU \ -DWITH_STATIC_LIB=$WITH_STATIC_LIB make -j - word2vec_model=${PADDLE_ROOT}'/build/python/paddle/fluid/tests/book/word2vec.inference.model' + word2vec_model=$DATA_DIR'/word2vec/word2vec.inference.model' if [ -d $word2vec_model ]; then for use_gpu in $use_gpu_list; do ./simple_on_word2vec \ @@ -67,14 +76,14 @@ for WITH_STATIC_LIB in ON OFF; do fi # ---------vis_demo--------- rm -rf * - cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \ + cmake .. -DPADDLE_LIB=${inference_install_dir} \ -DWITH_MKL=$TURN_ON_MKL \ -DDEMO_NAME=vis_demo \ -DWITH_GPU=$TEST_GPU_CPU \ -DWITH_STATIC_LIB=$WITH_STATIC_LIB make -j for use_gpu in $use_gpu_list; do - for vis_demo_name in $vis_demo_list; do + for vis_demo_name in $vis_demo_list; do ./vis_demo \ --modeldir=$DATA_DIR/$vis_demo_name/model \ --data=$DATA_DIR/$vis_demo_name/data.txt \ @@ -86,5 +95,23 @@ for WITH_STATIC_LIB in ON OFF; do fi done done + + # --------tensorrt mobilenet------ + if [ $USE_TENSORRT == ON -a $TEST_GPU_CPU == ON ]; then + rm -rf * + cmake .. -DPADDLE_LIB=${inference_install_dir} \ + -DWITH_MKL=$TURN_ON_MKL \ + -DDEMO_NAME=trt_mobilenet_demo \ + -DWITH_GPU=$TEST_GPU_CPU \ + -DWITH_STATIC_LIB=$WITH_STATIC_LIB \ + -DUSE_TENSORRT=$USE_TENSORRT \ + -DTENSORRT_INCLUDE_DIR=$TENSORRT_INCLUDE_DIR \ + -DTENSORRT_LIB_DIR=$TENSORRT_LIB_DIR + make -j + ./trt_mobilenet_demo \ + --modeldir=$DATA_DIR/mobilenet/model \ + --data=$DATA_DIR/mobilenet/data.txt \ + --refer=$DATA_DIR/mobilenet/result.txt + fi done set +x diff --git a/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc b/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc index 8058d7e881025b1d806efe187d4523adadff367d..5446fd4d4256c10442a53ea09a447cf308cbd681 100644 --- a/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc +++ b/paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc @@ -23,7 +23,7 @@ limitations under the License. */ #include #include //NOLINT -#include "paddle/fluid/inference/paddle_inference_api.h" +#include "paddle/include/paddle_inference_api.h" DEFINE_string(dirname, "", "Directory of the inference model."); DEFINE_bool(use_gpu, false, "Whether use gpu."); @@ -42,8 +42,7 @@ void Main(bool use_gpu) { config.use_gpu = use_gpu; config.fraction_of_gpu_memory = 0.15; config.device = 0; - auto predictor = - CreatePaddlePredictor(config); + auto predictor = CreatePaddlePredictor(config); for (int batch_id = 0; batch_id < 3; batch_id++) { //# 2. Prepare input. @@ -85,8 +84,7 @@ void MainThreads(int num_threads, bool use_gpu) { config.use_gpu = use_gpu; config.fraction_of_gpu_memory = 0.15; config.device = 0; - auto main_predictor = - CreatePaddlePredictor(config); + auto main_predictor = CreatePaddlePredictor(config); std::vector threads; for (int tid = 0; tid < num_threads; ++tid) { diff --git a/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc b/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc new file mode 100644 index 0000000000000000000000000000000000000000..6460514f3f80cac3c5e52560ab61b5cc7fd74636 --- /dev/null +++ b/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc @@ -0,0 +1,82 @@ +/* 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. */ + +/* + * This file contains demo of mobilenet for tensorrt. + */ + +#include +#include // use glog instead of CHECK to avoid importing other paddle header files. +#include "utils.h" // NOLINT + +DECLARE_double(fraction_of_gpu_memory_to_use); +DEFINE_string(modeldir, "", "Directory of the inference model."); +DEFINE_string(refer, "", "path to reference result for comparison."); +DEFINE_string( + data, "", + "path of data; each line is a record, format is " + "'\t predictor; + paddle::contrib::MixedRTConfig config; + config.param_file = FLAGS_modeldir + "/__params__"; + config.prog_file = FLAGS_modeldir + "/__model__"; + config.use_gpu = true; + config.device = 0; + config.max_batch_size = 1; + config.fraction_of_gpu_memory = 0.1; // set by yourself + predictor = CreatePaddlePredictor(config); + + VLOG(30) << "begin to process data"; + // Just a single batch of data. + std::string line; + std::ifstream file(FLAGS_data); + std::getline(file, line); + auto record = ProcessALine(line); + file.close(); + + // Inference. + PaddleTensor input; + input.shape = record.shape; + input.data = + PaddleBuf(record.data.data(), record.data.size() * sizeof(float)); + input.dtype = PaddleDType::FLOAT32; + + VLOG(30) << "run executor"; + std::vector output; + predictor->Run({input}, &output, 1); + + VLOG(30) << "output.size " << output.size(); + auto& tensor = output.front(); + VLOG(30) << "output: " << SummaryTensor(tensor); + + // compare with reference result + CheckOutput(FLAGS_refer, tensor); +} + +} // namespace demo +} // namespace paddle + +int main(int argc, char** argv) { + google::ParseCommandLineFlags(&argc, &argv, true); + paddle::demo::Main(); + return 0; +} diff --git a/paddle/fluid/inference/api/demo_ci/utils.h b/paddle/fluid/inference/api/demo_ci/utils.h index cb8990671162dff47228736e69617229528cc093..664b9d01c7810aa4f053cd6ebbff5f3f7619fd05 100644 --- a/paddle/fluid/inference/api/demo_ci/utils.h +++ b/paddle/fluid/inference/api/demo_ci/utils.h @@ -14,13 +14,20 @@ #pragma once #include +#include +#include #include #include -#include "paddle/fluid/inference/paddle_inference_api.h" +#include "paddle/include/paddle_inference_api.h" namespace paddle { namespace demo { +struct Record { + std::vector data; + std::vector shape; +}; + static void split(const std::string& str, char sep, std::vector* pieces) { pieces->clear(); @@ -39,6 +46,58 @@ static void split(const std::string& str, char sep, } } +Record ProcessALine(const std::string& line) { + VLOG(30) << "process a line"; + std::vector columns; + split(line, '\t', &columns); + CHECK_EQ(columns.size(), 2UL) + << "data format error, should be \t"; + + Record record; + std::vector data_strs; + split(columns[0], ' ', &data_strs); + for (auto& d : data_strs) { + record.data.push_back(std::stof(d)); + } + + std::vector shape_strs; + split(columns[1], ' ', &shape_strs); + for (auto& s : shape_strs) { + record.shape.push_back(std::stoi(s)); + } + VLOG(30) << "data size " << record.data.size(); + VLOG(30) << "data shape size " << record.shape.size(); + return record; +} + +void CheckOutput(const std::string& referfile, const PaddleTensor& output) { + std::string line; + std::ifstream file(referfile); + std::getline(file, line); + auto refer = ProcessALine(line); + file.close(); + + size_t numel = output.data.length() / PaddleDtypeSize(output.dtype); + VLOG(30) << "predictor output numel " << numel; + VLOG(30) << "reference output numel " << refer.data.size(); + CHECK_EQ(numel, refer.data.size()); + switch (output.dtype) { + case PaddleDType::INT64: { + for (size_t i = 0; i < numel; ++i) { + CHECK_EQ(static_cast(output.data.data())[i], refer.data[i]); + } + break; + } + case PaddleDType::FLOAT32: + for (size_t i = 0; i < numel; ++i) { + CHECK_LT( + fabs(static_cast(output.data.data())[i] - refer.data[i]), + 1e-5); + } + break; + } +} + /* * Get a summary of a PaddleTensor content. */ diff --git a/paddle/fluid/inference/api/demo_ci/vis_demo.cc b/paddle/fluid/inference/api/demo_ci/vis_demo.cc index fb59cea457027854a099574c867299450690e61c..d747f855803a6997d08957b5d35a56a0fe4160c5 100644 --- a/paddle/fluid/inference/api/demo_ci/vis_demo.cc +++ b/paddle/fluid/inference/api/demo_ci/vis_demo.cc @@ -18,11 +18,7 @@ limitations under the License. */ #include #include // use glog instead of CHECK to avoid importing other paddle header files. -#include -#include - -// #include "paddle/fluid/platform/enforce.h" -#include "paddle/fluid/inference/demo_ci/utils.h" +#include "utils.h" // NOLINT #ifdef PADDLE_WITH_CUDA DECLARE_double(fraction_of_gpu_memory_to_use); @@ -38,70 +34,13 @@ DEFINE_bool(use_gpu, false, "Whether use gpu."); namespace paddle { namespace demo { -struct Record { - std::vector data; - std::vector shape; -}; - -void split(const std::string& str, char sep, std::vector* pieces); - -Record ProcessALine(const std::string& line) { - VLOG(3) << "process a line"; - std::vector columns; - split(line, '\t', &columns); - CHECK_EQ(columns.size(), 2UL) - << "data format error, should be \t"; - - Record record; - std::vector data_strs; - split(columns[0], ' ', &data_strs); - for (auto& d : data_strs) { - record.data.push_back(std::stof(d)); - } - - std::vector shape_strs; - split(columns[1], ' ', &shape_strs); - for (auto& s : shape_strs) { - record.shape.push_back(std::stoi(s)); - } - VLOG(3) << "data size " << record.data.size(); - VLOG(3) << "data shape size " << record.shape.size(); - return record; -} - -void CheckOutput(const std::string& referfile, const PaddleTensor& output) { - std::string line; - std::ifstream file(referfile); - std::getline(file, line); - auto refer = ProcessALine(line); - file.close(); - - size_t numel = output.data.length() / PaddleDtypeSize(output.dtype); - VLOG(3) << "predictor output numel " << numel; - VLOG(3) << "reference output numel " << refer.data.size(); - CHECK_EQ(numel, refer.data.size()); - switch (output.dtype) { - case PaddleDType::INT64: { - for (size_t i = 0; i < numel; ++i) { - CHECK_EQ(static_cast(output.data.data())[i], refer.data[i]); - } - break; - } - case PaddleDType::FLOAT32: - for (size_t i = 0; i < numel; ++i) { - CHECK_LT( - fabs(static_cast(output.data.data())[i] - refer.data[i]), - 1e-5); - } - break; - } -} - +using contrib::AnalysisConfig; /* - * Use the native fluid engine to inference the demo. + * Use the native and analysis fluid engine to inference the demo. */ void Main(bool use_gpu) { - NativeConfig config; + std::unique_ptr predictor, analysis_predictor; + AnalysisConfig config; config.param_file = FLAGS_modeldir + "/__params__"; config.prog_file = FLAGS_modeldir + "/__model__"; config.use_gpu = use_gpu; @@ -110,11 +49,11 @@ void Main(bool use_gpu) { config.fraction_of_gpu_memory = 0.1; // set by yourself } - VLOG(3) << "init predictor"; - auto predictor = - CreatePaddlePredictor(config); + VLOG(30) << "init predictor"; + predictor = CreatePaddlePredictor(config); + analysis_predictor = CreatePaddlePredictor(config); - VLOG(3) << "begin to process data"; + VLOG(30) << "begin to process data"; // Just a single batch of data. std::string line; std::ifstream file(FLAGS_data); @@ -129,16 +68,20 @@ void Main(bool use_gpu) { PaddleBuf(record.data.data(), record.data.size() * sizeof(float)); input.dtype = PaddleDType::FLOAT32; - VLOG(3) << "run executor"; - std::vector output; - predictor->Run({input}, &output); + VLOG(30) << "run executor"; + std::vector output, analysis_output; + predictor->Run({input}, &output, 1); - VLOG(3) << "output.size " << output.size(); + VLOG(30) << "output.size " << output.size(); auto& tensor = output.front(); - VLOG(3) << "output: " << SummaryTensor(tensor); + VLOG(30) << "output: " << SummaryTensor(tensor); // compare with reference result CheckOutput(FLAGS_refer, tensor); + + // the analysis_output has some diff with native_output, + // TODO(luotao): add CheckOutput for analysis_output later. + analysis_predictor->Run({input}, &analysis_output, 1); } } // namespace demo @@ -146,9 +89,10 @@ void Main(bool use_gpu) { int main(int argc, char** argv) { google::ParseCommandLineFlags(&argc, &argv, true); - paddle::demo::Main(false /* use_gpu*/); if (FLAGS_use_gpu) { paddle::demo::Main(true /*use_gpu*/); + } else { + paddle::demo::Main(false /*use_gpu*/); } return 0; } diff --git a/paddle/fluid/inference/api/details/reset_tensor_array.cc b/paddle/fluid/inference/api/details/reset_tensor_array.cc new file mode 100644 index 0000000000000000000000000000000000000000..244b0b567b5df6735acd7f1bf3c2056f449be872 --- /dev/null +++ b/paddle/fluid/inference/api/details/reset_tensor_array.cc @@ -0,0 +1,50 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/api/details/reset_tensor_array.h" + +namespace paddle { +namespace details { + +// Should be called after the parameters are loaded. +void TensorArrayBatchCleaner::CollectTensorArrays(framework::Scope *scope) { + if (flag_) { + for (auto &var_name : scope->LocalVarNames()) { + auto *var = scope->FindVar(var_name); + // TODO(Superjomn) should avoid the case when a TensorArray is a + // parameter. + if (var_name == "feed" || var_name == "fetch") continue; + if (var->Type() == typeid(framework::LoDTensorArray)) { + VLOG(40) << "collect " << var_name; + arrays_.push_back(var->GetMutable()); + } + } + for (auto *kid : scope->kids()) { + CollectTensorArrays(kid); + } + + VLOG(30) << "Collect " << arrays_.size() << " arrays"; + flag_ = false; + } +} + +// Should be called when `Run` finished. +void TensorArrayBatchCleaner::ResetTensorArray() { + for (auto *arr : arrays_) { + arr->clear(); + } +} + +} // namespace details +} // namespace paddle diff --git a/paddle/fluid/inference/api/details/reset_tensor_array.h b/paddle/fluid/inference/api/details/reset_tensor_array.h new file mode 100644 index 0000000000000000000000000000000000000000..a39449ff0e67786815dfb8d2d30d79dcdba757d7 --- /dev/null +++ b/paddle/fluid/inference/api/details/reset_tensor_array.h @@ -0,0 +1,37 @@ +// 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 "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/framework/scope.h" + +namespace paddle { +namespace details { + +// Clean the TensorArray each batch to make the behavior the same with the +// training phase. +struct TensorArrayBatchCleaner { + // Fix the tensor array not clear in the inference scenarios. + void CollectTensorArrays(framework::Scope *scope); + void ResetTensorArray(); + + private: + bool flag_{true}; + std::vector arrays_; +}; + +} // namespace details +} // namespace paddle diff --git a/paddle/fluid/inference/api/helper.h b/paddle/fluid/inference/api/helper.h index 24f59cf43a9700ff1732e1ef6ad82e1a6294eede..af21c0095c28b26c0ef4afc83572a9681d49d497 100644 --- a/paddle/fluid/inference/api/helper.h +++ b/paddle/fluid/inference/api/helper.h @@ -16,13 +16,14 @@ #include #include +#include #include // NOLINT #include #include #include #include +#include "paddle/fluid/inference/api/paddle_inference_api.h" #include "paddle/fluid/string/printf.h" -#include "paddle_inference_api.h" namespace paddle { namespace inference { @@ -160,7 +161,8 @@ static void PrintTime(int batch_size, int repeat, int num_threads, int tid, double latency, int epoch = 1) { LOG(INFO) << "====== batch_size: " << batch_size << ", repeat: " << repeat << ", threads: " << num_threads << ", thread id: " << tid - << ", latency: " << latency << "ms ======"; + << ", latency: " << latency << "ms, fps: " << 1 / (latency / 1000.f) + << " ======"; if (epoch > 1) { int samples = batch_size * epoch; LOG(INFO) << "====== sample number: " << samples diff --git a/paddle/fluid/inference/api/paddle_inference_api.h b/paddle/fluid/inference/api/paddle_inference_api.h index d2876dc27c8826c2f27be21fa0b9fef92d03067a..a755ccb93bdee018dfeaf91157e7971b4d4cd832 100644 --- a/paddle/fluid/inference/api/paddle_inference_api.h +++ b/paddle/fluid/inference/api/paddle_inference_api.h @@ -124,7 +124,7 @@ class ZeroCopyTensor { std::vector> lod() const; protected: - ZeroCopyTensor(void* scope) : scope_{scope} {} + explicit ZeroCopyTensor(void* scope) : scope_{scope} {} void SetName(const std::string& name) { name_ = name; } void* FindTensor() const; @@ -263,6 +263,7 @@ struct AnalysisConfig : public NativeConfig { bool enable_ir_optim = true; // Manually determine the IR passes to run. IrPassMode ir_mode{IrPassMode::kExclude}; + // passes to be excluded/included std::vector ir_passes{"embedding_fc_lstm_fuse_pass"}; // NOT stable yet. diff --git a/paddle/fluid/inference/io.cc b/paddle/fluid/inference/io.cc index e246a06fd079d837ac321197914c9f70b528f2c8..bb749e8f8b0ba9d5cd82d91ce86c619f52f34c30 100644 --- a/paddle/fluid/inference/io.cc +++ b/paddle/fluid/inference/io.cc @@ -59,7 +59,8 @@ void ReadBinaryFile(const std::string& filename, std::string* contents) { bool IsPersistable(const framework::VarDesc* var) { if (var->Persistable() && var->GetType() != framework::proto::VarType::FEED_MINIBATCH && - var->GetType() != framework::proto::VarType::FETCH_LIST) { + var->GetType() != framework::proto::VarType::FETCH_LIST && + var->GetType() != framework::proto::VarType::RAW) { return true; } return false; @@ -77,7 +78,7 @@ void LoadPersistables(framework::Executor* executor, framework::Scope* scope, for (auto* var : global_block.AllVars()) { if (IsPersistable(var)) { - VLOG(3) << "persistable variable's name: " << var->Name(); + VLOG(30) << "persistable variable's name: " << var->Name(); framework::VarDesc* new_var = load_block->Var(var->Name()); new_var->SetShape(var->GetShape()); @@ -120,7 +121,7 @@ std::unique_ptr Load(framework::Executor* executor, const std::string& dirname) { std::string model_filename = dirname + "/__model__"; std::string program_desc_str; - VLOG(3) << "loading model from " << model_filename; + VLOG(30) << "loading model from " << model_filename; ReadBinaryFile(model_filename, &program_desc_str); std::unique_ptr main_program( diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index fac1babf6ec6131f84d3e3b9fc6efedd9f9f6cfc..0a35e10f6936313928ab21a6f17c40335e8fc882 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,7 +1,7 @@ # Add TRT tests nv_library(tensorrt_converter SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc -batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc +batch_norm_op.cc activation_op.cc softmax_op.cc concat_op.cc dropout_op.cc pad_op.cc DEPS tensorrt_engine operator scope framework_proto op_registry) nv_test(test_op_converter SRCS test_op_converter.cc DEPS @@ -26,6 +26,8 @@ nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL) nv_test(test_trt_concat_op SRCS test_concat_op.cc concat_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine concat_op SERIAL) - nv_test(test_trt_dropout_op SRCS test_dropout_op.cc dropout_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine dropout_op SERIAL) + +nv_test(test_trt_pad_op SRCS test_pad_op.cc pad_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine pad_op SERIAL) diff --git a/paddle/fluid/inference/tensorrt/convert/activation_op.cc b/paddle/fluid/inference/tensorrt/convert/activation_op.cc index e73c5bbf57501e4ff3c080a46d91685035652bfa..0b756534ec6fbf27a3e92bf39fb7544d9785ca48 100644 --- a/paddle/fluid/inference/tensorrt/convert/activation_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/activation_op.cc @@ -27,7 +27,7 @@ class ActivationOpConverter : public OpConverter { // Here the two nullptr looks strange, that's because the // framework::OpDesc's constructor is strange. framework::OpDesc op_desc(op, nullptr); - LOG(INFO) + VLOG(3) << "convert a fluid Activation op to tensorrt activation layer whose " "type is " << op_type_; diff --git a/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc b/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc index 3330af2da6c97ad153dcecd86be4b441eac62b5e..d017bac66dd99a4b54c44ec786de61d1e66b8981 100644 --- a/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc @@ -23,7 +23,7 @@ class BatchNormOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, const framework::Scope& scope, bool test_mode) override { - LOG(INFO) << "convert a fluid batch norm op to tensorrt batch_norm"; + VLOG(3) << "convert a fluid batch norm op to tensorrt batch_norm"; framework::OpDesc op_desc(op, nullptr); PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); diff --git a/paddle/fluid/inference/tensorrt/convert/concat_op.cc b/paddle/fluid/inference/tensorrt/convert/concat_op.cc index a11dfa1e8f2dacfad067d025678911200db500fb..b2e7c593e85974898012f8a353817a27ca212f4d 100644 --- a/paddle/fluid/inference/tensorrt/convert/concat_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/concat_op.cc @@ -25,7 +25,7 @@ class ConcatOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, const framework::Scope& scope, bool test_mode) override { - VLOG(4) << "convert a fluid mul op to tensorrt mul layer without bias"; + VLOG(3) << "convert a fluid mul op to tensorrt mul layer without bias"; framework::OpDesc op_desc(op, nullptr); // Declare inputs diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index 0a37d3968c39d2c244bbd82161afddf6330e421d..43950b8c048b4e1b8974956948caa639812b2f78 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -18,12 +18,26 @@ namespace paddle { namespace inference { namespace tensorrt { +bool to_skip_merging_optimize(TensorRTEngine* engine_, + const std::vector& filters, + const std::vector& strides, + const std::vector& paddings, + std::string input_name) { + if (engine_->itensor_quote_num[input_name] > 0) { + return true; + } + if (filters[0] == 1 && filters[1] == 1 && strides[0] == 1 && + strides[1] == 1 && paddings[0] == 0 && paddings[1] == 0) + engine_->itensor_quote_num[input_name] += 1; + + return false; +} + class Conv2dOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, const framework::Scope& scope, bool test_mode) override { - LOG(INFO) - << "convert a fluid conv2d op to tensorrt conv layer without bias"; + VLOG(3) << "convert a fluid conv2d op to tensorrt conv layer without bias"; framework::OpDesc op_desc(op, nullptr); PADDLE_ENFORCE_EQ(op_desc.Input("Input").size(), 1); @@ -31,6 +45,7 @@ class Conv2dOpConverter : public OpConverter { PADDLE_ENFORCE_EQ(op_desc.Output("Output").size(), 1); auto* X = engine_->GetITensor(op_desc.Input("Input").front()); + // Declare weights auto* Y_v = scope.FindVar(op_desc.Input("Filter").front()); PADDLE_ENFORCE_NOT_NULL(Y_v); @@ -83,7 +98,10 @@ class Conv2dOpConverter : public OpConverter { std::move(weight_tensor); layer->getOutput(0)->setName(output_name.c_str()); engine_->SetITensor(output_name, layer->getOutput(0)); - if (test_mode) { + + if (test_mode || + to_skip_merging_optimize(engine_, {filter_h, filter_w}, strides, + paddings, op_desc.Input("Input").front())) { engine_->DeclareOutput(output_name); } } diff --git a/paddle/fluid/inference/tensorrt/convert/dropout_op.cc b/paddle/fluid/inference/tensorrt/convert/dropout_op.cc index 9533ecbcfda4e2500fd201d8efc64fc5bd97169a..ddbc724e3b2a48b75df17f9bda691a1fd3883c32 100644 --- a/paddle/fluid/inference/tensorrt/convert/dropout_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/dropout_op.cc @@ -25,7 +25,7 @@ class DropoutOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, const framework::Scope& scope, bool test_mode) override { - VLOG(4) << "convert a fluid dropout op to tensorrt dropout layer"; + VLOG(3) << "convert a fluid dropout op to tensorrt dropout layer"; framework::OpDesc op_desc(op, nullptr); // Declare inputs auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); diff --git a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc index 0a6ce568f194f03c7259e1ebf28dd6ce4df2d594..671bcd8aa9a9fff34644a056499961cf6ab81287 100644 --- a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc @@ -26,7 +26,7 @@ class ElementwiseWeightOpConverter : public OpConverter { // Here the two nullptr looks strange, that's because the // framework::OpDesc's constructor is strange. framework::OpDesc op_desc(op, nullptr); - LOG(INFO) << "convert a fluid elementwise op to tensorrt IScaleLayer"; + VLOG(3) << "convert a fluid elementwise op to tensorrt IScaleLayer"; PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight @@ -108,7 +108,7 @@ class ElementwiseTensorOpConverter : public OpConverter { // Here the two nullptr looks strange, that's because the // framework::OpDesc's constructor is strange. framework::OpDesc op_desc(op, nullptr); - LOG(INFO) << "convert a fluid elementwise op to tensorrt IScaleLayer"; + VLOG(3) << "convert a fluid elementwise op to tensorrt IScaleLayer"; PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); PADDLE_ENFORCE_EQ(op_desc.Input("Y").size(), 1); // Y is a weight diff --git a/paddle/fluid/inference/tensorrt/convert/fc_op.cc b/paddle/fluid/inference/tensorrt/convert/fc_op.cc index 7c21ecd95da07b498eed2ab1bbdcc0e8cd184787..eef4fab4e86f05fa80bc614371f1aa43e433407e 100644 --- a/paddle/fluid/inference/tensorrt/convert/fc_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/fc_op.cc @@ -52,7 +52,7 @@ class FcOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, const framework::Scope& scope, bool test_mode) override { - VLOG(4) << "convert a fluid fc op to tensorrt fc layer without bias"; + VLOG(3) << "convert a fluid fc op to tensorrt fc layer without bias"; framework::OpDesc op_desc(op, nullptr); PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); diff --git a/paddle/fluid/inference/tensorrt/convert/mul_op.cc b/paddle/fluid/inference/tensorrt/convert/mul_op.cc index 514eb659a8da73b6e56b5d17148ec0cb2aeaa135..5b6aaad49833cedbd8d1ee0ec5d24c7f983190e6 100644 --- a/paddle/fluid/inference/tensorrt/convert/mul_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/mul_op.cc @@ -25,7 +25,7 @@ class MulOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, const framework::Scope& scope, bool test_mode) override { - VLOG(4) << "convert a fluid mul op to tensorrt mul layer without bias"; + VLOG(3) << "convert a fluid mul op to tensorrt mul layer without bias"; framework::OpDesc op_desc(op, nullptr); // Declare inputs diff --git a/paddle/fluid/inference/tensorrt/convert/pad_op.cc b/paddle/fluid/inference/tensorrt/convert/pad_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..4afcb0aecec9d07b52d2fd701fae8750067a6041 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/pad_op.cc @@ -0,0 +1,68 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +/* + * PadOp. + */ +class PadOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + VLOG(3) << "convert a fluid transpose op to tensorrt tranpose layer"; + + framework::OpDesc op_desc(op, nullptr); + // Declare inputs + auto* input = engine_->GetITensor(op_desc.Input("X")[0]); + + const std::vector paddings = + boost::get>(op_desc.GetAttr("paddings")); + const float pad_value = boost::get(op_desc.GetAttr("pad_value")); + + nvinfer1::Dims input_shape = input->getDimensions(); + int nbDims = input_shape.nbDims; + int pad_size = static_cast(paddings.size()); + PADDLE_ENFORCE_GE(nbDims, 2); + PADDLE_ENFORCE_EQ((nbDims + 1) * 2, pad_size); + PADDLE_ENFORCE(pad_value == 0.0, "The pad layer of TRT only support zero."); + + nvinfer1::DimsHW pre_pad(paddings[pad_size - 4], paddings[pad_size - 2]); + nvinfer1::DimsHW post_pad(paddings[pad_size - 3], paddings[pad_size - 1]); + + auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Padding, + *const_cast(input), + pre_pad, post_pad); + + PADDLE_ENFORCE(layer != nullptr); + auto output_name = op_desc.Output("Out")[0]; + engine_->SetITensor(output_name, layer->getOutput(0)); + layer->setName(("scale (Output: " + output_name + ")").c_str()); + layer->getOutput(0)->setName(output_name.c_str()); + if (test_mode) { // the test framework can not determine which is the + // output, so place the declaration inside. + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +REGISTER_TRT_OP_CONVERTER(pad, PadOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc index f9bb66a6e9f81a10368db7710108c319860e940a..48850020840a49bd309c007943f14b2f7eec5e2d 100644 --- a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc @@ -25,7 +25,7 @@ class Pool2dOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, const framework::Scope& scope, bool test_mode) override { - VLOG(4) + VLOG(3) << "convert a fluid pool2d op to tensorrt pool2d layer without bias"; framework::OpDesc op_desc(op, nullptr); // Declare inputs @@ -42,16 +42,22 @@ class Pool2dOpConverter : public OpConverter { boost::get>(op_desc.GetAttr("strides")); std::vector paddings = boost::get>(op_desc.GetAttr("paddings")); + bool ceil_mode = boost::get(op_desc.GetAttr("ceil_mode")); + nvinfer1::Dims input_shape = input1->getDimensions(); + int nbDims = input_shape.nbDims; nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]); + nvinfer1::DimsHW nv_strides(strides[0], strides[1]); + nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]); + if (global_pooling == true) { - nvinfer1::Dims input_shape = input1->getDimensions(); - int nbDims = input_shape.nbDims; nv_ksize.d[0] = input_shape.d[nbDims - 2]; nv_ksize.d[1] = input_shape.d[nbDims - 1]; + nv_strides.h() = 1; + nv_strides.w() = 1; + nv_paddings.h() = 0; + nv_paddings.w() = 0; } - const nvinfer1::DimsHW nv_strides(strides[0], strides[1]); - const nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]); PADDLE_ENFORCE_EQ(input1->getDimensions().nbDims, 3UL); @@ -64,6 +70,36 @@ class Pool2dOpConverter : public OpConverter { PADDLE_THROW("TensorRT unsupported pooling type!"); } + if (ceil_mode) { + nvinfer1::DimsHW pre_pad(0, 0); + nvinfer1::DimsHW post_pad(0, 0); + int input_height = input_shape.d[nbDims - 2]; + int input_width = input_shape.d[nbDims - 1]; + int floor_h_output_size = + (input_height - ksize[0] + 2 * paddings[0]) / strides[0] + 1; + int ceil_h_output_size = + (input_height - ksize[0] + 2 * paddings[0] + strides[0] - 1) / + strides[0] + + 1; + + int floor_w_output_size = + (input_width - ksize[1] + 2 * paddings[1]) / strides[1] + 1; + int ceil_w_output_size = + (input_width - ksize[1] + 2 * paddings[1] + strides[1] - 1) / + strides[1] + + 1; + if (floor_h_output_size != ceil_h_output_size) { + post_pad.h() = strides[0] - 1; + } + + if (floor_w_output_size != ceil_w_output_size) { + post_pad.w() = strides[1] - 1; + } + auto* layer = TRT_ENGINE_ADD_LAYER( + engine_, Padding, *const_cast(input1), pre_pad, + post_pad); + input1 = layer->getOutput(0); + } auto* layer = TRT_ENGINE_ADD_LAYER(engine_, Pooling, *const_cast(input1), nv_pool_type, nv_ksize); diff --git a/paddle/fluid/inference/tensorrt/convert/softmax_op.cc b/paddle/fluid/inference/tensorrt/convert/softmax_op.cc index 0064f90fd7944403c14d4d47616ea82f681ceb74..80bfb2d190a5637032e7c18fbac7f22b3a9e81e1 100644 --- a/paddle/fluid/inference/tensorrt/convert/softmax_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/softmax_op.cc @@ -25,7 +25,7 @@ class SoftMaxOpConverter : public OpConverter { public: void operator()(const framework::proto::OpDesc& op, const framework::Scope& scope, bool test_mode) override { - VLOG(4) + VLOG(3) << "convert a fluid softmax op to tensorrt softmax layer without bias"; framework::OpDesc op_desc(op, nullptr); // Declare inputs diff --git a/paddle/fluid/inference/tensorrt/convert/test_pad_op.cc b/paddle/fluid/inference/tensorrt/convert/test_pad_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ba35d7ddbb2f4e6062713bd82be277e7ad0cb341 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_pad_op.cc @@ -0,0 +1,52 @@ +/* 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 +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(PadConverter, main) { + framework::Scope scope; + std::unordered_set parameters; + TRTConvertValidation validator(10, parameters, scope, 1000); + validator.DeclInputVar("pad-X", nvinfer1::Dims3(3, 2, 2)); + validator.DeclOutputVar("pad-Out", nvinfer1::Dims3(3, 3, 5)); + + // Prepare Op description + framework::OpDesc desc; + desc.SetType("pad"); + desc.SetInput("X", {"pad-X"}); + desc.SetOutput("Out", {"pad-Out"}); + + std::vector paddings = {0, 0, 0, 0, 0, 1, 1, 2}; + float pad_value = 0.0; + desc.SetAttr("paddings", paddings); + desc.SetAttr("pad_value", pad_value); + + LOG(INFO) << "set OP"; + validator.SetOp(*desc.Proto()); + LOG(INFO) << "execute"; + + validator.Execute(2); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +USE_OP(pad); diff --git a/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc index aedd6b62df040eeee4e48f628128511cd8bf4439..ee597f8465c218c0fb6648374c128cabf7b033fb 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc @@ -20,18 +20,20 @@ namespace paddle { namespace inference { namespace tensorrt { -void test_pool2d(bool global_pooling) { +void test_pool2d(bool global_pooling, bool ceil_mode) { framework::Scope scope; std::unordered_set parameters; TRTConvertValidation validator(5, parameters, scope, 1 << 15); // The ITensor's Dims should not contain the batch size. // So, the ITensor's Dims of input and output should be C * H * W. - validator.DeclInputVar("pool2d-X", nvinfer1::Dims3(3, 4, 4)); + validator.DeclInputVar("pool2d-X", nvinfer1::Dims3(3, 13, 14)); if (global_pooling) validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 1, 1)); + else if (ceil_mode) + validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 6, 7)); else - validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 2, 2)); + validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 6, 6)); // Prepare Op description framework::OpDesc desc; @@ -39,7 +41,7 @@ void test_pool2d(bool global_pooling) { desc.SetInput("X", {"pool2d-X"}); desc.SetOutput("Out", {"pool2d-Out"}); - std::vector ksize({2, 2}); + std::vector ksize({3, 3}); std::vector strides({2, 2}); std::vector paddings({0, 0}); std::string pooling_t = "max"; @@ -49,6 +51,7 @@ void test_pool2d(bool global_pooling) { desc.SetAttr("strides", strides); desc.SetAttr("paddings", paddings); desc.SetAttr("global_pooling", global_pooling); + desc.SetAttr("ceil_mode", ceil_mode); LOG(INFO) << "set OP"; validator.SetOp(*desc.Proto()); @@ -57,9 +60,10 @@ void test_pool2d(bool global_pooling) { validator.Execute(3); } -TEST(Pool2dOpConverter, normal) { test_pool2d(false); } +TEST(Pool2dOpConverter, normal) { test_pool2d(false, false); } +TEST(Pool2dOpConverter, test_global_pooling) { test_pool2d(true, false); } -TEST(Pool2dOpConverter, test_global_pooling) { test_pool2d(true); } +TEST(Pool2dOpConverter, test_ceil_mode) { test_pool2d(false, true); } } // namespace tensorrt } // namespace inference diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index 14e9e14d33d637ee68e37593cc48721e5169499f..9e0f95844761db7571c5313726d34685a9aa66b2 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -133,6 +133,10 @@ void TensorRTEngine::DeclareOutput(const nvinfer1::ILayer *layer, int offset, buffer_sizes_[name] = 0; } +bool TensorRTEngine::HasDeclared(const std::string &name) { + return buffer_sizes_.count(name) > 0; +} + void TensorRTEngine::DeclareOutput(const std::string &name) { PADDLE_ENFORCE_EQ(0, buffer_sizes_.count(name), "duplicate output name %s", name); diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index bd3ba4cea6551a7f6651e311e2649de191a6faa1..828181200e300c370bbfa234c3c23ae44810878c 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -91,6 +91,8 @@ class TensorRTEngine : public EngineBase { const std::string& name); // Set the itensor_map_[name] as the network's output, and set its name. void DeclareOutput(const std::string& name); + // Check if the ITensor has been declared + bool HasDeclared(const std::string& name); // GPU memory address for an ITensor with specific name. One can operate on // these memory directly for acceleration, for example, output the converted @@ -132,6 +134,16 @@ class TensorRTEngine : public EngineBase { std::unordered_map> weight_map; + // TODO(NHZLX) + // In the normal case, the paddle-trt exists bug when runing the googlenet. + // When there are more than two convolutions of 1 * 1 with the same input, the + // paddle-tensorrt will do the merging optimization, which fuse those conv + // into + // one conv, and then trigger bug. So, We should use strategy to avoid this + // optimization for the time being. This bug will be fixed in the future. + std::unordered_map + itensor_quote_num; + private: // the max batch size int max_batch_; diff --git a/paddle/fluid/inference/tensorrt/helper.h b/paddle/fluid/inference/tensorrt/helper.h index b6e7968108403c9c9c192759c44eac040d1c5073..fc7ca7714e9325d2b6bce6189300aa339c81c2ba 100644 --- a/paddle/fluid/inference/tensorrt/helper.h +++ b/paddle/fluid/inference/tensorrt/helper.h @@ -52,7 +52,7 @@ class NaiveLogger : public nvinfer1::ILogger { void log(nvinfer1::ILogger::Severity severity, const char* msg) override { switch (severity) { case Severity::kINFO: - LOG(INFO) << msg; + VLOG(3) << msg; break; case Severity::kWARNING: LOG(WARNING) << msg; diff --git a/paddle/fluid/inference/tests/api/CMakeLists.txt b/paddle/fluid/inference/tests/api/CMakeLists.txt index c3dd1f433691e1c96e9f38ef7b595befad26408f..5287cd51cd2c339601b91b6a5e9ad4b9b1f5ee48 100644 --- a/paddle/fluid/inference/tests/api/CMakeLists.txt +++ b/paddle/fluid/inference/tests/api/CMakeLists.txt @@ -1,17 +1,9 @@ -set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com") -set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING - "A path setting inference demo download directories.") set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor) -function (inference_download install_dir url filename) - message(STATUS "Download inference test stuff from ${url}/${filename}") - execute_process(COMMAND bash -c "mkdir -p ${install_dir}") - execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}") - message(STATUS "finish downloading ${filename}") -endfunction() -function (inference_download_and_uncompress install_dir url filename) - inference_download(${install_dir} ${url} ${filename}) - execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}") +function(download_model install_dir model_name) + if (NOT EXISTS ${install_dir}) + inference_download_and_uncompress(${install_dir} ${INFERENCE_URL} ${model_name}) + endif() endfunction() function(download_model_and_data install_dir model_name data_name) @@ -27,6 +19,13 @@ function(inference_analysis_api_test target install_dir filename) ARGS --infer_model=${install_dir}/model --infer_data=${install_dir}/data.txt) endfunction() +function(inference_analysis_api_test_with_fake_data target install_dir filename model_name) + download_model(${install_dir} ${model_name}) + inference_analysis_test(${target} SRCS ${filename} + EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} + ARGS --infer_model=${install_dir}/model) +endfunction() + # RNN1 if(NOT APPLE) set(RNN1_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn1") @@ -43,6 +42,15 @@ set(RNN2_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/rnn2") download_model_and_data(${RNN2_INSTALL_DIR} "rnn2_model.tar.gz" "rnn2_data.txt.tar.gz") inference_analysis_api_test(test_analyzer_rnn2 ${RNN2_INSTALL_DIR} analyzer_rnn2_tester.cc) +# DAM +set(DAM_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/dam") +download_model_and_data(${DAM_INSTALL_DIR} "DAM_model.tar.gz" "DAM_data.txt.tar.gz") +inference_analysis_test(test_analyzer_dam SRCS analyzer_dam_tester.cc + EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} ARGS + --infer_model=${DAM_INSTALL_DIR}/model + --infer_data=${DAM_INSTALL_DIR}/data.txt + --use_analysis=0) + # chinese_ner set(CHINESE_NER_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/chinese_ner") download_model_and_data(${CHINESE_NER_INSTALL_DIR} "chinese_ner_model.tar.gz" "chinese_ner-data.txt.tar.gz") @@ -66,17 +74,13 @@ inference_analysis_api_test(test_analyzer_seq_conv1 ${SEQ_CONV1_INSTALL_DIR} ana # ocr set(OCR_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/ocr") if (NOT EXISTS ${OCR_INSTALL_DIR}) - inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz") + inference_download_and_uncompress(${OCR_INSTALL_DIR} "http://paddlemodels.cdn.bcebos.com/" "inference-vis-demos%2Focr.tar.gz") endif() inference_analysis_api_test(test_analyzer_ocr ${OCR_INSTALL_DIR} analyzer_vis_tester.cc) # resnet50 -set(RESNET50_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/resnet50") -if (NOT EXISTS ${RESNET50_INSTALL_DIR}) - inference_download_and_uncompress(${RESNET50_INSTALL_DIR} ${INFERENCE_URL} "resnet50_model.tar.gz") -endif() -inference_analysis_test(test_analyzer_resnet50 SRCS analyzer_resnet50_tester.cc - EXTRA_DEPS ${INFERENCE_EXTRA_DEPS} ARGS --infer_model=${RESNET50_INSTALL_DIR}/model) +inference_analysis_api_test_with_fake_data(test_analyzer_resnet50 + "${INFERENCE_DEMO_INSTALL_DIR}/resnet50" analyzer_resnet50_tester.cc "resnet50_model.tar.gz") # anakin if (WITH_ANAKIN AND WITH_MKL) # only needed in CI @@ -106,5 +110,5 @@ if(WITH_GPU AND TENSORRT_FOUND) endif() cc_test(test_trt_models SRCS trt_models_tester.cc ARGS --dirname=${TRT_MODEL_INSTALL_DIR}/trt_test_models - DEPS paddle_inference_tensorrt_subgraph_engine) + DEPS paddle_inference_tensorrt_subgraph_engine SERIAL) endif() diff --git a/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc b/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc new file mode 100644 index 0000000000000000000000000000000000000000..ceac5dc7e14365c77cf1cbbbc16e4bf3ebfced73 --- /dev/null +++ b/paddle/fluid/inference/tests/api/analyzer_dam_tester.cc @@ -0,0 +1,224 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/tests/api/tester_helper.h" + +namespace paddle { +namespace inference { +using contrib::AnalysisConfig; +#define MAX_TURN_NUM 9 +#define MAX_TURN_LEN 50 +static std::vector result_data; + +struct DataRecord { + std::vector> + turns[MAX_TURN_NUM]; // turns data : MAX_TURN_NUM + std::vector> + turns_mask[MAX_TURN_NUM]; // turns mask data : MAX_TURN_NUM + std::vector> response; // response data : 1 + std::vector> response_mask; // response mask data : 1 + size_t batch_iter{0}; + size_t batch_size{1}; + size_t num_samples; // total number of samples + DataRecord() = default; + explicit DataRecord(const std::string &path, int batch_size = 1) + : batch_size(batch_size) { + Load(path); + } + DataRecord NextBatch() { + DataRecord data; + size_t batch_end = batch_iter + batch_size; + // NOTE skip the final batch, if no enough data is provided. + if (batch_end <= response.size()) { + for (int i = 0; i < MAX_TURN_NUM; ++i) { + data.turns[i].assign(turns[i].begin() + batch_iter, + turns[i].begin() + batch_end); + } + for (int i = 0; i < MAX_TURN_NUM; ++i) { + data.turns_mask[i].assign(turns_mask[i].begin() + batch_iter, + turns_mask[i].begin() + batch_end); + } + data.response.assign(response.begin() + batch_iter, + response.begin() + batch_end); + data.response_mask.assign(response_mask.begin() + batch_iter, + response_mask.begin() + batch_end); + CHECK(!data.response.empty()); + CHECK(!data.response_mask.empty()); + CHECK_EQ(data.response.size(), data.response_mask.size()); + } + batch_iter += batch_size; + return data; + } + void Load(const std::string &path) { + std::ifstream file(path); + std::string line; + size_t num_lines = 0; + result_data.clear(); + while (std::getline(file, line)) { + num_lines++; + std::vector data; + split(line, ',', &data); + CHECK_EQ(data.size(), 2 * MAX_TURN_NUM + 3); + // load turn data + std::vector turns_tmp[MAX_TURN_NUM]; + for (int i = 0; i < MAX_TURN_NUM; ++i) { + split_to_int64(data[i], ' ', &turns_tmp[i]); + turns[i].push_back(std::move(turns_tmp[i])); + } + // load turn_mask data + std::vector turns_mask_tmp[MAX_TURN_NUM]; + for (int i = 0; i < MAX_TURN_NUM; ++i) { + split_to_float(data[MAX_TURN_NUM + i], ' ', &turns_mask_tmp[i]); + turns_mask[i].push_back(std::move(turns_mask_tmp[i])); + } + // load response data + std::vector response_tmp; + split_to_int64(data[2 * MAX_TURN_NUM], ' ', &response_tmp); + response.push_back(std::move(response_tmp)); + // load response_mask data + std::vector response_mask_tmp; + split_to_float(data[2 * MAX_TURN_NUM + 1], ' ', &response_mask_tmp); + response_mask.push_back(std::move(response_mask_tmp)); + // load result data + float result_tmp; + result_tmp = std::stof(data[2 * MAX_TURN_NUM + 2]); + result_data.push_back(result_tmp); + } + num_samples = num_lines; + } +}; + +void PrepareInputs(std::vector *input_slots, DataRecord *data, + int batch_size) { + PaddleTensor turns_tensor[MAX_TURN_NUM]; + PaddleTensor turns_mask_tensor[MAX_TURN_NUM]; + PaddleTensor response_tensor; + PaddleTensor response_mask_tensor; + std::string turn_pre = "turn_"; + std::string turn_mask_pre = "turn_mask_"; + + auto one_batch = data->NextBatch(); + int size = one_batch.response[0].size(); + CHECK_EQ(size, MAX_TURN_LEN); + // turn tensor assignment + for (int i = 0; i < MAX_TURN_NUM; ++i) { + turns_tensor[i].name = turn_pre + std::to_string(i); + turns_tensor[i].shape.assign({batch_size, size, 1}); + turns_tensor[i].dtype = PaddleDType::INT64; + TensorAssignData(&turns_tensor[i], one_batch.turns[i]); + } + // turn mask tensor assignment + for (int i = 0; i < MAX_TURN_NUM; ++i) { + turns_mask_tensor[i].name = turn_mask_pre + std::to_string(i); + turns_mask_tensor[i].shape.assign({batch_size, size, 1}); + turns_mask_tensor[i].dtype = PaddleDType::FLOAT32; + TensorAssignData(&turns_mask_tensor[i], one_batch.turns_mask[i]); + } + // response tensor assignment + response_tensor.name = "response"; + response_tensor.shape.assign({batch_size, size, 1}); + response_tensor.dtype = PaddleDType::INT64; + TensorAssignData(&response_tensor, one_batch.response); + // response mask tensor assignment + response_mask_tensor.name = "response_mask"; + response_mask_tensor.shape.assign({batch_size, size, 1}); + response_mask_tensor.dtype = PaddleDType::FLOAT32; + TensorAssignData(&response_mask_tensor, one_batch.response_mask); + + // Set inputs. + for (int i = 0; i < MAX_TURN_NUM; ++i) { + input_slots->push_back(std::move(turns_tensor[i])); + } + for (int i = 0; i < MAX_TURN_NUM; ++i) { + input_slots->push_back(std::move(turns_mask_tensor[i])); + } + input_slots->push_back(std::move(response_tensor)); + input_slots->push_back(std::move(response_mask_tensor)); +} + +void SetConfig(contrib::AnalysisConfig *cfg) { + cfg->prog_file = FLAGS_infer_model + "/__model__"; + cfg->param_file = FLAGS_infer_model + "/param"; + cfg->use_gpu = false; + cfg->device = 0; + cfg->specify_input_name = true; + cfg->enable_ir_optim = true; +} + +void SetInput(std::vector> *inputs) { + DataRecord data(FLAGS_infer_data, FLAGS_batch_size); + std::vector input_slots; + int test_batch_num = + FLAGS_test_all_data ? data.num_samples / FLAGS_batch_size : 1; + LOG(INFO) << "The number of samples to be test: " + << test_batch_num * FLAGS_batch_size; + for (int bid = 0; bid < test_batch_num; ++bid) { + input_slots.clear(); + PrepareInputs(&input_slots, &data, FLAGS_batch_size); + (*inputs).emplace_back(input_slots); + } +} + +// Easy for profiling independently. +TEST(Analyzer_dam, profile) { + contrib::AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector outputs; + std::vector> input_slots_all; + SetInput(&input_slots_all); + TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads); + + if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { + PADDLE_ENFORCE_GT(outputs.size(), 0); + size_t size = GetSize(outputs[0]); + PADDLE_ENFORCE_GT(size, 0); + float *result = static_cast(outputs[0].data.data()); + for (size_t i = 0; i < size; i++) { + EXPECT_NEAR(result[i], result_data[i], 1e-3); + } + } +} + +// Check the fuse status +TEST(Analyzer_dam, fuse_statis) { + contrib::AnalysisConfig cfg; + SetConfig(&cfg); + + if (FLAGS_use_analysis) { + int num_ops; + auto predictor = CreatePaddlePredictor(cfg); + auto fuse_statis = GetFuseStatis( + static_cast(predictor.get()), &num_ops); + ASSERT_TRUE(fuse_statis.count("fc_fuse")); + EXPECT_EQ(fuse_statis.at("fc_fuse"), 317); + EXPECT_EQ(num_ops, 2020); + } +} + +// Compare result of NativeConfig and AnalysisConfig +TEST(Analyzer_dam, compare) { + contrib::AnalysisConfig cfg; + SetConfig(&cfg); + + std::vector> input_slots_all; + SetInput(&input_slots_all); + + if (FLAGS_use_analysis) { + CompareNativeAndAnalysis(cfg, input_slots_all); + } +} + +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc b/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc index 577b97e271aacab5d6740de7c8bc00bc87ae54dd..d91f7c314d0a936da6f5b0c41920c905af5cd0ee 100644 --- a/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_ner_tester.cc @@ -20,7 +20,6 @@ using contrib::AnalysisConfig; struct DataRecord { std::vector> word_data_all, mention_data_all; - std::vector> rnn_word_datas, rnn_mention_datas; std::vector lod; // two inputs have the same lod info. size_t batch_iter{0}; size_t batch_size{1}; @@ -45,8 +44,6 @@ struct DataRecord { CHECK(!data.mention_data_all.empty()); CHECK_EQ(data.word_data_all.size(), data.mention_data_all.size()); for (size_t j = 0; j < data.word_data_all.size(); j++) { - data.rnn_word_datas.push_back(data.word_data_all[j]); - data.rnn_mention_datas.push_back(data.mention_data_all[j]); // calculate lod data.lod.push_back(data.lod.back() + data.word_data_all[j].size()); } @@ -87,8 +84,8 @@ void PrepareInputs(std::vector *input_slots, DataRecord *data, lod_mention_tensor.shape.assign({size, 1}); lod_mention_tensor.lod.assign({one_batch.lod}); // assign data - TensorAssignData(&lod_word_tensor, one_batch.rnn_word_datas); - TensorAssignData(&lod_mention_tensor, one_batch.rnn_mention_datas); + TensorAssignData(&lod_word_tensor, one_batch.word_data_all); + TensorAssignData(&lod_mention_tensor, one_batch.mention_data_all); // Set inputs. input_slots->assign({lod_word_tensor, lod_mention_tensor}); for (auto &tensor : *input_slots) { diff --git a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc index 290fb007d8ba94a2d121947fe67c6474586ac0e0..e5c8dfd22a006d5271248c5b083aab2c22417502 100644 --- a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc @@ -30,45 +30,26 @@ void SetConfig(AnalysisConfig *cfg) { } void SetInput(std::vector> *inputs) { - PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data."); - - PaddleTensor input; - // channel=3, height/width=318 - std::vector shape({FLAGS_batch_size, 3, 318, 318}); - input.shape = shape; - input.dtype = PaddleDType::FLOAT32; - - // fill input data, for profile easily, do not use random data here. - size_t size = FLAGS_batch_size * 3 * 318 * 318; - input.data.Resize(size * sizeof(float)); - float *input_data = static_cast(input.data.data()); - for (size_t i = 0; i < size; i++) { - *(input_data + i) = static_cast(i) / size; - } - - std::vector input_slots; - input_slots.assign({input}); - (*inputs).emplace_back(input_slots); + SetFakeImageInput(inputs, FLAGS_infer_model); } // Easy for profiling independently. -TEST(Analyzer_resnet50, profile) { +void profile(bool use_mkldnn = false) { AnalysisConfig cfg; SetConfig(&cfg); + cfg._use_mkldnn = use_mkldnn; std::vector outputs; std::vector> input_slots_all; SetInput(&input_slots_all); TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads); - - if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { - PADDLE_ENFORCE_EQ(outputs.size(), 1UL); - size_t size = GetSize(outputs[0]); - // output is a 512-dimension feature - EXPECT_EQ(size, 512 * FLAGS_batch_size); - } } +TEST(Analyzer_resnet50, profile) { profile(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_resnet50, profile_mkldnn) { profile(true /* use_mkldnn */); } +#endif + // Check the fuse status TEST(Analyzer_resnet50, fuse_statis) { AnalysisConfig cfg; @@ -77,20 +58,25 @@ TEST(Analyzer_resnet50, fuse_statis) { auto predictor = CreatePaddlePredictor(cfg); auto fuse_statis = GetFuseStatis( static_cast(predictor.get()), &num_ops); - ASSERT_TRUE(fuse_statis.count("fc_fuse")); - EXPECT_EQ(fuse_statis.at("fc_fuse"), 1); + LOG(INFO) << "num_ops: " << num_ops; } // Compare result of NativeConfig and AnalysisConfig -TEST(Analyzer_resnet50, compare) { +void compare(bool use_mkldnn = false) { AnalysisConfig cfg; SetConfig(&cfg); + cfg._use_mkldnn = use_mkldnn; std::vector> input_slots_all; SetInput(&input_slots_all); CompareNativeAndAnalysis(cfg, input_slots_all); } +TEST(Analyzer_resnet50, compare) { compare(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_resnet50, compare_mkldnn) { compare(true /* use_mkldnn */); } +#endif + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc index c76d72ccd99649913aefcb2aa57fe6061db8ca6d..e0416ff953b61f56a2ca1a45cb382d40a6cffa4a 100644 --- a/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc @@ -228,6 +228,7 @@ void SetInput(std::vector> *inputs) { TEST(Analyzer_rnn1, profile) { contrib::AnalysisConfig cfg; SetConfig(&cfg); + cfg.use_gpu = false; std::vector outputs; std::vector> input_slots_all; @@ -308,18 +309,13 @@ TEST(Analyzer_rnn1, ZeroCopy) { PaddlePlace place; int output_size{0}; - auto predictor = - CreatePaddlePredictor( - config); + auto predictor = CreatePaddlePredictor(config); config.use_feed_fetch_ops = true; - auto native_predictor = - CreatePaddlePredictor(config); + auto native_predictor = CreatePaddlePredictor(config); config.use_feed_fetch_ops = true; // the analysis predictor needs feed/fetch. - auto analysis_predictor = - CreatePaddlePredictor( - config); + auto analysis_predictor = CreatePaddlePredictor(config); #define NEW_TENSOR(name__) \ auto name__##_tensor = predictor->GetInputTensor(#name__); diff --git a/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc b/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc index ba04d030b94c0924311dcff5c6a34270a764f877..e0eb919bd896d73a557001982a436fc93f087a74 100644 --- a/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_rnn2_tester.cc @@ -18,12 +18,12 @@ namespace paddle { namespace inference { using namespace framework; // NOLINT +static std::vector result_data; struct DataRecord { std::vector>> link_step_data_all; std::vector lod; std::vector> rnn_link_data; - std::vector result_data; size_t num_samples; // total number of samples size_t batch_iter{0}; size_t batch_size{1}; @@ -57,6 +57,7 @@ struct DataRecord { std::ifstream file(path); std::string line; int num_lines = 0; + result_data.clear(); while (std::getline(file, line)) { num_lines++; std::vector data; @@ -135,13 +136,12 @@ TEST(Analyzer_rnn2, profile) { if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { // the first inference result - DataRecord data(FLAGS_infer_data, FLAGS_batch_size); PADDLE_ENFORCE_GT(outputs.size(), 0); size_t size = GetSize(outputs[0]); PADDLE_ENFORCE_GT(size, 0); float *result = static_cast(outputs[0].data.data()); for (size_t i = 0; i < size; i++) { - EXPECT_NEAR(result[i], data.result_data[i], 1e-3); + EXPECT_NEAR(result[i], result_data[i], 1e-3); } } } diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc index cb4671c4379b5f6f144bfd5330866aa38163f4d4..f590ef27967e47ffcb3a97e80dd147efdd1906e6 100644 --- a/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_seq_conv1_tester.cc @@ -183,7 +183,13 @@ TEST(Analyzer_seq_conv1, fuse_statis) { SetConfig(&cfg); int num_ops; auto predictor = CreatePaddlePredictor(cfg); - GetFuseStatis(predictor.get(), &num_ops); + + auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops); + ASSERT_TRUE(fuse_statis.count("fc_fuse")); + ASSERT_TRUE(fuse_statis.count("seqconv_eltadd_relu_fuse")); + EXPECT_EQ(fuse_statis.at("fc_fuse"), 2); + EXPECT_EQ(fuse_statis.at("seqconv_eltadd_relu_fuse"), 6); + EXPECT_EQ(num_ops, 32); } // Compare result of NativeConfig and AnalysisConfig diff --git a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc index 305b8bfe158150d5dfd8bdaee2c0a89afe264de4..b2cd49af9aa580482fad84b6b23cb19f954e22fc 100644 --- a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc @@ -27,7 +27,7 @@ struct Record { }; Record ProcessALine(const std::string &line) { - VLOG(3) << "process a line"; + VLOG(30) << "process a line"; std::vector columns; split(line, '\t', &columns); CHECK_EQ(columns.size(), 2UL) @@ -45,8 +45,8 @@ Record ProcessALine(const std::string &line) { for (auto &s : shape_strs) { record.shape.push_back(std::stoi(s)); } - VLOG(3) << "data size " << record.data.size(); - VLOG(3) << "data shape size " << record.shape.size(); + VLOG(30) << "data size " << record.data.size(); + VLOG(30) << "data shape size " << record.shape.size(); return record; } @@ -59,9 +59,6 @@ void SetConfig(AnalysisConfig *cfg) { cfg->specify_input_name = true; // TODO(TJ): fix fusion gru cfg->ir_passes.push_back("fc_gru_fuse_pass"); -#ifdef PADDLE_WITH_MKLDNN - cfg->_use_mkldnn = true; -#endif } void SetInput(std::vector> *inputs) { @@ -84,9 +81,10 @@ void SetInput(std::vector> *inputs) { // Easy for profiling independently. // ocr, mobilenet and se_resnext50 -TEST(Analyzer_vis, profile) { +void profile(bool use_mkldnn = false) { AnalysisConfig cfg; SetConfig(&cfg); + cfg._use_mkldnn = use_mkldnn; std::vector outputs; std::vector> input_slots_all; @@ -108,6 +106,12 @@ TEST(Analyzer_vis, profile) { } } +TEST(Analyzer_vis, profile) { profile(); } + +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_vis, profile_mkldnn) { profile(true /* use_mkldnn */); } +#endif + // Check the fuse status TEST(Analyzer_vis, fuse_statis) { AnalysisConfig cfg; @@ -118,15 +122,21 @@ TEST(Analyzer_vis, fuse_statis) { } // Compare result of NativeConfig and AnalysisConfig -TEST(Analyzer_vis, compare) { +void compare(bool use_mkldnn = false) { AnalysisConfig cfg; SetConfig(&cfg); + cfg._use_mkldnn = use_mkldnn; std::vector> input_slots_all; SetInput(&input_slots_all); CompareNativeAndAnalysis(cfg, input_slots_all); } +TEST(Analyzer_vis, compare) { compare(); } +#ifdef PADDLE_WITH_MKLDNN +TEST(Analyzer_vis, compare_mkldnn) { compare(true /* use_mkldnn */); } +#endif + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tests/api/tester_helper.h b/paddle/fluid/inference/tests/api/tester_helper.h index 8603d09cbd09e4cee254faf52c2ccbe8d661994d..8c5888d8da7b33eeca77311c10dd818648e8e524 100644 --- a/paddle/fluid/inference/tests/api/tester_helper.h +++ b/paddle/fluid/inference/tests/api/tester_helper.h @@ -25,6 +25,7 @@ #include "paddle/fluid/inference/api/analysis_predictor.h" #include "paddle/fluid/inference/api/helper.h" #include "paddle/fluid/inference/api/paddle_inference_pass.h" +#include "paddle/fluid/inference/tests/test_helper.h" #include "paddle/fluid/platform/profiler.h" DEFINE_string(infer_model, "", "model path"); @@ -50,7 +51,7 @@ void CompareResult(const std::vector &outputs, auto &ref_out = ref_outputs[i]; size_t size = VecReduceToInt(out.shape); size_t ref_size = VecReduceToInt(ref_out.shape); - EXPECT_GT(size, 0); + EXPECT_GT(size, 0UL); EXPECT_EQ(size, ref_size); EXPECT_EQ(out.dtype, ref_out.dtype); switch (out.dtype) { @@ -77,11 +78,9 @@ void CompareResult(const std::vector &outputs, std::unique_ptr CreateTestPredictor( const AnalysisConfig &config, bool use_analysis = true) { if (use_analysis) { - return CreatePaddlePredictor(config); + return CreatePaddlePredictor(config); } else { - return CreatePaddlePredictor( - config); + return CreatePaddlePredictor(config); } } @@ -107,6 +106,34 @@ std::unordered_map GetFuseStatis(PaddlePredictor *predictor, return fuse_statis; } +void SetFakeImageInput(std::vector> *inputs, + const std::string &dirname) { + // Set fake_image_data + PADDLE_ENFORCE_EQ(FLAGS_test_all_data, 0, "Only have single batch of data."); + std::vector> feed_target_shapes = + GetFeedTargetShapes(dirname, true, "model", "params"); + int dim1 = feed_target_shapes[0][1]; + int dim2 = feed_target_shapes[0][2]; + int dim3 = feed_target_shapes[0][3]; + + PaddleTensor input; + std::vector shape({FLAGS_batch_size, dim1, dim2, dim3}); + input.shape = shape; + input.dtype = PaddleDType::FLOAT32; + + // fill input data, for profile easily, do not use random data here. + size_t size = FLAGS_batch_size * dim1 * dim2 * dim3; + input.data.Resize(size * sizeof(float)); + float *input_data = static_cast(input.data.data()); + for (size_t i = 0; i < size; i++) { + *(input_data + i) = static_cast(i) / size; + } + + std::vector input_slots; + input_slots.assign({input}); + (*inputs).emplace_back(input_slots); +} + void TestOneThreadPrediction( const AnalysisConfig &config, const std::vector> &inputs, @@ -141,6 +168,9 @@ void TestMultiThreadPrediction( } for (int tid = 0; tid < num_threads; ++tid) { threads.emplace_back([&, tid]() { +#ifdef PADDLE_WITH_MKLDNN + platform::set_cur_thread_id(static_cast(tid) + 1); +#endif // Each thread should have local inputs and outputs. // The inputs of each thread are all the same. std::vector> inputs_tid = inputs; @@ -165,7 +195,8 @@ void TestPrediction(const AnalysisConfig &config, const std::vector> &inputs, std::vector *outputs, int num_threads, bool use_analysis = FLAGS_use_analysis) { - LOG(INFO) << "use_analysis: " << use_analysis; + LOG(INFO) << "use_analysis: " << use_analysis + << ", use_mkldnn: " << config._use_mkldnn; if (num_threads == 1) { TestOneThreadPrediction(config, inputs, outputs, use_analysis); } else { @@ -177,6 +208,7 @@ void TestPrediction(const AnalysisConfig &config, void CompareNativeAndAnalysis( const AnalysisConfig &config, const std::vector> &inputs) { + LOG(INFO) << "use_mkldnn: " << config._use_mkldnn; std::vector native_outputs, analysis_outputs; TestOneThreadPrediction(config, inputs, &native_outputs, false); TestOneThreadPrediction(config, inputs, &analysis_outputs, true); diff --git a/paddle/fluid/inference/tests/api/trt_models_tester.cc b/paddle/fluid/inference/tests/api/trt_models_tester.cc index bf320a0cbc2fff5f973c48768281e26d0fde232b..75840a9c437d956da4f542a38b2532ea20ee96c5 100644 --- a/paddle/fluid/inference/tests/api/trt_models_tester.cc +++ b/paddle/fluid/inference/tests/api/trt_models_tester.cc @@ -51,11 +51,8 @@ void CompareTensorRTWithFluid(int batch_size, std::string model_dirname) { config1.model_dir = model_dirname; config1.max_batch_size = batch_size; - auto predictor0 = - CreatePaddlePredictor(config0); - auto predictor1 = - CreatePaddlePredictor(config1); + auto predictor0 = CreatePaddlePredictor(config0); + auto predictor1 = CreatePaddlePredictor(config1); // Prepare inputs int height = 224; int width = 224; @@ -96,11 +93,16 @@ void CompareTensorRTWithFluid(int batch_size, std::string model_dirname) { } } -TEST(trt_models_test, main) { - std::vector infer_models = {"mobilenet", "resnet50", - "resnext50"}; - for (auto &model_dir : infer_models) { - CompareTensorRTWithFluid(1, FLAGS_dirname + "/" + model_dir); - } +TEST(trt_models_test, mobilenet) { + CompareTensorRTWithFluid(1, FLAGS_dirname + "/mobilenet"); +} + +TEST(trt_models_test, resnet50) { + CompareTensorRTWithFluid(1, FLAGS_dirname + "/resnet50"); } + +TEST(trt_models_test, resnext50) { + CompareTensorRTWithFluid(1, FLAGS_dirname + "/resnext50"); +} + } // namespace paddle diff --git a/paddle/fluid/inference/tests/test.cmake b/paddle/fluid/inference/tests/test.cmake new file mode 100644 index 0000000000000000000000000000000000000000..ab3a30ce6bba14a7d5ec700a159d90031e6b5dc7 --- /dev/null +++ b/paddle/fluid/inference/tests/test.cmake @@ -0,0 +1,31 @@ +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}") + execute_process(COMMAND bash -c "mkdir -p ${install_dir}") + execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}") + message(STATUS "finish downloading ${filename}") +endfunction() + +function (inference_download_and_uncompress install_dir url filename) + inference_download(${install_dir} ${url} ${filename}) + execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}") +endfunction() + +set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec") +if (NOT EXISTS ${WORD2VEC_INSTALL_DIR}) + inference_download_and_uncompress(${WORD2VEC_INSTALL_DIR} ${INFERENCE_URL} "word2vec.inference.model.tar.gz") +endif() +set(WORD2VEC_MODEL_DIR "${WORD2VEC_INSTALL_DIR}/word2vec.inference.model") + +function (inference_base_test TARGET) + set(options "") + set(oneValueArgs "") + set(multiValueArgs SRCS ARGS DEPS) + cmake_parse_arguments(base_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + if(WITH_GPU) + set(mem_opt "--fraction_of_gpu_memory_to_use=0.5") + endif() + cc_test(${TARGET} SRCS ${base_test_SRCS} DEPS ${base_test_DEPS} ARGS ${mem_opt} ${base_test_ARGS}) +endfunction() diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 94f0550df57e79fa68c135f5c9c4b7effe6ac156..2118fcfd4bb1589947617e462f09971fcc090b98 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -18,7 +18,6 @@ limitations under the License. */ #include #include -#include "paddle/fluid/framework/ir/graph_to_program_pass.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/inference/io.h" #include "paddle/fluid/platform/profiler.h" @@ -94,15 +93,15 @@ void CheckError(const paddle::framework::LoDTensor& output1, std::unique_ptr InitProgram( paddle::framework::Executor* executor, paddle::framework::Scope* scope, - const std::string& dirname, const bool is_combined = false) { + const std::string& dirname, const bool is_combined = false, + const std::string& prog_filename = "__model_combined__", + const std::string& param_filename = "__params_combined__") { std::unique_ptr inference_program; if (is_combined) { // All parameters are saved in a single file. // Hard-coding the file names of program and parameters in unittest. // The file names should be consistent with that used in Python API // `fluid.io.save_inference_model`. - std::string prog_filename = "__model_combined__"; - std::string param_filename = "__params_combined__"; inference_program = paddle::inference::Load(executor, scope, dirname + "/" + prog_filename, dirname + "/" + param_filename); @@ -115,12 +114,15 @@ std::unique_ptr InitProgram( } std::vector> GetFeedTargetShapes( - const std::string& dirname, const bool is_combined = false) { + const std::string& dirname, const bool is_combined = false, + const std::string& prog_filename = "__model_combined__", + const std::string& param_filename = "__params_combined__") { auto place = paddle::platform::CPUPlace(); auto executor = paddle::framework::Executor(place); auto* scope = new paddle::framework::Scope(); - auto inference_program = InitProgram(&executor, scope, dirname, is_combined); + auto inference_program = InitProgram(&executor, scope, dirname, is_combined, + prog_filename, param_filename); auto& global_block = inference_program->Block(0); const std::vector& feed_target_names = @@ -136,15 +138,6 @@ std::vector> GetFeedTargetShapes( return feed_target_shapes; } -void Compile(paddle::framework::ProgramDesc* program) { - std::unique_ptr g( - new paddle::framework::ir::Graph(*program)); - auto pass = paddle::framework::ir::PassRegistry::Instance().Get( - "graph_to_program_pass"); - pass->SetNotOwned("program", program); - pass->Apply(std::move(g)); -} - template void TestInference(const std::string& dirname, const std::vector& cpu_feeds, @@ -182,7 +175,6 @@ void TestInference(const std::string& dirname, paddle::platform::DeviceContextPool::Instance().Get(place)); inference_program = InitProgram(&executor, scope, dirname, is_combined); } - Compile(inference_program.get()); // Disable the profiler and print the timing information paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, @@ -261,5 +253,3 @@ void TestInference(const std::string& dirname, delete scope; } - -USE_PASS(graph_to_program_pass); diff --git a/paddle/fluid/memory/detail/buddy_allocator.cc b/paddle/fluid/memory/detail/buddy_allocator.cc index 26ef27c3caafadb4801b0ae52133f6175655ce0a..dd7ffaa26426edebd47ec3f6fb275ad5a2d23322 100644 --- a/paddle/fluid/memory/detail/buddy_allocator.cc +++ b/paddle/fluid/memory/detail/buddy_allocator.cc @@ -32,11 +32,11 @@ BuddyAllocator::BuddyAllocator( system_allocator_(std::move(system_allocator)) {} BuddyAllocator::~BuddyAllocator() { - VLOG(10) << "BuddyAllocator Disconstructor makes sure that all of these " - "have actually been freed"; + VLOG(100) << "BuddyAllocator Disconstructor makes sure that all of these " + "have actually been freed"; while (!pool_.empty()) { auto block = static_cast(std::get<2>(*pool_.begin())); - VLOG(10) << "Free from block (" << block << ", " << max_chunk_size_ << ")"; + VLOG(100) << "Free from block (" << block << ", " << max_chunk_size_ << ")"; system_allocator_->Free(block, max_chunk_size_, block->index(cache_)); cache_.invalidate(block); @@ -57,12 +57,12 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) { // acquire the allocator lock std::lock_guard lock(mutex_); - VLOG(10) << "Allocate " << unaligned_size << " bytes from chunk size " - << size; + VLOG(100) << "Allocate " << unaligned_size << " bytes from chunk size " + << size; // if the allocation is huge, send directly to the system allocator if (size > max_chunk_size_) { - VLOG(10) << "Allocate from system allocator."; + VLOG(100) << "Allocate from system allocator."; return SystemAlloc(size); } @@ -77,9 +77,9 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) { return nullptr; } } else { - VLOG(10) << "Allocation from existing memory block " << std::get<2>(*it) - << " at address " - << reinterpret_cast(std::get<2>(*it))->data(); + VLOG(100) << "Allocation from existing memory block " << std::get<2>(*it) + << " at address " + << reinterpret_cast(std::get<2>(*it))->data(); } total_used_ += size; @@ -96,10 +96,10 @@ void BuddyAllocator::Free(void* p) { // Acquire the allocator lock std::lock_guard lock(mutex_); - VLOG(10) << "Free from address " << block; + VLOG(100) << "Free from address " << block; if (block->type(cache_) == MemoryBlock::HUGE_CHUNK) { - VLOG(10) << "Free directly from system allocator"; + VLOG(100) << "Free directly from system allocator"; system_allocator_->Free(block, block->total_size(cache_), block->index(cache_)); @@ -116,8 +116,8 @@ void BuddyAllocator::Free(void* p) { // Trying to merge the right buddy if (block->has_right_buddy(cache_)) { - VLOG(10) << "Merging this block " << block << " with its right buddy " - << block->right_buddy(cache_); + VLOG(100) << "Merging this block " << block << " with its right buddy " + << block->right_buddy(cache_); auto right_buddy = block->right_buddy(cache_); @@ -134,8 +134,8 @@ void BuddyAllocator::Free(void* p) { // Trying to merge the left buddy if (block->has_left_buddy(cache_)) { - VLOG(10) << "Merging this block " << block << " with its left buddy " - << block->left_buddy(cache_); + VLOG(100) << "Merging this block " << block << " with its left buddy " + << block->left_buddy(cache_); auto left_buddy = block->left_buddy(cache_); @@ -151,8 +151,8 @@ void BuddyAllocator::Free(void* p) { } // Dumping this block into pool - VLOG(10) << "Inserting free block (" << block << ", " - << block->total_size(cache_) << ")"; + VLOG(100) << "Inserting free block (" << block << ", " + << block->total_size(cache_) << ")"; pool_.insert( IndexSizeAddress(block->index(cache_), block->total_size(cache_), block)); @@ -174,7 +174,7 @@ void* BuddyAllocator::SystemAlloc(size_t size) { size_t index = 0; void* p = system_allocator_->Alloc(&index, size); - VLOG(10) << "Allocated " << p << " from system allocator."; + VLOG(100) << "Allocated " << p << " from system allocator."; if (p == nullptr) return nullptr; @@ -200,8 +200,8 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() { if (p == nullptr) return pool_.end(); - VLOG(10) << "Creating and inserting new block " << p - << " from system allocator"; + VLOG(100) << "Creating and inserting new block " << p + << " from system allocator"; static_cast(p)->init(&cache_, MemoryBlock::FREE_CHUNK, index, max_chunk_size_, nullptr, nullptr); @@ -245,19 +245,19 @@ void* BuddyAllocator::SplitToAlloc(BuddyAllocator::PoolSet::iterator it, auto block = static_cast(std::get<2>(*it)); pool_.erase(it); - VLOG(10) << "Split block (" << block << ", " << block->total_size(cache_) - << ") into"; + VLOG(100) << "Split block (" << block << ", " << block->total_size(cache_) + << ") into"; block->split(&cache_, size); - VLOG(10) << "Left block (" << block << ", " << block->total_size(cache_) - << ")"; + VLOG(100) << "Left block (" << block << ", " << block->total_size(cache_) + << ")"; block->set_type(&cache_, MemoryBlock::ARENA_CHUNK); // the rest of memory if exist if (block->has_right_buddy(cache_)) { if (block->right_buddy(cache_)->type(cache_) == MemoryBlock::FREE_CHUNK) { - VLOG(10) << "Insert right block (" << block->right_buddy(cache_) << ", " - << block->right_buddy(cache_)->total_size(cache_) << ")"; + VLOG(100) << "Insert right block (" << block->right_buddy(cache_) << ", " + << block->right_buddy(cache_)->total_size(cache_) << ")"; pool_.insert( IndexSizeAddress(block->right_buddy(cache_)->index(cache_), @@ -284,7 +284,7 @@ void BuddyAllocator::CleanIdleFallBackAlloc() { return; } - VLOG(10) << "Return block " << block << " to fallback allocator."; + VLOG(100) << "Return block " << block << " to fallback allocator."; system_allocator_->Free(block, max_chunk_size_, block->index(cache_)); cache_.invalidate(block); @@ -320,7 +320,7 @@ void BuddyAllocator::CleanIdleNormalAlloc() { MemoryBlock* block = static_cast(std::get<2>(*pool)); - VLOG(10) << "Return block " << block << " to base allocator."; + VLOG(100) << "Return block " << block << " to base allocator."; system_allocator_->Free(block, max_chunk_size_, block->index(cache_)); cache_.invalidate(block); diff --git a/paddle/fluid/memory/detail/meta_cache.cc b/paddle/fluid/memory/detail/meta_cache.cc index b86e4f38c42a26e155f276f9b73cbed1d0d83f7d..152e4e7f9fa2e18a2b3e5b4042089660d291badf 100644 --- a/paddle/fluid/memory/detail/meta_cache.cc +++ b/paddle/fluid/memory/detail/meta_cache.cc @@ -29,7 +29,7 @@ MemoryBlock::Desc MetadataCache::load(const MemoryBlock* block) const { return existing_desc->second; } else { auto* desc = reinterpret_cast(block); - VLOG(10) << "Load MemoryBlock::Desc type=" << desc->type; + VLOG(100) << "Load MemoryBlock::Desc type=" << desc->type; PADDLE_ASSERT(desc->check_guards()); return *reinterpret_cast(block); } diff --git a/paddle/fluid/memory/malloc.cc b/paddle/fluid/memory/malloc.cc index 0f13a4ea9c1af175771f5cc201ea5c0a8a0f7555..ec87793b442058ddfc9e22fee47fb0aa5f430b93 100644 --- a/paddle/fluid/memory/malloc.cc +++ b/paddle/fluid/memory/malloc.cc @@ -71,18 +71,18 @@ struct NaiveAllocator { template <> void* Alloc(platform::CPUPlace place, size_t size) { - VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place); + VLOG(100) << "Allocate " << size << " bytes on " << platform::Place(place); void* p = GetCPUBuddyAllocator()->Alloc(size); if (FLAGS_init_allocated_mem) { memset(p, 0xEF, size); } - VLOG(10) << " pointer=" << p; + VLOG(100) << " pointer=" << p; return p; } template <> void Free(platform::CPUPlace place, void* p) { - VLOG(10) << "Free pointer=" << p << " on " << platform::Place(place); + VLOG(100) << "Free pointer=" << p << " on " << platform::Place(place); GetCPUBuddyAllocator()->Free(p); } @@ -110,12 +110,12 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { std::unique_ptr(new detail::GPUAllocator(i)), platform::GpuMinChunkSize(), platform::GpuMaxChunkSize()); - VLOG(10) << "\n\nNOTE: each GPU device use " - << FLAGS_fraction_of_gpu_memory_to_use * 100 - << "% of GPU memory.\n" - << "You can set GFlags environment variable '" - << "FLAGS_fraction_of_gpu_memory_to_use" - << "' to change the fraction of GPU usage.\n\n"; + VLOG(100) << "\n\nNOTE: each GPU device use " + << FLAGS_fraction_of_gpu_memory_to_use * 100 + << "% of GPU memory.\n" + << "You can set GFlags environment variable '" + << "FLAGS_fraction_of_gpu_memory_to_use" + << "' to change the fraction of GPU usage.\n\n"; } }); diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index 2ef13b72ed3ff6ae8ad8748ddea977e693615ac6..776bdfaee8ac24b066b95328fdb59d240f16a446 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -5,6 +5,8 @@ list(REMOVE_DUPLICATES GENERAL_OPS) set(DEPS_OPS "") set(pybind_file ${PADDLE_BINARY_DIR}/paddle/fluid/pybind/pybind.h) file(WRITE ${pybind_file} "// Generated by the paddle/fluid/operator/CMakeLists.txt. DO NOT EDIT!\n\n") + +set(PART_CUDA_KERNEL_FILES) function(op_library TARGET) # op_library is a function to create op library. The interface is same as # cc_library. But it handle split GPU/CPU code and link some common library @@ -37,6 +39,12 @@ function(op_library TARGET) if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu) list(APPEND cu_srcs ${TARGET}.cu) endif() + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu) + set(PART_CUDA_KERNEL_FILES ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu + ${PART_CUDA_KERNEL_FILES} PARENT_SCOPE) + list(APPEND cu_srcs ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.part.cu) + endif() + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu) list(APPEND hip_cu_srcs ${TARGET}.hip.cu) endif() @@ -86,7 +94,7 @@ function(op_library TARGET) # remove windows unsupported op, because windows has no nccl, no warpctc such ops. foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op" "warpctc_op" "hierarchical_sigmoid_op" "crf_decoding_op" "select_op" "lstmp_op" "gru_op" "fusion_gru_op" "lstm_op" "fusion_lstm_op" "cumsum_op" - "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op") + "fusion_seqconv_eltadd_relu_op" "channel_send_op" "channel_create_op" "channel_close_op" "channel_recv_op") if ("${TARGET}" STREQUAL "${windows_unsupport_op}") return() endif() @@ -230,7 +238,7 @@ if(WITH_DISTRIBUTE) op_library(${dist_op} DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(${dist_op}.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) endforeach() - + #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op # listen_and_serv_op sum_op executor SERIAL) @@ -268,6 +276,8 @@ if (WITH_GPU AND TENSORRT_FOUND) else() set(DEPS_OPS ${DEPS_OPS} tensorrt_engine_op) endif() +op_library(hash_op DEPS xxhash) +op_library(clip_by_norm_op DEPS selected_rows_functor selected_rows) op_library(sum_op DEPS selected_rows_functor) op_library(sgd_op DEPS selected_rows_functor) op_library(print_op DEPS lod_tensor) @@ -283,10 +293,10 @@ op_library(max_sequence_len_op DEPS lod_rank_table) op_library(sequence_conv_op DEPS context_project) op_library(sequence_pool_op DEPS sequence_pooling) if (NOT WIN32) -op_library(lstm_op DEPS sequence2batch lstm_compute) -op_library(hierarchical_sigmoid_op DEPS matrix_bit_code) -op_library(lstmp_op DEPS sequence2batch lstm_compute) -op_library(gru_op DEPS sequence2batch gru_compute) + op_library(lstm_op DEPS sequence2batch lstm_compute) + op_library(hierarchical_sigmoid_op DEPS matrix_bit_code) + op_library(lstmp_op DEPS sequence2batch lstm_compute) + op_library(gru_op DEPS sequence2batch gru_compute) endif(NOT WIN32) op_library(recurrent_op DEPS executor) op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale) @@ -294,16 +304,17 @@ op_library(cos_sim_op DEPS cos_sim_functor) op_library(parallel_do_op DEPS executor) op_library(unsqueeze_op DEPS reshape_op) op_library(squeeze_op DEPS reshape_op) -op_library(extract_rows_op DEPS memory) op_library(flatten_op DEPS reshape_op) op_library(sequence_pad_op DEPS sequence_padding) op_library(unstack_op DEPS stack_op) op_library(fake_quantize_op DEPS memory) -op_library(fusion_lstm_op DEPS cpu_lstm_compute) +op_library(crf_decoding_op DEPS jit_kernel) +op_library(fusion_lstm_op DEPS jit_kernel) if (WITH_GPU) op_library(conv_op DEPS vol2col depthwise_conv im2col) op_library(layer_norm_op DEPS cub) op_library(reduce_mean_op DEPS cub) + op_library(affine_channel_op DEPS cub) else() op_library(conv_op DEPS vol2col im2col) endif() @@ -314,7 +325,8 @@ op_library(save_op DEPS lod_tensor) op_library(load_op DEPS lod_tensor) op_library(save_combine_op DEPS lod_tensor) op_library(load_combine_op DEPS lod_tensor) -op_library(concat_op DEPS concat) +op_library(tensor_array_to_tensor_op DEPS concat_op) +op_library(concat_op DEPS concat_and_split) list(REMOVE_ITEM GENERAL_OPS ${DEPS_OPS}) @@ -323,6 +335,8 @@ foreach(src ${GENERAL_OPS}) endforeach() file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n") + + if (NOT WIN32) add_subdirectory(reader) endif(NOT WIN32) @@ -346,6 +360,17 @@ cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory) cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op) cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op) if(NOT WIN32) -nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) + nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) endif() nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor) + +if(WITH_GPU) + foreach(CUDA_KERNEL_FILE ${PART_CUDA_KERNEL_FILES}) + file(READ ${CUDA_KERNEL_FILE} TARGET_CONTENT) + string(REGEX MATCH "REGISTER_OP_CUDA_KERNEL\\(\\n?([^,]+),.*" MATCHED ${TARGET_CONTENT}) + if (MATCHED) + string(STRIP ${CMAKE_MATCH_1} MATCHED) + file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${MATCHED}, CUDA);\n") + endif() + endforeach() +endif() diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index bbf52bea1358c32596ab6f14eeaa419735d19fc6..ea260a3e92b775023085fd02eec33e6ecfaf2e81 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -28,7 +28,7 @@ using paddle::framework::Tensor; public: \ void Make() override { \ AddInput("X", "Input of " #OP_NAME " operator"); \ - AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X"); \ + AddOutput("Out", "Output of " #OP_NAME " operator"); \ AddAttr("use_mkldnn", \ "(bool, default false) Only used in mkldnn kernel") \ .SetDefault(false); \ @@ -91,16 +91,12 @@ class ActivationOp : public framework::OperatorWithKernel { } }; -class ActivationOpInferVarType : public framework::VarTypeInference { - public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto x_name = op_desc.Input("X")[0]; - auto out_name = op_desc.Output("Out")[0]; - auto& x = block->FindRecursiveOrCreateVar(x_name); - auto& out = block->FindRecursiveOrCreateVar(out_name); - out.SetType(x.GetType()); - out.SetDataType(x.GetDataType()); +class ActivationOpInferVarType + : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Out"}}; } }; diff --git a/paddle/fluid/operators/activation_op.cu b/paddle/fluid/operators/activation_op.cu index 27487b396ccf63d962defa6b270063ccb409164e..d3a7ceed466a9b5e4d773f1531d198adff97eac2 100644 --- a/paddle/fluid/operators/activation_op.cu +++ b/paddle/fluid/operators/activation_op.cu @@ -26,6 +26,8 @@ namespace plat = paddle::platform; act_type##_grad, ops::ActivationGradKernel>, \ ops::ActivationGradKernel>); + ops::grad_functor>, \ + ops::ActivationGradKernel>); FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL); diff --git a/paddle/fluid/operators/activation_op.h b/paddle/fluid/operators/activation_op.h index 2e31d1c9c708225135e27c93ba94722794c4b282..4ffc7f364bcb9bda5f94be5fe071c73bd5c40ca7 100644 --- a/paddle/fluid/operators/activation_op.h +++ b/paddle/fluid/operators/activation_op.h @@ -95,7 +95,7 @@ class ActivationGradKernel auto x = framework::EigenVector::Flatten(*X); functor(*place, x, out, dout, dx); } else { - VLOG(10) << " Inplace activation "; + VLOG(100) << " Inplace activation "; auto x = framework::EigenVector::Flatten(*dX); functor(*place, x, out, dout, dx); } @@ -333,8 +333,7 @@ struct SqrtGradFunctor : public BaseActivationFunctor { template void operator()(Device d, X x, Out out, dOut dout, dX dx) const { - const Out out_conj = Eigen::numext::conj(out); - dx.device(d) = static_cast(0.5) * dout / out_conj; + dx.device(d) = static_cast(0.5) * dout / out; } }; @@ -740,7 +739,7 @@ struct PowGradFunctor : public BaseActivationFunctor { typename dX> void operator()(Device d, X x, Out out, dOut dout, dX dx) const { dx.device(d) = dout * static_cast(factor) * - x.pow(static_cast(factor - static_cast(1))); + x.pow(static_cast(factor) - static_cast(1)); } }; diff --git a/paddle/fluid/operators/adadelta_op.cc b/paddle/fluid/operators/adadelta_op.cc index d1970515f58969948b1d2db5847e4344112f77f9..89a7a49e0fa8427826f5d91274912a68f2316b61 100644 --- a/paddle/fluid/operators/adadelta_op.cc +++ b/paddle/fluid/operators/adadelta_op.cc @@ -18,6 +18,7 @@ namespace paddle { namespace operators { using Tensor = framework::Tensor; + class AdadeltaOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -31,6 +32,16 @@ class AdadeltaOp : public framework::OperatorWithKernel { "Input(AvgSquaredGrad) of AdadeltaOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("AvgSquaredUpdate"), "Input(AvgSquaredUpdate) of AdadeltaOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of AdadeltaOp should not be null."); @@ -56,6 +67,7 @@ class AdadeltaOp : public framework::OperatorWithKernel { ctx->SetOutputDim("AvgSquaredGradOut", param_dim); ctx->SetOutputDim("AvgSquaredUpdateOut", param_dim); } + framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext &ctx) const override { auto input_data_type = diff --git a/paddle/fluid/operators/adadelta_op.h b/paddle/fluid/operators/adadelta_op.h index 822458daf663d99bbb38d99205f51163a0df4c4d..6c616aa03d9809e9b7725a700c7edd5ff5d6dc42 100644 --- a/paddle/fluid/operators/adadelta_op.h +++ b/paddle/fluid/operators/adadelta_op.h @@ -23,6 +23,17 @@ template class AdadeltaOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto avg_squared_grad_out_tensor = ctx.Output("AvgSquaredGradOut"); diff --git a/paddle/fluid/operators/adagrad_op.cc b/paddle/fluid/operators/adagrad_op.cc index a3ef9ad9f91f1f626bd33876693ecc17ad76b96b..c88297ff544ddb0e5a97452a8ad2e8f9f77825ba 100644 --- a/paddle/fluid/operators/adagrad_op.cc +++ b/paddle/fluid/operators/adagrad_op.cc @@ -119,8 +119,8 @@ struct SparseAdagradFunctor { auto* grad_merge_data = grad_merge.mutable_value()->template data(); // 2. m += g_m * g_m - math::scatter::Mul sqare_func; - auto grad_square = sqare_func(context, grad_merge, grad_merge); + auto grad_square = + SquareSelectedRows(context, grad_merge); math::SelectedRowsAddToTensor functor; functor(context, grad_square, moment); diff --git a/paddle/fluid/operators/adagrad_op.cu b/paddle/fluid/operators/adagrad_op.cu index b25268786d622bc7a94117849763833e528bef48..b99b33343d36fbb7f6b1a2928e142ca615b238b3 100644 --- a/paddle/fluid/operators/adagrad_op.cu +++ b/paddle/fluid/operators/adagrad_op.cu @@ -84,8 +84,8 @@ struct SparseAdagradFunctor { auto* grad_merge_data = grad_merge.mutable_value()->template data(); framework::Vector merge_rows(grad_merge.rows()); // 2. m += g_m * g_m - math::scatter::Mul sqare_func; - auto grad_square = sqare_func(context, grad_merge, grad_merge); + auto grad_square = + SquareSelectedRows(context, grad_merge); math::SelectedRowsAddToTensor functor; functor(context, grad_square, moment); diff --git a/paddle/fluid/operators/adagrad_op.h b/paddle/fluid/operators/adagrad_op.h index df520fcc898ff5514927dbdd845ecaecdcf3c147..9f6ef391696aa8718be71ae945e746b876813d94 100644 --- a/paddle/fluid/operators/adagrad_op.h +++ b/paddle/fluid/operators/adagrad_op.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once + #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" @@ -21,25 +22,45 @@ namespace operators { template struct SparseAdagradFunctor { - void operator()(const DeviceContext& context, - const framework::SelectedRows& grad, - const framework::Tensor& learning_rate, T epsilon, - framework::Tensor* moment, framework::Tensor* param); + void operator()(const DeviceContext &context, + const framework::SelectedRows &grad, + const framework::Tensor &learning_rate, T epsilon, + framework::Tensor *moment, framework::Tensor *param); }; +template +framework::SelectedRows SquareSelectedRows( + const DeviceContext &context, const framework::SelectedRows &input) { + framework::SelectedRows out; + out.set_rows(input.rows()); + out.set_height(input.height()); + out.mutable_value()->mutable_data(input.value().dims(), + context.GetPlace()); + auto e_out = framework::EigenVector::Flatten(*(out.mutable_value())); + auto e_in = framework::EigenVector::Flatten(input.value()); + e_out.device(*context.eigen_device()) = e_in.square(); + return out; +} + template class AdagradOpKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* param_out_tensor = ctx.Output("ParamOut"); - auto* moment_out_tensor = ctx.Output("MomentOut"); + void Compute(const framework::ExecutionContext &ctx) const override { + const auto *param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + + auto *param_out_tensor = ctx.Output("ParamOut"); + auto *moment_out_tensor = ctx.Output("MomentOut"); param_out_tensor->mutable_data(ctx.GetPlace()); moment_out_tensor->mutable_data(ctx.GetPlace()); T epsilon = static_cast(ctx.Attr("epsilon")); - auto* grad_var = ctx.InputVar("Grad"); + auto *grad_var = ctx.InputVar("Grad"); if (grad_var->IsType()) { auto param = framework::EigenVector::Flatten( *ctx.Input("Param")); @@ -47,16 +68,16 @@ class AdagradOpKernel : public framework::OpKernel { *ctx.Input("Grad")); auto moment = framework::EigenVector::Flatten( *ctx.Input("Moment")); - auto* learning_rate = ctx.Input("LearningRate"); + auto *learning_rate = ctx.Input("LearningRate"); auto param_out = framework::EigenVector::Flatten(*param_out_tensor); auto moment_out = framework::EigenVector::Flatten(*moment_out_tensor); - auto* place = ctx.template device_context().eigen_device(); + auto *place = ctx.template device_context().eigen_device(); moment_out.device(*place) = moment + grad * grad; Eigen::DSizes m_dsize(moment_out_tensor->numel()); if (platform::is_cpu_place(ctx.GetPlace())) { - auto* lr = learning_rate->data(); + auto *lr = learning_rate->data(); param_out.device(*place) = param - lr[0] * grad / (moment_out.sqrt() + epsilon); } else { @@ -66,10 +87,10 @@ class AdagradOpKernel : public framework::OpKernel { lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon); } } else if (grad_var->IsType()) { - auto* param_tensor = ctx.Input("Param"); + auto *param_tensor = ctx.Input("Param"); PADDLE_ENFORCE_EQ(param_tensor, param_out_tensor); - auto* moment_tensor = ctx.Input("Moment"); + auto *moment_tensor = ctx.Input("Moment"); PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor); SparseAdagradFunctor functor; diff --git a/paddle/fluid/operators/adam_op.cc b/paddle/fluid/operators/adam_op.cc index 5d670fe3b9d99a31a628ff707ff860564eca952e..f3717af630017eba18aa265f3dbb496e18280a57 100644 --- a/paddle/fluid/operators/adam_op.cc +++ b/paddle/fluid/operators/adam_op.cc @@ -92,9 +92,9 @@ class AdamOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Beta1Pow", "(Tensor) Input beta1 power accumulator"); AddInput("Beta2Pow", "(Tensor) Input beta2 power accumulator"); - AddOutput("ParamOut", "(Tensor) Output parameter").Reuse("Param"); - AddOutput("Moment1Out", "(Tensor) Output first moment").Reuse("Moment1"); - AddOutput("Moment2Out", "(Tensor) Output second moment").Reuse("Moment2"); + AddOutput("ParamOut", "(Tensor) Output parameter"); + AddOutput("Moment1Out", "(Tensor) Output first moment"); + AddOutput("Moment2Out", "(Tensor) Output second moment"); AddAttr("beta1", "(float, default 0.9) " diff --git a/paddle/fluid/operators/adam_op.h b/paddle/fluid/operators/adam_op.h index 4cb1f3a80e95bdda79e6451dc3cc87e899b11779..48e0448d09c64e2c2fa655d125064e7a6572e30e 100644 --- a/paddle/fluid/operators/adam_op.h +++ b/paddle/fluid/operators/adam_op.h @@ -18,6 +18,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" +#include "paddle/fluid/operators/math/algorithm.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/for_range.h" @@ -199,23 +200,9 @@ struct SparseAdamFunctor { row_numel_(row_numel), row_count_(row_count) {} - inline HOSTDEVICE int64_t BinarySearchInRows(int64_t row) const { - int64_t beg = 0, end = row_count_ - 1; - while (beg <= end) { - auto mid = ((beg + end) >> 1); - if (rows_[mid] == row) - return mid; - else if (rows_[mid] < row) - beg = mid + 1; - else - end = mid - 1; - } - return -1; - } - inline HOSTDEVICE void operator()(size_t i) const { - int64_t row = i / row_numel_; - auto row_idx = BinarySearchInRows(row); + auto row_idx = + math::BinarySearch(rows_, row_count_, i / row_numel_); T g = row_idx >= 0 ? grad_[row_idx * row_numel_ + i % row_numel_] : 0; // The following code is the same as dense @@ -244,6 +231,12 @@ template class AdamOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + using paddle::framework::LoDTensor; using paddle::operators::detail::Ref; @@ -304,7 +297,7 @@ class AdamOpKernel : public framework::OpKernel { auto& grad = Ref(ctx.Input("Grad"), "Must set Grad"); if (grad.rows().size() == 0) { - VLOG(3) << "grad row size is 0!!"; + VLOG(30) << "grad row size is 0!!"; return; } diff --git a/paddle/fluid/operators/adamax_op.cc b/paddle/fluid/operators/adamax_op.cc index 32062574bcf71ff96e451eaa6865b6bbfc3b1c80..d4aa4d338a2379adf985ba7f89b528bc402eda06 100644 --- a/paddle/fluid/operators/adamax_op.cc +++ b/paddle/fluid/operators/adamax_op.cc @@ -35,6 +35,16 @@ class AdamaxOp : public framework::OperatorWithKernel { "Input(LearningRate) of AdamaxOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Beta1Pow"), "Input(Beta1Pow) of AdamaxOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of AdamaxOp should not be null."); diff --git a/paddle/fluid/operators/adamax_op.h b/paddle/fluid/operators/adamax_op.h index de644676fd9c3fabdbf01d2fd9c69858c2627ed3..7137fbd9651b4523f6d1609a0595b30758aa40df 100644 --- a/paddle/fluid/operators/adamax_op.h +++ b/paddle/fluid/operators/adamax_op.h @@ -23,6 +23,17 @@ template class AdamaxOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto moment_out_tensor = ctx.Output("MomentOut"); auto inf_norm_out_tensor = ctx.Output("InfNormOut"); diff --git a/paddle/fluid/operators/add_position_encoding_op.cc b/paddle/fluid/operators/add_position_encoding_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..8127e554bed1aae7a5ce8837bcadf1b7f13f1ac2 --- /dev/null +++ b/paddle/fluid/operators/add_position_encoding_op.cc @@ -0,0 +1,97 @@ +/* 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/add_position_encoding_op.h" + +namespace paddle { +namespace operators { + +class AddPositionEncodingOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "X(Input) of add_position_encoding_op should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Out"), + "Out(Output) of add_position_encoding_op should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class AddPositionEncodingOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "X(Input) must not be null."); + PADDLE_ENFORCE(ctx->HasInput("Out"), "Out must not be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Out@GRAD must not be null."); + + auto out_dims = ctx->GetInputDim("Out"); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), out_dims); + } + } +}; + +class AddPositionEncodingOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "Input of AddPositionEncoding operator"); + AddOutput("Out", "Output of AddPositionEncoding operator"); + AddAttr("alpha", "The scale of Original Embedding.") + .SetDefault(1.0f) + .AddCustomChecker([](const float& alpha) { + PADDLE_ENFORCE(alpha >= 0.0f, "'alpha' must be above 0.0."); + }); + AddAttr("beta", "The scale of Position Embedding.") + .SetDefault(1.0f) + .AddCustomChecker([](const float& beta) { + PADDLE_ENFORCE(beta >= 0.0f, "'beta' must be between 0.0."); + }); + AddComment(R"DOC( + Add Position Encoding Operator. + + The add position encoding calculates the output based on the input, alpha, beta. + The size of each dimension of the parameters checked in the infer-shape. + )DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +namespace plt = paddle::platform; + +REGISTER_OPERATOR(add_position_encoding, ops::AddPositionEncodingOp, + ops::AddPositionEncodingOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(add_position_encoding_grad, ops::AddPositionEncodingOpGrad); + +REGISTER_OP_CPU_KERNEL( + add_position_encoding, + ops::AddPositionEncodingKernel, + ops::AddPositionEncodingKernel); + +REGISTER_OP_CPU_KERNEL( + add_position_encoding_grad, + ops::AddPositionEncodingGradKernel, + ops::AddPositionEncodingGradKernel); diff --git a/paddle/fluid/operators/add_position_encoding_op.h b/paddle/fluid/operators/add_position_encoding_op.h new file mode 100644 index 0000000000000000000000000000000000000000..0b40d3de890a02a9dbec2328f9f6388ffa35561b --- /dev/null +++ b/paddle/fluid/operators/add_position_encoding_op.h @@ -0,0 +1,106 @@ +/* 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 "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/safe_ref.h" + +namespace paddle { +namespace operators { + +template +class AddPositionEncodingKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* X = context.Input("X"); + auto& x_lod = X->lod(); + auto* src_ptr = X->data(); + + auto* Out = context.Output("Out"); + auto* dst_ptr = Out->mutable_data(context.GetPlace()); + + float alpha = context.Attr("alpha"); + float beta = context.Attr("beta"); + + auto x_dim = X->dims(); + int batch_size = 0; + int max_seq_len = 0; + int enc_size = 0; + + if (x_lod.empty()) { + PADDLE_ENFORCE( + x_dim.size() == 3UL, + "The input X of Add Position Encoding should be 3-D Tensor!"); + batch_size = x_dim[0]; + max_seq_len = x_dim[1]; + enc_size = x_dim[2]; + } else { + PADDLE_ENFORCE( + x_dim.size() == 2UL, + "The input X of Add Position Encoding should be 2-D LoDTensor!"); + PADDLE_ENFORCE( + x_lod.size() == 1UL, + "The Add Position Encoding Op only supports lod_level == 1!"); + batch_size = x_lod[0].size() - 1; + max_seq_len = -1; + enc_size = x_dim[1]; + } + + PADDLE_ENFORCE(enc_size % 2 == 0, "Only support even encode size!"); + + const int half_size = enc_size / 2; + for (int i = 0; i < batch_size; ++i) { + const int max_length = + x_lod.empty() ? max_seq_len : x_lod[0][i + 1] - x_lod[0][i]; + for (int j = 0; j < max_length; ++j) { + for (int k = 0; k < half_size; ++k) { + const double val = + (half_size > 1) + ? j / pow(10000.0, static_cast(k) / (half_size - 1)) + : j / 10000.0; + dst_ptr[k] = src_ptr[k] * alpha + sin(val) * beta; + dst_ptr[half_size + k] = + src_ptr[half_size + k] * alpha + cos(val) * beta; + } + src_ptr += enc_size; + dst_ptr += enc_size; + } + } + } +}; + +template +class AddPositionEncodingGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* dOut = + context.Input(framework::GradVarName("Out")); + auto dout = framework::EigenVector::Flatten(*dOut); + + auto* dX = + context.Output(framework::GradVarName("X")); + dX->mutable_data(context.GetPlace()); + auto dx = framework::EigenVector::Flatten(*dX); + + float alpha = context.Attr("alpha"); + + auto* place = + context.template device_context().eigen_device(); + dx.device(*place) = dout * static_cast(alpha); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/affine_channel_op.cc b/paddle/fluid/operators/affine_channel_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..8944a749674c3ba6c83526e4d66f449075716f43 --- /dev/null +++ b/paddle/fluid/operators/affine_channel_op.cc @@ -0,0 +1,255 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +Indicesou 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/framework/data_layout.h" +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +class AffineChannelOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor) Feature map input can be a 4D tensor with order NCHW " + "or NHWC. It also can be a 2D tensor and C is the second " + "dimension."); + AddInput("Scale", + "(Tensor) 1D input of shape (C), the c-th element " + "is the scale factor of the affine transformation " + "for the c-th channel of the input."); + AddInput("Bias", + "(Tensor) 1D input of shape (C), the c-th element " + "is the bias of the affine transformation for the " + "c-th channel of the input."); + AddAttr( + "data_layout", + "(string, default NCHW) Only used in " + "An optional string from: \"NHWC\", \"NCHW\". " + "Defaults to \"NHWC\". Specify the data format of the output data, " + "the input will be transformed automatically. ") + .SetDefault("AnyLayout"); + AddOutput("Out", "(Tensor) A tensor of the same shape and order with X."); + AddComment(R"DOC( + +Applies a separate affine transformation to each channel of the input. Useful +for replacing spatial batch norm with its equivalent fixed transformation. +The input also can be 2D tensor and applies a affine transformation in second +dimension. + +$$Out = Scale*X + Bias$$ + +)DOC"); + } +}; + +class AffineChannelOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of AffineChannelOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Scale"), + "Input(Scale) of AffineChannelOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Bias"), + "Input(Bias) of AffineChannelOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of AffineChannelOp should not be null."); + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", "Out"); + } +}; + +class AffineChannelOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) should not be null."); + if (ctx->HasOutput(framework::GradVarName("X"))) { + PADDLE_ENFORCE(ctx->HasInput("Scale"), + "Input(Scale) should not be null."); + ctx->SetOutputDim(framework::GradVarName("X"), + ctx->GetInputDim(framework::GradVarName("Out"))); + } + if (ctx->HasOutput(framework::GradVarName("Scale"))) { + // Scale@GRAD and Bias@GRAD must exist at the same time. + PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("Bias")), + "Output(Scale@GRAD) should not be null."); + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null."); + ctx->SetOutputDim(framework::GradVarName("Scale"), + ctx->GetInputDim("Scale")); + ctx->SetOutputDim(framework::GradVarName("Bias"), + ctx->GetInputDim("Scale")); + } + } +}; + +template +using EigenArrayMap = + Eigen::Map>; +template +using ConstEigenArrayMap = + Eigen::Map>; +template +using EigenVectorArrayMap = Eigen::Map>; +template +using ConstEigenVectorArrayMap = + Eigen::Map>; + +template +class AffineChannelKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* scale = ctx.Input("Scale"); + auto* bias = ctx.Input("Bias"); + + auto* y = ctx.Output("Out"); + y->mutable_data(ctx.GetPlace()); + + const framework::DataLayout layout = + framework::StringToDataLayout(ctx.Attr("data_layout")); + + auto dims = x->dims(); + int N = dims[0]; + int C = layout == framework::DataLayout::kNCHW ? dims[1] + : dims[dims.size() - 1]; + int HxW = x->numel() / N / C; + + auto* scale_d = scale->data(); + auto* bias_d = bias->data(); + ConstEigenVectorArrayMap a_e(scale_d, C); + ConstEigenVectorArrayMap b_e(bias_d, C); + + auto* x_d = x->data(); + auto* y_d = y->data(); + if (layout == framework::DataLayout::kNCHW) { + int stride = C * HxW; + for (int i = 0; i < N; i++) { + ConstEigenArrayMap x_e(x_d, HxW, C); + EigenArrayMap y_e(y_d, HxW, C); + y_e = (x_e.rowwise() * a_e.transpose()).rowwise() + b_e.transpose(); + x_d += stride; + y_d += stride; + } + } else { + int num = N * HxW; + ConstEigenArrayMap x_e(x_d, C, num); + EigenArrayMap y_e(y_d, C, num); + y_e = (x_e.colwise() * a_e).colwise() + b_e; + } + } +}; + +template +class AffineChannelGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* scale = ctx.Input("Scale"); + auto* dy = ctx.Input(framework::GradVarName("Out")); + + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dscale = + ctx.Output(framework::GradVarName("Scale")); + auto* dbias = ctx.Output(framework::GradVarName("Bias")); + + const framework::DataLayout layout = + framework::StringToDataLayout(ctx.Attr("data_layout")); + + auto dims = x->dims(); + int N = dims[0]; + int C = layout == framework::DataLayout::kNCHW ? dims[1] + : dims[dims.size() - 1]; + int HxW = x->numel() / N / C; + + auto* x_d = x->data(); + auto* dy_d = dy->data(); + auto* scale_d = scale->data(); + ConstEigenVectorArrayMap scale_e(scale_d, C); + + T* dx_d = dx ? dx->mutable_data(ctx.GetPlace()) : nullptr; + T* dscale_d = dscale ? dscale->mutable_data(ctx.GetPlace()) : nullptr; + T* dbias_d = dbias ? dbias->mutable_data(ctx.GetPlace()) : nullptr; + EigenVectorArrayMap dscale_e(dscale_d, C); + EigenVectorArrayMap dbias_e(dbias_d, C); + + if (layout == framework::DataLayout::kNCHW) { + // compute dx + int stride = C * HxW; + if (dx) { + for (int i = 0; i < N; i++) { + ConstEigenArrayMap dy_e(dy_d, HxW, C); + EigenArrayMap dx_e(dx_d, HxW, C); + dx_e = dy_e.rowwise() * scale_e.transpose(); + dy_d += stride; + dx_d += stride; + } + } + // compute dscale and dbias + if (dscale && dbias) { + dy_d = dy->data(); + for (int i = 0; i < N; i++) { + ConstEigenArrayMap x_e(x_d, HxW, C); + ConstEigenArrayMap dy_e(dy_d, HxW, C); + if (i == 0) { + dscale_e = (x_e * dy_e).colwise().sum(); + } else { + dscale_e += (x_e * dy_e).colwise().sum(); + } + if (i == 0) { + dbias_e = dy_e.colwise().sum(); + } else { + dbias_e += dy_e.colwise().sum(); + } + x_d += stride; + dy_d += stride; + } + } + } else { + int num = N * HxW; + ConstEigenArrayMap dy_e(dy_d, C, num); + // compute dx + if (dx) { + EigenArrayMap dx_e(dx_d, C, num); + dx_e = dy_e.colwise() * scale_e; + } + // compute dscale and dbias + if (dscale && dbias) { + ConstEigenArrayMap x_e(x_d, C, num); + dscale_e = (x_e * dy_e).rowwise().sum(); + dbias_e = dy_e.rowwise().sum(); + } + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +using CPU = paddle::platform::CPUDeviceContext; + +REGISTER_OPERATOR(affine_channel, ops::AffineChannelOp, + ops::AffineChannelOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(affine_channel_grad, ops::AffineChannelOpGrad); + +REGISTER_OP_CPU_KERNEL(affine_channel, ops::AffineChannelKernel, + ops::AffineChannelKernel); +REGISTER_OP_CPU_KERNEL(affine_channel_grad, + ops::AffineChannelGradKernel, + ops::AffineChannelGradKernel); diff --git a/paddle/fluid/operators/affine_channel_op.cu b/paddle/fluid/operators/affine_channel_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..2bebdb345ab324eb0a2dafd54c74833dd21bdb6d --- /dev/null +++ b/paddle/fluid/operators/affine_channel_op.cu @@ -0,0 +1,187 @@ +/* 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. +Indicesou 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 "cub/cub.cuh" +#include "paddle/fluid/framework/data_layout.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/cuda_primitives.h" + +namespace paddle { +namespace operators { + +template +__global__ void KeAffineChannelCUDA(const T* x, const T* scale, const T* bias, + const int C, const int HxW, const int num, + T* y) { + int gid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (int i = gid; i < num; i += stride) { + const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C; + if (HasBias) { + y[i] = scale[c] * x[i] + bias[c]; + } else { + y[i] = scale[c] * x[i]; + } + } +} + +template +class AffineChannelCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* scale = ctx.Input("Scale"); + auto* bias = ctx.Input("Bias"); + + auto* y = ctx.Output("Out"); + y->mutable_data(ctx.GetPlace()); + + const framework::DataLayout layout = + framework::StringToDataLayout(ctx.Attr("data_layout")); + auto& dev_ctx = ctx.template device_context(); + + auto dims = x->dims(); + const int num = x->numel(); + int N = dims[0]; + int C = layout == framework::DataLayout::kNCHW ? dims[1] + : dims[dims.size() - 1]; + int HxW = num / N / C; + + const T* x_d = x->data(); + const T* scale_d = scale->data(); + const T* bias_d = bias->data(); + T* y_d = y->data(); + + int block = 1024; + int grid = (num + block - 1) / block; + if (layout == framework::DataLayout::kNCHW) { + KeAffineChannelCUDA<<>>( + x_d, scale_d, bias_d, C, HxW, num, y_d); + } else { + KeAffineChannelCUDA<<>>( + x_d, scale_d, bias_d, C, HxW, num, y_d); + } + } +}; + +template +__global__ void AffineChannelScaleBiasGradientCUDAKernel( + const T* dy, const T* x, const int N, const int C, const int HxW, T* dscale, + T* dbias) { + const int outer_size = C; + const int inner_size = N * HxW; + typedef cub::BlockReduce BlockReduce; + __shared__ typename BlockReduce::TempStorage ds_storage; + __shared__ typename BlockReduce::TempStorage db_storage; + + for (int i = blockIdx.x; i < outer_size; i += gridDim.x) { + T ds_sum = 0; + T db_sum = 0; + for (int j = threadIdx.x; j < inner_size; j += blockDim.x) { + const int index = layout == framework::DataLayout::kNCHW + ? (j / HxW * C + i) * HxW + j % HxW + : j * outer_size + i; + ds_sum += dy[index] * x[index]; + db_sum += dy[index]; + } + ds_sum = BlockReduce(ds_storage).Reduce(ds_sum, cub::Sum()); + db_sum = BlockReduce(db_storage).Reduce(db_sum, cub::Sum()); + if (threadIdx.x == 0) { + dscale[i] = ds_sum; + dbias[i] = db_sum; + } + __syncthreads(); + } +} + +template +class AffineChannelGradCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* scale = ctx.Input("Scale"); + auto* bias = ctx.Input("Bias"); + auto* dy = ctx.Input(framework::GradVarName("Out")); + + auto* dx = ctx.Output(framework::GradVarName("X")); + auto* dscale = + ctx.Output(framework::GradVarName("Scale")); + auto* dbias = ctx.Output(framework::GradVarName("Bias")); + + const framework::DataLayout layout = + framework::StringToDataLayout(ctx.Attr("data_layout")); + auto& dev_ctx = ctx.template device_context(); + + auto dims = x->dims(); + const int num = x->numel(); + int N = dims[0]; + int C = layout == framework::DataLayout::kNCHW ? dims[1] + : dims[dims.size() - 1]; + int HxW = num / N / C; + + const T* x_d = x->data(); + const T* dy_d = dy->data(); + const T* s_d = scale->data(); + + T* dx_d = dx ? dx->mutable_data(ctx.GetPlace()) : nullptr; + T* ds_d = dscale ? dscale->mutable_data(ctx.GetPlace()) : nullptr; + T* db_d = dbias ? dbias->mutable_data(ctx.GetPlace()) : nullptr; + + const int block = 1024; + int max_threads = dev_ctx.GetMaxPhysicalThreadCount(); + const int max_blocks = std::max(max_threads / block, 1); + int grid1 = (num + block - 1) / block; + int grid2 = std::min(C, max_blocks); + if (layout == framework::DataLayout::kNCHW) { + if (dx) { + KeAffineChannelCUDA<<>>( + dy_d, s_d, nullptr, C, HxW, num, dx_d); + } + if (dscale && dbias) { + AffineChannelScaleBiasGradientCUDAKernel< + T, block, framework::DataLayout::kNCHW><<>>( + dy_d, x_d, N, C, HxW, ds_d, db_d); + } + } else { + if (dx) { + KeAffineChannelCUDA<<>>( + dy_d, s_d, nullptr, C, HxW, num, dx_d); + } + if (dscale && dbias) { + AffineChannelScaleBiasGradientCUDAKernel< + T, block, framework::DataLayout::kNHWC><<>>( + dy_d, x_d, N, C, HxW, ds_d, db_d); + } + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +using CUDA = paddle::platform::CUDADeviceContext; + +REGISTER_OP_CUDA_KERNEL(affine_channel, + ops::AffineChannelCUDAKernel, + ops::AffineChannelCUDAKernel); +REGISTER_OP_CUDA_KERNEL(affine_channel_grad, + ops::AffineChannelGradCUDAKernel, + ops::AffineChannelGradCUDAKernel); diff --git a/paddle/fluid/operators/affine_grid_cudnn_op.cu.cc b/paddle/fluid/operators/affine_grid_cudnn_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..ed71594ba5781590f3291d56c4ba1a4443003bd5 --- /dev/null +++ b/paddle/fluid/operators/affine_grid_cudnn_op.cu.cc @@ -0,0 +1,112 @@ +/* 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/framework/op_registry.h" +#include "paddle/fluid/platform/cudnn_helper.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using ScopedSpatialTransformerDescriptor = + platform::ScopedSpatialTransformerDescriptor; + +template +class CUDNNAffineGridOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace."); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto* theta = ctx.Input("Theta"); + auto* output = ctx.Output("Output"); + const T* theta_data = theta->data(); + + int n = theta->dims()[0]; + auto size_attr = ctx.Attr>("output_shape"); + Tensor h_sizes; + int* h_size_data; + if (size_attr.size() == 0) { + auto* output_shape = ctx.Input("OutputShape"); + framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes); + h_size_data = h_sizes.data(); + } else { + h_size_data = h_sizes.mutable_data({4}, platform::CPUPlace()); + h_size_data[0] = n; + h_size_data[1] = size_attr[1]; + h_size_data[2] = size_attr[2]; + h_size_data[3] = size_attr[3]; + } + + T* output_data = output->mutable_data( + {n, h_size_data[2], h_size_data[3], 2}, ctx.GetPlace()); + ScopedSpatialTransformerDescriptor st_desc; + cudnnSpatialTransformerDescriptor_t cudnn_st_desc = + st_desc.descriptor(4, h_size_data); + + PADDLE_ENFORCE(platform::dynload::cudnnSpatialTfGridGeneratorForward( + handle, cudnn_st_desc, theta_data, output_data)); + } +}; + +template +class CUDNNAffineGridGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace."); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto output_grad = ctx.Input(framework::GradVarName("Output")); + auto theta_grad = ctx.Output(framework::GradVarName("Theta")); + + int n = output_grad->dims()[0]; + auto size_attr = ctx.Attr>("output_shape"); + Tensor h_sizes; + int* h_size_data; + if (size_attr.size() == 0) { + auto* output_shape = ctx.Input("OutputShape"); + framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes); + h_size_data = h_sizes.data(); + } else { + h_size_data = h_sizes.mutable_data({4}, platform::CPUPlace()); + h_size_data[0] = n; + h_size_data[1] = size_attr[1]; + h_size_data[2] = size_attr[2]; + h_size_data[3] = size_attr[3]; + } + + ScopedSpatialTransformerDescriptor st_desc; + cudnnSpatialTransformerDescriptor_t cudnn_st_desc = + st_desc.descriptor(4, h_size_data); + + const T* output_grad_data = output_grad->data(); + T* theta_grad_data = theta_grad->mutable_data(ctx.GetPlace()); + + PADDLE_ENFORCE(platform::dynload::cudnnSpatialTfGridGeneratorBackward( + handle, cudnn_st_desc, output_grad_data, theta_grad_data)); + } +}; + +} // namespace operators +} // namespace paddle + +namespace plat = paddle::platform; +REGISTER_OP_KERNEL(affine_grid, CUDNN, plat::CUDAPlace, + paddle::operators::CUDNNAffineGridOpKernel, + paddle::operators::CUDNNAffineGridOpKernel); +REGISTER_OP_KERNEL(affine_grid_grad, CUDNN, plat::CUDAPlace, + paddle::operators::CUDNNAffineGridGradOpKernel, + paddle::operators::CUDNNAffineGridGradOpKernel); diff --git a/paddle/fluid/operators/affine_grid_op.cc b/paddle/fluid/operators/affine_grid_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..6f7da445fc84fc1f14b01a633af0e886aec6f8ed --- /dev/null +++ b/paddle/fluid/operators/affine_grid_op.cc @@ -0,0 +1,231 @@ +/* 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/affine_grid_op.h" +#include +#include "paddle/fluid/framework/op_registry.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/cudnn_helper.h" +#endif + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +struct Linspace { + void operator()(T start, T end, int count, framework::Tensor* numbers, + const framework::ExecutionContext& ctx) { + T* number_data = numbers->mutable_data({count}, platform::CPUPlace()); + T slice = (end - start) / (T)(count - 1); + for (int i = 0; i < count; ++i) { + number_data[i] = start + (T)i * slice; + } + } +}; + +class AffineGridOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Theta"), + "Input(Theta) of AffineGridOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Output"), + "Output(Output) of AffineGridOp should not be null."); + auto theta_dims = ctx->GetInputDim("Theta"); + PADDLE_ENFORCE(theta_dims.size() == 3, + "AffineGrid's Input(Theta) should be 3-D tensor."); + + auto output_shape = ctx->Attrs().Get>("output_shape"); + if (output_shape.size() == 0) { + PADDLE_ENFORCE(ctx->HasInput("OutputShape"), + "Input(OutputShape) of AffineGridOp should not be null if " + "attr(output_shape) is not configured."); + auto output_shape_dims = ctx->GetInputDim("OutputShape"); + PADDLE_ENFORCE(output_shape_dims.size() == 1, + "AffineGrid's Input(OutputShape) should be 1-D tensor."); + } else { + PADDLE_ENFORCE(output_shape.size() == 4, + "The size of attr(output_shape) should be 4."); + } + + PADDLE_ENFORCE(theta_dims[1] == 2, "Input(theta) dims[1] should be 2."); + PADDLE_ENFORCE(theta_dims[2] == 3, "Input(theta) dims[2] should be 3."); + // N * H * W * 2 + ctx->SetOutputDim("Output", + framework::make_ddim({theta_dims[0], -1, -1, 2})); + ctx->ShareLoD("Theta", "Output"); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::LibraryType library{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_CUDA + if (platform::CanCUDNNBeUsed(ctx)) { + library = framework::LibraryType::kCUDNN; + } +#endif + auto data_type = framework::ToDataType(ctx.Input("Theta")->type()); + return framework::OpKernelType(data_type, ctx.GetPlace(), + framework::DataLayout::kAnyLayout, library); + } +}; + +class AffineGridOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput( + "Theta", + "(Tensor) A batch of affine transform parameters with shape [N, 2, 3]. " + "It is used to transform coordinate (x_0, y_0) to coordinate (x_1, " + "y_1)."); + AddInput("OutputShape", + "(Tensor) The shape of target image with format [N, C, H, W].") + .AsDispensable(); + AddOutput("Output", "(Tensor) Output Tensor with shape [N, H, W, 2]."); + AddAttr( + "use_cudnn", + "(bool, default false) Only used in cudnn kernel, need install cudnn") + .SetDefault(true); + AddAttr>( + "output_shape", + "The target output image shape with format [N, C, H, W].") + .SetDefault(std::vector()); + + AddComment(R"DOC( + It generates a grid of (x,y) coordinates using the parameters of the + affine transformation that correspond to a set of points where the input + feature map should be sampled to produce the transformed output feature map. + + Given: + Theta = [[[x_11, x_12, x_13] + [x_14, x_15, x_16]] + [[x_21, x_22, x_23] + [x_24, x_25, x_26]]] + + OutputShape = [2, 3, 5, 5] + + Step 1: + + Generate relative coordinates according to OutputShape. + The values of relative coordinates are in the interval between -1 and 1. + The shape of the relative coordinates is [2, H, W] as below: + + C = [[[-1. -1. -1. -1. -1. ] + [-0.5 -0.5 -0.5 -0.5 -0.5] + [ 0. 0. 0. 0. 0. ] + [ 0.5 0.5 0.5 0.5 0.5] + [ 1. 1. 1. 1. 1. ]] + [[-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ]]] + C[0] is the coordinates in height axis and C[1] is the coordinates in width axis. + + Step2: + Tanspose and reshape C to shape [H * W, 2] and append ones to last dimension. The we get: + C_ = [[-1. -1. 1. ] + [-0.5 -1. 1. ] + [ 0. -1. 1. ] + [ 0.5 -1. 1. ] + [ 1. -1. 1. ] + [-1. -0.5 1. ] + [-0.5 -0.5 1. ] + [ 0. -0.5 1. ] + [ 0.5 -0.5 1. ] + [ 1. -0.5 1. ] + [-1. 0. 1. ] + [-0.5 0. 1. ] + [ 0. 0. 1. ] + [ 0.5 0. 1. ] + [ 1. 0. 1. ] + [-1. 0.5 1. ] + [-0.5 0.5 1. ] + [ 0. 0.5 1. ] + [ 0.5 0.5 1. ] + [ 1. 0.5 1. ] + [-1. 1. 1. ] + [-0.5 1. 1. ] + [ 0. 1. 1. ] + [ 0.5 1. 1. ] + [ 1. 1. 1. ]] + Step3: + Compute output by equation $$Output[i] = C_ * Theta[i]^T$$ + )DOC"); + } +}; + +class AffineGridOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + auto theta_dims = ctx->GetInputDim("Theta"); + if (ctx->HasOutput(framework::GradVarName("Theta"))) { + ctx->SetOutputDim(framework::GradVarName("Theta"), theta_dims); + } + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_CUDA + if (platform::CanCUDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kCUDNN; + } +#endif + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Theta")->type()), + ctx.GetPlace(), framework::DataLayout::kAnyLayout, library_); + } +}; + +class AffineGridGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("affine_grid_grad"); + op->SetInput("Theta", Input("Theta")); + op->SetInput("OutputShape", Input("OutputShape")); + op->SetInput(framework::GradVarName("Output"), OutputGrad("Output")); + + op->SetAttrMap(Attrs()); + + op->SetOutput(framework::GradVarName("Theta"), InputGrad("Theta")); + return std::unique_ptr(op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(affine_grid, ops::AffineGridOp, ops::AffineGridOpMaker, + ops::AffineGridGradMaker); +REGISTER_OPERATOR(affine_grid_grad, ops::AffineGridOpGrad); + +REGISTER_OP_CPU_KERNEL( + affine_grid, + ops::AffineGridOpKernel, + ops::AffineGridOpKernel); +REGISTER_OP_CPU_KERNEL( + affine_grid_grad, + ops::AffineGridGradOpKernel, + ops::AffineGridGradOpKernel); diff --git a/paddle/fluid/operators/affine_grid_op.h b/paddle/fluid/operators/affine_grid_op.h new file mode 100644 index 0000000000000000000000000000000000000000..87d23831486e658374d4c011412fdef57be1b994 --- /dev/null +++ b/paddle/fluid/operators/affine_grid_op.h @@ -0,0 +1,174 @@ +/* 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 "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenTensor = framework::EigenTensor; + +using Array1 = Eigen::DSizes; +using Array2 = Eigen::DSizes; +using Array3 = Eigen::DSizes; +using Array4 = Eigen::DSizes; + +/** + *Return a tensor with evenly spaced numbers over a specified interval. + */ +template +struct Linspace { + void operator()(T start, T end, int count, framework::Tensor* numbers, + const framework::ExecutionContext& ctx); +}; + +template +inline void GetIdxMap(int n, int h, int w, Tensor* grid, + const framework::ExecutionContext& ctx) { + auto& place = *ctx.template device_context().eigen_device(); + grid->mutable_data({n, h, w, 3}, ctx.GetPlace()); + auto grid_t = EigenTensor::From(*grid); + // Get indexes of height with shape [height, width, 1] + Tensor h_idx; + Linspace linspace; + linspace((T)-1, (T)1, h, &h_idx, ctx); + auto h_idx_t = EigenTensor::From(h_idx); + // Get indexes of width with shape [height, width, 1] + Tensor w_idx; + linspace((T)-1, (T)1, w, &w_idx, ctx); + auto w_idx_t = EigenTensor::From(w_idx); + // Get constant ones tensor with shape [height, width, 1] + Tensor ones; + ones.mutable_data({h, w, 1}, ctx.GetPlace()); + auto ones_t = EigenTensor::From(ones).setConstant((T)1); + // Get grid tensor with shape [n, h, w, 3] by concatenating h_idx, w_idx and + // ones + Tensor w_idx_map; + w_idx_map.mutable_data({h, w, 1}, ctx.GetPlace()); + auto w_idx_map_t = EigenTensor::From(w_idx_map); + Tensor h_idx_map; + h_idx_map.mutable_data({h, w, 1}, ctx.GetPlace()); + auto h_idx_map_t = EigenTensor::From(h_idx_map); + Tensor w_h_idx_map; + w_h_idx_map.mutable_data({h, w, 2}, ctx.GetPlace()); + auto w_h_idx_map_t = EigenTensor::From(w_h_idx_map); + Tensor w_h_one_idx_map; + w_h_one_idx_map.mutable_data({h, w, 3}, ctx.GetPlace()); + auto w_h_one_idx_map_t = EigenTensor::From(w_h_one_idx_map); + + w_idx_map_t.device(place) = w_idx_t.reshape(Array2(1, w)) + .broadcast(Array2(h, 1)) + .reshape(Array3(h, w, 1)); + + h_idx_map_t.device(place) = h_idx_t.reshape(Array2(1, h)) + .broadcast(Array2(w, 1)) + .shuffle(Array2(1, 0)) + .reshape(Array3(h, w, 1)); + + w_h_idx_map_t.device(place) = w_idx_map_t.concatenate(h_idx_map_t, 2); + w_h_one_idx_map_t.device(place) = w_h_idx_map_t.concatenate(ones_t, 2); + grid_t.device(place) = w_h_one_idx_map_t.reshape(Array4(1, h, w, 3)) + .broadcast(Array4(n, 1, 1, 1)); +} + +template +class AffineGridOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* theta = ctx.Input("Theta"); + int n = theta->dims()[0]; + auto size_attr = ctx.Attr>("output_shape"); + int h = 0; + int w = 0; + if (size_attr.size() == 0) { + auto* output_shape = ctx.Input("OutputShape"); + Tensor h_sizes; + framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes); + const int* h_size_data = h_sizes.data(); + h = h_size_data[2]; + w = h_size_data[3]; + } else { + h = size_attr[2]; + w = size_attr[3]; + } + auto* output = ctx.Output("Output"); + output->mutable_data({n, h, w, 2}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), output, + static_cast(0)); + Tensor grid; + GetIdxMap(n, h, w, &grid, ctx); + // output = grid * theta.T + // TODO(wanghaoshuang): Refine batched matrix multiply + auto blas = math::GetBlas(ctx); + for (int i = 0; i < n; ++i) { + Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3}); + Tensor sliced_theta = theta->Slice(i, i + 1).Resize({2, 3}); + Tensor sliced_out = output->Slice(i, i + 1).Resize({h * w, 2}); + blas.MatMul(sliced_grid, false, sliced_theta, true, T(1), &sliced_out, + T(0)); + } + } +}; + +template +class AffineGridGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto output_grad = ctx.Input(framework::GradVarName("Output")); + auto theta_grad = ctx.Output(framework::GradVarName("Theta")); + int n = output_grad->dims()[0]; + auto size_attr = ctx.Attr>("output_shape"); + int h = 0; + int w = 0; + if (size_attr.size() == 0) { + auto* output_shape = ctx.Input("OutputShape"); + Tensor h_sizes; + framework::TensorCopy(*output_shape, platform::CPUPlace(), &h_sizes); + const int* h_size_data = h_sizes.data(); + h = h_size_data[2]; + w = h_size_data[3]; + } else { + h = size_attr[2]; + w = size_attr[3]; + } + theta_grad->mutable_data({n, 2, 3}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), theta_grad, + static_cast(0)); + Tensor grid; + GetIdxMap(n, h, w, &grid, ctx); + // output = grid * theta.T + // TODO(wanghaoshuang): Refine batched matrix multiply + auto blas = math::GetBlas(ctx); + for (int i = 0; i < n; ++i) { + Tensor sliced_grid = grid.Slice(i, i + 1).Resize({h * w, 3}); + Tensor sliced_out_grad = output_grad->Slice(i, i + 1).Resize({h * w, 2}); + Tensor sliced_theta_grad = theta_grad->Slice(i, i + 1).Resize({2, 3}); + blas.MatMul(sliced_out_grad, true, sliced_grid, false, T(1), + &sliced_theta_grad, T(0)); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/array_operator.h b/paddle/fluid/operators/array_operator.h index 4309f0a5497456065e5c43bc8f7b265fa711f699..eddf34494bdab18c9d4ae1fb3d1e5d1a71fe590e 100644 --- a/paddle/fluid/operators/array_operator.h +++ b/paddle/fluid/operators/array_operator.h @@ -49,7 +49,7 @@ class ArrayOp : public framework::OperatorBase { } else { offset = static_cast(*i_tensor.data()); } - VLOG(10) << " Offset = " << offset; + VLOG(100) << " Offset = " << offset; return offset; } }; diff --git a/paddle/fluid/operators/array_to_lod_tensor_op.cc b/paddle/fluid/operators/array_to_lod_tensor_op.cc index b8b8b2290a0f002fd379032e28590b84a1da38e9..3c40135eca00f4e0bbff9b0f0f7cf2a4c85ec556 100644 --- a/paddle/fluid/operators/array_to_lod_tensor_op.cc +++ b/paddle/fluid/operators/array_to_lod_tensor_op.cc @@ -11,7 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include +#include #include #include "paddle/fluid/framework/lod_rank_table.h" @@ -148,8 +148,8 @@ class ArrayToLoDTensorOp : public framework::OperatorBase { size_t start_offset = lod_and_offset.second.first; size_t end_offset = lod_and_offset.second.second; - VLOG(10) << "idx=" << idx << " x_idx=" << x_idx << " [" - << ", " << end_offset << "]"; + VLOG(100) << "idx=" << idx << " x_idx=" << x_idx << " [" + << ", " << end_offset << "]"; // Copy data PADDLE_ENFORCE_GE(end_offset, start_offset); size_t len = end_offset - start_offset; diff --git a/paddle/fluid/operators/auc_op.cc b/paddle/fluid/operators/auc_op.cc index 0784920064a879963cd9725cd9acf4cec7b874ce..cb98bc514083ad113fdebfbac043a9516fd9435a 100644 --- a/paddle/fluid/operators/auc_op.cc +++ b/paddle/fluid/operators/auc_op.cc @@ -53,7 +53,7 @@ class AucOp : public framework::OperatorWithKernel { const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Predict")->type()), - ctx.device_context()); + platform::CPUPlace()); } }; diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index 5912a1a17cbd29c3ebd83f37133c044f0905c8bd..cf245f5038f5f5ad1b623542aa14686eff8aad32 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -135,15 +135,13 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Variance", "The global variance (for training) " "or estimated Variance (for testing)"); - AddOutput("Y", "result after normalization").Reuse("X"); + AddOutput("Y", "result after normalization"); AddOutput("MeanOut", "Share memory with Mean. " - "Store the global mean when training") - .Reuse("Mean"); + "Store the global mean when training"); AddOutput("VarianceOut", "Share memory with Variance. " - "Store the global Variance when training") - .Reuse("Variance"); + "Store the global Variance when training"); AddOutput("SavedMean", "Mean of the current mini batch, " "will apply to output when training") @@ -172,6 +170,15 @@ The required data format for this layer is one of the following: } }; +class BatchNormOpInferVarType + : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Y"}}; + } +}; + template class BatchNormKernel : public framework::OpKernel { @@ -527,7 +534,7 @@ class BatchNormGradMaker : public framework::SingleGradOpDescMaker { namespace ops = paddle::operators; REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker, - ops::BatchNormGradMaker); + ops::BatchNormOpInferVarType, ops::BatchNormGradMaker); REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp); REGISTER_OP_CPU_KERNEL( diff --git a/paddle/fluid/operators/batch_norm_op.cu.cc b/paddle/fluid/operators/batch_norm_op.cu.cc index ca6cd8669352fd5814f25a04433ca97fe4abe9ff..0609027c6940533483173209176f3243ccb36f8f 100644 --- a/paddle/fluid/operators/batch_norm_op.cu.cc +++ b/paddle/fluid/operators/batch_norm_op.cu.cc @@ -96,7 +96,7 @@ class BatchNormKernel mode_ = CUDNN_BATCHNORM_SPATIAL; #endif - VLOG(3) << "Setting descriptors."; + VLOG(30) << "Setting descriptors."; std::vector dims; std::vector strides; if (data_layout == DataLayout::kNCHW) { @@ -219,8 +219,8 @@ class BatchNormGradKernel auto *d_bias = ctx.Output(framework::GradVarName("Bias")); d_x->mutable_data(ctx.GetPlace()); - d_scale->mutable_data(ctx.GetPlace()); - d_bias->mutable_data(ctx.GetPlace()); + d_scale->mutable_data>(ctx.GetPlace()); + d_bias->mutable_data>(ctx.GetPlace()); auto &dev_ctx = ctx.template device_context(); if ((N * H * W * D) == 1) { @@ -272,8 +272,10 @@ class BatchNormGradKernel const auto *saved_mean = ctx.Input("SavedMean"); const auto *saved_var = ctx.Input("SavedVariance"); - const void *saved_mean_data = saved_mean->template data(); - const void *saved_var_data = saved_var->template data(); + const void *saved_mean_data = + saved_mean->template data>(); + const void *saved_var_data = + saved_var->template data>(); CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward( dev_ctx.cudnn_handle(), mode_, CudnnDataType::kOne(), @@ -281,10 +283,10 @@ class BatchNormGradKernel CudnnDataType::kZero(), data_desc_, x->template data(), data_desc_, d_y->template data(), data_desc_, d_x->template mutable_data(ctx.GetPlace()), bn_param_desc_, - scale->template data(), - d_scale->template mutable_data(ctx.GetPlace()), - d_bias->template mutable_data(ctx.GetPlace()), epsilon, - saved_mean_data, saved_var_data)); + scale->template data>(), + d_scale->template mutable_data>(ctx.GetPlace()), + d_bias->template mutable_data>(ctx.GetPlace()), + epsilon, saved_mean_data, saved_var_data)); // clean when exit. CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(data_desc_)); @@ -304,4 +306,5 @@ REGISTER_OP_CUDA_KERNEL( ops::BatchNormKernel); REGISTER_OP_CUDA_KERNEL( batch_norm_grad, ops::BatchNormGradKernel, - ops::BatchNormGradKernel); + ops::BatchNormGradKernel, + ops::BatchNormGradKernel); diff --git a/paddle/fluid/operators/beam_search_decode_op.cc b/paddle/fluid/operators/beam_search_decode_op.cc index b6cb935814e25b31d4104f9ce24fe952680cb491..0d32cae0e1e5ff274793df50e854283d8e2f7bf8 100644 --- a/paddle/fluid/operators/beam_search_decode_op.cc +++ b/paddle/fluid/operators/beam_search_decode_op.cc @@ -79,6 +79,9 @@ struct BeamSearchDecodeFunctor { bool tensor_on_gpu_; size_t beam_size_; int end_id_; + // TODO(Superjomn) Here might result serious performance issue in the + // concurrency + // scenarios. const LoDTensorArray& step_ids_origin_; const LoDTensorArray& step_scores_origin_; LoDTensorArray step_ids_ = LoDTensorArray(); diff --git a/paddle/fluid/operators/beam_search_op.cc b/paddle/fluid/operators/beam_search_op.cc index 62771d09f112785ca1ba741a0ba239b1f0234633..791f8a4d3be6780c584997113b7ffcfb7ab35667 100644 --- a/paddle/fluid/operators/beam_search_op.cc +++ b/paddle/fluid/operators/beam_search_op.cc @@ -33,11 +33,11 @@ void BeamSearch::operator()(const framework::LoDTensor &pre_ids, auto items = SelectTopBeamSizeItems(pre_ids, pre_scores); auto selected_items = ToMap(items, high_level.back()); - VLOG(3) << "selected_items:"; + VLOG(30) << "selected_items:"; for (size_t i = 0; i < selected_items.size(); ++i) { - VLOG(3) << "offset:" << i; + VLOG(30) << "offset:" << i; for (auto &item : selected_items[i]) { - VLOG(3) << ItemToString(item); + VLOG(30) << ItemToString(item); } } @@ -138,11 +138,11 @@ std::vector> BeamSearch::SelectTopBeamSizeItems( } result.emplace_back(items); } - VLOG(3) << "SelectTopBeamSizeItems result size " << result.size(); + VLOG(30) << "SelectTopBeamSizeItems result size " << result.size(); for (auto &items : result) { - VLOG(3) << "item set:"; + VLOG(30) << "item set:"; for (auto &item : items) { - VLOG(3) << ItemToString(item); + VLOG(30) << ItemToString(item); } } diff --git a/paddle/fluid/operators/bilinear_interp_op.cu b/paddle/fluid/operators/bilinear_interp_op.cu deleted file mode 100644 index 4c1971538495c6f111e9db18f4014786f6f0dd58..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/bilinear_interp_op.cu +++ /dev/null @@ -1,207 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#include "paddle/fluid/operators/bilinear_interp_op.h" -#include "paddle/fluid/platform/cuda_primitives.h" - -namespace paddle { -namespace operators { - -using framework::Tensor; - -template -__global__ void KeBilinearInterpFw( - const T* in, const size_t in_img_h, const size_t in_img_w, - const size_t input_h, const size_t input_w, T* out, const size_t out_img_h, - const size_t out_img_w, const size_t output_h, const size_t output_w, - const size_t num_channels, const T ratio_h, const T ratioW) { - int nthreads = output_h * output_w; - int tid = blockIdx.x * blockDim.x + threadIdx.x; - if (tid < nthreads) { - int out_id_h = tid / output_w; - int out_id_w = tid % output_w; - int in_img_size = input_w / num_channels; - int out_img_size = output_w / num_channels; - int channel_id = out_id_w / out_img_size; - - int out_img_idy = (out_id_w % out_img_size) / out_img_w; - int in_img_idy = ratio_h * out_img_idy; - int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0; - T h1lambda = ratio_h * out_img_idy - in_img_idy; - T h2lambda = 1.f - h1lambda; - - int out_img_idx = tid % out_img_w; - int in_img_idx = ratioW * out_img_idx; - int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0; - T w1lambda = ratioW * out_img_idx - in_img_idx; - T w2lambda = 1.f - w1lambda; - - const T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + - in_img_idy * in_img_w + in_img_idx]; - - // bilinear interpolation - out[out_id_h * output_w + out_id_w] = - h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[w_id]) + - h1lambda * (w2lambda * in_pos[h_id * in_img_w] + - w1lambda * in_pos[h_id * in_img_w + w_id]); - } -} - -template -__global__ void KeBilinearInterpBw( - T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h, - const size_t input_w, const T* out, const size_t out_img_h, - const size_t out_img_w, const size_t output_h, const size_t output_w, - const size_t num_channels, const T ratio_h, const T ratioW) { - int nthreads = output_h * output_w; - int tid = blockIdx.x * blockDim.x + threadIdx.x; - if (tid < nthreads) { - int out_id_h = tid / output_w; - int out_id_w = tid % output_w; - int in_img_size = input_w / num_channels; - int out_img_size = output_w / num_channels; - int channel_id = out_id_w / out_img_size; - - int out_img_idy = (out_id_w % out_img_size) / out_img_w; - int in_img_idy = ratio_h * out_img_idy; - int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0; - T h1lambda = ratio_h * out_img_idy - in_img_idy; - T h2lambda = 1.f - h1lambda; - - int out_img_idx = tid % out_img_w; - int in_img_idx = ratioW * out_img_idx; - int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0; - T w1lambda = ratioW * out_img_idx - in_img_idx; - T w2lambda = 1.f - w1lambda; - - T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + - in_img_idy * in_img_w + in_img_idx]; - const T* out_pos = &out[out_id_h * output_w + out_id_w]; - atomicAdd(&in_pos[0], h2lambda * w2lambda * out_pos[0]); - atomicAdd(&in_pos[w_id], h2lambda * w1lambda * out_pos[0]); - atomicAdd(&in_pos[h_id * in_img_w], h1lambda * w2lambda * out_pos[0]); - atomicAdd(&in_pos[h_id * in_img_w + w_id], - h1lambda * w1lambda * out_pos[0]); - } -} - -template -class BilinearInterpOpCUDAKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), - "This kernel only runs on GPU device."); - auto* input_t = ctx.Input("X"); // float tensor - auto* output_t = ctx.Output("Out"); // float tensor - auto* input = input_t->data(); - - int out_h = ctx.Attr("out_h"); - int out_w = ctx.Attr("out_w"); - auto out_dims = output_t->dims(); - auto out_size_t = ctx.Input("OutSize"); - if (out_size_t != nullptr) { - Tensor sizes; - framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes); - auto size_data = sizes.data(); - out_h = size_data[0]; - out_w = size_data[1]; - } - auto* output = output_t->mutable_data( - {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace()); - - int batch_size = input_t->dims()[0]; - int channels = input_t->dims()[1]; - int in_h = input_t->dims()[2]; - int in_w = input_t->dims()[3]; - - int in_hw = in_h * in_w; - int out_hw = out_h * out_w; - int in_chw = channels * in_hw; - int out_chw = channels * out_hw; - - T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; - - if (in_h == out_h && in_w == out_w) { - memcpy(output, input, input_t->numel() * sizeof(T)); - } else { - int threadNum = batch_size * out_chw; - int blocks = (threadNum + 1024 - 1) / 1024; - - KeBilinearInterpFw< - T><<>>( - input, in_h, in_w, batch_size, in_chw, output, out_h, out_w, - batch_size, out_chw, channels, ratio_h, ratio_w); - } - } -}; - -template -class BilinearInterpGradOpCUDAKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* d_input_t = ctx.Output(framework::GradVarName("X")); - auto* d_output_t = ctx.Input(framework::GradVarName("Out")); - auto* d_output = d_output_t->data(); - auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); - - auto& device_ctx = - ctx.template device_context(); - math::SetConstant zero; - zero(device_ctx, d_input_t, static_cast(0.0)); - - int out_h = ctx.Attr("out_h"); - int out_w = ctx.Attr("out_w"); - - auto out_size_t = ctx.Input("OutSize"); - if (out_size_t != nullptr) { - Tensor sizes; - framework::TensorCopy(*out_size_t, platform::CPUPlace(), &sizes); - auto size_data = sizes.data(); - out_h = size_data[0]; - out_w = size_data[1]; - } - - int batch_size = d_input_t->dims()[0]; - int channels = d_input_t->dims()[1]; - int in_h = d_input_t->dims()[2]; - int in_w = d_input_t->dims()[3]; - - int in_hw = in_h * in_w; - int out_hw = out_h * out_w; - int in_chw = channels * in_hw; - int out_chw = channels * out_hw; - - T ratio_h = (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - T ratio_w = (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; - - if (in_h == out_h && in_w == out_w) { - memcpy(d_input, d_output, d_input_t->numel() * sizeof(T)); - } else { - int threadNum = batch_size * out_chw; - int blocks = (threadNum + 1024 - 1) / 1024; - - KeBilinearInterpBw< - T><<>>( - d_input, in_h, in_w, batch_size, in_chw, d_output, out_h, out_w, - batch_size, out_chw, channels, ratio_h, ratio_w); - } - } -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(bilinear_interp, - ops::BilinearInterpOpCUDAKernel); -REGISTER_OP_CUDA_KERNEL(bilinear_interp_grad, - ops::BilinearInterpGradOpCUDAKernel); diff --git a/paddle/fluid/operators/bilinear_interp_op.h b/paddle/fluid/operators/bilinear_interp_op.h deleted file mode 100644 index 70847cb8c1abe2e94bc844ab8117d1f23fea533b..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/bilinear_interp_op.h +++ /dev/null @@ -1,163 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - Licensed under the Apache License, Version 2.0 (the "License"); - you may not use this file except in compliance with the License. - You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software - distributed under the License is distributed on an "AS IS" BASIS, - WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - See the License for the specific language governing permissions and - limitations under the License. */ - -#pragma once -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/math/math_function.h" - -namespace paddle { -namespace operators { - -using Tensor = framework::Tensor; - -template -class BilinearInterpKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* input_t = ctx.Input("X"); // float tensor - auto* output_t = ctx.Output("Out"); // float tensor - auto out_dims = output_t->dims(); - auto* input = input_t->data(); - int out_h = ctx.Attr("out_h"); - int out_w = ctx.Attr("out_w"); - auto out_size_t = ctx.Input("OutSize"); - if (out_size_t != nullptr) { - auto out_size_data = out_size_t->data(); - out_h = out_size_data[0]; - out_w = out_size_data[1]; - } - auto* output = output_t->mutable_data( - {out_dims[0], out_dims[1], out_h, out_w}, ctx.GetPlace()); - int batch_size = input_t->dims()[0]; - int channels = input_t->dims()[1]; - int in_h = input_t->dims()[2]; - int in_w = input_t->dims()[3]; - - int in_hw = in_h * in_w; - int out_hw = out_h * out_w; - int in_chw = channels * in_hw; - int out_chw = channels * out_hw; - - float ratio_h = - (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - float ratio_w = - (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; - - if (in_h == out_h && in_w == out_w) { - memcpy(output, input, input_t->numel() * sizeof(T)); - } else { - for (int k = 0; k < batch_size; ++k) { // loop for batches - for (int i = 0; i < out_h; ++i) { // loop for images - int h = ratio_h * i; - int hid = (h < in_h - 1) ? 1 : 0; - float h1lambda = ratio_h * i - h; - float h2lambda = 1.f - h1lambda; - - for (int j = 0; j < out_w; ++j) { - int w = ratio_w * j; - int wid = (w < in_w - 1) ? 1 : 0; - float w1lambda = ratio_w * j - w; - float w2lambda = 1.f - w1lambda; - // calculate four position for bilinear interpolation - const T* in_pos = &input[k * in_chw + h * in_w + w]; - T* out_pos = &output[k * out_chw + i * out_w + j]; - - for (int c = 0; c < channels; ++c) { // loop for channels - // bilinear interpolation - out_pos[0] = static_cast( - h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[wid]) + - h1lambda * (w2lambda * in_pos[hid * in_w] + - w1lambda * in_pos[hid * in_w + wid])); - in_pos += in_hw; - out_pos += out_hw; - } - } - } - } - } - } -}; - -template -class BilinearInterpGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* d_input_t = ctx.Output(framework::GradVarName("X")); - auto* d_output_t = ctx.Input(framework::GradVarName("Out")); - auto* d_output = d_output_t->data(); - auto* d_input = d_input_t->mutable_data(ctx.GetPlace()); - auto& device_ctx = - ctx.template device_context(); - math::SetConstant zero; - zero(device_ctx, d_input_t, static_cast(0.0)); - - int out_h = ctx.Attr("out_h"); - int out_w = ctx.Attr("out_w"); - - auto out_size_t = ctx.Input("OutSize"); - if (out_size_t != nullptr) { - auto out_size_data = out_size_t->data(); - out_h = out_size_data[0]; - out_w = out_size_data[1]; - } - - int batch_size = d_input_t->dims()[0]; - int channels = d_input_t->dims()[1]; - int in_h = d_input_t->dims()[2]; - int in_w = d_input_t->dims()[3]; - - int in_hw = in_h * in_w; - int out_hw = out_h * out_w; - int in_chw = channels * in_hw; - int out_chw = channels * out_hw; - - float ratio_h = - (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; - float ratio_w = - (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; - - if (in_h == out_h && in_w == out_w) { - memcpy(d_input, d_output, d_input_t->numel() * sizeof(T)); - } else { - for (int k = 0; k < batch_size; ++k) { // loop for batches - for (int i = 0; i < out_h; ++i) { // loop for images - int h = ratio_h * i; - int hid = (h < in_h - 1) ? 1 : 0; - float h1lambda = ratio_h * i - h; - float h2lambda = 1 - h1lambda; - - for (int j = 0; j < out_w; ++j) { - int w = ratio_w * j; - int wid = (w < in_w - 1) ? 1 : 0; - float w1lambda = ratio_w * j - w; - float w2lambda = 1 - w1lambda; - T* in_pos = &d_input[k * in_chw + h * in_w + w]; - const T* out_pos = &d_output[k * out_chw + i * out_w + j]; - - for (int c = 0; c < channels; ++c) { // loop for channels - in_pos[0] += static_cast(h2lambda * w2lambda * out_pos[0]); - in_pos[wid] += static_cast(h2lambda * w1lambda * out_pos[0]); - in_pos[hid * in_w] += - static_cast(h1lambda * w2lambda * out_pos[0]); - in_pos[hid * in_w + wid] += - static_cast(h1lambda * w1lambda * out_pos[0]); - in_pos += in_hw; - out_pos += out_hw; - } - } - } - } - } - } -}; - -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/checkpoint_notify_op.cc b/paddle/fluid/operators/checkpoint_notify_op.cc index 3a2527e407bb179c4873fa3ffe2e8f22fb47faf7..defa287bdb913e406aa7e2a383cefc3cca8c4d94 100644 --- a/paddle/fluid/operators/checkpoint_notify_op.cc +++ b/paddle/fluid/operators/checkpoint_notify_op.cc @@ -38,15 +38,16 @@ class CheckpointNotifyOp : public framework::OperatorBase { std::vector epmap = Attr>("epmap"); std::string dir = Attr("dir"); std::string lookup_table_name = Attr("lookup_table"); + int trainer_id = Attr("trainer_id"); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(trainer_id); for (size_t i = 0; i < epmap.size(); i++) { auto lookup_table_save_dir = string::Sprintf("%s/%s_%d", dir, lookup_table_name, i); rpc_client->AsyncCheckpointNotify(epmap[i], lookup_table_save_dir); - VLOG(3) << "checkpoint notify sending lookup table: " << lookup_table_name - << " and dir:" << dir << " to " << epmap[i]; + VLOG(30) << "checkpoint notify sending lookup table: " + << lookup_table_name << " and dir:" << dir << " to " << epmap[i]; } PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); } @@ -63,6 +64,7 @@ class CheckpointNotifyOpMaker : public framework::OpProtoAndCheckerMaker { "dir", "(string, default '') indicate the folder checkpoint will use"); AddAttr("lookup_table", "(string, default '') the lookup table name"); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddComment(R"DOC( CheckpointNotify operator diff --git a/paddle/fluid/operators/clip_by_norm_op.h b/paddle/fluid/operators/clip_by_norm_op.h index 5af0eb0b2ada66d5ae7d521d80e213f9e61f826f..855c4d70677395992e2bf685c910cbea2d37b20b 100644 --- a/paddle/fluid/operators/clip_by_norm_op.h +++ b/paddle/fluid/operators/clip_by_norm_op.h @@ -16,12 +16,15 @@ limitations under the License. */ #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/transform.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; +using SelectedRows = framework::SelectedRows; template using EigenVector = framework::EigenVector; @@ -31,9 +34,40 @@ class ClipByNormKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { auto max_norm = context.Attr("max_norm"); - auto* input = context.Input("X"); - auto* output = context.Output("Out"); - output->mutable_data(context.GetPlace()); + auto in_var = context.InputVar("X"); + + Tensor* output = nullptr; + const Tensor* input = nullptr; + if (in_var->IsType()) { + input = context.Input("X"); + + output = context.Output("Out"); + output->mutable_data(context.GetPlace()); + } else if (in_var->IsType()) { + auto* x = context.Input("X"); + + // merge ids in selected rows first + math::scatter::MergeAdd merge_func; + SelectedRows* merged_input = + const_cast(context.scope()) + .Var() + ->GetMutable(); + merge_func(context.template device_context(), *x, + merged_input); + input = &(merged_input->value()); + + SelectedRows* output_selected_rows = context.Output("Out"); + output_selected_rows->set_rows(merged_input->rows()); + output_selected_rows->set_height(merged_input->height()); + output = output_selected_rows->mutable_value(); + output->Resize(merged_input->value().dims()); + output->mutable_data(context.GetPlace()); + } else { + PADDLE_THROW("Unexpected branch, input variable type is %s", + in_var->Type().name()); + } + + PADDLE_ENFORCE_NOT_NULL(input); auto x = EigenVector::Flatten(*input); auto out = EigenVector::Flatten(*output); diff --git a/paddle/fluid/operators/concat_op.cc b/paddle/fluid/operators/concat_op.cc index 57817da71adfd80faad29a48b05ba2f326de6c07..093b0a9a1f9ac05cf4d72fc748fac827387e5dbe 100644 --- a/paddle/fluid/operators/concat_op.cc +++ b/paddle/fluid/operators/concat_op.cc @@ -37,7 +37,7 @@ class ConcatOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_GT(n, 0, "Input tensors count should > 0."); if (n == 1) { - VLOG(3) << "Warning: concat op have only one input, may waste memory"; + VLOG(30) << "Warning: concat op have only one input, may waste memory"; } auto out_dims = ins[0]; diff --git a/paddle/fluid/operators/concat_op.h b/paddle/fluid/operators/concat_op.h index b2c6495c442cd02679825425becc2160c303dcc6..bd474be0facb349c53a8766412311296383a86c5 100644 --- a/paddle/fluid/operators/concat_op.h +++ b/paddle/fluid/operators/concat_op.h @@ -17,7 +17,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/math/concat.h" +#include "paddle/fluid/operators/math/concat_and_split.h" #include "paddle/fluid/operators/strided_memcpy.h" namespace paddle { @@ -89,29 +89,17 @@ class ConcatGradKernel : public framework::OpKernel { outputs.push_back(nullptr); } } + auto& dev_ctx = ctx.template device_context(); // Sometimes direct copies will be faster, this maybe need deeply analysis. if (axis == 0 && outs.size() < 10) { - size_t input_offset = 0; - const auto in_stride = framework::stride_numel(out_grad->dims()); - - for (size_t i = 0; i < outs.size(); ++i) { - auto out_stride = framework::stride_numel(ins[i]->dims()); - auto* out = outputs[i]; - if (out != nullptr) { - StridedNumelCopyWithAxis( - ctx.device_context(), axis, out->data(), out_stride, - out_grad->data() + input_offset, in_stride, out_stride[axis]); - } - input_offset += out_stride[axis]; - } + std::vector ref_shape; + ref_shape.insert(ref_shape.begin(), ins.begin(), ins.end()); + StridedMemcpyWithAxis0(dev_ctx, *out_grad, ref_shape, &outputs); } else { - auto& dev_ctx = ctx.template device_context(); - paddle::operators::math::ConcatGradFunctor - concat_grad_functor; - concat_grad_functor(dev_ctx, *out_grad, - ctx.MultiInput("X"), - static_cast(axis), &outputs); + math::SplitFunctor split_functor; + split_functor(dev_ctx, *out_grad, ctx.MultiInput("X"), + static_cast(axis), &outputs); } } }; diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc index 4a7a6bcf7154d5680de751e3c933be46fb09fd74..3083e622c3066879e107f930a45bcec36d347f80 100644 --- a/paddle/fluid/operators/conv_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc @@ -15,15 +15,22 @@ limitations under the License. */ #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/memory/memory.h" +#include "paddle/fluid/operators/conv_cudnn_op_cache.h" #include "paddle/fluid/operators/conv_op.h" #include "paddle/fluid/platform/assert.h" #include "paddle/fluid/platform/cudnn_helper.h" #include "paddle/fluid/platform/float16.h" +#include "paddle/fluid/platform/profiler.h" DEFINE_bool(cudnn_deterministic, false, "Whether allow using an autotuning algorithm for convolution " "operator. The autotuning algorithm may be non-deterministic. If " "true, the algorithm is deterministic."); +DEFINE_uint64(conv_workspace_size_limit, 4096, + "cuDNN convolution workspace limit in MB unit."); +DEFINE_bool(cudnn_exhaustive_search, false, + "Whether enable exhaustive search for cuDNN convolution or " + "not, defalut is False."); namespace paddle { namespace operators { @@ -36,13 +43,25 @@ using DataLayout = platform::DataLayout; template using ScalingParamType = typename platform::CudnnDataType::ScalingParamType; +static constexpr char kCUDNNFwdAlgoCache[] = "kCUDNNFwdAlgoCache"; +static constexpr char kCUDNNBwdDataAlgoCache[] = "kCUDNNBwdDataAlgoCache"; +static constexpr char kCUDNNBwdFilterAlgoCache[] = "kCUDNNBwdFilterAlgoCache"; + static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES = static_cast(1024) * 1024 * 1024; +static constexpr size_t kNUM_CUDNN_FWD_ALGS = + CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT; +static constexpr size_t kNUM_CUDNN_BWD_FILTER_ALGS = + CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT; +static constexpr size_t kNUM_CUDNN_BWD_DATA_ALGS = + CUDNN_CONVOLUTION_BWD_DATA_ALGO_COUNT; + template class CUDNNConvOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + auto& dev_ctx = ctx.template device_context(); PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), "It must use CUDAPlace."); auto* input = ctx.Input("Input"); @@ -55,6 +74,8 @@ class CUDNNConvOpKernel : public framework::OpKernel { int groups = ctx.Attr("groups"); int64_t user_workspace_size = static_cast(ctx.Attr("workspace_size_MB")); + bool exhaustive_search = + FLAGS_cudnn_exhaustive_search || ctx.Attr("exhaustive_search"); const T* input_data = input->data(); const T* filter_data = filter->data(); @@ -120,19 +141,19 @@ class CUDNNConvOpKernel : public framework::OpKernel { // ------------------- cudnn conv workspace --------------------- size_t workspace_size_in_bytes; // final workspace to allocate. size_t workspace_size_limit = kCONV_CUDNN_WORKSPACE_LIMIT_BYTES; - if (user_workspace_size > 0) { - workspace_size_limit = user_workspace_size * 1024 * 1024; + if (FLAGS_conv_workspace_size_limit > 0 || user_workspace_size > 0) { + int64_t max_user_size = + std::max(static_cast(FLAGS_conv_workspace_size_limit), + user_workspace_size); + workspace_size_limit = max_user_size * 1024 * 1024; } + // ------------------- cudnn conv algorithm --------------------- cudnnConvolutionFwdAlgo_t algo; - auto& dev_ctx = ctx.template device_context(); auto handle = dev_ctx.cudnn_handle(); + auto workspace_handle = dev_ctx.cudnn_workspace_handle(); - CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( - handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, - cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, - workspace_size_limit, &algo)); - + bool half_float = false; #if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1) // Tensor core is supported since the volta GPU and // is only enabled when input and filter data are float16 @@ -143,12 +164,66 @@ class CUDNNConvOpKernel : public framework::OpKernel { cudnn_conv_desc, CUDNN_TENSOR_OP_MATH)); // Currently tensor core is only enabled using this algo algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM; + half_float = true; + VLOG(50) << "use cudnn_tensor_op_math"; } else { CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( cudnn_conv_desc, CUDNN_DEFAULT_MATH)); + VLOG(50) << "NOT use cudnn_tensor_op_math"; } #endif + auto x_dims = framework::vectorize(input->dims()); + auto f_dims = framework::vectorize(filter->dims()); + if ((!exhaustive_search) && (!half_float)) { + CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( + handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, + cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, + workspace_size_limit, &algo)); + VLOG(3) << "cuDNN forward algo " << algo; + } else if (exhaustive_search && (!half_float)) { + AlgorithmsCache* algo_cache = nullptr; + if (ctx.scope().FindVar(kCUDNNFwdAlgoCache)) { + algo_cache = + ctx.scope() + .FindVar(kCUDNNFwdAlgoCache) + ->GetMutable>(); + } else { + algo_cache = + const_cast(ctx.scope()) + .Var(kCUDNNFwdAlgoCache) + ->GetMutable>(); + } + algo = algo_cache->GetAlgorithm( + x_dims, f_dims, strides, paddings, dilations, 0, [&]() { + int returned_algo_count; + std::array + fwd_perf_stat; + auto cudnn_find_func = [&](void* cudnn_workspace) { + CUDNN_ENFORCE( + platform::dynload::cudnnFindConvolutionForwardAlgorithmEx( + handle, cudnn_input_desc, input_data, cudnn_filter_desc, + filter_data, cudnn_conv_desc, cudnn_output_desc, + output_data, kNUM_CUDNN_FWD_ALGS, &returned_algo_count, + fwd_perf_stat.data(), cudnn_workspace, + workspace_size_limit)); + }; + workspace_handle.RunFunc(cudnn_find_func, workspace_size_limit); + + VLOG(3) << "Perf result: (algo: stat, time, memory)"; + for (int i = 0; i < returned_algo_count; ++i) { + const auto& stat = fwd_perf_stat[i]; + VLOG(3) << stat.algo << ": " << stat.status << " " << stat.time + << " " << stat.memory; + } + return fwd_perf_stat[0].algo; + }); + VLOG(3) << "choose algo " << algo; + } else { + PADDLE_ENFORCE(half_float, + "cuDNN exhaustive search doesn't support half float."); + } + // get workspace size able to allocate CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize( handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, @@ -168,7 +243,7 @@ class CUDNNConvOpKernel : public framework::OpKernel { cudnn_conv_desc, algo, cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_output_desc, output_data + i * group_offset_out)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } }; @@ -177,6 +252,7 @@ template class CUDNNConvGradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + auto& dev_ctx = ctx.template device_context(); PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), "It must use CUDAPlace."); auto input = ctx.Input("Input"); @@ -195,6 +271,13 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { int groups = ctx.Attr("groups"); int64_t user_workspace_size = static_cast(ctx.Attr("workspace_size_MB")); + bool exhaustive_search = + FLAGS_cudnn_exhaustive_search || ctx.Attr("exhaustive_search"); + if (exhaustive_search && FLAGS_cudnn_deterministic) { + PADDLE_THROW( + "Cann't set exhaustive_search True and " + "FLAGS_cudnn_deterministic True at same time."); + } // ------------------- cudnn descriptors --------------------- ScopedTensorDescriptor input_desc; @@ -262,14 +345,66 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { cudnnConvolutionBwdFilterAlgo_t filter_algo; size_t workspace_size_in_bytes = 0, tmp_size = 0; size_t workspace_size_limit = kCONV_CUDNN_WORKSPACE_LIMIT_BYTES; - if (user_workspace_size > 0) { - workspace_size_limit = user_workspace_size * 1024 * 1024; + if (FLAGS_conv_workspace_size_limit > 0 || user_workspace_size > 0) { + int64_t max_user_size = + std::max(static_cast(FLAGS_conv_workspace_size_limit), + user_workspace_size); + workspace_size_limit = max_user_size * 1024 * 1024; } - auto& dev_ctx = ctx.template device_context(); + auto x_dims = framework::vectorize(input->dims()); + auto f_dims = framework::vectorize(filter->dims()); auto handle = dev_ctx.cudnn_handle(); + auto workspace_handle = dev_ctx.cudnn_workspace_handle(); if (input_grad) { - if (!FLAGS_cudnn_deterministic) { + T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); + if (exhaustive_search) { + AlgorithmsCache* data_algo_cache; + if (ctx.scope().FindVar(kCUDNNBwdDataAlgoCache)) { + data_algo_cache = + ctx.scope() + .FindVar(kCUDNNBwdDataAlgoCache) + ->GetMutable< + AlgorithmsCache>(); + } else { + data_algo_cache = + const_cast(ctx.scope()) + .Var(kCUDNNBwdDataAlgoCache) + ->GetMutable< + AlgorithmsCache>(); + } + data_algo = data_algo_cache->GetAlgorithm( + x_dims, f_dims, strides, paddings, dilations, 0, [&]() { + int returned_algo_count; + std::array + data_perf_stat; + auto cudnn_find_bd_data_func = [&](void* cudnn_workspace) { + CUDNN_ENFORCE( + platform::dynload:: + cudnnFindConvolutionBackwardDataAlgorithmEx( + handle, cudnn_filter_desc, filter_data, + cudnn_output_grad_desc, output_grad_data, + cudnn_conv_desc, cudnn_input_desc, input_grad_data, + kNUM_CUDNN_BWD_DATA_ALGS, &returned_algo_count, + data_perf_stat.data(), cudnn_workspace, + workspace_size_limit)); + }; + workspace_handle.RunFunc(cudnn_find_bd_data_func, + workspace_size_limit); + + VLOG(3) << "Perf result: (algo: stat, time, memory)"; + for (int i = 0; i < returned_algo_count; ++i) { + const auto& stat = data_perf_stat[i]; + VLOG(3) << stat.algo << ": " << stat.status << " " << stat.time + << " " << stat.memory; + } + return data_perf_stat[0].algo; + }); + VLOG(3) << "cuDNN backward data algo " << data_algo; + } else if (FLAGS_cudnn_deterministic) { + data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1; + } else { CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm( handle, cudnn_filter_desc, @@ -282,10 +417,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { cudnn_input_desc, CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, workspace_size_limit, &data_algo)); - } else { - data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1; } - CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( handle, cudnn_filter_desc, cudnn_output_grad_desc, @@ -294,17 +426,54 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { } if (filter_grad) { - if (!FLAGS_cudnn_deterministic) { + T* filter_grad_data = filter_grad->mutable_data(ctx.GetPlace()); + if (exhaustive_search) { + AlgorithmsCache* f_algo_cache; + if (ctx.scope().FindVar(kCUDNNBwdFilterAlgoCache)) { + f_algo_cache = + ctx.scope() + .FindVar(kCUDNNBwdFilterAlgoCache) + ->GetMutable< + AlgorithmsCache>(); + } else { + f_algo_cache = + const_cast(ctx.scope()) + .Var(kCUDNNBwdFilterAlgoCache) + ->GetMutable< + AlgorithmsCache>(); + } + filter_algo = f_algo_cache->GetAlgorithm( + x_dims, f_dims, strides, paddings, dilations, 0, [&]() { + int returned_algo_count; + std::array + filter_perf_stat; + auto cudnn_find_bd_f_func = [&](void* cudnn_workspace) { + CUDNN_ENFORCE( + platform::dynload:: + cudnnFindConvolutionBackwardFilterAlgorithmEx( + handle, cudnn_input_desc, input_data, + cudnn_output_grad_desc, output_grad_data, + cudnn_conv_desc, cudnn_filter_desc, + filter_grad_data, kNUM_CUDNN_BWD_FILTER_ALGS, + &returned_algo_count, filter_perf_stat.data(), + cudnn_workspace, workspace_size_limit)); + }; + workspace_handle.RunFunc(cudnn_find_bd_f_func, + workspace_size_limit); + return filter_perf_stat[0].algo; + }); + VLOG(3) << "cuDNN backward filter algo " << filter_algo; + } else if (FLAGS_cudnn_deterministic) { + filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1; + } else { CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm( handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc, cudnn_filter_desc, CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, workspace_size_limit, &filter_algo)); - } else { - filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1; } - CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( handle, cudnn_input_desc, cudnn_output_grad_desc, cudnn_conv_desc, @@ -327,7 +496,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { data_algo, cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc, input_grad_data + i * group_offset_in)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } // ------------------- cudnn conv backward filter --------------------- @@ -343,7 +512,7 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { filter_algo, cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_filter_desc, filter_grad_data + i * group_offset_filter)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } } @@ -359,7 +528,8 @@ REGISTER_OP_KERNEL(conv2d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel); REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvGradOpKernel, - paddle::operators::CUDNNConvGradOpKernel); + paddle::operators::CUDNNConvGradOpKernel, + paddle::operators::CUDNNConvGradOpKernel); REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel, diff --git a/paddle/fluid/operators/conv_cudnn_op_cache.h b/paddle/fluid/operators/conv_cudnn_op_cache.h new file mode 100644 index 0000000000000000000000000000000000000000..4b534321f746d5620005743eb8d45b71259156dd --- /dev/null +++ b/paddle/fluid/operators/conv_cudnn_op_cache.h @@ -0,0 +1,90 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include + +namespace paddle { +namespace operators { + +template +class AlgorithmsCache { + public: + // Caches the best algorithm for a given + // combination of tensor dimensions & compute data type. + TAlgorithm GetAlgorithm( + const std::vector& dims1, const std::vector& dims2, + const std::vector& strides, const std::vector& paddings, + const std::vector& dilations, + int algorithmFlags, // can set for different data type + std::function gen_func); + + private: + std::unordered_map hash_; + std::mutex mutex_; +}; + +template +TAlgorithm AlgorithmsCache::GetAlgorithm( + const std::vector& dims1, const std::vector& dims2, + const std::vector& strides, const std::vector& paddings, + const std::vector& dilations, int algorithmFlags, + std::function gen_func) { + std::lock_guard lock(mutex_); + int64_t seed = 0; + // Hash all of the inputs, use to try and look up a previously + // discovered algorithm, or fall back to generating a new one. + std::hash hashFn; + // do hash like boost + // https://stackoverflow.com/questions/2590677/how-do-i-combine-hash-values-in-c0x + for (const auto num : dims1) { + seed ^= hashFn(num) + 0x9e3779b9 + (seed << 6) + (seed >> 2); + } + + for (const auto num : dims2) { + seed ^= hashFn(num) + 0x9e3779b9 + (seed << 6) + (seed >> 2) + 1; + } + + for (const auto num : strides) { + seed ^= hashFn(static_cast(num)) + 0x9e3779b9 + (seed << 6) + + (seed >> 2) + 2; + } + + for (const auto num : paddings) { + seed ^= hashFn(static_cast(num)) + 0x9e3779b9 + (seed << 6) + + (seed >> 2) + 3; + } + + for (const auto num : dilations) { + seed ^= hashFn(static_cast(num)) + 0x9e3779b9 + (seed << 6) + + (seed >> 2) + 4; + } + + seed ^= hashFn(static_cast(algorithmFlags)) + 0x9e3779b9 + + (seed << 6) + (seed >> 2) + 5; + + if (seed == 0) return gen_func(); + + if (hash_.find(seed) == hash_.end()) { + TAlgorithm value = gen_func(); + hash_[seed] = value; + } + return hash_[seed]; +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc index eae65968285703f5882d910e29bc5d8e1511cba6..f2cc6642ee6c45cfd95fa3b5ccc58a4832fb8db4 100644 --- a/paddle/fluid/operators/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/conv_mkldnn_op.cc @@ -15,6 +15,8 @@ #include "paddle/fluid/operators/conv_op.h" #include "paddle/fluid/platform/mkldnn_helper.h" +#include "paddle/fluid/framework/data_layout_transform.h" + namespace paddle { namespace operators { @@ -57,6 +59,11 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler { return conv_pd_->dst_primitive_desc().get_size(); } + mkldnn::memory::format GetDstFormat() const { + return static_cast( + conv_pd_->dst_primitive_desc().desc().data.format); + } + size_t GetDiffWeightsMemorySize() const { return conv_bwd_weights_pd_->diff_weights_primitive_desc().get_size(); } @@ -108,6 +115,20 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler { "@data-weights_mem_p", pipeline); } + std::shared_ptr AcquireResidualDataMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_residual_data_mem_p"); + } + + std::shared_ptr AcquireDstMemoryFromResidualDataMemory( + const std::shared_ptr& user_residual_memory_p, + void* dst_ptr, + std::vector& pipeline) { // NOLINT + return this->AcquireMemory(user_residual_memory_p, + this->AcquireDstMemoryFromPrimitive(dst_ptr), + "@residual_data_mem_p", pipeline); + } + std::shared_ptr AcquireDiffSrcMemoryFromDataPrimitive( void* ptr) { return this->AcquireMemoryFromPrimitive( @@ -300,10 +321,10 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { std::vector paddings = ctx.Attr>("paddings"); std::vector dilations = ctx.Attr>("dilations"); bool fuse_relu = ctx.Attr("fuse_relu"); - bool fuse_eltwise = ctx.Attr("fuse_eltwise"); + bool fuse_residual_conn = ctx.Attr("fuse_residual_connection"); int groups = ctx.Attr("groups"); - // TODO: add support for dilation + // TODO(tpatejko): add support for dilation PADDLE_ENFORCE( dilations.size() == 2 && dilations[0] == 1 && dilations[1] == 1, "dilation in convolution is not implemented yet"); @@ -354,8 +375,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto src_md = platform::MKLDNNMemDesc( src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); auto weights_md = platform::MKLDNNMemDesc( - weights_tz, platform::MKLDNNGetDataType(), - (g == 1) ? chosen_memory_format : mkldnn::memory::format::goihw); + weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); std::vector bias_tz; // TODO(mgallus): avoid empty vector creation. // Currently used whenever bias is != nullptr. auto dst_md = platform::MKLDNNMemDesc( @@ -369,11 +389,11 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { bias_tz, platform::MKLDNNGetDataType(), memory::format::x); conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md, strides, paddings, mkldnn_engine, - fuse_relu, fuse_eltwise); + fuse_relu, fuse_residual_conn); } else { conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings, - mkldnn_engine, fuse_relu, fuse_eltwise); + mkldnn_engine, fuse_relu, fuse_residual_conn); } // Save conv_pd/src_memory/weights_memory for backward pass dev_ctx.SetBlob(key_conv_pd, conv_pd); @@ -386,15 +406,52 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto user_weights_memory_p = handler.AcquireWeightsMemory( user_weights_md, to_void_cast(filter_data)); - T* output_data = - output->mutable_data(ctx.GetPlace(), handler.GetDstMemorySize()); // create reorder primitive if the input format is not the preferred one auto src_memory_p = handler.AcquireSrcMemoryFromPrimitive(user_src_memory_p, pipeline); auto weights_memory_p = handler.AcquireWeightsMemoryFromPrimitive( user_weights_memory_p, pipeline, is_test); - auto dst_memory_p = - handler.AcquireDstMemoryFromPrimitive(to_void_cast(output_data)); + + std::shared_ptr dst_memory_p; + + if (fuse_residual_conn) { + auto residual_param = ctx.Input("ResidualData"); + auto residual_param_data = residual_param->data(); + + PADDLE_ENFORCE( + residual_param_data != nullptr, + "Provide data if you want MKLDNN conv+elementwise_add fusion"); + PADDLE_ENFORCE_EQ(output->dims(), residual_param->dims(), + "Output and elementwise parameter need to have the " + "same dimension sizes"); + + if (residual_param->format() != handler.GetDstFormat()) { + auto output_data = + output->mutable_data(ctx.GetPlace(), handler.GetDstMemorySize()); + auto residual_data_tz = + paddle::framework::vectorize2int(residual_param->dims()); + auto residual_data_type = + paddle::framework::ToMKLDNNDataType(residual_param->type()); + + auto user_residual_md = platform::MKLDNNMemDesc( + residual_data_tz, residual_data_type, residual_param->format()); + auto user_residual_memory_p = handler.AcquireResidualDataMemory( + user_residual_md, to_void_cast(residual_param_data)); + + dst_memory_p = handler.AcquireDstMemoryFromResidualDataMemory( + user_residual_memory_p, to_void_cast(output_data), pipeline); + } else { + output->ShareDataWith(*residual_param); + auto output_data = output->mutable_data(ctx.GetPlace()); + dst_memory_p = + handler.AcquireDstMemoryFromPrimitive(to_void_cast(output_data)); + } + } else { + auto output_data = + output->mutable_data(ctx.GetPlace(), handler.GetDstMemorySize()); + dst_memory_p = + handler.AcquireDstMemoryFromPrimitive(to_void_cast(output_data)); + } // create convolution op primitive std::shared_ptr conv_p; @@ -424,14 +481,15 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { private: mkldnn::primitive_attr CreatePostOps(bool fuse_relu, - bool fuse_eltwise) const { + bool fuse_residual_conn) const { mkldnn::primitive_attr conv_attr; mkldnn::post_ops post_operations; // Fusion with Elementwise layer relies on adding a sum post-operation with - // the scale parameter. It is assumed that when fuse_eltwise is true, the - // Output tensor contains the data coming from residual connection. The - // result of this post_op is: Output = scale * Output + Conv_Out. - if (fuse_eltwise) { + // the scale parameter. It is assumed that when fuse_residual_connection is + // true, the output tensor contains the data coming from residual + // connection. The result of this post_op is: + // Output = scale * Output + Conv_Out. + if (fuse_residual_conn) { post_operations.append_sum(1.0f); } // Fusion with ReLU layer is executed through the PostOps feature. Create a @@ -452,7 +510,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { const memory::desc& dst, const std::vector& strides, const std::vector& paddings, const mkldnn::engine& engine, const bool fuse_relu, - const bool fuse_eltwise) const { + const bool fuse_residual_conn) const { memory::dims stride_dims = {strides[0], strides[1]}; memory::dims padding_dims = {paddings[0], paddings[1]}; @@ -461,7 +519,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { dst, stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero); - mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise); + mkldnn::primitive_attr conv_attr = + CreatePostOps(fuse_relu, fuse_residual_conn); auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc( conv_desc, conv_attr, engine); @@ -476,7 +535,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { const std::vector& strides, const std::vector& paddings, const mkldnn::engine& engine, const bool fuse_relu, - const bool fuse_eltwise) const { + const bool fuse_residual_conn) const { memory::dims stride_dims = {strides[0], strides[1]}; memory::dims padding_dims = {paddings[0], paddings[1]}; @@ -485,7 +544,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { bias, dst, stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero); - mkldnn::primitive_attr conv_attr = CreatePostOps(fuse_relu, fuse_eltwise); + mkldnn::primitive_attr conv_attr = + CreatePostOps(fuse_relu, fuse_residual_conn); auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc( conv_desc, conv_attr, engine); diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 8f84bf71a7f77606bed6672f0830e3fc80165a42..4d370746382a4247f51aafa189e86eece941c320 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -130,8 +130,12 @@ void Conv2DOpMaker::Make() { .AsDispensable(); AddOutput("Output", "(Tensor) The output tensor of convolution operator. " - "The format of output tensor is also NCHW.") - .Reuse("Input"); + "The format of output tensor is also NCHW."); + AddInput("ResidualData", + "(Tensor) Tensor with residual data " + "to which convolution output will be added." + "Used with fuse_residual_connection fusion.") + .AsDispensable(); AddAttr>("strides", "(vector default:{1, 1}), the " "strides(h_stride, w_stride) of " @@ -164,10 +168,10 @@ void Conv2DOpMaker::Make() { .SetDefault(false); AddAttr("fuse_relu", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); - AddAttr("fuse_eltwise", + AddAttr("fuse_residual_connection", "(bool, default false) Only used in mkldnn kernel. Used " - "whenever convolution output is connected via skip connection " - "to a previous layer.") + "whenever convolution output is as an input to residual " + "connection.") .SetDefault(false); AddAttr( "data_format", @@ -185,6 +189,11 @@ void Conv2DOpMaker::Make() { "workspace size can increase performance but also requires " "better hardware. This size should be chosen carefully.") .SetDefault(4096); + AddAttr("exhaustive_search", + "(bool, default false) cuDNN has many algorithm to calculation " + "convolution, whether enable exhaustive search ", + "for cuDNN convolution or not, defalut is False.") + .SetDefault(false); AddComment(R"DOC( Convolution Operator. @@ -215,6 +224,15 @@ $$ )DOC"); } +class ConvOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{ + {"Input", /*->*/ "Output"}}; + } +}; + void Conv3DOpMaker::Make() { AddInput( "Input", @@ -233,8 +251,7 @@ void Conv3DOpMaker::Make() { "input image channels divided by the groups."); AddOutput("Output", "(Tensor) The output tensor of convolution operator." - "The format of output tensor is also NCDHW.") - .Reuse("Input"); + "The format of output tensor is also NCDHW."); AddAttr>("strides", "(vector, default:{1, 1, 1}), the " "strides(d_stride, h_stride, w_stride) of " @@ -280,7 +297,11 @@ void Conv3DOpMaker::Make() { "workspace size can increase performance but also requires " "better hardware. This size should be chosen carefully.") .SetDefault(4096); - + AddAttr("exhaustive_search", + "(bool, default false) cuDNN has many algorithm to calculation " + "convolution, whether enable exhaustive search ", + "for cuDNN convolution or not, defalut is False.") + .SetDefault(false); AddComment(R"DOC( Convolution3D Operator. @@ -353,6 +374,7 @@ framework::OpKernelType ConvOpGrad::GetExpectedKernelType( namespace ops = paddle::operators; REGISTER_OPERATOR(conv2d, ops::ConvOp, ops::Conv2DOpMaker, + ops::ConvOpInferVarType, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(conv2d_grad, ops::ConvOpGrad); @@ -360,7 +382,9 @@ REGISTER_OPERATOR(conv2d_grad, ops::ConvOpGrad); REGISTER_OPERATOR(depthwise_conv2d, ops::ConvOp, ops::Conv2DOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(depthwise_conv2d_grad, ops::ConvOpGrad); + REGISTER_OPERATOR(conv3d, ops::ConvOp, ops::Conv3DOpMaker, + ops::ConvOpInferVarType, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(conv3d_grad, ops::ConvOpGrad); diff --git a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc index 73831611d01b8c5b8d2d9f7f15634a0094e4a608..f44094ca6b7b7f23f2e7593ad79e4e2a6f0d3070 100644 --- a/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc @@ -104,6 +104,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { int output_offset = output->numel() / output->dims()[0] / groups; int filter_offset = filter->numel() / groups; T alpha = 1.0f, beta = 0.0f; + auto workspace_handle = dev_ctx.cudnn_workspace_handle(); for (int g = 0; g < groups; g++) { auto cudnn_func = [&](void* cudnn_workspace) { CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData( @@ -112,7 +113,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel { algo, cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_output_desc, output_data + output_offset * g)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } }; @@ -208,6 +209,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { output_grad->numel() / output_grad->dims()[0] / groups; int filter_offset = filter->numel() / groups; T alpha = 1.0f, beta = 0.0f; + auto workspace_handle = dev_ctx.cudnn_workspace_handle(); if (input_grad) { T* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); // Because beta is zero, it is unnecessary to reset input_grad. @@ -220,7 +222,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc, input_grad_data + input_offset * g)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } @@ -238,7 +240,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel { cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_filter_desc, filter_grad_data + filter_offset * g)); }; - dev_ctx.RunCudnnFuncWithWorkspace(cudnn_func, workspace_size_in_bytes); + workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes); } } } diff --git a/paddle/fluid/operators/crf_decoding_op.h b/paddle/fluid/operators/crf_decoding_op.h index 8181897c3d3844bda5574e85a08b2af038fcd664..e9d2e84a434d7084c526a6e75363a65577197262 100644 --- a/paddle/fluid/operators/crf_decoding_op.h +++ b/paddle/fluid/operators/crf_decoding_op.h @@ -16,6 +16,7 @@ limitations under the License. */ #include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/jit_kernel.h" #include "paddle/fluid/operators/math/math_function.h" namespace paddle { @@ -69,9 +70,6 @@ class CRFDecodingOpKernel : public framework::OpKernel { auto emission_dims = emission_weights.dims(); const size_t seq_len = emission_dims[0]; const size_t tag_num = emission_dims[1]; - - const size_t state_trans_base_idx = 2; - const T* x = emission_weights.data(); const T* w = transition_weights.data(); int64_t* path = decoded_path->data(); @@ -84,221 +82,10 @@ class CRFDecodingOpKernel : public framework::OpKernel { Tensor track; int* track_value = track.mutable_data(emission_dims, platform::CPUPlace()); - -#ifdef __AVX__ -// It use the AVX or AVX512 instruction to deal the data as the vector of 8 or -// 16 elements per iteration. Then it can implement the parallel processing. -// Only optimize for float type. -#ifdef __AVX512F__ - size_t step_size = 16; -#else - size_t step_size = 8; -#endif - if (std::is_same::value && (tag_num >= step_size)) { - size_t steps = tag_num / step_size; - size_t remain = tag_num % step_size; - int last_offset = static_cast(remain) - static_cast(step_size); - - // Setup the alpha initial value. - size_t i_offset = 0; - for (size_t i = 0; i <= steps; ++i) { -#ifdef __AVX512F__ - // Declare the variable for the content of weights, input and alpha - // values. - __m512 w_content, x_content, alpha_content; - - // Load the relevant data into the variables from un-aligned address. - w_content = _mm512_loadu_ps((const float*)(w + i_offset)); - x_content = _mm512_loadu_ps((const float*)(x + i_offset)); - alpha_content = _mm512_add_ps(w_content, x_content); - - // Save the alpha value. - _mm512_storeu_ps(reinterpret_cast(alpha_value + i_offset), - alpha_content); -#else - // Declare the variable for the content of weights, input and alpha - // values. - __m256 w_content, x_content, alpha_content; - - // Load the relevant data into the variables from un-aligned address. - w_content = _mm256_loadu_ps((const float*)(w + i_offset)); - x_content = _mm256_loadu_ps((const float*)(x + i_offset)); - alpha_content = _mm256_add_ps(w_content, x_content); - - // Save the alpha value. - _mm256_storeu_ps(reinterpret_cast(alpha_value + i_offset), - alpha_content); -#endif - i_offset += step_size; - if (i == steps - 1) { - if (remain > 0) { - i_offset += last_offset; - } else { - break; - } - } - } - - // Use the column-major strategy to get the location of maximum score. - size_t seq_offset = 0; - for (size_t k = 1; k < seq_len; ++k) { - size_t j_offset = 0; - for (size_t j = 0; j <= steps; ++j) { -#ifdef __AVX512F__ - // Initialize the variables of maximum score and location. - __m512 max_score = _mm512_set1_ps(-std::numeric_limits::max()); - __m512i max_j = _mm512_setzero_si512(); -#else - // Initialize the variables of maximum score and location. - __m256 max_score = _mm256_set1_ps(-std::numeric_limits::max()); - __m256i max_j = _mm256_set1_epi32(0); -#endif - // Calculate the offset of transition_weights. - size_t trans_offset = state_trans_base_idx * tag_num + j_offset; - for (size_t i = 0; i < tag_num; ++i) { -#ifdef __AVX512F__ - // Initalize the content of alpha variable with related offset. - __m512 alpha_content = - _mm512_set1_ps(*(const float*)(alpha_value + seq_offset + i)); - // Obtain the content of weights from un-aligned address. - __m512 w_content = - _mm512_loadu_ps((const float*)(w + trans_offset)); - - __m512 score_v = _mm512_add_ps(alpha_content, w_content); - - __mmask16 mask = _mm512_cmp_ps_mask(score_v, max_score, _CMP_GT_OS); - - // According to the mask value, it update the index of the max_score - // location. - max_j = _mm512_mask_set1_epi32(max_j, mask, i); - - // Update the max_score value. - max_score = _mm512_max_ps(max_score, score_v); -#else - // Initalize the content of alpha variable with related offset. - __m256 alpha_content = _mm256_broadcast_ss( - (const float*)(alpha_value + seq_offset + i)); - // Obtain the content of weights from un-aligned address. - __m256 w_content = - _mm256_loadu_ps((const float*)(w + trans_offset)); - __m256 score_v = _mm256_add_ps(alpha_content, w_content); - - __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); - -#ifdef __AVX2__ - // According to the mask value, it update the index of the max_score - // location. - max_j = _mm256_or_si256( - _mm256_andnot_si256((__m256i)mask, max_j), - _mm256_and_si256((__m256i)mask, _mm256_set1_epi32(i))); -#else - __m128i lo_max_j = _mm256_extractf128_si256(max_j, 0); - __m128i hi_max_j = _mm256_extractf128_si256(max_j, 1); - __m128i lo_mask = _mm256_extractf128_si256((__m256i)mask, 0); - __m128i hi_mask = _mm256_extractf128_si256((__m256i)mask, 1); - - lo_max_j = _mm_andnot_si128(lo_mask, lo_max_j); - hi_max_j = _mm_andnot_si128(hi_mask, hi_max_j); - lo_mask = _mm_and_si128(lo_mask, _mm_set1_epi32(i)); - hi_mask = _mm_and_si128(hi_mask, _mm_set1_epi32(i)); - - lo_max_j = _mm_or_si128(lo_mask, lo_max_j); - hi_max_j = _mm_or_si128(hi_mask, hi_max_j); - - // According to the mask value, it update the index of the max_score - // location. - max_j = _mm256_insertf128_si256(max_j, lo_max_j, 0); - max_j = _mm256_insertf128_si256(max_j, hi_max_j, 1); -#endif - - // Update the max_score value. - max_score = _mm256_max_ps(max_score, score_v); -#endif - trans_offset += tag_num; - } - -#ifdef __AVX512F__ - // Update the alpha and track values. - __m512 x_content = _mm512_loadu_ps( - (const float*)(x + seq_offset + tag_num + j_offset)); - max_score = _mm512_add_ps(max_score, x_content); - _mm512_storeu_ps(reinterpret_cast(alpha_value + seq_offset + - tag_num + j_offset), - max_score); - _mm512_storeu_si512( - reinterpret_cast<__m512i*>(track_value + seq_offset + tag_num + - j_offset), - max_j); -#else - // Update the alpha and track values. - __m256 x_content = _mm256_loadu_ps( - (const float*)(x + seq_offset + tag_num + j_offset)); - max_score = _mm256_add_ps(max_score, x_content); - _mm256_storeu_ps(reinterpret_cast(alpha_value + seq_offset + - tag_num + j_offset), - max_score); - _mm256_storeu_si256( - reinterpret_cast<__m256i*>(track_value + seq_offset + tag_num + - j_offset), - max_j); -#endif - - // Calculate the offset of next step - j_offset += step_size; - if (j == steps - 1) { - if (remain > 0) { - j_offset += last_offset; - } else { - break; - } - } - } - - seq_offset += tag_num; - } - } else { - for (size_t i = 0; i < tag_num; ++i) alpha_value[i] = w[i] + x[i]; - - for (size_t k = 1; k < seq_len; ++k) { - for (size_t i = 0; i < tag_num; ++i) { - T max_score = -std::numeric_limits::max(); - int max_j = 0; - for (size_t j = 0; j < tag_num; ++j) { - T score = alpha_value[(k - 1) * tag_num + j] + - w[(j + state_trans_base_idx) * tag_num + i]; - if (score > max_score) { - max_score = score; - max_j = j; - } - } - - alpha_value[k * tag_num + i] = max_score + x[k * tag_num + i]; - track_value[k * tag_num + i] = max_j; - } - } - } -#else - for (size_t i = 0; i < tag_num; ++i) alpha_value[i] = w[i] + x[i]; - - for (size_t k = 1; k < seq_len; ++k) { - for (size_t i = 0; i < tag_num; ++i) { - T max_score = -std::numeric_limits::max(); - int max_j = 0; - for (size_t j = 0; j < tag_num; ++j) { - T score = alpha_value[(k - 1) * tag_num + j] + - w[(j + state_trans_base_idx) * tag_num + i]; - if (score > max_score) { - max_score = score; - max_j = j; - } - } - - alpha_value[k * tag_num + i] = max_score + x[k * tag_num + i]; - track_value[k * tag_num + i] = max_j; - } - } - -#endif + const auto& ker = math::jitkernel::KernelPool::Instance() + .template Get>( + static_cast(tag_num)); + ker->Compute(static_cast(seq_len), x, w, alpha_value, track_value); T max_score = -std::numeric_limits::max(); int max_i = 0; for (size_t i = 0; i < tag_num; ++i) { diff --git a/paddle/fluid/operators/cross_entropy_op.cc b/paddle/fluid/operators/cross_entropy_op.cc index 66f19fe7ecfa51b2ce917f0c5fcb6d486f1a7307..a904dd91302c951560dc32ac107d4d73b6024c25 100644 --- a/paddle/fluid/operators/cross_entropy_op.cc +++ b/paddle/fluid/operators/cross_entropy_op.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/cross_entropy_op.h" +#include namespace paddle { namespace operators { @@ -179,6 +180,15 @@ or not. But the output only shares the LoD information with input X. )DOC"); } }; + +class CrossEntropyOpInferVarType + : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Y"}}; + } +}; } // namespace operators } // namespace paddle @@ -186,6 +196,7 @@ namespace ops = paddle::operators; using CPUCtx = paddle::platform::CPUDeviceContext; REGISTER_OPERATOR(cross_entropy, ops::CrossEntropyOp, ops::CrossEntropyOpMaker, + ops::CrossEntropyOpInferVarType, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(cross_entropy_grad, ops::CrossEntropyGradientOp); REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, diff --git a/paddle/fluid/operators/cross_entropy_op.cu b/paddle/fluid/operators/cross_entropy_op.cu index 30dbd5bd3d39dd2992c3dd91364003bb7715a2eb..fcd34383a85f6984a8f27ce0625364f8fd5e31d6 100644 --- a/paddle/fluid/operators/cross_entropy_op.cu +++ b/paddle/fluid/operators/cross_entropy_op.cu @@ -13,12 +13,17 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/cross_entropy_op.h" +#include "paddle/fluid/platform/float16.h" +namespace plat = paddle::platform; namespace ops = paddle::operators; using CUDACtx = paddle::platform::CUDADeviceContext; REGISTER_OP_CUDA_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, - ops::CrossEntropyOpKernel); -REGISTER_OP_CUDA_KERNEL(cross_entropy_grad, - ops::CrossEntropyGradientOpKernel, - ops::CrossEntropyGradientOpKernel); + ops::CrossEntropyOpKernel, + ops::CrossEntropyOpKernel); + +REGISTER_OP_CUDA_KERNEL( + cross_entropy_grad, ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel); diff --git a/paddle/fluid/operators/decayed_adagrad_op.cc b/paddle/fluid/operators/decayed_adagrad_op.cc index c0f2b49a04d9e88502c4b63bca493cd2b7ad1c5c..d73ae9e2721b388212cb6efa354eb4b480df9cad 100644 --- a/paddle/fluid/operators/decayed_adagrad_op.cc +++ b/paddle/fluid/operators/decayed_adagrad_op.cc @@ -32,6 +32,16 @@ class DecayedAdagradOp : public framework::OperatorWithKernel { PADDLE_ENFORCE( ctx->HasInput("LearningRate"), "Input(LearningRate) of DecayedAdagradOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of DecayedAdagradOp should not be null."); diff --git a/paddle/fluid/operators/decayed_adagrad_op.h b/paddle/fluid/operators/decayed_adagrad_op.h index a46af078e0c6b4bf1faca0570b6a97b026864f13..5df43d33ef9f720fd20d57c53ff37cc85440b24e 100644 --- a/paddle/fluid/operators/decayed_adagrad_op.h +++ b/paddle/fluid/operators/decayed_adagrad_op.h @@ -23,6 +23,17 @@ template class DecayedAdagradOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto param_out_tensor = ctx.Output("ParamOut"); auto moment_out_tensor = ctx.Output("MomentOut"); diff --git a/paddle/fluid/operators/delete_var_op.cc b/paddle/fluid/operators/delete_var_op.cc index d7a9bfbc437dbf4c723b9c87ff62ec6b62c38638..89416f7ab5d07ddac5b540b9bb361f831c1ef360 100644 --- a/paddle/fluid/operators/delete_var_op.cc +++ b/paddle/fluid/operators/delete_var_op.cc @@ -32,6 +32,11 @@ class DeleteVarOp : public framework::OperatorBase { } }; +class DeleteVarOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override {} +}; + class DeleteVarOpInfoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -48,4 +53,5 @@ It should not be configured by users directly. REGISTER_OPERATOR(delete_var, paddle::operators::DeleteVarOp, paddle::framework::EmptyGradOpMaker, - paddle::operators::DeleteVarOpInfoMaker); + paddle::operators::DeleteVarOpInfoMaker, + paddle::operators::DeleteVarOpShapeInference); diff --git a/paddle/fluid/operators/detection/CMakeLists.txt b/paddle/fluid/operators/detection/CMakeLists.txt index aa8ed502fc94bd0970dfe5dbf00ef090e799ad30..e5c3f0eeb385e1a15fdbb12a989603996420efe3 100644 --- a/paddle/fluid/operators/detection/CMakeLists.txt +++ b/paddle/fluid/operators/detection/CMakeLists.txt @@ -20,8 +20,9 @@ detection_library(box_coder_op SRCS box_coder_op.cc box_coder_op.cu) detection_library(iou_similarity_op SRCS iou_similarity_op.cc iou_similarity_op.cu) detection_library(mine_hard_examples_op SRCS mine_hard_examples_op.cc) -detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc) +detection_library(multiclass_nms_op SRCS multiclass_nms_op.cc poly_util.cc gpc.cc) detection_library(prior_box_op SRCS prior_box_op.cc prior_box_op.cu) +detection_library(density_prior_box_op SRCS density_prior_box_op.cc) detection_library(anchor_generator_op SRCS anchor_generator_op.cc anchor_generator_op.cu) detection_library(target_assign_op SRCS target_assign_op.cc diff --git a/paddle/fluid/operators/detection/density_prior_box_op.cc b/paddle/fluid/operators/detection/density_prior_box_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..99df15c3226b4305a28a3912398d6d1c766daa73 --- /dev/null +++ b/paddle/fluid/operators/detection/density_prior_box_op.cc @@ -0,0 +1,175 @@ +/*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/detection/density_prior_box_op.h" + +namespace paddle { +namespace operators { + +class DensityPriorBoxOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input(Input) of DensityPriorBoxOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Image"), + "Input(Image) of DensityPriorBoxOp should not be null."); + + auto image_dims = ctx->GetInputDim("Image"); + auto input_dims = ctx->GetInputDim("Input"); + PADDLE_ENFORCE(image_dims.size() == 4, "The layout of image is NCHW."); + PADDLE_ENFORCE(input_dims.size() == 4, "The layout of input is NCHW."); + + PADDLE_ENFORCE_LT(input_dims[2], image_dims[2], + "The height of input must smaller than image."); + + PADDLE_ENFORCE_LT(input_dims[3], image_dims[3], + "The width of input must smaller than image."); + auto variances = ctx->Attrs().Get>("variances"); + + auto fixed_sizes = ctx->Attrs().Get>("fixed_sizes"); + auto fixed_ratios = ctx->Attrs().Get>("fixed_ratios"); + auto densities = ctx->Attrs().Get>("densities"); + + PADDLE_ENFORCE_EQ(fixed_sizes.size(), densities.size(), + "The number of fixed_sizes and densities must be equal."); + size_t num_priors = 0; + if ((fixed_sizes.size() > 0) && (densities.size() > 0)) { + for (size_t i = 0; i < densities.size(); ++i) { + if (fixed_ratios.size() > 0) { + num_priors += (fixed_ratios.size()) * (pow(densities[i], 2)); + } + } + } + std::vector dim_vec(4); + dim_vec[0] = input_dims[2]; + dim_vec[1] = input_dims[3]; + dim_vec[2] = num_priors; + dim_vec[3] = 4; + ctx->SetOutputDim("Boxes", framework::make_ddim(dim_vec)); + ctx->SetOutputDim("Variances", framework::make_ddim(dim_vec)); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("Input")->type()), + platform::CPUPlace()); + } +}; + +class DensityPriorBoxOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput( + "Input", + "(Tensor, default Tensor), " + "the input feature data of DensityPriorBoxOp, the layout is NCHW."); + AddInput("Image", + "(Tensor, default Tensor), " + "the input image data of DensityPriorBoxOp, the layout is NCHW."); + AddOutput("Boxes", + "(Tensor, default Tensor), the output prior boxes of " + "DensityPriorBoxOp. The layout is [H, W, num_priors, 4]. " + "H is the height of input, W is the width of input, num_priors " + "is the box count of each position."); + AddOutput("Variances", + "(Tensor, default Tensor), the expanded variances of " + "DensityPriorBoxOp. The layout is [H, W, num_priors, 4]. " + "H is the height of input, W is the width of input, num_priors " + "is the box count of each position."); + AddAttr>("variances", + "(vector) List of variances to be " + "encoded in density prior boxes.") + .AddCustomChecker([](const std::vector& variances) { + PADDLE_ENFORCE_EQ(variances.size(), 4, + "Must and only provide 4 variance."); + for (size_t i = 0; i < variances.size(); ++i) { + PADDLE_ENFORCE_GT(variances[i], 0.0, + "variance[%d] must be greater than 0.", i); + } + }); + AddAttr("clip", "(bool) Whether to clip out-of-boundary boxes.") + .SetDefault(true); + + AddAttr( + "step_w", + "Density prior boxes step across width, 0.0 for auto calculation.") + .SetDefault(0.0) + .AddCustomChecker([](const float& step_w) { + PADDLE_ENFORCE_GE(step_w, 0.0, "step_w should be larger than 0."); + }); + AddAttr( + "step_h", + "Density prior boxes step across height, 0.0 for auto calculation.") + .SetDefault(0.0) + .AddCustomChecker([](const float& step_h) { + PADDLE_ENFORCE_GE(step_h, 0.0, "step_h should be larger than 0."); + }); + + AddAttr("offset", + "(float) " + "Density prior boxes center offset.") + .SetDefault(0.5); + AddAttr>("fixed_sizes", + "(vector) List of fixed sizes " + "of generated density prior boxes.") + .SetDefault(std::vector{}) + .AddCustomChecker([](const std::vector& fixed_sizes) { + for (size_t i = 0; i < fixed_sizes.size(); ++i) { + PADDLE_ENFORCE_GT(fixed_sizes[i], 0.0, + "fixed_sizes[%d] should be larger than 0.", i); + } + }); + + AddAttr>("fixed_ratios", + "(vector) List of fixed ratios " + "of generated density prior boxes.") + .SetDefault(std::vector{}) + .AddCustomChecker([](const std::vector& fixed_ratios) { + for (size_t i = 0; i < fixed_ratios.size(); ++i) { + PADDLE_ENFORCE_GT(fixed_ratios[i], 0.0, + "fixed_ratios[%d] should be larger than 0.", i); + } + }); + + AddAttr>("densities", + "(vector) List of densities " + "of generated density prior boxes.") + .SetDefault(std::vector{}) + .AddCustomChecker([](const std::vector& densities) { + for (size_t i = 0; i < densities.size(); ++i) { + PADDLE_ENFORCE_GT(densities[i], 0, + "densities[%d] should be larger than 0.", i); + } + }); + AddComment(R"DOC( + Density Prior box operator + Each position of the input produce N density prior boxes, N is determined by + the count of fixed_ratios, densities, the calculation of N is as follows: + for density in densities: + N += size(fixed_ratios)*density^2 + )DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(density_prior_box, ops::DensityPriorBoxOp, + ops::DensityPriorBoxOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL(density_prior_box, ops::DensityPriorBoxOpKernel, + ops::DensityPriorBoxOpKernel); diff --git a/paddle/fluid/operators/detection/density_prior_box_op.h b/paddle/fluid/operators/detection/density_prior_box_op.h new file mode 100644 index 0000000000000000000000000000000000000000..9a52077e9cf90b278549a077af161bd4e282d972 --- /dev/null +++ b/paddle/fluid/operators/detection/density_prior_box_op.h @@ -0,0 +1,146 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include +#include +#include "paddle/fluid/operators/detection/prior_box_op.h" + +namespace paddle { +namespace operators { + +template +class DensityPriorBoxOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("Input"); + auto* image = ctx.Input("Image"); + auto* boxes = ctx.Output("Boxes"); + auto* vars = ctx.Output("Variances"); + + auto variances = ctx.Attr>("variances"); + auto clip = ctx.Attr("clip"); + + auto fixed_sizes = ctx.Attr>("fixed_sizes"); + auto fixed_ratios = ctx.Attr>("fixed_ratios"); + auto densities = ctx.Attr>("densities"); + + T step_w = static_cast(ctx.Attr("step_w")); + T step_h = static_cast(ctx.Attr("step_h")); + T offset = static_cast(ctx.Attr("offset")); + + auto img_width = image->dims()[3]; + auto img_height = image->dims()[2]; + + auto feature_width = input->dims()[3]; + auto feature_height = input->dims()[2]; + + T step_width, step_height; + if (step_w == 0 || step_h == 0) { + step_width = static_cast(img_width) / feature_width; + step_height = static_cast(img_height) / feature_height; + } else { + step_width = step_w; + step_height = step_h; + } + int num_priors = 0; + if (fixed_sizes.size() > 0 && densities.size() > 0) { + for (size_t i = 0; i < densities.size(); ++i) { + if (fixed_ratios.size() > 0) { + num_priors += (fixed_ratios.size()) * (pow(densities[i], 2)); + } + } + } + + boxes->mutable_data(ctx.GetPlace()); + vars->mutable_data(ctx.GetPlace()); + auto e_boxes = framework::EigenTensor::From(*boxes).setConstant(0.0); + + int step_average = static_cast((step_width + step_height) * 0.5); + + for (int h = 0; h < feature_height; ++h) { + for (int w = 0; w < feature_width; ++w) { + T center_x = (w + offset) * step_width; + T center_y = (h + offset) * step_height; + int idx = 0; + // Generate density prior boxes with fixed sizes. + for (size_t s = 0; s < fixed_sizes.size(); ++s) { + auto fixed_size = fixed_sizes[s]; + int density = densities[s]; + // Generate density prior boxes with fixed ratios. + if (fixed_ratios.size() > 0) { + for (size_t r = 0; r < fixed_ratios.size(); ++r) { + float ar = fixed_ratios[r]; + int shift = step_average / density; + float box_width_ratio = fixed_size * sqrt(ar); + float box_height_ratio = fixed_size / sqrt(ar); + for (int di = 0; di < density; ++di) { + for (int dj = 0; dj < density; ++dj) { + float center_x_temp = + center_x - step_average / 2. + shift / 2. + dj * shift; + float center_y_temp = + center_y - step_average / 2. + shift / 2. + di * shift; + e_boxes(h, w, idx, 0) = + (center_x_temp - box_width_ratio / 2.) / img_width >= 0 + ? (center_x_temp - box_width_ratio / 2.) / img_width + : 0; + e_boxes(h, w, idx, 1) = + (center_y_temp - box_height_ratio / 2.) / img_height >= 0 + ? (center_y_temp - box_height_ratio / 2.) / img_height + : 0; + e_boxes(h, w, idx, 2) = + (center_x_temp + box_width_ratio / 2.) / img_width <= 1 + ? (center_x_temp + box_width_ratio / 2.) / img_width + : 1; + e_boxes(h, w, idx, 3) = + (center_y_temp + box_height_ratio / 2.) / img_height <= 1 + ? (center_y_temp + box_height_ratio / 2.) / img_height + : 1; + idx++; + } + } + } + } + } + } + } + if (clip) { + platform::Transform trans; + ClipFunctor clip_func; + trans(ctx.template device_context(), + boxes->data(), boxes->data() + boxes->numel(), + boxes->data(), clip_func); + } + framework::Tensor var_t; + var_t.mutable_data( + framework::make_ddim({1, static_cast(variances.size())}), + ctx.GetPlace()); + + auto var_et = framework::EigenTensor::From(var_t); + + for (size_t i = 0; i < variances.size(); ++i) { + var_et(0, i) = variances[i]; + } + + int box_num = feature_height * feature_width * num_priors; + auto var_dim = vars->dims(); + vars->Resize({box_num, static_cast(variances.size())}); + + auto e_vars = framework::EigenMatrix::From(*vars); + + e_vars = var_et.broadcast(Eigen::DSizes(box_num, 1)); + + vars->Resize(var_dim); + } +}; // namespace operators + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detection/generate_proposal_labels_op.cc b/paddle/fluid/operators/detection/generate_proposal_labels_op.cc index d7a53f1bef98ecda3ba7b36323678a11a632a15c..fddd6884017c35112ba48f245759f5d846b55f9a 100644 --- a/paddle/fluid/operators/detection/generate_proposal_labels_op.cc +++ b/paddle/fluid/operators/detection/generate_proposal_labels_op.cc @@ -16,7 +16,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detection/bbox_util.h" #include "paddle/fluid/operators/gather.h" -#include "paddle/fluid/operators/math/concat.h" +#include "paddle/fluid/operators/math/concat_and_split.h" #include "paddle/fluid/operators/math/math_function.h" namespace paddle { @@ -439,31 +439,88 @@ class GenerateProposalLabelsKernel : public framework::OpKernel { class GenerateProposalLabelsOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - // TODO(buxingyuan): Add Document - AddInput("RpnRois", "RpnRois."); - AddInput("GtClasses", "GtClasses."); - AddInput("IsCrowd", "IsCrowd."); - AddInput("GtBoxes", "GtBoxes."); - AddInput("ImInfo", "ImInfo."); - - AddOutput("Rois", "Rois."); - AddOutput("LabelsInt32", "LabelsInt32."); - AddOutput("BboxTargets", "BboxTargets."); - AddOutput("BboxInsideWeights", "BboxInsideWeights."); - AddOutput("BboxOutsideWeights", "BboxOutsideWeights."); - - AddAttr("batch_size_per_im", "batch_size_per_im"); - AddAttr("fg_fraction", "fg_fraction"); - AddAttr("fg_thresh", "fg_thresh"); - AddAttr("bg_thresh_hi", "bg_thresh_hi"); - AddAttr("bg_thresh_lo", "bg_thresh_lo"); - AddAttr>("bbox_reg_weights", "bbox_reg_weights"); - AddAttr("class_nums", "class_nums"); - AddAttr("use_random", "use_random").SetDefault(true); + AddInput( + "RpnRois", + "(LoDTensor), This input is a 2D LoDTensor with shape [N, 4]. " + "N is the number of the GenerateProposalOp's output, " + "each element is a bounding box with [xmin, ymin, xmax, ymax] format."); + AddInput("GtClasses", + "(LoDTensor), This input is a 2D LoDTensor with shape [M, 1]. " + "M is the number of groundtruth, " + "each element is a class label of groundtruth."); + AddInput( + "IsCrowd", + "(LoDTensor), This input is a 2D LoDTensor with shape [M, 1]. " + "M is the number of groundtruth, " + "each element is a flag indicates whether a groundtruth is crowd."); + AddInput( + "GtBoxes", + "(LoDTensor), This input is a 2D LoDTensor with shape [M, 4]. " + "M is the number of groundtruth, " + "each element is a bounding box with [xmin, ymin, xmax, ymax] format."); + AddInput("ImInfo", + "(Tensor), This input is a 2D Tensor with shape [B, 3]. " + "B is the number of input images, " + "each element consists of im_height, im_width, im_scale."); + + AddOutput( + "Rois", + "(LoDTensor), This output is a 2D LoDTensor with shape [P, 4]. " + "P usuall equal to batch_size_per_im * batch_size, " + "each element is a bounding box with [xmin, ymin, xmax, ymax] format."); + AddOutput("LabelsInt32", + "(LoDTensor), This output is a 2D LoDTensor with shape [P], " + "each element repersents a class label of a roi"); + AddOutput("BboxTargets", + "(LoDTensor), This output is a 2D LoDTensor with shape [P, 4 * " + "class_nums], " + "each element repersents a box label of a roi"); + AddOutput( + "BboxInsideWeights", + "(LoDTensor), This output is a 2D LoDTensor with shape [P, 4 * " + "class_nums], " + "each element indicates whether a box should contribute to loss."); + AddOutput( + "BboxOutsideWeights", + "(LoDTensor), This output is a 2D LoDTensor with shape [P, 4 * " + "class_nums], " + "each element indicates whether a box should contribute to loss."); + + AddAttr("batch_size_per_im", "Batch size of rois per images."); + AddAttr("fg_fraction", + "Foreground fraction in total batch_size_per_im."); + AddAttr( + "fg_thresh", + "Overlap threshold which is used to chose foreground sample."); + AddAttr("bg_thresh_hi", + "Overlap threshold upper bound which is used to chose " + "background sample."); + AddAttr("bg_thresh_lo", + "Overlap threshold lower bound which is used to chose " + "background sample."); + AddAttr>("bbox_reg_weights", "Box regression weights."); + AddAttr("class_nums", "Class number."); + AddAttr( + "use_random", + "Use random sampling to choose foreground and background boxes.") + .SetDefault(true); AddComment(R"DOC( -Generate Proposals Labels Operator. -)DOC"); +This operator can be, for given the GenerateProposalOp output bounding boxes and groundtruth, +to sample foreground boxes and background boxes, and compute loss target. + +RpnRois is the output boxes of RPN and was processed by generate_proposal_op, these boxes +were combined with groundtruth boxes and sampled according to batch_size_per_im and fg_fraction, +If an instance with a groundtruth overlap greater than fg_thresh, then it was considered as a foreground sample. +If an instance with a groundtruth overlap greater than bg_thresh_lo and lower than bg_thresh_hi, +then it was considered as a background sample. +After all foreground and background boxes are chosen (so called Rois), +then we apply random sampling to make sure +the number of foreground boxes is no more than batch_size_per_im * fg_fraction. + +For each box in Rois, we assign the classification (class label) and regression targets (box label) to it. +Finally BboxInsideWeights and BboxOutsideWeights are used to specify whether it would contribute to training loss. + )DOC"); } }; diff --git a/paddle/fluid/operators/detection/generate_proposals_op.cc b/paddle/fluid/operators/detection/generate_proposals_op.cc index 818d58ea9ee327fd99182ad2f8cbeed07e6aaea2..709c2dfc4b7c67d7d04074c58ce6da85b6e790fe 100644 --- a/paddle/fluid/operators/detection/generate_proposals_op.cc +++ b/paddle/fluid/operators/detection/generate_proposals_op.cc @@ -12,10 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include +#include #include #include #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/var_type.h" +#include "paddle/fluid/operators/detail/safe_ref.h" #include "paddle/fluid/operators/gather.h" #include "paddle/fluid/operators/math/math_function.h" @@ -25,21 +27,17 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -struct AppendProposalsFunctor { - LoDTensor *out_; - int64_t offset_; - Tensor *to_add_; +static const double kBBoxClipDefault = std::log(1000.0 / 16.0); - AppendProposalsFunctor(LoDTensor *out, int64_t offset, Tensor *to_add) - : out_(out), offset_(offset), to_add_(to_add) {} - - template - void apply() const { - auto *out_data = out_->data(); - auto *to_add_data = to_add_->data(); - memcpy(out_data + offset_, to_add_data, to_add_->numel() * sizeof(T)); - } -}; +static void AppendProposals(Tensor *dst, int64_t offset, const Tensor &src) { + auto *out_data = dst->data(); + auto *to_add_data = src.data(); + size_t size_of_t = framework::SizeOfType(src.type()); + offset *= size_of_t; + std::memcpy( + reinterpret_cast(reinterpret_cast(out_data) + offset), + to_add_data, src.numel() * size_of_t); +} class GenerateProposalsOp : public framework::OperatorWithKernel { public: @@ -75,8 +73,9 @@ class GenerateProposalsOp : public framework::OperatorWithKernel { }; template -void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors, - Tensor *bbox_deltas, Tensor *variances, Tensor *proposals) { +static inline void BoxCoder(const platform::DeviceContext &ctx, + Tensor *all_anchors, Tensor *bbox_deltas, + Tensor *variances, Tensor *proposals) { T *proposals_data = proposals->mutable_data(ctx.GetPlace()); int64_t row = all_anchors->dims()[0]; @@ -108,11 +107,11 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors, anchor_center_y; bbox_width = std::exp(std::min(variances_data[i * len + 2] * bbox_deltas_data[i * len + 2], - std::log(1000.0 / 16.0))) * + kBBoxClipDefault)) * anchor_width; bbox_height = std::exp(std::min(variances_data[i * len + 3] * bbox_deltas_data[i * len + 3], - std::log(1000.0 / 16.0))) * + kBBoxClipDefault)) * anchor_height; } else { bbox_center_x = @@ -120,10 +119,10 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors, bbox_center_y = bbox_deltas_data[i * len + 1] * anchor_height + anchor_center_y; bbox_width = std::exp(std::min(bbox_deltas_data[i * len + 2], - std::log(1000.0 / 16.0))) * + kBBoxClipDefault)) * anchor_width; bbox_height = std::exp(std::min(bbox_deltas_data[i * len + 3], - std::log(1000.0 / 16.0))) * + kBBoxClipDefault)) * anchor_height; } @@ -136,30 +135,32 @@ void BoxCoder(const platform::DeviceContext &ctx, Tensor *all_anchors, } template -void ClipTiledBoxes(const platform::DeviceContext &ctx, const Tensor &im_info, - Tensor *boxes) { +static inline void ClipTiledBoxes(const platform::DeviceContext &ctx, + const Tensor &im_info, Tensor *boxes) { T *boxes_data = boxes->mutable_data(ctx.GetPlace()); const T *im_info_data = im_info.data(); + T zero(0); for (int64_t i = 0; i < boxes->numel(); ++i) { if (i % 4 == 0) { boxes_data[i] = - std::max(std::min(boxes_data[i], im_info_data[1] - 1), 0.0f); + std::max(std::min(boxes_data[i], im_info_data[1] - 1), zero); } else if (i % 4 == 1) { boxes_data[i] = - std::max(std::min(boxes_data[i], im_info_data[0] - 1), 0.0f); + std::max(std::min(boxes_data[i], im_info_data[0] - 1), zero); } else if (i % 4 == 2) { boxes_data[i] = - std::max(std::min(boxes_data[i], im_info_data[1] - 1), 0.0f); + std::max(std::min(boxes_data[i], im_info_data[1] - 1), zero); } else { boxes_data[i] = - std::max(std::min(boxes_data[i], im_info_data[0] - 1), 0.0f); + std::max(std::min(boxes_data[i], im_info_data[0] - 1), zero); } } } template -void FilterBoxes(const platform::DeviceContext &ctx, Tensor *boxes, - float min_size, const Tensor &im_info, Tensor *keep) { +static inline void FilterBoxes(const platform::DeviceContext &ctx, + Tensor *boxes, float min_size, + const Tensor &im_info, Tensor *keep) { const T *im_info_data = im_info.data(); T *boxes_data = boxes->mutable_data(ctx.GetPlace()); T im_scale = im_info_data[2]; @@ -185,24 +186,24 @@ void FilterBoxes(const platform::DeviceContext &ctx, Tensor *boxes, keep->Resize({keep_len}); } -bool SortScorePairDescend(const std::pair &pair1, - const std::pair &pair2) { - return pair1.first > pair2.first; -} - template -void GetMaxScoreIndex(const std::vector &scores, - std::vector> *sorted_indices) { +static inline std::vector> GetSortedScoreIndex( + const std::vector &scores) { + std::vector> sorted_indices; + sorted_indices.reserve(scores.size()); for (size_t i = 0; i < scores.size(); ++i) { - sorted_indices->push_back(std::make_pair(scores[i], i)); + sorted_indices.emplace_back(scores[i], i); } // Sort the score pair according to the scores in descending order - std::stable_sort(sorted_indices->begin(), sorted_indices->end(), - SortScorePairDescend); + std::stable_sort(sorted_indices.begin(), sorted_indices.end(), + [](const std::pair &a, const std::pair &b) { + return a.first < b.first; + }); + return sorted_indices; } template -T BBoxArea(const T *box, const bool normalized) { +static inline T BBoxArea(const T *box, bool normalized) { if (box[2] < box[0] || box[3] < box[1]) { // If coordinate values are is invalid // (e.g. xmax < xmin or ymax < ymin), return 0. @@ -220,7 +221,7 @@ T BBoxArea(const T *box, const bool normalized) { } template -T JaccardOverlap(const T *box1, const T *box2, const bool normalized) { +static inline T JaccardOverlap(const T *box1, const T *box2, bool normalized) { if (box2[0] > box1[2] || box2[2] < box1[0] || box2[1] > box1[3] || box2[3] < box1[1]) { return static_cast(0.); @@ -229,8 +230,8 @@ T JaccardOverlap(const T *box1, const T *box2, const bool normalized) { const T inter_ymin = std::max(box1[1], box2[1]); const T inter_xmax = std::min(box1[2], box2[2]); const T inter_ymax = std::min(box1[3], box2[3]); - const T inter_w = std::max(0.0f, inter_xmax - inter_xmin + 1); - const T inter_h = std::max(0.0f, inter_ymax - inter_ymin + 1); + const T inter_w = std::max(T(0), inter_xmax - inter_xmin + 1); + const T inter_h = std::max(T(0), inter_ymax - inter_ymin + 1); const T inter_area = inter_w * inter_h; const T bbox1_area = BBoxArea(box1, normalized); const T bbox2_area = BBoxArea(box2, normalized); @@ -238,9 +239,21 @@ T JaccardOverlap(const T *box1, const T *box2, const bool normalized) { } } +template +static inline Tensor VectorToTensor(const std::vector &selected_indices, + int selected_num) { + Tensor keep_nms; + keep_nms.Resize({selected_num}); + auto *keep_data = keep_nms.mutable_data(platform::CPUPlace()); + for (int i = 0; i < selected_num; ++i) { + keep_data[i] = selected_indices[i]; + } + return keep_nms; +} + template -Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores, - const T nms_threshold, const float eta) { +static inline Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, + Tensor *scores, T nms_threshold, float eta) { PADDLE_ENFORCE_NOT_NULL(bbox); int64_t num_boxes = bbox->dims()[0]; // 4: [xmin ymin xmax ymax] @@ -248,20 +261,18 @@ Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores, std::vector scores_data(num_boxes); std::copy_n(scores->data(), num_boxes, scores_data.begin()); - std::vector> sorted_indices; - GetMaxScoreIndex(scores_data, &sorted_indices); + std::vector> sorted_indices = + GetSortedScoreIndex(scores_data); std::vector selected_indices; int selected_num = 0; T adaptive_threshold = nms_threshold; const T *bbox_data = bbox->data(); - bool flag; while (sorted_indices.size() != 0) { - int idx = sorted_indices.front().second; - flag = true; - for (size_t k = 0; k < selected_indices.size(); ++k) { + int idx = sorted_indices.back().second; + bool flag = true; + for (int kept_idx : selected_indices) { if (flag) { - const int kept_idx = selected_indices[k]; T overlap = JaccardOverlap(bbox_data + idx * box_size, bbox_data + kept_idx * box_size, false); flag = (overlap <= adaptive_threshold); @@ -271,32 +282,29 @@ Tensor NMS(const platform::DeviceContext &ctx, Tensor *bbox, Tensor *scores, } if (flag) { selected_indices.push_back(idx); - selected_num++; + ++selected_num; } - sorted_indices.erase(sorted_indices.begin()); + sorted_indices.erase(sorted_indices.end() - 1); if (flag && eta < 1 && adaptive_threshold > 0.5) { adaptive_threshold *= eta; } } - Tensor keep_nms; - keep_nms.Resize({selected_num}); - int *keep_data = keep_nms.mutable_data(ctx.GetPlace()); - for (int i = 0; i < selected_num; ++i) { - keep_data[i] = selected_indices[i]; - } - - return keep_nms; + return VectorToTensor(selected_indices, selected_num); } -template +template class GenerateProposalsKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext &context) const override { auto *scores = context.Input("Scores"); auto *bbox_deltas = context.Input("BboxDeltas"); auto *im_info = context.Input("ImInfo"); - auto *anchors = context.Input("Anchors"); - auto *variances = context.Input("Variances"); + auto anchors = detail::Ref(context.Input("Anchors"), + "Cannot find input Anchors(%s) in scope", + context.Inputs("Anchors")[0]); + auto variances = detail::Ref(context.Input("Variances"), + "Cannot find input Variances(%s) in scope", + context.Inputs("Variances")[0]); auto *rpn_rois = context.Output("RpnRois"); auto *rpn_roi_probs = context.Output("RpnRoiProbs"); @@ -307,15 +315,16 @@ class GenerateProposalsKernel : public framework::OpKernel { float min_size = context.Attr("min_size"); float eta = context.Attr("eta"); - auto &dev_ctx = context.template device_context(); + auto &dev_ctx = + context.template device_context(); - auto scores_dim = scores->dims(); + auto &scores_dim = scores->dims(); int64_t num = scores_dim[0]; int64_t c_score = scores_dim[1]; int64_t h_score = scores_dim[2]; int64_t w_score = scores_dim[3]; - auto bbox_dim = bbox_deltas->dims(); + auto &bbox_dim = bbox_deltas->dims(); int64_t c_bbox = bbox_dim[1]; int64_t h_bbox = bbox_dim[2]; int64_t w_bbox = bbox_dim[3]; @@ -330,17 +339,17 @@ class GenerateProposalsKernel : public framework::OpKernel { scores_swap.mutable_data({num, h_score, w_score, c_score}, dev_ctx.GetPlace()); - math::Transpose trans; + math::Transpose trans; std::vector axis = {0, 2, 3, 1}; trans(dev_ctx, *bbox_deltas, &bbox_deltas_swap, axis); trans(dev_ctx, *scores, &scores_swap, axis); framework::LoD lod; - std::vector lod0(1, 0); - Tensor *anchor = const_cast(anchors); - anchor->Resize({anchors->numel() / 4, 4}); - Tensor *var = const_cast(variances); - var->Resize({var->numel() / 4, 4}); + lod.resize(1); + auto &lod0 = lod[0]; + lod0.push_back(0); + anchors.Resize({anchors.numel() / 4, 4}); + variances.Resize({variances.numel() / 4, 4}); int64_t num_proposals = 0; for (int64_t i = 0; i < num; ++i) { @@ -352,24 +361,17 @@ class GenerateProposalsKernel : public framework::OpKernel { scores_slice.Resize({h_score * w_score * c_score, 1}); std::pair tensor_pair = - ProposalForOneImage(dev_ctx, im_info_slice, *anchor, *var, + ProposalForOneImage(dev_ctx, im_info_slice, anchors, variances, bbox_deltas_slice, scores_slice, pre_nms_top_n, post_nms_top_n, nms_thresh, min_size, eta); - Tensor proposals = tensor_pair.first; - Tensor scores = tensor_pair.second; - - framework::VisitDataType( - framework::ToDataType(rpn_rois->type()), - AppendProposalsFunctor(rpn_rois, 4 * num_proposals, &proposals)); - framework::VisitDataType( - framework::ToDataType(rpn_roi_probs->type()), - AppendProposalsFunctor(rpn_roi_probs, num_proposals, &scores)); + Tensor &proposals = tensor_pair.first; + Tensor &scores = tensor_pair.second; + AppendProposals(rpn_rois, 4 * num_proposals, proposals); + AppendProposals(rpn_roi_probs, num_proposals, scores); num_proposals += proposals.dims()[0]; - lod0.emplace_back(num_proposals); + lod0.push_back(num_proposals); } - - lod.emplace_back(lod0); rpn_rois->set_lod(lod); rpn_roi_probs->set_lod(lod); rpn_rois->Resize({num_proposals, 4}); @@ -377,7 +379,7 @@ class GenerateProposalsKernel : public framework::OpKernel { } std::pair ProposalForOneImage( - const DeviceContext &ctx, const Tensor &im_info_slice, + const platform::CPUDeviceContext &ctx, const Tensor &im_info_slice, const Tensor &anchors, const Tensor &variances, const Tensor &bbox_deltas_slice, // [M, 4] const Tensor &scores_slice, // [N, 1] @@ -392,10 +394,9 @@ class GenerateProposalsKernel : public framework::OpKernel { for (int i = 0; i < scores_slice.numel(); ++i) { index[i] = i; } - std::function compare = - [scores_data](const int64_t &i, const int64_t &j) { - return scores_data[i] > scores_data[j]; - }; + auto compare = [scores_data](const int64_t &i, const int64_t &j) { + return scores_data[i] > scores_data[j]; + }; if (pre_nms_top_n <= 0 || pre_nms_top_n >= scores_slice.numel()) { std::sort(index, index + scores_slice.numel(), compare); @@ -452,33 +453,45 @@ class GenerateProposalsKernel : public framework::OpKernel { class GenerateProposalsOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Scores", "The scores of anchors should be foreground."); - AddInput("BboxDeltas", "bbox_deltas."); - AddInput("ImInfo", "Information for image reshape."); - AddInput("Anchors", "All anchors."); - AddInput("Variances", " variances"); - - AddOutput("RpnRois", "Anchors."); - AddOutput("RpnRoiProbs", "Anchors."); - AddAttr("pre_nms_topN", "pre_nms_topN"); - AddAttr("post_nms_topN", "post_nms_topN"); - AddAttr("nms_thresh", "nms_thres"); - AddAttr("min_size", "min size"); + AddInput("Scores", + "(Tensor) The scores from conv is in shape (N, A, H, W), " + "N is batch size, A is number of anchors, " + "H and W are height and width of the feature map"); + AddInput("BboxDeltas", + "(Tensor) Bounding box deltas from conv is in " + "shape (N, 4*A, H, W)."); + AddInput("ImInfo", + "(Tensor) Information for image reshape is in shape (N, 3), " + "in format (height, width, scale)"); + AddInput("Anchors", + "(Tensor) Bounding box anchors from anchor_generator_op " + "is in shape (A, H, W, 4)."); + AddInput("Variances", + "(Tensor) Bounding box variances with same shape as `Anchors`."); + + AddOutput("RpnRois", + "(LoDTensor), Output proposals with shape (rois_num, 4)."); + AddOutput("RpnRoiProbs", + "(LoDTensor) Scores of proposals with shape (rois_num, 1)."); + AddAttr("pre_nms_topN", + "Number of top scoring RPN proposals to keep before " + "applying NMS."); + AddAttr("post_nms_topN", + "Number of top scoring RPN proposals to keep after " + "applying NMS"); + AddAttr("nms_thresh", "NMS threshold used on RPN proposals."); + AddAttr("min_size", + "Proposal height and width both need to be greater " + "than this min_size."); AddAttr("eta", "The parameter for adaptive NMS."); AddComment(R"DOC( -Generate Proposals OP - -This operator proposes rois according to each box with their probability to be a foreground object and -the box can be calculated by anchors. Bbox_deltais and scores are the output of RPN. Final proposals -could be used to train detection net. - -Scores is the probability for each box to be an object. In format of (N, A, H, W) where N is batch size, A is number -of anchors, H and W are height and width of the feature map. -BboxDeltas is the differece between predicted box locatoin and anchor location. In format of (N, 4*A, H, W) +This operator Generate bounding box proposals for Faster RCNN. +The propoasls are generated for a list of images based on image +score 'Scores', bounding box regression result 'BboxDeltas' as +well as predefined bounding box shapes 'anchors'. Greedy +non-maximum suppression is applied to generate the final bounding +boxes. -For generating proposals, this operator transposes and resizes scores and bbox_deltas in size of (H*W*A, 1) and (H*W*A, 4) and - calculate box locations as proposals candidates. Then clip boxes to image and remove predicted boxes with small area. -Finally, apply nms to get final proposals as output. )DOC"); } }; @@ -490,6 +503,5 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(generate_proposals, ops::GenerateProposalsOp, ops::GenerateProposalsOpMaker, paddle::framework::EmptyGradOpMaker); -REGISTER_OP_CPU_KERNEL( - generate_proposals, - ops::GenerateProposalsKernel); +REGISTER_OP_CPU_KERNEL(generate_proposals, ops::GenerateProposalsKernel, + ops::GenerateProposalsKernel); diff --git a/paddle/fluid/operators/detection/generate_proposals_op.cu b/paddle/fluid/operators/detection/generate_proposals_op.cu index 6146ff509d768c0317a5c65ed22af1a3075977a2..91213b3c4d9db54469ec151ff1dd8e56c3118fea 100644 --- a/paddle/fluid/operators/detection/generate_proposals_op.cu +++ b/paddle/fluid/operators/detection/generate_proposals_op.cu @@ -16,10 +16,13 @@ limitations under the License. */ #include #include #include "cub/cub.cuh" +#include "paddle/fluid/framework/mixed_vector.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/memory/memory.h" +#include "paddle/fluid/operators/detail/safe_ref.h" #include "paddle/fluid/operators/gather.cu.h" #include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { @@ -36,36 +39,38 @@ namespace { int const kThreadsPerBlock = sizeof(uint64_t) * 8; -template -__global__ void RangeInitKernel(const T start, const T delta, const int size, - T *out) { - CUDA_1D_KERNEL_LOOP(i, size) { out[i] = start + i * delta; } -} +static const double kBBoxClipDefault = std::log(1000.0 / 16.0); + +struct RangeInitFunctor { + int start_; + int delta_; + int *out_; + __device__ void operator()(size_t i) { out_[i] = start_ + i * delta_; } +}; template -void SortDescending(const platform::CUDADeviceContext &ctx, const Tensor &value, - Tensor *value_out, Tensor *index_out) { - int num = value.numel(); +static void SortDescending(const platform::CUDADeviceContext &ctx, + const Tensor &value, Tensor *value_out, + Tensor *index_out) { + int num = static_cast(value.numel()); Tensor index_in_t; int *idx_in = index_in_t.mutable_data({num}, ctx.GetPlace()); - int block = 512; - auto stream = ctx.stream(); - RangeInitKernel<<>>(0, 1, num, idx_in); + platform::ForRange for_range(ctx, num); + for_range(RangeInitFunctor{0, 1, idx_in}); + int *idx_out = index_out->mutable_data({num}, ctx.GetPlace()); const T *keys_in = value.data(); T *keys_out = value_out->mutable_data({num}, ctx.GetPlace()); // Determine temporary device storage requirements - void *d_temp_storage = NULL; size_t temp_storage_bytes = 0; cub::DeviceRadixSort::SortPairsDescending( - d_temp_storage, temp_storage_bytes, keys_in, keys_out, idx_in, idx_out, - num); + nullptr, temp_storage_bytes, keys_in, keys_out, idx_in, idx_out, num); // Allocate temporary storage auto place = boost::get(ctx.GetPlace()); - d_temp_storage = memory::Alloc(place, temp_storage_bytes); + void *d_temp_storage = memory::Alloc(place, temp_storage_bytes); // Run sorting operation cub::DeviceRadixSort::SortPairsDescending( @@ -76,22 +81,27 @@ void SortDescending(const platform::CUDADeviceContext &ctx, const Tensor &value, } template -__device__ __forceinline__ T Min(T x, T y) { - return x < y ? x : y; -} - -template -__device__ __forceinline__ T Max(T x, T y) { - return x > y ? x : y; -} - -template -__global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas, - const T *var, const int *index, - const T *im_info, const int num, - T *proposals) { - T kBBoxClipDefault = log(1000.0 / 16.0); - CUDA_1D_KERNEL_LOOP(i, num) { +struct BoxDecodeAndClipFunctor { + const T *anchor; + const T *deltas; + const T *var; + const int *index; + const T *im_info; + + T *proposals; + + BoxDecodeAndClipFunctor(const T *anchor, const T *deltas, const T *var, + const int *index, const T *im_info, T *proposals) + : anchor(anchor), + deltas(deltas), + var(var), + index(index), + im_info(im_info), + proposals(proposals) {} + + T bbox_clip_default{static_cast(kBBoxClipDefault)}; + + __device__ void operator()(size_t i) { int k = index[i] * 4; T axmin = anchor[k]; T aymin = anchor[k + 1]; @@ -108,17 +118,17 @@ __global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas, T dxmax = deltas[k + 2]; T dymax = deltas[k + 3]; - T d_cx = 0., d_cy = 0., d_w = 0., d_h = 0.; + T d_cx, d_cy, d_w, d_h; if (var) { d_cx = cx + dxmin * w * var[k]; d_cy = cy + dymin * h * var[k + 1]; - d_w = exp(Min(dxmax * var[k + 2], kBBoxClipDefault)) * w; - d_h = exp(Min(dymax * var[k + 3], kBBoxClipDefault)) * h; + d_w = exp(Min(dxmax * var[k + 2], bbox_clip_default)) * w; + d_h = exp(Min(dymax * var[k + 3], bbox_clip_default)) * h; } else { d_cx = cx + dxmin * w; d_cy = cy + dymin * h; - d_w = exp(Min(dxmax, kBBoxClipDefault)) * w; - d_h = exp(Min(dymax, kBBoxClipDefault)) * h; + d_w = exp(Min(dxmax, bbox_clip_default)) * w; + d_h = exp(Min(dymax, bbox_clip_default)) * h; } T oxmin = d_cx - d_w * 0.5; @@ -126,17 +136,21 @@ __global__ void BoxDecodeAndClipKernel(const T *anchor, const T *deltas, T oxmax = d_cx + d_w * 0.5 - 1.; T oymax = d_cy + d_h * 0.5 - 1.; - proposals[i * 4] = Max(Min(oxmin, im_info[1] - 1.), 0.); - proposals[i * 4 + 1] = Max(Min(oymin, im_info[0] - 1.), 0.); - proposals[i * 4 + 2] = Max(Min(oxmax, im_info[1] - 1.), 0.); - proposals[i * 4 + 3] = Max(Min(oymax, im_info[0] - 1.), 0.); + proposals[i * 4] = Max(Min(oxmin, im_info[1] - 1.), 0.); + proposals[i * 4 + 1] = Max(Min(oymin, im_info[0] - 1.), 0.); + proposals[i * 4 + 2] = Max(Min(oxmax, im_info[1] - 1.), 0.); + proposals[i * 4 + 3] = Max(Min(oymax, im_info[0] - 1.), 0.); } -} + + __device__ __forceinline__ T Min(T a, T b) const { return a > b ? b : a; } + + __device__ __forceinline__ T Max(T a, T b) const { return a > b ? a : b; } +}; template -__global__ void FilterBBoxes(const T *bboxes, const T *im_info, - const T min_size, const int num, int *keep_num, - int *keep) { +static __global__ void FilterBBoxes(const T *bboxes, const T *im_info, + const T min_size, const int num, + int *keep_num, int *keep) { T im_h = im_info[0]; T im_w = im_info[1]; T im_scale = im_info[2]; @@ -181,7 +195,7 @@ __global__ void FilterBBoxes(const T *bboxes, const T *im_info, } } -__device__ inline float IoU(const float *a, const float *b) { +static __device__ inline float IoU(const float *a, const float *b) { float left = max(a[0], b[0]), right = min(a[2], b[2]); float top = max(a[1], b[1]), bottom = min(a[3], b[3]); float width = max(right - left + 1, 0.f), height = max(bottom - top + 1, 0.f); @@ -191,8 +205,9 @@ __device__ inline float IoU(const float *a, const float *b) { return inter_s / (s_a + s_b - inter_s); } -__global__ void NMSKernel(const int n_boxes, const float nms_overlap_thresh, - const float *dev_boxes, uint64_t *dev_mask) { +static __global__ void NMSKernel(const int n_boxes, + const float nms_overlap_thresh, + const float *dev_boxes, uint64_t *dev_mask) { const int row_start = blockIdx.y; const int col_start = blockIdx.x; @@ -234,9 +249,9 @@ __global__ void NMSKernel(const int n_boxes, const float nms_overlap_thresh, } template -void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals, - const Tensor &sorted_indices, const T nms_threshold, - Tensor *keep_out) { +static void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals, + const Tensor &sorted_indices, const T nms_threshold, + Tensor *keep_out) { int boxes_num = proposals.dims()[0]; PADDLE_ENFORCE_EQ(boxes_num, sorted_indices.dims()[0]); @@ -247,13 +262,10 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals, const T *boxes = proposals.data(); auto place = boost::get(ctx.GetPlace()); - int size_bytes = boxes_num * col_blocks * sizeof(uint64_t); - uint64_t *d_mask = - reinterpret_cast(memory::Alloc(place, size_bytes)); - NMSKernel<<>>(boxes_num, nms_threshold, boxes, d_mask); - uint64_t *h_mask = reinterpret_cast( - memory::Alloc(platform::CPUPlace(), size_bytes)); - memory::Copy(platform::CPUPlace(), h_mask, place, d_mask, size_bytes, 0); + framework::Vector mask(boxes_num * col_blocks); + NMSKernel<<>>( + boxes_num, nms_threshold, boxes, + mask.CUDAMutableData(boost::get(ctx.GetPlace()))); std::vector remv(col_blocks); memset(&remv[0], 0, sizeof(uint64_t) * col_blocks); @@ -267,7 +279,7 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals, if (!(remv[nblock] & (1ULL << inblock))) { ++num_to_keep; keep_vec.push_back(i); - uint64_t *p = &h_mask[0] + i * col_blocks; + uint64_t *p = &mask[0] + i * col_blocks; for (int j = nblock; j < col_blocks; j++) { remv[j] |= p[j]; } @@ -276,12 +288,10 @@ void NMS(const platform::CUDADeviceContext &ctx, const Tensor &proposals, int *keep = keep_out->mutable_data({num_to_keep}, ctx.GetPlace()); memory::Copy(place, keep, platform::CPUPlace(), keep_vec.data(), sizeof(int) * num_to_keep, 0); - memory::Free(place, d_mask); - memory::Free(platform::CPUPlace(), h_mask); } template -std::pair ProposalForOneImage( +static std::pair ProposalForOneImage( const platform::CUDADeviceContext &ctx, const Tensor &im_info, const Tensor &anchors, const Tensor &variances, const Tensor &bbox_deltas, // [M, 4] @@ -300,18 +310,20 @@ std::pair ProposalForOneImage( // 2. box decode and clipping Tensor proposals; proposals.mutable_data({pre_nms_num, 4}, ctx.GetPlace()); - int block = 512; - auto stream = ctx.stream(); - BoxDecodeAndClipKernel<<>>( - anchors.data(), bbox_deltas.data(), variances.data(), - index_sort.data(), im_info.data(), pre_nms_num, - proposals.data()); + + { + platform::ForRange for_range(ctx, pre_nms_num); + for_range(BoxDecodeAndClipFunctor{ + anchors.data(), bbox_deltas.data(), variances.data(), + index_sort.data(), im_info.data(), proposals.data()}); + } // 3. filter Tensor keep_index, keep_num_t; keep_index.mutable_data({pre_nms_num}, ctx.GetPlace()); keep_num_t.mutable_data({1}, ctx.GetPlace()); min_size = std::max(min_size, 1.0f); + auto stream = ctx.stream(); FilterBBoxes<<<1, 512, 0, stream>>>( proposals.data(), im_info.data(), min_size, pre_nms_num, keep_num_t.data(), keep_index.data()); @@ -355,8 +367,12 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel { auto *scores = context.Input("Scores"); auto *bbox_deltas = context.Input("BboxDeltas"); auto *im_info = context.Input("ImInfo"); - auto *anchors = context.Input("Anchors"); - auto *variances = context.Input("Variances"); + auto anchors = detail::Ref(context.Input("Anchors"), + "Cannot find input Anchors(%s) in scope", + context.Inputs("Anchors")[0]); + auto variances = detail::Ref(context.Input("Variances"), + "Cannot find input Variances(%s) in scope", + context.Inputs("Variances")[0]); auto *rpn_rois = context.Output("RpnRois"); auto *rpn_roi_probs = context.Output("RpnRoiProbs"); @@ -392,10 +408,8 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel { trans(dev_ctx, *bbox_deltas, &bbox_deltas_swap, axis); trans(dev_ctx, *scores, &scores_swap, axis); - Tensor *anchor = const_cast(anchors); - anchor->Resize({anchors->numel() / 4, 4}); - Tensor *var = const_cast(variances); - var->Resize({var->numel() / 4, 4}); + anchors.Resize({anchors.numel() / 4, 4}); + variances.Resize({variances.numel() / 4, 4}); rpn_rois->mutable_data({bbox_deltas->numel() / 4, 4}, context.GetPlace()); @@ -417,12 +431,12 @@ class CUDAGenerateProposalsKernel : public framework::OpKernel { scores_slice.Resize({h_score * w_score * c_score, 1}); std::pair box_score_pair = - ProposalForOneImage(dev_ctx, im_info_slice, *anchor, *var, + ProposalForOneImage(dev_ctx, im_info_slice, anchors, variances, bbox_deltas_slice, scores_slice, pre_nms_top_n, post_nms_top_n, nms_thresh, min_size, eta); - Tensor proposals = box_score_pair.first; - Tensor scores = box_score_pair.second; + Tensor &proposals = box_score_pair.first; + Tensor &scores = box_score_pair.second; memory::Copy(place, rpn_rois_data + num_proposals * 4, place, proposals.data(), sizeof(T) * proposals.numel(), 0); diff --git a/paddle/fluid/operators/detection/gpc.cc b/paddle/fluid/operators/detection/gpc.cc new file mode 100644 index 0000000000000000000000000000000000000000..7c0823c0487d39eece5be08322e7d182b931ba3c --- /dev/null +++ b/paddle/fluid/operators/detection/gpc.cc @@ -0,0 +1,2201 @@ +// 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. + +/** + * @file src/gpc.cpp + * @author huhan02(com@baidu.com) + * @date 2015/12/18 14:17:30 + * @brief + * + * @modified by sunyipeng + * @email sunyipeng@baidu.com + * @date 2018/6/12 + **/ + +#include "paddle/fluid/operators/detection/gpc.h" + +namespace gpc { + +typedef struct lmt_shape { /* Local minima table */ + double y; /* Y coordinate at local minimum */ + edge_node *first_bound; /* Pointer to bound list */ + struct lmt_shape *next; /* Pointer to next local minimum */ +} lmt_node; + +typedef struct sbt_t_shape { /* Scanbeam tree */ + double y; /* Scanbeam node y value */ + struct sbt_t_shape *less; /* Pointer to nodes with lower y */ + struct sbt_t_shape *more; /* Pointer to nodes with higher y */ +} sb_tree; + +typedef struct it_shape { /* Intersection table */ + edge_node *ie[2]; /* Intersecting edge (bundle) pair */ + gpc_vertex point; /* Point of intersection */ + struct it_shape *next; /* The next intersection table node */ +} it_node; + +typedef struct st_shape { /* Sorted edge table */ + edge_node *edge; /* Pointer to AET edge */ + double xb; /* Scanbeam bottom x coordinate */ + double xt; /* Scanbeam top x coordinate */ + double dx; /* Change in x for a unit y increase */ + struct st_shape *prev; /* Previous edge in sorted list */ +} st_node; + +typedef struct bbox_shape { /* Contour axis-aligned bounding box */ + double xmin; /* Minimum x coordinate */ + double ymin; /* Minimum y coordinate */ + double xmax; /* Maximum x coordinate */ + double ymax; /* Maximum y coordinate */ +} bbox; + +/* +=========================================================================== + Global Data +=========================================================================== +*/ + +/* Horizontal edge state transitions within scanbeam boundary */ +const h_state next_h_state[3][6] = { + /* ABOVE BELOW CROSS */ + /* L R L R L R */ + /* NH */ + {BH, TH, TH, BH, NH, NH}, + /* BH */ + {NH, NH, NH, NH, TH, TH}, + /* TH */ + {NH, NH, NH, NH, BH, BH}}; + +/* +=========================================================================== + Private Functions +=========================================================================== +*/ + +static void reset_it(it_node **it) { + it_node *itn; + + while (*it) { + itn = (*it)->next; + gpc_free(*it); + *it = itn; + } +} + +static void reset_lmt(lmt_node **lmt) { + lmt_node *lmtn; + + while (*lmt) { + lmtn = (*lmt)->next; + gpc_free(*lmt); + *lmt = lmtn; + } +} + +static void insert_bound(edge_node **b, edge_node *e) { + edge_node *existing_bound = NULL; + + if (!*b) { + /* Link node e to the tail of the list */ + *b = e; + } else { + /* Do primary sort on the x field */ + if (e[0].bot.x < (*b)[0].bot.x) { + /* Insert a new node mid-list */ + existing_bound = *b; + *b = e; + (*b)->next_bound = existing_bound; + } else { + if (e[0].bot.x == (*b)[0].bot.x) { + /* Do secondary sort on the dx field */ + if (e[0].dx < (*b)[0].dx) { + /* Insert a new node mid-list */ + existing_bound = *b; + *b = e; + (*b)->next_bound = existing_bound; + } else { + /* Head further down the list */ + insert_bound(&((*b)->next_bound), e); + } + } else { + /* Head further down the list */ + insert_bound(&((*b)->next_bound), e); + } + } + } +} + +static edge_node **bound_list(lmt_node **lmt, double y) { + lmt_node *existing_node; + + if (!*lmt) { + /* Add node onto the tail end of the LMT */ + gpc_malloc(*lmt, sizeof(lmt_node), + const_cast("LMT insertion")); + (*lmt)->y = y; + (*lmt)->first_bound = NULL; + (*lmt)->next = NULL; + return &((*lmt)->first_bound); + } else if (y < (*lmt)->y) { + /* Insert a new LMT node before the current node */ + existing_node = *lmt; + gpc_malloc(*lmt, sizeof(lmt_node), + const_cast("LMT insertion")); + (*lmt)->y = y; + (*lmt)->first_bound = NULL; + (*lmt)->next = existing_node; + return &((*lmt)->first_bound); + } else { + if (y > (*lmt)->y) { + /* Head further up the LMT */ + return bound_list(&((*lmt)->next), y); + } else { + /* Use this existing LMT node */ + return &((*lmt)->first_bound); + } + } +} + +static void add_to_sbtree(int *entries, sb_tree **sbtree, double y) { + if (!*sbtree) { + /* Add a new tree node here */ + gpc_malloc(*sbtree, sizeof(sb_tree), + const_cast("scanbeam tree insertion")); + (*sbtree)->y = y; + (*sbtree)->less = NULL; + (*sbtree)->more = NULL; + (*entries)++; + } else { + if ((*sbtree)->y > y) { + /* Head into the 'less' sub-tree */ + add_to_sbtree(entries, &((*sbtree)->less), y); + } else { + if ((*sbtree)->y < y) { + /* Head into the 'more' sub-tree */ + add_to_sbtree(entries, &((*sbtree)->more), y); + } + } + } +} + +static void build_sbt(int *entries, double *sbt, sb_tree *sbtree) { + if (sbtree->less) { + build_sbt(entries, sbt, sbtree->less); + } + sbt[*entries] = sbtree->y; + (*entries)++; + if (sbtree->more) { + build_sbt(entries, sbt, sbtree->more); + } +} + +static void free_sbtree(sb_tree **sbtree) { + if (*sbtree) { + free_sbtree(&((*sbtree)->less)); + free_sbtree(&((*sbtree)->more)); + gpc_free(*sbtree); + } +} + +static int count_optimal_vertices(gpc_vertex_list c) { + int result = 0; + int i = 0; + + /* Ignore non-contributing contours */ + if (c.num_vertices > 0) { + for (i = 0; i < c.num_vertices; i++) { + /* Ignore superfluous vertices embedded in horizontal edges */ + if (gpc_optimal(c.vertex, i, c.num_vertices)) { + result++; + } + } + } + return result; +} + +static edge_node *build_lmt(lmt_node **lmt, sb_tree **sbtree, int *sbt_entries, + gpc_polygon *p, int type, gpc_op op) { + int c = 0; + int i = 0; + int min = 0; + int max = 0; + int num_edges = 0; + int v = 0; + int num_vertices = 0; + int total_vertices = 0; + int e_index = 0; + edge_node *e = NULL; + edge_node *edge_table = NULL; + + for (c = 0; c < p->num_contours; c++) { + total_vertices += count_optimal_vertices(p->contour[c]); + } + + /* Create the entire input polygon edge table in one go */ + gpc_malloc(edge_table, total_vertices * sizeof(edge_node), + const_cast("edge table creation")); + + for (c = 0; c < p->num_contours; c++) { + if (p->contour[c].num_vertices < 0) { + /* Ignore the non-contributing contour and repair the vertex count */ + p->contour[c].num_vertices = -p->contour[c].num_vertices; + } else { + /* Perform contour optimisation */ + num_vertices = 0; + for (i = 0; i < p->contour[c].num_vertices; i++) { + if (gpc_optimal(p->contour[c].vertex, i, p->contour[c].num_vertices)) { + edge_table[num_vertices].vertex.x = p->contour[c].vertex[i].x; + edge_table[num_vertices].vertex.y = p->contour[c].vertex[i].y; + + /* Record vertex in the scanbeam table */ + add_to_sbtree(sbt_entries, sbtree, edge_table[num_vertices].vertex.y); + + num_vertices++; + } + } + + /* Do the contour forward pass */ + for (min = 0; min < num_vertices; min++) { + /* If a forward local minimum... */ + if (gpc_fwd_min(edge_table, min, num_vertices)) { + /* Search for the next local maximum... */ + num_edges = 1; + max = gpc_next_index(min, num_vertices); + while (gpc_not_fmax(edge_table, max, num_vertices)) { + num_edges++; + max = gpc_next_index(max, num_vertices); + } + + /* Build the next edge list */ + e = &edge_table[e_index]; + e_index += num_edges; + v = min; + e[0].bstate[BELOW] = UNBUNDLED; + e[0].bundle[BELOW][CLIP] = 0; + e[0].bundle[BELOW][SUBJ] = 0; + for (i = 0; i < num_edges; i++) { + e[i].xb = edge_table[v].vertex.x; + e[i].bot.x = edge_table[v].vertex.x; + e[i].bot.y = edge_table[v].vertex.y; + + v = gpc_next_index(v, num_vertices); + + e[i].top.x = edge_table[v].vertex.x; + e[i].top.y = edge_table[v].vertex.y; + e[i].dx = (edge_table[v].vertex.x - e[i].bot.x) / + (e[i].top.y - e[i].bot.y); + e[i].type = type; + e[i].outp[ABOVE] = NULL; + e[i].outp[BELOW] = NULL; + e[i].next = NULL; + e[i].prev = NULL; + e[i].succ = + ((num_edges > 1) && (i < (num_edges - 1))) ? &(e[i + 1]) : NULL; + e[i].pred = ((num_edges > 1) && (i > 0)) ? &(e[i - 1]) : NULL; + e[i].next_bound = NULL; + e[i].bside[CLIP] = (op == GPC_DIFF) ? RIGHT : LEFT; + e[i].bside[SUBJ] = LEFT; + } + insert_bound(bound_list(lmt, edge_table[min].vertex.y), e); + } + } + + /* Do the contour reverse pass */ + for (min = 0; min < num_vertices; min++) { + /* If a reverse local minimum... */ + if (gpc_rev_min(edge_table, min, num_vertices)) { + /* Search for the previous local maximum... */ + num_edges = 1; + max = gpc_prev_index(min, num_vertices); + while (gpc_not_rmax(edge_table, max, num_vertices)) { + num_edges++; + max = gpc_prev_index(max, num_vertices); + } + + /* Build the previous edge list */ + e = &edge_table[e_index]; + e_index += num_edges; + v = min; + e[0].bstate[BELOW] = UNBUNDLED; + e[0].bundle[BELOW][CLIP] = 0; + e[0].bundle[BELOW][SUBJ] = 0; + for (i = 0; i < num_edges; i++) { + e[i].xb = edge_table[v].vertex.x; + e[i].bot.x = edge_table[v].vertex.x; + e[i].bot.y = edge_table[v].vertex.y; + + v = gpc_prev_index(v, num_vertices); + + e[i].top.x = edge_table[v].vertex.x; + e[i].top.y = edge_table[v].vertex.y; + e[i].dx = (edge_table[v].vertex.x - e[i].bot.x) / + (e[i].top.y - e[i].bot.y); + e[i].type = type; + e[i].outp[ABOVE] = NULL; + e[i].outp[BELOW] = NULL; + e[i].next = NULL; + e[i].prev = NULL; + e[i].succ = + ((num_edges > 1) && (i < (num_edges - 1))) ? &(e[i + 1]) : NULL; + e[i].pred = ((num_edges > 1) && (i > 0)) ? &(e[i - 1]) : NULL; + e[i].next_bound = NULL; + e[i].bside[CLIP] = (op == GPC_DIFF) ? RIGHT : LEFT; + e[i].bside[SUBJ] = LEFT; + } + insert_bound(bound_list(lmt, edge_table[min].vertex.y), e); + } + } + } + } + return edge_table; +} // NOLINT + +static void add_edge_to_aet(edge_node **aet, edge_node *edge, edge_node *prev) { + if (!*aet) { + /* Append edge onto the tail end of the AET */ + *aet = edge; + edge->prev = prev; + edge->next = NULL; + } else { + /* Do primary sort on the xb field */ + if (edge->xb < (*aet)->xb) { + /* Insert edge here (before the AET edge) */ + edge->prev = prev; + edge->next = *aet; + (*aet)->prev = edge; + *aet = edge; + } else { + if (edge->xb == (*aet)->xb) { + /* Do secondary sort on the dx field */ + if (edge->dx < (*aet)->dx) { + /* Insert edge here (before the AET edge) */ + edge->prev = prev; + edge->next = *aet; + (*aet)->prev = edge; + *aet = edge; + } else { + /* Head further into the AET */ + add_edge_to_aet(&((*aet)->next), edge, *aet); + } + } else { + /* Head further into the AET */ + add_edge_to_aet(&((*aet)->next), edge, *aet); + } + } + } +} + +static void add_intersection(it_node **it, edge_node *edge0, edge_node *edge1, + double x, double y) { + it_node *existing_node; + + if (!*it) { + /* Append a new node to the tail of the list */ + gpc_malloc(*it, sizeof(it_node), + const_cast("IT insertion")); + (*it)->ie[0] = edge0; + (*it)->ie[1] = edge1; + (*it)->point.x = x; + (*it)->point.y = y; + (*it)->next = NULL; + } else { + if ((*it)->point.y > y) { + /* Insert a new node mid-list */ + existing_node = *it; + gpc_malloc(*it, sizeof(it_node), + const_cast("IT insertion")); + (*it)->ie[0] = edge0; + (*it)->ie[1] = edge1; + (*it)->point.x = x; + (*it)->point.y = y; + (*it)->next = existing_node; + } else { + /* Head further down the list */ + add_intersection(&((*it)->next), edge0, edge1, x, y); + } + } +} + +static void add_st_edge(st_node **st, it_node **it, edge_node *edge, + double dy) { + st_node *existing_node; + double den = 0.0; + double r = 0.0; + double x = 0.0; + double y = 0.0; + + if (!*st) { + /* Append edge onto the tail end of the ST */ + gpc_malloc(*st, sizeof(st_node), + const_cast("ST insertion")); + (*st)->edge = edge; + (*st)->xb = edge->xb; + (*st)->xt = edge->xt; + (*st)->dx = edge->dx; + (*st)->prev = NULL; + } else { + den = ((*st)->xt - (*st)->xb) - (edge->xt - edge->xb); + + /* If new edge and ST edge don't cross */ + if ((edge->xt >= (*st)->xt) || (edge->dx == (*st)->dx) || + (fabs(den) <= DBL_EPSILON)) { + /* No intersection - insert edge here (before the ST edge) */ + existing_node = *st; + gpc_malloc(*st, sizeof(st_node), + const_cast("ST insertion")); + (*st)->edge = edge; + (*st)->xb = edge->xb; + (*st)->xt = edge->xt; + (*st)->dx = edge->dx; + (*st)->prev = existing_node; + } else { + /* Compute intersection between new edge and ST edge */ + r = (edge->xb - (*st)->xb) / den; + x = (*st)->xb + r * ((*st)->xt - (*st)->xb); + y = r * dy; + + /* Insert the edge pointers and the intersection point in the IT */ + add_intersection(it, (*st)->edge, edge, x, y); + + /* Head further into the ST */ + add_st_edge(&((*st)->prev), it, edge, dy); + } + } +} + +static void build_intersection_table(it_node **it, edge_node *aet, double dy) { + st_node *st; + st_node *stp; + edge_node *edge = NULL; + + /* Build intersection table for the current scanbeam */ + reset_it(it); + st = NULL; + + /* Process each AET edge */ + for (edge = aet; edge; edge = edge->next) { + if ((edge->bstate[ABOVE] == BUNDLE_HEAD) || edge->bundle[ABOVE][CLIP] || + edge->bundle[ABOVE][SUBJ]) { + add_st_edge(&st, it, edge, dy); + } + } + + /* Free the sorted edge table */ + while (st) { + stp = st->prev; + gpc_free(st); + st = stp; + } +} + +static int count_contours(polygon_node *polygon) { + int nc = 0; + int nv = 0; + vertex_node *v = NULL; + vertex_node *nextv = NULL; + + for (nc = 0; polygon; polygon = polygon->next) { + if (polygon->active) { + /* Count the vertices in the current contour */ + nv = 0; + for (v = polygon->proxy->v[LEFT]; v; v = v->next) { + nv++; + } + + /* Record valid vertex counts in the active field */ + if (nv > 2) { + polygon->active = nv; + nc++; + } else { + /* Invalid contour: just free the heap */ + for (v = polygon->proxy->v[LEFT]; v; v = nextv) { + nextv = v->next; + gpc_free(v); + } + polygon->active = 0; + } + } + } + return nc; +} + +static void add_left(polygon_node *p, double x, double y) { + vertex_node *nv = NULL; + + /* Create a new vertex node and set its fields */ + gpc_malloc(nv, sizeof(vertex_node), + const_cast("vertex node creation")); + nv->x = x; + nv->y = y; + + /* Add vertex nv to the left end of the polygon's vertex list */ + nv->next = p->proxy->v[LEFT]; + + /* Update proxy->[LEFT] to point to nv */ + p->proxy->v[LEFT] = nv; +} + +static void merge_left(polygon_node *p, polygon_node *q, polygon_node *list) { + polygon_node *target = NULL; + + /* Label contour as a hole */ + q->proxy->hole = 1; + + if (p->proxy != q->proxy) { + /* Assign p's vertex list to the left end of q's list */ + p->proxy->v[RIGHT]->next = q->proxy->v[LEFT]; + q->proxy->v[LEFT] = p->proxy->v[LEFT]; + + /* Redirect any p->proxy references to q->proxy */ + + for (target = p->proxy; list; list = list->next) { + if (list->proxy == target) { + list->active = 0; + list->proxy = q->proxy; + } + } + } +} + +static void add_right(polygon_node *p, double x, double y) { + vertex_node *nv = NULL; + + /* Create a new vertex node and set its fields */ + gpc_malloc(nv, sizeof(vertex_node), + const_cast("vertex node creation")); + nv->x = x; + nv->y = y; + nv->next = NULL; + + /* Add vertex nv to the right end of the polygon's vertex list */ + p->proxy->v[RIGHT]->next = nv; + + /* Update proxy->v[RIGHT] to point to nv */ + p->proxy->v[RIGHT] = nv; +} + +static void merge_right(polygon_node *p, polygon_node *q, polygon_node *list) { + polygon_node *target = NULL; + + /* Label contour as external */ + q->proxy->hole = 0; + + if (p->proxy != q->proxy) { + /* Assign p's vertex list to the right end of q's list */ + q->proxy->v[RIGHT]->next = p->proxy->v[LEFT]; + q->proxy->v[RIGHT] = p->proxy->v[RIGHT]; + + /* Redirect any p->proxy references to q->proxy */ + for (target = p->proxy; list; list = list->next) { + if (list->proxy == target) { + list->active = 0; + list->proxy = q->proxy; + } + } + } +} + +static void add_local_min(polygon_node **p, edge_node *edge, double x, + double y) { + polygon_node *existing_min = NULL; + vertex_node *nv = NULL; + + existing_min = *p; + + gpc_malloc(*p, sizeof(polygon_node), + const_cast("polygon node creation")); + + /* Create a new vertex node and set its fields */ + gpc_malloc(nv, sizeof(vertex_node), + const_cast("vertex node creation")); + nv->x = x; + nv->y = y; + nv->next = NULL; + + /* Initialise proxy to point to p itself */ + (*p)->proxy = (*p); + (*p)->active = 1; + (*p)->next = existing_min; + + /* Make v[LEFT] and v[RIGHT] point to new vertex nv */ + (*p)->v[LEFT] = nv; + (*p)->v[RIGHT] = nv; + + /* Assign polygon p to the edge */ + edge->outp[ABOVE] = *p; +} + +static int count_tristrips(polygon_node *tn) { + int total = 0; + + for (total = 0; tn; tn = tn->next) { + if (tn->active > 2) { + total++; + } + } + return total; +} + +void add_vertex(vertex_node **t, double x, double y) { + if (!(*t)) { + gpc_malloc(*t, sizeof(vertex_node), + const_cast("tristrip vertex creation")); + (*t)->x = x; + (*t)->y = y; + (*t)->next = NULL; + } else { + /* Head further down the list */ + add_vertex(&((*t)->next), x, y); + } +} + +void gpc_vertex_create(edge_node *e, int p, int s, double x, double y) { + add_vertex(&(e->outp[p]->v[s]), x, y); + e->outp[p]->active++; +} + +static void new_tristrip(polygon_node **tn, edge_node *edge, double x, + double y) { + if (!(*tn)) { + gpc_malloc(*tn, sizeof(polygon_node), + const_cast("tristrip node creation")); + (*tn)->next = NULL; + (*tn)->v[LEFT] = NULL; + (*tn)->v[RIGHT] = NULL; + (*tn)->active = 1; + add_vertex(&((*tn)->v[LEFT]), x, y); + edge->outp[ABOVE] = *tn; + } else { + /* Head further down the list */ + new_tristrip(&((*tn)->next), edge, x, y); + } +} + +static bbox *create_contour_bboxes(gpc_polygon *p) { + bbox *box; + int c = 0; + int v = 0; + + gpc_malloc(box, p->num_contours * sizeof(bbox), + const_cast("Bounding box creation")); + + /* Construct contour bounding boxes */ + for (c = 0; c < p->num_contours; c++) { + /* Initialise bounding box extent */ + box[c].xmin = DBL_MAX; + box[c].ymin = DBL_MAX; + box[c].xmax = -DBL_MAX; + box[c].ymax = -DBL_MAX; + + for (v = 0; v < p->contour[c].num_vertices; v++) { + /* Adjust bounding box */ + if (p->contour[c].vertex[v].x < box[c].xmin) { + box[c].xmin = p->contour[c].vertex[v].x; + } + if (p->contour[c].vertex[v].y < box[c].ymin) { + box[c].ymin = p->contour[c].vertex[v].y; + } + if (p->contour[c].vertex[v].x > box[c].xmax) { + box[c].xmax = p->contour[c].vertex[v].x; + } + if (p->contour[c].vertex[v].y > box[c].ymax) { + box[c].ymax = p->contour[c].vertex[v].y; + } + } + } + return box; +} + +static void minimax_test(gpc_polygon *subj, gpc_polygon *clip, gpc_op op) { + bbox *s_bbox; + bbox *c_bbox; + int s = 0; + int c = 0; + int *o_table = NULL; + int overlap = 0; + + s_bbox = create_contour_bboxes(subj); + c_bbox = create_contour_bboxes(clip); + + gpc_malloc(o_table, + subj->num_contours * clip->num_contours * sizeof(int), + const_cast("overlap table creation")); + + /* Check all subject contour bounding boxes against clip boxes */ + for (s = 0; s < subj->num_contours; s++) { + for (c = 0; c < clip->num_contours; c++) { + o_table[c * subj->num_contours + s] = + (!((s_bbox[s].xmax < c_bbox[c].xmin) || + (s_bbox[s].xmin > c_bbox[c].xmax))) && + (!((s_bbox[s].ymax < c_bbox[c].ymin) || + (s_bbox[s].ymin > c_bbox[c].ymax))); + } + } + + /* For each clip contour, search for any subject contour overlaps */ + for (c = 0; c < clip->num_contours; c++) { + overlap = 0; + for (s = 0; (!overlap) && (s < subj->num_contours); s++) { + overlap = o_table[c * subj->num_contours + s]; + } + + if (!overlap) { + /* Flag non contributing status by negating vertex count */ + clip->contour[c].num_vertices = -clip->contour[c].num_vertices; + } + } + + if (op == GPC_INT) { + /* For each subject contour, search for any clip contour overlaps */ + for (s = 0; s < subj->num_contours; s++) { + overlap = 0; + for (c = 0; (!overlap) && (c < clip->num_contours); c++) { + overlap = o_table[c * subj->num_contours + s]; + } + + if (!overlap) { + /* Flag non contributing status by negating vertex count */ + subj->contour[s].num_vertices = -subj->contour[s].num_vertices; + } + } + } + + gpc_free(s_bbox); + gpc_free(c_bbox); + gpc_free(o_table); +} + +/* +=========================================================================== + Public Functions +=========================================================================== +*/ + +void gpc_free_polygon(gpc_polygon *p) { + int c = 0; + + for (c = 0; c < p->num_contours; c++) { + gpc_free(p->contour[c].vertex); + } + gpc_free(p->hole); + gpc_free(p->contour); + p->num_contours = 0; +} + +/* +void gpc_read_polygon(FILE *fp, int read_hole_flags, gpc_polygon *p) { + int c = 0; + int v = 0; + + fscanf(fp, "%d", &(p->num_contours)); + gpc_malloc(p->hole, p->num_contours * sizeof(int), + (char *)"hole flag array creation"); + gpc_malloc(p->contour, + p->num_contours * sizeof(gpc_vertex_list), + (char *)"contour creation"); + for (c = 0; c < p->num_contours; c++) { + fscanf(fp, "%d", &(p->contour[c].num_vertices)); + + if (read_hole_flags) { + fscanf(fp, "%d", &(p->hole[c])); + } else { + p->hole[c] = 0; // Assume all contours to be external + } + + gpc_malloc(p->contour[c].vertex, + p->contour[c].num_vertices * sizeof(gpc_vertex), + (char *)"vertex creation"); + for (v = 0; v < p->contour[c].num_vertices; v++) { + fscanf(fp, "%lf %lf", &(p->contour[c].vertex[v].x), + &(p->contour[c].vertex[v].y)); + } + } +} + +void gpc_write_polygon(FILE *fp, int write_hole_flags, gpc_polygon *p) { + int c = 0; + int v = 0; + + fprintf(fp, "%d\n", p->num_contours); + for (c = 0; c < p->num_contours; c++) { + fprintf(fp, "%d\n", p->contour[c].num_vertices); + + if (write_hole_flags) { + fprintf(fp, "%d\n", p->hole[c]); + } + + for (v = 0; v < p->contour[c].num_vertices; v++) { + fprintf(fp, "% .*lf % .*lf\n", DBL_DIG, p->contour[c].vertex[v].x, + DBL_DIG, p->contour[c].vertex[v].y); + } + } +} +*/ + +void gpc_add_contour(gpc_polygon *p, gpc_vertex_list *new_contour, int hole) { + int *extended_hole = NULL; + int c = 0; + int v = 0; + gpc_vertex_list *extended_contour = NULL; + + /* Create an extended hole array */ + gpc_malloc(extended_hole, (p->num_contours + 1) * sizeof(int), + const_cast("contour hole addition")); + + /* Create an extended contour array */ + gpc_malloc(extended_contour, + (p->num_contours + 1) * sizeof(gpc_vertex_list), + const_cast("contour addition")); + + /* Copy the old contour and hole data into the extended arrays */ + for (c = 0; c < p->num_contours; c++) { + extended_hole[c] = p->hole[c]; + extended_contour[c] = p->contour[c]; + } + + /* Copy the new contour and hole onto the end of the extended arrays */ + c = p->num_contours; + extended_hole[c] = hole; + extended_contour[c].num_vertices = new_contour->num_vertices; + gpc_malloc(extended_contour[c].vertex, + new_contour->num_vertices * sizeof(gpc_vertex), + const_cast("contour addition")); + for (v = 0; v < new_contour->num_vertices; v++) { + extended_contour[c].vertex[v] = new_contour->vertex[v]; + } + + /* Dispose of the old contour */ + gpc_free(p->contour); + gpc_free(p->hole); + + /* Update the polygon information */ + p->num_contours++; + p->hole = extended_hole; + p->contour = extended_contour; +} + +// gpc_polygon_clip +void gpc_polygon_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip, + gpc_polygon *result) { + sb_tree *sbtree = NULL; + it_node *it = NULL; + it_node *intersect = NULL; + edge_node *edge = NULL; + edge_node *prev_edge = NULL; + edge_node *next_edge = NULL; + edge_node *succ_edge = NULL; + edge_node *e0 = NULL; + edge_node *e1 = NULL; + edge_node *aet = NULL; + edge_node *c_heap = NULL; + edge_node *s_heap = NULL; + lmt_node *lmt = NULL; + lmt_node *local_min = NULL; + polygon_node *out_poly = NULL; + polygon_node *p = NULL; + polygon_node *q = NULL; + polygon_node *poly = NULL; + polygon_node *npoly = NULL; + polygon_node *cf = NULL; + vertex_node *vtx = NULL; + vertex_node *nv = NULL; + h_state horiz[2]; + int in[2]; + int exists[2]; + int parity[2] = {LEFT, LEFT}; + int c = 0; + int v = 0; + int contributing = 0; + int search = 0; + int scanbeam = 0; + int sbt_entries = 0; + int vclass = 0; + int bl = 0; + int br = 0; + int tl = 0; + int tr = 0; + double *sbt = NULL; + double xb = 0.0; + double px = 0.0; + double yb = 0.0; + double yt = 0.0; + double dy = 0.0; + double ix = 0.0; + double iy = 0.0; + + /* Test for trivial NULL result cases */ + if (((subj->num_contours == 0) && (clip->num_contours == 0)) || + ((subj->num_contours == 0) && ((op == GPC_INT) || (op == GPC_DIFF))) || + ((clip->num_contours == 0) && (op == GPC_INT))) { + result->num_contours = 0; + result->hole = NULL; + result->contour = NULL; + return; + } + /* Identify potentialy contributing contours */ + if (((op == GPC_INT) || (op == GPC_DIFF)) && (subj->num_contours > 0) && + (clip->num_contours > 0)) { + minimax_test(subj, clip, op); + } + /* Build LMT */ + if (subj->num_contours > 0) { + s_heap = build_lmt(&lmt, &sbtree, &sbt_entries, subj, SUBJ, op); + } + if (clip->num_contours > 0) { + c_heap = build_lmt(&lmt, &sbtree, &sbt_entries, clip, CLIP, op); + } + /* Return a NULL result if no contours contribute */ + if (lmt == NULL) { + result->num_contours = 0; + result->hole = NULL; + result->contour = NULL; + reset_lmt(&lmt); + gpc_free(s_heap); + gpc_free(c_heap); + return; + } + + /* Build scanbeam table from scanbeam tree */ + gpc_malloc(sbt, sbt_entries * sizeof(double), + const_cast("sbt creation")); + build_sbt(&scanbeam, sbt, sbtree); + scanbeam = 0; + free_sbtree(&sbtree); + /* Allow pointer re-use without causing memory leak */ + if (subj == result) { + gpc_free_polygon(subj); + } + if (clip == result) { + gpc_free_polygon(clip); + } + /* Invert clip polygon for difference operation */ + if (op == GPC_DIFF) { + parity[CLIP] = RIGHT; + } + local_min = lmt; + + // Process each scanbeam + while (scanbeam < sbt_entries) { + /* Set yb and yt to the bottom and top of the scanbeam */ + yb = sbt[scanbeam++]; + if (scanbeam < sbt_entries) { + yt = sbt[scanbeam]; + dy = yt - yb; + } + /* === SCANBEAM BOUNDARY PROCESSING ================================ */ + /* If LMT node corresponding to yb exists */ + if (local_min) { + if (local_min->y == yb) { + /* Add edges starting at this local minimum to the AET */ + for (edge = local_min->first_bound; edge; edge = edge->next_bound) { + add_edge_to_aet(&aet, edge, NULL); + } + local_min = local_min->next; + } + } + /* Set dummy previous x value */ + px = -DBL_MAX; + /* Create bundles within AET */ + e0 = aet; + e1 = aet; + /* Set up bundle fields of first edge */ + aet->bundle[ABOVE][aet->type] = (aet->top.y != yb); + aet->bundle[ABOVE][!aet->type] = 0; + aet->bstate[ABOVE] = UNBUNDLED; + + for (next_edge = aet->next; next_edge; next_edge = next_edge->next) { + /* Set up bundle fields of next edge */ + next_edge->bundle[ABOVE][next_edge->type] = (next_edge->top.y != yb); + next_edge->bundle[ABOVE][!next_edge->type] = 0; + next_edge->bstate[ABOVE] = UNBUNDLED; + /* Bundle edges above the scanbeam boundary if they coincide */ + if (next_edge->bundle[ABOVE][next_edge->type]) { + if (gpc_eq(e0->xb, next_edge->xb) && gpc_eq(e0->dx, next_edge->dx) && + (e0->top.y != yb)) { + next_edge->bundle[ABOVE][next_edge->type] ^= + e0->bundle[ABOVE][next_edge->type]; + next_edge->bundle[ABOVE][!next_edge->type] = + e0->bundle[ABOVE][!next_edge->type]; + next_edge->bstate[ABOVE] = BUNDLE_HEAD; + e0->bundle[ABOVE][CLIP] = 0; + e0->bundle[ABOVE][SUBJ] = 0; + e0->bstate[ABOVE] = BUNDLE_TAIL; + } + e0 = next_edge; + } + } + horiz[CLIP] = NH; + horiz[SUBJ] = NH; + + // Process each edge at this scanbeam boundary + for (edge = aet; edge; edge = edge->next) { + exists[CLIP] = + edge->bundle[ABOVE][CLIP] + (edge->bundle[BELOW][CLIP] << 1); + exists[SUBJ] = + edge->bundle[ABOVE][SUBJ] + (edge->bundle[BELOW][SUBJ] << 1); + if (exists[CLIP] || exists[SUBJ]) { + /* Set bundle side */ + edge->bside[CLIP] = parity[CLIP]; + edge->bside[SUBJ] = parity[SUBJ]; + /* Determine contributing status and quadrant occupancies */ + switch (op) { + case GPC_DIFF: + case GPC_INT: + contributing = (exists[CLIP] && (parity[SUBJ] || horiz[SUBJ])) || + (exists[SUBJ] && (parity[CLIP] || horiz[CLIP])) || + (exists[CLIP] && exists[SUBJ] && + (parity[CLIP] == parity[SUBJ])); + br = (parity[CLIP]) && (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) && + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) && + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) && + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + case GPC_XOR: + contributing = exists[CLIP] || exists[SUBJ]; + br = (parity[CLIP]) ^ (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) ^ + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) ^ + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) ^ + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + case GPC_UNION: + contributing = (exists[CLIP] && (!parity[SUBJ] || horiz[SUBJ])) || + (exists[SUBJ] && (!parity[CLIP] || horiz[CLIP])) || + (exists[CLIP] && exists[SUBJ] && + (parity[CLIP] == parity[SUBJ])); + br = (parity[CLIP]) || (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) || + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) || + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) || + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + } + // Update parity + parity[CLIP] ^= edge->bundle[ABOVE][CLIP]; + parity[SUBJ] ^= edge->bundle[ABOVE][SUBJ]; + /* Update horizontal state */ + if (exists[CLIP]) { + horiz[CLIP] = next_h_state[horiz[CLIP]] + [((exists[CLIP] - 1) << 1) + parity[CLIP]]; + } + if (exists[SUBJ]) { + horiz[SUBJ] = next_h_state[horiz[SUBJ]] + [((exists[SUBJ] - 1) << 1) + parity[SUBJ]]; + } + vclass = tr + (tl << 1) + (br << 2) + (bl << 3); + if (contributing) { + xb = edge->xb; + switch (vclass) { + case EMN: + case IMN: + add_local_min(&out_poly, edge, xb, yb); + px = xb; + cf = edge->outp[ABOVE]; + break; + case ERI: + if (xb != px) { + add_right(cf, xb, yb); + px = xb; + } + edge->outp[ABOVE] = cf; + cf = NULL; + break; + case ELI: + add_left(edge->outp[BELOW], xb, yb); + px = xb; + cf = edge->outp[BELOW]; + break; + case EMX: + if (xb != px) { + add_left(cf, xb, yb); + px = xb; + } + merge_right(cf, edge->outp[BELOW], out_poly); + cf = NULL; + break; + case ILI: + if (xb != px) { + add_left(cf, xb, yb); + px = xb; + } + edge->outp[ABOVE] = cf; + cf = NULL; + break; + case IRI: + add_right(edge->outp[BELOW], xb, yb); + px = xb; + cf = edge->outp[BELOW]; + edge->outp[BELOW] = NULL; + break; + case IMX: + if (xb != px) { + add_right(cf, xb, yb); + px = xb; + } + merge_left(cf, edge->outp[BELOW], out_poly); + cf = NULL; + edge->outp[BELOW] = NULL; + break; + case IMM: + if (xb != px) { + add_right(cf, xb, yb); + px = xb; + } + merge_left(cf, edge->outp[BELOW], out_poly); + edge->outp[BELOW] = NULL; + add_local_min(&out_poly, edge, xb, yb); + cf = edge->outp[ABOVE]; + break; + case EMM: + if (xb != px) { + add_left(cf, xb, yb); + px = xb; + } + merge_right(cf, edge->outp[BELOW], out_poly); + edge->outp[BELOW] = NULL; + add_local_min(&out_poly, edge, xb, yb); + cf = edge->outp[ABOVE]; + break; + case LED: + if (edge->bot.y == yb) { + add_left(edge->outp[BELOW], xb, yb); + } + edge->outp[ABOVE] = edge->outp[BELOW]; + px = xb; + break; + case RED: + if (edge->bot.y == yb) { + add_right(edge->outp[BELOW], xb, yb); + } + edge->outp[ABOVE] = edge->outp[BELOW]; + px = xb; + break; + default: + break; + } /* End of switch */ + } /* End of contributing conditional */ + } /* End of edge exists conditional */ + } // End of AET loop + + /* Delete terminating edges from the AET, otherwise compute xt */ + for (edge = aet; edge; edge = edge->next) { + if (edge->top.y == yb) { + prev_edge = edge->prev; + next_edge = edge->next; + if (prev_edge) { + prev_edge->next = next_edge; + } else { + aet = next_edge; + } + if (next_edge) { + next_edge->prev = prev_edge; + } + /* Copy bundle head state to the adjacent tail edge if required */ + if ((edge->bstate[BELOW] == BUNDLE_HEAD) && prev_edge) { + if (prev_edge->bstate[BELOW] == BUNDLE_TAIL) { + prev_edge->outp[BELOW] = edge->outp[BELOW]; + prev_edge->bstate[BELOW] = UNBUNDLED; + if (prev_edge->prev) { + if (prev_edge->prev->bstate[BELOW] == BUNDLE_TAIL) { + prev_edge->bstate[BELOW] = BUNDLE_HEAD; + } + } + } + } + } else { + if (edge->top.y == yt) { + edge->xt = edge->top.x; + } else { + edge->xt = edge->bot.x + edge->dx * (yt - edge->bot.y); + } + } + } + + if (scanbeam < sbt_entries) { + /* === SCANBEAM INTERIOR PROCESSING ============================== */ + build_intersection_table(&it, aet, dy); + /* Process each node in the intersection table */ + for (intersect = it; intersect; intersect = intersect->next) { + e0 = intersect->ie[0]; + e1 = intersect->ie[1]; + /* Only generate output for contributing intersections */ + if ((e0->bundle[ABOVE][CLIP] || e0->bundle[ABOVE][SUBJ]) && + (e1->bundle[ABOVE][CLIP] || e1->bundle[ABOVE][SUBJ])) { + p = e0->outp[ABOVE]; + q = e1->outp[ABOVE]; + ix = intersect->point.x; + iy = intersect->point.y + yb; + + in[CLIP] = (e0->bundle[ABOVE][CLIP] && !e0->bside[CLIP]) || + (e1->bundle[ABOVE][CLIP] && e1->bside[CLIP]) || + (!e0->bundle[ABOVE][CLIP] && !e1->bundle[ABOVE][CLIP] && + e0->bside[CLIP] && e1->bside[CLIP]); + in[SUBJ] = (e0->bundle[ABOVE][SUBJ] && !e0->bside[SUBJ]) || + (e1->bundle[ABOVE][SUBJ] && e1->bside[SUBJ]) || + (!e0->bundle[ABOVE][SUBJ] && !e1->bundle[ABOVE][SUBJ] && + e0->bside[SUBJ] && e1->bside[SUBJ]); + + // Determine quadrant occupancies + switch (op) { + case GPC_DIFF: + case GPC_INT: + tr = (in[CLIP]) && (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + case GPC_XOR: + tr = (in[CLIP]) ^ (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + case GPC_UNION: + tr = (in[CLIP]) || (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + } + vclass = tr + (tl << 1) + (br << 2) + (bl << 3); + switch (vclass) { + case EMN: + add_local_min(&out_poly, e0, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + break; + case ERI: + if (p) { + add_right(p, ix, iy); + e1->outp[ABOVE] = p; + e0->outp[ABOVE] = NULL; + } + break; + case ELI: + if (q) { + add_left(q, ix, iy); + e0->outp[ABOVE] = q; + e1->outp[ABOVE] = NULL; + } + break; + case EMX: + if (p && q) { + add_left(p, ix, iy); + merge_right(p, q, out_poly); + e0->outp[ABOVE] = NULL; + e1->outp[ABOVE] = NULL; + } + break; + case IMN: + add_local_min(&out_poly, e0, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + break; + case ILI: + if (p) { + add_left(p, ix, iy); + e1->outp[ABOVE] = p; + e0->outp[ABOVE] = NULL; + } + break; + case IRI: + if (q) { + add_right(q, ix, iy); + e0->outp[ABOVE] = q; + e1->outp[ABOVE] = NULL; + } + break; + case IMX: + if (p && q) { + add_right(p, ix, iy); + merge_left(p, q, out_poly); + e0->outp[ABOVE] = NULL; + e1->outp[ABOVE] = NULL; + } + break; + case IMM: + if (p && q) { + add_right(p, ix, iy); + merge_left(p, q, out_poly); + add_local_min(&out_poly, e0, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + } + break; + case EMM: + if (p && q) { + add_left(p, ix, iy); + merge_right(p, q, out_poly); + add_local_min(&out_poly, e0, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + } + break; + default: + break; + } // End of switch + } /* End of contributing intersection conditional */ + + /* Swap bundle sides in response to edge crossing */ + if (e0->bundle[ABOVE][CLIP]) { + e1->bside[CLIP] = !e1->bside[CLIP]; + } + if (e1->bundle[ABOVE][CLIP]) { + e0->bside[CLIP] = !e0->bside[CLIP]; + } + if (e0->bundle[ABOVE][SUBJ]) { + e1->bside[SUBJ] = !e1->bside[SUBJ]; + } + if (e1->bundle[ABOVE][SUBJ]) { + e0->bside[SUBJ] = !e0->bside[SUBJ]; + } + + /* Swap e0 and e1 bundles in the AET */ + prev_edge = e0->prev; + next_edge = e1->next; + if (next_edge) { + next_edge->prev = e0; + } + if (e0->bstate[ABOVE] == BUNDLE_HEAD) { + search = 1; + while (search) { + prev_edge = prev_edge->prev; + if (prev_edge) { + if (prev_edge->bstate[ABOVE] != BUNDLE_TAIL) { + search = 0; + } + } else { + search = 0; + } + } + } + if (!prev_edge) { + aet->prev = e1; + e1->next = aet; + aet = e0->next; + } else { + prev_edge->next->prev = e1; + e1->next = prev_edge->next; + prev_edge->next = e0->next; + } + e0->next->prev = prev_edge; + e1->next->prev = e1; + e0->next = next_edge; + } /* End of IT loop*/ + + // Prepare for next scanbeam + for (edge = aet; edge; edge = next_edge) { + next_edge = edge->next; + succ_edge = edge->succ; + if ((edge->top.y == yt) && succ_edge) { + /* Replace AET edge by its successor */ + succ_edge->outp[BELOW] = edge->outp[ABOVE]; + succ_edge->bstate[BELOW] = edge->bstate[ABOVE]; + succ_edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP]; + succ_edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ]; + prev_edge = edge->prev; + if (prev_edge) { + prev_edge->next = succ_edge; + } else { + aet = succ_edge; + } + if (next_edge) { + next_edge->prev = succ_edge; + } + succ_edge->prev = prev_edge; + succ_edge->next = next_edge; + } else { + /* Update this edge */ + edge->outp[BELOW] = edge->outp[ABOVE]; + edge->bstate[BELOW] = edge->bstate[ABOVE]; + edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP]; + edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ]; + edge->xb = edge->xt; + } + edge->outp[ABOVE] = NULL; + } + } + } /* === END OF SCANBEAM PROCESSING ================================== */ + // Generate result polygon from out_poly + result->contour = NULL; + result->hole = NULL; + result->num_contours = count_contours(out_poly); + if (result->num_contours > 0) { + gpc_malloc(result->hole, result->num_contours * sizeof(int), + const_cast("hole flag table creation")); + gpc_malloc(result->contour, + result->num_contours * sizeof(gpc_vertex_list), + const_cast("contour creation")); + + c = 0; + for (poly = out_poly; poly; poly = npoly) { + npoly = poly->next; + if (poly->active) { + result->hole[c] = poly->proxy->hole; + result->contour[c].num_vertices = poly->active; + gpc_malloc( + result->contour[c].vertex, + result->contour[c].num_vertices * sizeof(gpc_vertex), + const_cast("vertex creation")); + + v = result->contour[c].num_vertices - 1; + for (vtx = poly->proxy->v[LEFT]; vtx; vtx = nv) { + nv = vtx->next; + result->contour[c].vertex[v].x = vtx->x; + result->contour[c].vertex[v].y = vtx->y; + gpc_free(vtx); + v--; + } + c++; + } + gpc_free(poly); + } + } else { + for (poly = out_poly; poly; poly = npoly) { + npoly = poly->next; + gpc_free(poly); + } + } + + // Tidy up + reset_it(&it); + reset_lmt(&lmt); + gpc_free(c_heap); + gpc_free(s_heap); + gpc_free(sbt); +} // NOLINT + +void gpc_free_tristrip(gpc_tristrip *t) { + int s = 0; + for (s = 0; s < t->num_strips; s++) { + gpc_free(t->strip[s].vertex); + } + gpc_free(t->strip); + t->num_strips = 0; +} + +void gpc_polygon_to_tristrip(gpc_polygon *s, gpc_tristrip *t) { + gpc_polygon c; + c.num_contours = 0; + c.hole = NULL; + c.contour = NULL; + gpc_tristrip_clip(GPC_DIFF, s, &c, t); +} + +// gpc_tristrip_clip +void gpc_tristrip_clip(gpc_op op, gpc_polygon *subj, gpc_polygon *clip, + gpc_tristrip *result) { + sb_tree *sbtree = NULL; + it_node *it = NULL; + it_node *intersect = NULL; + edge_node *edge = NULL; + edge_node *prev_edge = NULL; + edge_node *next_edge = NULL; + edge_node *succ_edge = NULL; + edge_node *e0 = NULL; + edge_node *e1 = NULL; + edge_node *aet = NULL; + edge_node *c_heap = NULL; + edge_node *s_heap = NULL; + edge_node *cf = NULL; + lmt_node *lmt = NULL; + lmt_node *local_min = NULL; + polygon_node *tlist = NULL; + polygon_node *tn = NULL; + polygon_node *tnn = NULL; + polygon_node *p = NULL; + polygon_node *q = NULL; + vertex_node *lt = NULL; + vertex_node *ltn = NULL; + vertex_node *rt = NULL; + vertex_node *rtn = NULL; + h_state horiz[2]; + vertex_type cft = NUL; + int in[2]; + int exists[2]; + int parity[2] = {LEFT, LEFT}; + int s = 0; + int v = 0; + int contributing = 0; + int search = 0; + int scanbeam = 0; + int sbt_entries = 0; + int vclass = 0; + int bl = 0; + int br = 0; + int tl = 0; + int tr = 0; + double *sbt = NULL; + double xb = 0.0; + double px = 0.0; + double nx = 0.0; + double yb = 0.0; + double yt = 0.0; + double dy = 0.0; + double ix = 0.0; + double iy = 0.0; + + /* Test for trivial NULL result cases */ + if (((subj->num_contours == 0) && (clip->num_contours == 0)) || + ((subj->num_contours == 0) && ((op == GPC_INT) || (op == GPC_DIFF))) || + ((clip->num_contours == 0) && (op == GPC_INT))) { + result->num_strips = 0; + result->strip = NULL; + return; + } + + /* Identify potentialy contributing contours */ + if (((op == GPC_INT) || (op == GPC_DIFF)) && (subj->num_contours > 0) && + (clip->num_contours > 0)) { + minimax_test(subj, clip, op); + } + /* Build LMT */ + if (subj->num_contours > 0) { + s_heap = build_lmt(&lmt, &sbtree, &sbt_entries, subj, SUBJ, op); + } + if (clip->num_contours > 0) { + c_heap = build_lmt(&lmt, &sbtree, &sbt_entries, clip, CLIP, op); + } + /* Return a NULL result if no contours contribute */ + if (lmt == NULL) { + result->num_strips = 0; + result->strip = NULL; + reset_lmt(&lmt); + gpc_free(s_heap); + gpc_free(c_heap); + return; + } + + /* Build scanbeam table from scanbeam tree */ + gpc_malloc(sbt, sbt_entries * sizeof(double), + const_cast("sbt creation")); + build_sbt(&scanbeam, sbt, sbtree); + scanbeam = 0; + free_sbtree(&sbtree); + + /* Invert clip polygon for difference operation */ + if (op == GPC_DIFF) { + parity[CLIP] = RIGHT; + } + local_min = lmt; + + // Process each scanbeam + while (scanbeam < sbt_entries) { + /* Set yb and yt to the bottom and top of the scanbeam */ + yb = sbt[scanbeam++]; + if (scanbeam < sbt_entries) { + yt = sbt[scanbeam]; + dy = yt - yb; + } + + /* === SCANBEAM BOUNDARY PROCESSING ================================ */ + /* If LMT node corresponding to yb exists */ + if (local_min) { + if (local_min->y == yb) { + /* Add edges starting at this local minimum to the AET */ + for (edge = local_min->first_bound; edge; edge = edge->next_bound) { + add_edge_to_aet(&aet, edge, NULL); + } + local_min = local_min->next; + } + } + /* Set dummy previous x value */ + /* Create bundles within AET */ + px = -DBL_MAX; + e0 = aet; + e1 = aet; + + /* Set up bundle fields of first edge */ + aet->bundle[ABOVE][aet->type] = (aet->top.y != yb); + aet->bundle[ABOVE][!aet->type] = 0; + aet->bstate[ABOVE] = UNBUNDLED; + + for (next_edge = aet->next; next_edge; next_edge = next_edge->next) { + /* Set up bundle fields of next edge */ + next_edge->bundle[ABOVE][next_edge->type] = (next_edge->top.y != yb); + next_edge->bundle[ABOVE][!next_edge->type] = 0; + next_edge->bstate[ABOVE] = UNBUNDLED; + + /* Bundle edges above the scanbeam boundary if they coincide */ + if (next_edge->bundle[ABOVE][next_edge->type]) { + if (gpc_eq(e0->xb, next_edge->xb) && gpc_eq(e0->dx, next_edge->dx) && + (e0->top.y != yb)) { + next_edge->bundle[ABOVE][next_edge->type] ^= + e0->bundle[ABOVE][next_edge->type]; + next_edge->bundle[ABOVE][!next_edge->type] = + e0->bundle[ABOVE][!next_edge->type]; + next_edge->bstate[ABOVE] = BUNDLE_HEAD; + e0->bundle[ABOVE][CLIP] = 0; + e0->bundle[ABOVE][SUBJ] = 0; + e0->bstate[ABOVE] = BUNDLE_TAIL; + } + e0 = next_edge; + } + } + horiz[CLIP] = NH; + horiz[SUBJ] = NH; + + /* Process each edge at this scanbeam boundary */ + for (edge = aet; edge; edge = edge->next) { + exists[CLIP] = + edge->bundle[ABOVE][CLIP] + (edge->bundle[BELOW][CLIP] << 1); + exists[SUBJ] = + edge->bundle[ABOVE][SUBJ] + (edge->bundle[BELOW][SUBJ] << 1); + + if (exists[CLIP] || exists[SUBJ]) { + /* Set bundle side */ + edge->bside[CLIP] = parity[CLIP]; + edge->bside[SUBJ] = parity[SUBJ]; + + /* Determine contributing status and quadrant occupancies */ + switch (op) { + case GPC_DIFF: + case GPC_INT: + contributing = (exists[CLIP] && (parity[SUBJ] || horiz[SUBJ])) || + (exists[SUBJ] && (parity[CLIP] || horiz[CLIP])) || + (exists[CLIP] && exists[SUBJ] && + (parity[CLIP] == parity[SUBJ])); + br = (parity[CLIP]) && (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) && + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) && + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) && + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + case GPC_XOR: + contributing = exists[CLIP] || exists[SUBJ]; + br = (parity[CLIP]) ^ (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) ^ + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) ^ + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) ^ + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + case GPC_UNION: + contributing = (exists[CLIP] && (!parity[SUBJ] || horiz[SUBJ])) || + (exists[SUBJ] && (!parity[CLIP] || horiz[CLIP])) || + (exists[CLIP] && exists[SUBJ] && + (parity[CLIP] == parity[SUBJ])); + br = (parity[CLIP]) || (parity[SUBJ]); + bl = (parity[CLIP] ^ edge->bundle[ABOVE][CLIP]) || + (parity[SUBJ] ^ edge->bundle[ABOVE][SUBJ]); + tr = (parity[CLIP] ^ (horiz[CLIP] != NH)) || + (parity[SUBJ] ^ (horiz[SUBJ] != NH)); + tl = (parity[CLIP] ^ (horiz[CLIP] != NH) ^ + edge->bundle[BELOW][CLIP]) || + (parity[SUBJ] ^ (horiz[SUBJ] != NH) ^ + edge->bundle[BELOW][SUBJ]); + break; + } + + // Update parity + parity[CLIP] ^= edge->bundle[ABOVE][CLIP]; + parity[SUBJ] ^= edge->bundle[ABOVE][SUBJ]; + + /* Update horizontal state */ + if (exists[CLIP]) { + horiz[CLIP] = next_h_state[horiz[CLIP]] + [((exists[CLIP] - 1) << 1) + parity[CLIP]]; + } + if (exists[SUBJ]) { + horiz[SUBJ] = next_h_state[horiz[SUBJ]] + [((exists[SUBJ] - 1) << 1) + parity[SUBJ]]; + } + vclass = tr + (tl << 1) + (br << 2) + (bl << 3); + + if (contributing) { + xb = edge->xb; + switch (vclass) { + case EMN: + new_tristrip(&tlist, edge, xb, yb); + cf = edge; + break; + case ERI: + edge->outp[ABOVE] = cf->outp[ABOVE]; + if (xb != cf->xb) { + gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb); + } + cf = NULL; + break; + case ELI: + gpc_vertex_create(edge, BELOW, LEFT, xb, yb); + edge->outp[ABOVE] = NULL; + cf = edge; + break; + case EMX: + if (xb != cf->xb) { + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + } + edge->outp[ABOVE] = NULL; + cf = NULL; + break; + case IMN: + if (cft == LED) { + if (cf->bot.y != yb) { + gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb); + } + new_tristrip(&tlist, cf, cf->xb, yb); + } + edge->outp[ABOVE] = cf->outp[ABOVE]; + gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb); + break; + case ILI: + new_tristrip(&tlist, edge, xb, yb); + cf = edge; + cft = ILI; + break; + case IRI: + if (cft == LED) { + if (cf->bot.y != yb) { + gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb); + } + new_tristrip(&tlist, cf, cf->xb, yb); + } + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + edge->outp[ABOVE] = NULL; + break; + case IMX: + gpc_vertex_create(edge, BELOW, LEFT, xb, yb); + edge->outp[ABOVE] = NULL; + cft = IMX; + break; + case IMM: + gpc_vertex_create(edge, BELOW, LEFT, xb, yb); + edge->outp[ABOVE] = cf->outp[ABOVE]; + if (xb != cf->xb) { + gpc_vertex_create(cf, ABOVE, RIGHT, xb, yb); + } + cf = edge; + break; + case EMM: + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + edge->outp[ABOVE] = NULL; + new_tristrip(&tlist, edge, xb, yb); + cf = edge; + break; + case LED: + if (edge->bot.y == yb) { + gpc_vertex_create(edge, BELOW, LEFT, xb, yb); + } + edge->outp[ABOVE] = edge->outp[BELOW]; + cf = edge; + cft = LED; + break; + case RED: + edge->outp[ABOVE] = cf->outp[ABOVE]; + if (cft == LED) { + if (cf->bot.y == yb) { + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + } else { + if (edge->bot.y == yb) { + gpc_vertex_create(cf, BELOW, LEFT, cf->xb, yb); + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + } + } + } else { + gpc_vertex_create(edge, BELOW, RIGHT, xb, yb); + gpc_vertex_create(edge, ABOVE, RIGHT, xb, yb); + } + cf = NULL; + break; + default: + break; + } /* End of switch */ + } /* End of contributing conditional */ + } /* End of edge exists conditional */ + } // End of AET loop + + /* Delete terminating edges from the AET, otherwise compute xt */ + for (edge = aet; edge; edge = edge->next) { + if (edge->top.y == yb) { + prev_edge = edge->prev; + next_edge = edge->next; + if (prev_edge) { + prev_edge->next = next_edge; + } else { + aet = next_edge; + } + if (next_edge) { + next_edge->prev = prev_edge; + } + + /* Copy bundle head state to the adjacent tail edge if required */ + if ((edge->bstate[BELOW] == BUNDLE_HEAD) && prev_edge) { + if (prev_edge->bstate[BELOW] == BUNDLE_TAIL) { + prev_edge->outp[BELOW] = edge->outp[BELOW]; + prev_edge->bstate[BELOW] = UNBUNDLED; + if (prev_edge->prev) { + if (prev_edge->prev->bstate[BELOW] == BUNDLE_TAIL) { + prev_edge->bstate[BELOW] = BUNDLE_HEAD; + } + } + } + } + } else { + if (edge->top.y == yt) { + edge->xt = edge->top.x; + } else { + edge->xt = edge->bot.x + edge->dx * (yt - edge->bot.y); + } + } + } + + if (scanbeam < sbt_entries) { + /* === SCANBEAM INTERIOR PROCESSING ============================== */ + build_intersection_table(&it, aet, dy); + /* Process each node in the intersection table */ + for (intersect = it; intersect; intersect = intersect->next) { + e0 = intersect->ie[0]; + e1 = intersect->ie[1]; + + /* Only generate output for contributing intersections */ + if ((e0->bundle[ABOVE][CLIP] || e0->bundle[ABOVE][SUBJ]) && + (e1->bundle[ABOVE][CLIP] || e1->bundle[ABOVE][SUBJ])) { + p = e0->outp[ABOVE]; + q = e1->outp[ABOVE]; + ix = intersect->point.x; + iy = intersect->point.y + yb; + + in[CLIP] = (e0->bundle[ABOVE][CLIP] && !e0->bside[CLIP]) || + (e1->bundle[ABOVE][CLIP] && e1->bside[CLIP]) || + (!e0->bundle[ABOVE][CLIP] && !e1->bundle[ABOVE][CLIP] && + e0->bside[CLIP] && e1->bside[CLIP]); + in[SUBJ] = (e0->bundle[ABOVE][SUBJ] && !e0->bside[SUBJ]) || + (e1->bundle[ABOVE][SUBJ] && e1->bside[SUBJ]) || + (!e0->bundle[ABOVE][SUBJ] && !e1->bundle[ABOVE][SUBJ] && + e0->bside[SUBJ] && e1->bside[SUBJ]); + + switch (op) { // Determine quadrant occupancies + case GPC_DIFF: + case GPC_INT: + tr = (in[CLIP]) && (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) && + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + case GPC_XOR: + tr = (in[CLIP]) ^ (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) ^ + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + case GPC_UNION: + tr = (in[CLIP]) || (in[SUBJ]); + tl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ]); + br = (in[CLIP] ^ e0->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e0->bundle[ABOVE][SUBJ]); + bl = (in[CLIP] ^ e1->bundle[ABOVE][CLIP] ^ + e0->bundle[ABOVE][CLIP]) || + (in[SUBJ] ^ e1->bundle[ABOVE][SUBJ] ^ + e0->bundle[ABOVE][SUBJ]); + break; + } + + vclass = tr + (tl << 1) + (br << 2) + (bl << 3); + switch (vclass) { + case EMN: + new_tristrip(&tlist, e1, ix, iy); + e0->outp[ABOVE] = e1->outp[ABOVE]; + break; + case ERI: + if (p) { + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + e0->outp[ABOVE] = NULL; + } + break; + case ELI: + if (q) { + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(e1, ABOVE, LEFT, ix, iy); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + e0->outp[ABOVE] = e1->outp[ABOVE]; + e1->outp[ABOVE] = NULL; + } + break; + case EMX: + if (p && q) { + gpc_vertex_create(e0, ABOVE, LEFT, ix, iy); + e0->outp[ABOVE] = NULL; + e1->outp[ABOVE] = NULL; + } + break; + case IMN: + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + new_tristrip(&tlist, prev_edge, px, iy); + e1->outp[ABOVE] = prev_edge->outp[ABOVE]; + gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy); + new_tristrip(&tlist, e0, ix, iy); + next_edge->outp[ABOVE] = e0->outp[ABOVE]; + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + break; + case ILI: + if (p) { + gpc_vertex_create(e0, ABOVE, LEFT, ix, iy); + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + e1->outp[ABOVE] = e0->outp[ABOVE]; + e0->outp[ABOVE] = NULL; + } + break; + case IRI: + if (q) { + gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy); + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + e0->outp[ABOVE] = e1->outp[ABOVE]; + e1->outp[ABOVE] = NULL; + } + break; + case IMX: + if (p && q) { + gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy); + gpc_vertex_create(e1, ABOVE, LEFT, ix, iy); + e0->outp[ABOVE] = NULL; + e1->outp[ABOVE] = NULL; + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + new_tristrip(&tlist, prev_edge, px, iy); + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + next_edge->outp[ABOVE] = prev_edge->outp[ABOVE]; + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + } + break; + case IMM: + if (p && q) { + gpc_vertex_create(e0, ABOVE, RIGHT, ix, iy); + gpc_vertex_create(e1, ABOVE, LEFT, ix, iy); + gpc_p_edge(prev_edge, e0, ABOVE); + gpc_vertex_create(prev_edge, ABOVE, LEFT, px, iy); + new_tristrip(&tlist, prev_edge, px, iy); + gpc_n_edge(next_edge, e1, ABOVE); + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + e1->outp[ABOVE] = prev_edge->outp[ABOVE]; + gpc_vertex_create(e1, ABOVE, RIGHT, ix, iy); + new_tristrip(&tlist, e0, ix, iy); + next_edge->outp[ABOVE] = e0->outp[ABOVE]; + gpc_vertex_create(next_edge, ABOVE, RIGHT, nx, iy); + } + break; + case EMM: + if (p && q) { + gpc_vertex_create(e0, ABOVE, LEFT, ix, iy); + new_tristrip(&tlist, e1, ix, iy); + e0->outp[ABOVE] = e1->outp[ABOVE]; + } + break; + default: + break; + } /* End of switch */ + } /* End of contributing intersection conditional */ + + // Swap bundle sides in response to edge crossing + if (e0->bundle[ABOVE][CLIP]) { + e1->bside[CLIP] = !e1->bside[CLIP]; + } + if (e1->bundle[ABOVE][CLIP]) { + e0->bside[CLIP] = !e0->bside[CLIP]; + } + if (e0->bundle[ABOVE][SUBJ]) { + e1->bside[SUBJ] = !e1->bside[SUBJ]; + } + if (e1->bundle[ABOVE][SUBJ]) { + e0->bside[SUBJ] = !e0->bside[SUBJ]; + } + + /* Swap e0 and e1 bundles in the AET */ + prev_edge = e0->prev; + next_edge = e1->next; + if (e1->next) { + e1->next->prev = e0; + } + + if (e0->bstate[ABOVE] == BUNDLE_HEAD) { + search = 1; + while (search) { + prev_edge = prev_edge->prev; + if (prev_edge) { + if (prev_edge->bundle[ABOVE][CLIP] || + prev_edge->bundle[ABOVE][SUBJ] || + (prev_edge->bstate[ABOVE] == BUNDLE_HEAD)) { + search = 0; + } + } else { + search = 0; + } + } + } + if (!prev_edge) { + e1->next = aet; + aet = e0->next; + } else { + e1->next = prev_edge->next; + prev_edge->next = e0->next; + } + e0->next->prev = prev_edge; + e1->next->prev = e1; + e0->next = next_edge; + } /* End of IT loop*/ + + /* Prepare for next scanbeam */ + for (edge = aet; edge; edge = next_edge) { + next_edge = edge->next; + succ_edge = edge->succ; + + if ((edge->top.y == yt) && succ_edge) { + /* Replace AET edge by its successor */ + succ_edge->outp[BELOW] = edge->outp[ABOVE]; + succ_edge->bstate[BELOW] = edge->bstate[ABOVE]; + succ_edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP]; + succ_edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ]; + prev_edge = edge->prev; + if (prev_edge) { + prev_edge->next = succ_edge; + } else { + aet = succ_edge; + } + if (next_edge) { + next_edge->prev = succ_edge; + } + succ_edge->prev = prev_edge; + succ_edge->next = next_edge; + } else { + /* Update this edge */ + edge->outp[BELOW] = edge->outp[ABOVE]; + edge->bstate[BELOW] = edge->bstate[ABOVE]; + edge->bundle[BELOW][CLIP] = edge->bundle[ABOVE][CLIP]; + edge->bundle[BELOW][SUBJ] = edge->bundle[ABOVE][SUBJ]; + edge->xb = edge->xt; + } + edge->outp[ABOVE] = NULL; + } + } + } /* === END OF SCANBEAM PROCESSING ================================== */ + + // Generate result tristrip from tlist + result->strip = NULL; + result->num_strips = count_tristrips(tlist); + if (result->num_strips > 0) { + gpc_malloc(result->strip, + result->num_strips * sizeof(gpc_vertex_list), + const_cast("tristrip list creation")); + + s = 0; + for (tn = tlist; tn; tn = tnn) { + tnn = tn->next; + if (tn->active > 2) { + /* Valid tristrip: copy the vertices and free the heap */ + result->strip[s].num_vertices = tn->active; + gpc_malloc(result->strip[s].vertex, + tn->active * sizeof(gpc_vertex), + const_cast("tristrip creation")); + v = 0; + if (0) { + lt = tn->v[RIGHT]; + rt = tn->v[LEFT]; + } else { + lt = tn->v[LEFT]; + rt = tn->v[RIGHT]; + } + while (lt || rt) { + if (lt) { + ltn = lt->next; + result->strip[s].vertex[v].x = lt->x; + result->strip[s].vertex[v].y = lt->y; + v++; + gpc_free(lt); + lt = ltn; + } + if (rt) { + rtn = rt->next; + result->strip[s].vertex[v].x = rt->x; + result->strip[s].vertex[v].y = rt->y; + v++; + gpc_free(rt); + rt = rtn; + } + } + s++; + } else { + /* Invalid tristrip: just free the heap */ + for (lt = tn->v[LEFT]; lt; lt = ltn) { + ltn = lt->next; + gpc_free(lt); + } + for (rt = tn->v[RIGHT]; rt; rt = rtn) { + rtn = rt->next; + gpc_free(rt); + } + } + gpc_free(tn); + } + } + // Tidy up + reset_it(&it); + reset_lmt(&lmt); + gpc_free(c_heap); + gpc_free(s_heap); + gpc_free(sbt); +} // NOLINT + +} // namespace gpc + +/* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */ diff --git a/paddle/fluid/operators/detection/gpc.h b/paddle/fluid/operators/detection/gpc.h new file mode 100644 index 0000000000000000000000000000000000000000..ee86262ef2c486e4eaeeeaf56c2392d2a1c5851b --- /dev/null +++ b/paddle/fluid/operators/detection/gpc.h @@ -0,0 +1,246 @@ +// 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. + +/*************************************************************************** + * + * Copyright (c) 2015 Baidu.com, Inc. All Rights Reserved + * + **************************************************************************/ + +/** + * @file include/gpc.h + * @author huhan02(com@baidu.com) + * @date 2015/12/18 13:52:10 + * @brief + * + * @modified by sunyipeng + * @email sunyipeng@baidu.com + * @date 2018/6/12 + **/ + +#ifndef PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ // GPC_H_ +#define PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ // GPC_H_ + +#include +#include +#include +#include + +namespace gpc { + +typedef enum { // Set operation type + GPC_DIFF, // Difference + GPC_INT, // Intersection + GPC_XOR, // Exclusive or + GPC_UNION // Union +} gpc_op; + +typedef struct { // Polygon vertex structure + double x; // Vertex x component + double y; // vertex y component +} gpc_vertex; + +typedef struct { // Vertex list structure + int num_vertices; // Number of vertices in list + gpc_vertex *vertex; // Vertex array pointer +} gpc_vertex_list; + +typedef struct { // Polygon set structure + int num_contours; // Number of contours in polygon + int *hole; // Hole external contour flags + gpc_vertex_list *contour; // Contour array pointer +} gpc_polygon; + +typedef struct { // Tristrip set structure + int num_strips; // Number of tristrips + gpc_vertex_list *strip; // Tristrip array pointer +} gpc_tristrip; + +typedef enum { LEFT, RIGHT } gpc_left_right; + +typedef enum { ABOVE, BELOW } gpc_above_below; + +typedef enum { CLIP, SUBJ } gpc_clip_subj; + +typedef enum { /* Edge intersection classes */ + NUL, /* Empty non-intersection */ + EMX, /* External maximum */ + ELI, /* External left intermediate */ + TED, /* Top edge */ + ERI, /* External right intermediate */ + RED, /* Right edge */ + IMM, /* Internal maximum and minimum */ + IMN, /* Internal minimum */ + EMN, /* External minimum */ + EMM, /* External maximum and minimum */ + LED, /* Left edge */ + ILI, /* Internal left intermediate */ + BED, /* Bottom edge */ + IRI, /* Internal right intermediate */ + IMX, /* Internal maximum */ + FUL /* Full non-intersection */ +} vertex_type; + +typedef enum { /* Horizontal edge states */ + NH, /* No horizontal edge */ + BH, /* Bottom horizontal edge */ + TH /* Top horizontal edge */ +} h_state; + +typedef enum { /* Edge bundle state */ + UNBUNDLED, /* Isolated edge not within a bundle */ + BUNDLE_HEAD, /* Bundle head node */ + BUNDLE_TAIL /* Passive bundle tail node */ +} bundle_state; + +typedef struct v_shape { /* Internal vertex list datatype */ + double x; /* X coordinate component */ + double y; /* Y coordinate component */ + struct v_shape *next; /* Pointer to next vertex in list */ +} vertex_node; + +typedef struct p_shape { /* Internal contour / tristrip type */ + int active; /* Active flag / vertex count */ + int hole; /* Hole / external contour flag */ + vertex_node *v[2]; /* Left and right vertex list ptrs */ + struct p_shape *next; /* Pointer to next polygon contour */ + struct p_shape *proxy; /* Pointer to actual structure used */ +} polygon_node; + +typedef struct edge_shape { + gpc_vertex vertex; /* Piggy-backed contour vertex data */ + gpc_vertex bot; /* Edge lower (x, y) coordinate */ + gpc_vertex top; /* Edge upper (x, y) coordinate */ + double xb; /* Scanbeam bottom x coordinate */ + double xt; /* Scanbeam top x coordinate */ + double dx; /* Change in x for a unit y increase */ + int type; /* Clip / subject edge flag */ + int bundle[2][2]; /* Bundle edge flags */ + int bside[2]; /* Bundle left / right indicators */ + bundle_state bstate[2]; /* Edge bundle state */ + polygon_node *outp[2]; /* Output polygon / tristrip pointer */ + struct edge_shape *prev; /* Previous edge in the AET */ + struct edge_shape *next; /* Next edge in the AET */ + struct edge_shape *pred; /* Edge connected at the lower end */ + struct edge_shape *succ; /* Edge connected at the upper end */ + struct edge_shape *next_bound; /* Pointer to next bound in LMT */ +} edge_node; + +inline bool gpc_eq(float a, float b) { return (fabs(a - b) <= 1e-6); } + +inline bool gpc_prev_index(float a, float b) { return (fabs(a - b) <= 1e-6); } + +inline int gpc_prev_index(int i, int n) { return ((i - 1 + n) % n); } + +inline int gpc_next_index(int i, int n) { return ((i + 1) % n); } + +inline int gpc_optimal(gpc_vertex *v, int i, int n) { + return (v[(i + 1) % n].y != v[i].y || v[(i - 1 + n) % n].y != v[i].y); +} + +inline int gpc_fwd_min(edge_node *v, int i, int n) { + return (v[(i + 1) % n].vertex.y > v[i].vertex.y && + v[(i - 1 + n) % n].vertex.y >= v[i].vertex.y); +} + +inline int gpc_not_fmax(edge_node *v, int i, int n) { + return (v[(i + 1) % n].vertex.y > v[i].vertex.y); +} + +inline int gpc_rev_min(edge_node *v, int i, int n) { + return (v[(i + 1) % n].vertex.y >= v[i].vertex.y && + v[(i - 1 + n) % n].vertex.y > v[i].vertex.y); +} + +inline int gpc_not_rmax(edge_node *v, int i, int n) { + return (v[(i - 1 + n) % n].vertex.y > v[i].vertex.y); +} + +// inline void gpc_p_edge(edge_node *d, edge_node *e, int p, double i, double j) +// { +inline void gpc_p_edge(edge_node *d, edge_node *e, int p) { + d = e; + do { + d = d->prev; + } while (!d->outp[p]); + // i = d->bot.x + d->dx * (j - d->bot.y); +} + +// inline void gpc_n_edge(edge_node *d, edge_node *e, int p, double i, double j) +// { +inline void gpc_n_edge(edge_node *d, edge_node *e, int p) { + d = e; + do { + d = d->next; + } while (!d->outp[p]); + // i = d->bot.x + d->dx * (j - d->bot.y); +} + +template +void gpc_malloc(T *&p, int b, char *s) { + if (b > 0) { + p = (T *)malloc(b); + + if (!p) { + fprintf(stderr, "gpc malloc failure: %s\n", s); + exit(0); + } + } else { + p = NULL; + } +} +template +void gpc_free(T *&p) { + if (p) { + free(p); + p = NULL; + } +} + +/* +=========================================================================== + Public Function Prototypes +=========================================================================== +*/ + +void add_vertex(vertex_node **t, double x, double y); + +void gpc_vertex_create(edge_node *e, int p, int s, double x, double y); + +/* +void gpc_read_polygon(FILE *infile_ptr, int read_hole_flags, + gpc_polygon *polygon); + +void gpc_write_polygon(FILE *outfile_ptr, int write_hole_flags, + gpc_polygon *polygon); +*/ +void gpc_add_contour(gpc_polygon *polygon, gpc_vertex_list *contour, int hole); + +void gpc_polygon_clip(gpc_op set_operation, gpc_polygon *subject_polygon, + gpc_polygon *clip_polygon, gpc_polygon *result_polygon); + +void gpc_tristrip_clip(gpc_op set_operation, gpc_polygon *subject_polygon, + gpc_polygon *clip_polygon, + gpc_tristrip *result_tristrip); + +void gpc_polygon_to_tristrip(gpc_polygon *polygon, gpc_tristrip *tristrip); + +void gpc_free_polygon(gpc_polygon *polygon); + +void gpc_free_tristrip(gpc_tristrip *tristrip); + +} // namespace gpc + +#endif // PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ +/* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */ diff --git a/paddle/fluid/operators/detection/multiclass_nms_op.cc b/paddle/fluid/operators/detection/multiclass_nms_op.cc index 60b93efdce810f8552374449fe5a6fc79b1a92c1..9e78b28a6011bb7bd299ca3438eb407f600d7000 100644 --- a/paddle/fluid/operators/detection/multiclass_nms_op.cc +++ b/paddle/fluid/operators/detection/multiclass_nms_op.cc @@ -9,10 +9,11 @@ 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/framework/op_registry.h" +#include "paddle/fluid/operators/detection/poly_util.h" namespace paddle { namespace operators { @@ -20,9 +21,6 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -constexpr int64_t kOutputDim = 6; -constexpr int64_t kBBoxSize = 4; - class MultiClassNMSOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -42,10 +40,15 @@ class MultiClassNMSOp : public framework::OperatorWithKernel { "The rank of Input(BBoxes) must be 3."); PADDLE_ENFORCE_EQ(score_dims.size(), 3, "The rank of Input(Scores) must be 3."); - PADDLE_ENFORCE_EQ(box_dims[2], 4, - "The 2nd dimension of Input(BBoxes) must be 4, " - "represents the layout of coordinate " - "[xmin, ymin, xmax, ymax]"); + PADDLE_ENFORCE(box_dims[2] == 4 || box_dims[2] == 8 || box_dims[2] == 16 || + box_dims[2] == 24 || box_dims[2] == 32, + "The 2nd dimension of Input(BBoxes) must be 4 or 8, " + "represents the layout of coordinate " + "[xmin, ymin, xmax, ymax] or " + "4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or " + "8 points: [xi, yi] i= 1,2,...,8 or " + "12 points: [xi, yi] i= 1,2,...,12 or " + "16 points: [xi, yi] i= 1,2,...,16"); PADDLE_ENFORCE_EQ(box_dims[1], score_dims[2], "The 1st dimensiong of Input(BBoxes) must be equal to " "3rd dimension of Input(Scores), which represents the " @@ -53,7 +56,7 @@ class MultiClassNMSOp : public framework::OperatorWithKernel { // Here the box_dims[0] is not the real dimension of output. // It will be rewritten in the computing kernel. - ctx->SetOutputDim("Out", {box_dims[1], 6}); + ctx->SetOutputDim("Out", {box_dims[1], box_dims[2] + 2}); } protected: @@ -128,6 +131,21 @@ static inline T JaccardOverlap(const T* box1, const T* box2, } } +template +T PolyIoU(const T* box1, const T* box2, const size_t box_size, + const bool normalized) { + T bbox1_area = PolyArea(box1, box_size, normalized); + T bbox2_area = PolyArea(box2, box_size, normalized); + T inter_area = PolyOverlapArea(box1, box2, box_size, normalized); + if (bbox1_area == 0 || bbox2_area == 0 || inter_area == 0) { + // If coordinate values are is invalid + // if area size <= 0, return 0. + return T(0.); + } else { + return inter_area / (bbox1_area + bbox2_area - inter_area); + } +} + template class MultiClassNMSKernel : public framework::OpKernel { public: @@ -137,6 +155,8 @@ class MultiClassNMSKernel : public framework::OpKernel { // The total boxes for each instance. int64_t num_boxes = bbox.dims()[0]; // 4: [xmin ymin xmax ymax] + // 8: [x1 y1 x2 y2 x3 y3 x4 y4] + // 16, 24, or 32: [x1 y1 x2 y2 ... xn yn], n = 8, 12 or 16 int64_t box_size = bbox.dims()[1]; std::vector scores_data(num_boxes); @@ -154,8 +174,19 @@ class MultiClassNMSKernel : public framework::OpKernel { for (size_t k = 0; k < selected_indices->size(); ++k) { if (keep) { const int kept_idx = (*selected_indices)[k]; - T overlap = JaccardOverlap(bbox_data + idx * box_size, + T overlap = T(0.); + // 4: [xmin ymin xmax ymax] + if (box_size == 4) { + overlap = JaccardOverlap(bbox_data + idx * box_size, bbox_data + kept_idx * box_size, true); + } + // 8: [x1 y1 x2 y2 x3 y3 x4 y4] or 16, 24, 32 + if (box_size == 8 || box_size == 16 || box_size == 24 || + box_size == 32) { + overlap = + PolyIoU(bbox_data + idx * box_size, + bbox_data + kept_idx * box_size, box_size, true); + } keep = overlap <= adaptive_threshold; } else { break; @@ -228,7 +259,9 @@ class MultiClassNMSKernel : public framework::OpKernel { void MultiClassOutput(const Tensor& scores, const Tensor& bboxes, const std::map>& selected_indices, Tensor* outs) const { - int predict_dim = scores.dims()[1]; + int64_t predict_dim = scores.dims()[1]; + int64_t box_size = bboxes.dims()[1]; + int64_t out_dim = bboxes.dims()[1] + 2; auto* scores_data = scores.data(); auto* bboxes_data = bboxes.data(); auto* odata = outs->data(); @@ -240,11 +273,11 @@ class MultiClassNMSKernel : public framework::OpKernel { const std::vector& indices = it.second; for (size_t j = 0; j < indices.size(); ++j) { int idx = indices[j]; - const T* bdata = bboxes_data + idx * kBBoxSize; - odata[count * kOutputDim] = label; // label - odata[count * kOutputDim + 1] = sdata[idx]; // score - // xmin, ymin, xmax, ymax - std::memcpy(odata + count * kOutputDim + 2, bdata, 4 * sizeof(T)); + const T* bdata = bboxes_data + idx * box_size; + odata[count * out_dim] = label; // label + odata[count * out_dim + 1] = sdata[idx]; // score + // xmin, ymin, xmax, ymax or multi-points coordinates + std::memcpy(odata + count * out_dim + 2, bdata, box_size * sizeof(T)); count++; } } @@ -261,6 +294,7 @@ class MultiClassNMSKernel : public framework::OpKernel { int64_t class_num = score_dims[1]; int64_t predict_dim = score_dims[2]; int64_t box_dim = boxes->dims()[2]; + int64_t out_dim = boxes->dims()[2] + 2; std::vector>> all_indices; std::vector batch_starts = {0}; @@ -283,7 +317,7 @@ class MultiClassNMSKernel : public framework::OpKernel { T* od = outs->mutable_data({1}, ctx.GetPlace()); od[0] = -1; } else { - outs->mutable_data({num_kept, kOutputDim}, ctx.GetPlace()); + outs->mutable_data({num_kept, out_dim}, ctx.GetPlace()); for (int64_t i = 0; i < batch_size; ++i) { Tensor ins_score = scores->Slice(i, i + 1); ins_score.Resize({class_num, predict_dim}); @@ -311,10 +345,11 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("BBoxes", - "(Tensor) A 3-D Tensor with shape [N, M, 4] represents the " + "(Tensor) A 3-D Tensor with shape " + "[N, M, 4 or 8 16 24 32] represents the " "predicted locations of M bounding bboxes, N is the batch size. " "Each bounding box has four coordinate values and the layout is " - "[xmin, ymin, xmax, ymax]."); + "[xmin, ymin, xmax, ymax], when box size equals to 4."); AddInput("Scores", "(Tensor) A 3-D Tensor with shape [N, C, M] represents the " "predicted confidence predictions. N is the batch size, C is the " @@ -351,8 +386,12 @@ class MultiClassNMSOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Out", "(LoDTensor) A 2-D LoDTensor with shape [No, 6] represents the " "detections. Each row has 6 values: " - "[label, confidence, xmin, ymin, xmax, ymax], No is the total " - "number of detections in this mini-batch. For each instance, " + "[label, confidence, xmin, ymin, xmax, ymax] or " + "(LoDTensor) A 2-D LoDTensor with shape [No, 10] represents the " + "detections. Each row has 10 values: " + "[label, confidence, x1, y1, x2, y2, x3, y3, x4, y4]. No is the " + "total number of detections in this mini-batch." + "For each instance, " "the offsets in first dimension are called LoD, the number of " "offset is N + 1, if LoD[i + 1] - LoD[i] == 0, means there is " "no detected bbox."); diff --git a/paddle/fluid/operators/detection/poly_util.cc b/paddle/fluid/operators/detection/poly_util.cc new file mode 100644 index 0000000000000000000000000000000000000000..1af2c95c6cf526d651b196b54614a21a9cddde8c --- /dev/null +++ b/paddle/fluid/operators/detection/poly_util.cc @@ -0,0 +1,132 @@ +/* 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. */ + +#ifndef POLY_UTIL_CC_ +#define POLY_UTIL_CC_ + +#include "paddle/fluid/operators/detection/poly_util.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using gpc::gpc_polygon_clip; +using gpc::gpc_free_polygon; + +template +void Array2PointVec(const T*& box, const size_t box_size, + std::vector>& vec) { + size_t pts_num = box_size / 2; + vec.resize(pts_num); + for (size_t i = 0; i < pts_num; i++) { + vec.at(i).x = box[2 * i]; + vec.at(i).y = box[2 * i + 1]; + } +} + +template +void Array2Poly(const T*& box, const size_t box_size, gpc::gpc_polygon& poly) { + size_t pts_num = box_size / 2; + poly.num_contours = 1; + poly.hole = (int*)malloc(sizeof(int)); + poly.hole[0] = 0; + poly.contour = (gpc::gpc_vertex_list*)malloc(sizeof(gpc::gpc_vertex_list)); + poly.contour->num_vertices = pts_num; + poly.contour->vertex = + (gpc::gpc_vertex*)malloc(sizeof(gpc::gpc_vertex) * pts_num); + for (size_t i = 0; i < pts_num; ++i) { + poly.contour->vertex[i].x = box[2 * i]; + poly.contour->vertex[i].y = box[2 * i + 1]; + } +} + +template +void PointVec2Poly(const std::vector>& vec, gpc::gpc_polygon& poly) { + int pts_num = vec.size(); + poly.num_contours = 1; + poly.hole = (int*)malloc(sizeof(int)); + poly.hole[0] = 0; + poly.contour = (gpc::gpc_vertex_list*)malloc(sizeof(gpc::gpc_vertex_list)); + poly.contour->num_vertices = pts_num; + poly.contour->vertex = + (gpc::gpc_vertex*)malloc(sizeof(gpc::gpc_vertex) * pts_num); + for (size_t i = 0; i < pts_num; ++i) { + poly.contour->vertex[i].x = vec[i].x; + poly.contour->vertex[i].y = vec[i].y; + } +} + +template +void Poly2PointVec(const gpc::gpc_vertex_list& contour, + std::vector>& vec) { + int pts_num = contour.num_vertices; + vec.resize(pts_num); + for (int i = 0; i < pts_num; i++) { + vec.at(i).x = contour.vertex[i].x; + vec.at(i).y = contour.vertex[i].y; + } +} + +template +T GetContourArea(std::vector>& vec) { + size_t pts_num = vec.size(); + if (pts_num < 3) return T(0.); + T area = T(0.); + for (size_t i = 0; i < pts_num; ++i) { + area += vec[i].x * vec[(i + 1) % pts_num].y - + vec[i].y * vec[(i + 1) % pts_num].x; + } + return std::fabs(area / 2.0); +} + +template +T PolyArea(const T* box, const size_t box_size, const bool normalized) { + // If coordinate values are is invalid + // if area size <= 0, return 0. + std::vector> vec; + Array2PointVec(box, box_size, vec); + return GetContourArea(vec); +} + +template +T PolyOverlapArea(const T* box1, const T* box2, const size_t box_size, + const bool normalized) { + gpc::gpc_polygon poly1; + gpc::gpc_polygon poly2; + Array2Poly(box1, box_size, poly1); + Array2Poly(box2, box_size, poly2); + gpc::gpc_polygon respoly; + gpc::gpc_op op = gpc::GPC_INT; + gpc::gpc_polygon_clip(op, &poly2, &poly1, &respoly); + + T inter_area = T(0.); + int contour_num = respoly.num_contours; + for (int i = 0; i < contour_num; ++i) { + std::vector> resvec; + Poly2PointVec(respoly.contour[i], resvec); + // inter_area += std::fabs(cv::contourArea(resvec)) + 0.5f * + // (cv::arcLength(resvec, true)); + inter_area += GetContourArea(resvec); + } + + gpc::gpc_free_polygon(&poly1); + gpc::gpc_free_polygon(&poly2); + gpc::gpc_free_polygon(&respoly); + return inter_area; +} + +} // namespace operators +} // namespace paddle + +#endif diff --git a/paddle/fluid/operators/detection/poly_util.h b/paddle/fluid/operators/detection/poly_util.h new file mode 100644 index 0000000000000000000000000000000000000000..f07baf72d9ff07b8fcb45dcfb2a35741fb1aeed0 --- /dev/null +++ b/paddle/fluid/operators/detection/poly_util.h @@ -0,0 +1,73 @@ +/* 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. */ + +#ifndef POLY_UTIL_H_ +#define POLY_UTIL_H_ + +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detection/gpc.h" + +namespace paddle { +namespace operators { + +template +class Point_ { + public: + // default constructor + Point_() {} + Point_(T _x, T _y) {} + Point_(const Point_& pt) {} + + Point_& operator=(const Point_& pt); + // conversion to another data type + // template operator Point_<_T>() const; + // conversion to the old-style C structures + // operator Vec() const; + + // checks whether the point is inside the specified rectangle + // bool inside(const Rect_& r) const; + T x; //!< x coordinate of the point + T y; //!< y coordinate of the point +}; + +template +void Array2PointVec(const T*& box, const size_t box_size, + std::vector>& vec); + +template +void Array2Poly(const T*& box, const size_t box_size, gpc::gpc_polygon& poly); + +template +void PointVec2Poly(const std::vector>& vec, gpc::gpc_polygon& poly); + +template +void Poly2PointVec(const gpc::gpc_vertex_list& contour, + std::vector>& vec); + +template +T GetContourArea(std::vector>& vec); + +template +T PolyArea(const T* box, const size_t box_size, const bool normalized); + +template +T PolyOverlapArea(const T* box1, const T* box2, const size_t box_size, + const bool normalized); +} // namespace operators +} // namespace paddle + +#include "paddle/fluid/operators/detection/poly_util.cc" + +#endif // POLY_UTIL_H_ diff --git a/paddle/fluid/operators/detection/polygon_box_transform_op.cc b/paddle/fluid/operators/detection/polygon_box_transform_op.cc index 568d50d457d838d5f11605710c0d3b987af01d10..4b3bc2edb58fe23393d906094c41b6ad62c71155 100644 --- a/paddle/fluid/operators/detection/polygon_box_transform_op.cc +++ b/paddle/fluid/operators/detection/polygon_box_transform_op.cc @@ -41,9 +41,9 @@ class PolygonBoxTransformCPUKernel : public framework::OpKernel { for (int id_w = 0; id_w < width; ++id_w) { id = id_n * height * width + width * id_h + id_w; if (id_n % 2 == 0) { - out_data[id] = id_w - in_data[id]; + out_data[id] = id_w * 4 - in_data[id]; } else { - out_data[id] = id_h - in_data[id]; + out_data[id] = id_h * 4 - in_data[id]; } } } diff --git a/paddle/fluid/operators/detection/polygon_box_transform_op.cu b/paddle/fluid/operators/detection/polygon_box_transform_op.cu index 6187ac6622c65d2bbc525c3fe2cb397cf74ac612..e1eaf084a3413dd1d13514e2d7b22572d21dd119 100644 --- a/paddle/fluid/operators/detection/polygon_box_transform_op.cu +++ b/paddle/fluid/operators/detection/polygon_box_transform_op.cu @@ -32,9 +32,9 @@ __global__ void PolygonBoxTransformKernel(const int n, const int h, const int w, if (id_n < n && id_h < h && id_w < w) { int id = id_n * h * w + w * id_h + id_w; if (id_n % 2 == 0) { - output[id] = id_w - input[id]; + output[id] = id_w * 4 - input[id]; } else { - output[id] = id_h - input[id]; + output[id] = id_h * 4 - input[id]; } } } diff --git a/paddle/fluid/operators/detection/rpn_target_assign_op.cc b/paddle/fluid/operators/detection/rpn_target_assign_op.cc index dda423efd35b96f5e1d7c55389818f46ef3d8694..46fff9d338b7759496faaf6dd9960d34887755ba 100644 --- a/paddle/fluid/operators/detection/rpn_target_assign_op.cc +++ b/paddle/fluid/operators/detection/rpn_target_assign_op.cc @@ -52,6 +52,9 @@ class RpnTargetAssignOp : public framework::OperatorWithKernel { PADDLE_ENFORCE( ctx->HasOutput("TargetBBox"), "Output(TargetBBox) of RpnTargetAssignOp should not be null"); + PADDLE_ENFORCE( + ctx->HasOutput("BBoxInsideWeight"), + "Output(BBoxInsideWeight) of RpnTargetAssignOp should not be null"); auto anchor_dims = ctx->GetInputDim("Anchor"); auto gt_boxes_dims = ctx->GetInputDim("GtBoxes"); @@ -68,6 +71,7 @@ class RpnTargetAssignOp : public framework::OperatorWithKernel { ctx->SetOutputDim("ScoreIndex", {-1}); ctx->SetOutputDim("TargetLabel", {-1, 1}); ctx->SetOutputDim("TargetBBox", {-1, 4}); + ctx->SetOutputDim("BBoxInsideWeight", {-1, 4}); } protected: @@ -169,6 +173,7 @@ void ScoreAssign(const T* anchor_by_gt_overlap_data, const float rpn_positive_overlap, const float rpn_negative_overlap, std::vector* fg_inds, std::vector* bg_inds, std::vector* tgt_lbl, + std::vector* fg_fake, std::vector* bbox_inside_weight, std::minstd_rand engine, bool use_random) { float epsilon = 0.00001; int anchor_num = anchor_to_gt_max.dims()[0]; @@ -201,12 +206,12 @@ void ScoreAssign(const T* anchor_by_gt_overlap_data, // Reservoir Sampling int fg_num = static_cast(rpn_fg_fraction * rpn_batch_size_per_im); ReservoirSampling(fg_num, &fg_inds_fake, engine, use_random); - fg_num = static_cast(fg_inds_fake.size()); - for (int64_t i = 0; i < fg_num; ++i) { + int fg_fake_num = static_cast(fg_inds_fake.size()); + for (int64_t i = 0; i < fg_fake_num; ++i) { target_label[fg_inds_fake[i]] = 1; } - int bg_num = rpn_batch_size_per_im - fg_num; + int bg_num = rpn_batch_size_per_im - fg_fake_num; for (int64_t i = 0; i < anchor_num; ++i) { if (anchor_to_gt_max_data[i] < rpn_negative_overlap) { bg_inds_fake.push_back(i); @@ -214,12 +219,28 @@ void ScoreAssign(const T* anchor_by_gt_overlap_data, } ReservoirSampling(bg_num, &bg_inds_fake, engine, use_random); bg_num = static_cast(bg_inds_fake.size()); + int fake_num = 0; for (int64_t i = 0; i < bg_num; ++i) { + // fg fake found + if (target_label[bg_inds_fake[i]] == 1) { + fake_num++; + fg_fake->emplace_back(fg_inds_fake[0]); + for (int j = 0; j < 4; ++j) { + bbox_inside_weight->emplace_back(T(0.)); + } + } target_label[bg_inds_fake[i]] = 0; } + for (int64_t i = 0; i < (fg_fake_num - fake_num) * 4; ++i) { + bbox_inside_weight->emplace_back(T(1.)); + } + for (int64_t i = 0; i < anchor_num; ++i) { - if (target_label[i] == 1) fg_inds->emplace_back(i); + if (target_label[i] == 1) { + fg_inds->emplace_back(i); + fg_fake->emplace_back(i); + } if (target_label[i] == 0) bg_inds->emplace_back(i); } fg_num = fg_inds->size(); @@ -248,7 +269,8 @@ std::vector SampleRpnFgBgGt(const platform::CPUDeviceContext& ctx, std::vector bg_inds; std::vector gt_inds; std::vector tgt_lbl; - + std::vector fg_fake; + std::vector bbox_inside_weight; // Calculate the max IoU between anchors and gt boxes // Map from anchor to gt box that has highest overlap auto place = ctx.GetPlace(); @@ -275,32 +297,37 @@ std::vector SampleRpnFgBgGt(const platform::CPUDeviceContext& ctx, // Follow the Faster RCNN's implementation ScoreAssign(anchor_by_gt_overlap_data, anchor_to_gt_max, gt_to_anchor_max, rpn_batch_size_per_im, rpn_fg_fraction, rpn_positive_overlap, - rpn_negative_overlap, &fg_inds, &bg_inds, &tgt_lbl, engine, - use_random); + rpn_negative_overlap, &fg_inds, &bg_inds, &tgt_lbl, &fg_fake, + &bbox_inside_weight, engine, use_random); int fg_num = fg_inds.size(); int bg_num = bg_inds.size(); - gt_inds.reserve(fg_num); - for (int i = 0; i < fg_num; ++i) { - gt_inds.emplace_back(argmax[fg_inds[i]]); + int fg_fake_num = fg_fake.size(); + gt_inds.reserve(fg_fake_num); + for (int i = 0; i < fg_fake_num; ++i) { + gt_inds.emplace_back(argmax[fg_fake[i]]); } - - Tensor loc_index_t, score_index_t, tgt_lbl_t, gt_inds_t; - int* loc_index_data = loc_index_t.mutable_data({fg_num}, place); + Tensor loc_index_t, score_index_t, tgt_lbl_t, gt_inds_t, bbox_inside_weight_t; + int* loc_index_data = loc_index_t.mutable_data({fg_fake_num}, place); int* score_index_data = score_index_t.mutable_data({fg_num + bg_num}, place); int* tgt_lbl_data = tgt_lbl_t.mutable_data({fg_num + bg_num}, place); - int* gt_inds_data = gt_inds_t.mutable_data({fg_num}, place); - std::copy(fg_inds.begin(), fg_inds.end(), loc_index_data); + int* gt_inds_data = gt_inds_t.mutable_data({fg_fake_num}, place); + T* bbox_inside_weight_data = + bbox_inside_weight_t.mutable_data({fg_fake_num, 4}, place); + std::copy(fg_fake.begin(), fg_fake.end(), loc_index_data); std::copy(fg_inds.begin(), fg_inds.end(), score_index_data); std::copy(bg_inds.begin(), bg_inds.end(), score_index_data + fg_num); std::copy(tgt_lbl.begin(), tgt_lbl.end(), tgt_lbl_data); std::copy(gt_inds.begin(), gt_inds.end(), gt_inds_data); + std::copy(bbox_inside_weight.begin(), bbox_inside_weight.end(), + bbox_inside_weight_data); std::vector loc_score_tgtlbl_gt; loc_score_tgtlbl_gt.emplace_back(loc_index_t); loc_score_tgtlbl_gt.emplace_back(score_index_t); loc_score_tgtlbl_gt.emplace_back(tgt_lbl_t); loc_score_tgtlbl_gt.emplace_back(gt_inds_t); + loc_score_tgtlbl_gt.emplace_back(bbox_inside_weight_t); return loc_score_tgtlbl_gt; } @@ -318,6 +345,7 @@ class RpnTargetAssignKernel : public framework::OpKernel { auto* score_index = context.Output("ScoreIndex"); auto* tgt_bbox = context.Output("TargetBBox"); auto* tgt_lbl = context.Output("TargetLabel"); + auto* bbox_inside_weight = context.Output("BBoxInsideWeight"); PADDLE_ENFORCE_EQ(gt_boxes->lod().size(), 1UL, "RpnTargetAssignOp gt_boxes needs 1 level of LoD"); @@ -340,7 +368,7 @@ class RpnTargetAssignKernel : public framework::OpKernel { score_index->mutable_data({max_num}, place); tgt_bbox->mutable_data({max_num, 4}, place); tgt_lbl->mutable_data({max_num, 1}, place); - + bbox_inside_weight->mutable_data({max_num, 4}, place); auto& dev_ctx = context.device_context(); std::random_device rnd; @@ -394,6 +422,7 @@ class RpnTargetAssignKernel : public framework::OpKernel { Tensor sampled_score_index = loc_score_tgtlbl_gt[1]; Tensor sampled_tgtlbl = loc_score_tgtlbl_gt[2]; Tensor sampled_gt_index = loc_score_tgtlbl_gt[3]; + Tensor sampled_bbox_inside_weight = loc_score_tgtlbl_gt[4]; int loc_num = sampled_loc_index.dims()[0]; int score_num = sampled_score_index.dims()[0]; @@ -432,6 +461,8 @@ class RpnTargetAssignKernel : public framework::OpKernel { AppendRpns(score_index, total_score_num, &sampled_score_index_unmap); AppendRpns(tgt_bbox, total_loc_num * 4, &sampled_tgt_bbox); AppendRpns(tgt_lbl, total_score_num, &sampled_tgtlbl); + AppendRpns(bbox_inside_weight, total_loc_num * 4, + &sampled_bbox_inside_weight); total_loc_num += loc_num; total_score_num += score_num; @@ -448,10 +479,12 @@ class RpnTargetAssignKernel : public framework::OpKernel { score_index->set_lod(loc_score); tgt_bbox->set_lod(lod_loc); tgt_lbl->set_lod(loc_score); + bbox_inside_weight->set_lod(lod_loc); loc_index->Resize({total_loc_num}); score_index->Resize({total_score_num}); tgt_bbox->Resize({total_loc_num, 4}); tgt_lbl->Resize({total_score_num, 1}); + bbox_inside_weight->Resize({total_loc_num, 4}); } }; @@ -514,6 +547,9 @@ class RpnTargetAssignOpMaker : public framework::OpProtoAndCheckerMaker { "TargetLabel", "(Tensor), The target labels of each anchor with shape " "[F + B, 1], F and B are sampled foreground and backgroud number."); + AddOutput("BBoxInsideWeight", + "(Tensor), The bbox inside weight with shape " + "[F, 4], F is the sampled foreground number."); AddComment(R"DOC( This operator can be, for a given set of ground truth bboxes and the anchors, to assign classification and regression targets to each prediction. diff --git a/paddle/fluid/operators/distributed/CMakeLists.txt b/paddle/fluid/operators/distributed/CMakeLists.txt index 56734b81e8716a0c0c37a11e35c9118ee7b55020..21db93958a4a586c74a1e060f1f04b5af1dcd889 100644 --- a/paddle/fluid/operators/distributed/CMakeLists.txt +++ b/paddle/fluid/operators/distributed/CMakeLists.txt @@ -20,7 +20,7 @@ if(WITH_GRPC) DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL) cc_test(rpc_server_test SRCS rpc_server_test.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_sparse_table_op SERIAL) - cc_test(varhandle_test SRCS varhandle_test.cc) + cc_test(varhandle_test SRCS varhandle_test.cc DEPS profiler) return() endif() diff --git a/paddle/fluid/operators/distributed/brpc_server.cc b/paddle/fluid/operators/distributed/brpc_server.cc index 862167f02084cfe81db1c0936bbfb0415fa85721..47a06dd0f378f6cc4f79aee52052717188d72420 100644 --- a/paddle/fluid/operators/distributed/brpc_server.cc +++ b/paddle/fluid/operators/distributed/brpc_server.cc @@ -133,10 +133,10 @@ void AsyncBRPCServer::StartServer() { void AsyncBRPCServer::ShutDownImpl() { server_.Stop(1000); } void AsyncBRPCServer::WaitServerReady() { - VLOG(3) << "AsyncGRPCServer is wait server ready"; + VLOG(30) << "AsyncGRPCServer is wait server ready"; std::unique_lock lock(this->mutex_ready_); condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); - VLOG(3) << "AsyncGRPCServer WaitSeverReady"; + VLOG(30) << "AsyncGRPCServer WaitSeverReady"; } }; // namespace distributed diff --git a/paddle/fluid/operators/distributed/grpc_client.cc b/paddle/fluid/operators/distributed/grpc_client.cc index 13682b78f0eccf049daa315f3a26aafd22e42a41..c28f86146d3040c6a26cabfb795eff67375d4b76 100644 --- a/paddle/fluid/operators/distributed/grpc_client.cc +++ b/paddle/fluid/operators/distributed/grpc_client.cc @@ -12,14 +12,12 @@ 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/distributed/grpc_client.h" - #include - #include #include "glog/logging.h" // For VLOG #include "paddle/fluid/framework/threadpool.h" +#include "paddle/fluid/operators/distributed/grpc_client.h" #include "paddle/fluid/operators/distributed/grpc_serde.h" #include "paddle/fluid/operators/distributed/request_handler.h" #include "paddle/fluid/platform/profiler.h" @@ -40,7 +38,7 @@ void GRPCClient::SendComplete() { std::unique_lock lk(completed_mutex_); if (!completed_) { for (auto& it : channels_) { - VLOG(3) << "send complete message to " << it.first; + VLOG(30) << "send complete message to " << it.first; this->AsyncSendComplete(it.first); } PADDLE_ENFORCE(this->Wait(), "internal grpc error"); @@ -73,24 +71,31 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, const framework::Scope* p_scope = &scope; const auto ch = GetChannel(ep_val); SendProcessor* s = new SendProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "Send", var_name_val, p_ctx, p_scope)); + const std::string method = "SendRPC"; + VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); - framework::AsyncIO([var_name_val, p_scope, p_ctx, s, this] { + framework::AsyncIO([var_name_val, p_scope, p_ctx, s, method, h, this] { auto* var = p_scope->FindVar(var_name_val); ::grpc::ByteBuffer req; - SerializeToByteBuffer(var_name_val, var, *p_ctx, &req); + SerializeToByteBuffer(var_name_val, var, *p_ctx, &req, "", trainer_id_); - VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; + VLOG(30) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = nullptr; + platform::RecordRPCEvent record_event(method, p_ctx); + auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } }); req_count_++; @@ -100,7 +105,10 @@ VarHandlePtr GRPCClient::AsyncSendVar(const std::string& ep, void ProcGetResponse(const VarHandle& var_h, const ::grpc::ByteBuffer& ret_msg) { framework::Variable* outvar = nullptr; - DeserializeFromByteBuffer(ret_msg, *var_h.ctx(), var_h.scope(), &outvar); + // get response's trainer_id is not used + int trainer_id; + DeserializeFromByteBuffer(ret_msg, *var_h.ctx(), var_h.scope(), &outvar, + &trainer_id); } template @@ -122,25 +130,33 @@ VarHandlePtr GRPCClient::AsyncGetVar(const std::string& ep, const framework::Scope* p_scope = &scope; const auto ch = GetChannel(ep_val); GetProcessor* s = new GetProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "Get", var_name_val, p_ctx, p_scope)); + const std::string method = "GetRPC"; + VarHandlePtr h(new VarHandle(ep, method, var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); - framework::AsyncIO([var_name_val, s, this] { + framework::AsyncIO([var_name_val, s, method, p_ctx, h, this] { // prepare input sendrecv::VariableMessage req; req.set_varname(var_name_val); + req.set_trainer_id(trainer_id_); ::grpc::ByteBuffer buf; RequestToByteBuffer(req, &buf); - VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; + VLOG(30) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = ProcGetResponse; + platform::RecordRPCEvent record_event(method, p_ctx); + auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } }); req_count_++; @@ -161,27 +177,35 @@ VarHandlePtr GRPCClient::AsyncPrefetchVar(const std::string& ep, const framework::Scope* p_scope = &scope; const auto ch = GetChannel(ep_val); GetProcessor* s = new GetProcessor(ch); - VarHandlePtr h( - new VarHandle(ep, "Prefetch", out_var_name_val, p_ctx, p_scope)); + + const std::string method = "PrefetchRPC"; + + VarHandlePtr h(new VarHandle(ep, method, out_var_name_val, p_ctx, p_scope)); s->Prepare(h, time_out); framework::AsyncIO([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, - s, this] { + s, method, h, this] { auto* var = p_scope->FindVar(in_var_name_val); ::grpc::ByteBuffer req; SerializeToByteBuffer(in_var_name_val, var, *p_ctx, &req, out_var_name_val); - VLOG(3) << s->GetVarHandlePtr()->String() << " begin"; + VLOG(30) << s->GetVarHandlePtr()->String() << " begin"; // stub context s->response_call_back_ = ProcGetResponse; + platform::RecordRPCEvent record_event(method, p_ctx); + auto call = s->stub_g_.PrepareUnaryCall( s->context_.get(), "/sendrecv.SendRecvService/PrefetchVariable", req, &cq_); call->StartCall(); call->Finish(&s->reply_, &s->status_, static_cast(s)); + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } }); req_count_++; @@ -193,15 +217,24 @@ VarHandlePtr GRPCClient::AsyncSendBatchBarrier(const std::string& ep, const auto ch = GetChannel(ep); BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "BatchBarrier", BATCH_BARRIER_MESSAGE, - nullptr, nullptr)); + const std::string method = "BatchBarrierRPC"; + VarHandlePtr h( + new VarHandle(ep, method, BATCH_BARRIER_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(BATCH_BARRIER_MESSAGE); + + platform::RecordRPCEvent record_event(method, nullptr); + auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + return h; } @@ -209,15 +242,24 @@ VarHandlePtr GRPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { const auto ch = GetChannel(ep); FetchBarrierProcessor* s = new FetchBarrierProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "FetchBarrier", FETCH_BARRIER_MESSAGE, - nullptr, nullptr)); + const std::string method = "FetchBarrierRPC"; + VarHandlePtr h( + new VarHandle(ep, method, FETCH_BARRIER_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(FETCH_BARRIER_MESSAGE); + + platform::RecordRPCEvent record_event(method, nullptr); + auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + return h; } @@ -226,15 +268,23 @@ VarHandlePtr GRPCClient::AsyncSendComplete(const std::string& ep, const auto ch = GetChannel(ep); BatchBarrierProcessor* s = new BatchBarrierProcessor(ch); - VarHandlePtr h( - new VarHandle(ep, "SendComplete", COMPLETE_MESSAGE, nullptr, nullptr)); + const std::string method = "SendCompleteRPC"; + VarHandlePtr h(new VarHandle(ep, method, COMPLETE_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(COMPLETE_MESSAGE); + + platform::RecordRPCEvent record_event(method, nullptr); + auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + return h; } @@ -244,17 +294,27 @@ VarHandlePtr GRPCClient::AsyncCheckpointNotify(const std::string& ep, const auto ch = GetChannel(ep); CheckpointNotifyProcessor* s = new CheckpointNotifyProcessor(ch); - VarHandlePtr h(new VarHandle(ep, "CheckPointNotify", CHECKPOINT_SAVE_MESSAGE, - nullptr, nullptr)); + + const std::string method = "CheckPointNotifyRPC"; + + VarHandlePtr h( + new VarHandle(ep, method, CHECKPOINT_SAVE_MESSAGE, nullptr, nullptr)); s->Prepare(h, time_out); sendrecv::VariableMessage req; req.set_varname(CHECKPOINT_SAVE_MESSAGE); req.set_out_varname(dir); + platform::RecordRPCEvent record_event(method, nullptr); + auto rpc = s->stub_->AsyncCheckpointNotify(s->context_.get(), req, &cq_); rpc->Finish(&s->reply_, &s->status_, reinterpret_cast(s)); req_count_++; + + if (UNLIKELY(platform::IsProfileEnabled())) { + h->Wait(); + } + return h; } @@ -268,17 +328,21 @@ void GRPCClient::Proceed() { void* tag = nullptr; bool ok = false; - VLOG(3) << "GRPCClient Proceed begin"; + VLOG(30) << "GRPCClient Proceed begin"; while (!stopped_ && cq_.Next(&tag, &ok)) { BaseProcessor* c = static_cast(tag); GPR_ASSERT(ok); PADDLE_ENFORCE(c); + if (c->status_.ok()) { - VLOG(3) << c->GetVarHandlePtr()->String() << " process"; + VLOG(30) << c->GetVarHandlePtr()->String() << " process"; c->Process(); } else if (c->status_.error_code() == grpc::StatusCode::DEADLINE_EXCEEDED) { + // FIXME(gongwb): parse error_details? LOG(ERROR) << c->GetVarHandlePtr()->String() - << " meets grpc error:" << c->status_.error_message(); + << " meets grpc error, error_code:" << c->status_.error_code() + << " error_message:" << c->status_.error_message() + << " error_details:" << c->status_.error_details(); { std::lock_guard lk(sync_mutex_); ok_ = false; @@ -286,7 +350,10 @@ void GRPCClient::Proceed() { c->Finish(false); } else { LOG(FATAL) << c->GetVarHandlePtr()->String() - << " meets grpc error:" << c->status_.error_message(); + << " meets grpc error, error_code:" << c->status_.error_code() + << " error_message:" << c->status_.error_message() + << " error_details:" << c->status_.error_details(); + c->Finish(false); } @@ -303,7 +370,7 @@ void GRPCClient::Proceed() { sync_cond_.notify_all(); } } - VLOG(3) << "GRPCClient Proceed end"; + VLOG(30) << "GRPCClient Proceed end"; } std::shared_ptr GRPCClient::GetChannel(const std::string& ep) { diff --git a/paddle/fluid/operators/distributed/grpc_serde.cc b/paddle/fluid/operators/distributed/grpc_serde.cc index 3f8796713a6b89a308113981614673e07e8d367f..b201c4a5763148165f517c719227d6317ecbe350 100644 --- a/paddle/fluid/operators/distributed/grpc_serde.cc +++ b/paddle/fluid/operators/distributed/grpc_serde.cc @@ -34,8 +34,9 @@ namespace distributed { void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, - ::grpc::ByteBuffer* msg, - const std::string& out_name) { + ::grpc::ByteBuffer* msg, const std::string& out_name, + const int trainer_id) { + platform::RecordRPCEvent record_event("serial", &ctx); // Default DestroyCallback does nothing, When using GPU // the CPU buffer need to be freed. DestroyCallback destroy_callback = [](void* backing) {}; @@ -44,6 +45,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, size_t payload_size; request.set_varname(name); + request.set_trainer_id(trainer_id); // Note: normally the profiler is enabled in 1 trainer, hence only // 1 trainer returns true for ShouldSendProfileState(). It tells PS // servers the trainer's profiling state so that PS can follow the @@ -146,10 +148,12 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, const framework::Scope* scope, - framework::Variable** var) { + framework::Variable** var, int* trainer_id) { + platform::RecordRPCEvent record_event("deserial", &ctx); operators::distributed::GRPCVariableResponse resp(scope, &ctx); PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!"); *var = resp.GetVar(); + *trainer_id = resp.GetTrainerId(); } } // namespace distributed diff --git a/paddle/fluid/operators/distributed/grpc_serde.h b/paddle/fluid/operators/distributed/grpc_serde.h index 450c41dcd6b1bf9a33d3bbef3a1c94a2f83ff322..7ec489e961630747ba00e68ad3603cacbb1aa485 100644 --- a/paddle/fluid/operators/distributed/grpc_serde.h +++ b/paddle/fluid/operators/distributed/grpc_serde.h @@ -38,12 +38,13 @@ typedef void (*DestroyCallback)(void*); void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, ::grpc::ByteBuffer* msg, - const std::string& out_varname = std::string()); + const std::string& out_varname = std::string(), + const int trainer_id = 0); void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, const framework::Scope* scope, - framework::Variable** var); + framework::Variable** var, int* trainer_id); } // namespace distributed } // namespace operators diff --git a/paddle/fluid/operators/distributed/grpc_server.cc b/paddle/fluid/operators/distributed/grpc_server.cc index 8edb00276df3ade1b320fbf2873e8b54ff3e1464..ffd2b1707bea6c9379dc09c629fa4c920dac8ed0 100644 --- a/paddle/fluid/operators/distributed/grpc_server.cc +++ b/paddle/fluid/operators/distributed/grpc_server.cc @@ -98,13 +98,14 @@ class RequestSend final : public RequestBase { void Process() override { std::string varname = GetReqName(); - VLOG(4) << "RequestSend var_name:" << varname; + VLOG(40) << "RequestSend var_name:" << varname; auto scope = request_->GetMutableLocalScope(); auto invar = request_->GetVar(); + int trainer_id = request_->GetTrainerId(); framework::Variable* outvar = nullptr; - request_handler_->Handle(varname, scope, invar, &outvar); + request_handler_->Handle(varname, scope, invar, &outvar, trainer_id); Finish(reply_, &responder_); } @@ -133,13 +134,14 @@ class RequestGet final : public RequestBase { void Process() override { // proc request. std::string varname = request_.varname(); - VLOG(4) << "RequestGet " << varname; + int trainer_id = request_.trainer_id(); + VLOG(40) << "RequestGet " << varname; auto scope = request_handler_->scope(); auto invar = scope->FindVar(varname); framework::Variable* outvar = nullptr; - request_handler_->Handle(varname, scope, invar, &outvar); + request_handler_->Handle(varname, scope, invar, &outvar, trainer_id); if (outvar) { SerializeToByteBuffer(varname, outvar, *request_handler_->dev_ctx(), @@ -179,15 +181,17 @@ class RequestPrefetch final : public RequestBase { // prefetch process... std::string in_var_name = request_->Varname(); std::string out_var_name = request_->OutVarname(); - VLOG(4) << "RequestPrefetch, in_var_name: " << in_var_name - << " out_var_name: " << out_var_name; + int trainer_id = request_->GetTrainerId(); + VLOG(40) << "RequestPrefetch, in_var_name: " << in_var_name + << " out_var_name: " << out_var_name; auto scope = request_->GetMutableLocalScope(); auto invar = scope->FindVar(in_var_name); // out var must be created in local scope! framework::Variable* outvar = scope->Var(out_var_name); - request_handler_->Handle(in_var_name, scope, invar, &outvar, out_var_name); + request_handler_->Handle(in_var_name, scope, invar, &outvar, trainer_id, + out_var_name); SerializeToByteBuffer(out_var_name, outvar, *request_handler_->dev_ctx(), &reply_); @@ -225,12 +229,13 @@ class RequestCheckpointNotify final : public RequestBase { std::string checkpoint_notify = request_->Varname(); std::string checkpoint_dir = request_->OutVarname(); + int trainer_id = request_->GetTrainerId(); - VLOG(4) << "RequestCheckpointNotify notify: " << checkpoint_notify - << ", dir: " << checkpoint_dir; + VLOG(40) << "RequestCheckpointNotify notify: " << checkpoint_notify + << ", dir: " << checkpoint_dir; request_handler_->Handle(checkpoint_notify, scope, nullptr, nullptr, - checkpoint_dir); + trainer_id, checkpoint_dir); Finish(reply_, &responder_); } @@ -241,10 +246,10 @@ class RequestCheckpointNotify final : public RequestBase { }; void AsyncGRPCServer::WaitServerReady() { - VLOG(4) << "AsyncGRPCServer is wait server ready"; + VLOG(40) << "AsyncGRPCServer is wait server ready"; std::unique_lock lock(this->mutex_ready_); condition_ready_.wait(lock, [=] { return this->ready_ == 1; }); - VLOG(4) << "AsyncGRPCServer WaitSeverReady"; + VLOG(40) << "AsyncGRPCServer WaitSeverReady"; } void AsyncGRPCServer::StartServer() { @@ -277,14 +282,15 @@ void AsyncGRPCServer::StartServer() { reqs.reserve(kRequestBufSize); for (int i = 0; i < kRequestBufSize; i++) { - VLOG(6) << "TryToRegisterNewOne on RPC NAME: " << rpc_name << " I: " << i; + VLOG(60) << "TryToRegisterNewOne on RPC NAME: " << rpc_name + << " I: " << i; TryToRegisterNewOne(rpc_name, i); } for (int i = 0; i < threadnum; i++) { rpc_threads_[rpc_name].emplace_back(new std::thread(std::bind( &AsyncGRPCServer::HandleRequest, this, cq.get(), rpc_name, f))); - VLOG(4) << t.first << " creates threads!"; + VLOG(40) << t.first << " creates threads!"; } } @@ -301,7 +307,7 @@ void AsyncGRPCServer::StartServer() { auto& threads = t.second; for (size_t i = 0; i < threads.size(); ++i) { threads[i]->join(); - VLOG(4) << t.first << " threads ends!"; + VLOG(40) << t.first << " threads ends!"; } } } @@ -309,7 +315,7 @@ void AsyncGRPCServer::StartServer() { void AsyncGRPCServer::ShutdownQueue() { for (auto& t : rpc_cq_) { t.second->Shutdown(); - VLOG(4) << t.first << " queue shutdown!"; + VLOG(40) << t.first << " queue shutdown!"; } } @@ -318,7 +324,7 @@ void AsyncGRPCServer::ShutDownImpl() { is_shut_down_ = true; ShutdownQueue(); - VLOG(4) << "server_ shutdown!"; + VLOG(40) << "server_ shutdown!"; server_->Shutdown(); } @@ -326,12 +332,12 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, int req_id) { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { - VLOG(4) << "shutdown, do not TryToRegisterNewSendOne"; + VLOG(40) << "shutdown, do not TryToRegisterNewSendOne"; return; } - VLOG(4) << "TryToRegisterNewOne on RPC NAME: " << rpc_name - << " REQ ID: " << req_id; + VLOG(40) << "TryToRegisterNewOne on RPC NAME: " << rpc_name + << " REQ ID: " << req_id; auto& reqs = rpc_reqs_[rpc_name]; auto& handler = rpc_call_map_[rpc_name]; @@ -352,7 +358,7 @@ void AsyncGRPCServer::TryToRegisterNewOne(const std::string& rpc_name, reqs[req_id] = b; - VLOG(4) << "Create RequestSend status:" << b->Status(); + VLOG(40) << "Create RequestSend status:" << b->Status(); } void AsyncGRPCServer::HandleRequest( @@ -362,15 +368,15 @@ void AsyncGRPCServer::HandleRequest( bool ok = false; while (true) { - VLOG(4) << "HandleRequest " << rpc_name << " wait next"; + VLOG(40) << "HandleRequest " << rpc_name << " wait next"; if (!cq->Next(&tag, &ok)) { - VLOG(3) << "CompletionQueue " << rpc_name << " shutdown!"; + VLOG(30) << "CompletionQueue " << rpc_name << " shutdown!"; break; } int req_id = static_cast(reinterpret_cast(tag)); - VLOG(4) << "HandleRequest " << rpc_name << ", req_id:" << req_id - << " get next"; + VLOG(40) << "HandleRequest " << rpc_name << ", req_id:" << req_id + << " get next"; auto& reqs = rpc_reqs_[rpc_name]; RequestBase* base = nullptr; @@ -380,7 +386,7 @@ void AsyncGRPCServer::HandleRequest( base = reqs[req_id]; } - VLOG(3) << base->Status2String(rpc_name); + VLOG(30) << base->Status2String(rpc_name); // reference: // https://github.com/tensorflow/tensorflow/issues/5596 diff --git a/paddle/fluid/operators/distributed/grpc_variable_response.cc b/paddle/fluid/operators/distributed/grpc_variable_response.cc index 34d47f3ec0f3025109447b66078b724607d2953a..d6d219d4369ba785e5c369538d4a18dc682952c1 100644 --- a/paddle/fluid/operators/distributed/grpc_variable_response.cc +++ b/paddle/fluid/operators/distributed/grpc_variable_response.cc @@ -286,13 +286,21 @@ int GRPCVariableResponse::Parse(Source* source) { platform::EnableProfiler(platform::ProfilerState::kCPU); } else if (profiling == platform::kDisableProfiler && platform::IsProfileEnabled()) { - // TODO(panyx0718): Should we allow to customize file dir. platform::DisableProfiler( platform::EventSortingKey::kDefault, - string::Sprintf("/tmp/profile_ps_%lld", listener_id)); + string::Sprintf("%s_%lld", FLAGS_rpc_server_profile_path, + listener_id)); } break; } + case sendrecv::VariableMessage::kTrainerIdFieldNumber: { + uint64_t trainer_id = 0; + if (!input.ReadVarint64(&trainer_id)) { + return tag; + } + meta_.set_trainer_id(trainer_id); + break; + } default: { // Unknown tag, return unknown error. return -1; diff --git a/paddle/fluid/operators/distributed/request_handler.h b/paddle/fluid/operators/distributed/request_handler.h index 5be7095acd3c5ac6f880a8a26c246f60a93643b5..3bcc59a47ba5f52da1374f220828a0f392e13d27 100644 --- a/paddle/fluid/operators/distributed/request_handler.h +++ b/paddle/fluid/operators/distributed/request_handler.h @@ -75,7 +75,7 @@ class VarHandle { wait_cond_.wait(lk, [this] { return status_ != kDefaultState; }); ret = status_; } - VLOG(7) << "VarHandle wait:" << ret; + VLOG(70) << "VarHandle wait:" << ret; return ret != kErrorState; } @@ -84,7 +84,7 @@ class VarHandle { std::unique_lock lk(sync_mutex_); status_ = ok ? kFinishState : kErrorState; } - VLOG(7) << "VarHandle finish:" << ok; + VLOG(70) << "VarHandle finish:" << ok; wait_cond_.notify_all(); } @@ -190,6 +190,7 @@ class RequestHandler { // } virtual bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") = 0; protected: diff --git a/paddle/fluid/operators/distributed/request_handler_impl.cc b/paddle/fluid/operators/distributed/request_handler_impl.cc index 849e412504eb9180b746db65fd4fa353ed0c05a1..dae56cc8436c2241bfc8ae37ba3cad4069a054bf 100644 --- a/paddle/fluid/operators/distributed/request_handler_impl.cc +++ b/paddle/fluid/operators/distributed/request_handler_impl.cc @@ -36,21 +36,21 @@ bool RequestSendHandler::Handle(const std::string& varname, framework::Scope* scope, framework::Variable* invar, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name) { - VLOG(4) << "RequestSendHandler:" << varname; + VLOG(40) << "RequestSendHandler:" << varname; // Sync if (varname == BATCH_BARRIER_MESSAGE) { - VLOG(3) << "sync: recv BATCH_BARRIER_MESSAGE"; + VLOG(30) << "sync: recv BATCH_BARRIER_MESSAGE"; rpc_server_->IncreaseBatchBarrier(kRequestSend); } else if (varname == COMPLETE_MESSAGE) { - VLOG(3) << "sync: recv complete message"; + VLOG(30) << "sync: recv complete message"; rpc_server_->Complete(); } else { // Async if (!sync_mode_) { - VLOG(3) << "async process var: " << varname; - rpc_server_->Profiler().OneStep(); + VLOG(30) << "async process var: " << varname; try { executor_->RunPreparedContext((*grad_to_prepared_ctx_)[varname].get(), scope); @@ -61,7 +61,7 @@ bool RequestSendHandler::Handle(const std::string& varname, return true; } else { // sync rpc_server_->WaitCond(kRequestSend); - VLOG(3) << "sync: processing received var: " << varname; + VLOG(30) << "sync: processing received var: " << varname; if (invar == nullptr) { LOG(FATAL) << "sync: Can not find server side var: " << varname; @@ -76,11 +76,12 @@ bool RequestGetHandler::Handle(const std::string& varname, framework::Scope* scope, framework::Variable* invar, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name) { - VLOG(4) << "RequestGetHandler:" << varname; + VLOG(40) << "RequestGetHandler:" << varname; if (sync_mode_) { if (varname == FETCH_BARRIER_MESSAGE) { - VLOG(3) << "sync: recv fetch barrier message"; + VLOG(30) << "sync: recv fetch barrier message"; rpc_server_->IncreaseBatchBarrier(kRequestGet); } else { rpc_server_->WaitCond(kRequestGet); @@ -88,6 +89,20 @@ bool RequestGetHandler::Handle(const std::string& varname, } } else { if (varname != FETCH_BARRIER_MESSAGE && varname != COMPLETE_MESSAGE) { + if (enable_dc_asgd_) { + // NOTE: the format is determined by distributed_transpiler.py + std::string param_bak_name = + string::Sprintf("%s.trainer_%d_bak", varname, trainer_id); + VLOG(30) << "getting " << param_bak_name << " trainer_id " + << trainer_id; + auto var = scope_->FindVar(varname); + auto t_orig = var->Get(); + auto param_bak = scope_->Var(param_bak_name); + auto t = param_bak->GetMutable(); + t->mutable_data(dev_ctx_->GetPlace(), t_orig.type()); + VLOG(30) << "copying " << varname << " to " << param_bak_name; + framework::TensorCopy(t_orig, dev_ctx_->GetPlace(), t); + } *outvar = scope_->FindVar(varname); } } @@ -98,8 +113,9 @@ bool RequestPrefetchHandler::Handle(const std::string& varname, framework::Scope* scope, framework::Variable* invar, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name) { - VLOG(4) << "RequestPrefetchHandler " << varname; + VLOG(40) << "RequestPrefetchHandler " << varname; auto var_desc = program_->Block(0).FindVar(out_var_name); InitializeVariable(*outvar, var_desc->GetType()); @@ -113,6 +129,7 @@ bool RequestCheckpointHandler::Handle(const std::string& varname, framework::Scope* scope, framework::Variable* invar, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name) { PADDLE_ENFORCE( checkpoint_notify_id != -1, @@ -122,8 +139,8 @@ bool RequestCheckpointHandler::Handle(const std::string& varname, auto* lt_var = scope_->FindVar(LOOKUP_TABLE_PATH)->GetMutable(); lt_var->clear(); lt_var->append(out_var_name); - VLOG(4) << "RequestCheckpointHandler update var kLookupTablePath to: " - << out_var_name; + VLOG(40) << "RequestCheckpointHandler update var kLookupTablePath to: " + << out_var_name; executor_->RunPreparedContext(checkpoint_prepared_ctx_.get(), scope_); return true; } diff --git a/paddle/fluid/operators/distributed/request_handler_impl.h b/paddle/fluid/operators/distributed/request_handler_impl.h index 8be5b21bb89a580f4091de19186fd2d7e5802478..c1afda9dd2445e492d8b93659c9ff13e6e1030b8 100644 --- a/paddle/fluid/operators/distributed/request_handler_impl.h +++ b/paddle/fluid/operators/distributed/request_handler_impl.h @@ -36,20 +36,34 @@ namespace distributed { class RequestSendHandler final : public RequestHandler { public: - explicit RequestSendHandler(bool sync_mode) : RequestHandler(sync_mode) {} + explicit RequestSendHandler(bool sync_mode, bool enable_dc_asgd = false) + : RequestHandler(sync_mode) { + enable_dc_asgd_ = enable_dc_asgd; + } virtual ~RequestSendHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") override; + + private: + bool enable_dc_asgd_; }; class RequestGetHandler final : public RequestHandler { public: - explicit RequestGetHandler(bool sync_mode) : RequestHandler(sync_mode) {} + explicit RequestGetHandler(bool sync_mode, bool enable_dc_asgd = false) + : RequestHandler(sync_mode) { + enable_dc_asgd_ = enable_dc_asgd; + } virtual ~RequestGetHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") override; + + private: + bool enable_dc_asgd_; }; class RequestPrefetchHandler final : public RequestHandler { @@ -58,6 +72,7 @@ class RequestPrefetchHandler final : public RequestHandler { virtual ~RequestPrefetchHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") override; }; @@ -70,6 +85,7 @@ class RequestCheckpointHandler final : public RequestHandler { virtual ~RequestCheckpointHandler() {} bool Handle(const std::string& varname, framework::Scope* scope, framework::Variable* var, framework::Variable** outvar, + const int trainer_id, const std::string& out_var_name = "") override; private: diff --git a/paddle/fluid/operators/distributed/rpc_client.cc b/paddle/fluid/operators/distributed/rpc_client.cc index b5ec9fe5367beb97b3cc7298102deff1e8ca4ec9..390e9af0f38c920f39e8cd3e1b3d28fd89b239fe 100644 --- a/paddle/fluid/operators/distributed/rpc_client.cc +++ b/paddle/fluid/operators/distributed/rpc_client.cc @@ -24,6 +24,7 @@ namespace distributed { std::once_flag RPCClient::init_flag_; std::unique_ptr RPCClient::rpc_client_(nullptr); +int RPCClient::trainer_id_ = 0; } // namespace distributed } // namespace operators diff --git a/paddle/fluid/operators/distributed/rpc_client.h b/paddle/fluid/operators/distributed/rpc_client.h index 3539ee5e459d6dfe0b6510806464bcc6817910bb..1983802e49506c79041112ac87d429e4c084ddfd 100644 --- a/paddle/fluid/operators/distributed/rpc_client.h +++ b/paddle/fluid/operators/distributed/rpc_client.h @@ -72,14 +72,15 @@ class RPCClient { virtual bool Wait() = 0; template - static RPCClient* GetInstance() { - std::call_once(init_flag_, &RPCClient::Init); + static RPCClient* GetInstance(int trainer_id) { + std::call_once(init_flag_, &RPCClient::Init, trainer_id); return rpc_client_.get(); } // Init is called by GetInstance. template - static void Init() { + static void Init(int trainer_id) { + trainer_id_ = trainer_id; if (rpc_client_.get() == nullptr) { rpc_client_.reset(new T()); rpc_client_->InitImpl(); @@ -88,6 +89,8 @@ class RPCClient { protected: virtual void InitImpl() {} + // each trainer have exact one trainer id, it should be static + static int trainer_id_; private: static std::once_flag init_flag_; diff --git a/paddle/fluid/operators/distributed/rpc_server.cc b/paddle/fluid/operators/distributed/rpc_server.cc index 084480ae48b8b9267ade1a840f6a70519cb28e48..4055091104f2f96070d0c4e806c6908da691d732 100644 --- a/paddle/fluid/operators/distributed/rpc_server.cc +++ b/paddle/fluid/operators/distributed/rpc_server.cc @@ -20,42 +20,10 @@ #include "paddle/fluid/operators/distributed/rpc_server.h" #include "paddle/fluid/platform/profiler.h" -DEFINE_int32(rpc_server_profile_period, 0, - "the period of listen_and_serv to do profile"); -DEFINE_string(rpc_server_profile_path, "/dev/null", - "the profile log file path"); - namespace paddle { namespace operators { namespace distributed { -RPCServerProfiler::RPCServerProfiler(int profile_period, - const std::string& profile_log_path) - : profile_period_(profile_period), profile_log_path_(profile_log_path) { - step_ = 0; -} - -void RPCServerProfiler::OneStep() { - PADDLE_ENFORCE_LE(step_, profile_period_, - "step_ should not be larger then " - "profile_period_"); - if (profile_period_ <= 0) { - return; - } - - if (step_ == 0) { - auto pf_state = paddle::platform::ProfilerState::kCPU; - paddle::platform::EnableProfiler(pf_state); - } - if (step_ == profile_period_) { - paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kTotal, - profile_log_path_); - step_ = 0; - } else { - step_++; - } -} - void RPCServer::ShutDown() { LOG(INFO) << "RPCServer ShutDown "; ShutDownImpl(); @@ -71,7 +39,7 @@ void RPCServer::SavePort() const { port_file.open(file_path); port_file << selected_port_; port_file.close(); - VLOG(4) << "selected port written to " << file_path; + VLOG(40) << "selected port written to " << file_path; } void RPCServer::WaitBarrier(const std::string& rpc_name) { @@ -81,12 +49,12 @@ void RPCServer::WaitBarrier(const std::string& rpc_name) { exit_flag_.load()); }); - VLOG(3) << "batch_barrier_: " << rpc_name << " " - << barrier_counter_[rpc_name]; + VLOG(30) << "batch_barrier_: " << rpc_name << " " + << barrier_counter_[rpc_name]; } void RPCServer::IncreaseBatchBarrier(const std::string rpc_name) { - VLOG(4) << "RPCServer begin IncreaseBatchBarrier " << rpc_name; + VLOG(40) << "RPCServer begin IncreaseBatchBarrier " << rpc_name; int b = 0; std::unique_lock lock(mutex_); b = ++barrier_counter_[rpc_name]; @@ -103,7 +71,7 @@ void RPCServer::Complete() { client_num_--; need_reset_all_vars_ = true; - VLOG(4) << "decrease client_num to: " << client_num_; + VLOG(40) << "decrease client_num to: " << client_num_; if (cur_cond_.load() == rpc_cond_map_[kRequestGet]) { barrier_counter_[kRequestGet]--; } @@ -122,7 +90,7 @@ int RPCServer::GetClientNum() { } void RPCServer::ResetBarrierCounter() { - VLOG(3) << "RPCServer ResetBarrierCounter "; + VLOG(30) << "RPCServer ResetBarrierCounter "; std::unique_lock lock(mutex_); for (auto& t : barrier_counter_) { t.second = 0; @@ -137,12 +105,12 @@ void RPCServer::RegisterRPC(const std::string& rpc_name, static int cond = -1; rpc_cond_map_[rpc_name] = ++cond; - VLOG(4) << "RegisterRPC rpc_name:" << rpc_name << ", handler:" << handler - << ", cond:" << rpc_cond_map_[rpc_name]; + VLOG(40) << "RegisterRPC rpc_name:" << rpc_name << ", handler:" << handler + << ", cond:" << rpc_cond_map_[rpc_name]; } void RPCServer::SetCond(const std::string& rpc_name) { - VLOG(3) << "RPCServer SetCond " << rpc_name; + VLOG(30) << "RPCServer SetCond " << rpc_name; { std::unique_lock lock(mutex_); cur_cond_ = rpc_cond_map_[rpc_name]; @@ -152,7 +120,7 @@ void RPCServer::SetCond(const std::string& rpc_name) { } void RPCServer::WaitCond(const std::string& rpc_name) { - VLOG(4) << "RPCServer WaitCond " << rpc_name; + VLOG(40) << "RPCServer WaitCond " << rpc_name; int cond = 0; { std::unique_lock lock(mutex_); diff --git a/paddle/fluid/operators/distributed/rpc_server.h b/paddle/fluid/operators/distributed/rpc_server.h index f3e61e1575ced0b9ffbad23e6973121daca9751b..c78c5007a7f262f15305b6c284e8c4fbddef42a0 100644 --- a/paddle/fluid/operators/distributed/rpc_server.h +++ b/paddle/fluid/operators/distributed/rpc_server.h @@ -23,30 +23,14 @@ #include "paddle/fluid/operators/distributed/request_handler.h" -DECLARE_int32(rpc_server_profile_period); -DECLARE_string(rpc_server_profile_path); - namespace paddle { namespace operators { namespace distributed { -class RPCServerProfiler { - public: - RPCServerProfiler(int profile_period, const std::string& profile_log_path); - void OneStep(); - - private: - const int profile_period_; - std::string profile_log_path_; - int step_; -}; - class RPCServer { public: explicit RPCServer(const std::string& address, int client_num) : cur_cond_(0), - profiler_(FLAGS_rpc_server_profile_period, - FLAGS_rpc_server_profile_path), bind_address_(address), exit_flag_(false), selected_port_(0), @@ -86,7 +70,6 @@ class RPCServer { void Complete(); void ResetBarrierCounter(); - RPCServerProfiler& Profiler() { return profiler_; } bool NeedResetAllVars(); @@ -101,7 +84,6 @@ class RPCServer { std::unordered_map rpc_cond_map_; std::atomic cur_cond_; std::condition_variable rpc_cond_; - RPCServerProfiler profiler_; protected: std::string bind_address_; diff --git a/paddle/fluid/operators/distributed/rpc_server_test.cc b/paddle/fluid/operators/distributed/rpc_server_test.cc index d6176e1443d2a441af7878e5efe99796d486bb7a..c3dd459fc4e8c4bd304c09d7a3ed4f456c4dc69f 100644 --- a/paddle/fluid/operators/distributed/rpc_server_test.cc +++ b/paddle/fluid/operators/distributed/rpc_server_test.cc @@ -125,7 +125,7 @@ TEST(PREFETCH, CPU) { g_req_handler.reset(new distributed::RequestPrefetchHandler(true)); g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 1)); distributed::RPCClient* client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(0); std::thread server_thread(StartServer, distributed::kRequestPrefetch); g_rpc_service->WaitServerReady(); @@ -165,7 +165,7 @@ TEST(COMPLETE, CPU) { g_req_handler.reset(new distributed::RequestSendHandler(true)); g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 2)); distributed::RPCClient* client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(0); PADDLE_ENFORCE(client != nullptr); std::thread server_thread(StartServer, distributed::kRequestSend); g_rpc_service->WaitServerReady(); diff --git a/paddle/fluid/operators/distributed/send_recv.proto.in b/paddle/fluid/operators/distributed/send_recv.proto.in index 8b0a09abe1d05dda10eda0030eb91cb9ca40683e..55820c980e8139625c1b589f9d2d68dfee74a212 100644 --- a/paddle/fluid/operators/distributed/send_recv.proto.in +++ b/paddle/fluid/operators/distributed/send_recv.proto.in @@ -79,6 +79,7 @@ message VariableMessage { // server stops profiling and generates a profile to /tmp/profile_ps_* // when profile switches from 1 to 2. int64 profile = 11; + int64 trainer_id = 12; } message VoidMessage {} diff --git a/paddle/fluid/operators/distributed/variable_response.cc b/paddle/fluid/operators/distributed/variable_response.cc index c4854d50b6371064003a10e18efc9e5f160d9a42..d1572ce01aa17273988955c27bdea5b2f40c27ea 100644 --- a/paddle/fluid/operators/distributed/variable_response.cc +++ b/paddle/fluid/operators/distributed/variable_response.cc @@ -16,6 +16,9 @@ #include #include "paddle/fluid/operators/distributed/sendrecvop_utils.h" +DEFINE_string(rpc_server_profile_path, "./profile_ps", + "the profile log file path"); + namespace paddle { namespace operators { namespace distributed { @@ -47,7 +50,7 @@ bool VariableResponse::ReadRaw(::google::protobuf::io::CodedInputStream* input, size_to_write = length - total_written; } // This log is useful to see how long a internal block size is of rpc. - VLOG(7) << "copy " << size_to_write << " data to CUDAPlace"; + VLOG(70) << "copy " << size_to_write << " data to CUDAPlace"; memory::Copy(boost::get(place), reinterpret_cast(p), cpu, data, size_to_write, gpu_dev_ctx.stream()); @@ -76,7 +79,7 @@ bool VariableResponse::ReadRaw(::google::protobuf::io::CodedInputStream* input, // TODO(gongwb): can we avoid copy? platform::CPUPlace cpu; // This log is useful to see how long a internal block size is of rpc. - VLOG(7) << "copy " << size_to_write << " data to CPUPlace"; + VLOG(70) << "copy " << size_to_write << " data to CPUPlace"; memory::Copy(cpu, reinterpret_cast(p), cpu, data, size_to_write); p += size_to_write; @@ -195,8 +198,8 @@ bool VariableResponse::ProcSerializedField( #endif } - VLOG(7) << "ProcSerializedField:" << meta_.varname() - << ", type:" << meta_.type() << std::endl; + VLOG(70) << "ProcSerializedField:" << meta_.varname() + << ", type:" << meta_.type() << std::endl; framework::DDim dims = GetDims(meta_.dims()); if (meta_.type() == sendrecv::LOD_TENSOR) { PADDLE_ENFORCE(meta_.lod_size() >= 0, "lod info should be got first!"); diff --git a/paddle/fluid/operators/distributed/variable_response.h b/paddle/fluid/operators/distributed/variable_response.h index 6aec52ca00f59a42ecca01da8df1680ce4eda432..4c7fcbbdfb305ce6b4fc9d1edd9738899b200ec6 100644 --- a/paddle/fluid/operators/distributed/variable_response.h +++ b/paddle/fluid/operators/distributed/variable_response.h @@ -27,6 +27,8 @@ #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/operators/distributed/send_recv.pb.h" +DECLARE_string(rpc_server_profile_path); + namespace paddle { namespace operators { namespace distributed { @@ -92,6 +94,8 @@ class VariableResponse { return scope_->FindVar(meta_.varname()); } + int GetTrainerId() { return static_cast(meta_.trainer_id()); } + protected: bool ReadRaw(::google::protobuf::io::CodedInputStream* input, const platform::DeviceContext& dev_ctx, platform::Place place, diff --git a/paddle/fluid/operators/dropout_op.cc b/paddle/fluid/operators/dropout_op.cc index 07322e720f26213ea777be3cd22f2fead28507f0..3c28ef30922e6d6ba09b96282619eef15867631e 100644 --- a/paddle/fluid/operators/dropout_op.cc +++ b/paddle/fluid/operators/dropout_op.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/dropout_op.h" +#include namespace paddle { namespace operators { @@ -57,6 +58,29 @@ class DropoutOpMaker : public framework::OpProtoAndCheckerMaker { "will be dropped.") .SetDefault(false); AddAttr("seed", "Dropout random seed.").SetDefault(0); + AddAttr( + "dropout_implementation", + "[\"downgrade_in_infer\"|\"upscale_in_train\"]" + "There are two kinds of ways to implement dropout" + "(the mask below is a tensor have the same shape with input" + "the value of mask is 0 or 1, the ratio of 0 is dropout_prob)" + "1. downgrade_in_infer(default), downgrade the outcome at inference " + "time" + " train: out = input * mask" + " inference: out = input * dropout_prob" + "2. upscale_in_train, upscale the outcome at training time, do nothing " + "in inference" + " train: out = input * mask / ( 1.0 - dropout_prob )" + " inference: out = input" + " dropout op can be removed from the program. the program will be " + "efficient") + .SetDefault("downgrade_in_infer") + .AddCustomChecker([](const std::string& type) { + PADDLE_ENFORCE( + type == "downgrade_in_infer" || type == "upscale_in_train", + "dropout_implementation can only be downgrade_in_infer or " + "upscale_in_train"); + }); AddComment(R"DOC( Dropout Operator. @@ -104,7 +128,9 @@ REGISTER_OPERATOR(dropout, ops::DropoutOp, ops::DropoutOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(dropout_grad, ops::DropoutOpGrad); REGISTER_OP_CPU_KERNEL( - dropout, ops::CPUDropoutKernel); + dropout, ops::CPUDropoutKernel, + ops::CPUDropoutKernel); REGISTER_OP_CPU_KERNEL( dropout_grad, - ops::DropoutGradKernel); + ops::DropoutGradKernel, + ops::DropoutGradKernel); diff --git a/paddle/fluid/operators/dropout_op.cu b/paddle/fluid/operators/dropout_op.cu index 1dd66e0280c46c0624ff70e822cb6fa6f06b7aa9..e011f47e086183a4ef3a3373c17acd6c21b6cf7e 100644 --- a/paddle/fluid/operators/dropout_op.cu +++ b/paddle/fluid/operators/dropout_op.cu @@ -17,6 +17,7 @@ limitations under the License. */ #include #include #include +#include #include "paddle/fluid/operators/dropout_op.h" #include "paddle/fluid/platform/float16.h" @@ -26,7 +27,8 @@ namespace operators { template __global__ void RandomGenerator(const size_t n, const int seed, const float dropout_prob, const T* src, - T* mask_data, T* dst) { + T* mask_data, T* dst, + bool is_upscale_in_train) { thrust::minstd_rand rng; rng.seed(seed); thrust::uniform_real_distribution dist(0, 1); @@ -47,7 +49,11 @@ __global__ void RandomGenerator(const size_t n, const int seed, if (dist(rng) < dropout_prob) { mask = static_cast(0); } else { - mask = static_cast(1); + if (is_upscale_in_train) { + mask = static_cast(1.0f / (1.0f - dropout_prob)); + } else { + mask = static_cast(1); + } } dest = s * mask; mask_data[idx] = mask; @@ -67,6 +73,8 @@ class GPUDropoutKernel : public framework::OpKernel { y->mutable_data(context.GetPlace()); float dropout_prob = context.Attr("dropout_prob"); + auto dropout_implementation = + context.Attr("dropout_implementation"); auto& place = *context.template device_context().eigen_device(); if (!context.Attr("is_test")) { auto* mask = context.Output("Mask"); @@ -83,11 +91,16 @@ class GPUDropoutKernel : public framework::OpKernel { int grid = (x->numel() + threads - 1) / threads; RandomGenerator< T><<>>( - size, seed, dropout_prob, x_data, mask_data, y_data); + size, seed, dropout_prob, x_data, mask_data, y_data, + (dropout_implementation == "upscale_in_train")); } else { auto X = EigenMatrix::Reshape(*x, 1); auto Y = EigenMatrix::Reshape(*y, 1); - Y.device(place) = X * static_cast(1.0f - dropout_prob); + if (dropout_implementation == "upscale_in_train") { + Y.device(place) = X; + } else { + Y.device(place) = X * static_cast(1.0f - dropout_prob); + } } } }; @@ -99,6 +112,8 @@ namespace ops = paddle::operators; namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( dropout, ops::GPUDropoutKernel, - ops::GPUDropoutKernel); -REGISTER_OP_CUDA_KERNEL(dropout_grad, - ops::DropoutGradKernel); + ops::GPUDropoutKernel, + ops::GPUDropoutKernel); +REGISTER_OP_CUDA_KERNEL( + dropout_grad, ops::DropoutGradKernel, + ops::DropoutGradKernel); diff --git a/paddle/fluid/operators/dropout_op.h b/paddle/fluid/operators/dropout_op.h index 0628b4b826d2730a8e3fb4842e4ae550b8c00569..6c629b7b6d255828023ed25680675ca104a33e12 100644 --- a/paddle/fluid/operators/dropout_op.h +++ b/paddle/fluid/operators/dropout_op.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once #include +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" @@ -36,6 +37,8 @@ class CPUDropoutKernel : public framework::OpKernel { auto* y_data = y->mutable_data(context.GetPlace()); float dropout_prob = context.Attr("dropout_prob"); + auto dropout_implementation = + context.Attr("dropout_implementation"); if (!context.Attr("is_test")) { auto* mask = context.Output("Mask"); auto* mask_data = mask->mutable_data(context.GetPlace()); @@ -49,14 +52,20 @@ class CPUDropoutKernel : public framework::OpKernel { engine.seed(seed); std::uniform_real_distribution dist(0, 1); + size_t size = framework::product(mask->dims()); for (size_t i = 0; i < size; ++i) { if (dist(engine) < dropout_prob) { mask_data[i] = 0; y_data[i] = 0; } else { - mask_data[i] = 1; - y_data[i] = x_data[i]; + if (dropout_implementation == "upscale_in_train") { + mask_data[i] = 1.0f / static_cast(1.0f - dropout_prob); + y_data[i] = x_data[i] / static_cast(1.0f - dropout_prob); + } else { + mask_data[i] = 1; + y_data[i] = x_data[i]; + } } } } else { @@ -64,7 +73,11 @@ class CPUDropoutKernel : public framework::OpKernel { auto Y = EigenMatrix::Reshape(*y, 1); auto& place = *context.template device_context().eigen_device(); - Y.device(place) = X * (1.0f - dropout_prob); + if (dropout_implementation == "upscale_in_train") { + Y.device(place) = X; + } else { + Y.device(place) = X * static_cast(1.0f - dropout_prob); + } } } }; diff --git a/paddle/fluid/operators/elementwise_add_op.cu b/paddle/fluid/operators/elementwise_add_op.cu index dfff518f170b56d180b6883c363effb8dbd677b6..f9f5c66d34fa1d73db00173e493f9953b8579518 100644 --- a/paddle/fluid/operators/elementwise_add_op.cu +++ b/paddle/fluid/operators/elementwise_add_op.cu @@ -30,4 +30,5 @@ REGISTER_OP_CUDA_KERNEL( ops::ElementwiseAddGradKernel, ops::ElementwiseAddGradKernel, ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel); + ops::ElementwiseAddGradKernel, + ops::ElementwiseAddGradKernel); diff --git a/paddle/fluid/operators/elementwise_add_op.h b/paddle/fluid/operators/elementwise_add_op.h index c60cb1f92e99329d52f6ed39dccde406a5f83563..9edbdbefe76600dc4bf937d95e70d11450206cd4 100644 --- a/paddle/fluid/operators/elementwise_add_op.h +++ b/paddle/fluid/operators/elementwise_add_op.h @@ -28,9 +28,9 @@ struct AddFunctor { }; template -void default_elementwise_add(const framework::ExecutionContext& ctx, - const framework::Tensor* x, - const framework::Tensor* y, framework::Tensor* z) { +void default_elementwise_add(const framework::ExecutionContext &ctx, + const framework::Tensor *x, + const framework::Tensor *y, framework::Tensor *z) { int axis = ctx.Attr("axis"); ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, AddFunctor(), z); @@ -40,9 +40,9 @@ template typename std::enable_if< std::is_floating_point::value && std::is_same::value>::type -elementwise_add(const framework::ExecutionContext& ctx, - const framework::Tensor* x, const framework::Tensor* y, - framework::Tensor* z) { +elementwise_add(const framework::ExecutionContext &ctx, + const framework::Tensor *x, const framework::Tensor *y, + framework::Tensor *z) { auto eigen_x = framework::EigenVector::Flatten(*x); auto eigen_y = framework::EigenVector::Flatten(*y); auto eigen_z = framework::EigenVector::Flatten(*z); @@ -55,21 +55,20 @@ template typename std::enable_if< !std::is_floating_point::value || !std::is_same::value>::type -elementwise_add(const framework::ExecutionContext& ctx, - const framework::Tensor* x, const framework::Tensor* y, - framework::Tensor* z) { +elementwise_add(const framework::ExecutionContext &ctx, + const framework::Tensor *x, const framework::Tensor *y, + framework::Tensor *z) { default_elementwise_add(ctx, x, y, z); } template class ElementwiseAddKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; + void Compute(const framework::ExecutionContext &ctx) const override { + auto *x = ctx.Input("X"); + auto *y = ctx.Input("Y"); + auto *z = ctx.Output("Out"); - const auto x = ctx.Input("X"); - const auto y = ctx.Input("Y"); - auto z = ctx.Output("Out"); z->mutable_data(ctx.GetPlace()); auto dims_equal = x->dims() == y->dims(); @@ -87,13 +86,13 @@ struct IdentityGrad { }; template -void default_elementwise_add_grad(const framework::ExecutionContext& ctx, - const framework::Tensor* x, - const framework::Tensor* y, - const framework::Tensor* out, - const framework::Tensor* dout, - framework::Tensor* dx, - framework::Tensor* dy) { +void default_elementwise_add_grad(const framework::ExecutionContext &ctx, + const framework::Tensor *x, + const framework::Tensor *y, + const framework::Tensor *out, + const framework::Tensor *dout, + framework::Tensor *dx, + framework::Tensor *dy) { int axis = ctx.Attr("axis"); ElemwiseExplicitGradCompute, @@ -106,11 +105,11 @@ template typename std::enable_if< std::is_floating_point::value && std::is_same::value>::type -elementwise_add_grad(const framework::ExecutionContext& ctx, - const framework::Tensor* x, const framework::Tensor* y, - const framework::Tensor* out, - const framework::Tensor* dout, framework::Tensor* dx, - framework::Tensor* dy) { +elementwise_add_grad(const framework::ExecutionContext &ctx, + const framework::Tensor *x, const framework::Tensor *y, + const framework::Tensor *out, + const framework::Tensor *dout, framework::Tensor *dx, + framework::Tensor *dy) { auto blas = math::GetBlas(ctx); if (dx) { @@ -128,27 +127,27 @@ template typename std::enable_if< !std::is_floating_point::value || !std::is_same::value>::type -elementwise_add_grad(const framework::ExecutionContext& ctx, - const framework::Tensor* x, const framework::Tensor* y, - const framework::Tensor* out, - const framework::Tensor* dout, framework::Tensor* dx, - framework::Tensor* dy) { +elementwise_add_grad(const framework::ExecutionContext &ctx, + const framework::Tensor *x, const framework::Tensor *y, + const framework::Tensor *out, + const framework::Tensor *dout, framework::Tensor *dx, + framework::Tensor *dy) { default_elementwise_add_grad(ctx, x, y, out, dout, dx, dy); } template class ElementwiseAddGradKernel : public ElemwiseGradKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { + void Compute(const framework::ExecutionContext &ctx) const override { ElemwiseGradKernel::Compute(ctx); using Tensor = framework::Tensor; - auto* dout = ctx.Input(framework::GradVarName("Out")); - auto* dx = ctx.Output(framework::GradVarName("X")); - auto* dy = ctx.Output(framework::GradVarName("Y")); + auto *dout = ctx.Input(framework::GradVarName("Out")); + auto *dx = ctx.Output(framework::GradVarName("X")); + auto *dy = ctx.Output(framework::GradVarName("Y")); // skip out, x, y - auto* out = dout; + auto *out = dout; auto *x = dout, *y = dout; if (platform::is_cpu_place(ctx.GetPlace()) && dx != nullptr && diff --git a/paddle/fluid/operators/elementwise_div_op.h b/paddle/fluid/operators/elementwise_div_op.h index 41a7950bf0c598507c0fda48c6a43f2fd38c41d2..cdb1264d298ef48d6b3da39d63ff1d09e1561aa4 100644 --- a/paddle/fluid/operators/elementwise_div_op.h +++ b/paddle/fluid/operators/elementwise_div_op.h @@ -28,11 +28,10 @@ template class ElementwiseDivKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); z->mutable_data(ctx.GetPlace()); int axis = ctx.Attr("axis"); ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, diff --git a/paddle/fluid/operators/elementwise_max_op.h b/paddle/fluid/operators/elementwise_max_op.h index bfb5c931958b4ca890ea720af42dad91d5625abb..367489dd563f7d8bdf430517cadf49d4ef2a0105 100644 --- a/paddle/fluid/operators/elementwise_max_op.h +++ b/paddle/fluid/operators/elementwise_max_op.h @@ -29,11 +29,10 @@ template class ElementwiseMaxKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); z->mutable_data(ctx.GetPlace()); int axis = ctx.Attr("axis"); ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, diff --git a/paddle/fluid/operators/elementwise_min_op.h b/paddle/fluid/operators/elementwise_min_op.h index db035ffb52e619b337c8190af4ed0e155aaac48d..1bd0a6279766c8eba92d1e3a76191c59410286b2 100644 --- a/paddle/fluid/operators/elementwise_min_op.h +++ b/paddle/fluid/operators/elementwise_min_op.h @@ -28,11 +28,10 @@ template class ElementwiseMinKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); z->mutable_data(ctx.GetPlace()); int axis = ctx.Attr("axis"); ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, diff --git a/paddle/fluid/operators/elementwise_mul_op.h b/paddle/fluid/operators/elementwise_mul_op.h index b870d08a1a28fd3e678aeb7211f7e3ec8b2c4c65..29e4ab7db1377b6aa80e94a26ab3cb8669f9154a 100644 --- a/paddle/fluid/operators/elementwise_mul_op.h +++ b/paddle/fluid/operators/elementwise_mul_op.h @@ -60,11 +60,10 @@ template class ElementwiseMulKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); z->mutable_data(ctx.GetPlace()); if (x->numel() == y->numel()) { elementwise_mul(ctx, x, y, z); diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h index 7e5975ead64ab39a9c618a33e300c4fce55a5b22..f01f67692e1e5dd040971cb0dd1dd793648da97a 100644 --- a/paddle/fluid/operators/elementwise_op.h +++ b/paddle/fluid/operators/elementwise_op.h @@ -13,10 +13,12 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once + #include #include "paddle/fluid/framework/data_layout.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" + #ifdef PADDLE_WITH_MKLDNN #include "paddle/fluid/platform/mkldnn_helper.h" #endif @@ -29,7 +31,8 @@ class ElementwiseOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; using Tensor = framework::Tensor; - void InferShape(framework::InferShapeContext* ctx) const override { + + void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of elementwise op should not be null."); PADDLE_ENFORCE(ctx->HasInput("Y"), @@ -37,6 +40,17 @@ class ElementwiseOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of elementwise op should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("X").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("X").front(), ctx->GetInputsVarType("X").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Y").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Y").front(), ctx->GetInputsVarType("Y").front()); + auto x_dim = ctx->GetInputDim("X"); auto y_dim = ctx->GetInputDim("Y"); PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(), @@ -47,9 +61,8 @@ class ElementwiseOp : public framework::OperatorWithKernel { } framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { - auto input_data_type = - framework::ToDataType(ctx.Input("X")->type()); + const framework::ExecutionContext &ctx) const override { + auto input_data_type = framework::GetDataTypeOfVar(ctx.InputVar("X")); #ifdef PADDLE_WITH_MKLDNN if (platform::CanMKLDNNBeUsed(ctx)) { @@ -62,16 +75,12 @@ class ElementwiseOp : public framework::OperatorWithKernel { } }; -class ElementwiseOpInferVarType : public framework::VarTypeInference { - public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto x_name = op_desc.Input("X")[0]; - auto out_name = op_desc.Output("Out")[0]; - auto& x = block->FindRecursiveOrCreateVar(x_name); - auto& out = block->FindRecursiveOrCreateVar(out_name); - out.SetType(x.GetType()); - out.SetDataType(x.GetDataType()); +class ElementwiseOpInferVarType + : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Out"}}; } }; @@ -80,8 +89,6 @@ class ElementwiseOpMaker : public framework::OpProtoAndCheckerMaker { void Make() final { AddInput("X", "(Tensor), The first input tensor of elementwise op."); AddInput("Y", "(Tensor), The second input tensor of elementwise op."); - // AddOutput("SavedShape", "(Tensor), save X, Y shape for grad to save - // memory.").AsIntermediate(); AddOutput("Out", "The output of elementwise op."); AddAttr("axis", "(int, default -1). The start dimension index " @@ -129,13 +136,12 @@ But the output only shares the LoD information with the input $X$. )DOC", GetName(), GetEquation())); - SetReuse(); } protected: virtual std::string GetName() const = 0; + virtual std::string GetEquation() const = 0; - virtual void SetReuse() {} }; class ElementwiseOpGrad : public framework::OperatorWithKernel { @@ -143,7 +149,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; using Tensor = framework::Tensor; - void InferShape(framework::InferShapeContext* ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null"); PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null"); PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), @@ -169,7 +175,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { } framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { + const framework::ExecutionContext &ctx) const override { auto input_data_type = framework::ToDataType( ctx.Input(framework::GradVarName("Out"))->type()); @@ -191,7 +197,7 @@ class ElementwiseOpExplicitGrad : public ElementwiseOpGrad { using operators::ElementwiseOpGrad::GetExpectedKernelType; using Tensor = framework::Tensor; - void InferShape(framework::InferShapeContext* ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), "Input(Out@GRAD) should not be null"); @@ -213,11 +219,11 @@ class ElementwiseOpExplicitGrad : public ElementwiseOpGrad { template class ElemwiseGradKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& context) const override { - auto* dx = + void Compute(const framework::ExecutionContext &context) const override { + auto *dx = context.Output(framework::GradVarName("X")); if (dx != nullptr) { - auto& dout = + auto &dout = *context.Input(framework::GradVarName("Out")); dx->set_lod(dout.lod()); } @@ -238,7 +244,7 @@ class ElemwiseGradKernel : public framework::OpKernel { \ protected: \ std::unique_ptr Apply() const override { \ - auto* op = new paddle::framework::OpDesc(); \ + auto *op = new paddle::framework::OpDesc(); \ op->SetType(#kernel_type "_grad"); \ op->SetInput("Y", Input("Y")); \ op->SetInput(::paddle::framework::GradVarName("Out"), \ @@ -269,7 +275,6 @@ class ElemwiseGradKernel : public framework::OpKernel { protected: \ virtual std::string GetName() const { return op_name; } \ virtual std::string GetEquation() const { return equation; } \ - virtual void SetReuse() { Reuse(__VA_ARGS__); } \ }; \ REGISTER_OPERATOR(op_type, ::paddle::operators::ElementwiseOp, \ __ElemwiseOp##op_type##Maker__, \ diff --git a/paddle/fluid/operators/elementwise_op_function.h b/paddle/fluid/operators/elementwise_op_function.h index 7c84a9d813948ab7347446872643c2e00823a5ad..93204216f947e5203863a3493005faa0c03ae4af 100644 --- a/paddle/fluid/operators/elementwise_op_function.h +++ b/paddle/fluid/operators/elementwise_op_function.h @@ -365,7 +365,7 @@ static __global__ void ElemwiseGradBroadcast1CUDAKernel( int j = blockIdx.x; int i = threadIdx.x; int tid = threadIdx.x; - T val = 0; + T val(0); do { int x_offset = i * w + j; @@ -433,7 +433,7 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel( int tid = threadIdx.x; int j = blockIdx.x; - T val = 0; + T val(0); int ttid = tid; while (true) { diff --git a/paddle/fluid/operators/elementwise_sub_op.h b/paddle/fluid/operators/elementwise_sub_op.h index 3385df0897700d37d60d8804a01db777ebc02a7e..7204c43464e0b81126148b86f64a36b0e299368b 100644 --- a/paddle/fluid/operators/elementwise_sub_op.h +++ b/paddle/fluid/operators/elementwise_sub_op.h @@ -28,11 +28,10 @@ template class ElementwiseSubKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { - using Tensor = framework::Tensor; + auto* x = ctx.Input("X"); + auto* y = ctx.Input("Y"); + auto* z = ctx.Output("Out"); - auto* x = ctx.Input("X"); - auto* y = ctx.Input("Y"); - auto* z = ctx.Output("Out"); z->mutable_data(ctx.GetPlace()); int axis = ctx.Attr("axis"); ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, diff --git a/paddle/fluid/operators/extract_rows_op.cc b/paddle/fluid/operators/extract_rows_op.cc deleted file mode 100644 index 3acae3bcdf4a509ab6e7e19f21c4b2ec4d72b7d7..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/extract_rows_op.cc +++ /dev/null @@ -1,103 +0,0 @@ -/* 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 -#include -#include "paddle/fluid/framework/op_registry.h" - -namespace paddle { -namespace operators { - -class ExtractRowsOpInferShape : public framework::InferShapeBase { - public: - void operator()(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of ExtractRowsOp should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of ExtractRowsOp should not be null."); - PADDLE_ENFORCE_EQ(ctx->GetInputsVarType("X")[0], - framework::proto::VarType::SELECTED_ROWS, - "The type of input(X) must be SelectedRows."); - auto in_dims = ctx->GetInputDim("X"); - - ctx->SetOutputDim( - "Out", framework::make_ddim(std::vector{in_dims[0], 1})); - } -}; - -class ExtractRowsOp : public framework::OperatorBase { - public: - ExtractRowsOp(const std::string &type, - const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : framework::OperatorBase(type, inputs, outputs, attrs) {} - - private: - void RunImpl(const framework::Scope &scope, - const platform::Place &place) const override { - auto &in = scope.FindVar(Input("X"))->Get(); - auto out = scope.FindVar(Output("Out"))->GetMutable(); - - auto &in_rows = in.rows(); - auto out_dim = framework::make_ddim( - std::vector{static_cast(in_rows.size()), 1}); - auto dst_ptr = out->mutable_data(out_dim, in.place()); - - if (paddle::platform::is_gpu_place(in.place())) { -#ifdef PADDLE_WITH_CUDA - platform::DeviceContextPool &pool = - platform::DeviceContextPool::Instance(); - auto *dev_ctx = pool.Get(in.place()); - auto src_ptr = in_rows.Data(in.place()); - auto stream = - reinterpret_cast(*dev_ctx) - .stream(); - memory::Copy(boost::get(out->place()), dst_ptr, - boost::get(in.place()), src_ptr, - in_rows.size() * sizeof(int64_t), stream); -#else - PADDLE_THROW("Not compiled with CUDA."); -#endif - } else { - memory::Copy(platform::CPUPlace(), dst_ptr, platform::CPUPlace(), - in_rows.data(), in_rows.size() * sizeof(int64_t)); - } - } -}; - -class ExtractRowsOpMaker : public framework::OpProtoAndCheckerMaker { - public: - void Make() override { - AddInput("X", - "(SelectedRows). The input tensor of extract_rows operator," - " and its type is SelectedRows."); - AddOutput("Out", "(Tensor). The the rows of input(X)."); - - AddComment(R"DOC( - ExtractRows Operator. - -The function of extract_rows_op is extracting the rows from the input(X) -whose type is SelectedRows. - - )DOC"); - } -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OPERATOR(extract_rows, ops::ExtractRowsOp, ops::ExtractRowsOpMaker, - ops::ExtractRowsOpInferShape); diff --git a/paddle/fluid/operators/fake_init_op.cc b/paddle/fluid/operators/fake_init_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..28ebdcb03ea83f3ec701106111a7cc5f0f7ed7dc --- /dev/null +++ b/paddle/fluid/operators/fake_init_op.cc @@ -0,0 +1,86 @@ +/* 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/framework/data_type.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +class FakeInitInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of FakeInitOp should not be null."); + auto &shape = ctx->Attrs().Get>("shape"); + ctx->SetOutputDim("Out", framework::make_ddim(shape)); + } +}; + +class FakeInitOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &dev_place) const override { + framework::Tensor *tensor = nullptr; + + auto &out_var = *scope.FindVar(Output("Out")); + + if (out_var.IsType()) { + tensor = out_var.GetMutable(); + tensor->Resize(framework::make_ddim(Attr>("shape"))); + } else if (out_var.IsType()) { + tensor = out_var.GetMutable()->mutable_value(); + tensor->Resize(framework::make_ddim(Attr>("shape"))); + } else { + PADDLE_THROW( + "fake init op's output only" + "supports SelectedRows and LoDTensor"); + } + } +}; + +class FakeInitOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override {} +}; + +class FakeInitOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddAttr>("shape", + "(vector) The shape of the output"); + AddOutput("Out", + "(Tensor) Tensor of specified shape will be filled " + "with the specified value"); + AddComment(R"DOC( +FakeInit Operator. + +Init an variable but not alloc memory for it, it is used for init the +table parameter at trainer side in distributed lookup table. + +)DOC"); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(fake_init, ops::FakeInitOp, ops::FakeInitInferShape, + ops::FakeInitOpMaker, paddle::framework::EmptyGradOpMaker, + ops::FakeInitOpVarTypeInference); diff --git a/paddle/fluid/operators/feed_op.cc b/paddle/fluid/operators/feed_op.cc index dc7ef664958238ddbd48745bd59cc7db28e49f5b..5da0a536d96e5184d51638bc6b374d2263b5e9eb 100644 --- a/paddle/fluid/operators/feed_op.cc +++ b/paddle/fluid/operators/feed_op.cc @@ -47,8 +47,8 @@ class FeedOp : public framework::OperatorBase { auto col = Attr("col"); - VLOG(3) << "Feed Var " << feed_var_name << "'s " << col << " column to var " - << out_name; + VLOG(30) << "Feed Var " << feed_var_name << "'s " << col + << " column to var " << out_name; auto &feed_list = feed_var->Get(); auto &feed_item = feed_list.at(static_cast(col)); diff --git a/paddle/fluid/operators/fetch_barrier_op.cc b/paddle/fluid/operators/fetch_barrier_op.cc index 9d7ac7ab6194593747548fac3cefc8d4ed3058d8..88a5e59ce7d6c0d14e480922bd328d632c9178e5 100644 --- a/paddle/fluid/operators/fetch_barrier_op.cc +++ b/paddle/fluid/operators/fetch_barrier_op.cc @@ -37,12 +37,13 @@ class FetchBarrierOp : public framework::OperatorBase { const platform::Place& place) const override { std::vector eps = Attr>("endpoints"); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); for (auto& ep : eps) { - VLOG(3) << "fetch barrier, ep: " << ep; + VLOG(30) << "fetch barrier, ep: " << ep; rpc_client->AsyncSendFetchBarrier(ep); } PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); @@ -61,6 +62,7 @@ This operator will send a send barrier signal to list_and_serv op, so that the Parameter Server would knew all variables have been sent. )DOC"); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr>("endpoints", "(string vector, default 127.0.0.1:6164)" "Server endpoints to send variables to.") diff --git a/paddle/fluid/operators/fetch_op.cc b/paddle/fluid/operators/fetch_op.cc index c197b45e8196a47def6465128e8ca39d8daefed6..c9e759ebff63948046e67def7fb94e0241029581 100644 --- a/paddle/fluid/operators/fetch_op.cc +++ b/paddle/fluid/operators/fetch_op.cc @@ -57,7 +57,7 @@ class FetchOp : public framework::OperatorBase { TensorCopySync(src_item, platform::CPUPlace(), &dst_item); dst_item.set_lod(src_item.lod()); - VLOG(3) << "Fetch variable " << fetch_var_name << " to " << out_name; + VLOG(30) << "Fetch variable " << fetch_var_name << " to " << out_name; } }; diff --git a/paddle/fluid/operators/fill_constant_op.cc b/paddle/fluid/operators/fill_constant_op.cc index 2826b82117db113d4d8c10095e89f610ca895775..252f313440296bd9e5eebf26f67b08bbe7decce8 100644 --- a/paddle/fluid/operators/fill_constant_op.cc +++ b/paddle/fluid/operators/fill_constant_op.cc @@ -24,7 +24,7 @@ class FillConstantInferShape : public framework::InferShapeBase { void operator()(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of FillConstantOp should not be null."); - auto &shape = ctx->Attrs().Get>("shape"); + auto &shape = ctx->Attrs().Get>("shape"); ctx->SetOutputDim("Out", framework::make_ddim(shape)); } }; @@ -47,10 +47,10 @@ class FillConstantOp : public framework::OperatorBase { if (out_var.IsType()) { tensor = out_var.GetMutable(); - tensor->Resize(framework::make_ddim(Attr>("shape"))); + tensor->Resize(framework::make_ddim(Attr>("shape"))); } else if (out_var.IsType()) { tensor = out_var.GetMutable()->mutable_value(); - tensor->Resize(framework::make_ddim(Attr>("shape"))); + tensor->Resize(framework::make_ddim(Attr>("shape"))); } else { PADDLE_THROW( "fill constant op's output only" @@ -70,6 +70,12 @@ class FillConstantOp : public framework::OperatorBase { } }; +class FillConstantOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override {} +}; + class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -77,7 +83,8 @@ class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker { "(int, default 5 (FP32)) " "Output data type") .SetDefault(framework::proto::VarType::FP32); - AddAttr>("shape", "(vector) The shape of the output"); + AddAttr>("shape", + "(vector) The shape of the output"); AddAttr("value", "(float, default 0) The value to be filled") .SetDefault(0.0f); AddAttr("force_cpu", @@ -102,4 +109,5 @@ Fill up a variable with specified constant value. namespace ops = paddle::operators; REGISTER_OPERATOR(fill_constant, ops::FillConstantOp, ops::FillConstantInferShape, ops::FillConstantOpMaker, - paddle::framework::EmptyGradOpMaker); + paddle::framework::EmptyGradOpMaker, + ops::FillConstantOpVarTypeInference); diff --git a/paddle/fluid/operators/ftrl_op.cc b/paddle/fluid/operators/ftrl_op.cc index 70ba25c213046cc934f46be067080d5fdbb42f9e..b77e12d6508eb07ae137b313ca91eac951afbcbe 100644 --- a/paddle/fluid/operators/ftrl_op.cc +++ b/paddle/fluid/operators/ftrl_op.cc @@ -34,6 +34,16 @@ class FTRLOp : public framework::OperatorWithKernel { "Input(Grad) of FTRL should not be null."); PADDLE_ENFORCE(ctx->HasInput("LearningRate"), "Input(LearningRate) of FTRL should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Grad").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Grad").front(), ctx->GetInputsVarType("Grad").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(ParamOut) of FTRL should not be null."); diff --git a/paddle/fluid/operators/ftrl_op.h b/paddle/fluid/operators/ftrl_op.h index 6f821e7e9944214fc5ebdf6bc7db8789b8ada6b9..8f812c9a037bfac8c1e29e32a5ad5b077c8153d1 100644 --- a/paddle/fluid/operators/ftrl_op.h +++ b/paddle/fluid/operators/ftrl_op.h @@ -28,6 +28,17 @@ template class FTRLOpKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + const auto* grad_var = ctx.InputVar("Grad"); + PADDLE_ENFORCE(grad_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Grad").front(), grad_var->Type().name()); + auto* param_out = ctx.Output("ParamOut"); auto* sq_accum_out = ctx.Output("SquaredAccumOut"); auto* lin_accum_out = ctx.Output("LinearAccumOut"); diff --git a/paddle/fluid/operators/fused_embedding_fc_lstm_op.cc b/paddle/fluid/operators/fused_embedding_fc_lstm_op.cc index 0b917a403620e2ffb2cbb4ca7856cce9584e1eef..fdc9cb4888b3468b85abfa0c693ed8ac5b0d450b 100644 --- a/paddle/fluid/operators/fused_embedding_fc_lstm_op.cc +++ b/paddle/fluid/operators/fused_embedding_fc_lstm_op.cc @@ -93,11 +93,7 @@ void FusedEmbeddingFCLSTMOp::InferShape( ctx->SetOutputDim("Cell", out_dims); ctx->ShareLoD("Ids", "Hidden"); ctx->ShareLoD("Ids", "Cell"); - int xx_width; - if (ctx->Attrs().Get("use_seq")) { - xx_width = wh_dims[1]; - } else { - xx_width = x_dims[1] > wh_dims[1] ? wh_dims[1] : x_dims[1]; + if (!ctx->Attrs().Get("use_seq")) { PADDLE_ENFORCE(ctx->HasOutput("BatchedInput"), "Assert only one Output(BatchedInput) of LSTM."); PADDLE_ENFORCE(ctx->HasOutput("BatchedHidden"), @@ -112,7 +108,7 @@ void FusedEmbeddingFCLSTMOp::InferShape( ctx->SetOutputDim("BatchedHidden", out_dims); ctx->SetOutputDim("BatchedCell", out_dims); } - ctx->SetOutputDim("XX", {x_dims[0], xx_width}); + ctx->SetOutputDim("XX", {x_dims[0], wh_dims[1]}); ctx->ShareLoD("Ids", "XX"); } @@ -435,8 +431,6 @@ class FusedEmbeddingFCLSTMKernel : public framework::OpKernel { INIT_VEC_FUNC INIT_BASE_INPUT_DATAS - // std::cout << "===> Batch Compute" << std::endl; - auto* reordered_h0 = ctx.Output("ReorderedH0"); auto* reordered_c0 = ctx.Output("ReorderedC0"); auto* batched_input = ctx.Output("BatchedInput"); diff --git a/paddle/fluid/operators/fusion_gru_op.cc b/paddle/fluid/operators/fusion_gru_op.cc index a04c1c1263fba659e2d3f623b607e9f476bb40ed..120b2ab440156f6020fd6005dd64a48e9a6918ec 100644 --- a/paddle/fluid/operators/fusion_gru_op.cc +++ b/paddle/fluid/operators/fusion_gru_op.cc @@ -16,10 +16,9 @@ limitations under the License. */ #include // for memcpy #include #include "paddle/fluid/operators/math/blas.h" -#include "paddle/fluid/operators/math/cpu_vec.h" #include "paddle/fluid/operators/math/fc_compute.h" +#include "paddle/fluid/operators/math/jit_kernel.h" #include "paddle/fluid/operators/math/sequence2batch.h" -#include "paddle/fluid/platform/cpu_info.h" namespace paddle { namespace operators { @@ -174,58 +173,44 @@ class FusionGRUKernel : public framework::OpKernel { } } -#define INIT_VEC_FUNC \ - std::function act_gate, act_state; \ - std::function cross; \ - auto& act_gate_str = ctx.Attr("gate_activation"); \ - auto& act_state_str = ctx.Attr("activation"); \ - if (platform::jit::MayIUse(platform::jit::avx)) { \ - math::VecActivations act_functor; \ - act_gate = act_functor(act_gate_str); \ - act_state = act_functor(act_state_str); \ - cross = math::vec_cross; \ - } else { \ - math::VecActivations act_functor; \ - act_gate = act_functor(act_gate_str); \ - act_state = act_functor(act_state_str); \ - cross = math::vec_cross; \ - } - -#define INIT_BASE_INPUT_OUTPUT \ - auto* h0 = ctx.Input("H0"); \ - auto* wx = ctx.Input("WeightX"); \ - auto* wh = ctx.Input("WeightH"); \ - auto* bias = ctx.Input("Bias"); \ - auto* xx = ctx.Output("XX"); \ - auto* hidden_out = ctx.Output("Hidden"); \ - bool is_reverse = ctx.Attr("is_reverse"); - -#define INIT_BASE_SIZES \ - auto x_dims = x->dims(); /* T x M*/ \ - auto wh_dims = wh->dims(); /* D x 3D*/ \ - const int total_T = x_dims[0]; \ - const int M = x_dims[1]; \ - const int D = wh_dims[0]; \ - const int D3 = wh_dims[1]; \ - const int D2 = D * 2; +#define INIT_BASE_DEFINES \ + auto* x = ctx.Input("X"); \ + auto* wh = ctx.Input("WeightH"); \ + auto* xx = ctx.Output("XX"); \ + auto x_lod = x->lod(); \ + auto x_dims = x->dims(); /* T x M*/ \ + auto wh_dims = wh->dims(); /* D x 3D*/ \ + const int total_T = x_dims[0]; \ + const int D3 = wh_dims[1] + +#define INIT_OTHER_DEFINES \ + auto* h0 = ctx.Input("H0"); \ + auto* wx = ctx.Input("WeightX"); \ + auto* bias = ctx.Input("Bias"); \ + auto* hidden_out = ctx.Output("Hidden"); \ + bool is_reverse = ctx.Attr("is_reverse"); \ + const int M = x_dims[1]; \ + const int D = wh_dims[0]; \ + const int D2 = D * 2; \ + const auto& ker = math::jitkernel::KernelPool::Instance() \ + .template Get, \ + const std::string&, const std::string&>( \ + ctx.Attr("gate_activation"), \ + ctx.Attr("activation"), D); \ + const T* x_data = x->data(); \ + const T* wx_data = wx->data(); \ + const T* wh_data = wh->data(); \ + auto place = ctx.GetPlace(); \ + T* xx_data = xx->mutable_data(place) void SeqCompute(const framework::ExecutionContext& ctx) const { using DeviceContext = paddle::platform::CPUDeviceContext; - auto* x = ctx.Input("X"); - INIT_BASE_INPUT_OUTPUT - INIT_BASE_SIZES - INIT_VEC_FUNC - - auto x_lod = x->lod(); + INIT_BASE_DEFINES; + INIT_OTHER_DEFINES; const int N = x_lod[0].size() - 1; - const T* x_data = x->data(); const T* h0_data = h0 ? h0->data() : nullptr; - const T* wx_data = wx->data(); - const T* wh_data = wh->data(); const T* wh_state_data = wh_data + D * D2; - T* xx_data = xx->mutable_data(ctx.GetPlace()); - T* hidden_out_data = hidden_out->mutable_data(ctx.GetPlace()); - + T* hidden_out_data = hidden_out->mutable_data(place); auto blas = math::GetBlas(ctx); math::FCCompute(blas, total_T, D3, M, x_data, wx_data, xx_data, @@ -252,14 +237,7 @@ class FusionGRUKernel : public framework::OpKernel { if (h0_data) { prev_hidden_data = h0_data + bid * D; } else { - // W: {W_update, W_reset; W_state} - // update gate - act_gate(D, xx_data, xx_data); - // state gate - act_state(D, xx_data + D2, xx_data + D2); - // out = a*b - blas.VMUL(D, xx_data, xx_data + D2, hidden_out_data); - // save prev + ker->ComputeH1(xx_data, hidden_out_data); prev_hidden_data = hidden_out_data; tstart = 1; move_step(); @@ -269,17 +247,12 @@ class FusionGRUKernel : public framework::OpKernel { blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D2, D, static_cast(1), prev_hidden_data, D, wh_data, D2, static_cast(1), xx_data, D3); - act_gate(D2, xx_data, xx_data); - // rt = rt*ht_1 inplace result - blas.VMUL(D, prev_hidden_data, xx_data + D, hidden_out_data); - + ker->ComputeHtPart1(xx_data, prev_hidden_data, hidden_out_data); // gemm rt * Ws blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D, D, static_cast(1), hidden_out_data, D, wh_state_data, D, static_cast(1), xx_data + D2, D3); - act_state(D, xx_data + D2, xx_data + D2); - // out = zt*ht~ + (1-zt)*ht_1 - cross(D, xx_data, xx_data + D2, prev_hidden_data, hidden_out_data); + ker->ComputeHtPart2(xx_data, prev_hidden_data, hidden_out_data); // save prev prev_hidden_data = hidden_out_data; move_step(); @@ -289,28 +262,19 @@ class FusionGRUKernel : public framework::OpKernel { void BatchCompute(const framework::ExecutionContext& ctx) const { using DeviceContext = paddle::platform::CPUDeviceContext; - auto* x = ctx.Input("X"); - INIT_BASE_INPUT_OUTPUT - INIT_BASE_SIZES - if (x->lod()[0].size() == 2) { + INIT_BASE_DEFINES; + if (x_lod[0].size() == 2) { xx->Resize({total_T, D3}); SeqCompute(ctx); return; } - INIT_VEC_FUNC - + INIT_OTHER_DEFINES; auto* reordered_h0 = ctx.Output("ReorderedH0"); auto* batched_input = ctx.Output("BatchedInput"); auto* batched_out = ctx.Output("BatchedOut"); - - const T* x_data = x->data(); - const T* wx_data = wx->data(); - const T* wh_data = wh->data(); - T* xx_data = xx->mutable_data(ctx.GetPlace()); - T* batched_input_data = batched_input->mutable_data(ctx.GetPlace()); - T* batched_out_data = batched_out->mutable_data(ctx.GetPlace()); - hidden_out->mutable_data(ctx.GetPlace()); - + T* batched_input_data = batched_input->mutable_data(place); + T* batched_out_data = batched_out->mutable_data(place); + hidden_out->mutable_data(place); auto& dev_ctx = ctx.template device_context(); auto blas = math::GetBlas(dev_ctx); math::LoDTensor2BatchFunctor to_batch; @@ -336,7 +300,7 @@ class FusionGRUKernel : public framework::OpKernel { T* prev_hidden_data = nullptr; if (h0) { // reorder h0 - T* reordered_h0_data = reordered_h0->mutable_data(ctx.GetPlace()); + T* reordered_h0_data = reordered_h0->mutable_data(place); const T* h0_data = h0->data(); prev_hidden_data = reordered_h0_data; size_t sz = sizeof(T) * D; @@ -350,12 +314,7 @@ class FusionGRUKernel : public framework::OpKernel { T* cur_out_data = batched_out_data; // W: {W_update, W_reset; W_state} for (int i = 0; i < max_bs; ++i) { - // update gate - act_gate(D, cur_in_data, cur_in_data); - // state gate - act_state(D, cur_in_data + D2, cur_in_data + D2); - // out = a*b - blas.VMUL(D, cur_in_data, cur_in_data + D2, cur_out_data); + ker->ComputeH1(cur_in_data, cur_out_data); // add offset cur_in_data += D3; cur_out_data += D; @@ -380,10 +339,8 @@ class FusionGRUKernel : public framework::OpKernel { T* cur_out_data = batched_out_data; T* cur_prev_hidden_data = prev_hidden_data; for (int i = 0; i < cur_bs; ++i) { - act_gate(D2, cur_batched_data, cur_batched_data); - // rt = rt*ht_1 inplace result - blas.VMUL(D, cur_prev_hidden_data, cur_batched_data + D, cur_out_data); - + ker->ComputeHtPart1(cur_batched_data, cur_prev_hidden_data, + cur_out_data); cur_batched_data += D3; cur_prev_hidden_data += D; cur_out_data += D; @@ -397,12 +354,8 @@ class FusionGRUKernel : public framework::OpKernel { cur_prev_hidden_data = prev_hidden_data; for (int i = 0; i < cur_bs; ++i) { - // ht~ = act_state(...) - act_state(D, cur_batched_data + D2, cur_batched_data + D2); - // out = zt*ht~ + (1-zt)*ht_1 - cross(D, cur_batched_data, cur_batched_data + D2, cur_prev_hidden_data, - cur_out_data); - + ker->ComputeHtPart2(cur_batched_data, cur_prev_hidden_data, + cur_out_data); cur_batched_data += D3; cur_prev_hidden_data += D; cur_out_data += D; @@ -416,9 +369,8 @@ class FusionGRUKernel : public framework::OpKernel { batched_out->set_lod(batched_lod); to_seq(dev_ctx, *batched_out, hidden_out); } -#undef INIT_VEC_FUNC -#undef INIT_BASE_SIZES -#undef INIT_BASE_INPUT_OUTPUT +#undef INIT_OTHER_DEFINES +#undef INIT_BASE_DEFINES }; } // namespace operators diff --git a/paddle/fluid/operators/fusion_lstm_op.cc b/paddle/fluid/operators/fusion_lstm_op.cc index ae1f6d8e489039667d861a69acabf2c632ef2061..067e6a3e7cccc1f15ebdd984f3a2441339a989ab 100644 --- a/paddle/fluid/operators/fusion_lstm_op.cc +++ b/paddle/fluid/operators/fusion_lstm_op.cc @@ -15,11 +15,9 @@ limitations under the License. */ #include "paddle/fluid/operators/fusion_lstm_op.h" #include #include "paddle/fluid/operators/math/blas.h" -#include "paddle/fluid/operators/math/cpu_lstm_compute.h" -#include "paddle/fluid/operators/math/cpu_vec.h" #include "paddle/fluid/operators/math/fc_compute.h" +#include "paddle/fluid/operators/math/jit_kernel.h" #include "paddle/fluid/operators/math/sequence2batch.h" -#include "paddle/fluid/platform/cpu_info.h" namespace paddle { namespace operators { @@ -219,121 +217,55 @@ This operator fuse the X into LSTM, more details can refer to LSTM op. template class FuisonLSTMKernel : public framework::OpKernel { public: -#define INIT_VEC_FUNC \ - std::function act_gate, act_cell, act_cand; \ - auto& act_gate_str = ctx.Attr("gate_activation"); \ - auto& act_cell_str = ctx.Attr("cell_activation"); \ - auto& act_cand_str = ctx.Attr("candidate_activation"); \ - if (platform::jit::MayIUse(platform::jit::avx)) { \ - math::VecActivations act_functor; \ - act_gate = act_functor(act_gate_str); \ - act_cell = act_functor(act_cell_str); \ - act_cand = act_functor(act_cand_str); \ - } else { \ - math::VecActivations act_functor; \ - act_gate = act_functor(act_gate_str); \ - act_cell = act_functor(act_cell_str); \ - act_cand = act_functor(act_cand_str); \ - } - -#define INIT_BASE_INPUT_OUTPUT \ - auto* x = ctx.Input("X"); \ - auto* h0 = ctx.Input("H0"); \ - auto* c0 = ctx.Input("C0"); \ - auto* wx = ctx.Input("WeightX"); \ - auto* wh = ctx.Input("WeightH"); \ - auto* bias = ctx.Input("Bias"); \ - auto* xx = ctx.Output("XX"); \ - auto* hidden_out = ctx.Output("Hidden"); \ - auto* cell_out = ctx.Output("Cell"); \ - bool is_reverse = ctx.Attr("is_reverse"); \ - bool use_peepholes = ctx.Attr("use_peepholes"); - -#define INIT_BASE_SIZES \ - auto x_dims = x->dims(); /* T x M*/ \ - auto wh_dims = wh->dims(); /* D x 4D*/ \ - const int M = x_dims[1]; \ - const int D = wh_dims[0]; \ - const int D2 = D * 2; \ - const int D3 = D * 3; \ - const int D4 = wh_dims[1]; - -#define INIT_BASE_INPUT_DATAS \ - const T* x_data = x->data(); \ - const T* wx_data = wx->data(); \ - const T* wh_data = wh->data(); \ - /* diagonal weight*/ \ - const T* wc_data = bias->data() + D4; \ - /* for peephole only*/ \ - T* checked_cell_data = nullptr; \ - auto place = ctx.GetPlace(); \ - if (use_peepholes) { \ - /* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/ \ - auto* checked_cell = ctx.Output("CheckedCell"); \ - checked_cell_data = checked_cell->mutable_data(place); \ - } - -/// Compute LSTM +#define INIT_BASE_DEFINES \ + using DeviceContext = paddle::platform::CPUDeviceContext; \ + auto* x = ctx.Input("X"); \ + auto* h0 = ctx.Input("H0"); \ + auto* c0 = ctx.Input("C0"); \ + auto* wx = ctx.Input("WeightX"); \ + auto* wh = ctx.Input("WeightH"); \ + auto* bias = ctx.Input("Bias"); \ + auto* xx = ctx.Output("XX"); \ + auto* hidden_out = ctx.Output("Hidden"); \ + auto* cell_out = ctx.Output("Cell"); \ + bool is_reverse = ctx.Attr("is_reverse"); \ + bool use_peepholes = ctx.Attr("use_peepholes"); \ + auto x_dims = x->dims(); /* T x M*/ \ + auto wh_dims = wh->dims(); /* D x 4D*/ \ + const int M = x_dims[1]; \ + const int D = wh_dims[0]; \ + const int D4 = wh_dims[1] + +#define INIT_OTHER_DEFINES \ + const T* x_data = x->data(); \ + const T* wx_data = wx->data(); \ + const T* wh_data = wh->data(); \ + /* diagonal weight*/ \ + const T* wp_data = bias->data() + D4; \ + /* for peephole only*/ \ + T* checked_cell_data = nullptr; \ + auto place = ctx.GetPlace(); \ + if (use_peepholes) { \ + /* w_ic * Ct-1, w_fc * Ct-1 ; w_oc * Ct => ih*/ \ + auto* checked_cell = ctx.Output("CheckedCell"); \ + checked_cell_data = checked_cell->mutable_data(place); \ + } \ + const auto& ker = \ + math::jitkernel::KernelPool::Instance() \ + .template Get, const std::string&, \ + const std::string&, const std::string&>( \ + ctx.Attr("gate_activation"), \ + ctx.Attr("candidate_activation"), \ + ctx.Attr("cell_activation"), D, use_peepholes) + +// Wh GEMM #define GEMM_WH_ADDON(bs, prev, out) \ blas.GEMM(CblasNoTrans, CblasNoTrans, bs, D4, D, static_cast(1), prev, D, \ wh_data, D4, static_cast(1), out, D4) -#define GET_Ct(ct_1, gates, ct) \ - /* C_t = C_t-1 * fgated + cand_gated * igated*/ \ - act_cand(D, gates, gates); \ - blas.VMUL(D, gates, gates + D, gates + D); \ - blas.VMUL(D, ct_1, gates + D2, gates + D2); \ - blas.VADD(D, gates + D, gates + D2, ct) - -#define GET_Ht(ct, gates, ht) \ - /* H_t = act_cell(C_t) * ogated */ \ - act_cell(D, ct, gates + D2); \ - blas.VMUL(D, gates + D2, gates + D3, ht) - -#define GET_Ct_NOH0C0(gates, ct) \ - /* C_t = igated * cgated*/ \ - act_gate(D, gates + D, gates + D); \ - act_cand(D, gates, gates); \ - blas.VMUL(D, gates, gates + D, ct) - -#define COMPUTE_CtHt_NOH0C0(gates, ct, ht) \ - GET_Ct_NOH0C0(gates, ct); \ - act_gate(D, gates + D3, gates + D3); \ - GET_Ht(ct, gates, ht) - -#define COMPUTE_CtHt_PEEPHOLE_NOH0C0(gates, ct, ht) \ - GET_Ct_NOH0C0(gates, ct); \ - /* get outgated, put W_oc * C_t on igated */ \ - blas.VMUL(D, wc_data + D2, ct, gates + D); \ - blas.VADD(D, gates + D, gates + D3, gates + D3); \ - act_gate(D, gates + D3, gates + D3); \ - GET_Ht(ct, gates, ht) - -#define COMPUTE_CtHt(gates, ct_1, ct, ht) \ - act_gate(D3, gates + D, gates + D); \ - GET_Ct(ct_1, gates, ct); \ - GET_Ht(ct, gates, ht) - -#define COMPUTE_CtHt_PEEPHOLE(gates, ct_1, ct, ht) \ - /* get fgated and igated*/ \ - blas.VMUL(D, wc_data, ct_1, checked_cell_data); \ - blas.VMUL(D, wc_data + D, ct_1, checked_cell_data + D); \ - blas.VADD(D2, checked_cell_data, gates + D, gates + D); \ - act_gate(D2, gates + D, gates + D); \ - GET_Ct(ct_1, gates, ct); \ - /* get ogated*/ \ - blas.VMUL(D, wc_data + D2, ct, gates + D); \ - blas.VADD(D, gates + D, gates + D3, gates + D3); \ - act_gate(D, gates + D3, gates + D3); \ - GET_Ht(ct, gates, ht) - void SeqCompute(const framework::ExecutionContext& ctx) const { - using DeviceContext = paddle::platform::CPUDeviceContext; - INIT_BASE_INPUT_OUTPUT - INIT_BASE_SIZES - INIT_VEC_FUNC - INIT_BASE_INPUT_DATAS - + INIT_BASE_DEFINES; + INIT_OTHER_DEFINES; auto x_lod = x->lod(); const int total_T = x_dims[0]; const int N = x_lod[0].size() - 1; @@ -357,89 +289,47 @@ class FuisonLSTMKernel : public framework::OpKernel { gate_offset = -D; } -#define MOVE_ONE_STEP \ - prev_h_data = h_out_data; \ - prev_c_data = c_out_data; \ - xx_data = xx_data + xx_offset; \ - h_out_data = h_out_data + gate_offset; \ - c_out_data = c_out_data + gate_offset - -#define PROCESS_H0C0_DEFINES \ - int bid = is_reverse ? N - 1 - i : i; \ - int seq_len = x_lod[0][bid + 1] - x_lod[0][bid]; \ - const T* prev_c_data = nullptr; \ - const T* prev_h_data = nullptr; \ - int tstart = 0 - -#define PROCESS_H0C0_PEEPHOLE \ - PROCESS_H0C0_DEFINES; \ - if (h0_data) { \ - prev_h_data = h0_data + bid * D; \ - prev_c_data = c0_data + bid * D; \ - } else { \ - COMPUTE_CtHt_PEEPHOLE_NOH0C0(xx_data, c_out_data, h_out_data); \ - MOVE_ONE_STEP; \ - tstart = 1; \ - } - -#define PROCESS_H0C0 \ - PROCESS_H0C0_DEFINES; \ - if (h0_data) { \ - prev_h_data = h0_data + bid * D; \ - prev_c_data = c0_data + bid * D; \ - } else { \ - COMPUTE_CtHt_NOH0C0(xx_data, c_out_data, h_out_data); \ - MOVE_ONE_STEP; \ - tstart = 1; \ - } - - if (use_peepholes) { - for (int i = 0; i < N; ++i) { - PROCESS_H0C0_PEEPHOLE - for (int step = tstart; step < seq_len; ++step) { - GEMM_WH_ADDON(1, prev_h_data, xx_data); - COMPUTE_CtHt_PEEPHOLE(xx_data, prev_c_data, c_out_data, h_out_data); - MOVE_ONE_STEP; - } - } - } else { - // TODO(TJ): unly workaround, clean me - std::function compute_ctht; - if (platform::jit::MayIUse(platform::jit::avx) && - act_gate_str == "sigmoid" && act_cand_str == "tanh" && - act_cell_str == "tanh" && D == 8) { - compute_ctht = math::lstm_compute_ctht; + for (int i = 0; i < N; ++i) { + int bid = is_reverse ? N - 1 - i : i; + int seq_len = x_lod[0][bid + 1] - x_lod[0][bid]; + const T* prev_c_data = nullptr; + const T* prev_h_data = nullptr; + int tstart = 0; + if (h0_data) { + prev_h_data = h0_data + bid * D; + prev_c_data = c0_data + bid * D; } else { - compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) { - COMPUTE_CtHt(gates, ct_1, ct, ht); - }; + ker->ComputeC1H1(xx_data, c_out_data, h_out_data, wp_data); + tstart = 1; + // move one step + prev_h_data = h_out_data; + prev_c_data = c_out_data; + xx_data = xx_data + xx_offset; + h_out_data = h_out_data + gate_offset; + c_out_data = c_out_data + gate_offset; } - for (int i = 0; i < N; ++i) { - PROCESS_H0C0 - for (int step = tstart; step < seq_len; ++step) { - GEMM_WH_ADDON(1, prev_h_data, xx_data); - compute_ctht(xx_data, prev_c_data, c_out_data, h_out_data); - MOVE_ONE_STEP; - } + for (int step = tstart; step < seq_len; ++step) { + GEMM_WH_ADDON(1, prev_h_data, xx_data); + ker->ComputeCtHt(xx_data, prev_c_data, c_out_data, h_out_data, wp_data, + checked_cell_data); + // move one step + prev_h_data = h_out_data; + prev_c_data = c_out_data; + xx_data = xx_data + xx_offset; + h_out_data = h_out_data + gate_offset; + c_out_data = c_out_data + gate_offset; } } -#undef PROCESS_H0C0_DEFINES -#undef PROCESS_H0C0_PEEPHOLE -#undef PROCESS_H0C0 -#undef MOVE_ONE_STEP } void BatchCompute(const framework::ExecutionContext& ctx) const { - using DeviceContext = platform::CPUDeviceContext; - INIT_BASE_INPUT_OUTPUT - INIT_BASE_SIZES + INIT_BASE_DEFINES; if (x->lod()[0].size() == 2) { xx->Resize({x_dims[0], D4}); SeqCompute(ctx); return; } - INIT_VEC_FUNC - INIT_BASE_INPUT_DATAS + INIT_OTHER_DEFINES; auto* reordered_h0 = ctx.Output("ReorderedH0"); auto* reordered_c0 = ctx.Output("ReorderedC0"); @@ -487,8 +377,8 @@ class FuisonLSTMKernel : public framework::OpKernel { prev_c_data = reordered_c0_data; size_t sz = sizeof(T) * D; for (int i = 0; i < max_bs; ++i) { - std::memcpy(reordered_h0_data, h0_data + seq_order[i] * D, sz); - std::memcpy(reordered_c0_data, c0_data + seq_order[i] * D, sz); + blas.VCOPY(sz, h0_data + seq_order[i] * D, reordered_h0_data); + blas.VCOPY(sz, c0_data + seq_order[i] * D, reordered_c0_data); reordered_h0_data += D; reordered_c0_data += D; } @@ -498,13 +388,7 @@ class FuisonLSTMKernel : public framework::OpKernel { T* cur_h_out_data = batched_h_out_data; T* cur_c_out_data = batched_c_out_data; for (int i = 0; i < max_bs; ++i) { - GET_Ct_NOH0C0(cur_in_data, cur_c_out_data); - if (use_peepholes) { - blas.VMUL(D, wc_data + D2, cur_c_out_data, cur_in_data + D); - blas.VADD(D, cur_in_data + D, cur_in_data + D3, cur_in_data + D3); - } - act_gate(D, cur_in_data + D3, cur_in_data + D3); - GET_Ht(cur_c_out_data, cur_in_data, cur_h_out_data); + ker->ComputeC1H1(cur_in_data, cur_c_out_data, cur_h_out_data, wp_data); cur_in_data += D4; cur_c_out_data += D; cur_h_out_data += D; @@ -513,71 +397,37 @@ class FuisonLSTMKernel : public framework::OpKernel { prev_h_data = batched_h_out_data; prev_c_data = batched_c_out_data; } + + // compute kernel part const auto& batch_starts = batched_lod[0]; const int max_seq_len = batch_starts.size() - 1; const int offset = tstart * max_bs * D; batched_input_data = batched_input_data + offset * 4; batched_h_out_data = batched_h_out_data + offset; batched_c_out_data = batched_c_out_data + offset; - -#define DEFINE_CUR \ - T* cur_in_data = batched_input_data; \ - T* cur_prev_c_data = prev_c_data; \ - T* cur_c_out_data = batched_c_out_data; \ - T* cur_h_out_data = batched_h_out_data - -#define MOVE_ONE_BATCH \ - cur_in_data += D4; \ - cur_prev_c_data += D; \ - cur_c_out_data += D; \ - cur_h_out_data += D - -#define MOVE_ONE_STEP \ - prev_c_data = batched_c_out_data; \ - prev_h_data = batched_h_out_data; \ - batched_c_out_data = cur_c_out_data; \ - batched_h_out_data = cur_h_out_data; \ - batched_input_data = cur_in_data - - if (use_peepholes) { - for (int step = tstart; step < max_seq_len; ++step) { - const int cur_bs = batch_starts[step + 1] - batch_starts[step]; - GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data); - DEFINE_CUR; - for (int i = 0; i < cur_bs; ++i) { - COMPUTE_CtHt_PEEPHOLE(cur_in_data, cur_prev_c_data, cur_c_out_data, - cur_h_out_data); - MOVE_ONE_BATCH; - } - MOVE_ONE_STEP; - } - } else { - // TODO(TJ): unly workaround, clean me - std::function compute_ctht; - if (platform::jit::MayIUse(platform::jit::avx) && - act_gate_str == "sigmoid" && act_cand_str == "tanh" && - act_cell_str == "tanh" && D == 8) { - compute_ctht = math::lstm_compute_ctht; - } else { - compute_ctht = [&](T* gates, const T* ct_1, T* ct, T* ht) { - COMPUTE_CtHt(gates, ct_1, ct, ht); - }; - } - for (int step = tstart; step < max_seq_len; ++step) { - const int cur_bs = batch_starts[step + 1] - batch_starts[step]; - GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data); - DEFINE_CUR; - for (int i = 0; i < cur_bs; ++i) { - compute_ctht(cur_in_data, cur_prev_c_data, cur_c_out_data, - cur_h_out_data); - MOVE_ONE_BATCH; - } - MOVE_ONE_STEP; + for (int step = tstart; step < max_seq_len; ++step) { + const int cur_bs = batch_starts[step + 1] - batch_starts[step]; + GEMM_WH_ADDON(cur_bs, prev_h_data, batched_input_data); + T* cur_in_data = batched_input_data; + T* cur_prev_c_data = prev_c_data; + T* cur_c_out_data = batched_c_out_data; + T* cur_h_out_data = batched_h_out_data; + for (int i = 0; i < cur_bs; ++i) { + ker->ComputeCtHt(cur_in_data, cur_prev_c_data, cur_c_out_data, + cur_h_out_data, wp_data, checked_cell_data); + // move one batch + cur_in_data += D4; + cur_prev_c_data += D; + cur_c_out_data += D; + cur_h_out_data += D; } + // move one step + prev_c_data = batched_c_out_data; + prev_h_data = batched_h_out_data; + batched_c_out_data = cur_c_out_data; + batched_h_out_data = cur_h_out_data; + batched_input_data = cur_in_data; } -#undef MOVE_ONE_STEP -#undef MOVE_ONE_BATCH -#undef DEFINE_CUR math::Batch2LoDTensorFunctor to_seq; batched_h_out->set_lod(batched_lod); @@ -594,18 +444,9 @@ class FuisonLSTMKernel : public framework::OpKernel { } } -#undef COMPUTE_CtHt_PEEPHOLE -#undef COMPUTE_CtHt -#undef GET_Ct_NOH0C0 -#undef COMPUTE_CtHt_NOH0C0 -#undef COMPUTE_CtHt_PEEPHOLE_NOH0C0 -#undef GET_Ht -#undef GET_Ct #undef GEMM_WH_ADDON -#undef INIT_BASE_INPUT_DATAS -#undef INIT_BASE_SIZES -#undef INIT_BASE_INPUT_OUTPUT -#undef INIT_VEC_FUNC +#undef INIT_OTHER_DEFINES +#undef INIT_BASE_DEFINES }; } // namespace operators diff --git a/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc b/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..b0910dc19edb246d9acfe3bdb15071c64cbdaba7 --- /dev/null +++ b/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.cc @@ -0,0 +1,229 @@ +/* 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/fusion_seqconv_eltadd_relu_op.h" +#include // for min, max +#include +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/fc_compute.h" + +namespace paddle { +namespace operators { + +void FusionSeqConvEltAddReluOp::InferShape( + framework::InferShapeContext* ctx) const { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of FusionSeqConvEltAddReluOp should not be null."); + PADDLE_ENFORCE( + ctx->HasInput("Filter"), + "Input(Filter) of FusionSeqConvEltAddReluOp should not be null."); + PADDLE_ENFORCE( + ctx->HasInput("Bias"), + "Input(Bias) of FusionSeqConvEltAddReluOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("Out"), + "Output(Out) of FusionSeqConvEltAddReluOp should not be null."); + PADDLE_ENFORCE( + ctx->HasOutput("ColMat"), + "Output(ColMat) of FusionSeqConvEltAddReluOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto w_dims = ctx->GetInputDim("Filter"); + int context_length = ctx->Attrs().Get("contextLength"); + PADDLE_ENFORCE( + ctx->Attrs().Get("contextStride") == 1, + "Currently, FusionSeqConvEltAddReluOp only supports contextStride=1."); + PADDLE_ENFORCE(x_dims.size() == 2 && w_dims.size() == 2, + "Input(X, Filter) should be 2-D tensor."); + PADDLE_ENFORCE(x_dims.size() == 2 && w_dims.size() == 2, + "Input(X, Filter) should be 2-D tensor."); + PADDLE_ENFORCE(w_dims[0] == context_length * x_dims[1], + "Filter's height should be context_length * " + "input_hidden_size ."); + PADDLE_ENFORCE_GT(context_length + ctx->Attrs().Get("contextStart"), 0, + "contextStart size should be smaller than contextLength."); + + ctx->SetOutputDim("Out", {x_dims[0], w_dims[1]}); + ctx->SetOutputDim("ColMat", {x_dims[0], w_dims[0]}); + ctx->ShareLoD("X", "Out"); +} + +framework::OpKernelType FusionSeqConvEltAddReluOp::GetExpectedKernelType( + const framework::ExecutionContext& ctx) const { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); +} + +void FusionSeqConvEltAddReluOpMaker::Make() { + AddInput("X", + "(LoDTensor) the input is a LodTensor, which support " + "variable-time length input sequence. The underlying tensor in " + "this LoDTensor is a matrix with shape (T X M), where T is the " + "total time steps in this mini-batch, M is the dim size of x."); + // PaddingData only support false yet, should be ensured at pass. + AddInput("Filter", + "(Tensor) same as the input(Filter) of sequence conv op is an " + "learnable parameter." + "This is a tensor with shape (K, N), where K is the " + "context_length * dim size of x, N is the output feature size."); + AddInput("Bias", + "(Tensor) the learnable weights. shape (1, N), where N is the " + "output feature size"); + AddOutput( + "Out", + "(LoDTensor) the output(Out) is a LodTensor, which support " + "variable-time length output sequence. The underlying tensor in " + "this LoDTensor is a matrix with shape (T, N), where, T is the " + "total time steps in this mini-batch, N is the output feature size."); + AddOutput("ColMat", + "(Tensor) (T, K), where T is where T is the " + "total time steps in this mini-batch, K is height of Filter") + .AsIntermediate(); + AddAttr("contextLength", + "(int) the contextLength of FusionSeqConvEltAddReluOp is the " + "height of the convolution kernel.") + .GreaterThan(0); + AddAttr("contextStart", + "(int, default:0) the contextStart of FusionSeqConvEltAddReluOp " + "represents the beginning of the convolution of the number of " + "rows of sequence, which can be negative. The negative number " + "means to pad contextStart time-steps of zeros or learnable " + "parameters at the beginning of each instance. The positive " + "number means to skip contextStart time-steps of each " + "instance.") + .SetDefault(0); + AddAttr( + "contextStride", + "(int, default:1) the contextStride of FusionSeqConvEltAddReluOp " + "represents the stride length of convolution kernel. " + "Currently, FusionSeqConvEltAddReluOp only supports" + "contextStride=1.") + .SetDefault(1) + .GreaterThan(0); + AddComment(R"DOC( +Fusion Sequence Conv and ElementwiseAdd Operator. +)DOC"); +} + +template +class FusionSeqConvEltAddReluKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + using DeviceContext = paddle::platform::CPUDeviceContext; + auto* x = ctx.Input("X"); + auto* w = ctx.Input("Filter"); + auto* b = ctx.Input("Bias"); + auto* y = ctx.Output("Out"); + auto* col = ctx.Output("ColMat"); + + auto x_lod = x->lod(); + auto x_dims = x->dims(); + auto w_dims = w->dims(); + PADDLE_ENFORCE_EQ(b->numel(), w_dims[1], + "bias size should be equal to output feature size."); + PADDLE_ENFORCE_EQ(x_lod.size(), 1UL, + "Only support one level sequence now."); + + const T* x_data = x->data(); + const T* w_data = w->data(); + const T* b_data = b->data(); + T* y_data = y->mutable_data(ctx.GetPlace()); + T* col_data = col->mutable_data(ctx.GetPlace()); + + int context_start = ctx.Attr("contextStart"); + int context_length = ctx.Attr("contextLength"); + int up_pad = std::max(0, -context_start); + int down_pad = std::max(0, context_start + context_length - 1); + // im2col + int src_mat_w = static_cast(x_dims[1]); + int src_mat_w_sz = src_mat_w * sizeof(T); + int col_mat_w = static_cast(w_dims[0]); + int col_mat_w_sz = col_mat_w * sizeof(T); + for (int i = 0; i < static_cast(x_lod[0].size()) - 1; ++i) { + int st = x_lod[0][i]; + int ed = x_lod[0][i + 1]; + const T* src_data = x_data + st * src_mat_w; + T* dst_data = col_data + st * col_mat_w; + int seq_len = ed - st; + if (seq_len > up_pad + down_pad) { + // zero all up_pad and fill data + std::memset(dst_data, 0, up_pad * col_mat_w_sz); + dst_data = dst_data + up_pad * src_mat_w; + int copy_size = col_mat_w_sz - up_pad * src_mat_w_sz; + for (int j = 0; j < up_pad; ++j) { + // blas.VCOPY? + std::memcpy(dst_data, src_data, copy_size); + dst_data += (col_mat_w - src_mat_w); + copy_size += src_mat_w_sz; + } + // fill data + for (int j = 0; j < seq_len - up_pad - down_pad; ++j) { + std::memcpy(dst_data, src_data, copy_size); + dst_data += col_mat_w; + src_data += src_mat_w; + } + // zero all down_pad and fill data + std::memset(dst_data, 0, down_pad * col_mat_w_sz); + copy_size -= src_mat_w_sz; + for (int j = 0; j < down_pad; ++j) { + std::memcpy(dst_data, src_data, copy_size); + dst_data += col_mat_w; + src_data += src_mat_w; + copy_size -= src_mat_w_sz; + } + } else { + PADDLE_ENFORCE_GE(context_length, up_pad + down_pad + 1); + std::memset(dst_data, 0, seq_len * col_mat_w_sz); + dst_data = dst_data + up_pad * src_mat_w; + int zero_sz = up_pad * src_mat_w_sz; + int cur_src_sz = seq_len * src_mat_w_sz; + for (int j = 0; j < std::min(up_pad, seq_len); ++j) { + int copy_size = std::min(cur_src_sz, col_mat_w_sz - zero_sz); + std::memcpy(dst_data, src_data, copy_size); + dst_data += (col_mat_w - src_mat_w); + zero_sz -= src_mat_w_sz; + } + // from bottom + dst_data = col_data + ed * col_mat_w; + src_data = x_data + st * src_mat_w; + zero_sz = down_pad * src_mat_w_sz; + for (int j = 1; j <= std::min(down_pad, seq_len); ++j) { + int copy_size = std::min(cur_src_sz, col_mat_w_sz - zero_sz); + std::memcpy(dst_data - (zero_sz + copy_size) / sizeof(T), + src_data + std::max(seq_len - j - up_pad, 0) * src_mat_w, + copy_size); + dst_data -= col_mat_w; + zero_sz -= src_mat_w_sz; + } + } + } + auto& dev_ctx = ctx.template device_context(); + auto blas = math::GetBlas(dev_ctx); + math::FCCompute(blas, x_dims[0], w_dims[1], w_dims[0], + col_data, w_data, y_data, b_data, true); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(fusion_seqconv_eltadd_relu, ops::FusionSeqConvEltAddReluOp, + ops::FusionSeqConvEltAddReluOpMaker, + paddle::framework::DefaultGradOpDescMaker); + +REGISTER_OP_CPU_KERNEL(fusion_seqconv_eltadd_relu, + ops::FusionSeqConvEltAddReluKernel, + ops::FusionSeqConvEltAddReluKernel); diff --git a/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h b/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h new file mode 100644 index 0000000000000000000000000000000000000000..028d79dc2a1ee8d789fe4b8724b320442041a71b --- /dev/null +++ b/paddle/fluid/operators/fusion_seqconv_eltadd_relu_op.h @@ -0,0 +1,42 @@ +/* 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 "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using Tensor = framework::Tensor; + +class FusionSeqConvEltAddReluOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override; + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override; +}; + +class FusionSeqConvEltAddReluOpMaker + : public framework::OpProtoAndCheckerMaker { + public: + void Make() override; +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc b/paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc index 0cd3d3887cf5167c779a8b20442fdb458cd7eab4..8d2f055d53a0c5bbef624ff3b01b01724d0b3a21 100644 --- a/paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc +++ b/paddle/fluid/operators/fusion_seqexpand_concat_fc_op.cc @@ -136,9 +136,9 @@ class FusionSeqExpandConcatFCOpKernel : public framework::OpKernel { // since infershape can not get lod info PADDLE_ENFORCE_EQ(ref_lod.size(), 1UL, "Only support input lod size is 1."); PADDLE_ENFORCE_EQ(in1_lod.size(), 1UL, "Only support input lod size is 1."); - PADDLE_ENFORCE_EQ(in1_lod[0].size() - 1, N, + PADDLE_ENFORCE_EQ(static_cast(in1_lod[0].size() - 1), N, "Batch size of all inputs should be equal."); - PADDLE_ENFORCE_EQ(in1_lod[0][N], N, + PADDLE_ENFORCE_EQ(static_cast(in1_lod[0][N]), N, "Seq_length of other inputs should be 1."); PADDLE_ENFORCE_EQ(in1_dims[0], N, "input height should be batch size."); for (size_t i = 2; i < ins.size(); ++i) { diff --git a/paddle/fluid/operators/gather.cu.h b/paddle/fluid/operators/gather.cu.h index d74d4db92528d69492ab7b90a98d3775dadc35d1..e4df59c5d51c390cf593add0c5562665c91f33f6 100644 --- a/paddle/fluid/operators/gather.cu.h +++ b/paddle/fluid/operators/gather.cu.h @@ -50,7 +50,9 @@ void GPUGather(const platform::DeviceContext& ctx, const Tensor& src, const Tensor& index, Tensor* output) { // PADDLE_ENFORCE(platform::is_gpu_place(place)); // check index of shape 1-D - PADDLE_ENFORCE(index.dims().size() == 1); + PADDLE_ENFORCE(index.dims().size() == 1 || + (index.dims().size() == 2 && index.dims()[1] == 1)); + int index_size = index.dims()[0]; auto src_dims = src.dims(); diff --git a/paddle/fluid/operators/gather.h b/paddle/fluid/operators/gather.h index d15cb55647ade2415041b11099974484835f55eb..dc08ee5efacde5e232d751b13aaf11f51237634a 100644 --- a/paddle/fluid/operators/gather.h +++ b/paddle/fluid/operators/gather.h @@ -38,12 +38,11 @@ void CPUGather(const platform::DeviceContext& ctx, const Tensor& src, const Tensor& index, Tensor* output) { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace())); // check index of shape 1-D - PADDLE_ENFORCE(index.dims().size() == 1); - int index_size = index.dims()[0]; + PADDLE_ENFORCE(index.dims().size() == 1 || + (index.dims().size() == 2 && index.dims()[1] == 1)); + int64_t index_size = index.dims()[0]; auto src_dims = src.dims(); - framework::DDim output_dims(src_dims); - output_dims[0] = index_size; const T* p_src = src.data(); const int* p_index = index.data(); @@ -55,7 +54,7 @@ void CPUGather(const platform::DeviceContext& ctx, const Tensor& src, const size_t slice_bytes = slice_size * sizeof(T); - for (int i = 0; i < index_size; ++i) { + for (int64_t i = 0; i < index_size; ++i) { int index_ = p_index[i]; memcpy(p_output + i * slice_size, p_src + index_ * slice_size, slice_bytes); } diff --git a/paddle/fluid/operators/gather_op.cc b/paddle/fluid/operators/gather_op.cc index 089b541a0a61adb5efda6b2e027c913d5808dff0..95aa9b573c795159079bdb5401b34d7a61252115 100644 --- a/paddle/fluid/operators/gather_op.cc +++ b/paddle/fluid/operators/gather_op.cc @@ -31,7 +31,8 @@ class GatherOp : public framework::OperatorWithKernel { "Output(Out) of GatherOp should not be null."); auto index_dims = ctx->GetInputDim("Index"); - PADDLE_ENFORCE(index_dims.size() == 1); + PADDLE_ENFORCE(index_dims.size() == 1 || + (index_dims.size() == 2 && index_dims[1] == 1)); int batch_size = ctx->GetInputDim("Index")[0]; framework::DDim output_dims(ctx->GetInputDim("X")); output_dims[0] = batch_size; @@ -53,6 +54,7 @@ class GatherGradOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*-->*/ framework::GradVarName("X")); } protected: @@ -75,7 +77,7 @@ Gather Operator. $Out = X[Index]$ -Out is obtained by gathering entries of the outer-most dimension +Out is obtained by gathering entries of the outer-most dimension of X indexed by Index and concatenate them together. Example: @@ -102,7 +104,9 @@ REGISTER_OPERATOR(gather, ops::GatherOp, ops::GatherOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(gather_grad, ops::GatherGradOp); REGISTER_OP_CPU_KERNEL(gather, ops::GatherOpKernel, - ops::GatherOpKernel, ops::GatherOpKernel); + ops::GatherOpKernel, ops::GatherOpKernel, + ops::GatherOpKernel); REGISTER_OP_CPU_KERNEL(gather_grad, ops::GatherGradientOpKernel, + ops::GatherGradientOpKernel, ops::GatherGradientOpKernel, - ops::GatherGradientOpKernel); + ops::GatherGradientOpKernel); diff --git a/paddle/fluid/operators/gather_op.cu b/paddle/fluid/operators/gather_op.cu index 7e014dd1cb47ee0575308dc13ba7bc7617baebff..9f4aef08cd58e72ce344a640e6564b9e360ce169 100644 --- a/paddle/fluid/operators/gather_op.cu +++ b/paddle/fluid/operators/gather_op.cu @@ -61,5 +61,11 @@ class GatherGradOpCUDAKernel : public framework::OpKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel); -REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel, + ops::GatherOpCUDAKernel, + ops::GatherOpCUDAKernel, + ops::GatherOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel, + ops::GatherGradOpCUDAKernel, + ops::GatherGradOpCUDAKernel, + ops::GatherGradOpCUDAKernel); diff --git a/paddle/fluid/operators/gaussian_random_op.cc b/paddle/fluid/operators/gaussian_random_op.cc index 1488aab1926b5b4ba7bceed582700f5a11fc6c93..c70d5b8bc7569c38cbc003aca7c62dc503df11cf 100644 --- a/paddle/fluid/operators/gaussian_random_op.cc +++ b/paddle/fluid/operators/gaussian_random_op.cc @@ -52,7 +52,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel { void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of GaussianRandomOp should not be null."); - auto shape = ctx->Attrs().Get>("shape"); + auto shape = ctx->Attrs().Get>("shape"); std::vector temp; temp.reserve(shape.size()); for (auto dim : shape) { @@ -88,9 +88,9 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddOutput("Out", "Output matrix of gaussian random op"); - AddAttr>("shape", - "(vector) " - "The dimension of random tensor."); + AddAttr>("shape", + "(vector) " + "The dimension of random tensor."); AddAttr("mean", "(float, default 0.0) " "mean of random tensor.") diff --git a/paddle/fluid/operators/gen_nccl_id_op.cc b/paddle/fluid/operators/gen_nccl_id_op.cc index 697c239e59d158428ae9ba9f7feded19637dff28..56ea165ff84291babc0e9ee56ada669cbbbe79fe 100644 --- a/paddle/fluid/operators/gen_nccl_id_op.cc +++ b/paddle/fluid/operators/gen_nccl_id_op.cc @@ -61,10 +61,10 @@ class GenNCCLIdOp : public framework::OperatorBase { std::vector endpoint_list = Attr>("endpoint_list"); distributed::RPCClient* client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(0); for (auto& ep : endpoint_list) { - VLOG(3) << "sending nccl id to " << ep; + VLOG(30) << "sending nccl id to " << ep; client->AsyncSendVar(ep, dev_ctx, *scope, NCCL_ID_VARNAME); } client->Wait(); @@ -72,7 +72,7 @@ class GenNCCLIdOp : public framework::OperatorBase { client->AsyncSendBatchBarrier(ep); } client->Wait(); - VLOG(3) << "sending completed..."; + VLOG(30) << "sending completed..."; } void GetIdByServer(framework::Scope* scope, @@ -99,11 +99,11 @@ class GenNCCLIdOp : public framework::OperatorBase { std::bind(&distributed::RPCServer::StartServer, rpc_service.get())); rpc_service->SetCond(distributed::kRequestSend); - VLOG(3) << "start getting nccl id from trainer 0..."; + VLOG(30) << "start getting nccl id from trainer 0..."; rpc_service->WaitBarrier(distributed::kRequestSend); - VLOG(3) << "got nccl id and stop server..."; + VLOG(30) << "got nccl id and stop server..."; rpc_service->ShutDown(); - VLOG(3) << "rpc server stopped"; + VLOG(30) << "rpc server stopped"; server_thread.join(); } }; diff --git a/paddle/fluid/operators/grid_sampler_cudnn_op.cu.cc b/paddle/fluid/operators/grid_sampler_cudnn_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..7cde7ca462fda9ae6ace7755af0a432afee28bba --- /dev/null +++ b/paddle/fluid/operators/grid_sampler_cudnn_op.cu.cc @@ -0,0 +1,132 @@ +/* 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/framework/op_registry.h" +#include "paddle/fluid/platform/cudnn_helper.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; +using ScopedTensorDescriptor = platform::ScopedTensorDescriptor; +using DataLayout = platform::DataLayout; +using ScopedSpatialTransformerDescriptor = + platform::ScopedSpatialTransformerDescriptor; +template +using CudnnDataType = platform::CudnnDataType; + +template +class CUDNNGridSampleOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace"); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto* input = ctx.Input("X"); + auto* grid = ctx.Input("Grid"); + auto* output = ctx.Output("Output"); + + int n = input->dims()[0]; + int c = input->dims()[1]; + int h = input->dims()[2]; + int w = input->dims()[3]; + const int size[4] = {n, c, h, w}; + + const T* input_data = input->data(); + const T* grid_data = grid->data(); + T* output_data = output->mutable_data({n, c, h, w}, ctx.GetPlace()); + + ScopedSpatialTransformerDescriptor st_desc; + cudnnSpatialTransformerDescriptor_t cudnn_st_desc = + st_desc.descriptor(4, size); + + ScopedTensorDescriptor input_desc; + ScopedTensorDescriptor output_desc; + cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(input->dims())); + cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(output->dims())); + + CUDNN_ENFORCE(platform::dynload::cudnnSpatialTfSamplerForward( + handle, cudnn_st_desc, CudnnDataType::kOne(), cudnn_input_desc, + input_data, grid_data, CudnnDataType::kZero(), cudnn_output_desc, + output_data)); + } +}; + +template +class CUDNNGridSampleGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "It must use CUDAPlace"); + auto& dev_ctx = ctx.template device_context(); + auto handle = dev_ctx.cudnn_handle(); + auto* input = ctx.Input("X"); + auto* grid = ctx.Input("Grid"); + auto* output_grad = ctx.Input(framework::GradVarName("Output")); + auto* input_grad = ctx.Output(framework::GradVarName("X")); + auto* grid_grad = ctx.Output(framework::GradVarName("Grid")); + + auto output_grad_dims = output_grad->dims(); + const int n = output_grad_dims[0]; + const int c = output_grad_dims[1]; + const int h = output_grad_dims[2]; + const int w = output_grad_dims[3]; + const int size[4] = {n, c, h, w}; + + ScopedSpatialTransformerDescriptor st_dest; + cudnnSpatialTransformerDescriptor_t cudnn_st_dest = + st_dest.descriptor(4, size); + + const T* input_data = input->data(); + const T* grid_data = grid->data(); + const T* output_grad_data = output_grad->data(); + T* input_grad_data = + input_grad->mutable_data(output_grad_dims, ctx.GetPlace()); + T* grid_grad_data = + grid_grad->mutable_data({n, h, w, 2}, ctx.GetPlace()); + + ScopedTensorDescriptor input_desc; + ScopedTensorDescriptor input_grad_desc; + ScopedTensorDescriptor output_grad_desc; + cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(input->dims())); + cudnnTensorDescriptor_t cudnn_input_grad_desc = + input_grad_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(input_grad->dims())); + cudnnTensorDescriptor_t cudnn_output_grad_desc = + output_grad_desc.descriptor( + DataLayout::kNCHW, framework::vectorize2int(output_grad->dims())); + + CUDNN_ENFORCE(platform::dynload::cudnnSpatialTfSamplerBackward( + handle, cudnn_st_dest, CudnnDataType::kOne(), cudnn_input_desc, + input_data, CudnnDataType::kZero(), cudnn_input_grad_desc, + input_grad_data, CudnnDataType::kOne(), cudnn_output_grad_desc, + output_grad_data, grid_data, CudnnDataType::kZero(), + grid_grad_data)); + } +}; + +} // namespace operators +} // namespace paddle + +namespace plat = paddle::platform; +REGISTER_OP_KERNEL(grid_sampler, CUDNN, plat::CUDAPlace, + paddle::operators::CUDNNGridSampleOpKernel, + paddle::operators::CUDNNGridSampleOpKernel); +REGISTER_OP_KERNEL(grid_sampler_grad, CUDNN, plat::CUDAPlace, + paddle::operators::CUDNNGridSampleGradOpKernel, + paddle::operators::CUDNNGridSampleGradOpKernel); diff --git a/paddle/fluid/operators/grid_sampler_op.cc b/paddle/fluid/operators/grid_sampler_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..e76eb6893b1f7b6a965682368860c02fa32f6330 --- /dev/null +++ b/paddle/fluid/operators/grid_sampler_op.cc @@ -0,0 +1,203 @@ +/* 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/grid_sampler_op.h" +#include "paddle/fluid/framework/op_registry.h" +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/cudnn_helper.h" +#endif + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +class GridSampleOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of GridSampleOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grid"), + "Input(Grid) of GridSampleOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Output"), + "Output(Output) of GridSampleOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + auto grid_dims = ctx->GetInputDim("Grid"); + PADDLE_ENFORCE(x_dims.size() == 4, + "Input(X) of GridSampleOp should be 4-D Tensor."); + PADDLE_ENFORCE(grid_dims.size() == 4, + "Input(Grid) of GridSampleOp should be 4-D Tensor."); + PADDLE_ENFORCE(grid_dims[3] == 2, "Input(Grid) dims[3] should be 2."); + PADDLE_ENFORCE_EQ(grid_dims[0], x_dims[0], + "Input(X) and Input(Grid) dims[0] should be equal."); + PADDLE_ENFORCE_EQ( + grid_dims[1], x_dims[2], + "Input(X) dims[2] and Input(Grid) dims[1] should be equal."); + PADDLE_ENFORCE_EQ( + grid_dims[2], x_dims[3], + "Input(X) dims[3] and Input(Grid) dims[2] should be equal."); + + ctx->SetOutputDim("Output", x_dims); + ctx->ShareLoD("X", "Output"); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_CUDA + if (platform::CanCUDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kCUDNN; + } +#endif + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + framework::DataLayout::kAnyLayout, library_); + } +}; + +class GridSampleOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor) The input data of GridSampleOp, " + "This is a 4-D tensor with shape of [N, C, H, W]"); + AddInput( + "Grid", + "(Tensor) The input grid of GridSampleOp generated by AffineGridOp, " + "This is a 4-D tensor with shape of [N, H, W, 2] is the concatenation " + "of x and y coordinates with shape [N, H, W] in last dimention"); + AddOutput("Output", "(Tensor) Output tensor with shape [N, C, H, W]"); + AddAttr( + "use_cudnn", + "(bool, default true) Only used in cudnn kernel, need install cudnn") + .SetDefault(true); + + AddComment(R"DOC( + This operation samples input X by using bilinear interpolation based on + flow field grid, which is usually gennerated by affine_grid. The grid of + shape [N, H, W, 2] is the concatenation of (grid_x, grid_y) coordinates + with shape [N, H, W] each, where grid_x is indexing the 4th dimension + (in width dimension) of input data x and grid_y is indexng the 3rd + dimention (in height dimension), finally results is the bilinear + interpolation value of 4 nearest corner points. + + Step 1: + Get (x, y) grid coordinates and scale to [0, H-1/W-1]. + + grid_x = 0.5 * (grid[:, :, :, 0] + 1) * (W - 1) + grid_y = 0.5 * (grid[:, :, :, 1] + 1) * (H - 1) + + Step 2: + Indices input data X with grid (x, y) in each [H, W] area, and bilinear + interpolate point value by 4 nearest points. + + wn ------- y_n ------- en + | | | + | d_n | + | | | + x_w --d_w-- grid--d_e-- x_e + | | | + | d_s | + | | | + ws ------- y_s ------- wn + + x_w = floor(x) // west side x coord + x_e = x_w + 1 // east side x coord + y_n = floor(y) // north side y coord + y_s = y_s + 1 // south side y coord + + d_w = grid_x - x_w // distance to west side + d_e = x_e - grid_x // distance to east side + d_n = grid_y - y_n // distance to north side + d_s = y_s - grid_y // distance to south side + + wn = X[:, :, y_n, x_w] // north-west point value + en = X[:, :, y_n, x_e] // north-east point value + ws = X[:, :, y_s, x_w] // south-east point value + es = X[:, :, y_s, x_w] // north-east point value + + output = wn * d_e * d_s + en * d_w * d_s + + ws * d_e * d_n + es * d_w * d_n + )DOC"); + } +}; + +class GridSampleOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + void InferShape(framework::InferShapeContext* ctx) const override { + auto input_dims = ctx->GetInputDim("X"); + auto grid_dims = ctx->GetInputDim("Grid"); + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), input_dims); + } + if (ctx->HasOutput(framework::GradVarName("Grid"))) { + ctx->SetOutputDim(framework::GradVarName("Grid"), grid_dims); + } + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_CUDA + if (platform::CanCUDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kCUDNN; + } +#endif + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + framework::DataLayout::kAnyLayout, library_); + } +}; + +class GridSampleGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto* op = new framework::OpDesc(); + op->SetType("grid_sampler_grad"); + op->SetInput("X", Input("X")); + op->SetInput("Grid", Input("Grid")); + op->SetInput(framework::GradVarName("Output"), OutputGrad("Output")); + + op->SetAttrMap(Attrs()); + + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + op->SetOutput(framework::GradVarName("Grid"), InputGrad("Grid")); + return std::unique_ptr(op); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(grid_sampler, ops::GridSampleOp, ops::GridSampleOpMaker, + ops::GridSampleGradMaker); +REGISTER_OPERATOR(grid_sampler_grad, ops::GridSampleOpGrad); + +REGISTER_OP_CPU_KERNEL( + grid_sampler, + ops::GridSampleOpKernel, + ops::GridSampleOpKernel); +REGISTER_OP_CPU_KERNEL( + grid_sampler_grad, + ops::GridSampleGradOpKernel, + ops::GridSampleGradOpKernel); diff --git a/paddle/fluid/operators/grid_sampler_op.h b/paddle/fluid/operators/grid_sampler_op.h new file mode 100644 index 0000000000000000000000000000000000000000..4e91a3dcd272c8d368cb8c43e7e1fb4c98265db4 --- /dev/null +++ b/paddle/fluid/operators/grid_sampler_op.h @@ -0,0 +1,329 @@ +/* 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 "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/gather.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/hostdevice.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +template +using EigenTensor = framework::EigenTensor; + +using Array3 = Eigen::DSizes; +using Array4 = Eigen::DSizes; + +template +static inline bool isInBound(T x, T y, T x_max, T y_max) { + if (x < 0 || x > x_max || y < 0 || y > y_max) { + return false; + } + return true; +} + +template +static void CalcGridLocations(const platform::CPUDeviceContext& ctx, + const Tensor& grid, Tensor* x_w, Tensor* x_e, + Tensor* y_n, Tensor* y_s, Tensor* d_w, + Tensor* d_e, Tensor* d_n, Tensor* d_s) { + auto& place = *ctx.eigen_device(); + const int n = grid.dims()[0]; + const int h = grid.dims()[1]; + const int w = grid.dims()[2]; + const T x_max = static_cast(w - 1); + const T y_max = static_cast(h - 1); + + // split grid with shape (n, h, w, 2) into (x, y) by the 3rd Dim + Tensor grid_x, grid_y; + T* grid_x_data = grid_x.mutable_data({n, h, w}, ctx.GetPlace()); + T* grid_y_data = grid_y.mutable_data({n, h, w}, ctx.GetPlace()); + const T* grid_data = grid.data(); + for (int i = 0; i < n * h * w; i++) { + grid_x_data[i] = grid_data[2 * i]; + grid_y_data[i] = grid_data[(2 * i) + 1]; + } + + Tensor ones; + ones.mutable_data({n, h, w}, ctx.GetPlace()); + auto ones_t = EigenTensor::From(ones).setConstant(1.0); + Tensor half_xmax, half_ymax; + half_xmax.mutable_data({n, h, w}, ctx.GetPlace()); + auto half_xmax_t = + EigenTensor::From(half_xmax).setConstant(0.5 * x_max); + half_ymax.mutable_data({n, h, w}, ctx.GetPlace()); + auto half_ymax_t = + EigenTensor::From(half_ymax).setConstant(0.5 * y_max); + + // scale grid to [0, h-1/w-1] + auto grid_x_t = EigenTensor::From(grid_x); + auto grid_y_t = EigenTensor::From(grid_y); + grid_x_t.device(place) = (grid_x_t + ones_t) * half_xmax_t; + grid_y_t.device(place) = (grid_y_t + ones_t) * half_ymax_t; + + // calculate coords of 4 corner points + x_w->mutable_data({n, h, w}, ctx.GetPlace()); + x_e->mutable_data({n, h, w}, ctx.GetPlace()); + y_n->mutable_data({n, h, w}, ctx.GetPlace()); + y_s->mutable_data({n, h, w}, ctx.GetPlace()); + auto x_w_t = EigenTensor::From(*x_w); + auto x_e_t = EigenTensor::From(*x_e); + auto y_n_t = EigenTensor::From(*y_n); + auto y_s_t = EigenTensor::From(*y_s); + x_w_t.device(place) = grid_x_t.floor(); + x_e_t.device(place) = x_w_t + ones_t; + y_n_t.device(place) = grid_y_t.floor(); + y_s_t.device(place) = y_n_t + ones_t; + + // calculate distances to 4 sides + d_w->mutable_data({n, h, w}, ctx.GetPlace()); + d_e->mutable_data({n, h, w}, ctx.GetPlace()); + d_n->mutable_data({n, h, w}, ctx.GetPlace()); + d_s->mutable_data({n, h, w}, ctx.GetPlace()); + auto d_w_t = EigenTensor::From(*d_w); + auto d_e_t = EigenTensor::From(*d_e); + auto d_n_t = EigenTensor::From(*d_n); + auto d_s_t = EigenTensor::From(*d_s); + d_w_t.device(place) = grid_x_t - x_w_t; + d_e_t.device(place) = x_e_t - grid_x_t; + d_n_t.device(place) = grid_y_t - y_n_t; + d_s_t.device(place) = y_s_t - grid_y_t; +} + +template +static void GetGridPointValue(const Tensor& input, Tensor* output, + const Tensor& x, const Tensor& y) { + const int n = input.dims()[0]; + const int c = input.dims()[1]; + const int h = input.dims()[2]; + const int w = input.dims()[3]; + auto x_t = EigenTensor::From(x); + auto y_t = EigenTensor::From(y); + auto output_t = EigenTensor::From(*output).setConstant((T)0); + auto input_t = EigenTensor::From(input); + + for (int i = 0; i < n; i++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + if (isInBound(x_t(i, k, l), y_t(i, k, l), (T)(w - 1), (T)(h - 1))) { + for (int j = 0; j < c; j++) { + output_t(i, j, k, l) = + input_t(i, j, static_cast(round(y_t(i, k, l))), + static_cast(round(x_t(i, k, l)))); + } + } + } + } + } +} + +template +static void GatherOutputGradToInputGrad(const Tensor& output_grad, + Tensor* input_grad, const Tensor& x, + const Tensor& y, const Tensor& d1, + const Tensor& d2) { + const int n = output_grad.dims()[0]; + const int c = output_grad.dims()[1]; + const int h = output_grad.dims()[2]; + const int w = output_grad.dims()[3]; + auto x_t = EigenTensor::From(x); + auto y_t = EigenTensor::From(y); + auto d1_t = EigenTensor::From(d1); + auto d2_t = EigenTensor::From(d2); + auto input_grad_t = EigenTensor::From(*input_grad); + auto output_grad_t = EigenTensor::From(output_grad); + + for (int i = 0; i < n; i++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + if (isInBound(x_t(i, k, l), y_t(i, k, l), (T)(w - 1), (T)(h - 1))) { + for (int j = 0; j < c; j++) { + input_grad_t(i, j, static_cast(round(y_t(i, k, l))), + static_cast(round(x_t(i, k, l)))) += + output_grad_t(i, j, k, l) * d1_t(i, k, l) * d2_t(i, k, l); + } + } + } + } + } +} + +template +class GridSampleOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto& place = *ctx.template device_context().eigen_device(); + auto* input = ctx.Input("X"); + auto* grid = ctx.Input("Grid"); + + const int n = input->dims()[0]; + const int c = input->dims()[1]; + const int h = input->dims()[2]; + const int w = input->dims()[3]; + + // calc locations and distances of 4 corner points + Tensor x_w, x_e, y_n, y_s; + Tensor d_w, d_e, d_n, d_s; + CalcGridLocations( + ctx.template device_context(), *grid, &x_w, + &x_e, &y_n, &y_s, &d_w, &d_e, &d_n, &d_s); + + auto* output = ctx.Output("Output"); + output->mutable_data({n, c, h, w}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), output, + static_cast(0)); + + // calc 4 corner points value + Tensor v_wn, v_en, v_ws, v_es; + v_wn.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_en.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_ws.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_es.mutable_data({n, c, h, w}, ctx.GetPlace()); + GetGridPointValue(*input, &v_wn, x_w, y_n); + GetGridPointValue(*input, &v_en, x_e, y_n); + GetGridPointValue(*input, &v_ws, x_w, y_s); + GetGridPointValue(*input, &v_es, x_e, y_s); + + auto d_w_t = EigenTensor::From(d_w); + auto d_e_t = EigenTensor::From(d_e); + auto d_n_t = EigenTensor::From(d_n); + auto d_s_t = EigenTensor::From(d_s); + auto d_w_scaled_t = + d_w_t.reshape(Array4(n, 1, h, w)).broadcast(Array4(1, c, 1, 1)); + auto d_e_scaled_t = + d_e_t.reshape(Array4(n, 1, h, w)).broadcast(Array4(1, c, 1, 1)); + auto d_n_scaled_t = + d_n_t.reshape(Array4(n, 1, h, w)).broadcast(Array4(1, c, 1, 1)); + auto d_s_scaled_t = + d_s_t.reshape(Array4(n, 1, h, w)).broadcast(Array4(1, c, 1, 1)); + auto v_wn_t = EigenTensor::From(v_wn); + auto v_en_t = EigenTensor::From(v_en); + auto v_ws_t = EigenTensor::From(v_ws); + auto v_es_t = EigenTensor::From(v_es); + auto output_t = EigenTensor::From(*output); + // bilinear interpolaetion by 4 corner points + output_t.device(place) = v_wn_t * d_e_scaled_t * d_s_scaled_t + + v_en_t * d_w_scaled_t * d_s_scaled_t + + v_ws_t * d_e_scaled_t * d_n_scaled_t + + v_es_t * d_w_scaled_t * d_n_scaled_t; + } +}; + +template +class GridSampleGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("X"); + auto* grid = ctx.Input("Grid"); + auto* output_grad = ctx.Input(framework::GradVarName("Output")); + + const int n = input->dims()[0]; + const int c = input->dims()[1]; + const int h = input->dims()[2]; + const int w = input->dims()[3]; + + auto* input_grad = ctx.Output(framework::GradVarName("X")); + input_grad->mutable_data({n, c, h, w}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), input_grad, + static_cast(0)); + auto* grid_grad = ctx.Output(framework::GradVarName("Grid")); + grid_grad->mutable_data({n, h, w, 2}, ctx.GetPlace()); + math::SetConstant()( + ctx.template device_context(), grid_grad, + static_cast(0)); + + Tensor x_w, x_e, y_n, y_s; + Tensor d_w, d_e, d_n, d_s; + CalcGridLocations( + ctx.template device_context(), *grid, &x_w, + &x_e, &y_n, &y_s, &d_w, &d_e, &d_n, &d_s); + + // gather output grad value to input grad by corner point coords and weight + GatherOutputGradToInputGrad(*output_grad, input_grad, x_w, y_n, d_e, + d_s); + GatherOutputGradToInputGrad(*output_grad, input_grad, x_w, y_s, d_e, + d_n); + GatherOutputGradToInputGrad(*output_grad, input_grad, x_e, y_n, d_w, + d_s); + GatherOutputGradToInputGrad(*output_grad, input_grad, x_e, y_s, d_w, + d_n); + + // calc 4 corner points value + Tensor v_wn, v_en, v_ws, v_es; + v_wn.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_en.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_ws.mutable_data({n, c, h, w}, ctx.GetPlace()); + v_es.mutable_data({n, c, h, w}, ctx.GetPlace()); + GetGridPointValue(*input, &v_wn, x_w, y_n); + GetGridPointValue(*input, &v_en, x_e, y_n); + GetGridPointValue(*input, &v_ws, x_w, y_s); + GetGridPointValue(*input, &v_es, x_e, y_s); + auto v_wn_t = EigenTensor::From(v_wn); + auto v_en_t = EigenTensor::From(v_en); + auto v_ws_t = EigenTensor::From(v_ws); + auto v_es_t = EigenTensor::From(v_es); + + auto d_w_t = EigenTensor::From(d_w); + auto d_e_t = EigenTensor::From(d_e); + auto d_n_t = EigenTensor::From(d_n); + auto d_s_t = EigenTensor::From(d_s); + + auto output_grad_t = EigenTensor::From(*output_grad); + + Tensor grid_grad_x, grid_grad_y; + grid_grad_x.mutable_data({n, h, w}, ctx.GetPlace()); + grid_grad_y.mutable_data({n, h, w}, ctx.GetPlace()); + auto grid_grad_x_t = EigenTensor::From(grid_grad_x).setConstant(0.0); + auto grid_grad_y_t = EigenTensor::From(grid_grad_y).setConstant(0.0); + for (int i = 0; i < n; i++) { + for (int j = 0; j < c; j++) { + for (int k = 0; k < h; k++) { + for (int l = 0; l < w; l++) { + grid_grad_x_t(i, k, l) += + ((v_en_t(i, j, k, l) - v_wn_t(i, j, k, l)) * d_s_t(i, k, l) + + (v_es_t(i, j, k, l) - v_ws_t(i, j, k, l)) * d_n_t(i, k, l)) * + output_grad_t(i, j, k, l); + grid_grad_y_t(i, k, l) += + ((v_ws_t(i, j, k, l) - v_wn_t(i, j, k, l)) * d_e_t(i, k, l) + + (v_es_t(i, j, k, l) - v_en_t(i, j, k, l)) * d_w_t(i, k, l)) * + output_grad_t(i, j, k, l); + } + } + } + } + const T x_max = static_cast(w - 1); + const T y_max = static_cast(h - 1); + grid_grad_x_t = grid_grad_x_t * (x_max / (T)2); + grid_grad_y_t = grid_grad_y_t * (y_max / (T)2); + + // gather grid_grad [x, y] in 3rd Dim + T* grid_grad_data = grid_grad->data(); + T* grid_grad_x_data = grid_grad_x.data(); + T* grid_grad_y_data = grid_grad_y.data(); + for (int i = 0; i < n * h * w; i++) { + grid_grad_data[2 * i] = grid_grad_x_data[i]; + grid_grad_data[2 * i + 1] = grid_grad_y_data[i]; + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/hash_op.cc b/paddle/fluid/operators/hash_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..b9ebe71a3d7ae270a10a45f4805652415078b363 --- /dev/null +++ b/paddle/fluid/operators/hash_op.cc @@ -0,0 +1,74 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/hash_op.h" +#include +#include + +namespace paddle { +namespace operators { + +class HashOp : public framework::OperatorWithKernel { + public: + HashOp(const std::string &type, const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of HashOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of HashOp should not be null."); + + auto dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_EQ(dims.size(), 2UL, + "The input of hash_op's dimensions must be 2"); + std::vector out_dims; + out_dims.reserve(dims.size() + 1); + // copy all dims except the last one + for (size_t i = 0u; i != dims.size() - 1; ++i) { + out_dims.emplace_back(dims[i]); + } + int num_hash = ctx->Attrs().Get("num_hash"); + out_dims.emplace_back(num_hash); + // keep the last dim to 1 + out_dims.emplace_back(1); + + ctx->SetOutputDim("Out", framework::make_ddim(out_dims)); + ctx->ShareLoD("X", /*->*/ "Out"); + } +}; + +class HashOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor) Input tensor of scale operator."); + AddOutput("Out", "(Tensor) Output tensor of scale operator."); + AddComment(R"DOC( +**Hash Operator** +$$Out = scale * X$$ +)DOC"); + AddAttr("num_hash", "").SetDefault(1); + AddAttr("mod_by", "").SetDefault(100000); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_WITHOUT_GRADIENT(hash, ops::HashOp, ops::HashOpMaker); +REGISTER_OP_CPU_KERNEL(hash, ops::HashKerel, ops::HashKerel); diff --git a/paddle/fluid/operators/hash_op.h b/paddle/fluid/operators/hash_op.h new file mode 100644 index 0000000000000000000000000000000000000000..9781bb0f453642cefb3eb59a05389c339a7de39d --- /dev/null +++ b/paddle/fluid/operators/hash_op.h @@ -0,0 +1,56 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +extern "C" { +#include +} +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +// template +template +class HashKerel : public framework::OpKernel { + public: + virtual void Compute(const framework::ExecutionContext& context) const { + auto* out_t = context.Output("Out"); + auto* in_t = context.Input("X"); + int mod_by = context.Attr("mod_by"); + int num_hash = context.Attr("num_hash"); + auto* output = out_t->mutable_data(context.GetPlace()); + + auto in_dims = in_t->dims(); + auto in_lod = in_t->lod(); + PADDLE_ENFORCE_EQ( + static_cast(in_dims[0]), in_lod[0].back(), + "The actual input data's size mismatched with LoD information."); + + auto seq_length = in_dims[0]; + auto last_dim = in_dims[in_dims.size() - 1]; + auto* input = in_t->data(); + for (int idx = 0; idx < seq_length; ++idx) { + for (int ihash = 0; ihash != num_hash; ++ihash) { + output[idx * num_hash + ihash] = + XXH64(input, sizeof(int) * last_dim, ihash) % mod_by; + } + input += last_dim; + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/bilinear_interp_op.cc b/paddle/fluid/operators/interpolate_op.cc similarity index 52% rename from paddle/fluid/operators/bilinear_interp_op.cc rename to paddle/fluid/operators/interpolate_op.cc index 2dc3399da183fbcf7664066f6f7ce12db3dc6d5e..8f979e05d31e5a85bc86784943f4588ab650f668 100644 --- a/paddle/fluid/operators/bilinear_interp_op.cc +++ b/paddle/fluid/operators/interpolate_op.cc @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at @@ -9,7 +9,8 @@ See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/bilinear_interp_op.h" +#include "paddle/fluid/operators/interpolate_op.h" +#include #include #include "paddle/fluid/framework/op_registry.h" @@ -18,27 +19,34 @@ namespace operators { using framework::Tensor; -class BilinearInterpOp : public framework::OperatorWithKernel { +class InterpolateOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; protected: void InferShape(framework::InferShapeContext* ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of BilinearInterOp should not be null."); + "Input(X) of InterpolateOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of BilinearInterOp should not be null."); + "Output(Out) of InterpolationOp should not be null."); + + auto interp_method = ctx->Attrs().Get("interp_method"); + PADDLE_ENFORCE( + "bilinear" == interp_method || "nearest" == interp_method, + "Interpolation method can only be \"bilinear\" or \"nearest\"."); auto dim_x = ctx->GetInputDim("X"); // NCHW format int out_h = ctx->Attrs().Get("out_h"); int out_w = ctx->Attrs().Get("out_w"); PADDLE_ENFORCE_EQ(dim_x.size(), 4, "X's dimension must be 4"); - if (ctx->HasInput("OutSize")) { + if (ctx->HasInput("OutSize") && ctx->IsRuntime()) { auto out_size_dim = ctx->GetInputDim("OutSize"); PADDLE_ENFORCE_EQ(out_size_dim.size(), 1, "OutSize's dimension size must be 1"); PADDLE_ENFORCE_EQ(out_size_dim[0], 2, "OutSize's dim[0] must be 2"); + ctx->ShareLoD("X", "Out"); + return; } std::vector dim_out({dim_x[0], dim_x[1], out_h, out_w}); ctx->SetOutputDim("Out", framework::make_ddim(dim_out)); @@ -52,35 +60,53 @@ class BilinearInterpOp : public framework::OperatorWithKernel { } }; -class BilinearInterpOpMaker : public framework::OpProtoAndCheckerMaker { +class InterpolateOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", - "The input tensor of bilinear interpolation, " - "This is a 4-D tensor with shape of (N x C x h x w)"); + "The input tensor of interpolate operator, " + "This is a 4-D tensor with shape of [N, C, H, w]."); AddInput("OutSize", - "This is a 1-D tensor with two number. " + "This is a 1-D tensor with two numbers to specify output size. " "The first number is height and the second number is width.") .AsDispensable(); - AddOutput("Out", "The dimension of output is (N x C x out_h x out_w)"); + AddOutput("Out", + "The output tensor of interpolate operator, " + "This is a 4-D tensor with shape of [N, C, H, W]."); - AddAttr("out_h", "output height of bilinear interpolation op."); - AddAttr("out_w", "output width of bilinear interpolation op."); + AddAttr("out_h", "output height of interpolate op."); + AddAttr("out_w", "output width of interpolate op."); + AddAttr( + "interp_method", + "(string), interpolation method, can be \"bilinear\" for " + "bilinear interpolation and \"nearest\" for nearest " + "neighbor interpolation."); AddComment(R"DOC( + This operator samples input X to given output shape by using specified + interpolation method, the interpolation methods can be \"nearest\" + for nearest neighbor interpolation and \"bilinear\" for bilinear + interpolation. + + Nearest neighbor interpolation is to perform nearest neighbor interpolation + in both the 3rd dimention(in height direction) and the 4th dimention(in width + direction) on input tensor. + Bilinear interpolation is an extension of linear interpolation for interpolating functions of two variables (e.g. H-direction and - W-direction in this op) on a rectilinear 2D grid. - - The key idea is to perform linear interpolation first in one - direction, and then again in the other direction. - - For details, please refer to Wikipedia: + W-direction in this op) on a rectilinear 2D grid. The key idea is + to perform linear interpolation first in one direction, and then + again in the other direction. + + For details of nearest neighbor interpolation, please refer to Wikipedia: + https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation + + For details of bilinear interpolation, please refer to Wikipedia: https://en.wikipedia.org/wiki/Bilinear_interpolation )DOC"); } }; -class BilinearInterpOpGrad : public framework::OperatorWithKernel { +class InterpolateOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -106,11 +132,11 @@ class BilinearInterpOpGrad : public framework::OperatorWithKernel { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(bilinear_interp, ops::BilinearInterpOp, - ops::BilinearInterpOpMaker, +REGISTER_OPERATOR(interpolate, ops::InterpolateOp, ops::InterpolateOpMaker, paddle::framework::DefaultGradOpDescMaker); -REGISTER_OPERATOR(bilinear_interp_grad, ops::BilinearInterpOpGrad); -REGISTER_OP_CPU_KERNEL(bilinear_interp, ops::BilinearInterpKernel, - ops::BilinearInterpKernel); -REGISTER_OP_CPU_KERNEL(bilinear_interp_grad, - ops::BilinearInterpGradKernel); +REGISTER_OPERATOR(interpolate_grad, ops::InterpolateOpGrad); +REGISTER_OP_CPU_KERNEL(interpolate, ops::InterpolateKernel, + ops::InterpolateKernel, + ops::InterpolateKernel); +REGISTER_OP_CPU_KERNEL(interpolate_grad, ops::InterpolateGradKernel, + ops::InterpolateGradKernel); diff --git a/paddle/fluid/operators/interpolate_op.cu b/paddle/fluid/operators/interpolate_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..190afbdac431f863c32e2a4a4b3ad83848e550fc --- /dev/null +++ b/paddle/fluid/operators/interpolate_op.cu @@ -0,0 +1,292 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include +#include "paddle/fluid/operators/interpolate_op.h" +#include "paddle/fluid/platform/cuda_primitives.h" + +namespace paddle { +namespace operators { + +using framework::Tensor; + +template +__global__ void KeNearestNeighborInterpFw( + const T* in, const size_t in_img_h, const size_t in_img_w, + const size_t input_h, const size_t input_w, T* out, const size_t out_img_h, + const size_t out_img_w, const size_t output_h, const size_t output_w, + const size_t num_channels, const float ratio_h, const float ratio_w) { + int nthreads = output_h * output_w; + int tid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (; tid < nthreads; tid += stride) { + int out_id_h = tid / output_w; + int out_id_w = tid % output_w; + int in_img_size = input_w / num_channels; + int out_img_size = output_w / num_channels; + int channel_id = out_id_w / out_img_size; + + int out_img_idy = (out_id_w % out_img_size) / out_img_w; + int in_img_idy = static_cast(ratio_h * out_img_idy + 0.5); + + int out_img_idx = tid % out_img_w; + int in_img_idx = static_cast(ratio_w * out_img_idx + 0.5); + + out[tid] = in[out_id_h * input_w + channel_id * in_img_size + + in_img_idy * in_img_w + in_img_idx]; + } +} + +template +__global__ void KeNearestNeighborInterpBw( + T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h, + const size_t input_w, const T* out, const size_t out_img_h, + const size_t out_img_w, const size_t output_h, const size_t output_w, + const size_t num_channels, const float ratio_h, const float ratio_w) { + int nthreads = output_h * output_w; + int tid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (; tid < nthreads; tid += stride) { + int out_id_h = tid / output_w; + int out_id_w = tid % output_w; + int in_img_size = input_w / num_channels; + int out_img_size = output_w / num_channels; + int channel_id = out_id_w / out_img_size; + + int out_img_idy = (out_id_w % out_img_size) / out_img_w; + int in_img_idy = static_cast(ratio_h * out_img_idy + 0.5); + + int out_img_idx = tid % out_img_w; + int in_img_idx = static_cast(ratio_w * out_img_idx + 0.5); + + T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + + in_img_idy * in_img_w + in_img_idx]; + const T out_pos = out[out_id_h * output_w + out_id_w]; + platform::CudaAtomicAdd(in_pos, out_pos); + } +} + +template +__global__ void KeBilinearInterpFw( + const T* in, const size_t in_img_h, const size_t in_img_w, + const size_t input_h, const size_t input_w, T* out, const size_t out_img_h, + const size_t out_img_w, const size_t output_h, const size_t output_w, + const size_t num_channels, const float ratio_h, const float ratio_w) { + int nthreads = output_h * output_w; + int tid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (; tid < nthreads; tid += stride) { + int out_id_h = tid / output_w; + int out_id_w = tid % output_w; + int in_img_size = input_w / num_channels; + int out_img_size = output_w / num_channels; + int channel_id = out_id_w / out_img_size; + + int out_img_idy = (out_id_w % out_img_size) / out_img_w; + int in_img_idy = ratio_h * out_img_idy; + int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0; + T h1lambda = ratio_h * out_img_idy - in_img_idy; + T h2lambda = 1.f - h1lambda; + + int out_img_idx = tid % out_img_w; + int in_img_idx = ratio_w * out_img_idx; + int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0; + T w1lambda = ratio_w * out_img_idx - in_img_idx; + T w2lambda = 1.f - w1lambda; + + const T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + + in_img_idy * in_img_w + in_img_idx]; + + // bilinear interpolation + out[out_id_h * output_w + out_id_w] = + h2lambda * (w2lambda * in_pos[0] + w1lambda * in_pos[w_id]) + + h1lambda * (w2lambda * in_pos[h_id * in_img_w] + + w1lambda * in_pos[h_id * in_img_w + w_id]); + } +} + +template +__global__ void KeBilinearInterpBw( + T* in, const size_t in_img_h, const size_t in_img_w, const size_t input_h, + const size_t input_w, const T* out, const size_t out_img_h, + const size_t out_img_w, const size_t output_h, const size_t output_w, + const size_t num_channels, const T ratio_h, const T ratio_w) { + int nthreads = output_h * output_w; + int tid = blockIdx.x * blockDim.x + threadIdx.x; + int stride = blockDim.x * gridDim.x; + for (; tid < nthreads; tid += stride) { + int out_id_h = tid / output_w; + int out_id_w = tid % output_w; + int in_img_size = input_w / num_channels; + int out_img_size = output_w / num_channels; + int channel_id = out_id_w / out_img_size; + + int out_img_idy = (out_id_w % out_img_size) / out_img_w; + int in_img_idy = ratio_h * out_img_idy; + int h_id = (in_img_idy < in_img_h - 1) ? 1 : 0; + T h1lambda = ratio_h * out_img_idy - in_img_idy; + T h2lambda = 1.f - h1lambda; + + int out_img_idx = tid % out_img_w; + int in_img_idx = ratio_w * out_img_idx; + int w_id = (in_img_idx < in_img_w - 1) ? 1 : 0; + T w1lambda = ratio_w * out_img_idx - in_img_idx; + T w2lambda = 1.f - w1lambda; + + T* in_pos = &in[out_id_h * input_w + channel_id * in_img_size + + in_img_idy * in_img_w + in_img_idx]; + const T* out_pos = &out[out_id_h * output_w + out_id_w]; + platform::CudaAtomicAdd(&in_pos[0], h2lambda * w2lambda * out_pos[0]); + platform::CudaAtomicAdd(&in_pos[w_id], h2lambda * w1lambda * out_pos[0]); + platform::CudaAtomicAdd(&in_pos[h_id * in_img_w], + h1lambda * w2lambda * out_pos[0]); + platform::CudaAtomicAdd(&in_pos[h_id * in_img_w + w_id], + h1lambda * w1lambda * out_pos[0]); + } +} + +template +class InterpolateOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()), + "This kernel only runs on GPU device."); + auto* input = ctx.Input("X"); + auto* output = ctx.Output("Out"); + auto* input_data = input->data(); + + auto interp_method = ctx.Attr("interp_method"); + int out_h = ctx.Attr("out_h"); + int out_w = ctx.Attr("out_w"); + auto out_size = ctx.Input("OutSize"); + if (out_size != nullptr) { + Tensor sizes; + framework::TensorCopy(*out_size, platform::CPUPlace(), &sizes); + auto size_data = sizes.data(); + out_h = size_data[0]; + out_w = size_data[1]; + } + + int n = input->dims()[0]; + int c = input->dims()[1]; + int in_h = input->dims()[2]; + int in_w = input->dims()[3]; + + auto* output_data = + output->mutable_data({n, c, out_h, out_w}, ctx.GetPlace()); + + int in_hw = in_h * in_w; + int out_hw = out_h * out_w; + int in_chw = c * in_hw; + int out_chw = c * out_hw; + + float ratio_h = + (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + float ratio_w = + (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + + if (in_h == out_h && in_w == out_w) { + framework::TensorCopy(*input, ctx.GetPlace(), output); + return; + } + + int pixelNum = n * out_chw; + int grid_dim = (pixelNum + 512 - 1) / 512; + grid_dim = grid_dim > 8 ? 8 : grid_dim; + + if ("nearest" == interp_method) { + KeNearestNeighborInterpFw< + T><<>>( + input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n, + out_chw, c, ratio_h, ratio_w); + } else if ("bilinear" == interp_method) { + KeBilinearInterpFw< + T><<>>( + input_data, in_h, in_w, n, in_chw, output_data, out_h, out_w, n, + out_chw, c, ratio_h, ratio_w); + } + } +}; + +template +class InterpolateGradOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input_grad = ctx.Output(framework::GradVarName("X")); + auto* output_grad = ctx.Input(framework::GradVarName("Out")); + auto* output_grad_data = output_grad->data(); + auto* input_grad_data = input_grad->mutable_data(ctx.GetPlace()); + + auto& device_ctx = + ctx.template device_context(); + math::SetConstant zero; + zero(device_ctx, input_grad, static_cast(0.0)); + + auto interp_method = ctx.Attr("interp_method"); + int out_h = ctx.Attr("out_h"); + int out_w = ctx.Attr("out_w"); + auto out_size = ctx.Input("OutSize"); + if (out_size != nullptr) { + Tensor sizes; + framework::TensorCopy(*out_size, platform::CPUPlace(), &sizes); + auto size_data = sizes.data(); + out_h = size_data[0]; + out_w = size_data[1]; + } + + int n = input_grad->dims()[0]; + int c = input_grad->dims()[1]; + int in_h = input_grad->dims()[2]; + int in_w = input_grad->dims()[3]; + + int in_hw = in_h * in_w; + int out_hw = out_h * out_w; + int in_chw = c * in_hw; + int out_chw = c * out_hw; + + float ratio_h = + (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + float ratio_w = + (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + + if (in_h == out_h && in_w == out_w) { + framework::TensorCopy(*output_grad, ctx.GetPlace(), input_grad); + return; + } + + int pixelNum = n * out_chw; + int grid_dim = (pixelNum + 512 - 1) / 512; + grid_dim = grid_dim > 8 ? 8 : grid_dim; + + if ("nearest" == interp_method) { + KeNearestNeighborInterpBw< + T><<>>( + input_grad_data, in_h, in_w, n, in_chw, output_grad_data, out_h, + out_w, n, out_chw, c, ratio_h, ratio_w); + } else if ("bilinear" == interp_method) { + KeBilinearInterpBw< + T><<>>( + input_grad_data, in_h, in_w, n, in_chw, output_grad_data, out_h, + out_w, n, out_chw, c, ratio_h, ratio_w); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL(interpolate, ops::InterpolateOpCUDAKernel, + ops::InterpolateOpCUDAKernel, + ops::InterpolateOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(interpolate_grad, + ops::InterpolateGradOpCUDAKernel, + ops::InterpolateGradOpCUDAKernel); diff --git a/paddle/fluid/operators/interpolate_op.h b/paddle/fluid/operators/interpolate_op.h new file mode 100644 index 0000000000000000000000000000000000000000..7fdb3e1f5a2ff82284d89dd0759e357978e1d873 --- /dev/null +++ b/paddle/fluid/operators/interpolate_op.h @@ -0,0 +1,236 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve. + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + http://www.apache.org/licenses/LICENSE-2.0 + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once +#include +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +template +using EigenTensor = framework::EigenTensor; +using Tensor = framework::Tensor; + +template +static void NearestNeighborInterpolate(const Tensor& input, Tensor* output, + const float ratio_h, const float ratio_w, + const int n, const int c, + const int out_h, const int out_w) { + auto input_t = EigenTensor::From(input); + auto output_t = EigenTensor::From(*output); + for (int k = 0; k < out_h; k++) { // loop for images + int in_k = static_cast(ratio_h * k + 0.5); + + for (int l = 0; l < out_w; l++) { + int in_l = static_cast(ratio_w * l + 0.5); + + for (int i = 0; i < n; i++) { // loop for batches + for (int j = 0; j < c; j++) { // loop for channels + output_t(i, j, k, l) = input_t(i, j, in_k, in_l); + } + } + } + } +} + +template +static void BilinearInterpolation(const Tensor& input, Tensor* output, + const float ratio_h, const float ratio_w, + const int in_h, const int in_w, const int n, + const int c, const int out_h, + const int out_w) { + auto input_t = EigenTensor::From(input); + auto output_t = EigenTensor::From(*output); + for (int k = 0; k < out_h; k++) { // loop for images + int y_n = static_cast(ratio_h * k); + int y_s = (y_n + 1) < (in_h - 1) ? (y_n + 1) : (in_h - 1); + float d_n = ratio_h * k - y_n; + float d_s = 1.f - d_n; + + for (int l = 0; l < out_w; l++) { + int x_w = static_cast(ratio_w * l); + int x_e = (x_w + 1) < (in_w - 1) ? (x_w + 1) : (in_w - 1); + float d_w = ratio_w * l - x_w; + float d_e = 1.f - d_w; + + for (int i = 0; i < n; i++) { // loop for batches + for (int j = 0; j < c; j++) { // loop for channels + // bilinear interpolation + output_t(i, j, k, l) = input_t(i, j, y_n, x_w) * d_s * d_e + + input_t(i, j, y_s, x_w) * d_n * d_e + + input_t(i, j, y_n, x_e) * d_s * d_w + + input_t(i, j, y_s, x_e) * d_n * d_w; + } + } + } + } +} + +template +static void NearestNeighborInterpolateGrad(const Tensor& output_grad, + Tensor* input_grad, + const float ratio_h, + const float ratio_w, const int n, + const int c, const int out_h, + const int out_w) { + auto input_grad_t = EigenTensor::From(*input_grad); + auto output_grad_t = EigenTensor::From(output_grad); + for (int k = 0; k < out_h; k++) { // loop for images + int in_k = static_cast(ratio_h * k + 0.5); + + for (int l = 0; l < out_w; l++) { + int in_l = static_cast(ratio_w * l + 0.5); + + for (int i = 0; i < n; i++) { // loop for batches + for (int j = 0; j < c; j++) { // loop for channels + input_grad_t(i, j, in_k, in_l) += output_grad_t(i, j, k, l); + } + } + } + } +} + +template +static void BilinearInterpolationGrad(const Tensor& output_grad, + Tensor* input_grad, const float ratio_h, + const float ratio_w, const int in_h, + const int in_w, const int n, const int c, + const int out_h, const int out_w) { + auto input_grad_t = EigenTensor::From(*input_grad); + auto output_grad_t = EigenTensor::From(output_grad); + for (int k = 0; k < out_h; k++) { // loop for images + int y_n = static_cast(ratio_h * k); + int y_s = (y_n + 1) < (in_h - 1) ? (y_n + 1) : (in_h - 1); + float d_n = ratio_h * k - y_n; + float d_s = 1.f - d_n; + + for (int l = 0; l < out_w; l++) { + int x_w = static_cast(ratio_w * l); + int x_e = (x_w + 1) < (in_w - 1) ? (x_w + 1) : (in_w - 1); + float d_w = ratio_w * l - x_w; + float d_e = 1.f - d_w; + + for (int i = 0; i < n; i++) { // loop for batches + for (int j = 0; j < c; j++) { // loop for channels + // bilinear interpolation grad + const T grad = output_grad_t(i, j, k, l); + input_grad_t(i, j, y_n, x_w) += static_cast(grad * d_s * d_e); + input_grad_t(i, j, y_s, x_w) += static_cast(grad * d_n * d_e); + input_grad_t(i, j, y_n, x_e) += static_cast(grad * d_s * d_w); + input_grad_t(i, j, y_s, x_e) += static_cast(grad * d_n * d_w); + } + } + } + } +} + +template +class InterpolateKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("X"); + auto* output = ctx.Output("Out"); + + std::string interp_method = ctx.Attr("interp_method"); + int out_h = ctx.Attr("out_h"); + int out_w = ctx.Attr("out_w"); + auto out_size = ctx.Input("OutSize"); + if (out_size != nullptr) { + auto out_size_data = out_size->data(); + out_h = out_size_data[0]; + out_w = out_size_data[1]; + } + + const int n = input->dims()[0]; + const int c = input->dims()[1]; + const int in_h = input->dims()[2]; + const int in_w = input->dims()[3]; + + output->mutable_data({n, c, out_h, out_w}, ctx.GetPlace()); + auto& device_ctx = + ctx.template device_context(); + math::SetConstant zero; + zero(device_ctx, output, static_cast(0.0)); + + if (in_h == out_h && in_w == out_w) { + framework::TensorCopy(*input, ctx.GetPlace(), output); + return; + } + + float ratio_h = + (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + float ratio_w = + (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + + if ("bilinear" == interp_method) { + BilinearInterpolation(*input, output, ratio_h, ratio_w, in_h, in_w, n, + c, out_h, out_w); + } else if ("nearest" == interp_method) { + NearestNeighborInterpolate(*input, output, ratio_h, ratio_w, n, c, + out_h, out_w); + } + } +}; + +template +class InterpolateGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* input = ctx.Input("X"); + auto* input_grad = ctx.Output(framework::GradVarName("X")); + auto* output_grad = ctx.Input(framework::GradVarName("Out")); + + std::string interp_method = ctx.Attr("interp_method"); + int out_h = ctx.Attr("out_h"); + int out_w = ctx.Attr("out_w"); + auto out_size = ctx.Input("OutSize"); + if (out_size != nullptr) { + auto out_size_data = out_size->data(); + out_h = out_size_data[0]; + out_w = out_size_data[1]; + } + + const int n = input->dims()[0]; + const int c = input->dims()[1]; + const int in_h = input->dims()[2]; + const int in_w = input->dims()[3]; + + input_grad->mutable_data({n, c, in_h, in_w}, ctx.GetPlace()); + auto& device_ctx = + ctx.template device_context(); + math::SetConstant zero; + zero(device_ctx, input_grad, static_cast(0.0)); + + if (in_h == out_h && in_w == out_w) { + framework::TensorCopy(*output_grad, ctx.GetPlace(), input_grad); + return; + } + + float ratio_h = + (out_h > 1) ? static_cast(in_h - 1) / (out_h - 1) : 0.f; + float ratio_w = + (out_w > 1) ? static_cast(in_w - 1) / (out_w - 1) : 0.f; + + if ("bilinear" == interp_method) { + BilinearInterpolationGrad(*output_grad, input_grad, ratio_h, ratio_w, + in_h, in_w, n, c, out_h, out_w); + } else if ("nearest" == interp_method) { + NearestNeighborInterpolateGrad(*output_grad, input_grad, ratio_h, + ratio_w, n, c, out_h, out_w); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/isfinite_op.cc b/paddle/fluid/operators/isfinite_op.cc index 248c7793560db99c0af06421bf74808422016061..7b42efd623b31a703bf51d2d157130b3120b42a4 100644 --- a/paddle/fluid/operators/isfinite_op.cc +++ b/paddle/fluid/operators/isfinite_op.cc @@ -60,7 +60,7 @@ class OverflowOpMaker : public framework::OpProtoAndCheckerMaker { "(Tensor) 1-dim tensor, contains a bool scalar. The output " "tensor of overflow operator."); AddComment(string::Sprintf(R"DOC( -Overflow operator. +Overflow %s operator. $$Out = any(X)$$ @@ -69,6 +69,8 @@ Out = Inf if any X contains Inf, Out = Nan if any X contains Nan, Out = 0 if no Inf/Nan detected. If X contains both Inf/Nan, it will return the first indicator it meeted. + +%s )DOC", GetName(), GetComments())); } diff --git a/paddle/fluid/operators/lars_momentum_op.cc b/paddle/fluid/operators/lars_momentum_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a8dda93902448fa1bd21b719ffd9c9b500caf755 --- /dev/null +++ b/paddle/fluid/operators/lars_momentum_op.cc @@ -0,0 +1,86 @@ +/* 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/lars_momentum_op.h" +#include "paddle/fluid/operators/momentum_op.h" + +namespace paddle { +namespace operators { + +class LarsMomentumOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("Param", + "(LoDTensor, default LoDTensor) " + "Input parameter that has to be updated"); + AddInput("Grad", + "(LoDTensor, default LoDTensor) " + "Input gradient of the parameter"); + AddInput("Velocity", + "(LoDTensor, default LoDTensor) " + "Input velocity (corresponding to the parameter) " + "that has to be updated"); + AddInput("LearningRate", + "(LoDTensor, default LoDTensor) " + "Input learning rate"); + + AddOutput("ParamOut", + "(LoDTensor) This output is updated parameter. " + "It shared memory with Input(Param)."); + AddOutput("VelocityOut", + "(LoDTensor) This output is updated velocity. " + "It shared memory with Input(Velocity)."); + + AddAttr("mu", "(float) Momentum coefficient"); + AddAttr("lars_coeff", "(float, default 0.001) LARS coefficient.") + .SetDefault(0.001); + AddAttr("lars_weight_decay", + "(float, default 0.0005) LARS weight decay") + .SetDefault(0.0005); + + AddComment(R"DOC( +Lars Momentum Optimizer. + +This optimizer use LARS (https://arxiv.org/abs/1708.03888) to optimize each +weight using a local learning rate: + +$$ +local\_lr = \eta * + \frac{\left \| param \right \|}{\left \| grad \right \| + \beta *\left \| param \right \|} \\ +velocity = mu * velocity + + local\_lr * (grad + \beta * param) \\ +param = param - velocity. \\ +$$ + +Note that we use lars_weight_decay here to decay weights, you may need not to +use L2 regularizers in case of using LARS. + +)DOC"); + } +}; + +class LarsMomentumOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override {} +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(lars_momentum, ops::MomentumOp, ops::LarsMomentumOpMaker, + paddle::framework::EmptyGradOpMaker, + ops::LarsMomentumOpVarTypeInference); +REGISTER_OP_CPU_KERNEL(lars_momentum, ops::LarsMomentumOpKernel, + ops::LarsMomentumOpKernel); diff --git a/paddle/fluid/operators/lars_momentum_op.cu b/paddle/fluid/operators/lars_momentum_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..eb346851a2f690fa05422c84ddcb08307539048f --- /dev/null +++ b/paddle/fluid/operators/lars_momentum_op.cu @@ -0,0 +1,94 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/lars_momentum_op.h" + +namespace paddle { +namespace operators { + +template +__global__ void MomentumLarsKernel(const T* p, const T* g, const T* v, + const T* learning_rate, const T mu, + const int64_t num, const T lars_coeff, + const T lars_weight_decay, const T* p_norm, + const T* g_norm, T* p_out, T* v_out) { + T lr = learning_rate[0]; + T local_lr = learning_rate[0]; + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; + i += blockDim.x * gridDim.x) { + if (p_norm[0] > 0 && g_norm[0] > 0) { + local_lr = lr * lars_coeff * p_norm[0] / + (g_norm[0] + lars_weight_decay * p_norm[0]); + } + T v_new = v[i] * mu + local_lr * (g[i] + lars_weight_decay * p[i]); + v_out[i] = v_new; + p_out[i] = p[i] - v_new; + } +} + +template +class LarsMomentumOpCUDAKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto param_out = ctx.Output("ParamOut"); + auto velocity_out = ctx.Output("VelocityOut"); + auto param = ctx.Input("Param"); + auto velocity = ctx.Input("Velocity"); + auto grad = ctx.Input("Grad"); + auto learning_rate = ctx.Input("LearningRate"); + + T* p_out = param_out->mutable_data(ctx.GetPlace()); + T* v_out = velocity_out->mutable_data(ctx.GetPlace()); + + T mu = static_cast(ctx.Attr("mu")); + T lars_coeff = ctx.Attr("lars_coeff"); + T lars_weight_decay = ctx.Attr("lars_weight_decay"); + + auto* p = param->data(); + auto* v = velocity->data(); + auto* g = grad->data(); + auto* lr = learning_rate->data(); + + int block = 512; + int grid = (param->numel() + block - 1) / block; + + auto eigen_p = framework::EigenVector::Flatten(*param); + auto eigen_g = framework::EigenVector::Flatten(*grad); + // calculate norms using eigein and launch the kernel. + framework::Tensor p_norm_t, g_norm_t; + p_norm_t.Resize({1}); + g_norm_t.Resize({1}); + auto* p_norm_data = p_norm_t.mutable_data(ctx.GetPlace()); + auto* g_norm_data = g_norm_t.mutable_data(ctx.GetPlace()); + auto ep_norm = framework::EigenScalar::From(p_norm_t); + auto eg_norm = framework::EigenScalar::From(g_norm_t); + + auto* place = ctx.template device_context().eigen_device(); + ep_norm.device(*place) = eigen_p.square().sum().sqrt(); + eg_norm.device(*place) = eigen_g.square().sum().sqrt(); + MomentumLarsKernel<<>>( + p, g, v, lr, mu, param->numel(), lars_coeff, lars_weight_decay, + p_norm_data, g_norm_data, p_out, v_out); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + lars_momentum, + ops::LarsMomentumOpCUDAKernel, + ops::LarsMomentumOpCUDAKernel); diff --git a/paddle/fluid/operators/lars_momentum_op.h b/paddle/fluid/operators/lars_momentum_op.h new file mode 100644 index 0000000000000000000000000000000000000000..e85be99fc42522e461a7915847d82144d8195a96 --- /dev/null +++ b/paddle/fluid/operators/lars_momentum_op.h @@ -0,0 +1,72 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class LarsMomentumOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto param_out = ctx.Output("ParamOut"); + auto velocity_out = ctx.Output("VelocityOut"); + auto param = ctx.Input("Param"); + auto velocity = ctx.Input("Velocity"); + auto learning_rate = ctx.Input("LearningRate"); + auto* grad_var = ctx.InputVar("Grad"); + // only support dense for now. + PADDLE_ENFORCE(grad_var->IsType()); + auto grad = ctx.Input("Grad"); + + param_out->mutable_data(ctx.GetPlace()); + velocity_out->mutable_data(ctx.GetPlace()); + + T mu = static_cast(ctx.Attr("mu")); + T lars_coeff = ctx.Attr("lars_coeff"); + T lars_weight_decay = ctx.Attr("lars_weight_decay"); + + auto p_out = framework::EigenVector::Flatten(*param_out); + auto v_out = framework::EigenVector::Flatten(*velocity_out); + + auto p = framework::EigenVector::Flatten(*param); + auto v = framework::EigenVector::Flatten(*velocity); + auto g = framework::EigenVector::Flatten(*grad); + auto* lr = learning_rate->data(); + + framework::Tensor p_norm_t, g_norm_t; + p_norm_t.Resize({1}); + g_norm_t.Resize({1}); + p_norm_t.mutable_data(ctx.GetPlace()); + g_norm_t.mutable_data(ctx.GetPlace()); + auto ep_norm = framework::EigenScalar::From(p_norm_t); + auto eg_norm = framework::EigenScalar::From(g_norm_t); + + ep_norm = p.square().sum().sqrt(); + eg_norm = g.square().sum().sqrt(); + T local_lr = lr[0]; + if (ep_norm(0) > 0 && eg_norm(0) > 0) { + local_lr = lr[0] * lars_coeff * ep_norm(0) / + (eg_norm(0) + lars_weight_decay * ep_norm(0)); + } + v_out = v * mu + local_lr * (g + lars_weight_decay * p); + p_out = p - v_out; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index dc008d16971bc762b401ddece56f9ec56f7a47d6..e3d09e2d14817fe0f2ccda18ed90c9436b399ae3 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -27,12 +27,16 @@ limitations under the License. */ #include "paddle/fluid/operators/distributed/request_handler_impl.h" #include "paddle/fluid/operators/listen_and_serv_op.h" +DEFINE_int32(rpc_send_thread_num, 5, "number of threads for rpc send"); +DEFINE_int32(rpc_get_thread_num, 5, "number of threads for rpc get"); +DEFINE_int32(rpc_prefetch_thread_num, 5, "number of threads for rpc prefetch"); + namespace paddle { namespace operators { void RunServer(std::shared_ptr service) { service->StartServer(); - VLOG(4) << "RunServer thread end"; + VLOG(40) << "RunServer thread end"; } static void split(const std::string &str, char sep, std::vector *pieces) { @@ -62,11 +66,11 @@ static void ParallelExecuteBlocks( fs.push_back(framework::Async([&executor, &prepared, &scope, idx]() { int run_block = idx; // thread local try { - VLOG(3) << "running server block: " << run_block - << "pointer: " << prepared[run_block].get(); + VLOG(30) << "running server block: " << run_block + << "pointer: " << prepared[run_block].get(); executor->RunPreparedContext(prepared[run_block].get(), scope); } catch (const std::exception &e) { - LOG(ERROR) << "run sub program error " << e.what(); + LOG(FATAL) << "run sub program:" << idx << " error " << e.what(); } })); } @@ -104,7 +108,7 @@ void ListenAndServOp::RunSyncLoop( framework::Scope *recv_scope, platform::DeviceContext *dev_ctx, const std::vector &prefetch_block_id_list, const int checkpoint_point_block_id) const { - VLOG(2) << "RunSyncLoop"; + VLOG(20) << "RunSyncLoop"; size_t num_blocks = program->Size(); auto optimize_blocks = Attr>(kOptimizeBlocks); @@ -130,7 +134,6 @@ void ListenAndServOp::RunSyncLoop( rpc_service_->ResetBarrierCounter(); while (true) { - rpc_service_->Profiler().OneStep(); // Get from multiple trainers, we don't care about the order in which // the gradients arrives, just add suffix 0~n and merge the gradient. rpc_service_->SetCond(distributed::kRequestSend); @@ -164,7 +167,7 @@ void ListenAndServOp::RunSyncLoop( } ParallelExecuteBlocks(parallel_blkids, executor, optimize_prepared, program, recv_scope); - VLOG(2) << "run all blocks spent " << GetTimestamp() - ts << "(ms)"; + VLOG(20) << "run all blocks spent " << GetTimestamp() - ts << "(ms)"; ResetReceivedVars(recv_scope, dev_ctx, rpc_service_->NeedResetAllVars()); @@ -180,11 +183,11 @@ void ListenAndServOp::ResetReceivedVars(framework::Scope *recv_scope, for (auto &varname : sparse_vars_) { auto var = recv_scope->FindVar(varname); if (var == nullptr) { - VLOG(2) << "can not find var " << varname << " in received scope"; + VLOG(20) << "can not find var " << varname << " in received scope"; continue; } if (var->IsType()) { - VLOG(3) << "reset sparse var: " << varname; + VLOG(30) << "reset sparse var: " << varname; var->GetMutable()->mutable_rows()->clear(); } else { PADDLE_THROW("The type of sparse var should be SelectedRows"); @@ -194,7 +197,7 @@ void ListenAndServOp::ResetReceivedVars(framework::Scope *recv_scope, for (auto &varname : dense_vars_) { auto var = recv_scope->FindVar(varname); if (var == nullptr) { - VLOG(2) << "can not find var " << varname << " in received scope"; + VLOG(20) << "can not find var " << varname << " in received scope"; continue; } if (var->IsType()) { @@ -213,24 +216,27 @@ void ListenAndServOp::ResetReceivedVars(framework::Scope *recv_scope, void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, framework::ProgramDesc *program, framework::Scope *recv_scope) const { - VLOG(2) << "RunAsyncLoop"; - // grad name to block id - std::unordered_map grad_to_block_id; - std::unordered_map id_to_grad; - + VLOG(20) << "RunAsyncLoop"; auto grad_to_block_id_str = Attr>("grad_to_block_id"); - for (const auto &grad_and_id : grad_to_block_id_str) { + DoubleFindMap grad_to_block_id; + + auto append_block_maps = [](DoubleFindMap *out_map, + const std::string &grad_and_id) { std::vector pieces; split(grad_and_id, ':', &pieces); - VLOG(3) << "after split, grad = " << pieces[0] << ", id=" << pieces[1]; + VLOG(30) << "after split, key = " << pieces[0] << ", id=" << pieces[1]; PADDLE_ENFORCE_EQ(pieces.size(), 2); - PADDLE_ENFORCE_EQ(grad_to_block_id.count(pieces[0]), 0); + PADDLE_ENFORCE_EQ(out_map->count(pieces[0]), 0); int block_id = std::stoi(pieces[1]); - grad_to_block_id[pieces[0]] = block_id; - id_to_grad[block_id] = pieces[0]; + (*out_map)[pieces[0]] = block_id; + }; + + for (const auto &grad_and_id : grad_to_block_id_str) { + append_block_maps(&grad_to_block_id, grad_and_id); } + size_t num_blocks = program->Size(); PADDLE_ENFORCE_GE(num_blocks, 2, "server program should have at least 2 blocks"); @@ -240,15 +246,22 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, block_list.push_back(blkid); } auto optimize_prepared = executor->Prepare(*program, block_list); - // execute global block if needed - if (block_list[0] == 1 && id_to_grad.count(1) == 0) { + // execute global block if needed, block id 1 in the program is global + // block if it's not bind to a grad var for it's update. + if (block_list[0] == 1 && + grad_to_block_id.find_value(static_cast(1)) == + grad_to_block_id.end()) { executor->RunPreparedContext(optimize_prepared[0].get(), recv_scope); } std::unordered_map> - grad_to_prepared_ctx; + grad_to_prepared_ctx, param_to_prepared_ctx; for (size_t i = 0; i < block_list.size(); ++i) { - grad_to_prepared_ctx[id_to_grad[block_list[i]]] = optimize_prepared[i]; + auto blkid = block_list[i]; + auto it = grad_to_block_id.find_value(blkid); + if (it != grad_to_block_id.end()) { + grad_to_prepared_ctx[it->first] = optimize_prepared[i]; + } } request_send_handler_->SetGradToPreparedCtx(&grad_to_prepared_ctx); @@ -257,7 +270,7 @@ void ListenAndServOp::RunAsyncLoop(framework::Executor *executor, while (true) { if (rpc_service_->IsExit()) { - VLOG(4) << "get exit!rpc_processor break!"; + VLOG(40) << "get exit!rpc_processor break!"; break; } @@ -311,6 +324,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, framework::Scope &recv_scope = scope.NewScope(); bool sync_mode = Attr("sync_mode"); + bool dc_sgd = Attr("dc_asgd"); auto fan_in = Attr("Fanin"); auto inputs = Inputs("X"); @@ -318,25 +332,30 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, std::string endpoint = Attr("endpoint"); int checkpoint_block_id = Attr(kCheckpointBlockId); - VLOG(4) << "sync_mode:" << sync_mode << ", fan_in:" << fan_in - << ", end_point:" << endpoint - << ", checkpoint_block_id: " << checkpoint_block_id; + VLOG(40) << "sync_mode:" << sync_mode << ", fan_in:" << fan_in + << ", end_point:" << endpoint + << ", checkpoint_block_id: " << checkpoint_block_id; rpc_service_.reset(new RPCSERVER_T(endpoint, fan_in)); - request_send_handler_.reset(new distributed::RequestSendHandler(sync_mode)); - request_get_handler_.reset(new distributed::RequestGetHandler(sync_mode)); + request_send_handler_.reset( + new distributed::RequestSendHandler(sync_mode, dc_sgd)); + request_get_handler_.reset( + new distributed::RequestGetHandler(sync_mode, dc_sgd)); request_prefetch_handler_.reset( new distributed::RequestPrefetchHandler(sync_mode)); request_checkpoint_handler_.reset(new distributed::RequestCheckpointHandler( sync_mode, checkpoint_block_id)); rpc_service_->RegisterRPC(distributed::kRequestSend, - request_send_handler_.get()); + request_send_handler_.get(), + FLAGS_rpc_send_thread_num); rpc_service_->RegisterRPC(distributed::kRequestGet, - request_get_handler_.get()); + request_get_handler_.get(), + FLAGS_rpc_get_thread_num); rpc_service_->RegisterRPC(distributed::kRequestPrefetch, - request_prefetch_handler_.get()); + request_prefetch_handler_.get(), + FLAGS_rpc_prefetch_thread_num); rpc_service_->RegisterRPC(distributed::kRequestCheckpoint, request_checkpoint_handler_.get()); @@ -364,8 +383,8 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, prefetch_var_name_to_block_id_str) { std::vector pieces; split(prefetch_var_name_and_id, ':', &pieces); - VLOG(3) << "after split, prefetch_var = " << pieces[0] - << ", id=" << pieces[1]; + VLOG(30) << "after split, prefetch_var = " << pieces[0] + << ", id=" << pieces[1]; PADDLE_ENFORCE_EQ(pieces.size(), 2); int block_id = std::stoi(pieces[1]); @@ -396,7 +415,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope, // start the server listening after all member initialized. server_thread_.reset(new std::thread(RunServer, rpc_service_)); - VLOG(3) << "wait server thread to become ready..."; + VLOG(30) << "wait server thread to become ready..."; rpc_service_->WaitServerReady(); // register SIGINT(from ctrl+C) and SIGTERM(from kill) signal handlers @@ -436,6 +455,8 @@ class ListenAndServOpMaker : public framework::OpProtoAndCheckerMaker { "a map from grad name to it's optimize block id") .SetDefault({}); AddAttr("sync_mode", "if works at sync_mode or not").SetDefault(true); + AddAttr("dc_asgd", "set to true will enable DC-ASGD training.") + .SetDefault(false); AddAttr>( kOptimizeBlocks, "Optimize blocks to run on server side.") .SetDefault({}); diff --git a/paddle/fluid/operators/listen_and_serv_op.h b/paddle/fluid/operators/listen_and_serv_op.h index 5f889793ab16249a4e06801090db087a089dbed1..9431978df836121baacc12ed4e1ee6b218cc7d7a 100644 --- a/paddle/fluid/operators/listen_and_serv_op.h +++ b/paddle/fluid/operators/listen_and_serv_op.h @@ -18,6 +18,7 @@ limitations under the License. */ #include #include #include +#include #include #include "paddle/fluid/framework/executor.h" @@ -37,6 +38,17 @@ constexpr char kCheckpointBlockId[] = "checkpint_block_id"; void RunServer(std::shared_ptr service); +template +class DoubleFindMap : public std::unordered_map { + public: + typename std::unordered_map::iterator find_value(TValue v) { + return std::find_if(this->begin(), this->end(), + [&v](const std::pair p) { + return p.second == v; + }); + } +}; + class ListenAndServOp : public framework::OperatorBase { public: ListenAndServOp(const std::string& type, diff --git a/paddle/fluid/operators/lod_rank_table_op.cc b/paddle/fluid/operators/lod_rank_table_op.cc index 166952fe23192799443ef9c9d1f7ba5056d19290..59ef9cb626d61f918c8ad1990a0f25030fb44ec6 100644 --- a/paddle/fluid/operators/lod_rank_table_op.cc +++ b/paddle/fluid/operators/lod_rank_table_op.cc @@ -30,9 +30,9 @@ class LoDRankTableOp : public framework::OperatorBase { auto x = scope.FindVar(Input("X"))->Get(); auto *out = scope.FindVar(Output("Out"))->GetMutable(); - VLOG(10) << "Level = " << static_cast(Attr("level")); + VLOG(100) << "Level = " << static_cast(Attr("level")); out->Reset(x.lod(), static_cast(Attr("level"))); - VLOG(10) << Input("X") << "'s lod information is " << *out; + VLOG(100) << Input("X") << "'s lod information is " << *out; } }; diff --git a/paddle/fluid/operators/lod_tensor_to_array_op.cc b/paddle/fluid/operators/lod_tensor_to_array_op.cc index 8eab83fcd247fcd099ae1fa5dab1e67c2081bf9c..e72337a3e6f7884c3a05372e8732647e5910f3e4 100644 --- a/paddle/fluid/operators/lod_tensor_to_array_op.cc +++ b/paddle/fluid/operators/lod_tensor_to_array_op.cc @@ -17,7 +17,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor_array.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" -#include "paddle/fluid/operators/math/concat.h" +#include "paddle/fluid/operators/math/concat_and_split.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/port.h" @@ -79,7 +79,7 @@ struct LoDTensorToArrayFunctor : public boost::static_visitor { template template void LoDTensorToArrayFunctorImpl::apply() { - math::ConcatGradFunctor func; + math::SplitFunctor func; func(*dev_ctx_, prev_functor_->input_, prev_functor_->ref_inputs_, 0, &prev_functor_->outputs_); } diff --git a/paddle/fluid/operators/lookup_table_op.cc b/paddle/fluid/operators/lookup_table_op.cc index b9ac54e446811889b647397ae1fbb11c28f46777..1878dfe8a897db1b8c948d325fa48a38ca224a2b 100644 --- a/paddle/fluid/operators/lookup_table_op.cc +++ b/paddle/fluid/operators/lookup_table_op.cc @@ -81,6 +81,12 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker { "Otherwise the given value indicates padding the output " "with zeros whenever lookup encounters it in Ids.") .SetDefault(kNoPadding); + // NOTE(minqiyang): grad_inplace is an temporal attribute, + // please do NOT set this attribute in python layer. + AddAttr("grad_inplace", + "(boolean, default false) " + "If the grad op reuse the input's variable.") + .SetDefault(false); AddComment(R"DOC( Lookup Table Operator. @@ -115,7 +121,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { - auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("W")); + auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Out")); return framework::OpKernelType(data_type, ctx.device_context()); } }; @@ -128,13 +134,13 @@ class LookupTableOpGradVarTypeInference : public framework::VarTypeInference { auto attr = op_desc.GetAttr("is_sparse"); bool is_sparse = boost::get(attr); if (is_sparse) { - VLOG(3) << "lookup_table_grad op " << framework::GradVarName("W") - << " is set to SelectedRows"; + VLOG(30) << "lookup_table_grad op " << framework::GradVarName("W") + << " is set to SelectedRows"; block->Var(out_var_name) ->SetType(framework::proto::VarType::SELECTED_ROWS); } else { - VLOG(3) << "lookup_table_grad op " << framework::GradVarName("W") - << " is set to LoDTensor"; + VLOG(30) << "lookup_table_grad op " << framework::GradVarName("W") + << " is set to LoDTensor"; block->Var(out_var_name)->SetType(framework::proto::VarType::LOD_TENSOR); } block->Var(out_var_name)->SetDataType(block->Var("W")->GetDataType()); diff --git a/paddle/fluid/operators/lookup_table_op.h b/paddle/fluid/operators/lookup_table_op.h index 58463dc4d6fd7cc3454de766814a947fee161070..e504c4f0cd5c0feaef4a251fad57b389a10a2ce7 100644 --- a/paddle/fluid/operators/lookup_table_op.h +++ b/paddle/fluid/operators/lookup_table_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/math/blas.h" namespace paddle { namespace operators { @@ -68,6 +69,7 @@ class LookupTableKernel : public framework::OpKernel { const auto *table = table_t.value().data(); auto *output = output_t->mutable_data(context.GetPlace()); + auto blas = math::GetBlas(context); for (int64_t i = 0; i < ids_numel; ++i) { if (padding_idx != kNoPadding && ids[i] == padding_idx) { memset(output + i * row_width, 0, row_width * sizeof(T)); @@ -75,8 +77,8 @@ class LookupTableKernel : public framework::OpKernel { PADDLE_ENFORCE_GE(ids[i], 0); auto id_index = table_t.Index(ids[i]); PADDLE_ENFORCE_GE(id_index, 0, "the input key should be exists."); - memcpy(output + i * row_width, table + id_index * row_width, - row_width * sizeof(T)); + blas.VCOPY(row_width, table + id_index * row_width, + output + i * row_width); } } } @@ -111,27 +113,37 @@ class LookupTableGradKernel : public framework::OpKernel { auto *ids_data = ids->data(); int64_t ids_num = ids->numel(); - framework::Vector new_rows; - new_rows.reserve(ids_num); - for (int64_t i = 0; i < ids_num; i++) { - new_rows.push_back(ids_data[i]); - } + std::vector new_rows; + new_rows.resize(ids_num); + std::memcpy(&new_rows[0], ids_data, ids_num * sizeof(int64_t)); d_table->set_rows(new_rows); auto *d_table_value = d_table->mutable_value(); d_table_value->Resize({ids_num, table_dim[1]}); - d_table_value->mutable_data(context.GetPlace()); - - d_table->set_height(table_dim[0]); - - auto *d_output_data = d_output->data(); - auto *d_table_data = d_table_value->data(); - - auto d_output_dims = d_output->dims(); - PADDLE_ENFORCE_EQ( - d_table_value->dims(), - framework::flatten_to_2d(d_output_dims, d_output_dims.size() - 1)); - memcpy(d_table_data, d_output_data, sizeof(T) * d_output->numel()); + // FIXME(minqiyang): + // memory optimization will NOT reuse Tensor with SelectedRows + // so we could just share the tensor here directly. + // However, the InferVarType method will infer the output SelectedRows + // to Tensor sometimes, which is a bug, so we will add an attribute + // here to indicate the inplace and remove this attribute after + // the InferVarType's bug was fixed + bool grad_inplace = context.Attr("grad_inplace"); + if (grad_inplace) { + d_table_value->ShareDataWith(*d_output); + } else { + d_table_value->mutable_data(context.GetPlace()); + + d_table->set_height(table_dim[0]); + + auto *d_output_data = d_output->data(); + auto *d_table_data = d_table_value->data(); + + auto d_output_dims = d_output->dims(); + PADDLE_ENFORCE_EQ( + d_table_value->dims(), + framework::flatten_to_2d(d_output_dims, d_output_dims.size() - 1)); + memcpy(d_table_data, d_output_data, sizeof(T) * d_output->numel()); + } } else { auto *ids = context.Input("Ids"); auto *d_output = context.Input(framework::GradVarName("Out")); diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt index 91101356436c26171eaca2fe01dfd4d937e71717..868a7a706471717ce0c8f268d5eaa6dc4789588c 100644 --- a/paddle/fluid/operators/math/CMakeLists.txt +++ b/paddle/fluid/operators/math/CMakeLists.txt @@ -1,10 +1,10 @@ if (NOT WIN32) -add_subdirectory(detail) + add_subdirectory(detail) endif(NOT WIN32) function(math_library TARGET) - # math_library is a function to create math library. - # The interface is the same as cc_library. + # math_library is a function to create math library. + # The interface is the same as cc_library. # But it handle split GPU/CPU code and link some common library. set(cc_srcs) set(cu_srcs) @@ -35,7 +35,7 @@ function(math_library TARGET) endfunction() # please add new math_library in alphabetical order -math_library(concat) +math_library(concat_and_split) math_library(context_project DEPS im2col math_function) math_library(cross_entropy) math_library(cos_sim_functor) @@ -43,24 +43,22 @@ math_library(depthwise_conv) math_library(im2col) if (NOT WIN32) # windows do not support avx functions yet. -math_library(gru_compute DEPS activation_functions math_function) -math_library(lstm_compute DEPS activation_functions) -# TODO(TJ): ugly workaround, clean me -cc_library(cpu_lstm_compute SRCS cpu_lstm_compute.cc DEPS activation_functions cblas cpu_info) + math_library(gru_compute DEPS activation_functions math_function) + math_library(lstm_compute DEPS activation_functions) endif (NOT WIN32) cc_library(blas SRCS blas.cc DEPS cblas framework_proto device_context) math_library(math_function DEPS blas) math_library(maxouting) math_library(pooling) -math_library(selected_rows_functor DEPS selected_rows math_function) +math_library(selected_rows_functor DEPS selected_rows math_function blas) math_library(sequence2batch) math_library(sequence_padding) math_library(sequence_pooling DEPS math_function) math_library(sequence_scale) math_library(softmax DEPS math_function) if (NOT WIN32) -math_library(matrix_bit_code) + math_library(matrix_bit_code) endif (NOT WIN32) math_library(unpooling) math_library(vol2col) @@ -70,9 +68,19 @@ cc_test(selected_rows_functor_test SRCS selected_rows_functor_test.cc DEPS selec cc_test(im2col_test SRCS im2col_test.cc DEPS im2col) cc_test(vol2col_test SRCS vol2col_test.cc DEPS vol2col) cc_test(sequence_padding_test SRCS sequence_padding_test.cc DEPS sequence_padding) +cc_test(sequence_pooling_test SRCS sequence_pooling_test.cc DEPS sequence_pooling) if(WITH_GPU) nv_test(math_function_gpu_test SRCS math_function_test.cu DEPS math_function) nv_test(selected_rows_functor_gpu_test SRCS selected_rows_functor_test.cu DEPS selected_rows_functor math_function) endif() -cc_test(concat_test SRCS concat_test.cc DEPS concat) +cc_test(concat_test SRCS concat_test.cc DEPS concat_and_split) cc_test(cpu_vec_test SRCS cpu_vec_test.cc DEPS blas cpu_info) + +set(JIT_KERNEL_SRCS jit_kernel.cc jit_kernel_blas.cc jit_kernel_exp.cc jit_kernel_rnn.cc jit_kernel_crf_decode.cc) +set(JIT_KERNEL_DEPS cpu_info cblas gflags enforce) +if(WITH_XBYAK) + list(APPEND JIT_KERNEL_SRCS jit_gen.cc jit_code.cc) + list(APPEND JIT_KERNEL_DEPS xbyak) +endif() +cc_library(jit_kernel SRCS ${JIT_KERNEL_SRCS} DEPS ${JIT_KERNEL_DEPS}) +cc_test(jit_kernel_test SRCS jit_kernel_test.cc DEPS jit_kernel) diff --git a/paddle/fluid/operators/math/algorithm.h b/paddle/fluid/operators/math/algorithm.h new file mode 100644 index 0000000000000000000000000000000000000000..2e75b6abce5e1f43742ee15bff1dac4801186cd4 --- /dev/null +++ b/paddle/fluid/operators/math/algorithm.h @@ -0,0 +1,90 @@ +// 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 // for int64_t +#include + +#include "paddle/fluid/platform/hostdevice.h" + +namespace paddle { +namespace operators { +namespace math { + +template +HOSTDEVICE inline int64_t BinarySearch(const T *x, int64_t num, const T &val) { + int64_t beg = 0, end = num - 1; + while (beg <= end) { + auto mid = ((beg + end) >> 1); + if (x[mid] == val) + return mid; + else if (x[mid] < val) + beg = mid + 1; + else + end = mid - 1; + } + return -1; +} + +template +HOSTDEVICE inline size_t LowerBound(const T *x, size_t num, const T &val) { +#ifdef __CUDA_ARCH__ + // The following code is from + // https://en.cppreference.com/w/cpp/algorithm/lower_bound + auto *first = x; + int64_t count = static_cast(num); + while (count > 0) { + int64_t step = (count >> 1); + auto *it = first + step; + if (*it < val) { + first = ++it; + count -= (step + 1); + } else { + count = step; + } + } + return static_cast(first - x); +#else + return static_cast(std::lower_bound(x, x + num, val) - x); +#endif +} + +template +HOSTDEVICE inline size_t UpperBound(const T *x, size_t num, const T &val) { +#ifdef __CUDA_ARCH__ + // The following code is from + // https://en.cppreference.com/w/cpp/algorithm/upper_bound + auto *first = x; + int64_t count = static_cast(num); + while (count > 0) { + auto step = (count >> 1); + auto *it = first + step; + if (val < *it) { + count = step; + } else { + first = ++it; + count -= (step + 1); + } + } + return static_cast(first - x); +#else + return static_cast(std::upper_bound(x, x + num, val) - x); +#endif +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/concat.cc b/paddle/fluid/operators/math/concat_and_split.cc similarity index 95% rename from paddle/fluid/operators/math/concat.cc rename to paddle/fluid/operators/math/concat_and_split.cc index 7b79f10e33d4474e279c6e46208722d6b52277fc..c6e17fd042f19bbeee3507e4cd64f49cff369682 100644 --- a/paddle/fluid/operators/math/concat.cc +++ b/paddle/fluid/operators/math/concat_and_split.cc @@ -12,7 +12,7 @@ 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/math/concat.h" +#include "paddle/fluid/operators/math/concat_and_split.h" #include namespace paddle { @@ -67,7 +67,7 @@ class ConcatFunctor { * each dimension must be the same, except the axis dimension. */ template -class ConcatGradFunctor { +class SplitFunctor { public: void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& input, @@ -111,7 +111,7 @@ class ConcatGradFunctor { }; #define DEFINE_FUNCTOR(type) \ template class ConcatFunctor; \ - template class ConcatGradFunctor; + template class SplitFunctor; FOR_ALL_TYPES(DEFINE_FUNCTOR); diff --git a/paddle/fluid/operators/math/concat.cu b/paddle/fluid/operators/math/concat_and_split.cu similarity index 90% rename from paddle/fluid/operators/math/concat.cu rename to paddle/fluid/operators/math/concat_and_split.cu index b59d86e661aff25eba8e770247e85845365d628b..760a065c1081d1e55901774b258ba524471b856b 100644 --- a/paddle/fluid/operators/math/concat.cu +++ b/paddle/fluid/operators/math/concat_and_split.cu @@ -15,7 +15,7 @@ limitations under the License. */ #include #include #include "paddle/fluid/framework/mixed_vector.h" -#include "paddle/fluid/operators/math/concat.h" +#include "paddle/fluid/operators/math/concat_and_split.h" #include "paddle/fluid/platform/cuda_primitives.h" #include "paddle/fluid/platform/float16.h" @@ -24,7 +24,7 @@ namespace operators { namespace math { template -__global__ void KernelConcat(T** inputs, const int* input_cols, int col_size, +__global__ void ConcatKernel(T** inputs, const int* input_cols, int col_size, const int output_rows, const int output_cols, T* output) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; @@ -50,7 +50,7 @@ __global__ void KernelConcat(T** inputs, const int* input_cols, int col_size, } template -__global__ void KernelConcat(T** inputs_data, const int fixed_in_col, +__global__ void ConcatKernel(T** inputs_data, const int fixed_in_col, const int out_rows, const int out_cols, T* output_data) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; @@ -67,9 +67,9 @@ __global__ void KernelConcat(T** inputs_data, const int fixed_in_col, } template -__global__ void KernelConcatGrad(const T* input_data, const int in_row, - const int in_col, const int* out_cols, - int out_cols_size, T** outputs_data) { +__global__ void SplitKernel(const T* input_data, const int in_row, + const int in_col, const int* out_cols, + int out_cols_size, T** outputs_data) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; int curr_segment = 0; int curr_offset = out_cols[0]; @@ -94,9 +94,9 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row, } template -__global__ void KernelConcatGrad(const T* input_data, const int in_row, - const int in_col, const int fixed_out_col, - T** outputs_data) { +__global__ void SplitKernel(const T* input_data, const int in_row, + const int in_col, const int fixed_out_col, + T** outputs_data) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) { int split = tid_x / fixed_out_col; @@ -170,11 +170,11 @@ class ConcatFunctor { dim3 grid_size = dim3(grid_cols, grid_rows, 1); if (sameShape) { - KernelConcat<<>>( + ConcatKernel<<>>( dev_ins_data, in_col, out_row, out_col, output->data()); } else { const int* dev_ins_col_data = inputs_col.CUDAData(context.GetPlace()); - KernelConcat<<>>( + ConcatKernel<<>>( dev_ins_data, dev_ins_col_data, static_cast(inputs_col.size()), out_row, out_col, output->data()); } @@ -189,7 +189,7 @@ class ConcatFunctor { * each dimension must be the same, except the axis dimension. */ template -class ConcatGradFunctor { +class SplitFunctor { public: void operator()(const platform::CUDADeviceContext& context, const framework::Tensor& input, @@ -248,11 +248,11 @@ class ConcatGradFunctor { dim3 grid_size = dim3(grid_cols, grid_rows, 1); if (sameShape) { - KernelConcatGrad<<>>( + SplitKernel<<>>( input.data(), in_row, in_col, out0_col, dev_out_gpu_data); } else { const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace()); - KernelConcatGrad<<>>( + SplitKernel<<>>( input.data(), in_row, in_col, dev_outs_col_data, static_cast(outputs_cols.size()), dev_out_gpu_data); } @@ -264,7 +264,7 @@ class ConcatGradFunctor { #define DEFINE_FUNCTOR(type) \ template class ConcatFunctor; \ - template class ConcatGradFunctor + template class SplitFunctor FOR_ALL_TYPES(DEFINE_FUNCTOR); diff --git a/paddle/fluid/operators/math/concat.h b/paddle/fluid/operators/math/concat_and_split.h similarity index 98% rename from paddle/fluid/operators/math/concat.h rename to paddle/fluid/operators/math/concat_and_split.h index 867a84fa873a2e90bdab7a5eecbb1755cb4b02d1..3a5eddcbf4af699a89ae1a21571337155699a1f3 100644 --- a/paddle/fluid/operators/math/concat.h +++ b/paddle/fluid/operators/math/concat_and_split.h @@ -54,7 +54,7 @@ class ConcatFunctor { * Output[1] = [[5,6]] */ template -class ConcatGradFunctor { +class SplitFunctor { public: void operator()(const DeviceContext& context, const framework::Tensor& input, const std::vector& ref_inputs, diff --git a/paddle/fluid/operators/math/concat_test.cc b/paddle/fluid/operators/math/concat_test.cc index a46f2d51ca64501a622b5b48b424dffa16efc5b4..8ba9e8e8ec1344edc3beaf7f4a58f99107cc0e9c 100644 --- a/paddle/fluid/operators/math/concat_test.cc +++ b/paddle/fluid/operators/math/concat_test.cc @@ -12,10 +12,10 @@ 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/math/concat.h" #include #include #include "paddle/fluid/framework/tensor_util.h" +#include "paddle/fluid/operators/math/concat_and_split.h" template void testConcat() { diff --git a/paddle/fluid/operators/math/cos_sim_functor.cu b/paddle/fluid/operators/math/cos_sim_functor.cu index 4e6ff5ee0a449b42762748ba1a103876beee01f2..537c7e47155fe9a12196869ceaed84fca198335b 100644 --- a/paddle/fluid/operators/math/cos_sim_functor.cu +++ b/paddle/fluid/operators/math/cos_sim_functor.cu @@ -51,7 +51,7 @@ struct CosSimDyFunctor { T* dy) const { const int block_size = 512; dim3 threads(block_size, 1); - dim3 grid(1, (rows + block_size - 1) / block_size); + dim3 grid((rows + block_size - 1) / block_size, 1); CosSimDyKernel<<>>( x_norm, y_norm, x, y, z, dz, rows, cols, dy); } diff --git a/paddle/fluid/operators/math/cpu_lstm_compute.cc b/paddle/fluid/operators/math/cpu_lstm_compute.cc deleted file mode 100644 index e96d1879331974e0873e13f171414bcfa8c45953..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/math/cpu_lstm_compute.cc +++ /dev/null @@ -1,43 +0,0 @@ -/* 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/math/cpu_lstm_compute.h" - -namespace paddle { -namespace operators { -namespace math { -#ifdef __AVX__ -template <> -void lstm_compute_ctht(float* gates, const float* ct_1, float* ct, - float* ht) { - namespace act = detail::forward::avx; - // gates: W_ch, W_ih, W_fh, W_oh - __m256 c, i, f, o; - c = _mm256_loadu_ps(gates); - i = _mm256_loadu_ps(gates + 8); - f = _mm256_loadu_ps(gates + 16); - o = _mm256_loadu_ps(gates + 24); - - /* C_t = C_t-1 * fgated + cand_gated * igated*/ - c = _mm256_mul_ps(act::Tanh(c), act::Sigmoid(i)); - i = _mm256_loadu_ps(ct_1); - f = _mm256_mul_ps(i, act::Sigmoid(f)); - f = _mm256_add_ps(c, f); - _mm256_storeu_ps(ct, f); - - /* H_t = act_cell(C_t) * ogated */ - o = _mm256_mul_ps(act::Tanh(f), act::Sigmoid(o)); - _mm256_storeu_ps(ht, o); -} -#endif -} // namespace math -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/math/cpu_lstm_compute.h b/paddle/fluid/operators/math/cpu_lstm_compute.h deleted file mode 100644 index 169a9e4b47f54851ad436428416eca879b78e186..0000000000000000000000000000000000000000 --- a/paddle/fluid/operators/math/cpu_lstm_compute.h +++ /dev/null @@ -1,64 +0,0 @@ -/* 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 "paddle/fluid/operators/math/cpu_vec.h" -#include "paddle/fluid/platform/cpu_info.h" -#ifdef __AVX__ -#include -#endif - -namespace paddle { -namespace operators { -namespace math { - -// TODO(TJ): ugly workaround, clean me -template -void lstm_compute_ctht(T* gates, const T* ct_1, T* ct, T* ht) { - // gates: W_ch, W_ih, W_fh, W_oh - vec_sigmoid(24, gates + 8, gates + 8); - vec_tanh(8, gates, gates); - const T *i = gates + 8, *f = gates + 16, *o = gates + 24; - const T min = SIGMOID_THRESHOLD_MIN; - const T max = SIGMOID_THRESHOLD_MAX; - for (int d = 0; d < 8; ++d) { - // C_t = C_t-1 * fgated + cand_gated * igated - ct[d] = ct_1[d] * f[d] + gates[d] * i[d]; - // H_t = act_cell(C_t) * ogated - T tmp = ct[d] * 2; - tmp = static_cast(0) - ((tmp < min) ? min : ((tmp > max) ? max : tmp)); - vec_exp(1, &tmp, &tmp); - tmp = static_cast(2) / (static_cast(1) + tmp) - static_cast(1); - ht[d] = tmp * o[d]; - } -} - -#ifdef __AVX__ -namespace detail { -namespace forward { -namespace avx { -__m256 Sigmoid(const __m256 a); -__m256 Tanh(const __m256 a); - -} // namespace avx -} // namespace forward -} // namespace detail - -template <> -void lstm_compute_ctht(float* gates, const float* ct_1, float* ct, - float* ht); - -#endif - -} // namespace math -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/math/cpu_vec.h b/paddle/fluid/operators/math/cpu_vec.h index 6a059968b79189458349e466079cc7a663a8e5ff..0aed253c80fc28560716cbcfa70f74ef9c84f9b6 100644 --- a/paddle/fluid/operators/math/cpu_vec.h +++ b/paddle/fluid/operators/math/cpu_vec.h @@ -125,10 +125,8 @@ inline void vec_scal(const int n, const float a, } template <> -inline void vec_scal(const int n, - const float a, - const float* x, - float* y) { +inline void vec_scal(const int n, const float a, + const float* x, float* y) { // TODO(TJ): enable me vec_scal(n, a, x, y); } @@ -181,10 +179,10 @@ inline void vec_bias_sub(const int n, const float a, } template <> -inline void vec_bias_sub(const int n, - const float a, - const float* x, - float* y) { +inline void vec_bias_sub(const int n, + const float a, + const float* x, + float* y) { // TODO(TJ): enable me vec_bias_sub(n, a, x, y); } @@ -242,7 +240,7 @@ inline void vec_cross(const int n, const float* x, } template <> -inline void vec_cross( +inline void vec_cross( const int n, const float* x, const float* y, const float* z, float* out) { // TODO(TJ): enable me vec_cross(n, x, y, z, out); @@ -296,10 +294,10 @@ inline void vec_add_bias(const int n, const float a, } template <> -inline void vec_add_bias(const int n, - const float a, - const float* x, - float* y) { +inline void vec_add_bias(const int n, + const float a, + const float* x, + float* y) { // TODO(TJ): enable me vec_add_bias(n, a, x, y); } @@ -390,9 +388,9 @@ inline void vec_sigmoid(const int n, const float* x, } template <> -inline void vec_sigmoid(const int n, - const float* x, - float* y) { +inline void vec_sigmoid(const int n, + const float* x, + float* y) { // TODO(TJ): enable me vec_sigmoid(n, x, y); } @@ -454,9 +452,8 @@ inline void vec_relu(const int n, const float* x, } template <> -inline void vec_relu(const int n, - const float* x, - float* y) { +inline void vec_relu(const int n, const float* x, + float* y) { // TODO(TJ): enable me vec_relu(n, x, y); } diff --git a/paddle/fluid/operators/math/cpu_vec_test.cc b/paddle/fluid/operators/math/cpu_vec_test.cc index 3ce66f49ed8354c49e8af26ca6eb48fef654a40b..18a586f8dd9f01357d9facca19c51ed5c293ffd2 100644 --- a/paddle/fluid/operators/math/cpu_vec_test.cc +++ b/paddle/fluid/operators/math/cpu_vec_test.cc @@ -96,8 +96,8 @@ void TestAndBench(const int n, std::function tgt, } auto et = GetCurrentUS(); - VLOG(3) << "Vec size " << n << ": refer takes: " << (et - mt) / repeat - << " us, tgt takes: " << (mt - st) / repeat; + VLOG(30) << "Vec size " << n << ": refer takes: " << (et - mt) / repeat + << " us, tgt takes: " << (mt - st) / repeat; for (int i = 0; i < n; ++i) { EXPECT_NEAR(ytgt_data[i], yref_data[i], 1e-3); } @@ -110,7 +110,7 @@ TEST(CpuVecTest, sigmoid) { TestAndBench(sz, vec_sigmoid, ref_sigmoid); TestAndBench(sz, vec_sigmoid, ref_sigmoid); TestAndBench(sz, vec_sigmoid, ref_sigmoid); - TestAndBench(sz, vec_sigmoid, + TestAndBench(sz, vec_sigmoid, ref_sigmoid); } TestAndBench(30, vec_sigmoid, ref_sigmoid); @@ -123,8 +123,7 @@ TEST(CpuVecTest, tanh) { TestAndBench(sz, vec_tanh, ref_tanh); TestAndBench(sz, vec_tanh, ref_tanh); TestAndBench(sz, vec_tanh, ref_tanh); - TestAndBench(sz, vec_tanh, - ref_tanh); + TestAndBench(sz, vec_tanh, ref_tanh); } TestAndBench(30, vec_tanh, ref_tanh); } @@ -136,8 +135,7 @@ TEST(CpuVecTest, relu) { TestAndBench(sz, vec_relu, ref_relu); TestAndBench(sz, vec_relu, ref_relu); TestAndBench(sz, vec_relu, ref_relu); - TestAndBench(sz, vec_relu, - ref_relu); + TestAndBench(sz, vec_relu, ref_relu); } TestAndBench(30, vec_relu, ref_relu); } @@ -170,7 +168,7 @@ TEST(CpuVecTest, inplace_sigmoid) { TestInplace(sz, vec_sigmoid, ref_sigmoid); TestInplace(sz, vec_sigmoid, ref_sigmoid); TestInplace(sz, vec_sigmoid, ref_sigmoid); - TestInplace(sz, vec_sigmoid, + TestInplace(sz, vec_sigmoid, ref_sigmoid); } TestInplace(30, vec_sigmoid, ref_sigmoid); @@ -183,8 +181,7 @@ TEST(CpuVecTest, inplace_tanh) { TestInplace(sz, vec_tanh, ref_tanh); TestInplace(sz, vec_tanh, ref_tanh); TestInplace(sz, vec_tanh, ref_tanh); - TestInplace(sz, vec_tanh, - ref_tanh); + TestInplace(sz, vec_tanh, ref_tanh); } TestInplace(30, vec_tanh, ref_tanh); } @@ -196,8 +193,7 @@ TEST(CpuVecTest, inplace_relu) { TestInplace(sz, vec_relu, ref_relu); TestInplace(sz, vec_relu, ref_relu); TestInplace(sz, vec_relu, ref_relu); - TestInplace(sz, vec_relu, - ref_relu); + TestInplace(sz, vec_relu, ref_relu); } TestInplace(30, vec_relu, ref_relu); } diff --git a/paddle/fluid/operators/math/cross_entropy.cu b/paddle/fluid/operators/math/cross_entropy.cu index c92341ea55ea21773acba33665e267b2f1c25fe3..cb200ec8d6ea533d546f3e01a16a48c88b14f677 100644 --- a/paddle/fluid/operators/math/cross_entropy.cu +++ b/paddle/fluid/operators/math/cross_entropy.cu @@ -21,6 +21,16 @@ namespace operators { namespace math { namespace { + +__device__ __forceinline__ float real_log(float x) { return logf(x); } + +__device__ __forceinline__ double real_log(double x) { return log(x); } + +__device__ __forceinline__ platform::float16 real_log( + const platform::float16& val) { + return static_cast(logf(static_cast(val))); +} + template __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label, const int N, const int D, @@ -29,8 +39,8 @@ __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label, i += blockDim.x * gridDim.x) { PADDLE_ASSERT(label[i] >= 0 && label[i] < D || label[i] == ignore_index); Y[i] = ignore_index == label[i] - ? 0 - : -math::TolerableValue()(log(X[i * D + label[i]])); + ? static_cast(0) + : -math::TolerableValue()(real_log(X[i * D + label[i]])); } } @@ -38,12 +48,12 @@ template __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, const int class_num) { int tid = threadIdx.x; - T val = 0; + T val(0); int idx = blockIdx.x * class_num + tid; int end = blockIdx.x * class_num + class_num; for (; idx < end; idx += blockDim.x) { - val += math::TolerableValue()(std::log(X[idx])) * label[idx]; + val += math::TolerableValue()(real_log(X[idx])) * label[idx]; } val = paddle::platform::reduceSum(val, tid, blockDim.x); @@ -53,8 +63,6 @@ __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, } } // namespace -using Tensor = framework::Tensor; - template class CrossEntropyFunctor { public: @@ -89,6 +97,8 @@ class CrossEntropyFunctor { template class CrossEntropyFunctor; template class CrossEntropyFunctor; +template class CrossEntropyFunctor; } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/math/cross_entropy.h b/paddle/fluid/operators/math/cross_entropy.h index e8aeb5d0575ac0f6b8761e97896df73578e8a103..99a4935186e1e6f9e3bf36eb029ce3d230510117 100644 --- a/paddle/fluid/operators/math/cross_entropy.h +++ b/paddle/fluid/operators/math/cross_entropy.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/platform/hostdevice.h" @@ -33,6 +34,26 @@ struct TolerableValue { } }; +// NOTE(dzh): float16 value clip behave different. +// 1. Our ValueClipping has a hardcore threshold 1e20 +// for float number. 1e20 will resulting in overflow in float16. +// 2. float16 should expose the the real number overflow to python. +// because mixed-training depends the inf/nan value to determine +// if the scale value will be adjusted. +// Also. In standard implementation of cross entropy, other +// framework not has the ValueClipping. +template <> +struct TolerableValue { + HOSTDEVICE platform::float16 operator()(const platform::float16& x) const { + if (platform::isfinite(x)) + return x; + else if (x > static_cast(0)) + return std::numeric_limits::max(); + else + return std::numeric_limits::min(); + } +}; + template class CrossEntropyFunctor { public: diff --git a/paddle/fluid/operators/math/depthwise_conv.cu b/paddle/fluid/operators/math/depthwise_conv.cu index 3be389912307f7aac6dda6d1018943eb8f08696d..66d37c3bf31ffa420cc527cb576dcdc5505a0960 100644 --- a/paddle/fluid/operators/math/depthwise_conv.cu +++ b/paddle/fluid/operators/math/depthwise_conv.cu @@ -46,17 +46,20 @@ __forceinline__ __device__ unsigned warp_id() { return ret; } +#define ARG_DEFINE_KernelDepthwiseConv \ + const T *const input_data, const T *const filter_data, const int batch_size, \ + const int output_channels, const int output_height, \ + const int output_width, const int input_channels, \ + const int input_height, const int input_width, \ + const int filter_multiplier, const int filter_height, \ + const int filter_width, const int stride_height, const int stride_width, \ + const int padding_height, const int padding_width, \ + const int dilate_height, const int dilate_width, T *const output_data + // A Cuda kernel to compute the depthwise convolution forward pass // in NCHW format. template -__device__ __inline__ void KernelDepthwiseConv( - const T* const input_data, const T* const filter_data, const int batch_size, - const int output_channels, const int output_height, const int output_width, - const int input_channels, const int input_height, const int input_width, - const int filter_multiplier, const int filter_height, - const int filter_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, const int dilate_height, - const int dilate_width, T* const output_data) { +__device__ __inline__ void KernelDepthwiseConv(ARG_DEFINE_KernelDepthwiseConv) { for (int w_out = threadIdx.x; w_out < output_width; w_out += blockDim.x) { for (int h_out = threadIdx.y; h_out < output_height; h_out += blockDim.y) { const int batch = blockIdx.y; @@ -97,42 +100,105 @@ __device__ __inline__ void KernelDepthwiseConv( } } -template -__global__ void KernelDepthwiseConvSp( - const T* const input_data, const T* const filter_data, const int batch_size, - const int output_channels, const int output_height, const int output_width, - const int input_channels, const int input_height, const int input_width, - const int filter_multiplier, const int filter_height, - const int filter_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, const int dilate_height, - const int dilate_width, T* const output_data) { - if (c_filter_multiplier == 0) - KernelDepthwiseConv(input_data, filter_data, batch_size, output_channels, - output_height, output_width, input_channels, - input_height, input_width, filter_multiplier, - filter_height, filter_width, stride_height, - stride_width, padding_height, padding_width, - dilate_height, dilate_width, output_data); +template +__device__ __inline__ void KernelDepthwiseConvCFilter( + ARG_DEFINE_KernelDepthwiseConv) { + const int kWeghtSize = c_filter * c_filter; + T r_weight[kWeghtSize]; + const int batch = blockIdx.y; + const int c_out = blockIdx.x; + const T* weight = filter_data + c_out * c_filter * c_filter; + for (int i = 0; i < c_filter * c_filter; i++) r_weight[i] = weight[i]; - else - KernelDepthwiseConv(input_data, filter_data, batch_size, output_channels, - output_height, output_width, input_channels, - input_height, input_width, c_filter_multiplier, - filter_height, filter_height, c_stride, c_stride, - padding_height, padding_width, dilate_height, - dilate_width, output_data); + for (int w_out = threadIdx.x; w_out < output_width; w_out += blockDim.x) { + for (int h_out = threadIdx.y; h_out < output_height; h_out += blockDim.y) { + const int batch = blockIdx.y; + const int c_out = blockIdx.x; + + const int c_in = c_out / filter_multiplier; + T value = 0; + const int h_in_start = -padding_height + h_out * stride_height; + const int w_in_start = -padding_width + w_out * stride_width; + const int h_in_end = h_in_start + c_filter * dilate_height; + const int w_in_end = w_in_start + c_filter * dilate_width; + + const int in_offset = + ((batch * input_channels + c_in) * input_height) * input_width; + + const int h_end = h_in_end < input_height ? h_in_end : input_height; + const int w_end = w_in_end < input_width ? w_in_end : input_width; + const int h_start = h_in_start > 0 ? h_in_start : 0; + const int w_start = w_in_start > 0 ? w_in_start : 0; + + for (int h_in = h_in_start, h_f = 0; h_f < c_filter; + h_in += dilate_height, h_f++) { + for (int w_in = w_in_start, w_f = 0; w_f < c_filter; + w_in += dilate_width, w_f++) { + if (h_in >= 0 && h_in < input_height && w_in >= 0 && + w_in < input_width) { + const int offset = in_offset + h_in * input_width + w_in; + value += r_weight[h_f * c_filter + w_f] * input_data[offset]; + } + } + } + int index = + ((batch * gridDim.x + c_out) * output_height + h_out) * output_width + + w_out; + output_data[index] = value; + } + } +} + +template +__global__ void KernelDepthwiseConvSp(ARG_DEFINE_KernelDepthwiseConv) { + if (c_filter_multiplier == 0) { + if (c_filter == -1) + KernelDepthwiseConv( + input_data, filter_data, batch_size, output_channels, output_height, + output_width, input_channels, input_height, input_width, + filter_multiplier, filter_height, filter_width, stride_height, + stride_width, padding_height, padding_width, dilate_height, + dilate_width, output_data); + else + KernelDepthwiseConvCFilter( + input_data, filter_data, batch_size, output_channels, output_height, + output_width, input_channels, input_height, input_width, + filter_multiplier, filter_height, filter_width, stride_height, + stride_width, padding_height, padding_width, dilate_height, + dilate_width, output_data); + } else { + if (c_filter == -1) + KernelDepthwiseConv(input_data, filter_data, batch_size, + output_channels, output_height, output_width, + input_channels, input_height, input_width, + c_filter_multiplier, filter_height, filter_height, + c_stride, c_stride, padding_height, padding_width, + dilate_height, dilate_width, output_data); + else + KernelDepthwiseConvCFilter( + input_data, filter_data, batch_size, output_channels, output_height, + output_width, input_channels, input_height, input_width, + c_filter_multiplier, filter_height, filter_height, c_stride, c_stride, + padding_height, padding_width, dilate_height, dilate_width, + output_data); + } } // CUDA kernel to compute the depthwise convolution backprop w.r.t input. +#define ARG_DEFINE_KernelDepthwiseConvInputGrad \ + const T *const output_grad_data, const T *const filter_data, \ + const int batch_size, const int output_channels, \ + const int output_height, const int output_width, \ + const int input_channels, const int input_height, const int input_width, \ + const int filter_multiplier, const int filter_height, \ + const int filter_width, const int stride_height, const int stride_width, \ + const int padding_height, const int padding_width, \ + const int dilate_height, const int dilate_width, \ + T *const input_grad_data + template __device__ __inline__ void KernelDepthwiseConvInputGrad( - const T* const output_grad_data, const T* const filter_data, - const int batch_size, const int output_channels, const int output_height, - const int output_width, const int input_channels, const int input_height, - const int input_width, const int filter_multiplier, const int filter_height, - const int filter_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, const int dilate_height, - const int dilate_width, T* const input_grad_data) { + ARG_DEFINE_KernelDepthwiseConvInputGrad) { for (int w_in = threadIdx.x; w_in < input_width; w_in += blockDim.x) { for (int h_in = threadIdx.y; h_in < input_height; h_in += blockDim.y) { const int batch = blockIdx.y; @@ -184,15 +250,67 @@ __device__ __inline__ void KernelDepthwiseConvInputGrad( } } -template +template +__device__ __inline__ void KernelDepthwiseConvInputGradCFilter( + ARG_DEFINE_KernelDepthwiseConvInputGrad) { + const int kWeghtSize = c_filter * c_filter * c_filter_multiplier + 1; + T r_weight[kWeghtSize]; + const int batch = blockIdx.y; + const int c_in = blockIdx.x; + + for (int c_i = 0; c_i < filter_multiplier; c_i++) { + int c_out = c_in * filter_multiplier + c_i; + const T* weight = filter_data + c_out * c_filter * c_filter; + for (int i = 0; i < c_filter * c_filter; i++) + r_weight[i + c_i * c_filter * c_filter] = + weight[c_filter * c_filter - i - 1]; + } + + for (int w_in = threadIdx.x; w_in < input_width; w_in += blockDim.x) { + for (int h_in = threadIdx.y; h_in < input_height; h_in += blockDim.y) { + const int batch = blockIdx.y; + const int c_in = blockIdx.x; + + int h_out_start = h_in - (c_filter - 1) * dilate_height + padding_height; + + int w_out_start = w_in - (c_filter - 1) * dilate_width + padding_width; + + T value = 0; + + for (int c_i = 0; c_i < filter_multiplier; c_i++) { + int c_out = c_in * filter_multiplier + c_i; + for (int h_out = h_out_start, h_f = 0; h_f < c_filter; + h_out += dilate_height, h_f++) { + for (int w_out = w_out_start, w_f = 0; w_f < c_filter; + w_out += dilate_width, w_f++) { + int s_h_out = h_out / stride_height; + int s_w_out = w_out / stride_width; + if (h_out % stride_height == 0 && w_out % stride_width == 0 && + s_h_out >= 0 && s_h_out < output_height && s_w_out >= 0 && + s_w_out < output_width) { + const int output_grad_offset = + ((batch * output_channels + c_out) * output_height + + s_h_out) * + output_width + + s_w_out; + value += + output_grad_data[output_grad_offset] * + r_weight[h_f * c_filter + w_f + c_i * c_filter * c_filter]; + } + } + } + } + int index = + ((batch * gridDim.x + c_in) * input_height + h_in) * input_width + + w_in; + input_grad_data[index] = value; + } + } +} + +template __global__ void KernelDepthwiseConvInputGradSp( - const T* const output_grad_data, const T* const filter_data, - const int batch_size, const int output_channels, const int output_height, - const int output_width, const int input_channels, const int input_height, - const int input_width, const int filter_multiplier, const int filter_height, - const int filter_width, const int stride_height, const int stride_width, - const int padding_height, const int padding_width, const int dilate_height, - const int dilate_width, T* const input_grad_data) { + ARG_DEFINE_KernelDepthwiseConvInputGrad) { if (c_filter_multiplier == 0) KernelDepthwiseConvInputGrad( output_grad_data, filter_data, batch_size, output_channels, @@ -200,13 +318,20 @@ __global__ void KernelDepthwiseConvInputGradSp( filter_multiplier, filter_height, filter_width, stride_height, stride_width, padding_height, padding_width, dilate_height, dilate_width, input_grad_data); - else + else if (c_filter == -1) KernelDepthwiseConvInputGrad( output_grad_data, filter_data, batch_size, output_channels, output_height, output_width, input_channels, input_height, input_width, c_filter_multiplier, filter_height, filter_width, c_stride, c_stride, padding_height, padding_width, dilate_height, dilate_width, input_grad_data); + else + KernelDepthwiseConvInputGradCFilter( + output_grad_data, filter_data, batch_size, output_channels, + output_height, output_width, input_channels, input_height, input_width, + c_filter_multiplier, filter_height, filter_width, c_stride, c_stride, + padding_height, padding_width, dilate_height, dilate_width, + input_grad_data); } // Cuda kernel to compute the depthwise convolution backprop w.r.t. filter. @@ -325,12 +450,14 @@ class DepthwiseConvFunctor { dim3 threads(std::min(output_width, thread), blocks, 1); dim3 grid(output_channels, batch_size, 1); int filter_multiplier = output_channels / input_channels; -#define check_case(c_filter_multiplier, c_stride) \ +#define check_case(c_filter_multiplier, c_stride, c_filter) \ if (c_filter_multiplier == 0 || \ filter_multiplier == c_filter_multiplier && \ - stride_height == stride_width && stride_height == c_stride) { \ - KernelDepthwiseConvSp<<>>( \ + stride_height == stride_width && stride_height == c_stride && \ + (ksize_height == ksize_width && ksize_height == c_filter || \ + c_filter == -1)) { \ + KernelDepthwiseConvSp<<>>( \ input_data, filter_data, batch_size, output_channels, output_height, \ output_width, input_channels, input_height, input_width, \ filter_multiplier, ksize_height, ksize_width, stride_height, \ @@ -338,11 +465,17 @@ class DepthwiseConvFunctor { dilate_width, output_data); \ return; \ } - check_case(1, 1); - check_case(1, 2); - // NOTE(liangdun): 0,0 for other case - // add other case if needed, e.g. check_case(2^n,1) - check_case(0, 0); + check_case(1, 1, 3); + check_case(1, 1, 5); + check_case(1, 1, -1); + check_case(1, 2, 3); + check_case(1, 2, 5); + check_case(1, 2, -1); + check_case(0, 0, 3); + check_case(0, 0, 5); + check_case(0, 0, -1); +// NOTE(liangdun): 0,0 for other case +// add other case if needed, e.g. check_case(2^n,1) #undef check_case } }; @@ -384,13 +517,15 @@ class DepthwiseConvInputGradFunctor { dim3 grid(input_channels, batch_size, 1); int filter_multiplier = output_channels / input_channels; -#define check_case(c_filter_multiplier, c_stride) \ +#define check_case(c_filter_multiplier, c_stride, c_filter) \ if (c_filter_multiplier == 0 || \ filter_multiplier == c_filter_multiplier && \ - stride_height == stride_width && stride_height == c_stride) { \ + stride_height == stride_width && stride_height == c_stride && \ + (ksize_height == ksize_width && ksize_height == c_filter || \ + c_filter == -1)) { \ KernelDepthwiseConvInputGradSp< \ - T, c_filter_multiplier, \ - c_stride><<>>( \ + T, c_filter_multiplier, c_stride, \ + c_filter><<>>( \ output_grad_data, filter_data, batch_size, output_channels, \ output_height, output_width, input_channels, input_height, \ input_width, filter_multiplier, ksize_height, ksize_width, \ @@ -398,11 +533,21 @@ class DepthwiseConvInputGradFunctor { dilate_height, dilate_width, input_grad_data); \ return; \ } - check_case(1, 1); - check_case(1, 2); - // NOTE(liangdun): 0,0 for other case - // add other case if needed, e.g. check_case(2^n,1) - check_case(0, 0); + check_case(1, 1, 3); + check_case(1, 1, 5); + check_case(1, 1, -1); + check_case(1, 2, 3); + check_case(1, 2, 5); + check_case(1, 2, -1); + check_case(2, 1, 3); + check_case(2, 1, 5); + check_case(2, 1, -1); + check_case(2, 2, 3); + check_case(2, 2, 5); + check_case(2, 2, -1); + check_case(0, 0, -1); +// NOTE(liangdun): 0,0 for other case +// add other case if needed, e.g. check_case(2^n,1) #undef check_case } }; diff --git a/paddle/fluid/operators/math/fc_compute.h b/paddle/fluid/operators/math/fc_compute.h index 1f5a49c0ab5a10b0d7dc1febd258ce76c467cb1c..b072b4c20a171d148bd892c162436d03da404fb9 100644 --- a/paddle/fluid/operators/math/fc_compute.h +++ b/paddle/fluid/operators/math/fc_compute.h @@ -15,6 +15,7 @@ limitations under the License. */ #pragma once #include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/jit_kernel.h" DECLARE_int32(paddle_num_threads); @@ -30,20 +31,25 @@ inline void FCCompute(const BlasT& blas, const int M, if (B == NULL) { return; } + if (relu) { + const auto& vaddrelu = jitkernel::KernelPool::Instance() + .template Get>(N); + for (int i = 0; i < M; i++) { + T* dst = Y + i * N; + vaddrelu->Compute(B, dst, dst, N); + } + } else { + const auto& vadd = jitkernel::KernelPool::Instance() + .template Get>(N); #ifdef PADDLE_WITH_MKLML #pragma omp parallel for if (FLAGS_paddle_num_threads > 1) #endif - for (int i = 0; i < M; i++) { - blas.AXPY(N, static_cast(1), B, Y + i * N); + for (int i = 0; i < M; i++) { + T* dst = Y + i * N; + vadd->Compute(B, dst, dst, N); + } } - - if (!relu) { - return; - } - - // TODO(TJ): fuse relu - LOG(FATAL) << "Not implemented!"; } } // namespace math diff --git a/paddle/fluid/operators/math/jit_code.cc b/paddle/fluid/operators/math/jit_code.cc new file mode 100644 index 0000000000000000000000000000000000000000..6b3eecfbd11471b5d95dcb10c91acc536f78cb85 --- /dev/null +++ b/paddle/fluid/operators/math/jit_code.cc @@ -0,0 +1,125 @@ +/* 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/math/jit_code.h" +#include "paddle/fluid/operators/math/jit_kernel.h" +#include "paddle/fluid/platform/cpu_info.h" + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace gen { + +using namespace platform::jit; // NOLINT + +bool VXXJitCode::init(int d, int scalar_index) { + // It's not necessary to use avx512 since it would slow down the frequency + // and this kernel is not compute bound. + return MayIUse(avx) && scalar_index >= 0 && scalar_index <= 2; +} + +void VXXJitCode::generate() { + // do not need push stack, and do not need save avx512reg if do not use avx512 + int offset = 0; + if (with_relu_) { + vxorps(ymm_zero, ymm_zero, ymm_zero); + } + if (scalar_index_ == 1) { + vbroadcastss(ymm_src1, ptr[param1]); + } else if (scalar_index_ == 2) { + vbroadcastss(ymm_src2, ptr[param2]); + } + for (int i = 0; i < num_ / AVX_FLOAT_BLOCK; ++i) { + if (scalar_index_ != 1) { + vmovups(ymm_src1, ptr[param1 + offset]); + } + if (scalar_index_ != 2) { + vmovups(ymm_src2, ptr[param2 + offset]); + } + if (type_ == operand_type::mul) { + vmulps(ymm_dst, ymm_src1, ymm_src2); + } else if (type_ == operand_type::add) { + vaddps(ymm_dst, ymm_src1, ymm_src2); + } + if (with_relu_) { + vmaxps(ymm_dst, ymm_zero, ymm_dst); + } + vmovups(ptr[param3 + offset], ymm_dst); + offset += sizeof(float) * AVX_FLOAT_BLOCK; + } + int rest = num_ % AVX_FLOAT_BLOCK; + if (rest >= 4) { + if (scalar_index_ != 1) { + vmovups(xmm_src1, ptr[param1 + offset]); + } + if (scalar_index_ != 2) { + vmovups(xmm_src2, ptr[param2 + offset]); + } + if (type_ == operand_type::mul) { + vmulps(xmm_dst, xmm_src1, xmm_src2); + } else if (type_ == operand_type::add) { + vaddps(xmm_dst, xmm_src1, xmm_src2); + } + if (with_relu_) { + vmaxps(xmm_dst, xmm_zero, xmm_dst); + } + vmovups(ptr[param3 + offset], xmm_dst); + offset += sizeof(float) * 4; + rest -= 4; + } + if (rest >= 2) { + if (scalar_index_ != 1) { + vmovups(xmm_src1, ptr[param1 + offset]); + } + if (scalar_index_ != 2) { + vmovups(xmm_src2, ptr[param2 + offset]); + } + if (type_ == operand_type::mul) { + vmulps(xmm_dst, xmm_src1, xmm_src2); + } else if (type_ == operand_type::add) { + vaddps(xmm_dst, xmm_src1, xmm_src2); + } + if (with_relu_) { + vmaxps(xmm_dst, xmm_zero, xmm_dst); + } + vmovq(ptr[param3 + offset], xmm_dst); + offset += sizeof(float) * 2; + rest -= 2; + } + if (rest > 0) { + if (scalar_index_ != 1) { + vmovups(xmm_src1, ptr[param1 + offset]); + } + if (scalar_index_ != 2) { + vmovups(xmm_src2, ptr[param2 + offset]); + } + if (type_ == operand_type::mul) { + vmulss(xmm_dst, xmm_src1, xmm_src2); + } else if (type_ == operand_type::add) { + vaddss(xmm_dst, xmm_src1, xmm_src2); + } + if (with_relu_) { + vmaxps(xmm_dst, xmm_zero, xmm_dst); + } + vmovss(ptr[param3 + offset], xmm_dst); + } + ret(); +} + +} // namespace gen +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_code.h b/paddle/fluid/operators/math/jit_code.h new file mode 100644 index 0000000000000000000000000000000000000000..aaedb0ae10323eeddfba9512d9e47c7a22320610 --- /dev/null +++ b/paddle/fluid/operators/math/jit_code.h @@ -0,0 +1,92 @@ +/* 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 "paddle/fluid/operators/math/jit_gen.h" +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace gen { + +using reg64_t = const Xbyak::Reg64; +using reg32_t = const Xbyak::Reg32; +using xmm_t = const Xbyak::Xmm; +using ymm_t = const Xbyak::Ymm; +using zmm_t = const Xbyak::Zmm; +using Label = Xbyak::Label; + +typedef enum { mul = 0, add } operand_type; + +// function: vec = Operand(vec(or scalar), vec(or scalar)) (maybe with relu) +class VXXJitCode : public JitCode { + public: + const char* name() const override { + std::string base = "VXXJitCode"; + if (scalar_index_ == 1) { + base += "_Scalar"; + } else { + base += "_Vec"; + } + if (type_ == operand_type::mul) { + base += "_Mul"; + } else if (type_ == operand_type::add) { + base += "_Add"; + } + if (scalar_index_ == 2) { + base += "_Scalar"; + } else { + base += "_Vec"; + } + base += (with_relu_ ? "_Relu" : ""); + return base.c_str(); + } + explicit VXXJitCode(int d, operand_type type, int scalar_index, + bool with_relu, size_t code_size = 256 * 1024, + void* code_ptr = nullptr) + : JitCode(code_size, code_ptr), + num_(d), + type_(type), + scalar_index_(scalar_index), + with_relu_(with_relu) {} + static bool init(int d, int scalar_index = 0); + void generate() override; + + private: + int num_; + operand_type type_; + int scalar_index_; + bool with_relu_; + reg64_t param1{abi_param1}; + reg64_t param2{abi_param2}; + reg64_t param3{abi_param3}; + + xmm_t xmm_src1 = xmm_t(0); + xmm_t xmm_src2 = xmm_t(1); + xmm_t xmm_dst = xmm_t(2); + xmm_t xmm_zero = xmm_t(3); + + ymm_t ymm_src1 = ymm_t(0); + ymm_t ymm_src2 = ymm_t(1); + ymm_t ymm_dst = ymm_t(2); + ymm_t ymm_zero = ymm_t(3); +}; + +} // namespace gen +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_gen.cc b/paddle/fluid/operators/math/jit_gen.cc new file mode 100644 index 0000000000000000000000000000000000000000..6af39518ed926554c8c839bba701d3827923dba0 --- /dev/null +++ b/paddle/fluid/operators/math/jit_gen.cc @@ -0,0 +1,90 @@ +/* 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/math/jit_gen.h" +#include +#include +#include +#include "paddle/fluid/platform/cpu_info.h" + +DEFINE_bool(dump_jitcode, false, "Whether to dump the jitcode to file"); + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace gen { + +constexpr Xbyak::Operand::Code g_abi_regs[] = { + Xbyak::Operand::RBX, Xbyak::Operand::RBP, Xbyak::Operand::R12, + Xbyak::Operand::R13, Xbyak::Operand::R14, Xbyak::Operand::R15}; + +constexpr int num_g_abi_regs = sizeof(g_abi_regs) / sizeof(g_abi_regs[0]); + +void JitCode::preCode() { + for (int i = 0; i < num_g_abi_regs; ++i) { + push(Xbyak::Reg64(g_abi_regs[i])); + } + if (platform::jit::MayIUse(platform::jit::avx512f)) { + mov(reg_EVEX_max_8b_offt, 2 * EVEX_max_8b_offt); + } +} + +void JitCode::postCode() { + for (int i = 0; i < num_g_abi_regs; ++i) { + pop(Xbyak::Reg64(g_abi_regs[num_g_abi_regs - 1 - i])); + } + ret(); +} + +void JitCode::dumpCode(const Xbyak::uint8 *code) const { + if (code) { + static int counter = 0; + std::ostringstream filename; + filename << "paddle_jitcode_" << name() << "." << counter << ".bin"; + counter++; + std::ofstream fout(filename.str(), std::ios::out); + if (fout.is_open()) { + fout.write(reinterpret_cast(code), getSize()); + fout.close(); + } + } +} + +Xbyak::Address JitCode::EVEX_compress_addr(Xbyak::Reg64 base, int offt, + bool bcast) { + int scale = 0; + if (EVEX_max_8b_offt <= offt && offt < 3 * EVEX_max_8b_offt) { + offt = offt - 2 * EVEX_max_8b_offt; + scale = 1; + } else if (3 * EVEX_max_8b_offt <= offt && offt < 5 * EVEX_max_8b_offt) { + offt = offt - 4 * EVEX_max_8b_offt; + scale = 2; + } + auto re = Xbyak::RegExp() + base + offt; + if (scale) { + re = re + reg_EVEX_max_8b_offt * scale; + } + if (bcast) { + return zword_b[re]; + } else { + return zword[re]; + } +} + +} // namespace gen +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_gen.h b/paddle/fluid/operators/math/jit_gen.h new file mode 100644 index 0000000000000000000000000000000000000000..6abf3434cc8d8f6ab2838ef822a4f6b948331802 --- /dev/null +++ b/paddle/fluid/operators/math/jit_gen.h @@ -0,0 +1,80 @@ +/* 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 "paddle/fluid/platform/macros.h" + +#define XBYAK_USE_MMAP_ALLOCATOR +#include "xbyak/xbyak.h" +#include "xbyak/xbyak_util.h" + +DECLARE_bool(dump_jitcode); + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace gen { + +#define DECLARE_JIT_CODE(codename) \ + const char *name() const override { return #codename; } + +// Application Binary Interface +constexpr Xbyak::Operand::Code abi_param1(Xbyak::Operand::RDI), + abi_param2(Xbyak::Operand::RSI), abi_param3(Xbyak::Operand::RDX), + abi_param4(Xbyak::Operand::RCX), abi_not_param1(Xbyak::Operand::RCX); + +class JitCode : public Xbyak::CodeGenerator { + public: + explicit JitCode(size_t code_size = 256 * 1024, void *code_ptr = nullptr) + : Xbyak::CodeGenerator(code_size, code_ptr) {} + + virtual ~JitCode() {} + virtual const char *name() const = 0; + virtual void generate() = 0; + + template + const FUNC getCode() { + this->generate(); + const Xbyak::uint8 *code = CodeGenerator::getCode(); + if (FLAGS_dump_jitcode) { + this->dumpCode(code); + } + return reinterpret_cast(code); + } + DISABLE_COPY_AND_ASSIGN(JitCode); + + protected: + Xbyak::Reg64 param1{abi_param1}; + const int EVEX_max_8b_offt = 0x200; + const Xbyak::Reg64 reg_EVEX_max_8b_offt = rbp; + + void preCode(); + void postCode(); + void dumpCode(const Xbyak::uint8 *code) const; + void L(const char *label) { Xbyak::CodeGenerator::L(label); } + void L(const Xbyak::Label &label) { Xbyak::CodeGenerator::L(label); } + // Enhanced vector extension + Xbyak::Address EVEX_compress_addr(Xbyak::Reg64 base, int offt, + bool bcast = false); +}; + +} // namespace gen +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel.cc b/paddle/fluid/operators/math/jit_kernel.cc new file mode 100644 index 0000000000000000000000000000000000000000..68b708b345334bc63b5e2e88c308d20ca6378e6b --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel.cc @@ -0,0 +1,41 @@ +/* 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/math/jit_kernel.h" +#include +#include + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +namespace jit = platform::jit; + +KernelPool& KernelPool::Instance() { + static thread_local KernelPool g_jit_kernels; + return g_jit_kernels; +} + +std::shared_ptr KernelPool::Get(const std::string& key) const { + if (kers_.find(key) == kers_.end()) { + return nullptr; + } + return kers_.at(key); +} + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel.h b/paddle/fluid/operators/math/jit_kernel.h new file mode 100644 index 0000000000000000000000000000000000000000..e9b259282cd00cc2afc46634423ec09590bf5dd3 --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel.h @@ -0,0 +1,166 @@ +/* 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 // for shared_ptr +#include +#include +#include "paddle/fluid/platform/cpu_info.h" +#include "paddle/fluid/platform/macros.h" + +// Note: Only support on CPU yet. +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +#define SIGMOID_THRESHOLD_MIN -40.0 +#define SIGMOID_THRESHOLD_MAX 13.0 +#define EXP_MAX_INPUT 40.0 +#define AVX_FLOAT_BLOCK 8 +#define AVX2_FLOAT_BLOCK 8 +#define AVX512_FLOAT_BLOCK 16 + +typedef enum { kLT8, kEQ8, kGT8LT16, kEQ16, kGT16 } jit_block; + +class Kernel { + public: + Kernel() = default; + virtual ~Kernel() = default; + // TODO(TJ): below members should be deprecated. + int num_{0}; + int end_{0}; + int rest_{0}; + DISABLE_COPY_AND_ASSIGN(Kernel); +}; + +class KernelPool { + public: + static KernelPool &Instance(); + + template + std::shared_ptr Get(ARGS... args); + + std::shared_ptr Get(const std::string &key) const; + + private: + KernelPool() = default; + std::unordered_map> kers_; + + DISABLE_COPY_AND_ASSIGN(KernelPool); +}; + +template +class VMulKernel : public Kernel { + public: + void (*Compute)(const T *, const T *, T *, int); +}; + +template +class VAddKernel : public Kernel { + public: + void (*Compute)(const T *, const T *, T *, int); +}; + +template +class VAddReluKernel : public Kernel { + public: + void (*Compute)(const T *, const T *, T *, int); +}; + +template +class VScalKernel : public Kernel { + public: + // y = a.*x + void (*Compute)(const T *, const T *, T *, int); +}; + +template +class VAddBiasKernel : public Kernel { + public: + // y = a.+x + void (*Compute)(const T *, const T *, T *, int); +}; + +template +class VActKernel : public Kernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VReluKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VIdentityKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VExpKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VSigmoidKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class VTanhKernel : public VActKernel { + public: + virtual void Compute(const T *x, T *y) const = 0; +}; + +template +class LSTMKernel : public Kernel { + public: + virtual void ComputeCtHt(T *gates, const T *ct_1, T *ct, T *ht, + /* below only used in peephole*/ + const T *wp_data = nullptr, + T *checked = nullptr) const = 0; + + // compute c1 and h1 without c0 or h0 + virtual void ComputeC1H1(T *gates, T *ct, T *ht, + /* below only used in peephole*/ + const T *wp_data = nullptr) const = 0; +}; + +template +class GRUKernel : public Kernel { + public: + // compute h1 without h0 + virtual void ComputeH1(T *gates, T *ht) const = 0; + virtual void ComputeHtPart1(T *gates, const T *ht_1, T *ht) const = 0; + virtual void ComputeHtPart2(T *gates, const T *ht_1, T *ht) const = 0; +}; + +template +class CRFDecodeKernel : public Kernel { + public: + virtual void Compute(const int seq_len, const T *x, const T *w, T *alpha, + int *track) const = 0; +}; + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_blas.cc b/paddle/fluid/operators/math/jit_kernel_blas.cc new file mode 100644 index 0000000000000000000000000000000000000000..c4bfbcf925a2bbdc39f8468049c58e126d3eba1b --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_blas.cc @@ -0,0 +1,470 @@ +/* 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/math/jit_kernel.h" +#include +#include "paddle/fluid/operators/math/jit_kernel_macro.h" +#include "paddle/fluid/platform/enforce.h" + +#ifdef PADDLE_WITH_XBYAK +#include "paddle/fluid/operators/math/jit_code.h" +#endif + +#ifdef PADDLE_WITH_MKLML +#include "paddle/fluid/platform/dynload/mklml.h" +#endif + +#ifdef __AVX__ +#include +#endif + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace jit = platform::jit; + +template +void VMulRefer(const T* x, const T* y, T* z, int n) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] * y[i]; + } +} + +template +void VAddRefer(const T* x, const T* y, T* z, int n) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] + y[i]; + } +} + +template +void VAddReluRefer(const T* x, const T* y, T* z, int n) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] + y[i]; + z[i] = z[i] > 0 ? z[i] : 0; + } +} + +template +void VScalRefer(const T* a, const T* x, T* y, int n) { + for (int i = 0; i < n; ++i) { + y[i] = a[0] * x[i]; + } +} + +template +void VAddBiasRefer(const T* a, const T* x, T* y, int n) { + for (int i = 0; i < n; ++i) { + y[i] = a[0] + x[i]; + } +} + +#ifdef PADDLE_WITH_MKLML +template +void VMulMKL(const T* x, const T* y, T* z, int n); + +template <> +void VMulMKL(const float* x, const float* y, float* z, int n) { + platform::dynload::vsMul(n, x, y, z); +} + +template <> +void VMulMKL(const double* x, const double* y, double* z, int n) { + platform::dynload::vdMul(n, x, y, z); +} + +template +void VAddMKL(const T* x, const T* y, T* z, int n); + +template <> +void VAddMKL(const float* x, const float* y, float* z, int n) { + platform::dynload::vsAdd(n, x, y, z); +} + +template <> +void VAddMKL(const double* x, const double* y, double* z, int n) { + platform::dynload::vdAdd(n, x, y, z); +} + +template +void VScalMKL(const T* a, const T* x, T* y, int n); + +template <> +void VScalMKL(const float* a, const float* x, float* y, int n) { + if (x == y) { + platform::dynload::cblas_sscal(n, *a, y, 1); + } else { + VScalRefer(a, x, y, n); + } +} + +template <> +void VScalMKL(const double* a, const double* x, double* y, int n) { + if (x == y) { + platform::dynload::cblas_dscal(n, *a, y, 1); + } else { + VScalRefer(a, x, y, n); + } +} + +#endif + +#define DECLARE_STATIC_FUNC \ + static inline std::string name(int d) { \ + PADDLE_THROW("DType should be either float or double"); \ + } \ + static inline bool useJIT(int d) { return false; } \ + static inline bool useMKL(int d) { return false; } + +/* VMUL JitKernel */ +template +class VMulKernelImpl : public VMulKernel { + public: + DECLARE_STATIC_FUNC; + explicit VMulKernelImpl(int d) : VMulKernel() { +#ifdef PADDLE_WITH_XBYAK + if (useJIT(d)) { + // roughly estimate the size of code + size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8; + jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::mul, 0, false, + sz > 4096 ? sz : 4096)); + this->Compute = + jitcode_->getCode(); + return; + } +#endif +#ifdef PADDLE_WITH_MKLML + if (useMKL(d)) { + this->Compute = VMulMKL; + return; + } +#endif + this->Compute = VMulRefer; + } + +#ifdef PADDLE_WITH_XBYAK + + private: + std::unique_ptr jitcode_{nullptr}; +#endif +}; + +#ifdef PADDLE_WITH_XBYAK +template <> +bool VMulKernelImpl::useJIT(int d) { + return gen::VXXJitCode::init(d); +} +#endif + +#ifdef PADDLE_WITH_MKLML +template <> +bool VMulKernelImpl::useMKL(int d) { + return jit::MayIUse(jit::avx512f) && d > 512; +} + +template <> +bool VMulKernelImpl::useMKL(int d) { + return true; +} +#endif + +/* VAdd JitKernel */ +template +class VAddKernelImpl : public VAddKernel { + public: + DECLARE_STATIC_FUNC; + explicit VAddKernelImpl(int d) : VAddKernel() { +#ifdef PADDLE_WITH_XBYAK + if (useJIT(d)) { + size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8; + jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 0, false, + sz > 4096 ? sz : 4096)); + this->Compute = + jitcode_->getCode(); + return; + } +#endif +#ifdef PADDLE_WITH_MKLML + if (useMKL(d)) { + this->Compute = VAddMKL; + return; + } +#endif + this->Compute = VAddRefer; + } +#ifdef PADDLE_WITH_XBYAK + + private: + std::unique_ptr jitcode_{nullptr}; +#endif +}; + +#ifdef PADDLE_WITH_XBYAK +template <> +bool VAddKernelImpl::useJIT(int d) { + return gen::VXXJitCode::init(d); +} +#endif + +#ifdef PADDLE_WITH_MKLML +template <> +bool VAddKernelImpl::useMKL(int d) { + return d > 512; +} + +template <> +bool VAddKernelImpl::useMKL(int d) { + return true; +} +#endif + +/* VAddRelu JitKernel */ +template +class VAddReluKernelImpl : public VAddReluKernel { + public: + DECLARE_STATIC_FUNC; + explicit VAddReluKernelImpl(int d) : VAddReluKernel() { +#ifdef PADDLE_WITH_XBYAK + if (useJIT(d)) { + size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8; + jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 0, true, + sz > 4096 ? sz : 4096)); + this->Compute = + jitcode_->getCode(); + return; + } +#endif + this->Compute = VAddReluRefer; + } +#ifdef PADDLE_WITH_XBYAK + + private: + std::unique_ptr jitcode_{nullptr}; +#endif +}; + +#ifdef PADDLE_WITH_XBYAK +template <> +bool VAddReluKernelImpl::useJIT(int d) { + return gen::VXXJitCode::init(d); +} +#endif + +/* VScal JitKernel */ +template +class VScalKernelImpl : public VScalKernel { + public: + DECLARE_STATIC_FUNC; + explicit VScalKernelImpl(int d) : VScalKernel() { +#ifdef PADDLE_WITH_XBYAK + if (useJIT(d)) { + size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8; + jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::mul, 1, false, + sz > 4096 ? sz : 4096)); + this->Compute = + jitcode_->getCode(); + return; + } +#endif +#ifdef PADDLE_WITH_MKLML + if (useMKL(d)) { + this->Compute = VScalMKL; + return; + } +#endif + this->Compute = VScalRefer; + } +#ifdef PADDLE_WITH_XBYAK + + private: + std::unique_ptr jitcode_{nullptr}; +#endif +}; + +#ifdef PADDLE_WITH_XBYAK +template <> +bool VScalKernelImpl::useJIT(int d) { + return gen::VXXJitCode::init(d, 1); +} +#endif + +#ifdef PADDLE_WITH_MKLML +template <> +bool VScalKernelImpl::useMKL(int d) { + return d > 512; +} +template <> +bool VScalKernelImpl::useMKL(int d) { + return true; +} +#endif + +/* VAddBias JitKernel */ +template +class VAddBiasKernelImpl : public VAddBiasKernel { + public: + DECLARE_STATIC_FUNC; + explicit VAddBiasKernelImpl(int d) : VAddBiasKernel() { +#ifdef PADDLE_WITH_XBYAK + if (useJIT(d)) { + size_t sz = 96 + d / AVX_FLOAT_BLOCK * 4 * 8; + jitcode_.reset(new gen::VXXJitCode(d, gen::operand_type::add, 1, false, + sz > 4096 ? sz : 4096)); + this->Compute = + jitcode_->getCode(); + return; + } +#endif + + this->Compute = VAddBiasRefer; + } +#ifdef PADDLE_WITH_XBYAK + + private: + std::unique_ptr jitcode_{nullptr}; +#endif +}; + +#ifdef PADDLE_WITH_XBYAK +template <> +bool VAddBiasKernelImpl::useJIT(int d) { + return gen::VXXJitCode::init(d, 1); +} +#endif + +#undef DECLARE_STATIC_FUNC + +REGISTER_JITKERNEL(vmul, VMulKernel); +REGISTER_JITKERNEL(vadd, VAddKernel); +REGISTER_JITKERNEL(vaddrelu, VAddReluKernel); +REGISTER_JITKERNEL(vscal, VScalKernel); +REGISTER_JITKERNEL(vaddbias, VAddBiasKernel); + +/* VRelu JitKernel */ +template +class VReluKernelImpl : public VReluKernel { + public: + explicit VReluKernelImpl(int d) : VReluKernel() { this->num_ = d; } + void Compute(const T* x, T* y) const override { + for (int i = 0; i < this->num_; ++i) { + y[i] = x[i] > 0 ? x[i] : 0; + } + } +}; + +#define INTRI8_FLOAT(isa) \ + template <> \ + void VReluKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + tmp = _mm256_max_ps(tmp, _mm256_setzero_ps()); \ + _mm256_storeu_ps(y, tmp); \ + } + +#define INTRI16_FLOAT(isa) \ + template <> \ + void VReluKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 zeros = _mm256_setzero_ps(); \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + tmp0 = _mm256_max_ps(tmp0, zeros); \ + tmp1 = _mm256_max_ps(tmp1, zeros); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#define INTRI_GT8LT16_FLOAT(isa) \ + template <> \ + VReluKernelImpl::VReluKernelImpl(int d) \ + : VReluKernel() { \ + this->num_ = d; \ + this->end_ = AVX_FLOAT_BLOCK; \ + this->rest_ = d - AVX_FLOAT_BLOCK; \ + } \ + template <> \ + void VReluKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 zeros = _mm256_setzero_ps(); \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + this->rest_); \ + tmp0 = _mm256_max_ps(tmp0, zeros); \ + tmp1 = _mm256_max_ps(tmp1, zeros); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + this->rest_, tmp1); \ + } + +#define INTRI_GT16_FLOAT(isa) \ + template <> \ + VReluKernelImpl::VReluKernelImpl(int d) \ + : VReluKernel() { \ + this->num_ = d; \ + this->end_ = d - d % AVX_FLOAT_BLOCK; \ + this->rest_ = d - AVX_FLOAT_BLOCK; \ + } \ + template <> \ + void VReluKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 zeros = _mm256_setzero_ps(); \ + for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \ + __m256 tmp = _mm256_loadu_ps(x + i); \ + tmp = _mm256_max_ps(tmp, zeros); \ + _mm256_storeu_ps(y + i, tmp); \ + } \ + __m256 tmp = _mm256_loadu_ps(x + this->rest_); \ + tmp = _mm256_max_ps(tmp, zeros); \ + _mm256_storeu_ps(y + this->rest_, tmp); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +INTRI16_FLOAT(jit::avx); +INTRI_GT8LT16_FLOAT(jit::avx); +INTRI_GT16_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +INTRI16_FLOAT(jit::avx2); +INTRI_GT8LT16_FLOAT(jit::avx2); +INTRI_GT16_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +// TODO(TJ): refine avx512 +INTRI8_FLOAT(jit::avx512f); +INTRI16_FLOAT(jit::avx512f); +INTRI_GT8LT16_FLOAT(jit::avx512f); +INTRI_GT16_FLOAT(jit::avx512f); +#endif + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef INTRI_GT8LT16_FLOAT +#undef INTRI_GT16_FLOAT + +/* An empty JitKernel */ +template +class VIdentityKernelImpl : public VIdentityKernel { + public: + explicit VIdentityKernelImpl(int d) : VIdentityKernel() { this->num_ = d; } + void Compute(const T* x, T* y) const override {} +}; + +REGISTER_JITKERNEL_DEPRECATED(vrelu, VReluKernel); +REGISTER_JITKERNEL_DEPRECATED(videntity, VIdentityKernel); + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_crf_decode.cc b/paddle/fluid/operators/math/jit_kernel_crf_decode.cc new file mode 100644 index 0000000000000000000000000000000000000000..a4861c347e44ad86a066861d3375b556302a84bc --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_crf_decode.cc @@ -0,0 +1,296 @@ +/* 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/math/jit_kernel.h" +#include +#include +#include "paddle/fluid/operators/math/jit_kernel_macro.h" +#ifdef __AVX__ +#include +#endif + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +namespace jit = platform::jit; + +/* CRF Decode JitKernel */ +template +class CRFDecodeKernelImpl : public CRFDecodeKernel { + public: + explicit CRFDecodeKernelImpl(int tag_num) : CRFDecodeKernel() { + this->num_ = tag_num; + } + void Compute(const int seq_len, const T* x, const T* w, T* alpha, + int* track) const override { + constexpr int state_trans_base_idx = 2; + for (int i = 0; i < this->num_; ++i) { + alpha[i] = w[i] + x[i]; + } + for (int k = 1; k < seq_len; ++k) { + for (int i = 0; i < this->num_; ++i) { + T max_score = -std::numeric_limits::max(); + int max_j = 0; + for (int j = 0; j < this->num_; ++j) { + T score = alpha[(k - 1) * this->num_ + j] + + w[(j + state_trans_base_idx) * this->num_ + i]; + if (score > max_score) { + max_score = score; + max_j = j; + } + } + alpha[k * this->num_ + i] = max_score + x[k * this->num_ + i]; + track[k * this->num_ + i] = max_j; + } + } + } +}; + +#define INIT_ALPHA(step_size) \ + /* Setup the alpha initial value.*/ \ + int i_offset = 0; \ + int last_offset = this->rest_ - step_size; \ + for (int i = 0; i <= this->end_; ++i) { \ + /* weights, input and alpha values. */ \ + __m256 w_content, x_content, alpha_content; \ + /* Load the relevant data into the variables from un-aligned address.*/ \ + w_content = _mm256_loadu_ps(w + i_offset); \ + x_content = _mm256_loadu_ps(x + i_offset); \ + alpha_content = _mm256_add_ps(w_content, x_content); \ + _mm256_storeu_ps(alpha + i_offset, alpha_content); \ + i_offset += step_size; \ + if (i == this->end_ - 1) { \ + if (this->rest_ > 0) { \ + i_offset += last_offset; \ + } else { \ + break; \ + } \ + } \ + } + +#define UPDATE_ALPHA(step_size) \ + /* Update the alpha and track values. */ \ + __m256 x_content = _mm256_loadu_ps(x + seq_offset + this->num_ + j_offset); \ + max_score = _mm256_add_ps(max_score, x_content); \ + _mm256_storeu_ps(alpha + seq_offset + this->num_ + j_offset, max_score); \ + _mm256_storeu_si256( \ + reinterpret_cast<__m256i*>(track + seq_offset + this->num_ + j_offset), \ + max_j); \ + /* Calculate the offset of next step*/ \ + j_offset += step_size; \ + if (j == this->end_ - 1) { \ + if (this->rest_ > 0) { \ + j_offset += last_offset; \ + } else { \ + break; \ + } \ + } + +#define INTRIAVX_FLOAT(block) \ + template <> \ + CRFDecodeKernelImpl::CRFDecodeKernelImpl( \ + int tag_num) \ + : CRFDecodeKernel() { \ + this->num_ = tag_num; \ + this->end_ = this->num_ / AVX_FLOAT_BLOCK; \ + this->rest_ = this->num_ % AVX_FLOAT_BLOCK; \ + } \ + template <> \ + void CRFDecodeKernelImpl::Compute( \ + const int seq_len, const float* x, const float* w, float* alpha, \ + int* track) const { \ + INIT_ALPHA(AVX_FLOAT_BLOCK) \ + /* Use the column-major strategy to get the location of maximum score.*/ \ + int seq_offset = 0; \ + constexpr int state_trans_base_idx = 2; \ + for (int k = 1; k < seq_len; ++k) { \ + int j_offset = 0; \ + for (int j = 0; j <= this->end_; ++j) { \ + /* Initialize the variables of maximum score and location.*/ \ + __m256 max_score = _mm256_set1_ps(-std::numeric_limits::max()); \ + __m256i max_j = _mm256_set1_epi32(0); \ + /* Calculate the offset of transition_weights.*/ \ + int trans_offset = state_trans_base_idx * this->num_ + j_offset; \ + for (int i = 0; i < this->num_; ++i) { \ + /* Initalize the content of alpha variable with related offset.*/ \ + __m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i); \ + /* Obtain the content of weights from un-aligned address.*/ \ + __m256 w_content = _mm256_loadu_ps(w + trans_offset); \ + __m256 score_v = _mm256_add_ps(alpha_content, w_content); \ + __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); \ + /* According to the mask value, update the index of the max_score.*/ \ + /* AVX instructions.*/ \ + __m128i lo_max_j = _mm256_extractf128_si256(max_j, 0); \ + __m128i hi_max_j = _mm256_extractf128_si256(max_j, 1); \ + __m128i lo_mask = _mm256_extractf128_si256((__m256i)mask, 0); \ + __m128i hi_mask = _mm256_extractf128_si256((__m256i)mask, 1); \ + lo_max_j = _mm_andnot_si128(lo_mask, lo_max_j); \ + hi_max_j = _mm_andnot_si128(hi_mask, hi_max_j); \ + lo_mask = _mm_and_si128(lo_mask, _mm_set1_epi32(i)); \ + hi_mask = _mm_and_si128(hi_mask, _mm_set1_epi32(i)); \ + lo_max_j = _mm_or_si128(lo_mask, lo_max_j); \ + hi_max_j = _mm_or_si128(hi_mask, hi_max_j); \ + max_j = _mm256_insertf128_si256(max_j, lo_max_j, 0); \ + max_j = _mm256_insertf128_si256(max_j, hi_max_j, 1); \ + /* AVX done*/ \ + /* Update the max_score value.*/ \ + max_score = _mm256_max_ps(max_score, score_v); \ + trans_offset += this->num_; \ + } \ + UPDATE_ALPHA(AVX_FLOAT_BLOCK) \ + } \ + seq_offset += this->num_; \ + } \ + } + +#define INTRIAVX2_FLOAT(isa, block) \ + template <> \ + CRFDecodeKernelImpl::CRFDecodeKernelImpl(int tag_num) \ + : CRFDecodeKernel() { \ + this->num_ = tag_num; \ + this->end_ = this->num_ / AVX2_FLOAT_BLOCK; \ + this->rest_ = this->num_ % AVX2_FLOAT_BLOCK; \ + } \ + template <> \ + void CRFDecodeKernelImpl::Compute( \ + const int seq_len, const float* x, const float* w, float* alpha, \ + int* track) const { \ + INIT_ALPHA(AVX2_FLOAT_BLOCK) \ + /* Use the column-major strategy to get the location of maximum score.*/ \ + int seq_offset = 0; \ + constexpr int state_trans_base_idx = 2; \ + for (int k = 1; k < seq_len; ++k) { \ + int j_offset = 0; \ + for (int j = 0; j <= this->end_; ++j) { \ + /* Initialize the variables of maximum score and location.*/ \ + __m256 max_score = _mm256_set1_ps(-std::numeric_limits::max()); \ + __m256i max_j = _mm256_set1_epi32(0); \ + /* Calculate the offset of transition_weights.*/ \ + int trans_offset = state_trans_base_idx * this->num_ + j_offset; \ + for (int i = 0; i < this->num_; ++i) { \ + /* Initalize the content of alpha variable with related offset.*/ \ + __m256 alpha_content = _mm256_broadcast_ss(alpha + seq_offset + i); \ + /* Obtain the content of weights from un-aligned address.*/ \ + __m256 w_content = _mm256_loadu_ps(w + trans_offset); \ + __m256 score_v = _mm256_add_ps(alpha_content, w_content); \ + __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); \ + /* According to the mask value, update the index of the max_score.*/ \ + /* AVX2 instructions.*/ \ + max_j = _mm256_or_si256( \ + _mm256_andnot_si256((__m256i)mask, max_j), \ + _mm256_and_si256((__m256i)mask, _mm256_set1_epi32(i))); \ + /* Update the max_score value.*/ \ + max_score = _mm256_max_ps(max_score, score_v); \ + trans_offset += this->num_; \ + } \ + UPDATE_ALPHA(AVX2_FLOAT_BLOCK) \ + } \ + seq_offset += this->num_; \ + } \ + } + +#define INTRIAVX512_FLOAT(block) \ + template <> \ + CRFDecodeKernelImpl::CRFDecodeKernelImpl( \ + int tag_num) \ + : CRFDecodeKernel() { \ + this->num_ = tag_num; \ + this->end_ = this->num_ / AVX512_FLOAT_BLOCK; \ + this->rest_ = this->num_ % AVX512_FLOAT_BLOCK; \ + } \ + template <> \ + void CRFDecodeKernelImpl::Compute( \ + const int seq_len, const float* x, const float* w, float* alpha, \ + int* track) const { \ + INIT_ALPHA(AVX512_FLOAT_BLOCK) \ + /* Use the column-major strategy to get the location of maximum score.*/ \ + int seq_offset = 0; \ + constexpr int state_trans_base_idx = 2; \ + for (int k = 1; k < seq_len; ++k) { \ + int j_offset = 0; \ + for (int j = 0; j <= this->end_; ++j) { \ + /* Initialize the variables of maximum score and location.*/ \ + __m512 max_score = _mm512_set1_ps(-std::numeric_limits::max()); \ + __m512i max_j = _mm512_setzero_si512(); \ + /* Calculate the offset of transition_weights.*/ \ + int trans_offset = state_trans_base_idx * this->num_ + j_offset; \ + for (int i = 0; i < this->num_; ++i) { \ + /* Initalize the content of alpha variable with related offset.*/ \ + __m512 alpha_content = _mm512_set1_ps(*(alpha + seq_offset + i)); \ + /* Obtain the content of weights from un-aligned address.*/ \ + __m512 w_content = _mm512_loadu_ps(w + trans_offset); \ + __m512 score_v = _mm512_add_ps(alpha_content, w_content); \ + __mmask16 mask = _mm512_cmp_ps_mask(score_v, max_score, _CMP_GT_OS); \ + /* AVX512 instructions.*/ \ + max_j = _mm512_mask_set1_epi32(max_j, mask, i); \ + /* Update the max_score value.*/ \ + max_score = _mm512_max_ps(max_score, score_v); \ + trans_offset += this->num_; \ + } \ + /* Update the alpha and track values.*/ \ + __m512 x_content = \ + _mm512_loadu_ps(x + seq_offset + this->num_ + j_offset); \ + max_score = _mm512_add_ps(max_score, x_content); \ + _mm512_storeu_ps(alpha + seq_offset + this->num_ + j_offset, \ + max_score); \ + _mm512_storeu_si512(reinterpret_cast<__m512i*>(track + seq_offset + \ + this->num_ + j_offset), \ + max_j); \ + /* Calculate the offset of next step*/ \ + j_offset += AVX512_FLOAT_BLOCK; \ + if (j == this->end_ - 1) { \ + if (this->rest_ > 0) { \ + j_offset += last_offset; \ + } else { \ + break; \ + } \ + } \ + } \ + seq_offset += this->num_; \ + } \ + } + +#ifdef __AVX__ +INTRIAVX_FLOAT(kEQ8); +INTRIAVX_FLOAT(kGT8LT16); +INTRIAVX_FLOAT(kEQ16); +INTRIAVX_FLOAT(kGT16); +#endif +#ifdef __AVX2__ +INTRIAVX2_FLOAT(jit::avx2, kEQ8); +INTRIAVX2_FLOAT(jit::avx2, kGT8LT16); +INTRIAVX2_FLOAT(jit::avx2, kEQ16); +INTRIAVX2_FLOAT(jit::avx2, kGT16); +#endif +#ifdef __AVX512F__ +INTRIAVX2_FLOAT(jit::avx512f, kEQ8); +INTRIAVX2_FLOAT(jit::avx512f, kGT8LT16); +INTRIAVX512_FLOAT(kEQ16); +INTRIAVX512_FLOAT(kGT16); +#endif + +#undef INTRIAVX512_FLOAT +#undef INTRIAVX2_FLOAT +#undef INTRIAVX_FLOAT +#undef INIT_ALPHA +#undef UPDATE_ALPHA + +REGISTER_JITKERNEL_DEPRECATED(crf_decode, CRFDecodeKernel); + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_exp.cc b/paddle/fluid/operators/math/jit_kernel_exp.cc new file mode 100644 index 0000000000000000000000000000000000000000..c55e54a13f539014c0f582436ca1a105d0b0fedd --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_exp.cc @@ -0,0 +1,544 @@ +/* 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/math/jit_kernel.h" +#include // for exp +#include +#include "paddle/fluid/operators/math/jit_kernel_macro.h" +#ifdef PADDLE_WITH_MKLML +#include "paddle/fluid/platform/dynload/mklml.h" +#endif + +#ifdef __AVX__ +#include +#endif + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace jit = platform::jit; + +/* VExp JitKernel */ +template +class VExpKernelImpl : public VExpKernel { + public: + explicit VExpKernelImpl(int d) : VExpKernel() { this->num_ = d; } + void Compute(const T* x, T* y) const override { + for (int i = 0; i < this->num_; ++i) { + y[i] = std::exp(x[i]); + } + } +}; + +#ifdef PADDLE_WITH_MKLML +#define MKL_FLOAT(isa, block) \ + template <> \ + void VExpKernelImpl::Compute(const float* x, float* y) \ + const { \ + platform::dynload::vsExp(this->num_, x, y); \ + } + +#define MKL_DOUBLE(isa, block) \ + template <> \ + void VExpKernelImpl::Compute(const double* x, double* y) \ + const { \ + platform::dynload::vdExp(this->num_, x, y); \ + } +FOR_EACH_ISA(MKL_FLOAT, kLT8); +FOR_EACH_ISA(MKL_FLOAT, kGT8LT16); +FOR_EACH_ISA(MKL_FLOAT, kGT16); +FOR_EACH_ISA_BLOCK(MKL_DOUBLE); +#endif + +namespace detail { + +#ifdef __AVX__ + +#define ALIGN32 __attribute__((aligned(32))) + +#define _PS256_CONST(Name, Val) \ + static const float _ps256_##Name[8] ALIGN32 = {Val, Val, Val, Val, \ + Val, Val, Val, Val} + +#define _PI256_CONST(Name, Val) \ + static const int _pi256_##Name[8] ALIGN32 = {Val, Val, Val, Val, \ + Val, Val, Val, Val} + +_PI256_CONST(0x7f, 0x7f); +_PS256_CONST(one, 1.f); +_PS256_CONST(0p5, 0.5f); +_PS256_CONST(exp_hi, 88.3762626647949f); +_PS256_CONST(exp_lo, -88.3762626647949f); +_PS256_CONST(cephes_LOG2EF, 1.44269504088896341); +_PS256_CONST(cephes_exp_C1, 0.693359375); +_PS256_CONST(cephes_exp_C2, -2.12194440e-4); +_PS256_CONST(cephes_exp_p0, 1.9875691500E-4); +_PS256_CONST(cephes_exp_p1, 1.3981999507E-3); +_PS256_CONST(cephes_exp_p2, 8.3334519073E-3); +_PS256_CONST(cephes_exp_p3, 4.1665795894E-2); +_PS256_CONST(cephes_exp_p4, 1.6666665459E-1); +_PS256_CONST(cephes_exp_p5, 5.0000001201E-1); + +typedef union imm_xmm_union { + __m256i imm; + __m128i xmm[2]; +} imm_xmm_union; + +#define COPY_IMM_TO_XMM(imm_, xmm0_, xmm1_) \ + { \ + imm_xmm_union u ALIGN32; \ + u.imm = imm_; \ + xmm0_ = u.xmm[0]; \ + xmm1_ = u.xmm[1]; \ + } + +#define COPY_XMM_TO_IMM(xmm0_, xmm1_, imm_) \ + { \ + imm_xmm_union u ALIGN32; \ + u.xmm[0] = xmm0_; \ + u.xmm[1] = xmm1_; \ + imm_ = u.imm; \ + } + +#define AVX2_BITOP_USING_SSE2(fn) \ + static inline __m256i avx2_mm256_##fn(__m256i x, int y) { \ + /* use SSE2 to perform the bitop AVX2 */ \ + __m128i x1, x2; \ + __m256i ret; \ + COPY_IMM_TO_XMM(x, x1, x2); \ + x1 = _mm_##fn(x1, y); \ + x2 = _mm_##fn(x2, y); \ + COPY_XMM_TO_IMM(x1, x2, ret); \ + return ret; \ + } + +#define AVX2_INTOP_USING_SSE2(fn) \ + static inline __m256i avx2_mm256_add_epi32(__m256i x, __m256i y) { \ + /* use SSE2 to perform the AVX2 integer operation */ \ + __m128i x1, x2; \ + __m128i y1, y2; \ + __m256i ret; \ + COPY_IMM_TO_XMM(x, x1, x2); \ + COPY_IMM_TO_XMM(y, y1, y2); \ + x1 = _mm_##fn(x1, y1); \ + x2 = _mm_##fn(x2, y2); \ + COPY_XMM_TO_IMM(x1, x2, ret); \ + return ret; \ + } + +AVX2_BITOP_USING_SSE2(slli_epi32); +AVX2_INTOP_USING_SSE2(add_epi32); + +#define AVXEXP_BASE \ + __m256 tmp = _mm256_setzero_ps(), fx; \ + __m256 one = *reinterpret_cast(_ps256_one); \ + __m256i imm0; \ + x = _mm256_min_ps(x, *reinterpret_cast(_ps256_exp_hi)); \ + x = _mm256_max_ps(x, *reinterpret_cast(_ps256_exp_lo)); \ + /* express exp(x) as exp(g + n*log(2)) */ \ + fx = _mm256_mul_ps(x, \ + *reinterpret_cast(_ps256_cephes_LOG2EF)); \ + fx = _mm256_add_ps(fx, *reinterpret_cast(_ps256_0p5)); \ + tmp = _mm256_floor_ps(fx); \ + /* if greater, substract 1 */ \ + __m256 mask = _mm256_cmp_ps(tmp, fx, _CMP_GT_OS); \ + mask = _mm256_and_ps(mask, one); \ + fx = _mm256_sub_ps(tmp, mask); \ + tmp = _mm256_mul_ps(fx, \ + *reinterpret_cast(_ps256_cephes_exp_C1)); \ + __m256 z = _mm256_mul_ps( \ + fx, *reinterpret_cast(_ps256_cephes_exp_C2)); \ + x = _mm256_sub_ps(x, tmp); \ + x = _mm256_sub_ps(x, z); \ + z = _mm256_mul_ps(x, x); \ + __m256 y = *reinterpret_cast(_ps256_cephes_exp_p0); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p1)); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p2)); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p3)); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p4)); \ + y = _mm256_mul_ps(y, x); \ + y = _mm256_add_ps(y, \ + *reinterpret_cast(_ps256_cephes_exp_p5)); \ + y = _mm256_mul_ps(y, z); \ + y = _mm256_add_ps(y, x); \ + y = _mm256_add_ps(y, one); \ + /* build 2^n */ \ + imm0 = _mm256_cvttps_epi32(fx) + +__m256 ExpAVX(__m256 x) { + AVXEXP_BASE; + // two AVX2 instructions using SSE2 + imm0 = avx2_mm256_add_epi32(imm0, + *reinterpret_cast(_pi256_0x7f)); + imm0 = avx2_mm256_slli_epi32(imm0, 23); + __m256 pow2n = _mm256_castsi256_ps(imm0); + y = _mm256_mul_ps(y, pow2n); + return y; +} +#endif + +#ifdef __AVX2__ +__m256 ExpAVX2(__m256 x) { + AVXEXP_BASE; + // two AVX2 instructions + imm0 = _mm256_add_epi32(imm0, *reinterpret_cast(_pi256_0x7f)); + imm0 = _mm256_slli_epi32(imm0, 23); + __m256 pow2n = _mm256_castsi256_ps(imm0); + y = _mm256_mul_ps(y, pow2n); + return y; +} +#endif + +} // namespace detail + +#define INTRI8_FLOAT(isa, expisa) \ + template <> \ + void VExpKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + _mm256_storeu_ps(y, expisa(tmp)); \ + } + +#define INTRI16_FLOAT(isa, expisa) \ + template <> \ + void VExpKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + tmp0 = expisa(tmp0); \ + tmp1 = expisa(tmp1); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx, detail::ExpAVX); +INTRI16_FLOAT(jit::avx, detail::ExpAVX); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx2, detail::ExpAVX2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); +#endif +// TODO(TJ): eq16 test and complete avx512 + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef MKL_FLOAT +#undef MKL_DOUBLE + +REGISTER_JITKERNEL_DEPRECATED(vexp, VExpKernel); + +/* VSigmoid JitKernel */ +template +class VSigmoidKernelImpl : public VSigmoidKernel { + public: + explicit VSigmoidKernelImpl(int d) : VSigmoidKernel() { + this->num_ = d; + vexp_ = KernelPool::Instance().template Get>(d); + } + void Compute(const T* x, T* y) const override { + const T min = SIGMOID_THRESHOLD_MIN; + const T max = SIGMOID_THRESHOLD_MAX; + for (int i = 0; i < this->num_; ++i) { + y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]); + y[i] = static_cast(0) - y[i]; + } + vexp_->Compute(y, y); + for (int i = 0; i < this->num_; ++i) { + y[i] = static_cast(1) / (static_cast(1) + y[i]); + } + } + + private: + std::shared_ptr> vexp_; +}; + +#define INTRI_SIGMOID(tmp, min, max, expisa) \ + tmp = _mm256_max_ps(tmp, min); \ + tmp = _mm256_min_ps(tmp, max); \ + tmp = _mm256_sub_ps(_mm256_set1_ps(0.0f), tmp); \ + tmp = expisa(tmp); \ + tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp); \ + tmp = _mm256_div_ps(_mm256_set1_ps(1.0f), tmp) + +#define INTRI8_FLOAT(isa, expisa) \ + template <> \ + void VSigmoidKernelImpl::Compute(const float* x, float* y) \ + const { \ + /* TODO(TJ): try to use static const*/ \ + __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \ + __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \ + __m256 tmp = _mm256_loadu_ps(x); \ + INTRI_SIGMOID(tmp, min, max, expisa); \ + _mm256_storeu_ps(y, tmp); \ + } + +#define INTRI16_FLOAT(isa, expisa) \ + template <> \ + void VSigmoidKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \ + __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + INTRI_SIGMOID(tmp0, min, max, expisa); \ + INTRI_SIGMOID(tmp1, min, max, expisa); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#define INTRI_GT8LT16_FLOAT(isa, expisa) \ + template <> \ + VSigmoidKernelImpl::VSigmoidKernelImpl(int d) \ + : VSigmoidKernel() { \ + this->num_ = d; \ + this->end_ = AVX_FLOAT_BLOCK; \ + this->rest_ = d - this->end_; \ + vexp_ = \ + KernelPool::Instance().template Get>(this->rest_); \ + } \ + template <> \ + void VSigmoidKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \ + __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \ + __m256 tmp = _mm256_loadu_ps(x); \ + INTRI_SIGMOID(tmp, min, max, expisa); \ + _mm256_storeu_ps(y, tmp); \ + const float min_ = SIGMOID_THRESHOLD_MIN; \ + const float max_ = SIGMOID_THRESHOLD_MAX; \ + for (int i = this->end_; i < this->num_; ++i) { \ + y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]); \ + y[i] = 0.f - y[i]; \ + } \ + vexp_->Compute(y + this->end_, y + this->end_); \ + for (int i = this->end_; i < this->num_; ++i) { \ + y[i] = 1.f / (1.f + y[i]); \ + } \ + } + +#define INTRI_GT16_FLOAT(isa, expisa) \ + template <> \ + VSigmoidKernelImpl::VSigmoidKernelImpl(int d) \ + : VSigmoidKernel() { \ + this->num_ = d; \ + this->rest_ = d % AVX_FLOAT_BLOCK; \ + this->end_ = d - this->rest_; \ + vexp_ = \ + KernelPool::Instance().template Get>(this->rest_); \ + } \ + template <> \ + void VSigmoidKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX); \ + __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN); \ + for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \ + __m256 tmp = _mm256_loadu_ps(x + i); \ + INTRI_SIGMOID(tmp, min, max, expisa); \ + _mm256_storeu_ps(y + i, tmp); \ + } \ + const float min_ = SIGMOID_THRESHOLD_MIN; \ + const float max_ = SIGMOID_THRESHOLD_MAX; \ + for (int i = this->end_; i < this->num_; ++i) { \ + y[i] = (x[i] < min_) ? min_ : ((x[i] > max_) ? max_ : x[i]); \ + y[i] = 0.f - y[i]; \ + } \ + vexp_->Compute(y + this->end_, y + this->end_); \ + for (int i = this->end_; i < this->num_; ++i) { \ + y[i] = 1.f / (1.f + y[i]); \ + } \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx, detail::ExpAVX); +INTRI16_FLOAT(jit::avx, detail::ExpAVX); +INTRI_GT8LT16_FLOAT(jit::avx, detail::ExpAVX); +INTRI_GT16_FLOAT(jit::avx, detail::ExpAVX); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx2, detail::ExpAVX2); +// maybe use avx at gt8lt16 and gt16 +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); +// maybe use avx2 at gt8lt16 and gt16 +#endif + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef INTRI_GT8LT16_FLOAT +#undef INTRI_GT16_FLOAT +#undef INTRI_VSIGMOID + +REGISTER_JITKERNEL_DEPRECATED(vsigmoid, VSigmoidKernel); + +/* VTanh JitKernel */ +template +class VTanhKernelImpl : public VTanhKernel { + public: + explicit VTanhKernelImpl(int d) : VTanhKernel() { + this->num_ = d; + vscal_ = KernelPool::Instance().template Get>(d); + vsigmoid_ = KernelPool::Instance().template Get>(d); + vaddbias_ = KernelPool::Instance().template Get>(d); + } + void Compute(const T* x, T* y) const override { + const T a = static_cast(2), b = static_cast(-1); + vscal_->Compute(&a, x, y, this->num_); + vsigmoid_->Compute(y, y); + vscal_->Compute(&a, y, y, this->num_); + vaddbias_->Compute(&b, y, y, this->num_); + } + + private: + std::shared_ptr> vscal_; + std::shared_ptr> vsigmoid_; + std::shared_ptr> vaddbias_; +}; + +#define INTRI_VTANH(tmp, expisa) \ + tmp = _mm256_mul_ps(_mm256_set1_ps(-2.0f), tmp); \ + tmp = _mm256_min_ps(tmp, _mm256_set1_ps(EXP_MAX_INPUT)); \ + tmp = expisa(tmp); \ + tmp = _mm256_add_ps(_mm256_set1_ps(1.0f), tmp); \ + tmp = _mm256_div_ps(_mm256_set1_ps(2.0f), tmp); \ + tmp = _mm256_sub_ps(tmp, _mm256_set1_ps(1.0f)) + +#define INTRI8_FLOAT(isa, expisa) \ + template <> \ + void VTanhKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + INTRI_VTANH(tmp, expisa); \ + _mm256_storeu_ps(y, tmp); \ + } + +#define INTRI16_FLOAT(isa, expisa) \ + template <> \ + void VTanhKernelImpl::Compute(const float* x, float* y) \ + const { \ + __m256 tmp0 = _mm256_loadu_ps(x); \ + __m256 tmp1 = _mm256_loadu_ps(x + 8); \ + INTRI_VTANH(tmp0, expisa); \ + INTRI_VTANH(tmp1, expisa); \ + _mm256_storeu_ps(y, tmp0); \ + _mm256_storeu_ps(y + 8, tmp1); \ + } + +#define INTRI_GT8LT16_FLOAT(isa, expisa) \ + template <> \ + VTanhKernelImpl::VTanhKernelImpl(int d) \ + : VTanhKernel() { \ + this->num_ = d; \ + this->end_ = AVX_FLOAT_BLOCK; \ + this->rest_ = d - this->end_; \ + vscal_ = \ + KernelPool::Instance().template Get>(this->rest_); \ + vsigmoid_ = KernelPool::Instance().template Get>( \ + this->rest_); \ + vaddbias_ = KernelPool::Instance().template Get>( \ + this->rest_); \ + } \ + template <> \ + void VTanhKernelImpl::Compute(const float* x, \ + float* y) const { \ + __m256 tmp = _mm256_loadu_ps(x); \ + INTRI_VTANH(tmp, expisa); \ + _mm256_storeu_ps(y, tmp); \ + x += AVX_FLOAT_BLOCK; \ + y += AVX_FLOAT_BLOCK; \ + const float a = 2.f, b = -1.f; \ + vscal_->Compute(&a, x, y, this->num_); \ + vsigmoid_->Compute(y, y); \ + vscal_->Compute(&a, y, y, this->num_); \ + vaddbias_->Compute(&b, y, y, this->num_); \ + } + +#define INTRI_GT16_FLOAT(isa, expisa) \ + template <> \ + VTanhKernelImpl::VTanhKernelImpl(int d) \ + : VTanhKernel() { \ + this->num_ = d; \ + this->rest_ = d % AVX_FLOAT_BLOCK; \ + this->end_ = d - this->rest_; \ + vscal_ = \ + KernelPool::Instance().template Get>(this->rest_); \ + vsigmoid_ = KernelPool::Instance().template Get>( \ + this->rest_); \ + vaddbias_ = KernelPool::Instance().template Get>( \ + this->rest_); \ + } \ + template <> \ + void VTanhKernelImpl::Compute(const float* x, float* y) \ + const { \ + for (int i = 0; i < this->end_; i += AVX_FLOAT_BLOCK) { \ + __m256 tmp = _mm256_loadu_ps(x + i); \ + INTRI_VTANH(tmp, expisa); \ + _mm256_storeu_ps(y + i, tmp); \ + } \ + x += this->end_; \ + y += this->end_; \ + const float a = 2.f, b = -1.f; \ + vscal_->Compute(&a, x, y, this->num_); \ + vsigmoid_->Compute(y, y); \ + vscal_->Compute(&a, y, y, this->num_); \ + vaddbias_->Compute(&b, y, y, this->num_); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx, detail::ExpAVX); +INTRI16_FLOAT(jit::avx, detail::ExpAVX); +INTRI_GT8LT16_FLOAT(jit::avx, detail::ExpAVX); +INTRI_GT16_FLOAT(jit::avx, detail::ExpAVX); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx2, detail::ExpAVX2); +// maybe use avx at gt8lt16 and gt16 +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f, detail::ExpAVX2); +INTRI16_FLOAT(jit::avx512f, detail::ExpAVX2); +// maybe use avx at gt8lt16 and gt16 +#endif + +#undef INTRI8_FLOAT +#undef INTRI16_FLOAT +#undef INTRI_GT8LT16_FLOAT +#undef INTRI_GT16_FLOAT +#undef INTRI_VTANH + +REGISTER_JITKERNEL_DEPRECATED(vtanh, VTanhKernel); + +#undef JITKERNEL_NEW_ACT_IMPL + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_macro.h b/paddle/fluid/operators/math/jit_kernel_macro.h new file mode 100644 index 0000000000000000000000000000000000000000..a8169ea48ae3eee5a8cba291be4496c4c6074221 --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_macro.h @@ -0,0 +1,172 @@ +/* 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 "paddle/fluid/platform/cpu_info.h" + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { + +#define JITKERNEL_DEFINE_NAME(ker_key, ker_class) \ + template <> \ + std::string ker_class##Impl::name(int d) { \ + std::string key(#ker_key "f"); \ + if (useJIT(d)) { \ + /* only jit code need record d*/ \ + return key + "jit" + std::to_string(d); \ + } else if (useMKL(d)) { \ + return key + "mkl"; \ + } else { \ + return key + "any"; \ + } \ + } \ + template <> \ + std::string ker_class##Impl::name(int d) { \ + std::string key(#ker_key "d"); \ + /* jit code do not support double yet*/ \ + if (useMKL(d)) { \ + return key + "mkl"; \ + } else { \ + return key + "any"; \ + } \ + } + +#define JITKERNEL_DECLARE(ker_class, ker_dtype) \ + template <> \ + std::shared_ptr> \ + KernelPool::Get, int>(int d) + +#define JITKERNEL_FIND_KEY(ker_class, ker_dtype) \ + std::string key = ker_class##Impl::name(d) + +#define JITKERNEL_IMPL(ker_class, ker_dtype) \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>(d)) + +#define REGISTER_JITKERNEL_WITH_DTYPE(ker_class, ker_dtype, marco_declare, \ + macro_find_key, macro_impl) \ + marco_declare(ker_class, ker_dtype) { \ + macro_find_key(ker_class, ker_dtype); \ + if (kers_.find(key) == kers_.end()) { \ + std::shared_ptr> p; \ + macro_impl(ker_class, ker_dtype); \ + kers_.insert({key, std::dynamic_pointer_cast(p)}); \ + return p; \ + } \ + return std::dynamic_pointer_cast>( \ + kers_.at(key)); \ + } + +#define REGISTER_JITKERNEL_ARGS(ker_key, ker_class, marco_define_name, \ + marco_declare, macro_find_key, macro_impl) \ + marco_define_name(ker_key, ker_class); \ + REGISTER_JITKERNEL_WITH_DTYPE(ker_class, float, JITKERNEL_DECLARE, \ + JITKERNEL_FIND_KEY, JITKERNEL_IMPL); \ + REGISTER_JITKERNEL_WITH_DTYPE(ker_class, double, JITKERNEL_DECLARE, \ + JITKERNEL_FIND_KEY, JITKERNEL_IMPL) + +#define REGISTER_JITKERNEL(ker_key, ker_class) \ + REGISTER_JITKERNEL_ARGS(ker_key, ker_class, JITKERNEL_DEFINE_NAME, \ + JITKERNEL_DECLARE, JITKERNEL_FIND_KEY, \ + JITKERNEL_IMPL) + +namespace jit = platform::jit; +// TODO(TJ): below defines are deprecated, would be remove recently +#define SEARCH_BLOCK(macro_, ker, dtype, isa) \ + if (d < AVX_FLOAT_BLOCK) { \ + macro_(ker, dtype, isa, kLT8); \ + } else if (d == AVX_FLOAT_BLOCK) { \ + macro_(ker, dtype, isa, kEQ8); \ + } else if (d > AVX_FLOAT_BLOCK && d < AVX512_FLOAT_BLOCK) { \ + macro_(ker, dtype, isa, kGT8LT16); \ + } else if (d == AVX512_FLOAT_BLOCK) { \ + macro_(ker, dtype, isa, kEQ16); \ + } else { \ + macro_(ker, dtype, isa, kGT16); \ + } + +#define SEARCH_ISA_BLOCK(macro_, ker, dtype) \ + if (jit::MayIUse(jit::avx512f)) { \ + SEARCH_BLOCK(macro_, ker, dtype, jit::avx512f); \ + } else if (jit::MayIUse(jit::avx2)) { \ + SEARCH_BLOCK(macro_, ker, dtype, jit::avx2); \ + } else if (jit::MayIUse(jit::avx)) { \ + SEARCH_BLOCK(macro_, ker, dtype, jit::avx); \ + } else { \ + SEARCH_BLOCK(macro_, ker, dtype, jit::isa_any); \ + } + +#define JITKERNEL_KEY(ker_key, dtype_key) \ + #ker_key #dtype_key + std::to_string(d) + +#define JITKERNEL_NEW_IMPL_DEPRECATED(ker, dtype, isa, k) \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>(d)) + +#define JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, ker_dtype, \ + dtype_key, marco_declare, macro_key, \ + macro_impl) \ + marco_declare(ker_class, ker_dtype) { \ + std::string key = macro_key(ker_key, dtype_key); \ + if (kers_.find(key) == kers_.end()) { \ + std::shared_ptr> p; \ + SEARCH_ISA_BLOCK(macro_impl, ker_class, ker_dtype); \ + kers_.insert({key, std::dynamic_pointer_cast(p)}); \ + return p; \ + } \ + return std::dynamic_pointer_cast>( \ + kers_.at(key)); \ + } + +#define REGISTER_JITKERNEL_DEPRECATED(ker_key, ker_class) \ + JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, float, f, \ + JITKERNEL_DECLARE, JITKERNEL_KEY, \ + JITKERNEL_NEW_IMPL_DEPRECATED); \ + JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, double, d, \ + JITKERNEL_DECLARE, JITKERNEL_KEY, \ + JITKERNEL_NEW_IMPL_DEPRECATED) + +#define REGISTER_JITKERNEL_ARGS_DEPRECATED(ker_key, ker_class, marco_declare, \ + macro_key, macro_impl) \ + JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, float, f, marco_declare, \ + macro_key, macro_impl); \ + JITKERNEL_WITH_DTYPE_DEPRECATED(ker_key, ker_class, double, d, \ + marco_declare, macro_key, macro_impl) + +#define FOR_EACH_ISA(macro_, block) \ + macro_(jit::avx512f, block); \ + macro_(jit::avx2, block); \ + macro_(jit::avx, block); \ + macro_(jit::isa_any, block) + +#define FOR_EACH_BLOCK(macro_, isa) \ + macro_(isa, kLT8); \ + macro_(isa, kEQ8); \ + macro_(isa, kGT8LT16); \ + macro_(isa, kEQ16); \ + macro_(isa, kGT16) + +#define FOR_EACH_ISA_BLOCK(macro_) \ + FOR_EACH_BLOCK(macro_, jit::avx512f); \ + FOR_EACH_BLOCK(macro_, jit::avx2); \ + FOR_EACH_BLOCK(macro_, jit::avx); \ + FOR_EACH_BLOCK(macro_, jit::isa_any) + +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_rnn.cc b/paddle/fluid/operators/math/jit_kernel_rnn.cc new file mode 100644 index 0000000000000000000000000000000000000000..ba3e917377cf12192a068a9d71238442e12d5e5e --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_rnn.cc @@ -0,0 +1,485 @@ +/* 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/math/jit_kernel.h" +#include +#include "paddle/fluid/operators/math/jit_kernel_macro.h" +#include "paddle/fluid/platform/enforce.h" +#include "paddle/fluid/platform/macros.h" + +#ifdef __AVX__ +#include +#endif + +namespace paddle { +namespace operators { +namespace math { +namespace jitkernel { +namespace detail { +#ifdef __AVX__ +__m256 ExpAVX(__m256 x); +#endif + +#ifdef __AVX2__ +__m256 ExpAVX2(__m256 x); +#endif + +} // namespace detail + +namespace jit = platform::jit; + +#ifdef __AVX__ +typedef enum { kSigmoid, kRelu, kTanh, kIdentity } act_type; + +class AVXAct { + public: + virtual ~AVXAct() = default; + virtual __m256 Compute(__m256 x) const = 0; +}; + +template +class AVXActImpl : public AVXAct { + public: + __m256 Compute(__m256 x) const override { PADDLE_THROW("Unkown type!"); } +}; + +#define AVX_SIGMOID(isa, expisa) \ + template <> \ + __m256 AVXActImpl::Compute(__m256 x) const { \ + __m256 ones = _mm256_set1_ps(1.0f); \ + x = _mm256_max_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MIN)); \ + x = _mm256_min_ps(x, _mm256_set1_ps(SIGMOID_THRESHOLD_MAX)); \ + x = _mm256_sub_ps(_mm256_set1_ps(0.0f), x); \ + x = expisa(x); \ + x = _mm256_add_ps(ones, x); \ + return _mm256_div_ps(ones, x); \ + } + +#define AVX_TANH(isa, expisa) \ + template <> \ + __m256 AVXActImpl::Compute(__m256 x) const { \ + __m256 ones = _mm256_set1_ps(1.0f); \ + x = _mm256_mul_ps(_mm256_set1_ps(-2.0f), x); \ + x = _mm256_min_ps(x, _mm256_set1_ps(EXP_MAX_INPUT)); \ + x = expisa(x); \ + x = _mm256_add_ps(ones, x); \ + x = _mm256_div_ps(_mm256_set1_ps(2.0f), x); \ + return _mm256_sub_ps(x, ones); \ + } + +#define AVX_RELU(isa) \ + template <> \ + __m256 AVXActImpl::Compute(__m256 x) const { \ + return _mm256_max_ps(x, _mm256_setzero_ps()); \ + } + +#define AVX_IDENTITY(isa) \ + template <> \ + __m256 AVXActImpl::Compute(__m256 x) const { \ + return x; \ + } + +#define FOR_EACH_AVX_ISA(macro_) \ + macro_(jit::avx); \ + macro_(jit::avx2); \ + macro_(jit::avx512f) + +FOR_EACH_AVX_ISA(AVX_RELU); +FOR_EACH_AVX_ISA(AVX_IDENTITY); + +AVX_SIGMOID(jit::avx, detail::ExpAVX); +AVX_TANH(jit::avx, detail::ExpAVX); + +#ifdef __AVX2__ +AVX_SIGMOID(jit::avx2, detail::ExpAVX2); +AVX_SIGMOID(jit::avx512f, detail::ExpAVX2); +AVX_TANH(jit::avx2, detail::ExpAVX2); +AVX_TANH(jit::avx512f, detail::ExpAVX2); +#endif + +#undef FOR_EACH_AVX_ISA +#undef AVX_IDENTITY +#undef AVX_RELU +#undef AVX_TANH +#undef AVX_SIGMOID + +#endif + +template +static std::shared_ptr> GetActKernel( + const std::string& type, int n) { + if (type == "sigmoid") { + return std::dynamic_pointer_cast>( + KernelPool::Instance().template Get>(n)); + } else if (type == "relu") { + return std::dynamic_pointer_cast>( + KernelPool::Instance().template Get>(n)); + } else if (type == "tanh") { + return std::dynamic_pointer_cast>( + KernelPool::Instance().template Get>(n)); + } else if (type == "identity" || type == "") { + return std::dynamic_pointer_cast>( + KernelPool::Instance().template Get>(n)); + } + PADDLE_THROW("Not support type: %s", type); + return nullptr; +} + +#ifdef __AVX__ +template +static std::unique_ptr GetAVXAct(const std::string& type) { + if (type == "sigmoid") { + return std::unique_ptr(new AVXActImpl()); + } else if (type == "relu") { + return std::unique_ptr(new AVXActImpl()); + } else if (type == "tanh") { + return std::unique_ptr(new AVXActImpl()); + } else if (type == "identity" || type == "") { + return std::unique_ptr(new AVXActImpl()); + } + PADDLE_THROW("Not support type: %s", type); + return nullptr; +} +#endif + +/* LSTM JitKernel */ +template +class LSTMKernelImpl : public LSTMKernel { + public: + explicit LSTMKernelImpl(const std::string& act_gate, + const std::string& act_cand, + const std::string& act_cell, int d) + : LSTMKernel() { + d_ = d; + d2_ = d * 2; + d3_ = d * 3; + act_gate_d3_ = GetActKernel(act_gate, d3_); + act_gate_d_ = GetActKernel(act_gate, d); + act_cand_d_ = GetActKernel(act_cand, d); + act_cell_d_ = GetActKernel(act_cell, d); + vmul_d_ = KernelPool::Instance().template Get>(d); + vadd_d_ = KernelPool::Instance().template Get>(d); + } + + void ComputeCtHt(T* gates, const T* ct_1, T* ct, T* ht, const T* wp_data, + T* checked) const override { + // gates: W_ch, W_ih, W_fh, W_oh + act_gate_d3_->Compute(gates + d_, gates + d_); + + /* C_t = C_t-1 * fgated + cand_gated * igated */ + act_cand_d_->Compute(gates, gates); + vmul_d_->Compute(gates, gates + d_, gates + d_, d_); + vmul_d_->Compute(ct_1, gates + d2_, gates + d2_, d_); + vadd_d_->Compute(gates + d_, gates + d2_, ct, d_); + + /* H_t = act_cell(C_t) * ogated */ + act_cell_d_->Compute(ct, gates + d2_); + vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_); + } + void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override { + /* C_t = igated * cgated*/ + act_gate_d_->Compute(gates + d_, gates + d_); + act_cand_d_->Compute(gates, gates); + vmul_d_->Compute(gates, gates + d_, ct, d_); + /* H_t = act_cell(C_t) * ogated */ + act_gate_d_->Compute(gates + d3_, gates + d3_); + act_cell_d_->Compute(ct, gates + d2_); + vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_); + } + + private: + int d_, d2_, d3_; + std::shared_ptr> act_gate_d3_, act_gate_d_, act_cand_d_, + act_cell_d_; + std::shared_ptr> vmul_d_; + std::shared_ptr> vadd_d_; +#ifdef __AVX__ + std::unique_ptr avx_act_gate_, avx_act_cand_, avx_act_cell_; +#endif +}; + +#define INTRI8_FLOAT(isa) \ + template <> \ + LSTMKernelImpl::LSTMKernelImpl( \ + const std::string& act_gate, const std::string& act_cand, \ + const std::string& act_cell, int d) \ + : LSTMKernel() { \ + avx_act_gate_ = GetAVXAct(act_gate); \ + avx_act_cand_ = GetAVXAct(act_cand); \ + avx_act_cell_ = GetAVXAct(act_cell); \ + } \ + template <> \ + void LSTMKernelImpl::ComputeCtHt( \ + float* gates, const float* ct_1, float* ct, float* ht, \ + const float* wp_data, float* checked) const { \ + /* gates: W_ch, W_ih, W_fh, W_oh */ \ + __m256 c, i, f, o; \ + c = _mm256_loadu_ps(gates); \ + i = _mm256_loadu_ps(gates + 8); \ + f = _mm256_loadu_ps(gates + 16); \ + o = _mm256_loadu_ps(gates + 24); \ + /* C_t = C_t-1 * fgated + cand_gated * igated*/ \ + c = _mm256_mul_ps(avx_act_cand_->Compute(c), avx_act_gate_->Compute(i)); \ + i = _mm256_loadu_ps(ct_1); \ + f = _mm256_mul_ps(i, avx_act_gate_->Compute(f)); \ + f = _mm256_add_ps(c, f); \ + _mm256_storeu_ps(ct, f); \ + /* H_t = act_cell(C_t) * ogated */ \ + o = _mm256_mul_ps(avx_act_cell_->Compute(f), avx_act_gate_->Compute(o)); \ + _mm256_storeu_ps(ht, o); \ + } \ + template <> \ + void LSTMKernelImpl::ComputeC1H1( \ + float* gates, float* ct, float* ht, const float* wp_data) const { \ + __m256 c, i, o; \ + c = _mm256_loadu_ps(gates); \ + i = _mm256_loadu_ps(gates + 8); \ + o = _mm256_loadu_ps(gates + 24); \ + /* C_t = igated * cgated*/ \ + c = _mm256_mul_ps(avx_act_gate_->Compute(i), avx_act_cand_->Compute(c)); \ + _mm256_storeu_ps(ct, c); \ + /* H_t = act_cell(C_t) * ogated */ \ + o = _mm256_mul_ps(avx_act_cell_->Compute(c), avx_act_gate_->Compute(o)); \ + _mm256_storeu_ps(ht, o); \ + } + +// TODO(TJ): optimize keq16 + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f); +#endif + +/* Peephole JitKernel */ +template +class PeepholeKernelImpl : public LSTMKernel { + public: + explicit PeepholeKernelImpl(const std::string& act_gate, + const std::string& act_cand, + const std::string& act_cell, int d) + : LSTMKernel() { + d_ = d; + d2_ = d * 2; + d3_ = d * 3; + act_gate_d_ = GetActKernel(act_gate, d); + act_cand_d_ = GetActKernel(act_cand, d); + act_cell_d_ = GetActKernel(act_cell, d); + vmul_d_ = KernelPool::Instance().template Get>(d); + vadd_d_ = KernelPool::Instance().template Get>(d); + vadd_d2_ = KernelPool::Instance().template Get>(d2_); + act_gate_d2_ = GetActKernel(act_gate, d2_); + } + + void ComputeCtHt(T* gates, const T* ct_1, T* ct, T* ht, const T* wp_data, + T* checked) const override { + /* get fgated and igated*/ + vmul_d_->Compute(wp_data, ct_1, checked, d_); + vmul_d_->Compute(wp_data + d_, ct_1, checked + d_, d_); + vadd_d2_->Compute(checked, gates + d_, gates + d_, d2_); + act_gate_d2_->Compute(gates + d_, gates + d_); + /* C_t = C_t-1 * fgated + cand_gated * igated*/ + act_cand_d_->Compute(gates, gates); + vmul_d_->Compute(gates, gates + d_, gates + d_, d_); + vmul_d_->Compute(ct_1, gates + d2_, gates + d2_, d_); + vadd_d_->Compute(gates + d_, gates + d2_, ct, d_); + /* get ogated*/ + vmul_d_->Compute(wp_data + d2_, ct, gates + d_, d_); + vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_, d_); + act_gate_d_->Compute(gates + d3_, gates + d3_); + /* H_t = act_cell(C_t) * ogated */ + act_cell_d_->Compute(ct, gates + d2_); + vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_); + } + + void ComputeC1H1(T* gates, T* ct, T* ht, const T* wp_data) const override { + /* C_t = igated * cgated*/ + act_gate_d_->Compute(gates + d_, gates + d_); + act_cand_d_->Compute(gates, gates); + vmul_d_->Compute(gates, gates + d_, ct, d_); + /* get outgated, put W_oc * C_t on igated */ + vmul_d_->Compute(wp_data + d2_, ct, gates + d_, d_); + vadd_d_->Compute(gates + d_, gates + d3_, gates + d3_, d_); + /* H_t = act_cell(C_t) * ogated */ + act_gate_d_->Compute(gates + d3_, gates + d3_); + act_cell_d_->Compute(ct, gates + d2_); + vmul_d_->Compute(gates + d2_, gates + d3_, ht, d_); + } + + private: + int d_, d2_, d3_; + std::shared_ptr> act_gate_d2_, act_gate_d_, act_cand_d_, + act_cell_d_; + std::shared_ptr> vmul_d_; + std::shared_ptr> vadd_d_, vadd_d2_; +}; + +#define JITKERNEL_DECLARE_LSTM(ker_class, ker_dtype) \ + template <> \ + std::shared_ptr> \ + KernelPool::Get, const std::string&, \ + const std::string&, const std::string&, int, bool>( \ + const std::string& act_gate, const std::string& act_cand, \ + const std::string& act_cell, int d, bool use_peephole) + +#define JITKERNEL_KEY_LSTM(ker_key, dtype_key) \ + #ker_key #dtype_key + std::to_string(d) + act_gate + act_cand + act_cell + \ + (use_peephole ? "p" : "n") + +#define JITKERNEL_NEW_LSTM_IMPL(ker, dtype, isa, k) \ + if (use_peephole) { \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>( \ + act_gate, act_cand, act_cell, d)); \ + } else { \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>(act_gate, act_cand, \ + act_cell, d)); \ + } + +REGISTER_JITKERNEL_ARGS_DEPRECATED(lstm, LSTMKernel, JITKERNEL_DECLARE_LSTM, + JITKERNEL_KEY_LSTM, JITKERNEL_NEW_LSTM_IMPL); + +#undef INTRI8_FLOAT +#undef JITKERNEL_DECLARE_LSTM +#undef JITKERNEL_KEY_LSTM +#undef JITKERNEL_NEW_LSTM_IMPL + +/* GRU JitKernel */ +template +class GRUKernelImpl : public GRUKernel { + public: + explicit GRUKernelImpl(const std::string& act_gate, + const std::string& act_state, int d) + : GRUKernel() { + d_ = d; + d2_ = d * 2; + act_gate_d2_ = GetActKernel(act_gate, d2_); + act_gate_d_ = GetActKernel(act_gate, d); + act_state_d_ = GetActKernel(act_state, d); + vmul_d_ = KernelPool::Instance().template Get>(d); + } + + void ComputeH1(T* gates, T* ht) const override { + act_gate_d_->Compute(gates, gates); + act_state_d_->Compute(gates + d2_, gates + d2_); + vmul_d_->Compute(gates, gates + d2_, ht, d_); + } + + void ComputeHtPart1(T* gates, const T* ht_1, T* ht) const override { + // W: {W_update, W_reset; W_state} + act_gate_d2_->Compute(gates, gates); + vmul_d_->Compute(ht_1, gates + d_, ht, d_); + } + + void ComputeHtPart2(T* gates, const T* ht_1, T* ht) const override { + T* y = gates + d2_; + act_state_d_->Compute(y, y); + // out = zt*ht~ + (1-zt)*ht_1 + for (int i = 0; i < d_; ++i) { + ht[i] = gates[i] * y[i] + (static_cast(1) - gates[i]) * ht_1[i]; + } + } + + private: + int d_, d2_; + std::shared_ptr> act_gate_d2_, act_gate_d_, act_state_d_; + std::shared_ptr> vmul_d_; +#ifdef __AVX__ + std::unique_ptr avx_act_gate_, avx_act_state_; +#endif +}; + +#define INTRI8_FLOAT(isa) \ + template <> \ + GRUKernelImpl::GRUKernelImpl( \ + const std::string& act_gate, const std::string& act_state, int d) \ + : GRUKernel() { \ + avx_act_gate_ = GetAVXAct(act_gate); \ + avx_act_state_ = GetAVXAct(act_state); \ + } \ + template <> \ + void GRUKernelImpl::ComputeH1(float* gates, float* ht) \ + const { \ + __m256 u, s; \ + /* W: {W_update, W_reset; W_state} */ \ + u = _mm256_loadu_ps(gates); \ + s = _mm256_loadu_ps(gates + 16); \ + s = _mm256_mul_ps(avx_act_gate_->Compute(u), avx_act_state_->Compute(s)); \ + _mm256_storeu_ps(ht, s); \ + } \ + template <> \ + void GRUKernelImpl::ComputeHtPart1( \ + float* gates, const float* ht_1, float* ht) const { \ + /* not exactly equal the any implementation */ \ + __m256 r, ht0; \ + r = _mm256_loadu_ps(gates + 8); \ + ht0 = _mm256_loadu_ps(ht_1); \ + r = _mm256_mul_ps(avx_act_gate_->Compute(r), ht0); \ + _mm256_storeu_ps(ht, r); \ + } \ + template <> \ + void GRUKernelImpl::ComputeHtPart2( \ + float* gates, const float* ht_1, float* ht) const { \ + /* not exactly equal the any implementation */ \ + __m256 u, s, ht0; \ + u = _mm256_loadu_ps(gates); \ + s = _mm256_loadu_ps(gates + 16); \ + ht0 = _mm256_loadu_ps(ht_1); \ + u = avx_act_gate_->Compute(u); \ + s = _mm256_mul_ps(u, avx_act_state_->Compute(s)); \ + u = _mm256_sub_ps(_mm256_set1_ps(1.f), u); \ + u = _mm256_mul_ps(u, ht0); \ + u = _mm256_add_ps(s, u); \ + _mm256_storeu_ps(ht, u); \ + } + +#ifdef __AVX__ +INTRI8_FLOAT(jit::avx); +#endif +#ifdef __AVX2__ +INTRI8_FLOAT(jit::avx2); +#endif +#ifdef __AVX512F__ +INTRI8_FLOAT(jit::avx512f); +#endif + +#define JITKERNEL_DECLARE_GRU(ker_class, ker_dtype) \ + template <> \ + std::shared_ptr> KernelPool::Get< \ + GRUKernel, const std::string&, const std::string&, int>( \ + const std::string& act_gate, const std::string& act_state, int d) + +#define JITKERNEL_KEY_GRU(ker_key, dtype_key) \ + #ker_key #dtype_key + std::to_string(d) + act_gate + act_state + +#define JITKERNEL_NEW_GRU_IMPL(ker, dtype, isa, k) \ + p = std::dynamic_pointer_cast>( \ + std::make_shared>(act_gate, act_state, d)); + +REGISTER_JITKERNEL_ARGS_DEPRECATED(gru, GRUKernel, JITKERNEL_DECLARE_GRU, + JITKERNEL_KEY_GRU, JITKERNEL_NEW_GRU_IMPL); + +#undef INTRI8_FLOAT +#undef JITKERNEL_NEW_GRU_IMPL +#undef JITKERNEL_KEY_GRU +#undef JITKERNEL_DECLARE_GRU +} // namespace jitkernel +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/jit_kernel_test.cc b/paddle/fluid/operators/math/jit_kernel_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..7dc3e600b564d95b46070ff4436b2d0de2f3e105 --- /dev/null +++ b/paddle/fluid/operators/math/jit_kernel_test.cc @@ -0,0 +1,823 @@ +/* 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/math/jit_kernel.h" +#include +#include // for exp +#include // for memcpy +#include +#include +#include +#include "gflags/gflags.h" +#include "glog/logging.h" +#include "gtest/gtest.h" + +#ifdef PADDLE_WITH_MKLML +#include "paddle/fluid/platform/dynload/mklml.h" +#endif + +#ifdef __AVX__ +#include +#endif + +constexpr int repeat = 20000; + +inline double GetCurrentUS() { + struct timeval time; + gettimeofday(&time, NULL); + return 1e+6 * time.tv_sec + time.tv_usec; +} + +template +void RandomVec(const int n, T* a, const T lower = static_cast(-20.f), + const T upper = static_cast(20.f)) { + static unsigned int seed = 100; + std::mt19937 rng(seed++); + std::uniform_real_distribution uniform_dist(0, 1); + for (int i = 0; i < n; ++i) { + a[i] = static_cast(uniform_dist(rng) * (upper - lower) + lower); + } +} + +void vrelu_ref(const int n, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = x[i] > 0.f ? x[i] : 0.f; + } +} + +#if defined __AVX__ || defined __AVX2__ +void vrelu_intri8(const int n, const float* x, float* y) { + __m256 tmp = _mm256_loadu_ps(x); + tmp = _mm256_max_ps(tmp, _mm256_setzero_ps()); + _mm256_storeu_ps(y, tmp); +} +#endif + +TEST(JitKernel, vrelu) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -10.f, 1.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vrelu_ref(d, x_data, zref_data); + } + auto trefe = GetCurrentUS(); +#if defined __AVX__ || defined __AVX2__ + if (d == 8) { + auto si0 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vrelu_intri8(d, x_data, zref_data); + } + auto si1 = GetCurrentUS(); + VLOG(30) << "Vec size 8 intr takes: " << (si1 - si0) / repeat; + } +#endif + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vaddbias_ref(const int n, const float a, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = x[i] + a; + } +} + +TEST(JitKernel, vaddbias) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 64, 100, 128, 256}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -2.f, 2.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float a = 2.f; + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vaddbias_ref(d, a, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(&a, x_data, ztgt_data, d); + } + auto ttgte = GetCurrentUS(); + + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vexp_ref(const int n, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = std::exp(x[i]); + } +} + +#ifdef PADDLE_WITH_MKLML +void vexp_mkl(const int n, const float* x, float* y) { + paddle::platform::dynload::vsExp(n, x, y); +} +#endif + +TEST(JitKernel, vexp) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 128, 256}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -2.f, 2.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vexp_ref(d, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + +#ifdef PADDLE_WITH_MKLML + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vexp_mkl(d, x_data, zref_data); + } + auto tmkle = GetCurrentUS(); +#endif + + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat +#ifdef PADDLE_WITH_MKLML + << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, " +#else + << " us, " +#endif + << "tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +inline float _sigmoid(float x) { + const float min = SIGMOID_THRESHOLD_MIN; + const float max = SIGMOID_THRESHOLD_MAX; + float tmp = (x < min) ? min : ((x > max) ? max : x); + return 1.f / (1.f + std::exp(-tmp)); +} + +void vsigmoid_ref(const int n, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = _sigmoid(x[i]); + } +} + +void vsigmoid_better( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VExpKernel>& vexp, + const int n, const float* x, float* y) { + const float min = SIGMOID_THRESHOLD_MIN; + const float max = SIGMOID_THRESHOLD_MAX; + for (int i = 0; i < n; ++i) { + y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]); + y[i] = 0.f - y[i]; + } + vexp->Compute(y, y); + for (int i = 0; i < n; ++i) { + y[i] = 1.f / (1.f + y[i]); + } +} + +TEST(JitKernel, vsigmoid) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -2.f, 2.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const auto& vexp = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vsigmoid_better(vexp, d, x_data, zref_data); + } + auto tmkle = GetCurrentUS(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vsigmoid_ref(d, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat + << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +inline float _tanh(float x) { return 2.f * _sigmoid(2.f * x) - 1.f; } + +void vtanh_ref(const int n, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = _tanh(x[i]); + } +} + +void vtanh_better( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VScalKernel>& vscal, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VSigmoidKernel>& + vsigmoid, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VAddBiasKernel>& + vaddbias, + const int n, const float* x, float* y) { + const float a = 2.f, b = -1.f; + vscal->Compute(&a, x, y, n); + vsigmoid->Compute(y, y); + vscal->Compute(&a, y, y, n); + vaddbias->Compute(&b, y, y, n); +} + +TEST(JitKernel, vtanh) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) { + std::vector x(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data(), -2.f, 2.f); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const auto& vscal = + jit::KernelPool::Instance().template Get>(d); + const auto& vsigmoid = + jit::KernelPool::Instance().template Get>(d); + const auto& vaddbias = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vtanh_better(vscal, vsigmoid, vaddbias, d, x_data, zref_data); + } + auto tmkle = GetCurrentUS(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vtanh_ref(d, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, ztgt_data); + } + auto ttgte = GetCurrentUS(); + + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat + << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void lstm_ctht_ref( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VSigmoidKernel>& + vsigmoid_3d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VTanhKernel>& vtanh_d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VExpKernel>& vexp_1, + const int d, float* gates, const float* ct_1, float* ct, float* ht) { + vsigmoid_3d->Compute(gates + d, gates + d); + vtanh_d->Compute(gates, gates); + const float *i = gates + d, *f = gates + d * 2, *o = gates + d * 3; + const float min = SIGMOID_THRESHOLD_MIN; + const float max = SIGMOID_THRESHOLD_MAX; + for (int k = 0; k < d; ++k) { + // C_t = C_t-1 * fgated + cand_gated * igated + ct[k] = ct_1[k] * f[k] + gates[k] * i[k]; + // H_t = act_cell(C_t) * ogated + float tmp = ct[k] * 2; + tmp = 0.f - ((tmp < min) ? min : ((tmp > max) ? max : tmp)); + vexp_1->Compute(&tmp, &tmp); + tmp = 2.f / (1.f + tmp) - 1.f; + ht[k] = tmp * o[k]; + } +} + +void lstm_ctht_better( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VSigmoidKernel>& + vsigmoid_3d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VTanhKernel>& vtanh_d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VMulKernel>& vmul_d, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VAddKernel>& vadd_d, + const int d, float* gates, const float* ct_1, float* ct, float* ht) { + int d2 = d * 2; + vsigmoid_3d->Compute(gates + d, gates + d); + vtanh_d->Compute(gates, gates); + vmul_d->Compute(gates, gates + d, gates + d, d); + vmul_d->Compute(ct_1, gates + d2, gates + d2, d); + vadd_d->Compute(gates + d, gates + d2, ct, d); + /* H_t = act_cell(C_t) * ogated */ + vtanh_d->Compute(ct, gates + d2); + vmul_d->Compute(gates + d2, gates + d * 3, ht, d); +} + +TEST(JitKernel, lstm) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 32, 64, 100}) { + int d4 = d * 4; + int d3 = d * 3; + std::vector x(d4), xref(d4); + std::vector ct_1(d), ct_tgt(d), ht_tgt(d); + std::vector ct_ref(d), ht_ref(d); + RandomVec(d4, x.data(), -2.f, 2.f); + RandomVec(d, ct_1.data(), -2.f, 2.f); + memcpy(xref.data(), x.data(), sizeof(float) * d4); + std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh"; + const auto& ker = + jit::KernelPool::Instance() + .template Get, const std::string&, + const std::string&, const std::string&>( + act_gate, act_cand, act_cell, d, false); + // below kernels are used to compute refer + const auto& vsigmoid_3d = + jit::KernelPool::Instance().template Get>( + d3); + const auto& vtanh_d = + jit::KernelPool::Instance().template Get>(d); + const auto& vexp_1 = + jit::KernelPool::Instance().template Get>(1); + const auto& vmul_d = + jit::KernelPool::Instance().template Get>(d); + const auto& vadd_d = + jit::KernelPool::Instance().template Get>(d); + + float* x_data = x.data(); + float* xref_data = xref.data(); + const float* ct_1_data = ct_1.data(); + float* ct_tgt_data = ct_tgt.data(); + float* ht_tgt_data = ht_tgt.data(); + float* ct_ref_data = ct_ref.data(); + float* ht_ref_data = ht_ref.data(); + // compute once to check correctness + lstm_ctht_ref(vsigmoid_3d, vtanh_d, vexp_1, d, xref_data, ct_1_data, + ct_ref_data, ht_ref_data); + ker->ComputeCtHt(x_data, ct_1_data, ct_tgt_data, ht_tgt_data); + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ct_tgt_data[i], ct_ref_data[i], 1e-3); + EXPECT_NEAR(ht_tgt_data[i], ht_ref_data[i], 1e-3); + } + + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + lstm_ctht_better(vsigmoid_3d, vtanh_d, vmul_d, vadd_d, d, xref_data, + ct_1_data, ct_ref_data, ht_ref_data); + } + auto tmkle = GetCurrentUS(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + lstm_ctht_ref(vsigmoid_3d, vtanh_d, vexp_1, d, xref_data, ct_1_data, + ct_ref_data, ht_ref_data); + } + auto trefe = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->ComputeCtHt(x_data, ct_1_data, ct_tgt_data, ht_tgt_data); + } + auto ttgte = GetCurrentUS(); + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat + << " us, better(jit) takes: " << (tmkle - tmkls) / repeat + << " us, tgt takes: " << (ttgte - ttgts) / repeat; + } +} + +void vscal_ref(const int n, const float a, const float* x, float* y) { + for (int i = 0; i < n; ++i) { + y[i] = a * x[i]; + } +} +void vscal_inp_ref(const int n, const float a, float* x) { + for (int i = 0; i < n; ++i) { + x[i] = a * x[i]; + } +} +#if defined __AVX__ || defined __AVX2__ +void vscal_intri8(const int n, const float a, const float* x, float* y) { + __m256 tmp; + __m256 scalar = _mm256_set1_ps(a); + tmp = _mm256_loadu_ps(x); + tmp = _mm256_mul_ps(tmp, scalar); + _mm256_storeu_ps(y, tmp); +} +void vscal_inp_intri8(const int n, const float a, float* x) { + __m256 tmp; + __m256 scalar = _mm256_set1_ps(a); + tmp = _mm256_loadu_ps(x); + tmp = _mm256_mul_ps(tmp, scalar); + _mm256_storeu_ps(x, tmp); +} +#endif + +#ifdef PADDLE_WITH_MKLML +void vscal_inp_mkl(const int n, const float a, float* x) { + paddle::platform::dynload::cblas_sscal(n, a, x, 1); +} +#endif + +TEST(JitKernel, vscal) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d), y(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data()); + std::memcpy(y.data(), x.data(), sizeof(float) * d); + float a = 2.f; + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + float* y_data = y.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_ref(d, a, x_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto trefs1 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_inp_ref(d, a, y_data); + } + auto trefe1 = GetCurrentUS(); + +#ifdef PADDLE_WITH_MKLML + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_inp_mkl(d, a, y_data); + } + auto tmkle = GetCurrentUS(); +#endif + +#if defined __AVX__ || defined __AVX2__ + if (d == 8) { + auto si0 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_intri8(d, a, x_data, zref_data); + } + auto si1 = GetCurrentUS(); + auto si2 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vscal_inp_intri8(d, a, y_data); + } + auto si3 = GetCurrentUS(); + VLOG(30) << "Vec size 8 intr takes: " << (si1 - si0) / repeat + << " us, inplace: " << (si3 - si2) / repeat; + } +#endif + + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(&a, x_data, ztgt_data, d); + } + auto ttgte = GetCurrentUS(); + auto ttgts1 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(&a, y_data, y_data, d); + } + auto ttgte1 = GetCurrentUS(); + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat + << " us, inplace takes: " << (trefe1 - trefs1) / repeat +#ifdef PADDLE_WITH_MKLML + << " us, mkl inplace takes: " << (tmkle - tmkls) / repeat + << " us, " +#else + << " us, " +#endif + << "tgt takes: " << (ttgte - ttgts) / repeat + << "us, tgt inplace takes: " << (ttgte1 - ttgts1) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vmul_ref(const int n, const float* x, const float* y, float* z) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] * y[i]; + } +} + +#if defined __AVX__ || defined __AVX2__ +void vmul_intri8(const int n, const float* x, const float* y, float* z) { + __m256 tmpx, tmpy; + tmpx = _mm256_loadu_ps(x); + tmpy = _mm256_loadu_ps(y); + tmpx = _mm256_mul_ps(tmpx, tmpy); + _mm256_storeu_ps(z, tmpx); +} +#endif + +#ifdef PADDLE_WITH_MKLML +void vmul_mkl(const int n, const float* x, const float* y, float* z) { + paddle::platform::dynload::vsMul(n, x, y, z); +} +#endif + +TEST(JitKernel, vmul) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 20, 30, 256, 512, 1000, 1024}) { + std::vector x(d), y(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data()); + RandomVec(d, y.data()); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + const float* y_data = y.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vmul_ref(d, x_data, y_data, zref_data); + } + auto trefe = GetCurrentUS(); + +#ifdef PADDLE_WITH_MKLML + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vmul_mkl(d, x_data, y_data, zref_data); + } + auto tmkle = GetCurrentUS(); +#endif + +#if defined __AVX__ || defined __AVX2__ + if (d == 8) { + auto si0 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vmul_intri8(d, x_data, y_data, zref_data); + } + auto si1 = GetCurrentUS(); + VLOG(30) << "Vec size 8 intr takes: " << (si1 - si0) / repeat; + } +#endif + + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, y_data, ztgt_data, d); + } + auto ttgte = GetCurrentUS(); + + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat +#ifdef PADDLE_WITH_MKLML + << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, " +#else + << " us, " +#endif + << "tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vadd_ref(const int n, const float* x, const float* y, float* z) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] + y[i]; + } +} + +#if defined __AVX__ || defined __AVX2__ +void vadd_intri8(const int n, const float* x, const float* y, float* z) { + __m256 tmpx, tmpy; + tmpx = _mm256_loadu_ps(x); + tmpy = _mm256_loadu_ps(y); + tmpx = _mm256_add_ps(tmpx, tmpy); + _mm256_storeu_ps(z, tmpx); +} +#endif + +#ifdef PADDLE_WITH_MKLML +void vadd_mkl(const int n, const float* x, const float* y, float* z) { + paddle::platform::dynload::vsAdd(n, x, y, z); +} +#endif + +TEST(JitKernel, vadd) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d), y(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data()); + RandomVec(d, y.data()); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + const float* y_data = y.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vadd_ref(d, x_data, y_data, zref_data); + } + auto trefe = GetCurrentUS(); + +#ifdef PADDLE_WITH_MKLML + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vadd_mkl(d, x_data, y_data, zref_data); + } + auto tmkle = GetCurrentUS(); +#endif + +#if defined __AVX__ || defined __AVX2__ + if (d == 8) { + auto si0 = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vadd_intri8(d, x_data, y_data, zref_data); + } + auto si1 = GetCurrentUS(); + VLOG(30) << "Vec size 8 intr takes: " << (si1 - si0) / repeat; + } +#endif + + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, y_data, ztgt_data, d); + } + auto ttgte = GetCurrentUS(); + + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat +#ifdef PADDLE_WITH_MKLML + << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, " +#else + << " us, " +#endif + << "tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +void vaddrelu_ref(const int n, const float* x, const float* y, float* z) { + for (int i = 0; i < n; ++i) { + z[i] = x[i] + y[i]; + z[i] = z[i] > 0 ? z[i] : 0; + } +} +void vaddrelu_better( + const std::shared_ptr< + const paddle::operators::math::jitkernel::VAddKernel>& vadd, + const std::shared_ptr< + const paddle::operators::math::jitkernel::VReluKernel>& vrelu, + const float* x, const float* y, float* z, int d) { + vadd->Compute(x, y, z, d); + vrelu->Compute(z, z); +} + +TEST(JitKernel, vaddrelu) { + namespace jit = paddle::operators::math::jitkernel; + for (int d : {7, 8, 15, 16, 30, 256, 512}) { + std::vector x(d), y(d); + std::vector zref(d), ztgt(d); + RandomVec(d, x.data()); + RandomVec(d, y.data()); + const auto& ker = + jit::KernelPool::Instance().template Get>(d); + const auto& vadd = + jit::KernelPool::Instance().template Get>(d); + const auto& vrelu = + jit::KernelPool::Instance().template Get>(d); + const float* x_data = x.data(); + const float* y_data = y.data(); + float* ztgt_data = ztgt.data(); + float* zref_data = zref.data(); + auto trefs = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vadd_ref(d, x_data, y_data, zref_data); + } + auto trefe = GetCurrentUS(); + auto tmkls = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + vaddrelu_better(vadd, vrelu, x_data, y_data, zref_data, d); + } + auto tmkle = GetCurrentUS(); + auto ttgts = GetCurrentUS(); + for (int i = 0; i < repeat; ++i) { + ker->Compute(x_data, y_data, ztgt_data, d); + } + auto ttgte = GetCurrentUS(); + VLOG(30) << "Vec size " << d + << ": refer takes: " << (trefe - trefs) / repeat + << " us, better takes: " << (tmkle - tmkls) / repeat << " us, " + << "tgt takes: " << (ttgte - ttgts) / repeat; + for (int i = 0; i < d; ++i) { + EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3); + } + } +} + +TEST(JitKernel, pool) { + namespace jit = paddle::operators::math::jitkernel; + const int frame_size = 4; + std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh"; + const auto& plstm1 = + jit::KernelPool::Instance() + .template Get, const std::string&, + const std::string&, const std::string&>( + act_gate, act_cand, act_cell, frame_size, false); + const auto& plstm2 = + jit::KernelPool::Instance() + .template Get, const std::string&, + const std::string&, const std::string&>( + act_gate, act_cand, act_cell, frame_size, false); + const auto& peephole = + jit::KernelPool::Instance() + .template Get, const std::string&, + const std::string&, const std::string&>( + act_gate, act_cand, act_cell, frame_size, true); + EXPECT_TRUE(plstm1 != peephole); + + const auto& pvmul_f = + jit::KernelPool::Instance().template Get>(4); + EXPECT_TRUE(std::dynamic_pointer_cast(plstm2) != + std::dynamic_pointer_cast(pvmul_f)); + + const auto& pvmul_d = + jit::KernelPool::Instance().template Get>(4); + EXPECT_TRUE(std::dynamic_pointer_cast(pvmul_f) != + std::dynamic_pointer_cast(pvmul_d)); + + const auto& pvmul_from_key = jit::KernelPool::Instance().Get("vmulfjit4"); +#if defined(__APPLE__) || defined(__OSX__) || defined(_WIN32) + EXPECT_EQ(pvmul_from_key, nullptr); +#else + EXPECT_EQ(pvmul_from_key, pvmul_f); +#endif + const auto& pvmul_from_key2 = jit::KernelPool::Instance().Get("vmulfjit"); + EXPECT_TRUE(pvmul_from_key2 == nullptr); +} diff --git a/paddle/fluid/operators/math/pooling.cc b/paddle/fluid/operators/math/pooling.cc index b871851798e48e6b598cb4ab8e2e42db478a3820..8df43bb616179e2487534e0acabb71b09b87e1af 100644 --- a/paddle/fluid/operators/math/pooling.cc +++ b/paddle/fluid/operators/math/pooling.cc @@ -31,7 +31,7 @@ class Pool2dFunctor { const framework::Tensor& input, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* output) { + bool exclusive, framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_height = input.dims()[2]; const int input_width = input.dims()[3]; @@ -68,7 +68,8 @@ class Pool2dFunctor { pool_process.compute(input_data[h * input_width + w], &ele); } } - int pool_size = (hend - hstart) * (wend - wstart); + int pool_size = exclusive ? (hend - hstart) * (wend - wstart) + : ksize_height * ksize_width; pool_process.finalize(static_cast(pool_size), &ele); output_data[ph * output_width + pw] = ele; } @@ -93,7 +94,7 @@ class Pool2dGradFunctor { const framework::Tensor& output, const framework::Tensor& output_grad, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_grad_process, - framework::Tensor* input_grad) { + bool exclusive, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_height = input.dims()[2]; const int input_width = input.dims()[3]; @@ -124,7 +125,8 @@ class Pool2dGradFunctor { int wstart = pw * stride_width - padding_width; int wend = std::min(wstart + ksize_width, input_width); wstart = std::max(wstart, 0); - int pool_size = (hend - hstart) * (wend - wstart); + int pool_size = exclusive ? (hend - hstart) * (wend - wstart) + : ksize_height * ksize_width; float scale = 1.0 / pool_size; for (int h = hstart; h < hend; ++h) { for (int w = wstart; w < wend; ++w) { @@ -249,7 +251,7 @@ class Pool3dFunctor { const framework::Tensor& input, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* output) { + bool exclusive, framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_depth = input.dims()[2]; const int input_height = input.dims()[3]; @@ -300,7 +302,9 @@ class Pool3dFunctor { } } int pool_size = - (dend - dstart) * (hend - hstart) * (wend - wstart); + exclusive + ? (dend - dstart) * (hend - hstart) * (wend - wstart) + : ksize_depth * ksize_height * ksize_width; pool_process.finalize(static_cast(pool_size), &ele); output_data[output_idx] = ele; } @@ -326,7 +330,7 @@ class Pool3dGradFunctor { const framework::Tensor& output, const framework::Tensor& output_grad, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_grad_process, - framework::Tensor* input_grad) { + bool exclusive, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_depth = input.dims()[2]; const int input_height = input.dims()[3]; @@ -369,7 +373,9 @@ class Pool3dGradFunctor { wstart = std::max(wstart, 0); int pool_size = - (dend - dstart) * (hend - hstart) * (wend - wstart); + exclusive + ? (dend - dstart) * (hend - hstart) * (wend - wstart) + : ksize_depth * ksize_height * ksize_width; float scale = 1.0 / pool_size; for (int d = dstart; d < dend; ++d) { for (int h = hstart; h < hend; ++h) { diff --git a/paddle/fluid/operators/math/pooling.cu b/paddle/fluid/operators/math/pooling.cu index b1c76350d1724629bae175abf47e6671a1532242..a689eb42242e551caa3470f34f7e8d7e80b6dfbe 100644 --- a/paddle/fluid/operators/math/pooling.cu +++ b/paddle/fluid/operators/math/pooling.cu @@ -29,7 +29,7 @@ __global__ void KernelPool2D(const int nthreads, const T* input_data, const int ksize_width, const int stride_height, const int stride_width, const int padding_height, const int padding_width, PoolProcess pool_process, - T* output_data) { + bool exclusive, T* output_data) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int pw = index % output_width; @@ -52,7 +52,8 @@ __global__ void KernelPool2D(const int nthreads, const T* input_data, pool_process.compute(input_data[h * input_width + w], &ele); } } - int pool_size = (hend - hstart) * (wend - wstart); + int pool_size = exclusive ? (hend - hstart) * (wend - wstart) + : ksize_height * ksize_width; pool_process.finalize(static_cast(pool_size), &ele); output_data[index] = ele; } @@ -65,7 +66,7 @@ __global__ void KernelPool2DGrad( const int input_width, const int output_height, const int output_width, const int ksize_height, const int ksize_width, const int stride_height, const int stride_width, const int padding_height, const int padding_width, - PoolProcess pool_process, T* input_grad) { + PoolProcess pool_process, bool exclusive, T* input_grad) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int offsetW = index % input_width + padding_width; @@ -95,7 +96,8 @@ __global__ void KernelPool2DGrad( int wend = min(wstart + ksize_width, input_width); hstart = max(hstart, 0); wstart = max(wstart, 0); - int pool_size = (hend - hstart) * (wend - wstart); + int pool_size = exclusive ? (hend - hstart) * (wend - wstart) + : ksize_height * ksize_width; int output_sub_idx = ph * output_width + pw; pool_process.compute(input, output_data[output_sub_idx], output_grad[output_sub_idx], @@ -163,7 +165,7 @@ class Pool2dFunctor { const framework::Tensor& input, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* output) { + bool exclusive, framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_height = input.dims()[2]; @@ -189,7 +191,8 @@ class Pool2dFunctor { KernelPool2D<<>>( nthreads, input_data, input_channels, input_height, input_width, output_height, output_width, ksize_height, ksize_width, stride_height, - stride_width, padding_height, padding_width, pool_process, output_data); + stride_width, padding_height, padding_width, pool_process, exclusive, + output_data); } }; @@ -208,7 +211,7 @@ class Pool2dGradFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* input_grad) { + bool exclusive, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_height = input.dims()[2]; @@ -236,7 +239,7 @@ class Pool2dGradFunctor { nthreads, input_data, output_data, output_grad_data, input_channels, input_height, input_width, output_height, output_width, ksize_height, ksize_width, stride_height, stride_width, padding_height, padding_width, - pool_process, input_grad_data); + pool_process, exclusive, input_grad_data); } }; @@ -313,16 +316,14 @@ template class Pool2dGradFunctor; template -__global__ void KernelPool3D(const int nthreads, const T* input_data, - const int channels, const int input_depth, - const int input_height, const int input_width, - const int output_depth, const int output_height, - const int output_width, const int ksize_depth, - const int ksize_height, const int ksize_width, - const int stride_depth, const int stride_height, - const int stride_width, const int padding_depth, - const int padding_height, const int padding_width, - PoolProcess pool_process, T* output_data) { +__global__ void KernelPool3D( + const int nthreads, const T* input_data, const int channels, + const int input_depth, const int input_height, const int input_width, + const int output_depth, const int output_height, const int output_width, + const int ksize_depth, const int ksize_height, const int ksize_width, + const int stride_depth, const int stride_height, const int stride_width, + const int padding_depth, const int padding_height, const int padding_width, + PoolProcess pool_process, bool exclusive, T* output_data) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int pw = index % output_width; @@ -351,7 +352,9 @@ __global__ void KernelPool3D(const int nthreads, const T* input_data, } } } - int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + int pool_size = exclusive + ? (dend - dstart) * (hend - hstart) * (wend - wstart) + : ksize_depth * ksize_height * ksize_width; pool_process.finalize(static_cast(pool_size), &ele); output_data[index] = ele; } @@ -366,7 +369,7 @@ __global__ void KernelPool3DGrad( const int ksize_height, const int ksize_width, const int stride_depth, const int stride_height, const int stride_width, const int padding_depth, const int padding_height, const int padding_width, PoolProcess pool_process, - T* input_grad) { + bool exclusive, T* input_grad) { for (int index = blockIdx.x * blockDim.x + threadIdx.x; index < nthreads; index += blockDim.x * gridDim.x) { int offsetW = index % input_width + padding_width; @@ -409,7 +412,9 @@ __global__ void KernelPool3DGrad( dstart = max(dstart, 0); hstart = max(hstart, 0); wstart = max(wstart, 0); - int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + int pool_size = + exclusive ? (dend - dstart) * (hend - hstart) * (wend - wstart) + : ksize_depth * ksize_height * ksize_width; int output_sub_idx = (pd * output_height + ph) * output_width + pw; pool_process.compute(input, output_data[output_sub_idx], output_grad[output_sub_idx], @@ -484,7 +489,7 @@ class Pool3dFunctor { const framework::Tensor& input, const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* output) { + bool exclusive, framework::Tensor* output) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_depth = input.dims()[2]; @@ -517,7 +522,7 @@ class Pool3dFunctor { nthreads, input_data, input_channels, input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, stride_width, - padding_depth, padding_height, padding_width, pool_process, + padding_depth, padding_height, padding_width, pool_process, exclusive, output_data); } }; @@ -537,7 +542,7 @@ class Pool3dGradFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_process, - framework::Tensor* input_grad) { + bool exclusive, framework::Tensor* input_grad) { const int batch_size = input.dims()[0]; const int input_channels = input.dims()[1]; const int input_depth = input.dims()[2]; @@ -573,7 +578,7 @@ class Pool3dGradFunctor { input_depth, input_height, input_width, output_depth, output_height, output_width, ksize_depth, ksize_height, ksize_width, stride_depth, stride_height, stride_width, padding_depth, padding_height, - padding_width, pool_process, input_grad_data); + padding_width, pool_process, exclusive, input_grad_data); } }; diff --git a/paddle/fluid/operators/math/pooling.h b/paddle/fluid/operators/math/pooling.h index 120f5919803806e0d3b7dc8eaf530ae89819b84d..0f64e321bf01eea69767af020ed8c1a75e31acb5 100644 --- a/paddle/fluid/operators/math/pooling.h +++ b/paddle/fluid/operators/math/pooling.h @@ -89,7 +89,7 @@ class Pool2dFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_compute, - framework::Tensor* output); + bool exclusive, framework::Tensor* output); }; template @@ -101,7 +101,7 @@ class Pool2dGradFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_compute, - framework::Tensor* input_grad); + bool exclusive, framework::Tensor* input_grad); }; template @@ -123,7 +123,7 @@ class Pool3dFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_compute, - framework::Tensor* output); + bool exclusive, framework::Tensor* output); }; template @@ -135,7 +135,7 @@ class Pool3dGradFunctor { const std::vector& ksize, const std::vector& strides, const std::vector& paddings, PoolProcess pool_compute, - framework::Tensor* input_grad); + bool exclusive, framework::Tensor* input_grad); }; template diff --git a/paddle/fluid/operators/math/selected_rows_functor.cc b/paddle/fluid/operators/math/selected_rows_functor.cc index 8e8baf49b2330e95ff1a868b0b0a03bc10d84484..9577a4cb9d275df9604b7578f8685e4d2938a5e9 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cc +++ b/paddle/fluid/operators/math/selected_rows_functor.cc @@ -13,9 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #include -#include +#include -#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" namespace paddle { @@ -150,6 +150,45 @@ template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; +template +struct SelectedRowsSumTo { + void operator()(const platform::CPUDeviceContext& context, + const std::vector& input1, + const std::vector& input2_offsets, + framework::SelectedRows* input2) { + // Ensure all selected rows have the same height + size_t size = 0u; + for (auto iter = input1.begin(); iter != input1.end(); ++iter) { + auto& in_rows = (*iter)->rows(); + size += in_rows.end() - in_rows.begin(); + auto in1_height = (*iter)->height(); + PADDLE_ENFORCE_EQ(in1_height, input2->height()); + } + // concat rows + std::vector in2_rows; + in2_rows.reserve(in2_rows.size() + size); + for (auto iter = input1.begin(); iter != input1.end(); ++iter) { + const framework::Vector& in_rows = (*iter)->rows(); + in2_rows.insert(in2_rows.end(), in_rows.begin(), in_rows.end()); + } + input2->set_rows(in2_rows); + + auto* in2_value = input2->mutable_value(); + auto* in2_data = in2_value->data(); + auto blas = math::GetBlas(context); + size_t offset = 0u; + for (size_t i = 0u; i != input1.size(); ++i) { + auto& in_value = input1[i]->value(); + const auto* in_data = in_value.data(); + offset += input2_offsets[i]; + blas.VCOPY(in_value.numel(), in_data, in2_data + offset); + } + } +}; + +template struct SelectedRowsSumTo; +template struct SelectedRowsSumTo; + template struct SelectedRowsAddToTensor { void operator()(const platform::CPUDeviceContext& context, @@ -190,8 +229,24 @@ template struct SelectedRowsAddToTensor; // add or mul. namespace scatter { -size_t FindPos(const std::vector& rows, int64_t value) { - return std::find(rows.begin(), rows.end(), value) - rows.begin(); +template +typename std::enable_if< + std::is_floating_point::value && + std::is_same::value>::type +elementwise_add_to(const DeviceContext& ctx, BlasT* blas, + size_t data_len, const T* in, T* out) { + blas->AXPY(data_len, 1., in, out); +} + +template +typename std::enable_if< + !std::is_floating_point::value && + std::is_same::value>::type +elementwise_add_to(const DeviceContext& ctx, BlasT* blas, + size_t data_len, const T* in, T* out) { + for (int64_t i = 0; i < data_len; i++) { + out[i] += in[i]; + } } template @@ -206,14 +261,52 @@ struct MergeAdd { void operator()(const platform::CPUDeviceContext& context, const framework::SelectedRows& input, framework::SelectedRows* output) { - framework::SelectedRows& out = *output; - auto input_rows = input.rows(); - std::set row_set(input_rows.begin(), input_rows.end()); - std::vector merge_rows(row_set.begin(), row_set.end()); + std::vector inputs; + inputs.push_back(&input); + (*this)(context, inputs, output); + } - auto input_width = input.value().dims()[1]; + void operator()(const platform::CPUDeviceContext& context, + const std::vector& inputs, + framework::SelectedRows* output) { + if (inputs.size() == 0) { + VLOG(30) << "no input! return"; + return; + } + const framework::SelectedRows* has_value_input = nullptr; + for (auto* in : inputs) { + if (in->rows().size() > 0) { + has_value_input = in; + break; + } + } + if (has_value_input == nullptr) { + VLOG(30) << "no input has value! just return" << std::endl; + return; + } + auto input_width = has_value_input->value().dims()[1]; + auto input_height = has_value_input->height(); + framework::SelectedRows& out = *output; + std::set merged_row_set; + for (auto* input : inputs) { + if (input->rows().size() == 0) { + continue; + } + PADDLE_ENFORCE_EQ(input_width, input->value().dims()[1], + "all input should have same " + "dimension except for the first one"); + PADDLE_ENFORCE_EQ(input_height, input->height(), + "all input should have same height"); + merged_row_set.insert(input->rows().begin(), input->rows().end()); + } + std::vector merge_rows(merged_row_set.begin(), + merged_row_set.end()); + std::unordered_map rows_to_id; + for (size_t i = 0; i < merge_rows.size(); ++i) { + rows_to_id[merge_rows[i]] = i; + } out.set_rows(merge_rows); - out.set_height(input.height()); + out.set_height(input_height); out.mutable_value()->mutable_data( framework::make_ddim( {static_cast(merge_rows.size()), input_width}), @@ -223,21 +316,29 @@ struct MergeAdd { constant_functor(context, out.mutable_value(), 0.0); auto* out_data = out.mutable_value()->data(); - auto* input_data = input.value().data(); - for (size_t i = 0; i < input_rows.size(); i++) { - size_t out_i = FindPos(merge_rows, input_rows[i]); - for (int64_t j = 0; j < input_width; j++) { - out_data[out_i * input_width + j] += input_data[i * input_width + j]; + auto blas = math::GetBlas(context); + for (auto* input : inputs) { + if (input->rows().size() == 0) { + continue; + } + auto* input_data = input->value().data(); + auto& input_rows = input->rows(); + + for (size_t i = 0; i < input_rows.size(); i++) { + size_t out_i = rows_to_id[input_rows[i]]; + elementwise_add_to( + context, &blas, static_cast(input_width), + &input_data[i * input_width], &out_data[out_i * input_width]); } } } }; -template struct MergeAdd; -template struct MergeAdd; template struct MergeAdd; template struct MergeAdd; +template struct MergeAdd; +template struct MergeAdd; template struct UpdateToTensor { diff --git a/paddle/fluid/operators/math/selected_rows_functor.cu b/paddle/fluid/operators/math/selected_rows_functor.cu index ba8eccf82042b679f69a32f9d053f05ac8fb9a99..74b9659cfd38076bf1948b5c664817a6753b7090 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cu +++ b/paddle/fluid/operators/math/selected_rows_functor.cu @@ -18,6 +18,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/cuda_primitives.h" +#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { @@ -80,7 +81,7 @@ template __global__ void SelectedRowsAddTensorKernel(const T* selected_rows, const int64_t* rows, T* tensor_out, int64_t row_numel) { - const int ty = blockIdx.y; + const int ty = blockIdx.x; int tid = threadIdx.x; selected_rows += ty * row_numel; @@ -118,11 +119,11 @@ struct SelectedRowsAddTensor { auto* out_data = output->data(); SetConstant functor; - functor(context, output, 0.0); + functor(context, output, static_cast(0)); const int block_size = 256; dim3 threads(block_size, 1); - dim3 grid(1, in1_rows.size()); + dim3 grid(in1_rows.size(), 1); SelectedRowsAddTensorKernel< T, block_size><<>>( in1_data, in1_rows.CUDAData(context.GetPlace()), out_data, @@ -136,6 +137,9 @@ struct SelectedRowsAddTensor { template struct SelectedRowsAddTensor; template struct SelectedRowsAddTensor; +template struct SelectedRowsAdd; +template struct SelectedRowsAddTensor; template struct SelectedRowsAddTo { @@ -175,6 +179,8 @@ template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; +template struct SelectedRowsAddTo; namespace { template @@ -182,7 +188,7 @@ __global__ void SelectedRowsAddToTensorKernel(const T* selected_rows, const int64_t* rows, T* tensor_out, int64_t row_numel) { - const int ty = blockIdx.y; + const int ty = blockIdx.x; int tid = threadIdx.x; selected_rows += ty * row_numel; @@ -215,7 +221,7 @@ struct SelectedRowsAddToTensor { auto* in2_data = input2->data(); const int block_size = 256; dim3 threads(block_size, 1); - dim3 grid(1, in1_rows.size()); + dim3 grid(in1_rows.size(), 1); SelectedRowsAddToTensorKernel< T, block_size><<>>( in1_data, in1_rows.CUDAData(context.GetPlace()), in2_data, @@ -227,6 +233,8 @@ template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; +template struct SelectedRowsAddToTensor; namespace scatter { @@ -267,10 +275,15 @@ struct MergeAdd { void operator()(const platform::CUDADeviceContext& context, const framework::SelectedRows& input, framework::SelectedRows* output) { - framework::SelectedRows& out = *output; framework::Vector input_rows(input.rows()); + if (input_rows.size() == 0) { + return; + } + + framework::SelectedRows& out = *output; std::set row_set(input_rows.begin(), input_rows.end()); - std::vector merge_rows(row_set.begin(), row_set.end()); + std::vector merge_rows_cpu(row_set.begin(), row_set.end()); + framework::Vector merge_rows(merge_rows_cpu); auto input_width = input.value().dims()[1]; @@ -282,7 +295,7 @@ struct MergeAdd { context.GetPlace()); math::SetConstant constant_functor; - constant_functor(context, out.mutable_value(), 0.0); + constant_functor(context, out.mutable_value(), static_cast(0)); auto* out_data = out.mutable_value()->data(); auto* input_data = input.value().data(); @@ -296,18 +309,86 @@ struct MergeAdd { out.mutable_rows()->CUDAMutableData(context.GetPlace()), out.rows().size(), input_width); } + + void operator()(const platform::CUDADeviceContext& context, + const std::vector& inputs, + framework::SelectedRows* output) { + if (inputs.size() == 0) { + VLOG(30) << "no input! return"; + return; + } + const framework::SelectedRows* has_value_input = nullptr; + for (auto* in : inputs) { + if (in->rows().size() > 0) { + has_value_input = in; + break; + } + } + if (has_value_input == nullptr) { + VLOG(30) << "no input has value! just return" << std::endl; + return; + } + auto input_width = has_value_input->value().dims()[1]; + auto input_height = has_value_input->height(); + framework::SelectedRows& out = *output; + std::set merged_row_set; + for (auto* input : inputs) { + if (input->rows().size() == 0) { + continue; + } + PADDLE_ENFORCE_EQ(input_width, input->value().dims()[1], + "all input should have same " + "dimension except for the first one"); + PADDLE_ENFORCE_EQ(input_height, input->height(), + "all input should have same height"); + merged_row_set.insert(input->rows().begin(), input->rows().end()); + } + std::vector merge_rows_cpu(merged_row_set.begin(), + merged_row_set.end()); + framework::Vector merge_rows(merge_rows_cpu); + + out.set_rows(merge_rows); + out.set_height(input_height); + out.mutable_value()->mutable_data( + framework::make_ddim( + {static_cast(merge_rows.size()), input_width}), + context.GetPlace()); + + math::SetConstant constant_functor; + constant_functor(context, out.mutable_value(), static_cast(0)); + + auto* out_data = out.mutable_value()->data(); + + const int block_size = 256; + dim3 threads(block_size, 1); + + for (auto* input : inputs) { + if (input->rows().size() == 0) { + continue; + } + auto* input_data = input->value().data(); + auto& input_rows = input->rows(); + dim3 grid1(input_rows.size(), 1); + + MergeAddKernel<<>>( + input_data, input_rows.CUDAData(context.GetPlace()), out_data, + out.mutable_rows()->CUDAMutableData(context.GetPlace()), + out.rows().size(), input_width); + } + } }; template struct MergeAdd; template struct MergeAdd; template struct MergeAdd; template struct MergeAdd; +template struct MergeAdd; template __global__ void UpdateToTensorKernel(const T* selected_rows, const int64_t* rows, const ScatterOps& op, T* tensor_out, int64_t row_numel) { - const int ty = blockIdx.y; + const int ty = blockIdx.x; int tid = threadIdx.x; selected_rows += ty * row_numel; @@ -376,7 +457,7 @@ struct UpdateToTensor { auto* in2_data = input2->data(); dim3 threads(platform::PADDLE_CUDA_NUM_THREADS, 1); - dim3 grid(1, in1_rows.size()); + dim3 grid(in1_rows.size(), 1); UpdateToTensorKernel<<< grid, threads, 0, context.stream()>>>(in1_data, in1_rows.cuda_data(), op, in2_data, in1_row_numel); diff --git a/paddle/fluid/operators/math/selected_rows_functor.h b/paddle/fluid/operators/math/selected_rows_functor.h index aa419f74fcd2a53cdd734ec270bc154b78c9f2ff..6d146d39d6d07678e859b82b25ba60ed7661546d 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.h +++ b/paddle/fluid/operators/math/selected_rows_functor.h @@ -12,8 +12,14 @@ 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 "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/platform/device_context.h" #define INLINE_FOR2(sizei, sizej) \ @@ -49,6 +55,17 @@ struct SelectedRowsAddTo { const int64_t input2_offset, framework::SelectedRows* input2); }; +// input2 = [all input in input1] + input2 +template +struct SelectedRowsSumTo { + void operator()(const DeviceContext& context, + const std::vector& input1, + const std::vector& input2_offsets, + framework::SelectedRows* input2); +}; + +// FIXME: The result of SelectedRowsAddToTensor maybe non deterministic, +// because it uses CudaAtomicAdd. // input2 = input1 + input2 template struct SelectedRowsAddToTensor { @@ -68,57 +85,9 @@ struct MergeAdd { void operator()(const DeviceContext& context, const framework::SelectedRows& input, framework::SelectedRows* output); -}; - -template -struct Add { - framework::SelectedRows operator()(const DeviceContext& context, - const framework::SelectedRows& input1, - const framework::SelectedRows& input2) { - framework::SelectedRows out; - out.set_rows(input1.rows()); - out.set_height(input1.height()); - out.mutable_value()->mutable_data(input1.value().dims(), - context.GetPlace()); - auto e_out = framework::EigenVector::Flatten(*(out.mutable_value())); - auto e_in1 = framework::EigenVector::Flatten(input1.value()); - auto e_in2 = framework::EigenVector::Flatten(input2.value()); - e_out.device(*context.eigen_device()) = e_in1 + e_in2; - return out; - } -}; - -template -struct Mul { - // multiply two SelectedRows - framework::SelectedRows operator()(const DeviceContext& context, - const framework::SelectedRows& input1, - const framework::SelectedRows& input2) { - framework::SelectedRows out; - out.set_rows(input1.rows()); - out.set_height(input1.height()); - out.mutable_value()->mutable_data(input1.value().dims(), - context.GetPlace()); - auto e_out = framework::EigenVector::Flatten(*(out.mutable_value())); - auto e_in1 = framework::EigenVector::Flatten(input1.value()); - auto e_in2 = framework::EigenVector::Flatten(input2.value()); - e_out.device(*context.eigen_device()) = e_in1 * e_in2; - return out; - } - // multiply scalar to SelectedRows - framework::SelectedRows operator()(const DeviceContext& context, - const framework::SelectedRows& input1, - const T input2) { - framework::SelectedRows out; - out.set_rows(input1.rows()); - out.set_height(input1.height()); - out.mutable_value()->mutable_data(input1.value().dims(), - context.GetPlace()); - auto e_out = framework::EigenVector::Flatten(*(out.mutable_value())); - auto e_in1 = framework::EigenVector::Flatten(input1.value()); - e_out.device(*context.eigen_device()) = input2 * e_in1; - return out; - } + void operator()(const DeviceContext& context, + const std::vector& inputs, + framework::SelectedRows* output); }; enum class ScatterOps { ASSIGN, ADD, SUB, SUBBY, MUL, DIV, DIVBY }; diff --git a/paddle/fluid/operators/math/selected_rows_functor_test.cc b/paddle/fluid/operators/math/selected_rows_functor_test.cc index 70bed820ee58885861fa8c5535c931f258625572..f15b37a1e3f0ae9c7612c4f74470472393ff4ad6 100644 --- a/paddle/fluid/operators/math/selected_rows_functor_test.cc +++ b/paddle/fluid/operators/math/selected_rows_functor_test.cc @@ -219,3 +219,234 @@ TEST(selected_rows_functor, cpu_add_to) { // row9: 2.0 + 3.0 EXPECT_EQ(tensor1_data[9 * row_numel + 6], 5.0); } + +TEST(selected_rows_functor, cpu_merge_add_float) { + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + functor; + int64_t height = 10; + int64_t row_numel = 10; + + std::vector rows{0, 4, 4, 7}; + std::unique_ptr selected_rows{ + new paddle::framework::SelectedRows(rows, height)}; + auto* in_value = selected_rows->mutable_value(); + in_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows.size()), row_numel}), + cpu_place); + functor(ctx, in_value, 1.0); + + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + + paddle::operators::math::scatter::MergeAdd + merge_add_functor; + merge_add_functor(ctx, *selected_rows, output.get()); + + auto out_height = output->height(); + EXPECT_EQ(out_height, height); + + auto& out_rows = output->rows(); + EXPECT_EQ(out_rows[0], 0); + EXPECT_EQ(out_rows[1], 4); + EXPECT_EQ(out_rows[2], 7); + + auto* out_data = output->value().data(); + + EXPECT_EQ(out_data[0 * row_numel], 1.0); + EXPECT_EQ(out_data[1 * row_numel], 2.0); + EXPECT_EQ(out_data[2 * row_numel], 1.0); +} + +TEST(selected_rows_functor, cpu_merge_add_int) { + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + functor; + int64_t height = 10; + int64_t row_numel = 10; + + std::vector rows{0, 4, 4, 7}; + std::unique_ptr selected_rows{ + new paddle::framework::SelectedRows(rows, height)}; + auto* in_value = selected_rows->mutable_value(); + in_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows.size()), row_numel}), + cpu_place); + functor(ctx, in_value, 1); + + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + + paddle::operators::math::scatter::MergeAdd + merge_add_functor; + merge_add_functor(ctx, *selected_rows, output.get()); + + auto out_height = output->height(); + EXPECT_EQ(out_height, height); + + auto& out_rows = output->rows(); + EXPECT_EQ(out_rows[0], 0); + EXPECT_EQ(out_rows[1], 4); + EXPECT_EQ(out_rows[2], 7); + + auto* out_data = output->value().data(); + + EXPECT_EQ(out_data[0 * row_numel], 1); + EXPECT_EQ(out_data[1 * row_numel], 2); + EXPECT_EQ(out_data[2 * row_numel], 1); +} + +TEST(selected_rows_functor, cpu_merge_add_multi) { + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + set_const; + + int64_t height = 10; + int64_t row_numel = 8; + + std::vector rows1{5, 2, 5, 3, 5}; + std::unique_ptr selected_rows1{ + new paddle::framework::SelectedRows(rows1, height)}; + auto* in1_value = selected_rows1->mutable_value(); + in1_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows1.size()), row_numel}), + cpu_place); + set_const(ctx, in1_value, 1.0); + + std::vector rows2{2, 5, 3, 5, 3}; + std::unique_ptr selected_rows2{ + new paddle::framework::SelectedRows(rows2, height)}; + auto* in2_value = selected_rows2->mutable_value(); + in2_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows2.size()), row_numel}), + cpu_place); + set_const(ctx, in2_value, 1.0); + + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + output->set_height(height); + paddle::operators::math::scatter::MergeAdd + merge_add_functor; + + std::vector inputs; + inputs.push_back(selected_rows1.get()); + inputs.push_back(selected_rows2.get()); + merge_add_functor(ctx, inputs, output.get()); + + EXPECT_EQ(output->height(), height); + EXPECT_EQ(output->value().dims(), + paddle::framework::make_ddim({3, row_numel})); + + std::vector ret_rows{2, 3, 5}; + EXPECT_EQ(output->rows(), ret_rows); + + auto* out_data = output->value().data(); + for (size_t i = 0; i < ret_rows.size(); ++i) { + for (size_t j = 0; j < row_numel; ++j) { + EXPECT_EQ(out_data[i * row_numel + j], ret_rows[i]); + } + } +} + +TEST(selected_rows_functor, cpu_sum_to) { + paddle::platform::CPUPlace cpu_place; + paddle::platform::CPUDeviceContext ctx(cpu_place); + paddle::operators::math::SetConstant + functor; + int64_t height = 10; + int64_t row_numel = 10; + std::vector rows1{0, 4, 7}; + std::unique_ptr selected_rows1{ + new paddle::framework::SelectedRows(rows1, height)}; + auto* in1_value = selected_rows1->mutable_value(); + in1_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows1.size()), row_numel}), + cpu_place); + + functor(ctx, in1_value, 1.0); + std::vector rows2{0, 5, 7, 9}; + std::unique_ptr selected_rows2{ + new paddle::framework::SelectedRows(rows2, height)}; + auto* in2_value = selected_rows2->mutable_value(); + in2_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows2.size()), row_numel}), + cpu_place); + + functor(ctx, in2_value, 2.0); + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + output->set_height(height); + auto* out_value = output->mutable_value(); + // simplely concat two SelectedRows + out_value->mutable_data(paddle::framework::make_ddim({7, 10}), + cpu_place); + paddle::operators::math::SelectedRowsSumTo + sum_to_functor; + sum_to_functor(ctx, std::vector( + {selected_rows1.get(), selected_rows2.get()}), + std::vector({0, in1_value->numel()}), output.get()); + auto out_height = output->height(); + EXPECT_EQ(out_height, height); + auto& out_rows = output->rows(); + // input1 rows + EXPECT_EQ(out_rows[0], 0); + EXPECT_EQ(out_rows[1], 4); + EXPECT_EQ(out_rows[2], 7); + // input2 rows + EXPECT_EQ(out_rows[3], 0); + EXPECT_EQ(out_rows[4], 5); + EXPECT_EQ(out_rows[5], 7); + EXPECT_EQ(out_rows[6], 9); + auto* out_data = output->value().data(); + // input1 value + EXPECT_EQ(out_data[0 * row_numel + 0], 1.0); + EXPECT_EQ(out_data[0 * row_numel + 8], 1.0); + EXPECT_EQ(out_data[1 * row_numel + 1], 1.0); + EXPECT_EQ(out_data[2 * row_numel + 6], 1.0); + // input2 value + EXPECT_EQ(out_data[3 * row_numel + 3], 2.0); + EXPECT_EQ(out_data[3 * row_numel + 8], 2.0); + EXPECT_EQ(out_data[4 * row_numel + 4], 2.0); + EXPECT_EQ(out_data[5 * row_numel + 7], 2.0); + EXPECT_EQ(out_data[6 * row_numel + 9], 2.0); + std::unique_ptr tensor1{ + new paddle::framework::Tensor()}; + tensor1->mutable_data( + paddle::framework::make_ddim({height, row_numel}), cpu_place); + functor(ctx, tensor1.get(), 3.0); + paddle::operators::math::SelectedRowsAddToTensor< + paddle::platform::CPUDeviceContext, float> + add_to_tensor_functor; + add_to_tensor_functor(ctx, *output, tensor1.get()); + auto* tensor1_data = tensor1->data(); + // row0: 1.0 + 2.0 + 3.0 + EXPECT_EQ(tensor1_data[0 * row_numel + 0], 6.0); + // row1: 3.0 + EXPECT_EQ(tensor1_data[1 * row_numel + 1], 3.0); + // row4 : 1.0 + 3.0 + EXPECT_EQ(tensor1_data[4 * row_numel + 6], 4.0); + // row5: 2.0 + 3.0 + EXPECT_EQ(tensor1_data[5 * row_numel + 7], 5.0); + // row6: 3.0 + EXPECT_EQ(tensor1_data[6 * row_numel + 1], 3.0); + // row7: 1.0 + 2.0 + 3.0 + EXPECT_EQ(tensor1_data[7 * row_numel + 3], 6.0); + // row9: 2.0 + 3.0 + EXPECT_EQ(tensor1_data[9 * row_numel + 6], 5.0); +} diff --git a/paddle/fluid/operators/math/selected_rows_functor_test.cu b/paddle/fluid/operators/math/selected_rows_functor_test.cu index 5fc50aba25d8e69480a17f0f80877b0d03e17276..17af3e3999ca688c584f636f4c00386f886f9bbf 100644 --- a/paddle/fluid/operators/math/selected_rows_functor_test.cu +++ b/paddle/fluid/operators/math/selected_rows_functor_test.cu @@ -241,3 +241,67 @@ TEST(selected_rows_functor, gpu_add_to) { // row9: 2.0 + 3.0 EXPECT_EQ(tensor1_cpu_data[9 * row_numel + 6], 5.0); } + +TEST(selected_rows_functor, gpu_merge_add) { + paddle::platform::CUDAPlace gpu_place(0); + paddle::platform::CPUPlace cpu_place; + paddle::platform::CUDADeviceContext& ctx = + *reinterpret_cast( + paddle::platform::DeviceContextPool::Instance().Get(gpu_place)); + paddle::operators::math::SetConstant + set_const; + + int64_t height = 10; + int64_t row_numel = 8; + + std::vector rows1{5, 2, 5, 3, 5}; + std::unique_ptr selected_rows1{ + new paddle::framework::SelectedRows(rows1, height)}; + auto* in1_value = selected_rows1->mutable_value(); + in1_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows1.size()), row_numel}), + gpu_place); + set_const(ctx, in1_value, 1.0); + + std::vector rows2{2, 5, 3, 5, 3}; + std::unique_ptr selected_rows2{ + new paddle::framework::SelectedRows(rows2, height)}; + auto* in2_value = selected_rows2->mutable_value(); + in2_value->mutable_data( + paddle::framework::make_ddim( + {static_cast(rows2.size()), row_numel}), + gpu_place); + set_const(ctx, in2_value, 1.0); + + std::unique_ptr output{ + new paddle::framework::SelectedRows()}; + output->set_height(height); + paddle::operators::math::scatter::MergeAdd< + paddle::platform::CUDADeviceContext, float> + merge_add_functor; + + std::vector inputs; + inputs.push_back(selected_rows1.get()); + inputs.push_back(selected_rows2.get()); + merge_add_functor(ctx, inputs, output.get()); + + paddle::framework::Tensor output_cpu; + paddle::framework::TensorCopy(output->value(), cpu_place, ctx, &output_cpu); + ctx.Wait(); + + EXPECT_EQ(output->height(), height); + EXPECT_EQ(output->value().dims(), + paddle::framework::make_ddim({3, row_numel})); + + std::vector ret_rows{2, 3, 5}; + EXPECT_EQ(output->rows(), ret_rows); + + auto* out_data = output_cpu.data(); + for (size_t i = 0; i < ret_rows.size(); ++i) { + for (size_t j = 0; j < row_numel; ++j) { + EXPECT_EQ(out_data[i * row_numel + j], ret_rows[i]); + } + } +} diff --git a/paddle/fluid/operators/math/sequence_pooling.cc b/paddle/fluid/operators/math/sequence_pooling.cc index 69318a6598c8c69eceab7216df6382537153d34f..6d491dbf1ed162ef07fda4c07e95cc57108486fd 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cc +++ b/paddle/fluid/operators/math/sequence_pooling.cc @@ -12,9 +12,11 @@ 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/math/sequence_pooling.h" #include + +#include "paddle/fluid/operators/math/blas.h" #include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/sequence_pooling.h" namespace paddle { namespace operators { @@ -29,7 +31,7 @@ template using EigenMatrix = framework::EigenMatrix; -template +template class MaxSeqPoolFunctor { public: void operator()(const platform::CPUDeviceContext& context, @@ -68,7 +70,41 @@ class MaxSeqPoolFunctor { } } }; +// Instantisation of Max Sequence Pooling for test phase eg. no need to fill +// index buffer +template +class MaxSeqPoolFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::LoDTensor& input, framework::Tensor* output, + framework::Tensor* index) { + auto in_dims = input.dims(); + auto out_dims = output->dims(); + PADDLE_ENFORCE_GT(in_dims.size(), 1); + PADDLE_ENFORCE_GT(out_dims.size(), 1); + for (int64_t i = 1; i < in_dims.size(); ++i) { + PADDLE_ENFORCE_EQ(in_dims[i], out_dims[i]); + } + auto starts = input.lod()[0]; + const T* in_data = input.data(); + T* out_data = output->data(); + + int64_t num_seq = out_dims[0]; + int64_t dim = output->numel() / num_seq; + for (int64_t i = 0; i < num_seq; ++i) { + std::memcpy(&out_data[i * dim], &in_data[starts[i] * dim], + dim * sizeof(T)); + for (size_t j = starts[i] + 1; j < starts[i + 1]; ++j) { + for (int64_t k = 0; k < dim; ++k) { + if (in_data[j * dim + k] > out_data[i * dim + k]) { + out_data[i * dim + k] = in_data[j * dim + k]; + } + } + } + } + } +}; template class MaxSeqPoolGradFunctor { public: @@ -155,17 +191,47 @@ class FirstSeqPoolFunctor { } }; +template +class SumSeqPoolGradFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& out_grad, + framework::LoDTensor* in_grad) { + auto lod = in_grad->lod()[0]; + int64_t out_w = out_grad.numel() / out_grad.dims()[0]; + int64_t in_w = in_grad->numel() / in_grad->dims()[0]; + PADDLE_ENFORCE(in_w == out_w); + const T* out_g_data = out_grad.data(); + T* in_g_data = in_grad->mutable_data(context.GetPlace()); + auto blas = math::GetBlas(context); + for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { + int64_t h = static_cast(lod[i + 1] - lod[i]); + int64_t in_offset = lod[i] * in_w; + const T* out_pos = out_g_data + i * out_w; + T* in_pos = in_g_data + in_offset; + for (int r = 0; r != h; ++r) { + blas.VCOPY(in_w, out_pos, in_pos + r * in_w); + } + } + } +}; + template class SequencePoolFunctor { public: /* max pool has index output */ void operator()(const platform::CPUDeviceContext& context, const std::string pooltype, const framework::LoDTensor& input, - framework::Tensor* output, + framework::Tensor* output, bool is_test, framework::Tensor* index = nullptr) { if (pooltype == "MAX") { - math::MaxSeqPoolFunctor max_pool; - max_pool(context, input, output, index); + if (is_test) { + math::MaxSeqPoolFunctor max_pool; + max_pool(context, input, output, index); + } else { + math::MaxSeqPoolFunctor max_pool; + max_pool(context, input, output, index); + } return; } if (pooltype == "LAST") { @@ -173,6 +239,7 @@ class SequencePoolFunctor { last_pool(context, input, output); return; } + if (pooltype == "FIRST") { math::FirstSeqPoolFunctor first_pool; first_pool(context, input, output); @@ -180,6 +247,7 @@ class SequencePoolFunctor { } auto lod = input.lod()[0]; auto& place = *context.eigen_device(); + auto blas = math::GetBlas(context); for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { Tensor in_t = input.Slice(static_cast(lod[i]), static_cast(lod[i + 1])); @@ -191,7 +259,14 @@ class SequencePoolFunctor { if (pooltype == "AVERAGE") { out_e.device(place) = in_e.mean(Eigen::array({{0}})); } else if (pooltype == "SUM") { - out_e.device(place) = in_e.sum(Eigen::array({{0}})); + if (h > 0) { + const T* in_data = in_t.data(); + T* out_data = out_t.mutable_data(context.GetPlace()); + blas.VCOPY(w, in_data, out_data); + for (int64_t r = 1; r != h; ++r) { + blas.AXPY(w, 1., in_data + r * w, out_data); + } + } } else if (pooltype == "SQRT") { out_e.device(place) = in_e.sum(Eigen::array({{0}})) / std::sqrt(static_cast(h)); @@ -221,6 +296,13 @@ class SequencePoolGradFunctor { math::SetConstant functor; functor(context, in_grad, 0); } + + if (pooltype == "SUM") { + math::SumSeqPoolGradFunctor sum_pool_grad; + sum_pool_grad(context, out_grad, in_grad); + return; + } + auto lod = in_grad->lod()[0]; auto& place = *context.eigen_device(); for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { @@ -236,8 +318,6 @@ class SequencePoolGradFunctor { if (pooltype == "AVERAGE") { in_g_e.device(place) = (out_g_e / static_cast(h)).broadcast(bcast); - } else if (pooltype == "SUM") { - in_g_e.device(place) = (out_g_e).broadcast(bcast); } else if (pooltype == "SQRT") { in_g_e.device(place) = (out_g_e / std::sqrt(static_cast(h))).broadcast(bcast); diff --git a/paddle/fluid/operators/math/sequence_pooling.cu b/paddle/fluid/operators/math/sequence_pooling.cu index a92aef805a0434f2ebcbc62d4e5eaef0cfb21bfa..0015fafbc892912424dfa6dbd1778438d384ca19 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cu +++ b/paddle/fluid/operators/math/sequence_pooling.cu @@ -133,7 +133,7 @@ class SequencePoolFunctor { public: void operator()(const platform::CUDADeviceContext& context, const std::string pooltype, const framework::LoDTensor& input, - framework::Tensor* output, + framework::Tensor* output, bool is_test, framework::Tensor* index = nullptr) { auto& lod = input.lod()[0]; const size_t item_dim = output->numel() / output->dims()[0]; diff --git a/paddle/fluid/operators/math/sequence_pooling.h b/paddle/fluid/operators/math/sequence_pooling.h index 8dcbee65d0b63a137e5f422ec8667cc950641b4a..a1046ea2160d0ae9c2251612c97d3f2640b0aad1 100644 --- a/paddle/fluid/operators/math/sequence_pooling.h +++ b/paddle/fluid/operators/math/sequence_pooling.h @@ -28,7 +28,7 @@ class SequencePoolFunctor { /* max pool has index output */ void operator()(const DeviceContext& context, const std::string pooltype, const framework::LoDTensor& input, framework::Tensor* output, - framework::Tensor* index = nullptr); + bool is_test = false, framework::Tensor* index = nullptr); }; template diff --git a/paddle/fluid/operators/math/sequence_pooling_test.cc b/paddle/fluid/operators/math/sequence_pooling_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..2bc008dd34ffcfe93a00bd4a8cde61626d91e235 --- /dev/null +++ b/paddle/fluid/operators/math/sequence_pooling_test.cc @@ -0,0 +1,126 @@ +/* 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/math/sequence_pooling.h" +#include +#include + +template +void TestSequencePoolingSum(const paddle::framework::LoD& lod) { + paddle::framework::LoDTensor cpu_out_grad; + paddle::framework::LoDTensor cpu_in_grad; + paddle::framework::LoDTensor out_grad; + paddle::framework::LoDTensor in_grad; + const size_t second_dim = 128u; + + // construct out_grad's tensor in cpu + const size_t out_first_dim = lod[0].size() - 1; + auto out_dims = paddle::framework::make_ddim( + {static_cast(out_first_dim), static_cast(second_dim)}); + + cpu_out_grad.mutable_data(out_dims, paddle::platform::CPUPlace()); + for (int64_t i = 0; i < cpu_out_grad.numel(); ++i) { + cpu_out_grad.data()[i] = static_cast(i); + } + + // copy to dst out_grad + auto* place = new Place(); + DeviceContext* context = new DeviceContext(*place); + if (paddle::platform::is_cpu_place(*place)) { + out_grad = cpu_out_grad; + } else { + TensorCopySync(cpu_out_grad, *place, &out_grad); + } + + // construct in_grad + in_grad.set_lod(lod); + auto in_dims = paddle::framework::make_ddim( + {static_cast(lod[0].back()), static_cast(second_dim)}); + in_grad.mutable_data(in_dims, context->GetPlace()); + + // check tensor contruction result + PADDLE_ENFORCE_EQ(in_grad.dims().size(), out_grad.dims().size()); + for (int64_t i = 1; i < out_grad.dims().size(); ++i) { + PADDLE_ENFORCE_EQ(in_grad.dims()[i], out_grad.dims()[i]); + } + + // call functor + paddle::operators::math::SequencePoolGradFunctor()( + *context, "SUM", out_grad, &in_grad); + + if (paddle::platform::is_cpu_place(*place)) { + cpu_in_grad = in_grad; + } else { + TensorCopySync(in_grad, paddle::platform::CPUPlace(), &cpu_in_grad); + cpu_in_grad.set_lod(in_grad.lod()); + } + + EXPECT_EQ(in_grad.numel(), lod[0].back() * second_dim); + EXPECT_EQ(in_grad.lod(), lod); + + if (paddle::platform::is_cpu_place(*place)) { + for (int64_t i = 0; i < in_grad.lod()[0].size() - 1; ++i) { + int64_t begin = in_grad.lod()[0][i]; + int64_t end = in_grad.lod()[0][i + 1]; + paddle::framework::Tensor tmp = in_grad.Slice(begin, end); + for (int64_t j = 0; j != tmp.numel() / second_dim; ++j) { + for (int64_t m = 0; m != second_dim; ++m) { + EXPECT_EQ(tmp.data()[m + j * second_dim], + out_grad.data()[m + i * second_dim]); + } + } + } + } else { + for (int64_t i = 0; i < cpu_in_grad.lod()[0].size() - 1; ++i) { + int64_t begin = cpu_in_grad.lod()[0][i]; + int64_t end = cpu_in_grad.lod()[0][i + 1]; + paddle::framework::Tensor tmp = cpu_in_grad.Slice(begin, end); + for (int64_t j = 0; j != tmp.numel() / second_dim; ++j) { + for (int64_t m = 0; m != second_dim; ++m) { + EXPECT_EQ(tmp.data()[m + j * second_dim], + cpu_out_grad.data()[m + i * second_dim]); + } + } + } + } + + delete place; + delete context; +} + +TEST(SequencePoolingGrad, CPU_SUM) { + paddle::framework::LoD lod1; + lod1.push_back(std::vector{0, 10}); + TestSequencePoolingSum(lod1); + + paddle::framework::LoD lod2; + lod2.push_back(std::vector{0, 2, 7, 10}); + TestSequencePoolingSum(lod2); +} + +#ifdef PADDLE_WITH_CUDA +TEST(SequencePoolingGrad, CUDA_SUM) { + paddle::framework::LoD lod1; + lod1.push_back(std::vector{0, 10}); + TestSequencePoolingSum(lod1); + + paddle::framework::LoD lod2; + lod2.push_back(std::vector{0, 2, 7, 10}); + TestSequencePoolingSum(lod2); +} +#endif diff --git a/paddle/fluid/operators/math/softmax.cu b/paddle/fluid/operators/math/softmax.cu index 3effe776258cb541dbba32f63eda457d917011f4..ce183ed3649055aab31eb6e3f44f2224475957e9 100644 --- a/paddle/fluid/operators/math/softmax.cu +++ b/paddle/fluid/operators/math/softmax.cu @@ -96,12 +96,15 @@ template class SoftmaxCUDNNFunctor; template class SoftmaxCUDNNFunctor; template class SoftmaxGradCUDNNFunctor; template class SoftmaxGradCUDNNFunctor; +template class SoftmaxGradCUDNNFunctor; template class SoftmaxFunctor; template class SoftmaxFunctor; template class SoftmaxFunctor; template class SoftmaxGradFunctor; template class SoftmaxGradFunctor; +template class SoftmaxGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/fluid/operators/mean_op.cc b/paddle/fluid/operators/mean_op.cc index 9e0bebd17c02a3ce010b77142757b8789cfbcdd9..820636defad0be9fb2e6decefc938658ae70ea9b 100644 --- a/paddle/fluid/operators/mean_op.cc +++ b/paddle/fluid/operators/mean_op.cc @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/mean_op.h" - +#include namespace paddle { namespace operators { @@ -34,7 +34,7 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", "(Tensor) The input of mean op"); - AddOutput("Out", "(Tensor) The output of mean op").Reuse("X"); + AddOutput("Out", "(Tensor) The output of mean op"); AddComment(R"DOC( Mean Operator calculates the mean of all elements in X. @@ -42,6 +42,14 @@ Mean Operator calculates the mean of all elements in X. } }; +class MeanOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Out"}}; + } +}; + class MeanGradOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -50,6 +58,14 @@ class MeanGradOp : public framework::OperatorWithKernel { ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); ctx->ShareLoD("X", framework::GradVarName("X")); } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto input_data_type = + framework::ToDataType(ctx.Input("X")->type()); + + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } }; class MeanGradMaker : public framework::SingleGradOpDescMaker { @@ -71,7 +87,8 @@ class MeanGradMaker : public framework::SingleGradOpDescMaker { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanGradMaker); +REGISTER_OPERATOR(mean, ops::MeanOp, ops::MeanOpMaker, ops::MeanOpInferVarType, + ops::MeanGradMaker); REGISTER_OPERATOR(mean_grad, ops::MeanGradOp); REGISTER_OP_CPU_KERNEL( mean, ops::MeanKernel, diff --git a/paddle/fluid/operators/mean_op.cu b/paddle/fluid/operators/mean_op.cu index 91e0ab28efc21d4376524c8ecf66b429d51d8847..413b8ace67bd0a36849373812950834523b62216 100644 --- a/paddle/fluid/operators/mean_op.cu +++ b/paddle/fluid/operators/mean_op.cu @@ -15,11 +15,15 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/mean_op.h" +#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( mean, ops::MeanKernel, - ops::MeanKernel); + ops::MeanKernel, + ops::MeanKernel); REGISTER_OP_CUDA_KERNEL( mean_grad, ops::MeanGradKernel, - ops::MeanGradKernel); + ops::MeanGradKernel, + ops::MeanGradKernel); diff --git a/paddle/fluid/operators/mean_op.h b/paddle/fluid/operators/mean_op.h index 362e9f9ae8b2f0f77198e3f3939211ae1117b27b..360b2f68a749f630d3c7ed009c16cb51ec150581 100644 --- a/paddle/fluid/operators/mean_op.h +++ b/paddle/fluid/operators/mean_op.h @@ -55,8 +55,7 @@ class MeanGradKernel : public framework::OpKernel { IG->mutable_data(context.GetPlace()); T ig_size = static_cast(IG->numel()); - Eigen::DSizes bcast(ig_size); - + Eigen::DSizes bcast(static_cast(ig_size)); EigenVector::Flatten(*IG).device( *context.template device_context().eigen_device()) = (EigenVector::From(*OG) / ig_size).broadcast(bcast); diff --git a/paddle/fluid/operators/merge_ids_op.cc b/paddle/fluid/operators/merge_ids_op.cc index c6ec4ab047d5e91625e646fd26108d2e477cdce5..6e0e13698097ade36449f2e8ff6ab981a1b24311 100644 --- a/paddle/fluid/operators/merge_ids_op.cc +++ b/paddle/fluid/operators/merge_ids_op.cc @@ -20,13 +20,16 @@ namespace operators { class MergeIdsOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); - AddInput( - "X", - "(LoDTensors) multi input tensor with shape{batch_num, N}, N is the " - "size of embedding table") + AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}") + .AsDuplicable(); + AddInput("Rows", "(LoDTensor) the input ids with shape{row_size, 1}, ") + .AsDuplicable(); + AddInput("X", + "(LoDTensors) multi input tensor with shape{Rows, N}, N is the " + "size of embedding table") + .AsDuplicable(); + AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors.") .AsDuplicable(); - AddOutput("Out", "(LoDTensor) The merged outputs of the input tensors."); AddComment(R"DOC( Merge multi LoDTensor's into one according to Ids's shard num. @@ -79,15 +82,19 @@ class MergeIdsOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Ids"), "MergeIdsOp must has input Ids."); - PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has input X."); - PADDLE_ENFORCE(ctx->HasOutput("Out"), "MergeIdsOp must has output Out."); + PADDLE_ENFORCE(ctx->HasInputs("Ids"), + "MergeIdsOp must has multi input Ids."); + PADDLE_ENFORCE(ctx->HasInputs("Rows"), + "MergeIdsOp must has multi input Rows."); + PADDLE_ENFORCE(ctx->HasInputs("X"), "MergeIdsOp must has multi input X."); + PADDLE_ENFORCE(ctx->HasOutputs("Out"), + "MergeIdsOp must has multi output Out."); auto ids_var_type = ctx->GetInputsVarType("Ids").front(); - auto ids_dims = ctx->GetInputDim("Ids"); + auto ids_dims = ctx->GetInputsDim("Ids"); if (ids_var_type == framework::proto::VarType::LOD_TENSOR) { - PADDLE_ENFORCE_EQ(ids_dims.size(), 2); - PADDLE_ENFORCE_EQ(ids_dims[1], 1); + PADDLE_ENFORCE_EQ(ids_dims[0].size(), 2); + PADDLE_ENFORCE_EQ(ids_dims[0][1], 1); } auto x_var_type = ctx->GetInputsVarType("X"); for (auto &var_type : x_var_type) { diff --git a/paddle/fluid/operators/merge_ids_op.h b/paddle/fluid/operators/merge_ids_op.h index 83712a8519c6817151e1922c606c0fdd4682a2db..fef9e023d02f45e21ec409ad398ba7d9bdd36880 100644 --- a/paddle/fluid/operators/merge_ids_op.h +++ b/paddle/fluid/operators/merge_ids_op.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once +#include +#include #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/tensor_util.h" @@ -30,59 +32,70 @@ class MergeIdsOpKernel : public framework::OpKernel { if (!platform::is_cpu_place(place)) { PADDLE_THROW("MergeIds do not support GPU kernel"); } - VLOG(3) << "run in MergeIdsOpKernel"; - const auto *ids_var = ctx.InputVar("Ids"); - PADDLE_ENFORCE(ids_var->IsType(), - "only support to merge Ids of LoDTensor"); + const auto ids = ctx.MultiInput("Ids"); + const auto row_ids = ctx.MultiInput("Rows"); + const auto x_tensors = ctx.MultiInput("X"); + auto outs = ctx.MultiOutput("Out"); - const auto &ids_tensor = ids_var->Get(); - const auto &ids_dims = ids_tensor.dims(); - const int64_t *ids = ids_tensor.data(); + PADDLE_ENFORCE_EQ(row_ids.size(), x_tensors.size(), + "the number of Rows and X should be the same"); + PADDLE_ENFORCE_EQ(ids.size(), outs.size(), + "the number of Ids and Out should be the same"); - auto x_tensors = ctx.MultiInput("X"); + int row_ids_size = 0; + int row_size = 0; + int embedding_size = 0; - auto *out = ctx.Output("Out"); + for (int i = 0; i < x_tensors.size(); ++i) { + const auto *x_tensor = x_tensors[i]; + const auto *row_id = row_ids[i]; - int batch_size = 0; - int embedding_size = 0; - for (auto &input : x_tensors) { - if (framework::product(input->dims()) != 0) { - if (embedding_size == 0) { - embedding_size = input->dims()[1]; - } - PADDLE_ENFORCE_EQ(embedding_size, input->dims()[1], - "embedding size of all input should be the same"); - batch_size += input->dims()[0]; + if (embedding_size == 0) { + embedding_size = x_tensor->dims()[1]; } + PADDLE_ENFORCE_EQ(embedding_size, x_tensor->dims()[1], + "embedding size of all input should be the same"); + row_size += x_tensor->dims()[0]; + row_ids_size += row_id->dims()[0]; } + PADDLE_ENFORCE_EQ( - batch_size, ids_dims[0], - "the batch size of ids and merged embedding value should be the same"); + row_size, row_ids_size, + "the merged X dim[0] and merged Rows dim[0] should be the same"); + + std::unordered_map> + selected_rows_idx_map; + for (int i = 0; i < x_tensors.size(); ++i) { + const auto *row_id = row_ids[i]; + + for (int j = 0; j < row_id->numel(); ++j) { + int64_t key = row_id->data()[j]; + std::tuple val = std::make_tuple(i, j); + selected_rows_idx_map.insert(std::make_pair(key, val)); + } + } + PADDLE_ENFORCE_EQ(row_ids_size, selected_rows_idx_map.size(), + "the rows and tensor map size should be the same"); + + for (int i = 0; i < outs.size(); ++i) { + auto *out_ids = ids[i]; + auto *out = outs[i]; - const size_t shard_num = x_tensors.size(); + out->set_lod(out_ids->lod()); - if (shard_num == 1) { - VLOG(3) << "only one shard, we can copy the data directly"; - TensorCopy(*x_tensors[0], place, out); - } else { - std::vector in_indexs(shard_num, 0); + int nums = static_cast(out_ids->dims()[0]); auto *out_data = out->mutable_data( - framework::make_ddim({batch_size, embedding_size}), place); - // copy data from ins[shard_num] to out. - for (int i = 0; i < ids_dims[0]; ++i) { - int64_t id = ids[i]; - size_t shard_id = static_cast(id) % shard_num; - int index = in_indexs[shard_id]; - memcpy(out_data + embedding_size * i, - x_tensors[shard_id]->data() + index * embedding_size, + framework::make_ddim({nums, embedding_size}), place); + for (int j = 0; j < nums; ++j) { + int id = out_ids->data()[j]; + auto row_tuple = selected_rows_idx_map[id]; + int64_t row_idx = std::get<1>(row_tuple); + const auto *x_tensor = x_tensors[std::get<0>(row_tuple)]; + + memcpy(out_data + embedding_size * j, + x_tensor->data() + row_idx * embedding_size, sizeof(T) * embedding_size); - in_indexs[shard_id] += 1; - } - - for (size_t i = 0; i < shard_num; ++i) { - PADDLE_ENFORCE_EQ(in_indexs[i], x_tensors[i]->dims()[0], - "after merge, all data in x_tensor should be used"); } } } diff --git a/paddle/fluid/operators/momentum_op.cc b/paddle/fluid/operators/momentum_op.cc index 5f43c5810812260c4384349bdb709716c9a182f5..7f0b51580aa2591ac7338ad7c29ee4756d909925 100644 --- a/paddle/fluid/operators/momentum_op.cc +++ b/paddle/fluid/operators/momentum_op.cc @@ -19,44 +19,25 @@ namespace operators { using Tensor = framework::Tensor; -class MomentumOp : public framework::OperatorWithKernel { +class MomentumOpInferVarType : public framework::VarTypeInference { public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Param"), - "Input(param) of Momentum should not be null."); - PADDLE_ENFORCE(ctx->HasInput("Grad"), - "Input(grad) of Momentum should not be null."); - PADDLE_ENFORCE(ctx->HasInput("Velocity"), - "Input(velocity) of Momentum should not be null."); - PADDLE_ENFORCE(ctx->HasInput("LearningRate"), - "Input(LearningRate) of Momentum should not be null."); - - PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), - "Output(ParamOut) of Momentum should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("VelocityOut"), - "Output(VelocityOut) of Momentum should not be null."); - - auto param_dim = ctx->GetInputDim("Param"); - PADDLE_ENFORCE_EQ( - param_dim, ctx->GetInputDim("Grad"), - "Param and Grad input of MomentumOp should have the same dimension."); - PADDLE_ENFORCE_EQ( - param_dim, ctx->GetInputDim("Velocity"), - "Param and Velocity of MomentumOp should have the same dimension."); - PADDLE_ENFORCE_EQ(framework::product(ctx->GetInputDim("LearningRate")), 1, - "Learning_rate should be a scalar"); - - ctx->SetOutputDim("ParamOut", param_dim); - ctx->SetOutputDim("VelocityOut", param_dim); - } - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext &ctx) const override { - auto input_data_type = - framework::ToDataType(ctx.Input("Param")->type()); - return framework::OpKernelType(input_data_type, ctx.GetPlace()); + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto input_var = op_desc.Input("Param")[0]; + for (auto& out_var : op_desc.Output("ParamOut")) { + if (block->FindRecursiveOrCreateVar(input_var).GetType() == + framework::proto::VarType::SELECTED_ROWS) { + block->FindRecursiveOrCreateVar(out_var).SetType( + framework::proto::VarType::SELECTED_ROWS); + } else if (block->FindRecursiveOrCreateVar(input_var).GetType() == + framework::proto::VarType::LOD_TENSOR) { + block->FindRecursiveOrCreateVar(out_var).SetType( + framework::proto::VarType::LOD_TENSOR); + } else { + PADDLE_THROW( + "Only support LodTensor and SelectedRows, Unexpected Input Type."); + } + } } }; @@ -110,6 +91,9 @@ $$ } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(momentum, ops::MomentumOp, ops::MomentumOpMaker); -REGISTER_OP_CPU_KERNEL(momentum, ops::MomentumOpKernel, - ops::MomentumOpKernel); +REGISTER_OPERATOR(momentum, ops::MomentumOp, ops::MomentumOpMaker, + paddle::framework::EmptyGradOpMaker, + ops::MomentumOpInferVarType); +REGISTER_OP_CPU_KERNEL( + momentum, ops::MomentumOpKernel, + ops::MomentumOpKernel); diff --git a/paddle/fluid/operators/momentum_op.cu b/paddle/fluid/operators/momentum_op.cu index a3932db1f3a50305d585cd3d5e86fa1b527df78b..b68fec34d43f0dee834f1045f192d5c6089d9356 100644 --- a/paddle/fluid/operators/momentum_op.cu +++ b/paddle/fluid/operators/momentum_op.cu @@ -15,65 +15,7 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/momentum_op.h" -namespace paddle { -namespace operators { - -template -__global__ void MomentumKernel(const T* p, const T* g, const T* v, - const T* learning_rate, const T mu, - const int64_t num, bool use_nesterov, T* p_out, - T* v_out) { - T lr = learning_rate[0]; - if (use_nesterov) { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; - i += blockDim.x * gridDim.x) { - T g_val = g[i]; - T v_new = v[i] * mu + g_val; - v_out[i] = v_new; - p_out[i] = p[i] - (g_val + v_new * mu) * lr; - } - } else { - for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; - i += blockDim.x * gridDim.x) { - T v_new = v[i] * mu + g[i]; - v_out[i] = v_new; - p_out[i] = p[i] - lr * v_new; - } - } -} - -template -class MomentumOpCUDAKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto param_out = ctx.Output("ParamOut"); - auto velocity_out = ctx.Output("VelocityOut"); - auto param = ctx.Input("Param"); - auto velocity = ctx.Input("Velocity"); - auto grad = ctx.Input("Grad"); - auto learning_rate = ctx.Input("LearningRate"); - - T* p_out = param_out->mutable_data(ctx.GetPlace()); - T* v_out = velocity_out->mutable_data(ctx.GetPlace()); - - T mu = static_cast(ctx.Attr("mu")); - bool use_nesterov = ctx.Attr("use_nesterov"); - - auto* p = param->data(); - auto* v = velocity->data(); - auto* g = grad->data(); - auto* lr = learning_rate->data(); - - int block = 512; - int grid = (param->numel() + block - 1) / block; - MomentumKernel<<>>( - p, g, v, lr, mu, param->numel(), use_nesterov, p_out, v_out); - } -}; - -} // namespace operators -} // namespace paddle - namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL(momentum, ops::MomentumOpCUDAKernel, - ops::MomentumOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL( + momentum, ops::MomentumOpKernel, + ops::MomentumOpKernel); diff --git a/paddle/fluid/operators/momentum_op.h b/paddle/fluid/operators/momentum_op.h index 264726040fb566a52b8c0cdee0a1524197d2a675..e5b756b4fa637f2d4136f8c8a87bf34c6c04413a 100644 --- a/paddle/fluid/operators/momentum_op.h +++ b/paddle/fluid/operators/momentum_op.h @@ -13,29 +13,96 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/algorithm.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { -template -class MomentumOpKernel : public framework::OpKernel { +using framework::Tensor; +using framework::SelectedRows; +struct NoNesterov; +struct UseNesterov; + +class MomentumOp : public framework::OperatorWithKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto param_out = ctx.Output("ParamOut"); - auto velocity_out = ctx.Output("VelocityOut"); - auto param = ctx.Input("Param"); - auto velocity = ctx.Input("Velocity"); - auto grad = ctx.Input("Grad"); - auto learning_rate = ctx.Input("LearningRate"); + using framework::OperatorWithKernel::OperatorWithKernel; - param_out->mutable_data(ctx.GetPlace()); - velocity_out->mutable_data(ctx.GetPlace()); + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Param"), + "Input(param) of Momentum should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Grad"), + "Input(grad) of Momentum should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Velocity"), + "Input(velocity) of Momentum should not be null."); + PADDLE_ENFORCE(ctx->HasInput("LearningRate"), + "Input(LearningRate) of Momentum should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); - T mu = static_cast(ctx.Attr("mu")); - bool use_nesterov = ctx.Attr("use_nesterov"); + PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), + "Output(ParamOut) of Momentum should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("VelocityOut"), + "Output(VelocityOut) of Momentum should not be null."); + auto param_dim = ctx->GetInputDim("Param"); + if (ctx->GetInputsVarType("Grad")[0] == + framework::proto::VarType::LOD_TENSOR) { + PADDLE_ENFORCE_EQ( + param_dim, ctx->GetInputDim("Grad"), + "Param and Grad input of MomentumOp should have the same dimension."); + PADDLE_ENFORCE_EQ( + param_dim, ctx->GetInputDim("Velocity"), + "Param and Velocity of MomentumOp should have the same dimension."); + } + PADDLE_ENFORCE_EQ(framework::product(ctx->GetInputDim("LearningRate")), 1, + "Learning_rate should be a scalar"); + + ctx->SetOutputDim("ParamOut", param_dim); + ctx->SetOutputDim("VelocityOut", param_dim); + } + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto input_data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param")); + return framework::OpKernelType(input_data_type, ctx.GetPlace()); + } +}; + +template +class CPUDenseMomentumFunctor { + private: + const Tensor* param; + const Tensor* grad; + const Tensor* velocity; + const Tensor* learning_rate; + const T mu; + const T use_nesterov; + Tensor* param_out; + Tensor* velocity_out; + + public: + CPUDenseMomentumFunctor(const Tensor* param, const Tensor* grad, + const Tensor* velocity, const Tensor* learning_rate, + const T mu, const bool use_nesterov, + Tensor* param_out, Tensor* velocity_out) + : param(param), + grad(grad), + velocity(velocity), + learning_rate(learning_rate), + mu(mu), + use_nesterov(use_nesterov), + param_out(param_out), + velocity_out(velocity_out) {} + + inline void operator()() { auto p_out = framework::EigenVector::Flatten(*param_out); auto v_out = framework::EigenVector::Flatten(*velocity_out); @@ -53,5 +120,283 @@ class MomentumOpKernel : public framework::OpKernel { } }; +template +class DenseMomentumFunctor; + +// NOTE(dzh) for performance. +// avoid if/else in inside kernel, implement GPU UseNesterov/NoNesterov as two +// functor. +template +class DenseMomentumFunctor { + private: + const T* p_; + const T* g_; + const T* v_; + const T* lr_; + const T mu_; + const int64_t num_; + T* p_out_; + T* v_out_; + + public: + DenseMomentumFunctor(const T* p, const T* g, const T* v, + const T* learning_rate, const T mu, const int64_t num, + T* p_out, T* v_out) + : p_(p), + g_(g), + v_(v), + lr_(learning_rate), + mu_(mu), + num_(num), + p_out_(p_out), + v_out_(v_out) {} + inline HOSTDEVICE void operator()(size_t i) const { + // put memory access in register + const T p = p_[i]; + const T g = g_[i]; + const T lr = lr_[0]; + const T v = v_[i]; + T v_out = v * mu_ + g; + T p_out = p - (g + v_out * mu_) * lr; + // write reigster to memory + v_out_[i] = v_out; + p_out_[i] = p_out; + } +}; + +template +class DenseMomentumFunctor { + private: + const T* p_; + const T* g_; + const T* v_; + const T* lr_; + const T mu_; + const int64_t num_; + T* p_out_; + T* v_out_; + + public: + DenseMomentumFunctor(const T* p, const T* g, const T* v, + const T* learning_rate, const T mu, const int64_t num, + T* p_out, T* v_out) + : p_(p), + g_(g), + v_(v), + lr_(learning_rate), + mu_(mu), + num_(num), + p_out_(p_out), + v_out_(v_out) {} + inline HOSTDEVICE void operator()(size_t i) const { + // put memory access in register + const T p = p_[i]; + const T g = g_[i]; + const T lr = lr_[0]; + const T v = v_[i]; + T v_out = v * mu_ + g; + T p_out = p - lr * v_out; + // write reigster to memory + v_out_[i] = v_out; + p_out_[i] = p_out; + } +}; + +template +class SparseMomentumFunctor; + +template +class SparseMomentumFunctor { + private: + const T* p_; + const T* g_; + const T* v_; + const T* lr_; + const T mu_; + const int64_t* rows_; + const int64_t row_numel_; + const int64_t row_height_; + T* p_out_; + T* v_out_; + + public: + SparseMomentumFunctor(const T* p, const T* g, const T* v, const T* lr, + const T mu, const int64_t* rows, int64_t row_numel, + int64_t row_height, T* p_out, T* v_out) + : p_(p), + g_(g), + v_(v), + lr_(lr), + mu_(mu), + rows_(rows), + row_numel_(row_numel), + row_height_(row_height), + p_out_(p_out), + v_out_(v_out) {} + + inline HOSTDEVICE void operator()(size_t i) { + auto row_idx = + math::BinarySearch(rows_, row_height_, i / row_numel_); + T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0; + // put memory access in register + const T p = p_[i]; + const T lr = lr_[0]; + const T v = v_[i]; + T v_out = v * mu_ + g; + T p_out = p - (g + v_out * mu_) * lr; + // write reigster to memory + v_out_[i] = v_out; + p_out_[i] = p_out; + } +}; + +template +class SparseMomentumFunctor { + private: + const T* p_; + const T* g_; + const T* v_; + const T* lr_; + const T mu_; + const int64_t* rows_; + const int64_t row_numel_; + const int64_t row_height_; + T* p_out_; + T* v_out_; + + public: + SparseMomentumFunctor(const T* p, const T* g, const T* v, const T* lr, + const T mu, const int64_t* rows, int64_t row_numel, + int64_t row_height, T* p_out, T* v_out) + : p_(p), + g_(g), + v_(v), + lr_(lr), + mu_(mu), + rows_(rows), + row_numel_(row_numel), + row_height_(row_height), + p_out_(p_out), + v_out_(v_out) {} + + inline HOSTDEVICE void operator()(size_t i) { + auto row_idx = + math::BinarySearch(rows_, row_height_, i / row_numel_); + T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0; + // put memory access in register + const T p = p_[i]; + const T lr = lr_[0]; + const T v = v_[i]; + T v_out = v * mu_ + g; + T p_out = p - v_out * lr; + // write reigster to memory + v_out_[i] = v_out; + p_out_[i] = p_out; + } +}; + +template +class MomentumOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + T mu = static_cast(ctx.Attr("mu")); + bool use_nesterov = ctx.Attr("use_nesterov"); + + auto learning_rate = ctx.Input("LearningRate"); + auto param = ctx.Input("Param"); + auto param_out = ctx.Output("ParamOut"); + auto* velocity = ctx.Input("Velocity"); + auto velocity_out = ctx.Output("VelocityOut"); + param_out->mutable_data(ctx.GetPlace()); + velocity_out->mutable_data(ctx.GetPlace()); + + auto* grad_var = ctx.InputVar("Grad"); + if (grad_var->IsType()) { + auto grad = ctx.Input("Grad"); + if (platform::is_cpu_place(ctx.GetPlace())) { + CPUDenseMomentumFunctor functor(param, grad, velocity, learning_rate, + mu, use_nesterov, param_out, + velocity_out); + functor(); + } else if (platform::is_gpu_place(ctx.GetPlace())) { + platform::ForRange for_range( + static_cast(ctx.device_context()), + param->numel()); + if (use_nesterov) { + DenseMomentumFunctor functor( + param->data(), grad->data(), velocity->data(), + learning_rate->data(), mu, param->numel(), + param_out->mutable_data(ctx.GetPlace()), + velocity_out->mutable_data(ctx.GetPlace())); + for_range(functor); + + } else { + DenseMomentumFunctor functor( + param->data(), grad->data(), velocity->data(), + learning_rate->data(), mu, param->numel(), + param_out->mutable_data(ctx.GetPlace()), + velocity_out->mutable_data(ctx.GetPlace())); + for_range(functor); + } + } + + } else if (grad_var->IsType()) { + // sparse update embedding with selectedrows + auto grad = ctx.Input("Grad"); + + // sparse update maybe empty. + if (grad->rows().size() == 0) { + VLOG(30) << "Grad SelectedRows contains no data!"; + return; + } + auto* merged_grad = const_cast(ctx.scope()) + .Var() + ->GetMutable(); + math::scatter::MergeAdd merge_func; + merge_func(ctx.template device_context(), *grad, + merged_grad); + + const int64_t* rows = nullptr; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + rows = merged_grad->rows().CUDAData(ctx.GetPlace()); + } else { +#endif + rows = merged_grad->rows().data(); +#ifdef PADDLE_WITH_CUDA + } +#endif + int64_t row_numel = + merged_grad->value().numel() / merged_grad->rows().size(); + platform::ForRange for_range( + static_cast(ctx.device_context()), + param->numel()); + if (use_nesterov) { + SparseMomentumFunctor functor( + param->data(), merged_grad->value().data(), + velocity->data(), learning_rate->data(), mu, rows, row_numel, + static_cast(merged_grad->rows().size()), + param_out->mutable_data(ctx.GetPlace()), + velocity_out->mutable_data(ctx.GetPlace())); + for_range(functor); + + } else { + SparseMomentumFunctor functor( + param->data(), merged_grad->value().data(), + velocity->data(), learning_rate->data(), mu, rows, row_numel, + static_cast(merged_grad->rows().size()), + param_out->mutable_data(ctx.GetPlace()), + velocity_out->mutable_data(ctx.GetPlace())); + for_range(functor); + } + } else { + PADDLE_THROW( + string::Sprintf("MomentumOp only supports LoDTensor or SelectedRows " + "gradient, but the received Variable Type is %s", + grad_var->Type().name())); + } + } +}; + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/mul_op.cc b/paddle/fluid/operators/mul_op.cc index 363abfb0e0c96e8a4d82124dff168f28e339a9ae..08f2949d4a3774894912ae5251806b46e6240702 100644 --- a/paddle/fluid/operators/mul_op.cc +++ b/paddle/fluid/operators/mul_op.cc @@ -38,9 +38,9 @@ class MulOp : public framework::OperatorWithKernel { int x_num_col_dims = ctx->Attrs().Get("x_num_col_dims"); int y_num_col_dims = ctx->Attrs().Get("y_num_col_dims"); - VLOG(3) << "mul operator x.shape=" << x_dims << " y.shape=" << y_dims - << " x_num_col_dims=" << x_num_col_dims - << " y_num_col_dims=" << y_num_col_dims; + VLOG(30) << "mul operator x.shape=" << x_dims << " y.shape=" << y_dims + << " x_num_col_dims=" << x_num_col_dims + << " y_num_col_dims=" << y_num_col_dims; PADDLE_ENFORCE_GT( x_dims.size(), x_num_col_dims, @@ -126,6 +126,14 @@ or not. But the output only shares the LoD information with input $X$. } }; +class MulOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Out"}}; + } +}; + class MulGradOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -178,7 +186,8 @@ class MulOpGradMaker : public framework::SingleGradOpDescMaker { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(mul, ops::MulOp, ops::MulOpMaker, ops::MulOpGradMaker); +REGISTER_OPERATOR(mul, ops::MulOp, ops::MulOpMaker, ops::MulOpInferVarType, + ops::MulOpGradMaker); REGISTER_OPERATOR(mul_grad, ops::MulGradOp); REGISTER_OP_CPU_KERNEL( mul, ops::MulKernel, diff --git a/paddle/fluid/operators/mul_op.cu.cc b/paddle/fluid/operators/mul_op.cu.cc index 81f3e42bf412fa4d2cb48405f2f8ee49b6aa0b67..6c5a83c6a50c463502171f09bbf18e17e43917b5 100644 --- a/paddle/fluid/operators/mul_op.cu.cc +++ b/paddle/fluid/operators/mul_op.cu.cc @@ -20,6 +20,7 @@ namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel, ops::MulKernel, ops::MulKernel); -REGISTER_OP_CUDA_KERNEL(mul_grad, - ops::MulGradKernel, - ops::MulGradKernel); +REGISTER_OP_CUDA_KERNEL( + mul_grad, ops::MulGradKernel, + ops::MulGradKernel, + ops::MulGradKernel); diff --git a/paddle/fluid/operators/nccl_op.cu.cc b/paddle/fluid/operators/nccl_op.cu.cc index 8de974bc2b333fb6ccc5b5f0bb1af86533139925..9db0031a6934537a7d991b775ecac688ae6b66e9 100644 --- a/paddle/fluid/operators/nccl_op.cu.cc +++ b/paddle/fluid/operators/nccl_op.cu.cc @@ -63,16 +63,16 @@ class NCCLAllReduceKernel : public framework::OpKernel { // device id int gpu_id = boost::get(ctx.GetPlace()).GetDeviceId(); int idx = comm->GetCommId(gpu_id); - VLOG(3) << "gpu : " - << " invoke allreduce. send " << x->numel() << " recv " - << out->numel(); + VLOG(30) << "gpu : " + << " invoke allreduce. send " << x->numel() << " recv " + << out->numel(); PADDLE_ENFORCE(platform::dynload::ncclAllReduce( x->data(), out->mutable_data(ctx.GetPlace()), out->numel(), NCCLTypeWrapper::type, reduction_op_, comm->comms().at(idx), ctx.cuda_device_context().stream())); - VLOG(3) << "gpu : " - << " finished allreduce. send " << x->numel() << " recv " - << out->numel(); + VLOG(30) << "gpu : " + << " finished allreduce. send " << x->numel() << " recv " + << out->numel(); } }; @@ -109,14 +109,14 @@ class NCCLReduceKernel : public framework::OpKernel { } else { out->Resize(framework::make_ddim({0})); } - VLOG(3) << "gpu : " << gpu_id << " invoke reduce. send " << x->numel() - << " recv " << out->numel(); + VLOG(30) << "gpu : " << gpu_id << " invoke reduce. send " << x->numel() + << " recv " << out->numel(); PADDLE_ENFORCE(platform::dynload::ncclReduce( x->data(), recvbuffer, x->numel(), NCCLTypeWrapper::type, reduction_op_, root, comm->comms().at(idx), ctx.cuda_device_context().stream())); - VLOG(3) << "gpu : " << gpu_id << " finished reduce. send " << x->numel() - << " recv " << out->numel(); + VLOG(30) << "gpu : " << gpu_id << " finished reduce. send " << x->numel() + << " recv " << out->numel(); } }; @@ -133,21 +133,22 @@ class NCCLBcastKernel : public framework::OpKernel { int idx = comm->GetCommId(gpu_id); if (idx == root) { auto* x = ctx.Input("X"); - VLOG(3) << "gpu : " << gpu_id << " invoke Bcast. send " << x->numel(); + VLOG(30) << "gpu : " << gpu_id << " invoke Bcast. send " << x->numel(); PADDLE_ENFORCE(platform::dynload::ncclBcast( reinterpret_cast(const_cast(x->data())), x->numel(), NCCLTypeWrapper::type, root, comm->comms().at(idx), ctx.cuda_device_context().stream())); - VLOG(3) << "gpu : " << gpu_id << " finished Bcast."; + VLOG(30) << "gpu : " << gpu_id << " finished Bcast."; } else { auto* out = ctx.Output("Out"); - VLOG(3) << "gpu : " << gpu_id << " invoke Bcast. recv buffer " - << framework::product(out->dims()); + VLOG(30) << "gpu : " << gpu_id << " invoke Bcast. recv buffer " + << framework::product(out->dims()); PADDLE_ENFORCE(platform::dynload::ncclBcast( out->mutable_data(ctx.GetPlace()), out->numel(), NCCLTypeWrapper::type, root, comm->comms().at(idx), ctx.cuda_device_context().stream())); - VLOG(3) << "gpu : " << gpu_id << " finished Bcast. recv " << out->numel(); + VLOG(30) << "gpu : " << gpu_id << " finished Bcast. recv " + << out->numel(); } } }; diff --git a/paddle/fluid/operators/nccl_op_test.cu.cc b/paddle/fluid/operators/nccl_op_test.cu.cc index d5fb7a12e5d9757f3e639f6de7f0129bd531e2a1..f48ccdd97fa5adb475013cf26e7544c2729b4457 100644 --- a/paddle/fluid/operators/nccl_op_test.cu.cc +++ b/paddle/fluid/operators/nccl_op_test.cu.cc @@ -86,9 +86,9 @@ class NCCLTester : public ::testing::Test { (*p_scopes).resize(gpu_list_.size()); auto op = f::OpRegistry::CreateOp(*op1); - VLOG(1) << "invoke NCCLInitOp."; + VLOG(10) << "invoke NCCLInitOp."; op->Run(g_scope_, cpu_place); - VLOG(1) << "NCCLInitOp finished."; + VLOG(10) << "NCCLInitOp finished."; } int GetGPUData(int gpu_id) { return gpu_id + 42; } @@ -109,7 +109,7 @@ class NCCLTester : public ::testing::Test { std::vector send_vector(f::product(kDims), GetGPUData(gpu_id)); paddle::framework::TensorFromVector(send_vector, *ctx, send_tensor); - VLOG(1) << "Send Tensor filled with elements " << send_tensor->numel(); + VLOG(10) << "Send Tensor filled with elements " << send_tensor->numel(); } lk.unlock(); @@ -119,11 +119,11 @@ class NCCLTester : public ::testing::Test { auto op = f::OpRegistry::CreateOp(*op1); - VLOG(1) << "Device : " << gpu_id << " invoke " << op_desc.Type(); - VLOG(1) << " send_tensor : " << send_tensor->numel() - << " recv_tensor : " << recv_tensor->numel(); + VLOG(10) << "Device : " << gpu_id << " invoke " << op_desc.Type(); + VLOG(10) << " send_tensor : " << send_tensor->numel() + << " recv_tensor : " << recv_tensor->numel(); op->Run(*scope, place); - VLOG(1) << "Device : " << gpu_id << " finished " << op_desc.Type(); + VLOG(10) << "Device : " << gpu_id << " finished " << op_desc.Type(); } public: diff --git a/paddle/fluid/operators/nce_op.cc b/paddle/fluid/operators/nce_op.cc index e471f04662a1fa3e8e77a2db37f0da4521682018..877c9a0528441a7d5b1306c3f8f8be1a5aea577a 100644 --- a/paddle/fluid/operators/nce_op.cc +++ b/paddle/fluid/operators/nce_op.cc @@ -69,7 +69,7 @@ class NCEOp : public framework::OperatorWithKernel { const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), - ctx.GetPlace()); + platform::CPUPlace()); } }; @@ -174,7 +174,7 @@ class NCEOpGrad : public framework::OperatorWithKernel { const framework::ExecutionContext& ctx) const override { return framework::OpKernelType( framework::ToDataType(ctx.Input("Input")->type()), - ctx.GetPlace()); + platform::CPUPlace()); } }; diff --git a/paddle/fluid/operators/parallel_do_op.cc b/paddle/fluid/operators/parallel_do_op.cc index 97c36a83fc5eff421725d05f66fca05f5169d1bb..c795d4bdd10c0ffbf30a4849fc773335036e34c2 100644 --- a/paddle/fluid/operators/parallel_do_op.cc +++ b/paddle/fluid/operators/parallel_do_op.cc @@ -48,7 +48,7 @@ static void SplitTensorAndMoveTensorToScopes( auto lod_tensors = tensor.SplitLoDTensor(places); for (auto &lod : lod_tensors) { - VLOG(3) << lod.dims(); + VLOG(30) << lod.dims(); } if (num_sub_scopes == 0) { num_sub_scopes = lod_tensors.size(); @@ -263,7 +263,7 @@ class ParallelDoGradOp : public framework::OperatorBase { if (s == framework::kEmptyVarName) { continue; } - VLOG(3) << "Moving " << s; + VLOG(30) << "Moving " << s; CopyOrShare(*sub_scopes[0]->FindVar(s), place, scope.FindVar(s)); } WaitOnPlaces(places); @@ -277,7 +277,7 @@ class ParallelDoGradOp : public framework::OperatorBase { if (s == framework::kEmptyVarName) { continue; } - VLOG(3) << "Accumulating " << s; + VLOG(30) << "Accumulating " << s; if (s == framework::kEmptyVarName) continue; std::string tmp_name; auto *tmp = sub_scopes[0]->Var(&tmp_name); @@ -289,7 +289,7 @@ class ParallelDoGradOp : public framework::OperatorBase { auto sum_op = framework::OpRegistry::CreateOp( "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}}, framework::AttributeMap{{"use_mkldnn", {false}}}); - VLOG(10) << sum_op->DebugStringEx(sub_scopes[0]); + VLOG(100) << sum_op->DebugStringEx(sub_scopes[0]); sum_op->Run(*sub_scopes[0], places[0]); WaitOnPlace(places[0]); } @@ -316,7 +316,7 @@ class ParallelDoGradOpDescMaker : public framework::SingleGradOpDescMaker { auto *grad = new framework::OpDesc(); grad->SetType("parallel_do_grad"); for (auto &input_param : this->InputNames()) { - VLOG(3) << input_param; + VLOG(30) << input_param; grad->SetInput(input_param, this->Input(input_param)); if (input_param != kPlaces) { grad->SetOutput(framework::GradVarName(input_param), @@ -397,6 +397,24 @@ class ParallelDoGradOpShapeInference : public framework::InferShapeBase { } }; +class ParallelDoGradOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + framework::BlockDesc *sub_block = + boost::get(op_desc.GetAttr(kParallelBlock)); + for (auto &out_vars : op_desc.Outputs()) { + for (auto &out_var : out_vars.second) { + auto &var = block->FindRecursiveOrCreateVar(out_var); + auto sub_var = sub_block->FindRecursiveOrCreateVar(out_var); + if (sub_var.GetType() != var.GetType()) { + var.SetType(sub_var.GetType()); + } + } + } + } +}; + } // namespace operators } // namespace paddle @@ -404,4 +422,5 @@ REGISTER_OPERATOR(parallel_do, paddle::operators::ParallelDoOp, paddle::operators::ParallelDoOpProtoMaker, paddle::operators::ParallelDoGradOpDescMaker); REGISTER_OPERATOR(parallel_do_grad, paddle::operators::ParallelDoGradOp, - paddle::operators::ParallelDoGradOpShapeInference); + paddle::operators::ParallelDoGradOpShapeInference, + paddle::operators::ParallelDoGradOpVarTypeInference); diff --git a/paddle/fluid/operators/pool_cudnn_op.cu.cc b/paddle/fluid/operators/pool_cudnn_op.cu.cc index 31f083565fddee66aea1485ed71f41b6199f4502..4a332ce10b59b21d2518684237ce0bbf1bbfa75a 100644 --- a/paddle/fluid/operators/pool_cudnn_op.cu.cc +++ b/paddle/fluid/operators/pool_cudnn_op.cu.cc @@ -41,6 +41,7 @@ class PoolCUDNNOpKernel : public framework::OpKernel { T *output_data = output->mutable_data(ctx.GetPlace()); std::string pooling_type = ctx.Attr("pooling_type"); + bool exclusive = ctx.Attr("exclusive"); std::vector ksize = ctx.Attr>("ksize"); std::vector strides = ctx.Attr>("strides"); std::vector paddings = ctx.Attr>("paddings"); @@ -72,7 +73,8 @@ class PoolCUDNNOpKernel : public framework::OpKernel { if (pooling_type == "max") { pooling_mode = PoolingMode::kMaximum; } else { - pooling_mode = PoolingMode::kAverage; + pooling_mode = exclusive ? PoolingMode::kAverageExclusive + : PoolingMode::kAverageInclusive; } cudnnPoolingDescriptor_t cudnn_pool_desc = @@ -101,6 +103,7 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { Tensor *input_grad = ctx.Output(framework::GradVarName("X")); std::string pooling_type = ctx.Attr("pooling_type"); + bool exclusive = ctx.Attr("exclusive"); std::vector ksize = ctx.Attr>("ksize"); std::vector strides = ctx.Attr>("strides"); std::vector paddings = ctx.Attr>("paddings"); @@ -141,7 +144,8 @@ class PoolCUDNNGradOpKernel : public framework::OpKernel { pooling_mode = PoolingMode::kMaximum; } } else { - pooling_mode = PoolingMode::kAverage; + pooling_mode = exclusive ? PoolingMode::kAverageExclusive + : PoolingMode::kAverageInclusive; } cudnnPoolingDescriptor_t cudnn_pool_desc = @@ -174,7 +178,8 @@ REGISTER_OP_KERNEL(pool2d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel); REGISTER_OP_KERNEL(pool2d_grad, CUDNN, plat::CUDAPlace, ops::PoolCUDNNGradOpKernel, - ops::PoolCUDNNGradOpKernel); + ops::PoolCUDNNGradOpKernel, + ops::PoolCUDNNGradOpKernel); REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel, diff --git a/paddle/fluid/operators/pool_op.cc b/paddle/fluid/operators/pool_op.cc index f8ad63690e84339da0390d4ddd2db45f25db385a..46a95350a7293c18313811ba9b367fd65955145a 100644 --- a/paddle/fluid/operators/pool_op.cc +++ b/paddle/fluid/operators/pool_op.cc @@ -40,7 +40,7 @@ int PoolOutputSize(int input_size, int filter_size, int padding, int stride, return output_size; } -void PoolOp::InferShape(framework::InferShapeContext *ctx) const { +void PoolOp::InferShape(framework::InferShapeContext* ctx) const { PADDLE_ENFORCE(ctx->HasInput("X"), "X(Input) of Pooling should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Out(Output) of Pooling should not be null."); @@ -81,7 +81,7 @@ void PoolOp::InferShape(framework::InferShapeContext *ctx) const { } framework::OpKernelType PoolOp::GetExpectedKernelType( - const framework::ExecutionContext &ctx) const { + const framework::ExecutionContext& ctx) const { framework::LibraryType library_{framework::LibraryType::kPlain}; std::string data_format = ctx.Attr("data_format"); framework::DataLayout layout_ = framework::StringToDataLayout(data_format); @@ -104,7 +104,7 @@ framework::OpKernelType PoolOp::GetExpectedKernelType( layout_, library_); } -void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const { +void PoolOpGrad::InferShape(framework::InferShapeContext* ctx) const { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null."); PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")), "Input(X@GRAD) should not be null."); @@ -112,7 +112,7 @@ void PoolOpGrad::InferShape(framework::InferShapeContext *ctx) const { } framework::OpKernelType PoolOpGrad::GetExpectedKernelType( - const framework::ExecutionContext &ctx) const { + const framework::ExecutionContext& ctx) const { framework::LibraryType library_{framework::LibraryType::kPlain}; std::string data_format = ctx.Attr("data_format"); framework::DataLayout layout_ = framework::StringToDataLayout(data_format); @@ -151,8 +151,7 @@ void Pool2dOpMaker::Make() { "The format of output tensor is also NCHW, " "where N is batch size, C is the number of channels, " "H is the height of the feature, " - "and W is the width of the feature.") - .Reuse("X"); + "and W is the width of the feature."); AddAttr("pooling_type", "(string), pooling type, can be \"max\" for max-pooling " @@ -181,6 +180,12 @@ void Pool2dOpMaker::Make() { "operator." "If global_pooling = true, paddings and ksize will be ignored.") .SetDefault({0, 0}); + AddAttr( + "exclusive", + "(bool, default True) When true, will exclude the zero-padding in the " + "averaging calculating, otherwise, include the zero-padding. Note, it " + "is only used when pooling_type is avg. The defalut is True.") + .SetDefault(true); AddAttr( "use_cudnn", "(bool, default false) Only used in cudnn kernel, need install cudnn") @@ -237,9 +242,34 @@ Example: W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1 $$ + For exclusive = true: + $$ + hstart = i * strides[0] - paddings[0] + hend = hstart + ksize[0] + wstart = j * strides[1] - paddings[1] + wend = wstart + ksize[1] + Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{ksize[0] * ksize[1]} + $$ + For exclusive = false: + $$ + hstart = max(0, i * strides[0] - paddings[0]) + hend = min(H, hstart + ksize[0]) + wstart = max(0, j * strides[1] - paddings[1]) + wend = min(W, wstart + ksize[1]) + Output(i ,j) = \\frac{sum(Input[hstart:hend, wstart:wend])}{(hend - hstart) * (wend - wstart)} + $$ + )DOC"); } +class PoolOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Out"}}; + } +}; + void Pool3dOpMaker::Make() { AddInput("X", "(Tensor) The input tensor of pooling operator. " @@ -252,8 +282,7 @@ void Pool3dOpMaker::Make() { "The format of output tensor is also NCDHW, " "where N is batch size, C is " "the number of channels, and D, H and W is the depth, height and " - "width of the feature, respectively.") - .Reuse("X"); + "width of the feature, respectively."); AddAttr("pooling_type", "(string) Pooling type, can be \"max\" for max-pooling " @@ -285,6 +314,12 @@ void Pool3dOpMaker::Make() { "If global_pooling = true, ksize and paddings will be ignored.") .SetDefault({0, 0, 0}); // TODO(Chengduo): Add checker. (Currently, // TypedAttrChecker don't support vector type.) + AddAttr( + "exclusive", + "(bool, default True) When true, will exclude the zero-padding in the " + "averaging calculating, otherwise, include the zero-padding. Note, it " + "is only used when pooling_type is avg. The defalut is True.") + .SetDefault(true); AddAttr( "use_cudnn", @@ -345,6 +380,7 @@ Example: namespace ops = paddle::operators; REGISTER_OPERATOR(pool2d, ops::PoolOp, ops::Pool2dOpMaker, + ops::PoolOpInferVarType, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(pool2d_grad, ops::PoolOpGrad); @@ -356,6 +392,7 @@ REGISTER_OP_CPU_KERNEL( ops::PoolGradKernel); REGISTER_OPERATOR(pool3d, ops::PoolOp, ops::Pool3dOpMaker, + ops::PoolOpInferVarType, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(pool3d_grad, ops::PoolOpGrad); diff --git a/paddle/fluid/operators/pool_op.h b/paddle/fluid/operators/pool_op.h index a63963ca926bb94ff99e5cfe6dbcb2b15075bcb8..c0594b7e3cc5602a44bb01951a22c2135ba5c7ce 100644 --- a/paddle/fluid/operators/pool_op.h +++ b/paddle/fluid/operators/pool_op.h @@ -69,6 +69,7 @@ class PoolKernel : public framework::OpKernel { std::vector ksize = context.Attr>("ksize"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); + bool exclusive = context.Attr("exclusive"); if (context.Attr("global_pooling")) { for (size_t i = 0; i < ksize.size(); ++i) { paddings[i] = 0; @@ -84,7 +85,7 @@ class PoolKernel : public framework::OpKernel { pool2d_forward; paddle::operators::math::MaxPool pool_process; pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, - out); + true, out); } else if (pooling_type == "avg") { paddle::operators::math::Pool2dFunctor< @@ -92,7 +93,7 @@ class PoolKernel : public framework::OpKernel { pool2d_forward; paddle::operators::math::AvgPool pool_process; pool2d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, - out); + exclusive, out); } } break; case 3: { @@ -102,14 +103,14 @@ class PoolKernel : public framework::OpKernel { pool3d_forward; paddle::operators::math::MaxPool pool_process; pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, - out); + true, out); } else if (pooling_type == "avg") { paddle::operators::math::Pool3dFunctor< DeviceContext, paddle::operators::math::AvgPool, T> pool3d_forward; paddle::operators::math::AvgPool pool_process; pool3d_forward(dev_ctx, *in_x, ksize, strides, paddings, pool_process, - out); + exclusive, out); } } break; default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } @@ -131,6 +132,7 @@ class PoolGradKernel : public framework::OpKernel { std::vector ksize = context.Attr>("ksize"); std::vector strides = context.Attr>("strides"); std::vector paddings = context.Attr>("paddings"); + bool exclusive = context.Attr("exclusive"); if (context.Attr("global_pooling")) { for (size_t i = 0; i < ksize.size(); ++i) { @@ -157,7 +159,7 @@ class PoolGradKernel : public framework::OpKernel { pool2d_backward; paddle::operators::math::AvgPoolGrad pool_process; pool2d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, - paddings, pool_process, in_x_grad); + paddings, pool_process, exclusive, in_x_grad); } } break; case 3: { @@ -172,7 +174,7 @@ class PoolGradKernel : public framework::OpKernel { pool3d_backward; paddle::operators::math::AvgPoolGrad pool_process; pool3d_backward(dev_ctx, *in_x, *out, *out_grad, ksize, strides, - paddings, pool_process, in_x_grad); + paddings, pool_process, exclusive, in_x_grad); } } break; default: { PADDLE_THROW("Pool op only supports 2D and 3D input."); } diff --git a/paddle/fluid/operators/prefetch_op.cc b/paddle/fluid/operators/prefetch_op.cc index 0519c15e13aac99802ff0f95b975712b36b44246..55853d25460bf6e3d07c829d686e71cc9367118c 100644 --- a/paddle/fluid/operators/prefetch_op.cc +++ b/paddle/fluid/operators/prefetch_op.cc @@ -42,17 +42,18 @@ class PrefetchOp : public framework::OperatorBase { auto& ctx = *pool.Get(place); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); std::vector rets; for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { - VLOG(3) << "sending " << ins[i] << " to " << epmap[i] << " to get " - << outs[i] << " back"; + VLOG(30) << "sending " << ins[i] << " to " << epmap[i] << " to get " + << outs[i] << " back"; rets.push_back(rpc_client->AsyncPrefetchVar(epmap[i], ctx, scope, ins[i], outs[i])); } else { - VLOG(3) << "don't send no-initialied variable: " << ins[i]; + VLOG(30) << "don't send no-initialied variable: " << ins[i]; } } for (size_t i = 0; i < rets.size(); i++) { @@ -69,6 +70,7 @@ class PrefetchOpMaker : public framework::OpProtoAndCheckerMaker { "(LoDTensor) result " "to be fetched from parameter server") .AsDuplicable(); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr>( "epmap", "(string vector, default 127.0.0.1:6164)" diff --git a/paddle/fluid/operators/random_crop_op.h b/paddle/fluid/operators/random_crop_op.h index d68ba9d661698bb0d33b139f5748daec2ead6595..5f1a48b6de01550978638917e3c66ef2851ee2ed 100644 --- a/paddle/fluid/operators/random_crop_op.h +++ b/paddle/fluid/operators/random_crop_op.h @@ -155,8 +155,8 @@ class RandomCropKernel : public framework::OpKernel { seed = *cpu_seed.data(); } } else { - VLOG(5) << "WARNING: The input 'Seed' is not initialized, use attribute " - "'startup_seed' instead."; + VLOG(50) << "WARNING: The input 'Seed' is not initialized, use attribute " + "'startup_seed' instead."; seed = ctx.Attr("startup_seed"); } auto shape = ctx.Attr>("shape"); diff --git a/paddle/fluid/operators/read_op.cc b/paddle/fluid/operators/read_op.cc index a0d640b2020958af53a4405ae886eadb2a1e117e..a0b70938d354cbb3bf10a9c8c589ba5153624f45 100644 --- a/paddle/fluid/operators/read_op.cc +++ b/paddle/fluid/operators/read_op.cc @@ -33,6 +33,19 @@ class ReadInferShape : public framework::InferShapeBase { reader_dims.size(), out_names.size(), "The reader's dim number doesn't match the output number."); ctx->SetOutputsDim("Out", reader_dims); + if (!ctx->IsRuntime()) { + auto in_desc = + boost::get(ctx->GetInputVarPtrs("Reader")[0]); + auto in_lod_levels = in_desc->GetLoDLevels(); + auto out_var_ptrs = ctx->GetOutputVarPtrs("Out"); + PADDLE_ENFORCE_EQ(in_lod_levels.size(), out_var_ptrs.size(), + "LoDLevels of Input(Reader) must be the same as the " + "number of Outputs(Out)."); + for (size_t i = 0; i < out_var_ptrs.size(); ++i) { + auto* out_desc = boost::get(out_var_ptrs[i]); + out_desc->SetLoDLevel(in_lod_levels[i]); + } + } } }; diff --git a/paddle/fluid/operators/reader/blocking_queue.h b/paddle/fluid/operators/reader/blocking_queue.h index 28cc91a5ed5d74994e5b960a0a4dd3c6a5e6cdcc..618248f87298d62078aeccfa135b853b9d2b1744 100644 --- a/paddle/fluid/operators/reader/blocking_queue.h +++ b/paddle/fluid/operators/reader/blocking_queue.h @@ -31,8 +31,8 @@ class BlockingQueue { // is a workaround and a simplified version of framework::Channel as it // doesn't support GPU and it implements on buffered blocking queue. public: - explicit BlockingQueue(size_t capacity) - : capacity_(capacity), closed_(false) { + explicit BlockingQueue(size_t capacity, bool speed_test_mode = false) + : capacity_(capacity), speed_test_mode_(speed_test_mode), closed_(false) { PADDLE_ENFORCE_GT( capacity_, 0, "The capacity of a reader::BlockingQueue must be greater than 0."); @@ -42,7 +42,7 @@ class BlockingQueue { std::unique_lock lock(mutex_); send_cv_.wait(lock, [&] { return queue_.size() < capacity_ || closed_; }); if (closed_) { - VLOG(5) + VLOG(50) << "WARNING: Sending an element to a closed reader::BlokcingQueue."; return false; } @@ -56,7 +56,7 @@ class BlockingQueue { std::unique_lock lock(mutex_); send_cv_.wait(lock, [&] { return queue_.size() < capacity_ || closed_; }); if (closed_) { - VLOG(5) + VLOG(50) << "WARNING: Sending an element to a closed reader::BlokcingQueue."; return false; } @@ -72,7 +72,9 @@ class BlockingQueue { if (!queue_.empty()) { PADDLE_ENFORCE_NOT_NULL(elem); *elem = queue_.front(); - queue_.pop_front(); + if (LIKELY(!speed_test_mode_)) { + queue_.pop_front(); + } send_cv_.notify_one(); return true; } else { @@ -114,6 +116,7 @@ class BlockingQueue { private: size_t capacity_; + bool speed_test_mode_; bool closed_; std::deque queue_; diff --git a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc index 3f72890a7cee1453585d50afa04fa62a9b059dc3..3fe4e9e7adee071fd56cf9f3d2560829f096ba9b 100644 --- a/paddle/fluid/operators/reader/create_shuffle_reader_op.cc +++ b/paddle/fluid/operators/reader/create_shuffle_reader_op.cc @@ -26,7 +26,7 @@ class ShuffleReader : public framework::DecoratedReader { ShuffleReader(const std::shared_ptr& reader, size_t buffer_size, size_t seed = 0) : DecoratedReader(reader), buffer_size_(buffer_size), seed_(seed) { - VLOG(10) << "Create shuffle reader of " << reader_; + VLOG(100) << "Create shuffle reader of " << reader_; if (seed_ == 0) { std::random_device device; seed_ = device(); @@ -37,7 +37,7 @@ class ShuffleReader : public framework::DecoratedReader { void ReadNextImpl(std::vector* out) override { out->clear(); if (iteration_pos_ >= buffer_.size()) { - VLOG(10) << "Resetting shuffle buffer"; + VLOG(100) << "Resetting shuffle buffer"; ReloadBuffer(); if (buffer_.empty()) { return; @@ -73,7 +73,7 @@ class ShuffleReader : public framework::DecoratedReader { std::mt19937 g(seed_); std::shuffle(buffer_.begin(), buffer_.end(), g); seed_ = g(); // update seed_; - VLOG(10) << "random buffer size = " << buffer_.size(); + VLOG(100) << "random buffer size = " << buffer_.size(); } size_t buffer_size_; diff --git a/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h b/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h index 4f7cfc24ec035349f3c85e84d876ad9b5b5493a6..3f041ff7e4e32b407729a22aab25d3aab199fee0 100644 --- a/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h +++ b/paddle/fluid/operators/reader/lod_tensor_blocking_queue.h @@ -33,8 +33,9 @@ class LoDTensorBlockingQueue { private: LoDTensorBlockingQueue(size_t capacity, - const std::vector& dims) - : queue_(capacity), dims_(dims) {} + const std::vector& dims, + bool speed_test_mode = false) + : queue_(capacity, speed_test_mode), dims_(dims) {} public: bool Push(const std::vector& lod_tensor_vec) { @@ -69,11 +70,12 @@ class LoDTensorBlockingQueue { class LoDTensorBlockingQueueHolder { public: - void InitOnce(size_t capacity, const std::vector& dims) { + void InitOnce(size_t capacity, const std::vector& dims, + bool speed_test_mode = false) { PADDLE_ENFORCE( queue_ == nullptr, "LoDTensorBlockingQueueHolder::InitOnce() can only be called once"); - queue_.reset(new LoDTensorBlockingQueue(capacity, dims)); + queue_.reset(new LoDTensorBlockingQueue(capacity, dims, speed_test_mode)); } inline const std::shared_ptr& GetQueue() const { diff --git a/paddle/fluid/operators/reader/reader_blocking_queue_test.cc b/paddle/fluid/operators/reader/reader_blocking_queue_test.cc index 7d1b381d56c8cdc1e79e594b18c1a1ed59ab5284..dc0940ac0b78d295b5088cb6ae26300da1dc883d 100644 --- a/paddle/fluid/operators/reader/reader_blocking_queue_test.cc +++ b/paddle/fluid/operators/reader/reader_blocking_queue_test.cc @@ -217,3 +217,27 @@ TEST(BlockingQueue, MyClassTest) { q.Receive(&b); EXPECT_EQ(a.val_, b.val_); } + +TEST(BlockingQueue, speed_test_mode) { + size_t queue_size = 10; + BlockingQueue q1(queue_size, false); + for (size_t i = 0; i < queue_size; ++i) { + q1.Send(i); + } + size_t b; + for (size_t i = 0; i < queue_size; ++i) { + q1.Receive(&b); + EXPECT_EQ(b, i); + } + EXPECT_EQ(q1.Size(), 0UL); + + BlockingQueue q2(queue_size, true); + for (size_t i = 0; i < queue_size; ++i) { + q2.Send(i); + } + for (size_t i = 0; i < queue_size; ++i) { + q2.Receive(&b); + EXPECT_EQ(b, 0UL); + } + EXPECT_EQ(q2.Size(), queue_size); +} diff --git a/paddle/fluid/operators/recurrent_op.cc b/paddle/fluid/operators/recurrent_op.cc index 162bfcbb0844d29385d0f8ad5d25a3f8de6bd41b..283dce93212ac91fc4a3276598c1f32cfd36d1e7 100644 --- a/paddle/fluid/operators/recurrent_op.cc +++ b/paddle/fluid/operators/recurrent_op.cc @@ -160,7 +160,7 @@ class RecurrentBase : public framework::OperatorBase { Callback callback) { PADDLE_ENFORCE_EQ(src_vars.size(), dst_vars.size()); for (size_t i = 0; i < dst_vars.size(); ++i) { - VLOG(10) << "Link " << src_vars[i] << " to " << dst_vars[i]; + VLOG(100) << "Link " << src_vars[i] << " to " << dst_vars[i]; AccessTensor(src_scope, src_vars[i], dst_scope, dst_vars[i], callback); } } @@ -176,7 +176,7 @@ class RecurrentBase : public framework::OperatorBase { Callback callback) { PADDLE_ENFORCE_EQ(src_vars.size(), dst_vars.size()); for (size_t i = 0; i < dst_vars.size(); ++i) { - VLOG(10) << "Link " << src_vars[i] << " to " << dst_vars[i]; + VLOG(100) << "Link " << src_vars[i] << " to " << dst_vars[i]; AccessTensor(src_scope, src_vars[i], dst_scope, dst_vars[i], callback); } } @@ -230,7 +230,7 @@ class RecurrentOp : public RecurrentBase { void RunImpl(const framework::Scope &scope, const platform::Place &place) const override { auto seq_len = static_cast(this->GetSequenceLength(scope)); - VLOG(3) << "Static RNN input sequence length = " << seq_len; + VLOG(30) << "Static RNN input sequence length = " << seq_len; StepScopes scopes = CreateStepScopes(scope, seq_len); auto reverse = Attr(kReverse); @@ -241,7 +241,7 @@ class RecurrentOp : public RecurrentBase { for (size_t i = 0; i < seq_len; ++i) { size_t seq_offset = reverse ? seq_len - i - 1 : i; - VLOG(3) << "Recurrent operate at the time step " << seq_offset; + VLOG(30) << "Recurrent operate at the time step " << seq_offset; auto &cur_scope = scopes.CurScope(); @@ -334,7 +334,7 @@ class RecurrentGradOp : public RecurrentBase { for (size_t step_id = 0; step_id < seq_len; ++step_id) { size_t seq_offset = reverse ? step_id : seq_len - step_id - 1; - VLOG(3) << "Recurrent backward operate at the time step " << seq_offset; + VLOG(30) << "Recurrent backward operate at the time step " << seq_offset; auto &cur_scope = scopes.CurScope(); // Link outside::output_grads --> inside::output_grads // inside::output_grad = outside::output_grad[seq_offset:seq_offset+1] @@ -348,11 +348,11 @@ class RecurrentGradOp : public RecurrentBase { }); auto og_set = List2Set(Inputs(kOutputGrads)); - if (VLOG_IS_ON(10)) { + if (VLOG_IS_ON(100)) { std::ostringstream sout; std::copy(og_set.begin(), og_set.end(), std::ostream_iterator(sout, ",")); - VLOG(10) << " RNN output gradients = [" << sout.str() << "]"; + VLOG(100) << " RNN output gradients = [" << sout.str() << "]"; } // Link states @@ -374,7 +374,7 @@ class RecurrentGradOp : public RecurrentBase { auto &ex_tensor = ex_scope.FindVar(ex_grad)->Get(); - VLOG(10) << " RNN link " << cur_grad << " from " << ex_grad; + VLOG(100) << " RNN link " << cur_grad << " from " << ex_grad; auto *cur_grad_var = cur_scope.Var(cur_grad); auto cur_grad_tensor = cur_grad_var->GetMutable(); @@ -382,12 +382,12 @@ class RecurrentGradOp : public RecurrentBase { } } - VLOG(5) << "Recurrent memory linking finished "; + VLOG(50) << "Recurrent memory linking finished "; // Run step block with cur_scope executor.Run(*program, &cur_scope, block->ID(), false /*create_local_scope*/); - VLOG(5) << "executor.Run finished "; + VLOG(50) << "executor.Run finished "; auto local_var_names = LocalVarNames(cur_scope); @@ -436,7 +436,7 @@ class RecurrentGradOp : public RecurrentBase { cur_scope.Rename(new_inside_name, inside_grad_name); } } - VLOG(5) << "Accumulate Parameter finished "; + VLOG(50) << "Accumulate Parameter finished "; // Copy input gradient from inside to outside // outside::input_grad[seq_offset: seq_offset + 1] = inside::input_grad @@ -455,7 +455,7 @@ class RecurrentGradOp : public RecurrentBase { auto dst = outside->Slice(seq_offset, seq_offset + 1); framework::TensorCopy(inside, place, dev_ctx, &dst); }); - VLOG(5) << "Link outside gradient finished "; + VLOG(50) << "Link outside gradient finished "; if (step_id + 1 == seq_len) { // at_end // copy initialize states gradient from inside to outside @@ -468,7 +468,7 @@ class RecurrentGradOp : public RecurrentBase { outside->mutable_data(place, inside.type()); framework::TensorCopy(inside, place, dev_ctx, outside); }); - VLOG(5) << "Link initialize state gradient finished "; + VLOG(50) << "Link initialize state gradient finished "; } scopes.Next(); } diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc index 4d34b8a1686efb1fc30020f0d27e9a3c3a6c0866..fbbd86502bfc61c004f88971526195f6a083d5a9 100644 --- a/paddle/fluid/operators/recv_op.cc +++ b/paddle/fluid/operators/recv_op.cc @@ -42,11 +42,12 @@ class RecvOp : public framework::OperatorBase { auto& ctx = *pool.Get(place); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); std::vector rets; for (size_t i = 0; i < outs.size(); i++) { - VLOG(3) << "getting " << outs[i] << " from " << epmap[i]; + VLOG(30) << "getting " << outs[i] << " from " << epmap[i]; rets.push_back(rpc_client->AsyncGetVar(epmap[i], ctx, scope, outs[i])); } if (sync_mode) { @@ -73,6 +74,7 @@ This operator can get variables from server side. "Server endpoints in the order of input " "variables for mapping") .SetDefault({}); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr("sync_mode", "(int, default 0)" "sync recv or async recv.") diff --git a/paddle/fluid/operators/reduce_max_op.cu b/paddle/fluid/operators/reduce_max_op.cu index 0d86b3127e42f7ee14ba57b1c762e8128a0f2d54..b21da178f3eeaafa41bde5f64cc4abcf7944b032 100644 --- a/paddle/fluid/operators/reduce_max_op.cu +++ b/paddle/fluid/operators/reduce_max_op.cu @@ -23,12 +23,3 @@ REGISTER_OP_CUDA_KERNEL(reduce_max, int, ops::MaxFunctor>, ops::ReduceKernel); -REGISTER_OP_CUDA_KERNEL( - reduce_max_grad, ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_max_op.part.cu b/paddle/fluid/operators/reduce_max_op.part.cu new file mode 100644 index 0000000000000000000000000000000000000000..6954c8d744faee6f8f0b715d6e4c8e3bcda7fb83 --- /dev/null +++ b/paddle/fluid/operators/reduce_max_op.part.cu @@ -0,0 +1,25 @@ +// 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/reduce_min_max_op.h" + +REGISTER_OP_CUDA_KERNEL( + reduce_max_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_mean_op.cu b/paddle/fluid/operators/reduce_mean_op.cu index 59b30244839849d79e3e531953134633503c4090..4408200d2d052c2f68c2dd35619de6ed67f07f6e 100644 --- a/paddle/fluid/operators/reduce_mean_op.cu +++ b/paddle/fluid/operators/reduce_mean_op.cu @@ -69,13 +69,3 @@ REGISTER_OP_CUDA_KERNEL(reduce_mean, ops::ReduceMeanKernel, ops::ReduceMeanKernel, ops::ReduceMeanKernel, ops::ReduceMeanKernel); - -REGISTER_OP_CUDA_KERNEL( - reduce_mean_grad, ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_mean_op.part.cu b/paddle/fluid/operators/reduce_mean_op.part.cu new file mode 100644 index 0000000000000000000000000000000000000000..4b663bcdca7c20f8802d962a362f429d8eafe9af --- /dev/null +++ b/paddle/fluid/operators/reduce_mean_op.part.cu @@ -0,0 +1,26 @@ +// 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. + +// .part used to speed up nvcc compile +#include "paddle/fluid/operators/reduce_mean_op.h" + +REGISTER_OP_CUDA_KERNEL( + reduce_mean_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_min_op.cu b/paddle/fluid/operators/reduce_min_op.cu index da466f805eff4709dc23471baef03e94052ee6c1..5a04a12b79444dcea30d3c1140d9708a98b55fe3 100644 --- a/paddle/fluid/operators/reduce_min_op.cu +++ b/paddle/fluid/operators/reduce_min_op.cu @@ -23,12 +23,3 @@ REGISTER_OP_CUDA_KERNEL(reduce_min, int, ops::MinFunctor>, ops::ReduceKernel); -REGISTER_OP_CUDA_KERNEL( - reduce_min_grad, ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_min_op.part.cu b/paddle/fluid/operators/reduce_min_op.part.cu new file mode 100644 index 0000000000000000000000000000000000000000..5b8f061b2d03eb76863401905ac87044fd5ea778 --- /dev/null +++ b/paddle/fluid/operators/reduce_min_op.part.cu @@ -0,0 +1,25 @@ +// 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/reduce_min_max_op.h" + +REGISTER_OP_CUDA_KERNEL( + reduce_min_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_prod_op.cu b/paddle/fluid/operators/reduce_prod_op.cu index d62e677d92cffecf629d1684026b0c7bcfec29e3..d8692afb96e4d5d3206210060684dd12fb4d79a7 100644 --- a/paddle/fluid/operators/reduce_prod_op.cu +++ b/paddle/fluid/operators/reduce_prod_op.cu @@ -23,12 +23,3 @@ REGISTER_OP_CUDA_KERNEL(reduce_prod, int, ops::ProdFunctor>, ops::ReduceKernel); -REGISTER_OP_CUDA_KERNEL( - reduce_prod_grad, ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_prod_op.part.cu b/paddle/fluid/operators/reduce_prod_op.part.cu new file mode 100644 index 0000000000000000000000000000000000000000..486c578c64b9a2d80abc940a7c4266ef5fd23c7f --- /dev/null +++ b/paddle/fluid/operators/reduce_prod_op.part.cu @@ -0,0 +1,25 @@ +// 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/reduce_prod_op.h" + +REGISTER_OP_CUDA_KERNEL( + reduce_prod_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_sum_op.cu b/paddle/fluid/operators/reduce_sum_op.cu index 53cd9e9419dd9aecee730917ae21d7a4ab332ffc..2b031e8df99768c9208146640bddbe51149b2614 100644 --- a/paddle/fluid/operators/reduce_sum_op.cu +++ b/paddle/fluid/operators/reduce_sum_op.cu @@ -64,13 +64,3 @@ class ReduceSumKernel : public framework::OpKernel { REGISTER_OP_CUDA_KERNEL(reduce_sum, ops::ReduceSumKernel, ops::ReduceSumKernel, ops::ReduceSumKernel, ops::ReduceSumKernel); - -REGISTER_OP_CUDA_KERNEL( - reduce_sum_grad, ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel, - ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/reduce_sum_op.part.cu b/paddle/fluid/operators/reduce_sum_op.part.cu new file mode 100644 index 0000000000000000000000000000000000000000..525633f62a95b2d0d677fcbebe551b75cb2a180d --- /dev/null +++ b/paddle/fluid/operators/reduce_sum_op.part.cu @@ -0,0 +1,26 @@ +// 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/cub_reduce.h" +#include "paddle/fluid/operators/reduce_sum_op.h" + +REGISTER_OP_CUDA_KERNEL( + reduce_sum_grad, ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel, + ops::ReduceGradKernel); diff --git a/paddle/fluid/operators/ref_by_trainer_id_op.cc b/paddle/fluid/operators/ref_by_trainer_id_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..6cb651af6dc3d8e301365968787c199acc4c60ee --- /dev/null +++ b/paddle/fluid/operators/ref_by_trainer_id_op.cc @@ -0,0 +1,79 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/ref_by_trainer_id_op.h" +#include + +namespace paddle { +namespace operators { + +class RefByTrainerIdOp : public framework::OperatorWithKernel { + public: + RefByTrainerIdOp(const std::string &type, + const framework::VariableNameMap &inputs, + const framework::VariableNameMap &outputs, + const framework::AttributeMap &attrs) + : OperatorWithKernel(type, inputs, outputs, attrs) {} + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInputs("X"), + "Input(X) of RefByTrainerIdOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("TrainerId"), + "Input(TrainerId) of RefByTrainerIdOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of RefByTrainerIdOp should not be null."); + PADDLE_ENFORCE_EQ(ctx->GetInputDim("TrainerId").size(), 1, + "TrainerId should be a scalar."); + // Out's shape is determined at runtime. + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType( + ctx.MultiInput("X")[0]->type()), + ctx.GetPlace()); + } +}; + +class RefByTrainerIdOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "(Tensor) Input tensor list.").AsDuplicable(); + AddInput("TrainerId", "(Tensor) Scalar int, the trainer id runtime value."); + AddOutput("Out", "(Tensor) Return one tensor reference of X[trainer_id]"); + AddComment(R"DOC( +**RefByTrainerId operator** + +Return a reference of a tensor, using trainer_id as the index to find from the input. + +$$Out = X[TrainerId]$$ +)DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_WITHOUT_GRADIENT(ref_by_trainer_id, ops::RefByTrainerIdOp, + ops::RefByTrainerIdOpMaker); +REGISTER_OP_CPU_KERNEL( + ref_by_trainer_id, + ops::RefByTrainerIdKernel, + ops::RefByTrainerIdKernel, + ops::RefByTrainerIdKernel, + ops::RefByTrainerIdKernel); diff --git a/paddle/fluid/operators/ref_by_trainer_id_op.cu.cc b/paddle/fluid/operators/ref_by_trainer_id_op.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..b98e2b5c9c7341f2a424fb4b32ff1e8bc45a056c --- /dev/null +++ b/paddle/fluid/operators/ref_by_trainer_id_op.cu.cc @@ -0,0 +1,26 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/ref_by_trainer_id_op.h" + +REGISTER_OP_CUDA_KERNEL( + ref_by_trainer_id, + paddle::operators::RefByTrainerIdKernel, + paddle::operators::RefByTrainerIdKernel, + paddle::operators::RefByTrainerIdKernel, + paddle::operators::RefByTrainerIdKernel); diff --git a/paddle/fluid/operators/ref_by_trainer_id_op.h b/paddle/fluid/operators/ref_by_trainer_id_op.h new file mode 100644 index 0000000000000000000000000000000000000000..2ce577544ae2437b9297da2190fd09b435d5173c --- /dev/null +++ b/paddle/fluid/operators/ref_by_trainer_id_op.h @@ -0,0 +1,48 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +template +class RefByTrainerIdKernel : public framework::OpKernel { + public: + virtual void Compute(const framework::ExecutionContext& context) const { + auto* out = context.Output("Out"); + auto in_list = context.MultiInput("X"); + auto* trainer_id_t = context.Input("TrainerId"); + int64_t trainer_id = 0; + auto* trainer_id_data = trainer_id_t->data(); + if (platform::is_gpu_place(context.GetPlace())) { +#ifdef PADDLE_WITH_CUDA + auto stream = context.cuda_device_context().stream(); + memory::Copy<>(platform::CPUPlace(), &trainer_id, + boost::get(context.GetPlace()), + trainer_id_data, sizeof(int64_t), stream); +#endif + } else { + trainer_id = *trainer_id_data; + } + PADDLE_ENFORCE_LT(trainer_id, in_list.size()); + out->mutable_data(context.GetPlace()); + out->ShareDataWith(*(in_list[trainer_id])); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc index d72f85f2c44db2fa887732cfc05e1376a6a79e4a..500d86fec33830fc2cfb0412f1f2c7780d08eb02 100644 --- a/paddle/fluid/operators/reshape_op.cc +++ b/paddle/fluid/operators/reshape_op.cc @@ -164,7 +164,7 @@ dimension value will be copied from Input(X) at runtime. Note that the index of [2, 3, 4], Attr(shape) = [2, 3, 2, 0] is an invalid input. 3. Input(Shape) has a higher priority than Attr(shape) if it is provided, while -Attr(shape) still should be set correctly to gurantee shape inference in +Attr(shape) still should be set correctly to gurantee shape inference in compile-time. )DOC"); @@ -259,7 +259,6 @@ class Reshape2Op : public ReshapeOp { : ReshapeOp(type, inputs, outputs, attrs) {} void InferShape(framework::InferShapeContext *ctx) const override { - ReshapeOp::InferShape(ctx); PADDLE_ENFORCE(ctx->HasOutput("XShape"), "Output(XShape) of ReshapeOp should not be null."); const auto &x_dims = ctx->GetInputDim("X"); @@ -270,6 +269,8 @@ class Reshape2Op : public ReshapeOp { } ctx->SetOutputDim("XShape", framework::make_ddim(xshape_dims)); ctx->ShareLoD("X", /*->*/ "XShape"); + + ReshapeOp::InferShape(ctx); } }; diff --git a/paddle/fluid/operators/rmsprop_op.cc b/paddle/fluid/operators/rmsprop_op.cc index 2f773f222e50a440801b06a4fd997bf237b34772..f06f87e61d3a4d1fc8b864b9dd84e697fb12a006 100644 --- a/paddle/fluid/operators/rmsprop_op.cc +++ b/paddle/fluid/operators/rmsprop_op.cc @@ -32,6 +32,11 @@ class RmspropOp : public framework::OperatorWithKernel { "Input(Grad) of RmspropOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Moment"), "Input(Moment) of RmspropOp should not be null."); + PADDLE_ENFORCE( + ctx->GetInputsVarType("Param").front() == + framework::proto::VarType::LOD_TENSOR, + "The input var's type should be LoDTensor, but the received is %s", + ctx->Inputs("Param").front(), ctx->GetInputsVarType("Param").front()); PADDLE_ENFORCE(ctx->HasOutput("ParamOut"), "Output(param_out) of RmspropOp should not be null."); diff --git a/paddle/fluid/operators/rmsprop_op.h b/paddle/fluid/operators/rmsprop_op.h index 25ed32c5ebb2ff5be962ac1e3e38c970623d705c..389c84d2464090ff9bd9e8b471cd0103c86a347a 100644 --- a/paddle/fluid/operators/rmsprop_op.h +++ b/paddle/fluid/operators/rmsprop_op.h @@ -13,66 +13,254 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/algorithm.h" +#include "paddle/fluid/operators/math/selected_rows_functor.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { -using Tensor = framework::Tensor; template using EigenVector = framework::EigenVector; +template +struct DenseRmspropGradFunctor { + inline explicit DenseRmspropGradFunctor(const T *grad) : grad_(grad) {} + + HOSTDEVICE inline T operator()(int64_t idx) const { return grad_[idx]; } + + const T *grad_; +}; + +template +struct SparseRmspropGradFunctor { + inline SparseRmspropGradFunctor(const T *grad, const int64_t *rows, + int64_t row_numel, int64_t row_count) + : grad_(grad), + rows_(rows), + row_numel_(row_numel), + row_count_(row_count) {} + + HOSTDEVICE inline T operator()(int64_t idx) const { + auto row_idx = math::BinarySearch(rows_, row_count_, idx / row_numel_); + return row_idx >= 0 ? grad_[row_idx * row_numel_ + idx % row_numel_] : 0; + } + + const T *grad_; + const int64_t *rows_; + int64_t row_numel_; + int64_t row_count_; +}; + +template +struct UncenteredRmspropFunctor { + UncenteredRmspropFunctor(T *param, T *ms, T *mom, const T *lr, T rho, + T epsilon, T momentum, + const GradFunctor &grad_functor) + : param_(param), + ms_(ms), + mom_(mom), + lr_(lr), + rho_(rho), + epsilon_(epsilon), + momentum_(momentum), + grad_functor_(grad_functor) {} + + HOSTDEVICE inline void operator()(int64_t idx) const { + T g = grad_functor_(idx); + T ms_out = rho_ * ms_[idx] + (1 - rho_) * g * g; + T mom_out = momentum_ * mom_[idx] + lr_[0] * g / sqrt(ms_out + epsilon_); + param_[idx] -= mom_out; + ms_[idx] = ms_out; + mom_[idx] = mom_out; + } + + T *param_; + T *ms_; + T *mom_; + const T *lr_; + T rho_; + T epsilon_; + T momentum_; + GradFunctor grad_functor_; +}; + +template +struct CenteredRmspropFunctor { + CenteredRmspropFunctor(T *param, T *ms, T *mom, T *mean_grad, const T *lr, + T rho, T epsilon, T momentum, + const GradFunctor &grad_functor) + : param_(param), + ms_(ms), + mom_(mom), + mean_grad_(mean_grad), + lr_(lr), + rho_(rho), + epsilon_(epsilon), + momentum_(momentum), + grad_functor_(grad_functor) {} + + HOSTDEVICE inline void operator()(int64_t idx) const { + T g = grad_functor_(idx); + T ms_out = rho_ * ms_[idx] + (1 - rho_) * g * g; + T mg_out = rho_ * mean_grad_[idx] + (1 - rho_) * g; + T mom_out = momentum_ * mom_[idx] + + lr_[0] * g / sqrt(ms_out - mg_out * mg_out + epsilon_); + param_[idx] -= mom_out; + ms_[idx] = ms_out; + mom_[idx] = mom_out; + mean_grad_[idx] = mg_out; + } + + T *param_; + T *ms_; + T *mom_; + T *mean_grad_; + const T *lr_; + T rho_; + T epsilon_; + T momentum_; + GradFunctor grad_functor_; +}; + template class RmspropOpKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto* param_out = ctx.Output("ParamOut"); - auto* moment_out = ctx.Output("MomentOut"); - auto* mean_square_out = ctx.Output("MeanSquareOut"); + void Compute(const framework::ExecutionContext &ctx) const override { + using LoDTensor = framework::LoDTensor; + auto *grad_var = ctx.InputVar("Grad"); + auto *param_out = ctx.Output("ParamOut"); + auto *moment_out = ctx.Output("MomentOut"); + auto *mean_square_out = ctx.Output("MeanSquareOut"); - auto grad = ctx.Input("Grad"); + auto epsilon = static_cast(ctx.Attr("epsilon")); + auto rho = static_cast(ctx.Attr("decay")); + auto momentum = static_cast(ctx.Attr("momentum")); + bool centered = ctx.Attr("centered"); - param_out->mutable_data(ctx.GetPlace()); - moment_out->mutable_data(ctx.GetPlace()); - mean_square_out->mutable_data(ctx.GetPlace()); + auto &p_tensor = *ctx.Input("Param"); + auto &ms_tensor = *ctx.Input("MeanSquare"); + auto &lr_tensor = *ctx.Input("LearningRate"); + auto &mom_tensor = *ctx.Input("Moment"); - float epsilon = ctx.Attr("epsilon"); - float rho = ctx.Attr("decay"); - float momentum = ctx.Attr("momentum"); - bool centered = ctx.Attr("centered"); + PADDLE_ENFORCE_EQ(&p_tensor, param_out, + "Param and ParamOut must be the same Tensor"); + PADDLE_ENFORCE_EQ(&mom_tensor, moment_out, + "Moment and MomentOut must be the same Tensor"); + PADDLE_ENFORCE_EQ(&ms_tensor, mean_square_out, + "MeanSquare and MeanSquareOut must be the same Tensor"); + + auto &dev_ctx = ctx.template device_context(); + size_t limit = static_cast(ms_tensor.numel()); + + if (grad_var->IsType()) { + auto &grad_tensor = grad_var->Get(); + + if (std::is_same::value) { + auto &place = + *ctx.template device_context().eigen_device(); + auto lr_value = lr_tensor.data()[0]; + + auto p = EigenVector::Flatten(p_tensor); + auto ms = EigenVector::Flatten(ms_tensor); + auto g = EigenVector::Flatten(grad_tensor); + auto mom = EigenVector::Flatten(mom_tensor); + + auto p_out = EigenVector::Flatten(*param_out); + auto mom_out = EigenVector::Flatten(*moment_out); + auto ms_out = EigenVector::Flatten(*mean_square_out); + + ms_out.device(place) = rho * ms + (1 - rho) * g * g; + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto mg = EigenVector::Flatten(mg_tensor); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE_EQ(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + auto mg_out = EigenVector::Flatten(*mean_grad_out); + + mg_out.device(place) = rho * mg + (1 - rho) * g; + mom_out.device(place) = + momentum * mom + + lr_value * g / (ms_out - mg_out.square() + epsilon).sqrt(); + } else { + mom_out.device(place) = + momentum * mom + lr_value * g / (ms_out + epsilon).sqrt(); + } + p_out.device(place) = p - mom_out; + } else { + DenseRmspropGradFunctor grad_func(grad_tensor.data()); + platform::ForRange for_range(dev_ctx, limit); + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE_EQ(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + for_range(CenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), + mean_grad_out->mutable_data(ctx.GetPlace()), + lr_tensor.data(), rho, epsilon, momentum, grad_func)); + } else { + for_range(UncenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } + } + } else if (grad_var->IsType()) { + auto &grad = grad_var->Get(); + auto *merged_grad = const_cast(ctx.scope()) + .Var() + ->GetMutable(); + + math::scatter::MergeAdd merge_func; + merge_func(dev_ctx, grad, merged_grad); + + platform::ForRange for_range(dev_ctx, limit); + const int64_t *rows; +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + rows = merged_grad->rows().CUDAData(ctx.GetPlace()); + } else { +#endif + rows = merged_grad->rows().data(); +#ifdef PADDLE_WITH_CUDA + } +#endif + auto &merged_tensor = merged_grad->value(); + int64_t row_count = merged_grad->rows().size(); + int64_t row_numel = merged_tensor.numel() / row_count; + SparseRmspropGradFunctor grad_func(merged_tensor.data(), rows, + row_numel, row_count); - auto p = EigenVector::Flatten(*ctx.Input("Param")); - auto ms = EigenVector::Flatten(*ctx.Input("MeanSquare")); - auto lr = EigenVector::Flatten(*ctx.Input("LearningRate")); - auto g = EigenVector::Flatten(*grad); - auto mom = EigenVector::Flatten(*ctx.Input("Moment")); - - auto p_out = EigenVector::Flatten(*param_out); - auto mom_out = EigenVector::Flatten(*moment_out); - auto ms_out = EigenVector::Flatten(*mean_square_out); - auto& place = *ctx.template device_context().eigen_device(); - - Eigen::DSizes grad_dsize(static_cast(grad->numel())); - - ms_out.device(place) = rho * ms + (1 - rho) * g * g; - if (centered) { - auto mg = EigenVector::Flatten(*ctx.Input("MeanGrad")); - auto* mean_grad_out = ctx.Output("MeanGradOut"); - mean_grad_out->mutable_data(ctx.GetPlace()); - auto mg_out = EigenVector::Flatten(*mean_grad_out); - - mg_out.device(place) = rho * mg + (1 - rho) * g; - mom_out.device(place) = momentum * mom + - lr.broadcast(grad_dsize) * g / - (ms_out - mg_out.square() + epsilon).sqrt(); + if (centered) { + auto &mg_tensor = *ctx.Input("MeanGrad"); + auto *mean_grad_out = ctx.Output("MeanGradOut"); + PADDLE_ENFORCE_EQ(&mg_tensor, mean_grad_out, + "MeanGrad and MeanGradOut must be the same Tensor"); + for_range(CenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), + mean_grad_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } else { + for_range(UncenteredRmspropFunctor>( + param_out->mutable_data(ctx.GetPlace()), + mean_square_out->mutable_data(ctx.GetPlace()), + moment_out->mutable_data(ctx.GetPlace()), lr_tensor.data(), + rho, epsilon, momentum, grad_func)); + } } else { - mom_out.device(place) = - momentum * mom + - lr.broadcast(grad_dsize) * g / (ms_out + epsilon).sqrt(); + PADDLE_THROW("RMSProp only supports LoDTensor or SelectedRows gradient"); } - p_out.device(place) = p - mom_out; } }; diff --git a/paddle/fluid/operators/rnn_memory_helper_op.cc b/paddle/fluid/operators/rnn_memory_helper_op.cc index 0fb7776fd9dbf437673820c7cf9411644272626c..b840e690960cf77a37895f5b3d83c4cdbc2fca35 100644 --- a/paddle/fluid/operators/rnn_memory_helper_op.cc +++ b/paddle/fluid/operators/rnn_memory_helper_op.cc @@ -93,7 +93,7 @@ class RNNMemoryHelperGradOp : public framework::OperatorBase { in_grad_var_name); if (out_grad_var == nullptr) { - VLOG(5) << "Using fill constant 0 as starting gradient"; + VLOG(50) << "Using fill constant 0 as starting gradient"; auto in_var_name = Input("X"); auto *in_var = scope.FindVar(in_var_name); auto &in_var_tensor = in_var->Get(); diff --git a/paddle/fluid/operators/roi_align_op.cc b/paddle/fluid/operators/roi_align_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..c57a34c3a745e8fc03ca57dce478ecf60058a9a9 --- /dev/null +++ b/paddle/fluid/operators/roi_align_op.cc @@ -0,0 +1,166 @@ +/* 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/roi_align_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +class ROIAlignOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of ROIAlignOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("ROIs"), + "Input(ROIs) of ROIAlignOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ROIAlignOp should not be null."); + auto input_dims = ctx->GetInputDim("X"); + auto rois_dims = ctx->GetInputDim("ROIs"); + + PADDLE_ENFORCE(input_dims.size() == 4, + "The format of input tensor is NCHW."); + PADDLE_ENFORCE(rois_dims.size() == 2, + "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]."); + PADDLE_ENFORCE(rois_dims[1] == 4, + "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]."); + int pooled_height = ctx->Attrs().Get("pooled_height"); + int pooled_width = ctx->Attrs().Get("pooled_width"); + float spatial_scale = ctx->Attrs().Get("spatial_scale"); + + PADDLE_ENFORCE_GT(pooled_height, 0, + "The pooled output height must greater than 0"); + PADDLE_ENFORCE_GT(pooled_width, 0, + "The pooled output width must greater than 0"); + PADDLE_ENFORCE_GT(spatial_scale, 0.0f, + "The spatial scale must greater than 0"); + + auto out_dims = input_dims; + out_dims[0] = rois_dims[0]; + out_dims[1] = input_dims[1]; + out_dims[2] = pooled_height; + out_dims[3] = pooled_width; + + ctx->SetOutputDim("Out", out_dims); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + +class ROIAlignGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "The GRAD@Out of ROIAlignGradOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName("X")), + "The GRAD@X of ROIAlignGradOp should not be null."); + ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X")); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + +class ROIAlignOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor), " + "The input of ROIAlignOp. " + "The format of input tensor is NCHW. Where N is batch size, " + "C is the number of input channels, " + "H is the height of the feature, and " + "W is the width of the feature."); + AddInput("ROIs", + "(LoDTensor), " + "ROIs (Regions of Interest) to pool over. " + "should be a 2-D LoDTensor of shape (num_rois, 4)" + "given as [[x1, y1, x2, y2], …]. " + "(x1, y1) is the top left coordinates, and " + "(x2, y2) is the bottom right coordinates."); + AddOutput("Out", + "(Tensor), " + "The output of ROIAlignOp is a 4-D tensor with shape " + "(num_rois, channels, pooled_h, pooled_w)."); + AddAttr("spatial_scale", + "(float, default 1.0), " + "Multiplicative spatial scale factor " + "to translate ROI coords from their input scale " + "to the scale used when pooling.") + .SetDefault(1.0); + AddAttr("pooled_height", + "(int, default 1), " + "The pooled output height.") + .SetDefault(1); + AddAttr("pooled_width", + "(int, default 1), " + "The pooled output width.") + .SetDefault(1); + AddAttr("sampling_ratio", + "(int,default -1)," + "number of sampling points in the interpolation grid" + "If <=0, then grid points are adaptive to roi_width " + "and pooled_w, likewise for height") + .SetDefault(-1); + AddComment(R"DOC( +**RoIAlign Operator** + +Region of interest align (also known as RoI align) is to perform +bilinear interpolation on inputs of nonuniform sizes to obtain +fixed-size feature maps (e.g. 7*7) + +Dividing each region proposal into equal-sized sections with +the pooled_width and pooled_height. Location remains the origin +result. + +In each ROI bin, the value of the four regularly sampled locations +are computed directly through bilinear interpolation. The output is +the mean of four locations. +Thus avoid the misaligned problem. + )DOC"); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(roi_align, ops::ROIAlignOp, ops::ROIAlignOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(roi_align_grad, ops::ROIAlignGradOp); +REGISTER_OP_CPU_KERNEL( + roi_align, + ops::CPUROIAlignOpKernel, + ops::CPUROIAlignOpKernel); +REGISTER_OP_CPU_KERNEL( + roi_align_grad, + ops::CPUROIAlignGradOpKernel, + ops::CPUROIAlignGradOpKernel); diff --git a/paddle/fluid/operators/roi_align_op.cu b/paddle/fluid/operators/roi_align_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..bcec6f3563df7f4e1e48554cc891d596f9e56024 --- /dev/null +++ b/paddle/fluid/operators/roi_align_op.cu @@ -0,0 +1,353 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/roi_align_op.h" +#include "paddle/fluid/platform/cuda_primitives.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +static constexpr int kNumCUDAThreads = 512; +static constexpr int kNumMaxinumNumBlocks = 4096; + +static inline int NumBlocks(const int N) { + return std::min((N + kNumCUDAThreads - 1) / kNumCUDAThreads, + kNumMaxinumNumBlocks); +} + +#define CUDA_1D_KERNEL_LOOP(i, n) \ + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < (n); \ + i += blockDim.x * gridDim.x) + +template +__device__ T BilinearInterpolate(const T* input_data, const int height, + const int width, T y, T x) { + if (y < -1.0 || y > height || x < -1.0 || x > width) { + return 0; + } + y = y <= 0 ? 0 : y; + x = x <= 0 ? 0 : x; + int y_low = static_cast(y); + int x_low = static_cast(x); + int y_high; + int x_high; + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = static_cast(y_low); + } else { + y_high = y_low + 1; + } + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = static_cast(x_low); + } else { + x_high = x_low + 1; + } + T ly = y - y_low, lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + + T v1 = input_data[y_low * width + x_low]; + T v2 = input_data[y_low * width + x_high]; + T v3 = input_data[y_high * width + x_low]; + T v4 = input_data[y_high * width + x_high]; + T w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + + T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4); + return val; +} + +template +__device__ void BilinearInterpolateGradient(const int height, const int width, + T y, T x, T* w1, T* w2, T* w3, + T* w4, int* x_low, int* x_high, + int* y_low, int* y_high) { + if (y < -1.0 || y > height || x < -1.0 || x > width) { + return; + } + + y = y <= 0 ? 0 : y; + x = x <= 0 ? 0 : x; + *y_low = static_cast(y); + *x_low = static_cast(x); + if (*y_low >= height - 1) { + *y_high = *y_low = height - 1; + y = static_cast(*y_low); + } else { + *y_high = *y_low + 1; + } + if (*x_low >= width - 1) { + *x_high = *x_low = width - 1; + x = static_cast(*x_low); + } else { + *x_high = *x_low + 1; + } + T ly = y - *y_low, lx = x - *x_low; + T hy = 1. - ly, hx = 1. - lx; + *w1 = hy * hx, *w2 = hy * lx, *w3 = ly * hx, *w4 = ly * lx; + + return; +} + +template +__global__ void GPUROIAlignForward( + const int nthreads, const T* input_data, const T* input_rois, + const float spatial_scale, const int channels, const int height, + const int width, const int pooled_height, const int pooled_width, + const int sampling_ratio, int* roi_batch_id_data, T* output_data) { + CUDA_1D_KERNEL_LOOP(i, nthreads) { + int pw = i % pooled_width; + int ph = (i / pooled_width) % pooled_height; + int c = (i / pooled_width / pooled_height) % channels; + int n = i / pooled_width / pooled_height / channels; + + const T* offset_input_rois = input_rois + n * kROISize; + int roi_batch_ind = roi_batch_id_data[n]; + + T roi_xmin = offset_input_rois[0] * spatial_scale; + T roi_ymin = offset_input_rois[1] * spatial_scale; + T roi_xmax = offset_input_rois[2] * spatial_scale; + T roi_ymax = offset_input_rois[3] * spatial_scale; + + T roi_width = max(roi_xmax - roi_xmin, static_cast(1.)); + T roi_height = max(roi_ymax - roi_ymin, static_cast(1.)); + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + + const T* offset_input_data = + input_data + (roi_batch_ind * channels + c) * height * width; + + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); + int roi_bin_grid_w = + (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); + const T count = roi_bin_grid_h * roi_bin_grid_w; + T output_val = 0; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + const T y = roi_ymin + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T x = roi_xmin + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + T val = BilinearInterpolate(offset_input_data, height, width, y, x); + output_val += val; + } + } + output_val /= count; + output_data[i] = output_val; + } +} + +template +__global__ void GPUROIAlignBackward(const int nthreads, const T* input_rois, + const T* out_grad, const int num_rois, + const float spatial_scale, + const int channels, const int height, + const int width, const int pooled_height, + const int pooled_width, + const int sampling_ratio, + int* roi_batch_id_data, T* input_grad) { + CUDA_1D_KERNEL_LOOP(i, nthreads) { + int pw = i % pooled_width; + int ph = (i / pooled_width) % pooled_height; + int c = (i / pooled_width / pooled_height) % channels; + int n = i / pooled_width / pooled_height / channels; + const T* offset_input_rois = input_rois + n * kROISize; + int roi_batch_ind = roi_batch_id_data[n]; + + T roi_xmin = offset_input_rois[0] * spatial_scale; + T roi_ymin = offset_input_rois[1] * spatial_scale; + T roi_xmax = offset_input_rois[2] * spatial_scale; + T roi_ymax = offset_input_rois[3] * spatial_scale; + + T roi_width = max(roi_xmax - roi_xmin, static_cast(1.)); + T roi_height = max(roi_ymax - roi_ymin, static_cast(1.)); + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + + T* offset_input_grad = + input_grad + (roi_batch_ind * channels + c) * height * width; + + const T* offset_out_grad = + out_grad + (n * channels + c) * pooled_height * pooled_width; + const T out_grad_this_bin = offset_out_grad[ph * pooled_width + pw]; + + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); + int roi_bin_grid_w = + (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width); + + const T count = roi_bin_grid_h * roi_bin_grid_w; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + const T y = roi_ymin + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T x = roi_xmin + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + T w1 = 0, w2 = 0, w3 = 0, w4 = 0; + int x_low = -1, x_high = -1, y_low = -1, y_high = -1; + BilinearInterpolateGradient(height, width, y, x, &w1, &w2, &w3, &w4, + &x_low, &x_high, &y_low, &y_high); + T diff1 = out_grad_this_bin * w1 / count; + T diff2 = out_grad_this_bin * w2 / count; + T diff3 = out_grad_this_bin * w3 / count; + T diff4 = out_grad_this_bin * w4 / count; + if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) { + platform::CudaAtomicAdd(offset_input_grad + y_low * width + x_low, + diff1); + platform::CudaAtomicAdd(offset_input_grad + y_low * width + x_high, + diff2); + platform::CudaAtomicAdd(offset_input_grad + y_high * width + x_low, + diff3); + platform::CudaAtomicAdd(offset_input_grad + y_high * width + x_high, + diff4); + } + } + } + } +} + +template +class GPUROIAlignOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* rois = ctx.Input("ROIs"); + auto* out = ctx.Output("Out"); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + auto sampling_ratio = ctx.Attr("sampling_ratio"); + + auto in_dims = in->dims(); + int batch_size = in_dims[0]; + int channels = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + + int rois_num = rois->dims()[0]; + + if (rois_num == 0) return; + + int output_size = out->numel(); + int blocks = NumBlocks(output_size); + int threads = kNumCUDAThreads; + + Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(platform::CPUPlace()); + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + PADDLE_ENFORCE_EQ( + rois_batch_size, batch_size, + "The rois_batch_size and imgs batch_size must be the same."); + int rois_num_with_lod = rois_lod[rois_batch_size]; + PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod, + "The rois_num from input and lod must be the same."); + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + Tensor roi_batch_id_list_gpu; + framework::TensorCopySync(roi_batch_id_list, ctx.GetPlace(), + &roi_batch_id_list_gpu); + GPUROIAlignForward< + T><<>>( + output_size, in->data(), rois->data(), spatial_scale, channels, + height, width, pooled_height, pooled_width, sampling_ratio, + roi_batch_id_list_gpu.data(), + out->mutable_data(ctx.GetPlace())); + } +}; + +template +class GPUROIAlignGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* rois = ctx.Input("ROIs"); + + auto* out_grad = ctx.Input(framework::GradVarName("Out")); + auto* in_grad = ctx.Output(framework::GradVarName("X")); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + auto sampling_ratio = ctx.Attr("sampling_ratio"); + + int rois_num = rois->dims()[0]; + int channels = in->dims()[1]; + int height = in->dims()[2]; + int width = in->dims()[3]; + + if (!in_grad) { + return; + } + Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(platform::CPUPlace()); + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + Tensor roi_batch_id_list_gpu; + framework::TensorCopySync(roi_batch_id_list, ctx.GetPlace(), + &roi_batch_id_list_gpu); + + in_grad->mutable_data(ctx.GetPlace()); + math::SetConstant set_zero; + set_zero(ctx.cuda_device_context(), in_grad, static_cast(0)); + + int output_grad_size = out_grad->numel(); + int blocks = NumBlocks(output_grad_size); + int threads = kNumCUDAThreads; + + if (output_grad_size > 0) { + GPUROIAlignBackward< + T><<>>( + output_grad_size, rois->data(), out_grad->data(), rois_num, + spatial_scale, channels, height, width, pooled_height, pooled_width, + sampling_ratio, roi_batch_id_list_gpu.data(), + in_grad->mutable_data(ctx.GetPlace())); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + roi_align, + ops::GPUROIAlignOpKernel, + ops::GPUROIAlignOpKernel); +REGISTER_OP_CUDA_KERNEL( + roi_align_grad, + ops::GPUROIAlignGradOpKernel, + ops::GPUROIAlignGradOpKernel); diff --git a/paddle/fluid/operators/roi_align_op.h b/paddle/fluid/operators/roi_align_op.h new file mode 100644 index 0000000000000000000000000000000000000000..a18aee1b86283cbb48f0b804ccfc476d7cd78f3b --- /dev/null +++ b/paddle/fluid/operators/roi_align_op.h @@ -0,0 +1,332 @@ +/* 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 "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; + +static constexpr int kROISize = 4; + +template +void PreCalcForBilinearInterpolate( + const platform::DeviceContext& ctx, const int height, const int width, + const int pooled_height, const int pooled_width, const int iy_upper, + const int ix_upper, T roi_ymin, T roi_xmin, T bin_size_h, T bin_size_w, + int roi_bin_grid_h, int roi_bin_grid_w, Tensor* pre_pos, Tensor* pre_w) { + int pre_calc_index = 0; + int* pre_pos_data = pre_pos->mutable_data(ctx.GetPlace()); + T* pre_w_data = pre_w->mutable_data(ctx.GetPlace()); + for (int ph = 0; ph < pooled_height; ph++) { + for (int pw = 0; pw < pooled_width; pw++) { + for (int iy = 0; iy < iy_upper; iy++) { + // calculate y of sample points + T y = roi_ymin + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); + // calculate x of samle points + for (int ix = 0; ix < ix_upper; ix++) { + T x = roi_xmin + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + // deal with elements out of map + if (y < -1.0 || y > height || x < -1.0 || x > width) { + for (int i = 0; i < kROISize; ++i) { + pre_pos_data[i + pre_calc_index * kROISize] = 0; + pre_w_data[i + pre_calc_index * kROISize] = 0; + } + pre_calc_index += 1; + continue; + } + y = y <= 0 ? 0 : y; + x = x <= 0 ? 0 : x; + + int y_low = static_cast(y); + int x_low = static_cast(x); + int y_high; + int x_high; + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = static_cast(y_low); + } else { + y_high = y_low + 1; + } + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = static_cast(x_low); + } else { + x_high = x_low + 1; + } + T ly = y - y_low, lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + pre_pos_data[pre_calc_index * kROISize] = y_low * width + x_low; + pre_pos_data[pre_calc_index * kROISize + 1] = y_low * width + x_high; + pre_pos_data[pre_calc_index * kROISize + 2] = y_high * width + x_low; + pre_pos_data[pre_calc_index * kROISize + 3] = y_high * width + x_high; + pre_w_data[pre_calc_index * kROISize] = hy * hx; + pre_w_data[pre_calc_index * kROISize + 1] = hy * lx; + pre_w_data[pre_calc_index * kROISize + 2] = ly * hx; + pre_w_data[pre_calc_index * kROISize + 3] = ly * lx; + pre_calc_index += 1; + } + } + } + } +} + +template +void bilinear_interpolate_gradient(const int height, const int width, T y, T x, + const T out_grad_this_bin, const T count, + T* batch_grad_data) { + int x_low, y_low, x_high, y_high; + T w1, w2, w3, w4; + if (y < -1.0 || y > height || x < -1.0 || x > width) { + w1 = w2 = w3 = w4 = 0; + x_low = x_high = y_low = y_high = -1; + return; + } + y = y <= 0 ? 0 : y; + x = x <= 0 ? 0 : x; + y_low = static_cast(y); + x_low = static_cast(x); + if (y_low >= height - 1) { + y_high = y_low = height - 1; + y = static_cast(y_low); + } else { + y_high = y_low + 1; + } + + if (x_low >= width - 1) { + x_high = x_low = width - 1; + x = static_cast(x_low); + } else { + x_high = x_low + 1; + } + + T ly = y - y_low, lx = x - x_low; + T hy = 1. - ly, hx = 1. - lx; + w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx; + T diff1 = out_grad_this_bin * w1 / count; + T diff2 = out_grad_this_bin * w2 / count; + T diff3 = out_grad_this_bin * w3 / count; + T diff4 = out_grad_this_bin * w4 / count; + if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) { + *(batch_grad_data + y_low * width + x_low) += diff1; + *(batch_grad_data + y_low * width + x_high) += diff2; + *(batch_grad_data + y_high * width + x_low) += diff3; + *(batch_grad_data + y_high * width + x_high) += diff4; + } +} + +template +class CPUROIAlignOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* rois = ctx.Input("ROIs"); + auto* out = ctx.Output("Out"); + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + auto sampling_ratio = ctx.Attr("sampling_ratio"); + + auto& dev_ctx = ctx.template device_context(); + + auto in_dims = in->dims(); + int batch_size = in_dims[0]; + int channels = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + int rois_num = rois->dims()[0]; + + auto in_stride = framework::stride(in_dims); + auto roi_stride = framework::stride(rois->dims()); + auto out_stride = framework::stride(out->dims()); + + const T* input_data = in->data(); + framework::Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(ctx.GetPlace()); + + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + PADDLE_ENFORCE_EQ( + rois_batch_size, batch_size, + "The rois_batch_size and imgs batch_size must be the same."); + int rois_num_with_lod = rois_lod[rois_batch_size]; + PADDLE_ENFORCE_EQ(rois_num, rois_num_with_lod, + "The rois_num from input and lod must be the same."); + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + T* output_data = out->mutable_data(ctx.GetPlace()); + const T* rois_data = rois->data(); + for (int n = 0; n < rois_num; ++n) { + int roi_batch_id = roi_batch_id_data[n]; + T roi_xmin = rois_data[0] * spatial_scale; + T roi_ymin = rois_data[1] * spatial_scale; + T roi_xmax = rois_data[2] * spatial_scale; + T roi_ymax = rois_data[3] * spatial_scale; + + T roi_width = std::max(roi_xmax - roi_xmin, static_cast(1.)); + T roi_height = std::max(roi_ymax - roi_ymin, static_cast(1.)); + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + const T* batch_data = input_data + roi_batch_id * in_stride[0]; + + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); + int roi_bin_grid_w = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_width / pooled_width); + const T count = roi_bin_grid_h * roi_bin_grid_w; + Tensor pre_pos; + Tensor pre_w; + int pre_size = count * out_stride[1]; + pre_pos.Resize({pre_size, kROISize}); + pre_w.Resize({pre_size, kROISize}); + + PreCalcForBilinearInterpolate( + dev_ctx, height, width, pooled_height, pooled_width, roi_bin_grid_h, + roi_bin_grid_w, roi_ymin, roi_xmin, bin_size_h, bin_size_w, + roi_bin_grid_h, roi_bin_grid_w, &pre_pos, &pre_w); + const int* pre_pos_data = pre_pos.data(); + const T* pre_w_data = pre_w.data(); + for (int c = 0; c < channels; c++) { + int pre_calc_index = 0; + for (int ph = 0; ph < pooled_height; ph++) { + for (int pw = 0; pw < pooled_width; pw++) { + const int pool_index = ph * pooled_width + pw; + T output_val = 0; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + for (int i = 0; i < kROISize; i++) { + int pos = pre_pos_data[pre_calc_index * kROISize + i]; + T w = pre_w_data[pre_calc_index * kROISize + i]; + output_val += w * batch_data[pos]; + } + pre_calc_index += 1; + } + } + output_val /= count; + output_data[pool_index] = output_val; + } + } + batch_data += in_stride[1]; + output_data += out_stride[1]; + } + rois_data += roi_stride[0]; + } + } +}; + +template +class CPUROIAlignGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* in = ctx.Input("X"); + auto* rois = ctx.Input("ROIs"); + auto* out_grad = + ctx.Input(framework::GradVarName("Out")); + auto* in_grad = ctx.Output(framework::GradVarName("X")); + + auto pooled_height = ctx.Attr("pooled_height"); + auto pooled_width = ctx.Attr("pooled_width"); + auto spatial_scale = ctx.Attr("spatial_scale"); + auto sampling_ratio = ctx.Attr("sampling_ratio"); + auto in_dims = in->dims(); + if (!in_grad) { + return; + } + int channels = in_dims[1]; + int height = in_dims[2]; + int width = in_dims[3]; + int rois_num = rois->dims()[0]; + Tensor roi_batch_id_list; + roi_batch_id_list.Resize({rois_num}); + int* roi_batch_id_data = + roi_batch_id_list.mutable_data(ctx.GetPlace()); + + auto rois_lod = rois->lod().back(); + int rois_batch_size = rois_lod.size() - 1; + for (int n = 0; n < rois_batch_size; ++n) { + for (size_t i = rois_lod[n]; i < rois_lod[n + 1]; ++i) { + roi_batch_id_data[i] = n; + } + } + + const T* rois_data = rois->data(); + const T* out_grad_data = out_grad->data(); + T* in_grad_data = in_grad->mutable_data(ctx.GetPlace()); + + auto in_stride = framework::stride(in->dims()); + auto roi_stride = framework::stride(rois->dims()); + auto out_stride = framework::stride(out_grad->dims()); + + for (int n = 0; n < rois_num; ++n) { + int roi_batch_idx = roi_batch_id_data[n]; + T roi_xmin = rois_data[0] * spatial_scale; + T roi_ymin = rois_data[1] * spatial_scale; + T roi_xmax = rois_data[2] * spatial_scale; + T roi_ymax = rois_data[3] * spatial_scale; + T roi_width = std::max(roi_xmax - roi_xmin, static_cast(1.)); + T roi_height = std::max(roi_ymax - roi_ymin, static_cast(1.)); + T bin_size_h = static_cast(roi_height) / static_cast(pooled_height); + T bin_size_w = static_cast(roi_width) / static_cast(pooled_width); + for (int c = 0; c < channels; ++c) { + T* batch_grad_data = + in_grad_data + roi_batch_idx * in_stride[0] + c * in_stride[1]; + const T* batch_out_grad_data = + out_grad_data + n * out_stride[0] + c * out_stride[1]; + for (int ph = 0; ph < pooled_height; ++ph) { + for (int pw = 0; pw < pooled_width; ++pw) { + int pool_index = ph * pooled_width + pw; + T out_grad_this_bin = batch_out_grad_data[pool_index]; + int roi_bin_grid_h = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_height / pooled_height); + int roi_bin_grid_w = (sampling_ratio > 0) + ? sampling_ratio + : ceil(roi_width / pooled_width); + T count = roi_bin_grid_h * roi_bin_grid_w; + for (int iy = 0; iy < roi_bin_grid_h; iy++) { + const T y = roi_ymin + ph * bin_size_h + + static_cast(iy + .5f) * bin_size_h / + static_cast(roi_bin_grid_h); + for (int ix = 0; ix < roi_bin_grid_w; ix++) { + const T x = roi_xmin + pw * bin_size_w + + static_cast(ix + .5f) * bin_size_w / + static_cast(roi_bin_grid_w); + bilinear_interpolate_gradient(height, width, y, x, + out_grad_this_bin, count, + batch_grad_data); + } + } + } + } + } + rois_data += roi_stride[0]; + } + } +}; +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/roi_pool_op.cc b/paddle/fluid/operators/roi_pool_op.cc index d6d209d5de041500a9b4893d70800a58e8ee1e1d..8e29761ec208764e263e357a0b3c9456c932d093 100644 --- a/paddle/fluid/operators/roi_pool_op.cc +++ b/paddle/fluid/operators/roi_pool_op.cc @@ -174,4 +174,4 @@ REGISTER_OP_CPU_KERNEL( REGISTER_OP_CPU_KERNEL( roi_pool_grad, ops::CPUROIPoolGradOpKernel, - ops::CPUROIPoolOpKernel); + ops::CPUROIPoolGradOpKernel); diff --git a/paddle/fluid/operators/roi_pool_op.cu b/paddle/fluid/operators/roi_pool_op.cu index 46e20285db6d7acd39dead3994409645adddf494..75c3dd6bc498e35c6249f79a1c24cfe17316670e 100644 --- a/paddle/fluid/operators/roi_pool_op.cu +++ b/paddle/fluid/operators/roi_pool_op.cu @@ -249,4 +249,4 @@ REGISTER_OP_CUDA_KERNEL( REGISTER_OP_CUDA_KERNEL( roi_pool_grad, ops::GPUROIPoolGradOpKernel, - ops::GPUROIPoolOpKernel); + ops::GPUROIPoolGradOpKernel); diff --git a/paddle/fluid/operators/save_op.cc b/paddle/fluid/operators/save_op.cc index e79cffcf498c52ed14db235f6221cfdf08399c9d..0dcf3f0e372f07370078553465973edfd7c96e07 100644 --- a/paddle/fluid/operators/save_op.cc +++ b/paddle/fluid/operators/save_op.cc @@ -110,7 +110,7 @@ class SaveOp : public framework::OperatorBase { lt_var != nullptr, "Can not find variable kLookupTablePath for SaveSelectedRows"); std::string filename = lt_var->data(); - VLOG(4) << "SaveSelectedRows get File name: " << filename; + VLOG(40) << "SaveSelectedRows get File name: " << filename; MkDirRecursively(DirName(filename).c_str()); diff --git a/paddle/fluid/operators/scale_op.cu b/paddle/fluid/operators/scale_op.cu index 04c802da12958a53626f533833c2709110531136..349f39360b8e3100a7f844d3e2d3768053c37c58 100644 --- a/paddle/fluid/operators/scale_op.cu +++ b/paddle/fluid/operators/scale_op.cu @@ -13,6 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/scale_op.h" +#include "paddle/fluid/platform/float16.h" +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( scale, @@ -20,4 +22,6 @@ REGISTER_OP_CUDA_KERNEL( paddle::operators::ScaleKernel, paddle::operators::ScaleKernel, paddle::operators::ScaleKernel); + int64_t>, + paddle::operators::ScaleKernel); diff --git a/paddle/fluid/operators/scale_op.h b/paddle/fluid/operators/scale_op.h index d8a199bc2b860515645b4954b49d8eb59fbd02dc..96b8b00b429df72569ef2a292c8a600c56159f19 100644 --- a/paddle/fluid/operators/scale_op.h +++ b/paddle/fluid/operators/scale_op.h @@ -24,19 +24,13 @@ class ScaleKernel : public framework::OpKernel { public: virtual void Compute(const framework::ExecutionContext& ctx) const { auto* in_var = ctx.InputVar("X"); - auto* in = ctx.Input("X"); - - auto* out_var = ctx.OutputVar("Out"); - auto* out = ctx.Output("Out"); - out->mutable_data(in->place()); - - PADDLE_ENFORCE_EQ(in->dims(), out->dims(), - "in and out should have the same dim"); + auto* in = framework::GetLoDTensorOrSelectedRowsValueFromVar(*in_var); auto scale = static_cast(ctx.Attr("scale")); auto bias = static_cast(ctx.Attr("bias")); auto bias_after_scale = ctx.Attr("bias_after_scale"); + auto* out_var = ctx.OutputVar("Out"); if (in_var->IsType() && in_var != out_var) { auto& in_slr = in_var->Get(); auto* out_slr = out_var->GetMutable(); @@ -44,6 +38,13 @@ class ScaleKernel : public framework::OpKernel { out_slr->set_height(in_slr.height()); } + auto* out = + framework::GetMutableLoDTensorOrSelectedRowsValueFromVar(out_var); + out->mutable_data(in->place()); + + PADDLE_ENFORCE_EQ(in->dims(), out->dims(), + "in and out should have the same dim"); + auto eigen_out = framework::EigenVector::Flatten(*out); auto eigen_in = framework::EigenVector::Flatten(*in); auto& dev = *ctx.template device_context().eigen_device(); diff --git a/paddle/fluid/operators/scatter.cu.h b/paddle/fluid/operators/scatter.cu.h index ac7d69bfb549fd98c76fcf834e8d3ad9bec2ef23..b2e79f6c82bb748293f4219845e6798347c8c46e 100644 --- a/paddle/fluid/operators/scatter.cu.h +++ b/paddle/fluid/operators/scatter.cu.h @@ -51,7 +51,8 @@ void GPUScatterAssign(const platform::DeviceContext& ctx, const Tensor& src, const Tensor& index, Tensor* output) { // PADDLE_ENFORCE(platform::is_gpu_place(place)); // check index of shape 1-D - PADDLE_ENFORCE(index.dims().size() == 1); + PADDLE_ENFORCE(index.dims().size() == 1 || + (index.dims().size() == 2 && index.dims()[1] == 1)); int index_size = index.dims()[0]; auto src_dims = src.dims(); diff --git a/paddle/fluid/operators/scatter.h b/paddle/fluid/operators/scatter.h index 39af717615c01f5c121e32b176b74d05be738531..8bae6606c94620ab4fa8ae34f69236e7e87e9670 100644 --- a/paddle/fluid/operators/scatter.h +++ b/paddle/fluid/operators/scatter.h @@ -37,7 +37,8 @@ void ScatterAssign(const platform::DeviceContext& ctx, const Tensor& src, const Tensor& index, Tensor* output) { PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace())); // check index of shape 1-D - PADDLE_ENFORCE(index.dims().size() == 1); + PADDLE_ENFORCE(index.dims().size() == 1 || + (index.dims().size() == 2 && index.dims()[1] == 1)); int index_size = index.dims()[0]; auto src_dims = src.dims(); diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index 40404295266899c6ac2f7b1e08fdf7db40958794..02ca107ca35348df1827805e40730acd39f39e87 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -39,14 +39,15 @@ class SendBarrierOp : public framework::OperatorBase { std::vector eps = Attr>("endpoints"); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); - VLOG(3) << "SendBarrierOp sync"; + VLOG(30) << "SendBarrierOp sync"; // need to wait before sending send_barrier message PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); for (auto& ep : eps) { - VLOG(3) << "send barrier, ep: " << ep; + VLOG(30) << "send barrier, ep: " << ep; rpc_client->AsyncSendBatchBarrier(ep); } PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); @@ -67,6 +68,7 @@ This operator will send a send barrier signal to list_and_serv op, so that the Parameter Server would knew all variables have been sent. )DOC"); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr>("endpoints", "(string vector, default 127.0.0.1:6164)" "Server endpoints to send variables to.") diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index 48322ac7fd54a2e4cc3405a2c4dcddfc273f5a66..0ad43d56d3cd7500290dc1e386a2dbaf4453a191 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -44,15 +44,16 @@ class SendOp : public framework::OperatorBase { auto& ctx = *pool.Get(place); distributed::RPCClient* rpc_client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance( + Attr("trainer_id")); std::vector rets; for (size_t i = 0; i < ins.size(); i++) { if (NeedSend(scope, ins[i])) { - VLOG(3) << "sending " << ins[i] << " to " << epmap[i]; + VLOG(30) << "sending " << ins[i] << " to " << epmap[i]; rets.push_back(rpc_client->AsyncSendVar(epmap[i], ctx, scope, ins[i])); } else { - VLOG(3) << "don't send no-initialied variable: " << ins[i]; + VLOG(30) << "don't send no-initialied variable: " << ins[i]; } } if (sync_send) { @@ -79,6 +80,7 @@ This operator will send variables to listen_and_serve op at the parameter server "(int, default 0)" "sync send or async send.") .SetDefault(0); + AddAttr("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0); AddAttr>("epmap", "(string vector, default 127.0.0.1:6164)" "Server endpoints in the order of input " diff --git a/paddle/fluid/operators/send_recv_op_test.cc b/paddle/fluid/operators/send_recv_op_test.cc index aee6180add5708d31f7ce927b37c4524a291fe3c..d79b16e3cca714d44c88834082cea9367480da9a 100644 --- a/paddle/fluid/operators/send_recv_op_test.cc +++ b/paddle/fluid/operators/send_recv_op_test.cc @@ -120,7 +120,7 @@ void AddOp(const std::string &type, const f::VariableNameMap &inputs, void StartServerNet(bool is_sparse, std::atomic *initialized) { f::Scope scope; p::CPUPlace place; - VLOG(4) << "before init tensor"; + VLOG(40) << "before init tensor"; if (is_sparse) { InitSelectedRowsInScope(place, &scope); } else { @@ -146,7 +146,7 @@ void StartServerNet(bool is_sparse, std::atomic *initialized) { attrs.insert({"PrefetchBlock", prefetch_block}); attrs.insert({"grad_to_block_id", std::vector({""})}); attrs.insert({"sync_mode", true}); - VLOG(4) << "before init op"; + VLOG(40) << "before init op"; listen_and_serv_op = f::OpRegistry::CreateOp("listen_and_serv", {{"X", {"x1"}}}, {}, attrs); *initialized = true; diff --git a/paddle/fluid/operators/sequence_concat_op.cc b/paddle/fluid/operators/sequence_concat_op.cc index 397a3182953e3f1afaeadeff6d53a4f22fb95d26..3234b60861da3d0c6a8434eb11fd0488a95e171f 100644 --- a/paddle/fluid/operators/sequence_concat_op.cc +++ b/paddle/fluid/operators/sequence_concat_op.cc @@ -90,11 +90,13 @@ REGISTER_OPERATOR(sequence_concat, paddle::framework::OperatorWithKernel, paddle::framework::DefaultGradOpDescMaker); template using Kernel = op::SeqConcatKernel; -REGISTER_OP_CPU_KERNEL(sequence_concat, Kernel, Kernel); +REGISTER_OP_CPU_KERNEL(sequence_concat, Kernel, Kernel, + Kernel); + REGISTER_OPERATOR(sequence_concat_grad, paddle::framework::OperatorWithKernel, op::SeqConcatGradShapeInferer); template using GradKernel = op::SeqConcatGradKernel; REGISTER_OP_CPU_KERNEL(sequence_concat_grad, GradKernel, - GradKernel); + GradKernel, GradKernel); diff --git a/paddle/fluid/operators/sequence_concat_op.h b/paddle/fluid/operators/sequence_concat_op.h index 33e9babff274af888b84d33c991cc0a5b70333ae..ff035f421c4907ba940b973b3fd2a9421ed2dbae 100644 --- a/paddle/fluid/operators/sequence_concat_op.h +++ b/paddle/fluid/operators/sequence_concat_op.h @@ -17,7 +17,7 @@ #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" -#include "paddle/fluid/operators/math/concat.h" +#include "paddle/fluid/operators/math/concat_and_split.h" namespace paddle { namespace operators { @@ -106,7 +106,7 @@ class SeqConcatGradKernel : public framework::OpKernel { } } - math::ConcatGradFunctor functor; + math::SplitFunctor functor; std::vector sliced_x_ptr; std::vector sliced_dx_ptr; for (auto &x : sliced_x) { diff --git a/paddle/fluid/operators/sequence_mask_op.h b/paddle/fluid/operators/sequence_mask_op.h index 18acb735cecabd1e01f7821c880fd8ed5e52971f..7ff68f9c715e4c7243afe9de84af9474e7e4e260 100644 --- a/paddle/fluid/operators/sequence_mask_op.h +++ b/paddle/fluid/operators/sequence_mask_op.h @@ -127,7 +127,7 @@ class SequenceMaskKernel : public framework::OpKernel { auto x_numel = x->numel(); if (maxlen < 0) { #ifdef __NVCC__ - VLOG(10) + VLOG(100) << "SequenceMaskOp on GPU may be slow when maxlen is not provided."; maxlen = static_cast( thrust::reduce(thrust::device_pointer_cast(x_data), diff --git a/paddle/fluid/operators/sequence_pool_op.cc b/paddle/fluid/operators/sequence_pool_op.cc index 15d3f064eb7b025dc9a85b2aabad24186061cbd4..217bb1610fd3f02f0f72d3b7750ebcdfad243f48 100644 --- a/paddle/fluid/operators/sequence_pool_op.cc +++ b/paddle/fluid/operators/sequence_pool_op.cc @@ -47,6 +47,7 @@ class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker { "(Tensor) This tensor is used for the sequence max-pooling " "to record the max indexes.") .AsIntermediate(); + AddAttr("is_test", "").SetDefault(false); AddAttr( "pooltype", "(string, default 'AVERAGE') the pooling pooltype of SequencePoolOp.") diff --git a/paddle/fluid/operators/sequence_pool_op.h b/paddle/fluid/operators/sequence_pool_op.h index 2aa20792f24305a106c500a3d7a6e3d363bc31d8..f2e4a55dee49664b2fc09813f6dba5f68aaf11d5 100644 --- a/paddle/fluid/operators/sequence_pool_op.h +++ b/paddle/fluid/operators/sequence_pool_op.h @@ -32,10 +32,6 @@ class SequencePoolKernel : public framework::OpKernel { auto* in = context.Input("X"); auto* out = context.Output("Out"); std::string pooltype = context.Attr("pooltype"); - Tensor* index = nullptr; - if (pooltype == "MAX") { - index = context.Output("MaxIndex"); - } auto dims = in->dims(); auto lod = in->lod(); @@ -48,13 +44,22 @@ class SequencePoolKernel : public framework::OpKernel { dims[0] = lod[0].size() - 1; out->Resize({dims}); out->mutable_data(context.GetPlace()); - if (pooltype == "MAX") { + Tensor* index = nullptr; + + const bool is_test = context.Attr("is_test"); + + // Do not create index buffer for inference (is_test) mode + // TODO(jczaja): Skip index buffer creation for other devices eg. GPU + if (pooltype == "MAX" && + (is_test == false || + platform::is_cpu_place(context.GetPlace()) == false)) { + index = context.Output("MaxIndex"); index->Resize({dims}); index->mutable_data(context.GetPlace()); } math::SequencePoolFunctor pool; pool(context.template device_context(), pooltype, *in, out, - index); + is_test, index); } }; diff --git a/paddle/fluid/operators/sequence_reverse_op.cc b/paddle/fluid/operators/sequence_reverse_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..1428cca1a6bf6150594f9cb72dbf00cd0eff7df5 --- /dev/null +++ b/paddle/fluid/operators/sequence_reverse_op.cc @@ -0,0 +1,29 @@ +// 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/sequence_reverse_op.h" + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(sequence_reverse, ops::SequenceReverseOp, + ops::SequenceReverseOpMaker, + ops::SequenceReverseGradOpDescMaker); + +REGISTER_OP_CPU_KERNEL( + sequence_reverse, + ops::SequenceReverseOpKernel, + ops::SequenceReverseOpKernel, + ops::SequenceReverseOpKernel, + ops::SequenceReverseOpKernel, + ops::SequenceReverseOpKernel); diff --git a/paddle/fluid/operators/sequence_reverse_op.cu b/paddle/fluid/operators/sequence_reverse_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..ce65f4799e8661adca60d212eaa9c3f0f92c4c29 --- /dev/null +++ b/paddle/fluid/operators/sequence_reverse_op.cu @@ -0,0 +1,25 @@ +// 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/sequence_reverse_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_CUDA_KERNEL( + sequence_reverse, + ops::SequenceReverseOpKernel, + ops::SequenceReverseOpKernel, + ops::SequenceReverseOpKernel, + ops::SequenceReverseOpKernel, + ops::SequenceReverseOpKernel); diff --git a/paddle/fluid/operators/sequence_reverse_op.h b/paddle/fluid/operators/sequence_reverse_op.h new file mode 100644 index 0000000000000000000000000000000000000000..39dad2311b2bcf29f808723caf7bfaef4c88cef2 --- /dev/null +++ b/paddle/fluid/operators/sequence_reverse_op.h @@ -0,0 +1,157 @@ +// 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 "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/algorithm.h" +#include "paddle/fluid/platform/for_range.h" + +namespace paddle { +namespace operators { + +class SequenceReverseOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must exist"); + PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) must exist"); + + auto x_dim = ctx->GetInputDim("X"); + PADDLE_ENFORCE_GE(x_dim.size(), 2, + "Rank of Input(X) must be not less than 2."); + + ctx->SetOutputDim("Y", x_dim); + ctx->ShareLoD("X", "Y"); + } +}; + +class SequenceReverseOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "The input LoDTensor of sequence_reverse op."); + AddOutput("Y", "The output LoDTensor of sequence_reverse op."); + AddComment(R"DOC( +SequenceReverse Operator. + +Reverse each sequence in input X along dim 0. + +Assuming X is a LoDTensor with dims [5, 4] and lod [[0, 2, 5]], where: + +X.data() = [ + [1, 2, 3, 4], + [5, 6, 7, 8], # the 0-th sequence with length 2 + [9, 10, 11, 12], + [13, 14, 15, 16], + [17, 18, 19, 20] # the 1-st sequence with length 3 +] + +The output Y would be a LoDTensor sharing the same dims and lod with input X, +and: + +Y.data() = [ + [5, 6, 7, 8], + [1, 2, 3, 4], # the reversed 0-th sequence with length 2 + [17, 18, 19, 20], + [13, 14, 15, 16], + [9, 10, 11, 12] # the reversed 1-st sequence with length 3 +] + +This Operator is useful to build a reverse dynamic RNN network. + +This Operator only supports one-level lod currently. + )DOC"); + } +}; + +template +struct SequenceReverseFunctor { + SequenceReverseFunctor(const T *x, T *y, const size_t *lod, size_t lod_count, + size_t row_numel) + : x_(x), y_(y), lod_(lod), lod_count_(lod_count), row_numel_(row_numel) {} + + HOSTDEVICE void operator()(size_t idx_x) const { + auto row_idx_x = idx_x / row_numel_; + auto lod_idx = math::UpperBound(lod_, lod_count_, row_idx_x); + auto row_idx_y = lod_[lod_idx - 1] + (lod_[lod_idx] - 1 - row_idx_x); + auto idx_y = row_idx_y * row_numel_ + idx_x % row_numel_; + y_[idx_y] = x_[idx_x]; + } + + const T *x_; + T *y_; + const size_t *lod_; + size_t lod_count_; + size_t row_numel_; +}; + +template +class SequenceReverseOpKernel : public framework::OpKernel { + using LoDTensor = framework::LoDTensor; + + public: + void Compute(const framework::ExecutionContext &ctx) const override { + auto &x = *ctx.Input("X"); + auto *y = ctx.Output("Y"); + + PADDLE_ENFORCE_EQ(x.lod().size(), 1, + "SequenceReverse Op only support one level lod."); + + auto &dev_ctx = ctx.template device_context(); + const size_t *lod; + size_t lod_count = x.lod()[0].size(); + +#ifdef PADDLE_WITH_CUDA + if (platform::is_gpu_place(ctx.GetPlace())) { + lod = x.lod()[0].CUDAData(ctx.GetPlace()); + } else { +#endif + lod = x.lod()[0].data(); +#ifdef PADDLE_WITH_CUDA + } +#endif + + size_t limit = static_cast(x.numel()); + size_t row_numel = static_cast(limit / x.dims()[0]); + auto *x_data = x.data(); + auto *y_data = y->mutable_data(ctx.GetPlace()); + + PADDLE_ENFORCE_NE(x_data, y_data, + "SequenceReverse Op does not support in-place operation"); + + SequenceReverseFunctor functor(x_data, y_data, lod, lod_count, + row_numel); + platform::ForRange for_range(dev_ctx, limit); + for_range(functor); + } +}; + +class SequenceReverseGradOpDescMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + std::unique_ptr op(new framework::OpDesc()); + op->SetType("sequence_reverse"); + op->SetInput("X", OutputGrad("Y")); + op->SetOutput("Y", InputGrad("X")); + op->SetAttrMap(Attrs()); + return op; + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/sequence_unpad_op.cc b/paddle/fluid/operators/sequence_unpad_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..e633e378a226ece8adea2e150cc6c1e9aa874331 --- /dev/null +++ b/paddle/fluid/operators/sequence_unpad_op.cc @@ -0,0 +1,153 @@ +/* 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/sequence_unpad_op.h" + +namespace paddle { +namespace operators { + +class SequenceUnpadOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + protected: + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceUnpadOp should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Length"), + "Input(Length) of SequenceUnpadOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SequenceUnpadOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_GE(x_dims.size(), 2, + "The rank of Input(X) can't be less than 2."); + + auto len_dims = ctx->GetInputDim("Length"); + PADDLE_ENFORCE(len_dims.size() == 2 && len_dims[1] == 1, + "The shape of Input(Length) should be [batch_size, 1]."); + PADDLE_ENFORCE( + len_dims[0] == x_dims[0], + "Input(X) and Input(Length) should have the same first dimension."); + + int64_t out_dim_0 = -1; + if (ctx->IsRuntime()) { + out_dim_0 = x_dims[0] * x_dims[1]; + } + + std::vector out_dims_vec{out_dim_0}; + if (x_dims.size() == 2) { + out_dims_vec.push_back(1); + } else { + for (int i = 2; i < x_dims.size(); ++i) { + out_dims_vec.push_back(x_dims[i]); + } + } + ctx->SetOutputDim("Out", framework::make_ddim(out_dims_vec)); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("X")); + return framework::OpKernelType(data_type, ctx.device_context()); + } +}; + +class SequenceUnpadOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(LoDTensor, default LoDTensor) Input tensor which " + "contains the padded sequences with equal length."); + AddInput("Length", + "(LoDTensor) The input tensor which specifies the actual ength of " + "sequences after unpadding."); + AddOutput( + "Out", + "(LoDTensor) The output tensor which contains unpadded sequences."); + AddComment(R"DOC( + Sequence Unpad Operator + + This operator removes the padding data in the input sequences and convert + them into sequences with actual length as output, identitied by lod + information. + + Example: + + Given input tensor Input(X): + X.data = [[ 1.0, 2.0, 3.0, 4.0, 5.0], + [ 6.0, 7.0, 8.0, 9.0, 10.0], + [11.0, 12.0, 13.0, 14.0, 15.0]], +` + in which there are 3 sequences padded to length 5, and the acutal length + specified by Input(Length): + + Length.data = [[2], [3], [4]], + + after unpadding, Output(Out) will be: + + Out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]] + Out.lod = [[0, 2, 5, 9]] + + )DOC"); + } +}; + +class SequenceUnpadGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SequenceUnpadGradOp should not be null."); + PADDLE_ENFORCE( + ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) of SequenceUnpadGradOp should not be null."); + + if (ctx->HasOutput(framework::GradVarName("X"))) { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ framework::GradVarName("X")); + } + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("X")); + return framework::OpKernelType(data_type, ctx.device_context()); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(sequence_unpad, ops::SequenceUnpadOp, + ops::SequenceUnpadOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(sequence_unpad_grad, ops::SequenceUnpadGradOp); +REGISTER_OP_CPU_KERNEL( + sequence_unpad, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel); +REGISTER_OP_CPU_KERNEL( + sequence_unpad_grad, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel); diff --git a/paddle/fluid/operators/sequence_unpad_op.cu b/paddle/fluid/operators/sequence_unpad_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..75248372237ec2cb23122f6b16e64f6ce750ebf9 --- /dev/null +++ b/paddle/fluid/operators/sequence_unpad_op.cu @@ -0,0 +1,30 @@ +/* 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/sequence_unpad_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + sequence_unpad, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel, + ops::SequenceUnpadOpKernel); +REGISTER_OP_CUDA_KERNEL( + sequence_unpad_grad, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel, + ops::SequenceUnpadGradOpKernel); diff --git a/paddle/fluid/operators/sequence_unpad_op.h b/paddle/fluid/operators/sequence_unpad_op.h new file mode 100644 index 0000000000000000000000000000000000000000..07df3dca831d7e646050ae57402c1a493c2e50e9 --- /dev/null +++ b/paddle/fluid/operators/sequence_unpad_op.h @@ -0,0 +1,104 @@ +/* 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 "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/memory/memcpy.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/operators/math/sequence_padding.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using LoD = framework::LoD; + +template +class SequenceUnpadOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x_t = ctx.Input("X"); + auto* len_t = ctx.Input("Length"); + auto* out_t = ctx.Output("Out"); + out_t->mutable_data(ctx.GetPlace()); + + const int64_t* seq_len_ptr = nullptr; + if (platform::is_gpu_place(ctx.GetPlace())) { + LoDTensor seq_len_cpu; + seq_len_cpu.Resize(len_t->dims()); + seq_len_ptr = seq_len_cpu.mutable_data(platform::CPUPlace()); + framework::TensorCopy(*len_t, platform::CPUPlace(), + ctx.template device_context(), + &seq_len_cpu); + } else { + seq_len_ptr = len_t->data(); + } + + size_t batch_size = x_t->dims()[0]; + std::vector out_lod0(batch_size + 1, 0); + for (size_t i = 0; i < batch_size; ++i) { + out_lod0[i + 1] = out_lod0[i] + seq_len_ptr[i]; + } + + framework::LoD out_lod; + out_lod.push_back(out_lod0); + out_t->set_lod(out_lod); + + std::vector out_dims_vec{static_cast(out_lod0.back())}; + if (x_t->dims().size() == 2) { + out_dims_vec.push_back(1); + } else { + for (int i = 2; i < x_t->dims().size(); ++i) { + out_dims_vec.push_back(x_t->dims()[i]); + } + } + out_t->Resize(framework::make_ddim(out_dims_vec)); + + int64_t padded_length = x_t->dims()[1]; + math::UnpaddingLoDTensorFunctor()( + ctx.template device_context(), *x_t, out_t, + padded_length, 0, false, math::kBatchLengthWidth); + } +}; + +template +class SequenceUnpadGradOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* d_x = ctx.Output(framework::GradVarName("X")); + if (d_x) { + const auto* d_out = ctx.Input(framework::GradVarName("Out")); + const auto* x_t = ctx.Input("X"); + d_x->mutable_data(ctx.GetPlace()); + + int padded_length = x_t->dims()[1]; + + LoDTensor zero_pads; + zero_pads.Resize({1, 1}); + zero_pads.mutable_data(ctx.GetPlace()); + math::SetConstant set_zero; + auto& dev_ctx = ctx.template device_context(); + set_zero(dev_ctx, &zero_pads, static_cast(0)); + + math::PaddingLoDTensorFunctor()( + ctx.template device_context(), *d_out, d_x, zero_pads, + padded_length, 0, false, math::kBatchLengthWidth); + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/sgd_op.cc b/paddle/fluid/operators/sgd_op.cc index fef230e42d07a5ed73b7a7a6ab682694675bb9d2..ea62acd08c5009556abf05c91726111870d1a462 100644 --- a/paddle/fluid/operators/sgd_op.cc +++ b/paddle/fluid/operators/sgd_op.cc @@ -21,7 +21,7 @@ class SGDOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("Param"), "Input(Param) of SGDOp should not be null."); PADDLE_ENFORCE(ctx->HasInput("Grad"), @@ -42,7 +42,7 @@ class SGDOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { + const framework::ExecutionContext &ctx) const override { auto data_type = framework::GetDataTypeOfVar(ctx.InputVar("Param")); return framework::OpKernelType(data_type, ctx.device_context()); } @@ -50,17 +50,20 @@ class SGDOp : public framework::OperatorWithKernel { class SGDOpInferVarType : public framework::VarTypeInference { public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { - auto input_var = op_desc.Input("Param")[0]; - for (auto& out_var : op_desc.Output("ParamOut")) { - if (block->FindRecursiveOrCreateVar(input_var).GetType() == - framework::proto::VarType::SELECTED_ROWS) { - block->FindRecursiveOrCreateVar(out_var).SetType( - framework::proto::VarType::SELECTED_ROWS); - } else { - block->FindRecursiveOrCreateVar(out_var).SetType( - framework::proto::VarType::LOD_TENSOR); + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + auto input_var_n = op_desc.Input("Param")[0]; + auto in_var_type = block->FindRecursiveOrCreateVar(input_var_n).GetType(); + PADDLE_ENFORCE(in_var_type == framework::proto::VarType::SELECTED_ROWS || + in_var_type == framework::proto::VarType::LOD_TENSOR, + "The input Var's type should be LoDtensor or SelectedRows," + " but the received var(%s)'s type is %s", + input_var_n, in_var_type); + + for (auto &out_var_n : op_desc.Output("ParamOut")) { + auto &out_var = block->FindRecursiveOrCreateVar(out_var_n); + if (out_var.GetType() != in_var_type) { + out_var.SetType(in_var_type); } } } @@ -74,8 +77,7 @@ class SGDOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("Grad", "(Tensor or SelectedRows) Input gradient"); AddOutput("ParamOut", "(Tensor or SelectedRows, same with Param) " - "Output parameter, should share the same memory with Param") - .Reuse("Param"); + "Output parameter, should share the same memory with Param"); AddComment(R"DOC( SGD operator diff --git a/paddle/fluid/operators/sgd_op.cu b/paddle/fluid/operators/sgd_op.cu index 243609075713305a90dc162991166ba24d54e835..d3f4eba3b24ec1ac0328ef270256cdf3abe499db 100644 --- a/paddle/fluid/operators/sgd_op.cu +++ b/paddle/fluid/operators/sgd_op.cu @@ -56,6 +56,12 @@ template class SGDOpCUDAKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { + const auto* param_var = ctx.InputVar("Param"); + PADDLE_ENFORCE(param_var->IsType(), + "The Var(%s)'s type should be LoDTensor, " + "but the received is %s", + ctx.Inputs("Param").front(), param_var->Type().name()); + auto* param = ctx.Input("Param"); auto* param_out = ctx.Output("ParamOut"); auto* learning_rate = ctx.Input("LearningRate"); diff --git a/paddle/fluid/operators/sgd_op.h b/paddle/fluid/operators/sgd_op.h index d8b0165b2a89b04bd55671a37d96ee4ba275b2eb..2e206c963ea009b436bb03433d30683a29fe83aa 100644 --- a/paddle/fluid/operators/sgd_op.h +++ b/paddle/fluid/operators/sgd_op.h @@ -98,10 +98,10 @@ class SGDOpKernel : public framework::OpKernel { auto param_row_width = param.value().dims()[1]; auto grad_row_width = grad.value().dims()[1]; - VLOG(4) << " param rows: " << param.rows().size() - << " param memory rows: " << param.value().dims()[0] - << " grad rows: " << grad.rows().size() - << " grad memory rows: " << grad.value().dims()[0]; + VLOG(40) << " param rows: " << param.rows().size() + << " param memory rows: " << param.value().dims()[0] + << " grad rows: " << grad.rows().size() + << " grad memory rows: " << grad.value().dims()[0]; PADDLE_ENFORCE_EQ(param_row_width, grad_row_width, "param_row should have the same size with grad_row"); diff --git a/paddle/fluid/operators/sign_op.cc b/paddle/fluid/operators/sign_op.cc index f3985dcc027f974e0213a73ea9a21e268d77615f..6837856a6da804e27af2cd6c83052c04f17140d8 100644 --- a/paddle/fluid/operators/sign_op.cc +++ b/paddle/fluid/operators/sign_op.cc @@ -67,4 +67,5 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(sign, ops::SignOp, ops::SignOpMaker, ops::SignGradMaker); REGISTER_OP_CPU_KERNEL( - sign, ops::SignKernel); + sign, ops::SignKernel, + ops::SignKernel); diff --git a/paddle/fluid/operators/sign_op.cu b/paddle/fluid/operators/sign_op.cu index e0d7a87e6485a74dd1cdee1a05abc42eef460990..817e0fbbd511462f161633242d28e63062676eb9 100644 --- a/paddle/fluid/operators/sign_op.cu +++ b/paddle/fluid/operators/sign_op.cu @@ -13,7 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/sign_op.h" +#include "paddle/fluid/platform/float16.h" REGISTER_OP_CUDA_KERNEL( sign, - paddle::operators::SignKernel); + paddle::operators::SignKernel, + paddle::operators::SignKernel, + paddle::operators::SignKernel); diff --git a/paddle/fluid/operators/similarity_focus_op.cc b/paddle/fluid/operators/similarity_focus_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..9612f82b6d45dc4e08bfe288ddd1c7790875ee4d --- /dev/null +++ b/paddle/fluid/operators/similarity_focus_op.cc @@ -0,0 +1,87 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/similarity_focus_op.h" + +namespace paddle { +namespace operators { +class SimilarityFocusOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor, default Tensor), a 4-D tensor with shape," + " [BatchSize, X, Y, Z]"); + AddOutput("Out", + "(Tensor, default Tensor), the similarity focus mask" + " with the same shape of input X."); + AddAttr("axis", + "(int32), indicating the dimension to be select. It can" + " only be 1, 2, or 3."); + AddAttr>("indexes", + "(std::vector), indicating the indexes" + " of the selected dimension."); + AddComment(R"DOC( +SimilarityFocus Operator. + +Generate a similarity focus mask with the same shape of input using the following method: +1. Extract the 3-D tensor(here the first dimension is BatchSize) corresponding + to the axis according to the indexes. For example, if axis=1 and indexes=[a], + it will get the matrix T=X[:, a, :, :]. In this case, if the shape of input X + is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C). +2. For each index, find the largest numbers in the tensor T, so that the same + row and same column has at most one number(what it means is that if the + largest number has been found in the i-th row and the j-th column, then + the numbers in the i-th row or j-th column will be skipped. And then the + next largest number will be selected from the remaining numbers. Obviously + there will be min(B, C) numbers), and mark the corresponding position of the + 3-D similarity focus mask as 1, otherwise as 0. Do elementwise-or for + each index. +3. Broadcast the 3-D similarity focus mask to the same shape of input X. + +Refer to `Similarity Focus Layer `_ +)DOC"); + } +}; + +class SimilarityFocusOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should be not null."); + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_EQ(x_dims.size(), 4, "Input(X)'s rank should be 4."); + ctx->SetOutputDim("Out", x_dims); + ctx->ShareLoD("X", /*->*/ "Out"); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + platform::CPUPlace()); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(similarity_focus, ops::SimilarityFocusOp, + ops::SimilarityFocusOpMaker, + paddle::framework::EmptyGradOpMaker); +REGISTER_OP_CPU_KERNEL(similarity_focus, ops::SimilarityFocusKernel, + ops::SimilarityFocusKernel); diff --git a/paddle/fluid/operators/similarity_focus_op.h b/paddle/fluid/operators/similarity_focus_op.h new file mode 100644 index 0000000000000000000000000000000000000000..bf3fed2aaf2cf92d5619ae5bce6dd70d9dfe9621 --- /dev/null +++ b/paddle/fluid/operators/similarity_focus_op.h @@ -0,0 +1,168 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include +#include +#include +#include +#include "paddle/fluid/framework/eigen.h" +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { +using Tensor = framework::Tensor; + +template +class SimilarityFocusKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + Tensor* out = context.Output("Out"); + const Tensor* x = context.Input("X"); + T* out_data = out->mutable_data(context.GetPlace()); + const T* x_data = x->data(); + + int axis = context.Attr("axis"); + std::vector indexes = context.Attr>("indexes"); + + int64_t batch_size = x->dims()[0]; + int64_t dim[4]; + for (int i = 1; i <= 3; ++i) { + dim[i] = x->dims()[i]; + } + + if (indexes.size() < 1) { + PADDLE_THROW("Indexes' size can not be 0."); + } + for (auto index : indexes) { + if (dim[axis] < index) { + PADDLE_THROW("Index exceeds tensor shape limit."); + } + } + + int64_t array_size = 1; + for (int i = 1; i <= 3; ++i) { + if (i != axis) { + array_size *= dim[i]; + } + } + + std::vector> array(array_size); + + bool (*cmp)(std::pair, std::pair) = []( + std::pair x, std::pair y) { + return x.first > y.first; + }; + + int64_t (*compute_index)(int64_t*, int, int, int, int) = []( + int64_t* dim, int d1, int d2, int d3, int d4) { + return d1 * dim[1] * dim[2] * dim[3] + d2 * dim[2] * dim[3] + + d3 * dim[3] + d4; + }; + + memset(out_data, 0, sizeof(T) * batch_size * dim[1] * dim[2] * dim[3]); + for (int i = 0; i < batch_size; ++i) { + for (auto index : indexes) { + if (axis == 1) { + for (int j = 0; j < dim[2]; ++j) { + for (int k = 0; k < dim[3]; ++k) { + array[j * dim[3] + k] = std::make_pair( + x_data[compute_index(dim, i, index, j, k)], j * dim[3] + k); + } + } + + std::sort(array.begin(), array.end(), cmp); + int tag_num = 0; + std::vector tag2(dim[2]), tag3(dim[3]); + for (auto x : array) { + int idx2 = x.second / dim[3]; + int idx3 = x.second % dim[3]; + if (tag2[idx2] || tag3[idx3]) { + continue; + } + tag_num++; + tag2[idx2] = true; + tag3[idx3] = true; + for (int j = 0; j < dim[1]; ++j) { + out_data[compute_index(dim, i, j, idx2, idx3)] = 1; + } + if (tag_num == std::min(dim[2], dim[3])) { + break; + } + } + } else if (axis == 2) { + for (int j = 0; j < dim[1]; ++j) { + for (int k = 0; k < dim[3]; ++k) { + array[j * dim[3] + k] = std::make_pair( + x_data[compute_index(dim, i, j, index, k)], j * dim[3] + k); + } + } + + std::sort(array.begin(), array.end(), cmp); + int tag_num = 0; + std::vector tag1(dim[1]), tag3(dim[3]); + for (auto x : array) { + int idx1 = x.second / dim[3]; + int idx3 = x.second % dim[3]; + if (tag1[idx1] || tag3[idx3]) { + continue; + } + tag_num++; + tag1[idx1] = true; + tag3[idx3] = true; + for (int j = 0; j < dim[2]; ++j) { + out_data[compute_index(dim, i, idx1, j, idx3)] = 1; + } + if (tag_num == std::min(dim[1], dim[3])) { + break; + } + } + } else if (axis == 3) { + for (int j = 0; j < dim[1]; ++j) { + for (int k = 0; k < dim[2]; ++k) { + array[j * dim[2] + k] = std::make_pair( + x_data[compute_index(dim, i, j, k, index)], j * dim[2] + k); + } + } + + std::sort(array.begin(), array.end(), cmp); + int tag_num = 0; + std::vector tag1(dim[1]), tag2(dim[2]); + for (auto x : array) { + int idx1 = x.second / dim[2]; + int idx2 = x.second % dim[2]; + if (tag1[idx1] || tag2[idx2]) { + continue; + } + tag_num++; + tag1[idx1] = true; + tag2[idx2] = true; + for (int j = 0; j < dim[3]; ++j) { + out_data[compute_index(dim, i, idx1, idx2, j)] = 1; + } + if (tag_num == std::min(dim[1], dim[2])) { + break; + } + } + } else { + PADDLE_THROW("Axis must be 1 or 2 or 3"); + } + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/softmax_cudnn_op.cu.cc b/paddle/fluid/operators/softmax_cudnn_op.cu.cc index 2bdb23e999621b10799b5163f326bc4b66a437e6..ad3e5543f10ae05865565110ba2231c897c205b8 100644 --- a/paddle/fluid/operators/softmax_cudnn_op.cu.cc +++ b/paddle/fluid/operators/softmax_cudnn_op.cu.cc @@ -76,6 +76,9 @@ namespace ops = paddle::operators; namespace plat = paddle::platform; REGISTER_OP_KERNEL(softmax, CUDNN, plat::CUDAPlace, ops::SoftmaxCUDNNKernel, + ops::SoftmaxCUDNNKernel, ops::SoftmaxCUDNNKernel); REGISTER_OP_KERNEL(softmax_grad, CUDNN, plat::CUDAPlace, - ops::SoftmaxGradCUDNNKernel); + ops::SoftmaxGradCUDNNKernel, + ops::SoftmaxGradCUDNNKernel, + ops::SoftmaxGradCUDNNKernel); diff --git a/paddle/fluid/operators/softmax_op.cc b/paddle/fluid/operators/softmax_op.cc index bb081238820b9ee3ae095442d21cfce11f7b41e5..9e21b6c824bfd7d1c1090e5ba3ba2f6aa9bdb230 100644 --- a/paddle/fluid/operators/softmax_op.cc +++ b/paddle/fluid/operators/softmax_op.cc @@ -80,8 +80,7 @@ class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker { AddInput("X", "The input tensor of softmax, " "whose last dimension is the input_feature_dimensions."); - AddOutput("Out", "The normalized values with the same shape as X.") - .Reuse("X"); + AddOutput("Out", "The normalized values with the same shape as X."); AddAttr( "use_cudnn", "(bool, default false) Only used in cudnn kernel, need install cudnn") @@ -125,6 +124,14 @@ For each row $i$ and each column $j$ in the matrix, we have: } }; +class SoftmaxOpInferVarType : public framework::PassInDtypeAndVarTypeToOutput { + protected: + std::unordered_map GetInputOutputWithSameType() + const override { + return std::unordered_map{{"X", /*->*/ "Out"}}; + } +}; + class SoftmaxOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -197,7 +204,7 @@ class SoftmaxOpGradMaker : public framework::SingleGradOpDescMaker { namespace ops = paddle::operators; REGISTER_OPERATOR(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker, - ops::SoftmaxOpGradMaker); + ops::SoftmaxOpInferVarType, ops::SoftmaxOpGradMaker); REGISTER_OPERATOR(softmax_grad, ops::SoftmaxOpGrad); REGISTER_OP_CPU_KERNEL( softmax, ops::SoftmaxKernel, diff --git a/paddle/fluid/operators/softmax_op.cu.cc b/paddle/fluid/operators/softmax_op.cu.cc index 5fb4f011d9b47cebc4a23bcce47eada825263343..19359b7eef5126d84f0707d39095a74ae4561186 100644 --- a/paddle/fluid/operators/softmax_op.cu.cc +++ b/paddle/fluid/operators/softmax_op.cu.cc @@ -23,4 +23,5 @@ REGISTER_OP_CUDA_KERNEL( ops::SoftmaxKernel); REGISTER_OP_CUDA_KERNEL( softmax_grad, ops::SoftmaxGradKernel, - ops::SoftmaxGradKernel); + ops::SoftmaxGradKernel, + ops::SoftmaxGradKernel); diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cc b/paddle/fluid/operators/softmax_with_cross_entropy_op.cc index 1a9324ec862fc3dd7ce669c5fed94527cac22b8f..2900221485e6ec097796ac38936ce31f8382c86a 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cc +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cc @@ -44,6 +44,12 @@ class SoftmaxWithCrossEntropyOpMaker "(bool, default: false), A flag to indicate whether to interpretate " "the given labels as soft labels.") .SetDefault(false); + AddAttr( + "numeric_stable_mode", + "(bool, default: false), A flag to indicate whether to use more " + "numerically stable algorithm. This flag is only valid when " + "soft_label is false and GPU is used.") + .SetDefault(false); AddAttr( "ignore_index", "(int, default -100), Specifies a target value that is ignored and" diff --git a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu index a07c17348ebb3f768d1c8be65c2d31e3c130bd23..6d48796191dd13a45f0c7267bfaf05489f528a9d 100644 --- a/paddle/fluid/operators/softmax_with_cross_entropy_op.cu +++ b/paddle/fluid/operators/softmax_with_cross_entropy_op.cu @@ -17,6 +17,7 @@ limitations under the License. */ #include #include "paddle/fluid/operators/math/cross_entropy.h" #include "paddle/fluid/operators/softmax_with_cross_entropy_op.h" +#include "paddle/fluid/platform/for_range.h" namespace paddle { namespace operators { @@ -117,8 +118,8 @@ using BlockReduceTempStorage = typename BlockReduce::TempStorage; // Make sure that BlockDim <= feature_size // This kernel is used to calculate the max element of each row template -__global__ void RowReductionForMax(const T* logits_data, T* max_data, - int feature_size) { +static __global__ void RowReductionForMax(const T* logits_data, T* max_data, + int feature_size) { __shared__ BlockReduceTempStorage temp_storage; auto beg_idx = feature_size * blockIdx.x + threadIdx.x; @@ -141,9 +142,10 @@ __global__ void RowReductionForMax(const T* logits_data, T* max_data, } // Make sure that BlockDim <= feature_size -template -__global__ void RowReductionForDiffMaxSum(const T* logits_data, T* max_data, - T* softmax, int feature_size) { +template +static __global__ void RowReductionForDiffMaxSum(const T* logits_data, + T* max_data, T* softmax, + int feature_size) { __shared__ BlockReduceTempStorage temp_storage; auto beg_idx = feature_size * blockIdx.x + threadIdx.x; @@ -153,24 +155,34 @@ __global__ void RowReductionForDiffMaxSum(const T* logits_data, T* max_data, softmax[beg_idx] = logits_data[beg_idx] - block_max; T diff_max_sum = real_exp(softmax[beg_idx]); - beg_idx += BlockDim; - while (beg_idx < end_idx) { - softmax[beg_idx] = logits_data[beg_idx] - block_max; - diff_max_sum += real_exp(softmax[beg_idx]); - beg_idx += BlockDim; + auto idx = beg_idx + BlockDim; + while (idx < end_idx) { + softmax[idx] = logits_data[idx] - block_max; + diff_max_sum += real_exp(softmax[idx]); + idx += BlockDim; } diff_max_sum = BlockReduce(temp_storage).Reduce(diff_max_sum, cub::Sum()); if (threadIdx.x == 0) max_data[blockIdx.x] = real_log(diff_max_sum); + + if (!CalculateLogSoftmax) return; + __syncthreads(); + diff_max_sum = max_data[blockIdx.x]; + softmax[beg_idx] -= diff_max_sum; + beg_idx += BlockDim; + while (beg_idx < end_idx) { + softmax[beg_idx] -= diff_max_sum; + beg_idx += BlockDim; + } + if (threadIdx.x == 0) max_data[blockIdx.x] = 0; } // Make sure that BlockDim <= feature_size template -__global__ void RowReductionForSoftmaxAndCrossEntropy(const T* logits_data, - const T* labels_data, - T* loss_data, T* softmax, - int feature_size) { +static __global__ void RowReductionForSoftmaxAndCrossEntropy( + const T* logits_data, const T* labels_data, T* loss_data, T* softmax, + int feature_size) { __shared__ BlockReduceTempStorage temp_storage; auto beg_idx = feature_size * blockIdx.x + threadIdx.x; @@ -194,11 +206,134 @@ __global__ void RowReductionForSoftmaxAndCrossEntropy(const T* logits_data, } template -__global__ void SetSoftmaxToOneWhenFeatureSizeIsOne(T* out, int batch_size) { +struct HardLabelSoftmaxWithCrossEntropyFunctor { + public: + HardLabelSoftmaxWithCrossEntropyFunctor(const T* logits, + const int64_t* labels, T* loss, + T* log_softmax, int feature_size) + : logits_(logits), + labels_(labels), + loss_(loss), + log_softmax_(log_softmax), + feature_size_(feature_size) {} + + __device__ void operator()(int idx) const { + auto row_idx = idx / feature_size_; + auto col_idx = idx % feature_size_; + if (col_idx != labels_[row_idx]) { + log_softmax_[idx] = real_exp(log_softmax_[idx]); + } else { + auto softmax = log_softmax_[idx]; + log_softmax_[idx] = real_exp(softmax); + loss_[row_idx] = -softmax; + } + } + + private: + const T* logits_; + const int64_t* labels_; + T* loss_; + T* log_softmax_; + int feature_size_; +}; + +template +struct HardLabelSoftmaxWithCrossEntropyFunctorWithIgnoreIdx { + public: + HardLabelSoftmaxWithCrossEntropyFunctorWithIgnoreIdx(const T* logits, + const int64_t* labels, + T* loss, T* log_softmax, + int feature_size, + int ignore_idx) + : logits_(logits), + labels_(labels), + loss_(loss), + log_softmax_(log_softmax), + feature_size_(feature_size), + ignore_idx_(ignore_idx) {} + + __device__ void operator()(int idx) const { + auto row_idx = idx / feature_size_; + auto col_idx = idx % feature_size_; + if (col_idx != labels_[row_idx] || col_idx == ignore_idx_) { + log_softmax_[idx] = real_exp(log_softmax_[idx]); + } else { + auto softmax = log_softmax_[idx]; + log_softmax_[idx] = real_exp(softmax); + loss_[row_idx] = -softmax; + } + } + + private: + const T* logits_; + const int64_t* labels_; + T* loss_; + T* log_softmax_; + int feature_size_; + int ignore_idx_; +}; + +template +static __global__ void SetSoftmaxToOneWhenFeatureSizeIsOne(T* out, + int batch_size) { auto idx = threadIdx.x + blockIdx.x * blockDim.x; if (idx < batch_size) out[idx] = static_cast(1); } +template +static void HardLabelSoftmaxWithCrossEntropy( + const platform::CUDADeviceContext& ctx, const T* logits_data, + const int64_t* labels_data, T* loss_data, T* softmax_data, int batch_size, + int feature_size, int ignore_idx) { + constexpr int kMaxBlockDim = 512; + int block_dim = feature_size >= kMaxBlockDim + ? kMaxBlockDim + : (1 << static_cast(std::log2(feature_size))); + auto stream = ctx.stream(); + +#define CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(BlockDim) \ + case BlockDim: { \ + RowReductionForMax<<>>( \ + logits_data, loss_data, feature_size); \ + RowReductionForDiffMaxSum<<>>( \ + logits_data, loss_data, softmax_data, feature_size); \ + platform::ForRange for_range( \ + ctx, batch_size* feature_size); \ + if (ignore_idx >= 0 && ignore_idx < feature_size) { \ + for_range(HardLabelSoftmaxWithCrossEntropyFunctorWithIgnoreIdx( \ + logits_data, labels_data, loss_data, softmax_data, feature_size, \ + ignore_idx)); \ + } else { \ + for_range(HardLabelSoftmaxWithCrossEntropyFunctor( \ + logits_data, labels_data, loss_data, softmax_data, feature_size)); \ + } \ + } break + + switch (block_dim) { + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(512); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(256); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(128); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(64); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(32); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(16); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(8); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(4); + CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL(2); + case 1: + SetSoftmaxToOneWhenFeatureSizeIsOne<<<(batch_size + kMaxBlockDim - 1) / + kMaxBlockDim, + kMaxBlockDim, 0, stream>>>( + softmax_data, batch_size); + cudaMemsetAsync(loss_data, 0, batch_size * sizeof(T), stream); + break; + default: + PADDLE_THROW("BlockDim must be 2^n in softmax_with_cross_entropy_op"); + break; + } +#undef CALL_HARD_LABEL_SOFTMAX_WITH_CROSS_ENTROPY_FUSED_KERNEL +} + template static void SoftmaxWithCrossEntropyFusedKernel(const T* logits_data, const T* labels_data, @@ -237,7 +372,7 @@ static void SoftmaxWithCrossEntropyFusedKernel(const T* logits_data, kMaxBlockDim, kMaxBlockDim, 0, stream>>>( softmax_data, batch_size); - cudaMemsetAsync(loss_data, 0, batch_size, stream); + cudaMemsetAsync(loss_data, 0, batch_size * sizeof(T), stream); break; default: PADDLE_THROW("BlockDim must be 2^n in softmax_with_cross_entropy_op"); @@ -272,11 +407,21 @@ class SoftmaxWithCrossEntropyCUDAKernel : public framework::OpKernel { logits_data, labels_data, softmax_data, loss_data, batch_size, feature_size, context.cuda_device_context().stream()); } else { - math::SoftmaxCUDNNFunctor()(context.cuda_device_context(), logits, - softmax); - math::CrossEntropyFunctor()( - context.cuda_device_context(), loss, softmax, labels, false, - ignore_index); + if (!context.Attr("numeric_stable_mode")) { + math::SoftmaxCUDNNFunctor()(context.cuda_device_context(), logits, + softmax); + math::CrossEntropyFunctor()( + context.cuda_device_context(), loss, softmax, labels, false, + ignore_index); + } else { + int batch_size = logits->dims()[0]; + int feature_size = logits->dims()[1]; + auto* logits_data = logits->data(); + auto* labels_data = labels->data(); + HardLabelSoftmaxWithCrossEntropy( + context.cuda_device_context(), logits_data, labels_data, loss_data, + softmax_data, batch_size, feature_size, ignore_index); + } } } }; diff --git a/paddle/fluid/operators/space_to_depth_op.cc b/paddle/fluid/operators/space_to_depth_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..f109dd685c87ab1b0776a855bb5f510eab1f5526 --- /dev/null +++ b/paddle/fluid/operators/space_to_depth_op.cc @@ -0,0 +1,131 @@ +/* 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/space_to_depth_op.h" +#include +#include + +namespace paddle { +namespace operators { + +class SpaceToDepthOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SpaceToDepthOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SpaceToDepthOp should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_EQ(x_dims.size(), 4, "input should be a 4D tensor"); + auto blocksize = ctx->Attrs().Get("blocksize"); + + PADDLE_ENFORCE_GT(blocksize, 1, "The blocksize should be Greater than 1"); + PADDLE_ENFORCE_GT(x_dims[1], 0, "input channel should be Greater than 0"); + PADDLE_ENFORCE_GT(x_dims[2], 0, "input Height should be Greater than 0"); + PADDLE_ENFORCE_GT(x_dims[3], 0, "input Width should be Greater than 0"); + + PADDLE_ENFORCE_EQ(x_dims[1] % (blocksize * blocksize), 0, + "input channel should be divisible of the square of " + "SpaceToDepthOp blocksize"); + PADDLE_ENFORCE_EQ(x_dims[2] % (blocksize), 0, + "input Height should be divisible of the square of " + "SpaceToDepthOp blocksize"); + PADDLE_ENFORCE_EQ(x_dims[3] % (blocksize), 0, + "input Width should be divisible of the square of " + "SpaceToDepthOp blocksize"); + + VLOG(3) << "SpaceToDepthOp operator x.shape=" << x_dims + << "Attribute blocksize" << blocksize << std::endl; + + std::vector output_shape(4, 0); // [B,C,H,W] + output_shape[0] = x_dims[0]; + output_shape[1] = x_dims[1] * blocksize * blocksize; + output_shape[2] = x_dims[2] / blocksize; + output_shape[3] = x_dims[3] / blocksize; + + auto out_dims = framework::make_ddim(output_shape); + + ctx->SetOutputDim("Out", out_dims); + + if (x_dims[0] == out_dims[0]) { + // Only pass LoD when the first dimension of output and Input(X) + // are the same. + ctx->ShareLoD("X", /*->*/ "Out"); + } + } +}; + +class SpaceToDepthOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "(Tensor). The input should be a 4D tensor B * C * W * H of " + "SpaceToDepthOp " + "operator."); + AddOutput("Out", + "(Tensor), The output should be a 4D tensor B * C2 * W2 * H2 of " + "SpaceToDepthOp operator."); + AddAttr( + "blocksize", + "(int64_t, default 2) blocksize used to do change Space To Depth.") + .SetDefault(2) + .GreaterThan(1); + AddComment(R"DOC( + reorg operator used in Yolo v2. + The equation is: C2 = C1/blocksize * blocksize, W2 = W1 ∗ blocksize + offset % blocksize, H2 = H1 ∗ blocksize + offset / blocksize, + + Reshape Input(X) into the shape according to Attr(blocksize). The + data in Input(X) are unchanged. + + Examples: + + 1. Given a 4-D tensor Input(X) with a shape [128, 2048, 26, 26], and the blocksize is 2, the reorg operator will transform Input(X) + into a 4-D tensor with shape [128, 2048, 13, 13] and leaving Input(X)'s data unchanged. + + )DOC"); + } +}; + +class SpaceToDepthGradOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) shouldn't be null."); + PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")), + "Input(Out@GRAD) shouldn't be null."); + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(space_to_depth, ops::SpaceToDepthOp, ops::SpaceToDepthOpMaker, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(space_to_depth_grad, ops::SpaceToDepthGradOp); +REGISTER_OP_CPU_KERNEL( + space_to_depth, + ops::SpaceToDepthKernel, + ops::SpaceToDepthKernel, + ops::SpaceToDepthKernel); +REGISTER_OP_CPU_KERNEL( + space_to_depth_grad, + ops::SpaceToDepthGradKernel, + ops::SpaceToDepthGradKernel, + ops::SpaceToDepthGradKernel); diff --git a/paddle/fluid/operators/space_to_depth_op.cu b/paddle/fluid/operators/space_to_depth_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..38d0a662733222386b8ecd68d064f3d1abe56c3b --- /dev/null +++ b/paddle/fluid/operators/space_to_depth_op.cu @@ -0,0 +1,30 @@ +// 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/space_to_depth_op.h" + +namespace plat = paddle::platform; +namespace ops = paddle::operators; + +REGISTER_OP_CUDA_KERNEL( + space_to_depth, + ops::SpaceToDepthKernel, + ops::SpaceToDepthKernel, + ops::SpaceToDepthKernel); + +REGISTER_OP_CUDA_KERNEL( + space_to_depth_grad, + ops::SpaceToDepthGradKernel, + ops::SpaceToDepthGradKernel, + ops::SpaceToDepthGradKernel); diff --git a/paddle/fluid/operators/space_to_depth_op.h b/paddle/fluid/operators/space_to_depth_op.h new file mode 100644 index 0000000000000000000000000000000000000000..a71662b4813ab27b65f5c7a918e2bb6fb15a1993 --- /dev/null +++ b/paddle/fluid/operators/space_to_depth_op.h @@ -0,0 +1,127 @@ +/* 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. */ +#ifndef PADDLE_FLUID_OPERATORS_SPACE_TO_DEPTH_OP_H_ +#define PADDLE_FLUID_OPERATORS_SPACE_TO_DEPTH_OP_H_ +#endif // PADDLE_FLUID_OPERATORS_SPACE_TO_DEPTH_OP_H_ + +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/platform/for_range.h" + +namespace paddle { +namespace operators { + +template +class space_to_depth_compute { + public: + HOSTDEVICE space_to_depth_compute(const T *x, int64_t w, int64_t h, int64_t c, + int64_t batch, int64_t blocksize, + int64_t forward, T *out) + : x_(x), + w_(w), + h_(h), + c_(c), + batch_(batch), + blocksize_(blocksize), + forward_(forward), + out_(out) {} + + HOSTDEVICE void operator()(int64_t in_index) { + int64_t out_c = c_ / (blocksize_ * blocksize_); + // calculate each dim position with index of tensor + int64_t b = in_index / (c_ * h_ * w_); + int64_t k = (in_index % (c_ * h_ * w_)) / (h_ * w_); + int64_t j = ((in_index % (c_ * h_ * w_)) % (h_ * w_)) / w_; + int64_t i = ((in_index % (c_ * h_ * w_)) % (h_ * w_)) % w_; + + int64_t c2 = k % out_c; + int64_t offset = k / out_c; + int64_t w2 = i * blocksize_ + offset % blocksize_; + int64_t h2 = j * blocksize_ + offset / blocksize_; + int64_t out_index = + w2 + w_ * blocksize_ * (h2 + h_ * blocksize_ * (c2 + out_c * b)); + if (forward_) + out_[out_index] = x_[in_index]; + else + out_[in_index] = x_[out_index]; + } + + private: + const T *x_; + int64_t w_, h_, c_, batch_, blocksize_, forward_; + T *out_; +}; + +template +class SpaceToDepthKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto *out = context.Output("Out"); + auto *x = context.Input("X"); + auto blocksize = context.Attr("blocksize"); + auto in_dims = x->dims(); + out->mutable_data(context.GetPlace(), x->type()); + + auto out_dims = out->dims(); + auto B = in_dims[0]; + auto C = in_dims[1]; + auto H = in_dims[2]; + auto W = in_dims[3]; + platform::ForRange for_range( + context.template device_context(), + static_cast(x->numel())); + + auto *x_data = x->data(); + auto *out_data = out->data(); + paddle::operators::space_to_depth_compute computer( + x_data, W, H, C, B, blocksize, 1, out_data); + for_range(computer); + + out->Resize(out_dims); + } +}; + +template +class SpaceToDepthGradKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext &context) const override { + auto *d_out = + context.Input(framework::GradVarName("Out")); + auto *d_x = + context.Output(framework::GradVarName("X")); + auto blocksize = context.Attr("blocksize"); + auto in_dims = d_x->dims(); + d_x->mutable_data(context.GetPlace(), d_out->type()); + + auto B = in_dims[0]; + auto C = in_dims[1]; + auto H = in_dims[2]; + auto W = in_dims[3]; + + platform::ForRange for_range( + context.template device_context(), + static_cast(d_x->numel())); + + auto *dx_data = d_x->data(); + auto *dout_data = d_out->data(); + + paddle::operators::space_to_depth_compute computer( + dout_data, W, H, C, B, blocksize, 0, dx_data); + for_range(computer); + + d_x->Resize(in_dims); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/split_byref_op.h b/paddle/fluid/operators/split_byref_op.h index fedd7218dd6cc9481e94a92a3820cafbe4157bd0..3b7ae6fc91e0a9e08406e38b9a557cab442c2560 100644 --- a/paddle/fluid/operators/split_byref_op.h +++ b/paddle/fluid/operators/split_byref_op.h @@ -32,7 +32,7 @@ class SplitByrefOpKernel : public framework::OpKernel { for (size_t i = 0; i < outs.size(); ++i) { // NOTE: no need to call mutable_data here to allocate memory. auto* out = outs[i]; - VLOG(3) << "spliting by ref: " << row_offset << " " << out->dims()[0]; + VLOG(30) << "spliting by ref: " << row_offset << " " << out->dims()[0]; *out = in->Slice(row_offset, row_offset + out->dims()[0]); row_offset += out->dims()[0]; } diff --git a/paddle/fluid/operators/split_ids_op.cc b/paddle/fluid/operators/split_ids_op.cc index c867c46873ae7ddbdbda280351e4ab28235bcc08..01d432e13068f7b718d08dc15d8cc99a7fbb0afe 100644 --- a/paddle/fluid/operators/split_ids_op.cc +++ b/paddle/fluid/operators/split_ids_op.cc @@ -20,20 +20,27 @@ namespace operators { class SplitIdsOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); - AddOutput("Out", "(LoDTensor) The outputs of the input Ids.") + AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}") + .AsDuplicable(); + + AddOutput("Out", "(LoDTensors) The outputs of the input Ids.") .AsDuplicable(); AddComment(R"DOC( Split a LoDTensor of Ids into multi LoDTensors, the number is pserver's number Example: Input: - X = [1,2,3,4,5,6] + X = [[1,2,3,4,5,6],[2,3]] Out(3 output): - out0 = [3, 6] - out1 = [1, 4] - out2 = [2, 5] + if compress is True: + out0 = [3, 3, 6] + out1 = [1, 4] + out2 = [2, 2, 5] + else: + out0 = [3, 6] + out1 = [1, 4] + out2 = [2, 5] )DOC"); } }; @@ -43,16 +50,23 @@ class SplitIdsOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext *ctx) const override { - PADDLE_ENFORCE(ctx->HasInput("Ids"), "SplitIdsOp must has input Ids."); + PADDLE_ENFORCE(ctx->HasInputs("Ids"), "SplitIdsOp must has input Ids."); PADDLE_ENFORCE(ctx->HasOutputs("Out"), "SplitIdsOp must has output Out."); auto ids_var_type = ctx->GetInputsVarType("Ids").front(); - auto ids_dims = ctx->GetInputDim("Ids"); + auto ids_dims = ctx->GetInputsDim("Ids"); if (ids_var_type == framework::proto::VarType::LOD_TENSOR) { - PADDLE_ENFORCE_EQ(ids_dims.size(), 2); - PADDLE_ENFORCE_EQ(ids_dims[1], 1); + PADDLE_ENFORCE_EQ(ids_dims[0].size(), 2); } } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::GetDataTypeOfVar(ctx.MultiInputVar("Ids").front()), + ctx.GetPlace()); + } }; class SplitIdsOpInferVarType : public framework::VarTypeInference { @@ -66,12 +80,28 @@ class SplitIdsOpInferVarType : public framework::VarTypeInference { } }; +class SplitIdsOpGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto grad = new framework::OpDesc(); + grad->SetType("concat"); + grad->SetInput("X", OutputGrad("Out")); + grad->SetOutput("Out", InputGrad("Ids")); + grad->SetAttr("axis", 0); + return std::unique_ptr(grad); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; REGISTER_OPERATOR(split_ids, ops::SplitIdsOp, ops::SplitIdsOpMaker, - ops::SplitIdsOpInferVarType); + ops::SplitIdsOpGradMaker, ops::SplitIdsOpInferVarType); + REGISTER_OP_CPU_KERNEL( split_ids, ops::SplitIdsOpKernel, ops::SplitIdsOpKernel); diff --git a/paddle/fluid/operators/split_ids_op.h b/paddle/fluid/operators/split_ids_op.h index c4af5a65fc5f81c1af7c1fdcca637ca37c940637..6dbada3da8826f0e7cb07a9642d327e5ee38c309 100644 --- a/paddle/fluid/operators/split_ids_op.h +++ b/paddle/fluid/operators/split_ids_op.h @@ -14,6 +14,8 @@ limitations under the License. */ #pragma once +#include +#include #include #include #include "paddle/fluid/framework/op_registry.h" @@ -31,19 +33,39 @@ class SplitIdsOpKernel : public framework::OpKernel { PADDLE_THROW("SplitIds do not support GPU kernel"); } - const auto *ids_var = ctx.InputVar("Ids"); + const auto ids_vars = ctx.MultiInputVar("Ids"); + + PADDLE_ENFORCE_GT(ids_vars.size(), 0, "The number of Ids should > 0"); + auto *ids_var = ids_vars[0]; + if (ids_var->IsType()) { - const auto &ids_dims = ctx.Input("Ids")->dims(); - const T *ids = ctx.Input("Ids")->data(); + int batch_size = 0; + const auto ids_tensors = ctx.MultiInput("Ids"); + for (size_t i = 0; i < ids_tensors.size(); ++i) { + batch_size += ids_tensors[i]->dims()[0]; + } + VLOG(40) << "Get Total BatchSize is: " << batch_size; + + std::vector all_ids(batch_size); + int offset = 0; + for (size_t i = 0; i < ids_tensors.size(); ++i) { + const auto *ids = ids_tensors[i]; + std::memcpy(all_ids.data() + offset, ids->data(), + ids->numel() * sizeof(T)); + offset += ids->numel(); + } + + std::set st(all_ids.begin(), all_ids.end()); + all_ids.assign(st.begin(), st.end()); + auto outs = ctx.MultiOutput("Out"); const size_t shard_num = outs.size(); - std::vector> out_ids; out_ids.resize(outs.size()); // split id by their shard_num. - for (int i = 0; i < ids_dims[0]; ++i) { - T id = ids[i]; + for (int i = 0; i < all_ids.size(); ++i) { + T id = all_ids[i]; size_t shard_id = static_cast(id) % shard_num; out_ids[shard_id].push_back(id); } @@ -64,7 +86,7 @@ class SplitIdsOpKernel : public framework::OpKernel { PADDLE_ENFORCE_EQ(ids_dims[0], static_cast(ids_selected_rows->rows().size()), ""); - const T *ids = ids_selected_rows->value().data(); + const T *ids_data = ids_selected_rows->value().data(); const auto &ids_rows = ids_selected_rows->rows(); auto outs = ctx.MultiOutput("Out"); const size_t shard_num = outs.size(); @@ -87,10 +109,14 @@ class SplitIdsOpKernel : public framework::OpKernel { T *output = out->mutable_value()->mutable_data(ddim, place); for (int64_t i = 0; i < ddim[0]; ++i) { memcpy(output + i * row_width, - ids + id_to_index[out->rows()[i]] * row_width, + ids_data + id_to_index[out->rows()[i]] * row_width, row_width * sizeof(T)); } } + } else { + PADDLE_THROW( + "% should be LoDTensor or SelectedRows, but the received type is %s", + ctx.Inputs("Ids")[0], ids_var->Type().name()); } } }; diff --git a/paddle/fluid/operators/split_op.cc b/paddle/fluid/operators/split_op.cc index d661b276bc31bf0c3ab181d706ffdccec89f0632..a05582ae09e16ee17194d299d713d321f28ccace 100644 --- a/paddle/fluid/operators/split_op.cc +++ b/paddle/fluid/operators/split_op.cc @@ -111,11 +111,10 @@ Example: } // namespace paddle namespace ops = paddle::operators; -USE_CPU_ONLY_OP(concat); REGISTER_OPERATOR(split, ops::SplitOp, ops::SplitOpMaker, ops::SplitGradMaker); -REGISTER_OP_CPU_KERNEL(split, - ops::SplitOpKernel, - ops::SplitOpKernel, - ops::SplitOpKernel, - ops::SplitOpKernel); +REGISTER_OP_CPU_KERNEL( + split, ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel, + ops::SplitOpKernel); diff --git a/paddle/fluid/operators/split_op.h b/paddle/fluid/operators/split_op.h index f0c417c70521b1bb3816f884d6ab7393473999e4..6f4a25ab5ed86937f2f5db532a9eba22b5a2c5be 100644 --- a/paddle/fluid/operators/split_op.h +++ b/paddle/fluid/operators/split_op.h @@ -17,6 +17,7 @@ limitations under the License. */ #include // NOLINT #include #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/concat_and_split.h" #include "paddle/fluid/operators/strided_memcpy.h" namespace paddle { @@ -28,18 +29,22 @@ class SplitOpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* in = ctx.Input("X"); auto outs = ctx.MultiOutput("Out"); - auto in_stride = framework::stride_numel(in->dims()); - int64_t axis = static_cast(ctx.Attr("axis")); + int axis = ctx.Attr("axis"); auto place = ctx.GetPlace(); - size_t input_offset = 0; - for (auto& out : outs) { - out->mutable_data(ctx.GetPlace()); - auto out_stride = framework::stride_numel(out->dims()); - StridedNumelCopyWithAxis(ctx.device_context(), axis, out->data(), - out_stride, in->data() + input_offset, - in_stride, out_stride[axis]); - input_offset += out_stride[axis]; + std::vector shape_refer; + for (size_t j = 0; j < outs.size(); ++j) { + outs[j]->mutable_data(ctx.GetPlace()); + shape_refer.emplace_back(outs[j]); + } + + auto& dev_ctx = ctx.template device_context(); + // Sometimes direct copies will be faster, this maybe need deeply analysis. + if (axis == 0 && outs.size() < 10) { + StridedMemcpyWithAxis0(dev_ctx, *in, shape_refer, &outs); + } else { + math::SplitFunctor functor; + functor(dev_ctx, *in, shape_refer, axis, &outs); } } }; diff --git a/paddle/fluid/operators/split_selected_rows_op.cc b/paddle/fluid/operators/split_selected_rows_op.cc index 76615a9405d7a8e3fa9dba8d01a956209e02ae8f..0e7b1463d1ba81aed53e0e3f3a90d2a1fbf0ffbc 100644 --- a/paddle/fluid/operators/split_selected_rows_op.cc +++ b/paddle/fluid/operators/split_selected_rows_op.cc @@ -22,9 +22,9 @@ class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "The input SelectedRows."); AddOutput("Out", "The outputs of the input SelectedRows.").AsDuplicable(); - AddAttr>("height_sections", - "Height for each output SelectedRows.") - .SetDefault(std::vector({})); + AddAttr>("height_sections", + "Height for each output SelectedRows.") + .SetDefault(std::vector({})); AddComment(R"DOC( Split a SelectedRows with a specified rows section. diff --git a/paddle/fluid/operators/split_selected_rows_op.h b/paddle/fluid/operators/split_selected_rows_op.h index 0e9ce165b98845f4745ee70b028513ea31cc6657..af64607fafc6544047714e731846a2440be219b8 100644 --- a/paddle/fluid/operators/split_selected_rows_op.h +++ b/paddle/fluid/operators/split_selected_rows_op.h @@ -21,7 +21,7 @@ limitations under the License. */ namespace paddle { namespace operators { -static int FindOutIdx(int row, const std::vector& abs_sections) { +static int FindOutIdx(int row, const std::vector& abs_sections) { for (size_t i = 1; i < abs_sections.size(); ++i) { if (row < abs_sections[i]) { return i - 1; @@ -30,9 +30,9 @@ static int FindOutIdx(int row, const std::vector& abs_sections) { return abs_sections.size() - 1; } -static std::vector ToAbsoluteSection( - const std::vector& height_sections) { - std::vector abs_sections; +static std::vector ToAbsoluteSection( + const std::vector& height_sections) { + std::vector abs_sections; abs_sections.resize(height_sections.size()); abs_sections[0] = 0; for (size_t i = 1; i < height_sections.size(); ++i) { @@ -47,7 +47,7 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& ctx) const override { auto* x = ctx.Input("X"); auto outs = ctx.MultiOutput("Out"); - auto height_sections = ctx.Attr>("height_sections"); + auto height_sections = ctx.Attr>("height_sections"); auto abs_sections = ToAbsoluteSection(height_sections); diff --git a/paddle/fluid/operators/spp_op.h b/paddle/fluid/operators/spp_op.h index 08cb7849d20443862b66ea6096c095b294c7242c..35d9737ee01fe1505cbe30e8ed735e6b92cb8df2 100644 --- a/paddle/fluid/operators/spp_op.h +++ b/paddle/fluid/operators/spp_op.h @@ -56,12 +56,14 @@ class SppKernel : public framework::OpKernel { math::Pool2dFunctor, T> pool_forward; math::MaxPool max_process; pool_forward(context.template device_context(), *in_x, - kernel_size, strides, paddings, max_process, &out_level); + kernel_size, strides, paddings, max_process, true, + &out_level); } else if (pooling_type == "avg") { math::Pool2dFunctor, T> pool_forward; math::AvgPool avg_process; pool_forward(context.template device_context(), *in_x, - kernel_size, strides, paddings, avg_process, &out_level); + kernel_size, strides, paddings, avg_process, true, + &out_level); } // flatten pooling output shape int output_flatten_w = in_x->dims()[1] * bins * bins; @@ -154,7 +156,7 @@ class SppGradKernel : public framework::OpKernel { math::AvgPoolGrad avg_process; pool_backward(context.template device_context(), *in_x, *&out_level, *&outgrad_level, kernel_size, strides, - paddings, avg_process, in_x_grad); + paddings, avg_process, true, in_x_grad); } } } diff --git a/paddle/fluid/operators/stack_op.cc b/paddle/fluid/operators/stack_op.cc index 3f4b48bc7391def082c82ed451fc5a752009a2f1..9345b495415d203728238c19621a20f446c40bf5 100644 --- a/paddle/fluid/operators/stack_op.cc +++ b/paddle/fluid/operators/stack_op.cc @@ -21,8 +21,12 @@ REGISTER_OPERATOR(stack, ops::StackOp, ops::StackOpMaker, REGISTER_OPERATOR(stack_grad, ops::StackOpGrad); REGISTER_OP_CPU_KERNEL(stack, ops::StackKernel, - ops::StackKernel); + ops::StackKernel, + ops::StackKernel, + ops::StackKernel); REGISTER_OP_CPU_KERNEL(stack_grad, ops::StackGradKernel, - ops::StackGradKernel); + ops::StackGradKernel, + ops::StackGradKernel, + ops::StackGradKernel); diff --git a/paddle/fluid/operators/stack_op.cu b/paddle/fluid/operators/stack_op.cu index 92c1bde2bcf089e5c715e90e564408e6ad37ba17..bf2a9e5b3d22996e688621727cb280dc9aed7859 100644 --- a/paddle/fluid/operators/stack_op.cu +++ b/paddle/fluid/operators/stack_op.cu @@ -18,8 +18,12 @@ namespace plat = paddle::platform; namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL(stack, ops::StackKernel, - ops::StackKernel); + ops::StackKernel, + ops::StackKernel, + ops::StackKernel); REGISTER_OP_CUDA_KERNEL(stack_grad, ops::StackGradKernel, - ops::StackGradKernel); + ops::StackGradKernel, + ops::StackGradKernel, + ops::StackGradKernel); diff --git a/paddle/fluid/operators/strided_memcpy.h b/paddle/fluid/operators/strided_memcpy.h index 7a10218e1556698f3e0a1828db5de8851dd1c90b..c3d83a06f23a34ec8cf27d585863135ebfd56a4f 100644 --- a/paddle/fluid/operators/strided_memcpy.h +++ b/paddle/fluid/operators/strided_memcpy.h @@ -13,8 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include +#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/operators/detail/strided_memcpy.h" - namespace paddle { namespace operators { @@ -98,5 +99,26 @@ inline void StridedNumelCopyWithAxis(const platform::DeviceContext& ctx, } } +template +inline void StridedMemcpyWithAxis0( + const platform::DeviceContext& dev_ctx, const framework::Tensor& input, + const std::vector& shape_refer, + std::vector* outputs) { + const framework::DDim in_stride = stride_numel(input.dims()); + const int axis = 0; + size_t input_offset = 0; + + for (size_t i = 0; i < outputs->size(); ++i) { + auto out_stride = stride_numel(shape_refer[i]->dims()); + auto out = outputs->at(i); + if (out != nullptr) { + StridedNumelCopyWithAxis(dev_ctx, axis, out->data(), out_stride, + input.data() + input_offset, in_stride, + out_stride[axis]); + } + input_offset += out_stride[axis]; + } +} + } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/sum_mkldnn_op.cc b/paddle/fluid/operators/sum_mkldnn_op.cc index f9a16ef35ecb9eeb6c8eda9d124ecb17e7f9d5ce..2ae5c17bf6465874572e80da54e40fbe22403660 100644 --- a/paddle/fluid/operators/sum_mkldnn_op.cc +++ b/paddle/fluid/operators/sum_mkldnn_op.cc @@ -186,7 +186,7 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel { } if (in_dim.empty()) { - VLOG(3) << "WARNING: all the inputs are empty"; + VLOG(30) << "WARNING: all the inputs are empty"; in_dim = framework::vectorize(get_selected_row(N - 1).value().dims()); } else { in_dim[0] = static_cast(first_dim); diff --git a/paddle/fluid/operators/sum_op.cc b/paddle/fluid/operators/sum_op.cc index fe7c7039c7dec714e265ede1b7167fd800ddc2f7..c67b694283cd8f0203021c0329f5ac16ae7854a5 100644 --- a/paddle/fluid/operators/sum_op.cc +++ b/paddle/fluid/operators/sum_op.cc @@ -45,7 +45,7 @@ class SumOp : public framework::OperatorWithKernel { size_t N = x_dims.size(); PADDLE_ENFORCE_GT(N, 0, "Input tensors count should > 0."); if (N == 1) { - VLOG(3) << "Warning: sum have only one input, may waste memory"; + VLOG(30) << "Warning: sum have only one input, may waste memory"; } framework::DDim in_dim({0}); @@ -67,6 +67,7 @@ class SumOp : public framework::OperatorWithKernel { framework::OpKernelType GetExpectedKernelType( const framework::ExecutionContext& ctx) const override { auto x_vars = ctx.MultiInputVar("X"); + auto x_vars_name = ctx.Inputs("X"); framework::LibraryType library{framework::LibraryType::kPlain}; framework::DataLayout layout{framework::DataLayout::kAnyLayout}; @@ -81,15 +82,18 @@ class SumOp : public framework::OperatorWithKernel { if (x_vars[0]->IsType()) { int dtype = -1; - for (auto& x_var : x_vars) { - auto& lod_tensor = x_var->Get(); - if (lod_tensor.numel() == 0) { + for (size_t idx = 0; idx < x_vars.size(); ++idx) { + PADDLE_ENFORCE(x_vars[idx] != nullptr, + "Input var[%s] should not be nullptr", x_vars_name[idx]); + auto tensor = + framework::GetLoDTensorOrSelectedRowsValueFromVar(*x_vars[idx]); + if (tensor->numel() == 0) { continue; } if (dtype == -1) { - dtype = framework::ToDataType(lod_tensor.type()); + dtype = framework::ToDataType(tensor->type()); } else { - PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(lod_tensor.type())); + PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(tensor->type())); } } PADDLE_ENFORCE_NE(dtype, -1, @@ -132,7 +136,7 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "(vector) The input tensors of sum operator.") .AsDuplicable(); - AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X"); + AddOutput("Out", "(Tensor) The output tensor of sum operator."); AddAttr("use_mkldnn", "(bool, default false) Only used in mkldnn kernel") .SetDefault(false); @@ -153,8 +157,8 @@ class SumOpVarTypeInference : public framework::VarTypeInference { auto& inputs = op_desc.Input("X"); auto var_type = framework::proto::VarType::SELECTED_ROWS; for (auto& name : op_desc.Input("X")) { - VLOG(10) << name << " " - << block->FindRecursiveOrCreateVar(name).GetType(); + VLOG(100) << name << " " + << block->FindRecursiveOrCreateVar(name).GetType(); } bool any_input_is_lod_tensor = std::any_of( diff --git a/paddle/fluid/operators/sum_op.cu b/paddle/fluid/operators/sum_op.cu index 89bcd1bbc86dc29cb7b98cbef3057a8f98c74555..db4c2d6c115f04b436db00854ca4b02fea09866b 100644 --- a/paddle/fluid/operators/sum_op.cu +++ b/paddle/fluid/operators/sum_op.cu @@ -11,10 +11,13 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/sum_op.h" +#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; +namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( sum, ops::SumKernel, ops::SumKernel, ops::SumKernel, - ops::SumKernel); + ops::SumKernel, + ops::SumKernel); diff --git a/paddle/fluid/operators/sum_op.h b/paddle/fluid/operators/sum_op.h index 34403c7a7aa717cca470be2931009e219e00e3ae..19b2c68c823adbed82319f7b04992baedd5d41f9 100644 --- a/paddle/fluid/operators/sum_op.h +++ b/paddle/fluid/operators/sum_op.h @@ -43,17 +43,31 @@ class SumKernel : public framework::OpKernel { out->mutable_data(context.GetPlace()); } auto result = EigenVector::Flatten(*out); + auto &place = + *context.template device_context().eigen_device(); + int start = in_place ? 1 : 0; if (!in_place) { - math::SetConstant constant_functor; - constant_functor(context.template device_context(), out, - 0.0); + if ((in_num >= 2) && in_vars[0]->IsType() && + in_vars[1]->IsType()) { + auto &in_0 = in_vars[0]->Get(); + auto &in_1 = in_vars[1]->Get(); + if (in_0.numel() && in_1.numel()) { + auto in_0_e = EigenVector::Flatten(in_0); + auto in_1_e = EigenVector::Flatten(in_1); + result.device(place) = in_0_e + in_1_e; + start = 2; + } + } + if (start != 2) { + math::SetConstant constant_functor; + constant_functor(context.template device_context(), + out, static_cast(0)); + } } math::SelectedRowsAddToTensor functor; - auto &place = - *context.template device_context().eigen_device(); // If in_place, just skip the first tensor - for (size_t i = in_place ? 1 : 0; i < in_num; i++) { + for (size_t i = start; i < in_num; i++) { if (in_vars[i]->IsType()) { auto &in_t = in_vars[i]->Get(); if (in_t.numel() == 0) { @@ -69,79 +83,54 @@ class SumKernel : public framework::OpKernel { } } } else if (out_var->IsType()) { - std::unique_ptr in0; - if (in_place) { - // If is in_place, we store the input[0] to in0 - auto &in_sel0 = in_vars[0]->Get(); - auto &rows = in_sel0.rows(); -#ifdef PADDLE_WITH_CUDA - std::vector rows_in_cpu; - rows_in_cpu.reserve(rows.size()); - for (auto item : rows) { - rows_in_cpu.push_back(item); - } - in0.reset(new framework::SelectedRows(rows_in_cpu, in_sel0.height())); -#else - in0.reset(new framework::SelectedRows(rows, in_sel0.height())); -#endif - in0->mutable_value()->ShareDataWith(in_sel0.value()); + if (in_place && in_vars.size() < 2) { + return; } - auto get_selected_row = [&](size_t i) -> const SelectedRows & { - if (i == 0 && in0) { - return *in0.get(); - } else { - return in_vars[i]->Get(); + std::vector inputs; + SelectedRows temp_in0; + + if (in_place) { + auto &in0 = in_vars[0]->Get(); + temp_in0.set_height(in0.height()); + temp_in0.set_rows(in0.rows()); + framework::TensorCopy(in0.value(), in0.place(), + context.device_context(), + temp_in0.mutable_value()); + inputs.push_back(&temp_in0); + for (size_t i = 1; i < in_vars.size(); ++i) { + auto &in = in_vars[i]->Get(); + if (in.rows().size() > 0) { + inputs.push_back(&in); + } } - }; + } else { + for (auto &in_var : in_vars) { + auto &in = in_var->Get(); + if (in.rows().size() > 0) { + inputs.push_back(&in_var->Get()); + } + } + } auto *out = context.Output("Out"); out->mutable_rows()->clear(); - auto *out_value = out->mutable_value(); - // Runtime InferShape - size_t first_dim = 0; - for (size_t i = 0; i < in_num; i++) { - auto &sel_row = get_selected_row(i); - first_dim += sel_row.rows().size(); - } - - std::vector in_dim; - for (size_t i = 0; i < in_num; i++) { - auto &sel_row = get_selected_row(i); - if (sel_row.rows().size() > 0) { - in_dim = framework::vectorize(sel_row.value().dims()); + bool has_data = false; + for (auto &in : inputs) { + if (in->rows().size() > 0) { + has_data = true; break; } } - if (in_dim.empty()) { - VLOG(3) << "WARNING: all the inputs are empty"; - in_dim = - framework::vectorize(get_selected_row(in_num - 1).value().dims()); + if (has_data) { + math::scatter::MergeAdd merge_add; + merge_add(context.template device_context(), inputs, + out); } else { - in_dim[0] = static_cast(first_dim); - } - - out_value->Resize(framework::make_ddim(in_dim)); - out_value->mutable_data(context.GetPlace()); - // if all the input sparse vars are empty, no need to - // merge these vars. - if (first_dim == 0UL) { - return; - } - - math::SelectedRowsAddTo functor; - - int64_t offset = 0; - for (size_t i = 0; i < in_num; i++) { - auto &sel_row = get_selected_row(i); - if (sel_row.rows().size() == 0) { - continue; - } - PADDLE_ENFORCE_EQ(out->height(), sel_row.height()); - functor(context.template device_context(), sel_row, - offset, out); - offset += sel_row.value().numel(); + // no data, just set a empty out tensor. + out->mutable_value()->mutable_data(framework::make_ddim({0}), + context.GetPlace()); } } else if (out_var->IsType()) { auto &out_array = *out_var->GetMutable(); diff --git a/paddle/fluid/operators/tensor_array_read_write_op.cc b/paddle/fluid/operators/tensor_array_read_write_op.cc index a2d44284e9de1ace42cabbce82e0b45929432d7b..484160aeb8de573c6a6c1bb2ea5da23600d2d287 100644 --- a/paddle/fluid/operators/tensor_array_read_write_op.cc +++ b/paddle/fluid/operators/tensor_array_read_write_op.cc @@ -34,8 +34,8 @@ class WriteToArrayOp : public ArrayOp { auto *out = scope.FindVar(Output("Out"))->GetMutable(); if (offset >= out->size()) { - VLOG(10) << "Resize " << Output("Out") << " from " << out->size() - << " to " << offset + 1; + VLOG(100) << "Resize " << Output("Out") << " from " << out->size() + << " to " << offset + 1; out->resize(offset + 1); } auto *out_tensor = &out->at(offset); @@ -47,9 +47,9 @@ class WriteToArrayOp : public ArrayOp { TensorCopy(x_tensor, place, dev_ctx, out_tensor); } else { - VLOG(10) << "WARNING: The input tensor 'x_tensor' holds no memory, so " - "nothing has been written to output array[" - << offset << "]."; + VLOG(100) << "WARNING: The input tensor 'x_tensor' holds no memory, so " + "nothing has been written to output array[" + << offset << "]."; } } }; @@ -104,7 +104,7 @@ class WriteToArrayInferVarType : public framework::VarTypeInference { framework::BlockDesc *block) const override { auto x_name = op_desc.Input("X")[0]; auto out_name = op_desc.Output("Out")[0]; - VLOG(10) << "Set Variable " << out_name << " as LOD_TENSOR_ARRAY"; + VLOG(100) << "Set Variable " << out_name << " as LOD_TENSOR_ARRAY"; auto &out = block->FindRecursiveOrCreateVar(out_name); out.SetType(framework::proto::VarType::LOD_TENSOR_ARRAY); auto *x = block->FindVarRecursive(x_name); @@ -139,7 +139,7 @@ class ReadFromArrayOp : public ArrayOp { framework::TensorCopy(x_array[offset], place, dev_ctx, out_tensor); out_tensor->set_lod(x_array[offset].lod()); } else { - VLOG(10) << "offset " << offset << " >= " << x_array.size(); + VLOG(100) << "offset " << offset << " >= " << x_array.size(); } } }; diff --git a/paddle/fluid/operators/tensor_array_to_tensor_op.cc b/paddle/fluid/operators/tensor_array_to_tensor_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..96dc123f6a36e1a2b6ae04e0d97dffe1e10ac4ea --- /dev/null +++ b/paddle/fluid/operators/tensor_array_to_tensor_op.cc @@ -0,0 +1,246 @@ +/* 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 +#include + +#include "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/variable.h" + +namespace paddle { +namespace operators { +using framework::Tensor; + +void LodTensorArray2LodTensorVector(const framework::Scope &scope, + const std::string &base_name, + const std::string &lod_tensor_array_name, + std::vector *res_names) { + auto &inx = + scope.FindVar(lod_tensor_array_name)->Get(); + for (size_t i = 0; i < inx.size(); i++) { + std::string var_name = base_name + std::to_string(i); + framework::Variable *g_feed_value = + const_cast(scope).Var(var_name); + auto &feed_input = + *(g_feed_value->GetMutable()); + feed_input.ShareDataWith(inx[i]); + res_names->push_back(var_name); + } +} + +void LodTensorVectorResizeFromLodTensorArray( + const framework::Scope &scope, const std::string &base_name, + const std::string &lod_tensor_array_name, + std::vector *res_names) { + auto &inx = + scope.FindVar(lod_tensor_array_name)->Get(); + for (size_t i = 0; i < inx.size(); i++) { + std::string var_name = base_name + std::to_string(i); + framework::Variable *g_feed_value = + const_cast(scope).Var(var_name); + auto &feed_input = + *(g_feed_value->GetMutable()); + auto dims = inx[i].dims(); + feed_input.Resize(dims); + res_names->push_back(var_name); + } +} + +void LodTensorArrayCreateFromLodTensorArray( + const framework::Scope &scope, + const std::string &input_lod_tensor_array_name, + const std::string &output_lod_tensor_array_name) { + auto &inx = scope.FindVar(input_lod_tensor_array_name) + ->Get(); + auto &grad_inx = *scope.FindVar(output_lod_tensor_array_name) + ->GetMutable(); + + for (size_t i = 0; i < inx.size(); i++) { + std::string var_name = output_lod_tensor_array_name + std::to_string(i); + framework::Variable *g_feed_value = + const_cast(scope).Var(var_name); + auto &feed_input = + *(g_feed_value->GetMutable()); + grad_inx.push_back(feed_input); + } +} + +class LoDTensorArray2TensorOp : public framework::OperatorBase { + public: + using OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto axis = Attr("axis"); + + framework::AttributeMap attrs; + attrs["axis"] = axis; + + auto &inx = scope.FindVar(Input("X"))->Get(); + auto &out = + *scope.FindVar(Output("Out"))->GetMutable(); + auto &out_inx = + *scope.FindVar(Output("OutIndex"))->GetMutable(); + + const size_t n = inx.size(); + PADDLE_ENFORCE_GT(n, 0, "Input tensorarray size should > 0."); + + std::string base_name = Inputs("X")[0]; + std::vector names; + + // get the input tensorarray items' dim in out_inx + auto out_inx_dim = out_inx.dims(); + out_inx_dim[0] = inx.size(); + out_inx.Resize(out_inx_dim); + + std::string var_name = "out_index"; + framework::Variable *tmp_index_var = + const_cast(scope).Var(var_name); + auto &tmp_index_tensor = + *(tmp_index_var->GetMutable()); + tmp_index_tensor.Resize(out_inx_dim); + int *tmp_index_data = + tmp_index_tensor.mutable_data(platform::CPUPlace()); + + auto out_dims = inx[0].dims(); + size_t out_dim_sum = 0; + for (size_t index = 0; index < inx.size(); index++) { + auto inx_dims = inx[index].dims(); + out_dim_sum += inx_dims[axis]; + tmp_index_data[index] = inx_dims[axis]; + } + out_inx.ShareDataWith(tmp_index_tensor); + + // get input array items' dims + out_dims[axis] = out_dim_sum; + out.Resize(out_dims); + + LodTensorArray2LodTensorVector(scope, base_name, Input("X"), &names); + // Invoke Reshape Op + auto concat_op = framework::OpRegistry::CreateOp( + "concat", {{"X", names}}, {{"Out", {Output("Out")}}}, attrs); + + concat_op->Run(scope, place); + } +}; + +class LoDTensorArray2TensorOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", "Input LoDTensorArray of tensor_array_to_tensor operator."); + AddOutput("Out", "Output tensor of tensor_array_to_tensor operator."); + AddOutput("OutIndex", + "Output input LoDTensorArray items' dims of " + "tensor_array_to_tensor operator."); + AddAttr("axis", + "The axis along which the input tensors will be concatenated.") + .SetDefault(0); + AddComment(R"DOC( +tensor_array_to_tensor Operator. + +Concatenate the input LoDTensorArray along dimension axis to the output Tensor. +Examples: + Input = {[1,2], [3,4], [5,6]} + axis = 0 + Output = [[1,2], + [3,4], + [5,6]] + OutputIndex = [1,1,1] + +)DOC"); + } +}; + +class LoDTensorArray2TensorOpInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *ctx) const override {} +}; + +class LoDTensorArray2TensorGradInferShape : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext *context) const override {} +}; + +class LoDTensorArray2TensorGradInferVarType + : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + for (auto &out_var : op_desc.Output(framework::GradVarName("X"))) { + block->Var(out_var)->SetType(framework::proto::VarType::LOD_TENSOR_ARRAY); + } + } +}; + +class LoDTensorArray2TensorGradOp : public framework::OperatorBase { + public: + using OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &place) const override { + auto axis = Attr("axis"); + framework::AttributeMap attrs; + attrs["axis"] = axis; + + auto &inx = scope.FindVar(Input("X"))->Get(); + const size_t n = inx.size(); + PADDLE_ENFORCE_GT(n, 0, "Input tensorarray size should > 0."); + + std::string base_name = Inputs("X")[0]; + std::vector names; + + LodTensorArray2LodTensorVector(scope, base_name, Input("X"), &names); + + // grad + auto dx_name = Output(framework::GradVarName("X")); + auto dout_name = Input(framework::GradVarName("Out")); + + std::vector grad_names; + + LodTensorVectorResizeFromLodTensorArray(scope, "grad_name", Input("X"), + &grad_names); + + auto concat_grad_op = framework::OpRegistry::CreateOp( + "concat_grad", {{"X", names}, {"Out@GRAD", {dout_name}}}, + {{"X@GRAD", grad_names}}, attrs); + + concat_grad_op->Run(scope, place); + + LodTensorArrayCreateFromLodTensorArray(scope, Input("X"), dx_name); + auto &grad_inx = + *scope.FindVar(dx_name)->GetMutable(); + + for (size_t i = 0; i < grad_names.size(); i++) { + std::string var_name = grad_names[i]; + auto &feed_input = scope.FindVar(var_name)->Get(); + grad_inx[i].ShareDataWith(feed_input); + } + } +}; + +} // namespace operators +} // namespace paddle +USE_OP(concat); + +namespace ops = paddle::operators; +REGISTER_OPERATOR(tensor_array_to_tensor, ops::LoDTensorArray2TensorOp, + ops::LoDTensorArray2TensorOpMaker, + ops::LoDTensorArray2TensorOpInferShape, + paddle::framework::DefaultGradOpDescMaker); +REGISTER_OPERATOR(tensor_array_to_tensor_grad, ops::LoDTensorArray2TensorGradOp, + ops::LoDTensorArray2TensorGradInferShape, + ops::LoDTensorArray2TensorGradInferVarType); diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index d4ba0f9c33c91811647f9d19a332f139c16b0eb2..3af9376da1d3fa096b277e6b5a9d1a8de197d6f1 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -34,7 +34,7 @@ namespace operators { using FluidDT = framework::proto::VarType_Type; using TRT_DT = nvinfer1::DataType; -namespace { +namespace { // NOLINT TRT_DT FluidDataType2TRT(FluidDT type) { switch (type) { @@ -60,7 +60,7 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector& shape) { return nvinfer1::DimsCHW(shape[1], 1, 1); } -} // namespace +} // namespace // NOLINT using inference::Singleton; using inference::tensorrt::TRT_EngineManager; @@ -127,9 +127,9 @@ class TensorRTEngineKernel : public framework::OpKernel { // Convert output tensor from engine to fluid int output_index = 0; - VLOG(4) << "TensorRT Engine Op Outputs:"; + VLOG(40) << "TensorRT Engine Op Outputs:"; for (const auto& y : context.Outputs("Ys")) { - VLOG(4) << y; + VLOG(40) << y; // convert output and copy to fluid. nvinfer1::ITensor* trt_t = engine->GetITensor(output_maps[output_index]); auto dims = trt_t->getDimensions(); @@ -167,7 +167,7 @@ class TensorRTEngineKernel : public framework::OpKernel { protected: void Prepare(const framework::ExecutionContext& context) const { - VLOG(4) << "Prepare engine"; + VLOG(40) << "Prepare engine"; // Get the ProgramDesc and pass to convert. framework::proto::BlockDesc block_desc; block_desc.ParseFromString(context.Attr("subgraph")); @@ -192,12 +192,12 @@ class TensorRTEngineKernel : public framework::OpKernel { engine->InitNetwork(); framework::BlockDesc block(nullptr /*programdesc*/, &block_desc); - VLOG(4) << "parsed var size " << block.AllVars().size(); + VLOG(40) << "parsed var size " << block.AllVars().size(); // Add inputs - VLOG(4) << "declare inputs"; + VLOG(40) << "declare inputs"; for (auto& input : context.Inputs("Xs")) { if (parameters.count(input)) continue; - VLOG(4) << "declare input " << input; + VLOG(40) << "declare input " << input; auto* var = block.FindVar(input); // TensorRT engine need to create parameters. The parameter's description // should be set in @@ -223,7 +223,9 @@ class TensorRTEngineKernel : public framework::OpKernel { // Add outputs for (auto& output : output_maps) { - engine->DeclareOutput(output); + if (!engine->HasDeclared(output)) { + engine->DeclareOutput(output); + } } engine->FreezeNetwork(); diff --git a/paddle/fluid/operators/test_send_nccl_id.cc b/paddle/fluid/operators/test_send_nccl_id.cc index e2b7b6b8e447381229e4ad594b7974bc0aa159d5..b5426e17aac19dc07ee56545fac8472d9ef0d93c 100644 --- a/paddle/fluid/operators/test_send_nccl_id.cc +++ b/paddle/fluid/operators/test_send_nccl_id.cc @@ -92,7 +92,7 @@ TEST(SendNcclId, RPCServer) { std::string ep = string::Sprintf("127.0.0.1:%d", port); distributed::RPCClient* client = - distributed::RPCClient::GetInstance(); + distributed::RPCClient::GetInstance(0); LOG(INFO) << "connect to server" << ep; client->AsyncSendVar(ep, dev_ctx, scope, NCCL_ID_VARNAME); diff --git a/paddle/fluid/operators/top_k_op.cc b/paddle/fluid/operators/top_k_op.cc index 4a8ac441cfaf642fde58ee30865a22e83c065498..c17d1afc309c65035063348d4934ea1783b018ed 100644 --- a/paddle/fluid/operators/top_k_op.cc +++ b/paddle/fluid/operators/top_k_op.cc @@ -50,7 +50,7 @@ class TopkOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { AddInput("X", "(Tensor) The input of Topk op"); - AddOutput("Out", "(Tensor) The output tensor of Topk op").Reuse("X"); + AddOutput("Out", "(Tensor) The output tensor of Topk op"); AddOutput("Indices", "(Tensor) The indices of Topk elements of input"); AddComment(R"DOC( Top K operator diff --git a/paddle/fluid/operators/top_k_op.cu b/paddle/fluid/operators/top_k_op.cu index 8e4a07556fb51dbb15ef948fcee120e2f68e089a..0cad224ca8860b0e4bc2e3f2bc1659235aadfe2d 100644 --- a/paddle/fluid/operators/top_k_op.cu +++ b/paddle/fluid/operators/top_k_op.cu @@ -262,31 +262,31 @@ __global__ void KeMatrixTopK(T* output, int output_stride, int64_t* indices, const T* src, int lds, int dim, int k, int grid_dim, int num) { __shared__ Pair sh_topk[BlockSize]; - __shared__ int maxid[BlockSize / 2]; const int tid = threadIdx.x; const int warp = threadIdx.x / 32; const int bid = blockIdx.x; for (int i = bid; i < num; i += grid_dim) { - output += i * output_stride; - indices += i * k; - + int top_num = k; + __shared__ int maxid[BlockSize / 2]; + T* out = output + i * output_stride; + int64_t* inds = indices + i * k; Pair topk[MaxLength]; int beam = MaxLength; Pair max; bool is_empty = false; bool firststep = true; - for (int k = 0; k < MaxLength; k++) { - topk[k].set(-INFINITY, -1); + for (int j = 0; j < MaxLength; j++) { + topk[j].set(-INFINITY, -1); } - while (k) { + while (top_num) { ThreadGetTopK( topk, &beam, k, src + i * lds, &firststep, &is_empty, &max, dim, tid); sh_topk[tid] = topk[0]; - BlockReduce(sh_topk, maxid, topk, &output, - &indices, &beam, &k, tid, warp); + BlockReduce(sh_topk, maxid, topk, &out, &inds, + &beam, &top_num, tid, warp); } } } @@ -327,13 +327,15 @@ class TopkOpCUDAKernel : public framework::OpKernel { size_t k = static_cast(ctx.Attr("k")); const T* input_data = input->data(); - T* output_data = output->mutable_data(ctx.GetPlace()); // FIXME(typhoonzero): data is always converted to type T? int64_t* indices_data = indices->mutable_data(ctx.GetPlace()); - size_t input_height = input->dims()[0]; - size_t input_width = input->dims()[1]; + framework::DDim inputdims = input->dims(); + const size_t input_height = framework::product( + framework::slice_ddim(inputdims, 0, inputdims.size() - 1)); + const size_t input_width = inputdims[inputdims.size() - 1]; + if (k > input_width) k = input_width; // NOTE: pass lds and dim same to input width. @@ -342,14 +344,12 @@ class TopkOpCUDAKernel : public framework::OpKernel { const int kMaxHeight = 2048; int gridx = input_height < kMaxHeight ? input_height : kMaxHeight; auto& dev_ctx = ctx.cuda_device_context(); - switch (GetDesiredBlockDim(input_width)) { FIXED_BLOCK_DIM( KeMatrixTopK<<>>( - output_data, output->dims()[1], indices_data, input_data, - input_width, input_width, static_cast(k), gridx, - input_height)); + output_data, k, indices_data, input_data, input_width, + input_width, static_cast(k), gridx, input_height)); default: PADDLE_THROW("Error"); } diff --git a/paddle/fluid/operators/top_k_op.h b/paddle/fluid/operators/top_k_op.h index 054dd481994d03f71b0ed5dc73e103085f6c91aa..76ece57b39919148da04caecaa43ea9d2b9d95df 100644 --- a/paddle/fluid/operators/top_k_op.h +++ b/paddle/fluid/operators/top_k_op.h @@ -34,7 +34,6 @@ class TopkKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& ctx) const override { // Get the top k elements of each row of input tensor - // FIXME: only deal with matrix(2d tensor). auto* input = ctx.Input("X"); auto* output = ctx.Output("Out"); auto* indices = ctx.Output("Indices"); @@ -44,8 +43,6 @@ class TopkKernel : public framework::OpKernel { T* output_data = output->mutable_data(ctx.GetPlace()); int64_t* indices_data = indices->mutable_data(ctx.GetPlace()); - auto eg_input = EigenMatrix::From(*input); - // reshape input to a flattern matrix(like flat_inner_dims) framework::DDim inputdims = input->dims(); const size_t row = framework::product( @@ -53,7 +50,7 @@ class TopkKernel : public framework::OpKernel { const size_t col = inputdims[inputdims.size() - 1]; Eigen::DSizes flat2dims(row, col); // NOTE: eigen shape doesn't affect paddle tensor. - eg_input.reshape(flat2dims); + auto eg_input = EigenMatrix::Reshape(*input, inputdims.size() - 1); #ifdef PADDLE_WITH_MKLML #pragma omp parallel for diff --git a/paddle/fluid/operators/transpose_op.cc b/paddle/fluid/operators/transpose_op.cc index 6a9fc6611a8f8eaa6749aefac0673ccabaebbcfe..bbd71db6062107f6ba40343c84d942b54b3958e6 100644 --- a/paddle/fluid/operators/transpose_op.cc +++ b/paddle/fluid/operators/transpose_op.cc @@ -210,18 +210,21 @@ REGISTER_OPERATOR(transpose, ops::TransposeOp, ops::TransposeOpMaker, REGISTER_OPERATOR(transpose_grad, ops::TransposeOpGrad); REGISTER_OP_CPU_KERNEL( - transpose, ops::TransposeKernel); + transpose, ops::TransposeKernel, + ops::TransposeKernel); REGISTER_OP_CPU_KERNEL( transpose_grad, - ops::TransposeGradKernel); + ops::TransposeGradKernel, + ops::TransposeGradKernel); REGISTER_OPERATOR(transpose2, ops::Transpose2Op, ops::Transpose2OpMaker, ops::Transpose2GradMaker); REGISTER_OPERATOR(transpose2_grad, ops::Transpose2OpGrad); REGISTER_OP_CPU_KERNEL( - transpose2, - ops::TransposeKernel); + transpose2, ops::TransposeKernel, + ops::TransposeKernel); REGISTER_OP_CPU_KERNEL( transpose2_grad, - ops::TransposeGradKernel); + ops::TransposeGradKernel, + ops::TransposeGradKernel); diff --git a/paddle/fluid/operators/transpose_op.cu.cc b/paddle/fluid/operators/transpose_op.cu.cc index c1b5a8b31be243fab3af06a18c8e51986c953700..b4025350fa9f3610bde43eee91cd059f3063813f 100644 --- a/paddle/fluid/operators/transpose_op.cu.cc +++ b/paddle/fluid/operators/transpose_op.cu.cc @@ -16,15 +16,18 @@ limitations under the License. */ namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( - transpose, - ops::TransposeKernel); + transpose, ops::TransposeKernel, + ops::TransposeKernel); REGISTER_OP_CUDA_KERNEL( transpose_grad, - ops::TransposeGradKernel); + ops::TransposeGradKernel, + ops::TransposeGradKernel); REGISTER_OP_CUDA_KERNEL( transpose2, - ops::TransposeKernel); + ops::TransposeKernel, + ops::TransposeKernel); REGISTER_OP_CUDA_KERNEL( transpose2_grad, - ops::TransposeGradKernel); + ops::TransposeGradKernel, + ops::TransposeGradKernel); diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.cc b/paddle/fluid/operators/truncated_gaussian_random_op.cc index d854e2803975543b51c50ea2bc173322d3c3ca5e..1e8708f2648d7dd3c10319bd0a4be193d2458d53 100644 --- a/paddle/fluid/operators/truncated_gaussian_random_op.cc +++ b/paddle/fluid/operators/truncated_gaussian_random_op.cc @@ -148,7 +148,7 @@ struct TruncatedNormal { T operator()(T value) const { auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value; - return (std::sqrt(2.0) * Erfinv(2 * p - 1) + mean) * std; + return std::sqrt(2.0) * Erfinv(2 * p - 1) * std + mean; } }; diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.cu b/paddle/fluid/operators/truncated_gaussian_random_op.cu index ad2a9021bfe344d838dff2040b3fb9371274e218..5a3510babe4d57b9e80f0e7898df98033834ca15 100644 --- a/paddle/fluid/operators/truncated_gaussian_random_op.cu +++ b/paddle/fluid/operators/truncated_gaussian_random_op.cu @@ -42,7 +42,7 @@ struct TruncatedNormal { rng.discard(n); T value = dist(rng); auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value; - return (std::sqrt(2.0) * erfinvf(2 * p - 1) + mean) * std; + return std::sqrt(2.0) * erfinvf(2 * p - 1) * std + mean; } }; @@ -52,6 +52,7 @@ class GPUTruncatedGaussianRandomKernel : public framework::OpKernel { void Compute(const framework::ExecutionContext& context) const override { auto* tensor = context.Output("Out"); T* data = tensor->mutable_data(context.GetPlace()); + unsigned int seed = static_cast(context.Attr("seed")); if (seed == 0) { std::random_device rd; diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index 763bb403588d13c15271d26b09813dddf3a5dd8c..e3132ae76f624f3338d749e4fcebbd0ecd7ffe79 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -23,14 +23,14 @@ namespace operators { template class CPUUniformRandomKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const override { - framework::Tensor* tensor = nullptr; + void Compute(const framework::ExecutionContext &ctx) const override { + framework::Tensor *tensor = nullptr; auto out_var = ctx.OutputVar("Out"); if (out_var->IsType()) { tensor = out_var->GetMutable(); } else if (out_var->IsType()) { - auto shape = ctx.Attr>("shape"); - auto* selected_rows = out_var->GetMutable(); + auto shape = ctx.Attr>("shape"); + auto *selected_rows = out_var->GetMutable(); tensor = selected_rows->mutable_value(); tensor->Resize(framework::make_ddim(shape)); selected_rows->mutable_rows()->reserve(shape[0]); @@ -39,7 +39,7 @@ class CPUUniformRandomKernel : public framework::OpKernel { "uniform_random_op's output only" "supports SelectedRows and LoDTensor"); } - T* data = tensor->mutable_data(ctx.GetPlace()); + T *data = tensor->mutable_data(ctx.GetPlace()); unsigned int seed = static_cast(ctx.Attr("seed")); std::minstd_rand engine; if (seed == 0) { @@ -60,14 +60,14 @@ class UniformRandomOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; - void InferShape(framework::InferShapeContext* ctx) const override { + void InferShape(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of UniformRandomOp should not be null."); PADDLE_ENFORCE( ctx->Attrs().Get("min") < ctx->Attrs().Get("max"), "uniform_random's min must less then max"); - auto& shape = ctx->Attrs().Get>("shape"); + auto &shape = ctx->Attrs().Get>("shape"); std::vector temp; temp.reserve(shape.size()); for (auto dim : shape) { @@ -78,7 +78,7 @@ class UniformRandomOp : public framework::OperatorWithKernel { protected: framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { + const framework::ExecutionContext &ctx) const override { return framework::OpKernelType( static_cast(ctx.Attr("dtype")), ctx.GetPlace()); @@ -94,7 +94,7 @@ This operator initializes a tensor with random values sampled from a uniform distribution. The random result is in set [min, max]. )DOC"); - AddAttr>("shape", "The shape of the output tensor"); + AddAttr>("shape", "The shape of the output tensor"); AddAttr("min", "Minimum value of uniform random. [default -1.0].") .SetDefault(-1.0f); AddAttr("max", "Maximun value of uniform random. [default 1.0].") @@ -112,17 +112,17 @@ uniform distribution. The random result is in set [min, max]. class UniformRandomOpVarTypeInference : public framework::VarTypeInference { public: - void operator()(const framework::OpDesc& op_desc, - framework::BlockDesc* block) const override { + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { auto out_var_name = op_desc.Output("Out").front(); - if (block->FindRecursiveOrCreateVar(out_var_name).GetType() == - framework::proto::VarType::SELECTED_ROWS) { - block->FindRecursiveOrCreateVar(out_var_name) - .SetType(framework::proto::VarType::SELECTED_ROWS); - } else { - block->FindRecursiveOrCreateVar(out_var_name) - .SetType(framework::proto::VarType::LOD_TENSOR); + auto var_data_type = static_cast( + boost::get(op_desc.GetAttr("dtype"))); + + auto out_var = block->FindRecursiveOrCreateVar(out_var_name); + if (out_var.GetType() != framework::proto::VarType::SELECTED_ROWS) { + out_var.SetType(framework::proto::VarType::LOD_TENSOR); } + out_var.SetDataType(var_data_type); } }; diff --git a/paddle/fluid/operators/uniform_random_op.cu b/paddle/fluid/operators/uniform_random_op.cu index bbb692b0ddfc18e8a62c0d2a6bac88f9932f6704..2bb0ecc139f7096d1b61150e0a2d4fb095338749 100644 --- a/paddle/fluid/operators/uniform_random_op.cu +++ b/paddle/fluid/operators/uniform_random_op.cu @@ -48,7 +48,7 @@ class GPUUniformRandomKernel : public framework::OpKernel { if (out_var->IsType()) { tensor = out_var->GetMutable(); } else if (out_var->IsType()) { - auto shape = context.Attr>("shape"); + auto shape = context.Attr>("shape"); tensor = out_var->GetMutable()->mutable_value(); tensor->Resize(framework::make_ddim(shape)); } else { diff --git a/paddle/fluid/operators/while_op.cc b/paddle/fluid/operators/while_op.cc index 3c8a01b6e47459760b05b5ca7fa4fa5e1d37d112..aa6af055decc4856fcf2036d324af6b1ff3a5de0 100644 --- a/paddle/fluid/operators/while_op.cc +++ b/paddle/fluid/operators/while_op.cc @@ -129,15 +129,15 @@ class WhileGradOp : public framework::OperatorBase { for (auto cur_scope_iter = step_scopes->rbegin(); cur_scope_iter != step_scopes->rend(); ++cur_scope_iter) { - VLOG(3) << "Start backward at time_step " - << cur_scope_iter - step_scopes->rbegin(); + VLOG(30) << "Start backward at time_step " + << cur_scope_iter - step_scopes->rbegin(); framework::Scope &cur_scope = **cur_scope_iter; // Link OG from outside to inside for (size_t i = 0; i < outside_og_names.size(); ++i) { auto outside_og_name = outside_og_names[i]; auto inside_og_name = inside_og_names[i]; - VLOG(8) << "Linking outside " << outside_og_name << " --> inside " - << inside_og_name; + VLOG(80) << "Linking outside " << outside_og_name << " --> inside " + << inside_og_name; if (scope.FindVar(outside_og_name) == nullptr) { continue; } @@ -159,11 +159,11 @@ class WhileGradOp : public framework::OperatorBase { auto &outside_array = og_outside.Get(); auto &inside_array = detail::Ref(og_inside.GetMutable()); - VLOG(8) << outside_og_name << " size = " << outside_array.size(); + VLOG(80) << outside_og_name << " size = " << outside_array.size(); inside_array.resize(outside_array.size()); for (size_t j = 0; j < inside_array.size(); ++j) { - VLOG(8) << j << " " << outside_array[j].numel(); + VLOG(80) << j << " " << outside_array[j].numel(); if (outside_array[j].numel() != 0) { inside_array[j].set_lod(outside_array[j].lod()); inside_array[j].ShareDataWith(outside_array[j]); @@ -289,7 +289,7 @@ class WhileGradOpDescMaker : public framework::SingleGradOpDescMaker { auto igs = InputGrad(kX, /*do not drop empty gradient*/ false); for (auto &each_ig : igs) { if (inner_op_outputs.find(each_ig) == inner_op_outputs.end()) { - VLOG(8) << "Ignore " << each_ig; + VLOG(80) << "Ignore " << each_ig; each_ig = framework::kEmptyVarName; } } @@ -353,8 +353,8 @@ class WhileGradOpVarTypeInference : public framework::VarTypeInference { auto &p_var = detail::Ref(block->FindVarRecursive(p_names[i])); auto *g_var = block->FindVarRecursive(pg_ig_names[i]); if (g_var != nullptr) { // Gradient could be @EMPTY@ - VLOG(5) << "Setting " << pg_ig_names[i] << " following " << p_names[i] - << " type: " << p_var.GetType(); + VLOG(50) << "Setting " << pg_ig_names[i] << " following " << p_names[i] + << " type: " << p_var.GetType(); g_var->SetType(p_var.GetType()); g_var->SetDataType(p_var.GetDataType()); } diff --git a/paddle/fluid/platform/cpu_info.cc b/paddle/fluid/platform/cpu_info.cc index 2880c09263f10e9c624e11b77188171f48d9db28..b5f472d20f40fa182a4aa55ff384b0954e4ba9e3 100644 --- a/paddle/fluid/platform/cpu_info.cc +++ b/paddle/fluid/platform/cpu_info.cc @@ -128,7 +128,7 @@ bool MayIUse(const cpu_isa_t cpu_isa) { return cpu.has(Cpu::tAVX); case avx2: return cpu.has(Cpu::tAVX2); - case avx512_common: + case avx512f: return cpu.has(Cpu::tAVX512F); case avx512_core: return true && cpu.has(Cpu::tAVX512F) && cpu.has(Cpu::tAVX512BW) && diff --git a/paddle/fluid/platform/cpu_info.h b/paddle/fluid/platform/cpu_info.h index 30c8fbcfce92a8b06a175ddf198cde572f72b2a4..6810a1651a14cdb2080af846b21cad242b70bf35 100644 --- a/paddle/fluid/platform/cpu_info.h +++ b/paddle/fluid/platform/cpu_info.h @@ -43,7 +43,7 @@ typedef enum { sse42, avx, avx2, - avx512_common, + avx512f, avx512_core, avx512_core_vnni, avx512_mic, diff --git a/paddle/fluid/platform/cudnn_helper.h b/paddle/fluid/platform/cudnn_helper.h index bb8b14bb9fa41942c3aa653ca224c0842fbf9a00..07bb02be1962f758e50cab1f27de43e89f3953c3 100644 --- a/paddle/fluid/platform/cudnn_helper.h +++ b/paddle/fluid/platform/cudnn_helper.h @@ -76,8 +76,9 @@ enum class DataLayout { // Not use enum class PoolingMode { kMaximum, - kAverage, kMaximumDeterministic, + kAverageExclusive, + kAverageInclusive, }; #if CUDNN_VERSION < 6000 @@ -91,8 +92,10 @@ inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) { switch (mode) { case PoolingMode::kMaximumDeterministic: return CUDNN_POOLING_MAX; - case PoolingMode::kAverage: + case PoolingMode::kAverageExclusive: return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + case PoolingMode::kAverageInclusive: + return CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING; case PoolingMode::kMaximum: return CUDNN_POOLING_MAX; default: @@ -105,8 +108,10 @@ inline cudnnPoolingMode_t GetPoolingMode(const PoolingMode& mode) { switch (mode) { case PoolingMode::kMaximumDeterministic: return CUDNN_POOLING_MAX_DETERMINISTIC; - case PoolingMode::kAverage: + case PoolingMode::kAverageExclusive: return CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING; + case PoolingMode::kAverageInclusive: + return CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING; case PoolingMode::kMaximum: return CUDNN_POOLING_MAX; default: @@ -341,6 +346,28 @@ class ScopedPoolingDescriptor { DISABLE_COPY_AND_ASSIGN(ScopedPoolingDescriptor); }; +class ScopedSpatialTransformerDescriptor { + public: + ScopedSpatialTransformerDescriptor() { + PADDLE_ENFORCE(dynload::cudnnCreateSpatialTransformerDescriptor(&desc_)); + } + ~ScopedSpatialTransformerDescriptor() { + PADDLE_ENFORCE(dynload::cudnnDestroySpatialTransformerDescriptor(desc_)); + } + + template + inline cudnnSpatialTransformerDescriptor_t descriptor(const int nbDims, + const int dimA[]) { + PADDLE_ENFORCE(dynload::cudnnSetSpatialTransformerNdDescriptor( + desc_, CUDNN_SAMPLER_BILINEAR, CudnnDataType::type, nbDims, dimA)); + return desc_; + } + + private: + cudnnSpatialTransformerDescriptor_t desc_; + DISABLE_COPY_AND_ASSIGN(ScopedSpatialTransformerDescriptor); +}; + inline bool CanCUDNNBeUsed(const framework::ExecutionContext& ctx) { bool use_cudnn = ctx.Attr("use_cudnn"); use_cudnn &= paddle::platform::is_gpu_place(ctx.GetPlace()); diff --git a/paddle/fluid/platform/device_context.cc b/paddle/fluid/platform/device_context.cc index dfc079e986e93c7f02f17b299e5d6293edbedd05..f5541014af5170488efbb10f6e7e331ef015a848 100644 --- a/paddle/fluid/platform/device_context.cc +++ b/paddle/fluid/platform/device_context.cc @@ -32,13 +32,25 @@ platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) { "'Place' is not supported, Please re-compile with WITH_GPU " "option"); } - return it->second.get(); + return it->second.get().get(); +} + +template +inline void EmplaceDeviceContext( + std::map>>* + map_ptr, + platform::Place p) { + using PtrType = std::unique_ptr; + map_ptr->emplace(p, std::async(std::launch::deferred, [=] { + // lazy evaluation. i.e., only create device context at + // first `Get` + return PtrType(new DevCtx(boost::get(p))); + })); } DeviceContextPool::DeviceContextPool( const std::vector& places) { PADDLE_ENFORCE_GT(places.size(), 0); - using PtrType = std::unique_ptr; std::set set; for (auto& p : places) { set.insert(p); @@ -47,16 +59,13 @@ DeviceContextPool::DeviceContextPool( for (auto& p : set) { if (platform::is_cpu_place(p)) { #ifdef PADDLE_WITH_MKLDNN - device_contexts_.emplace( - p, PtrType(new MKLDNNDeviceContext(boost::get(p)))); + EmplaceDeviceContext(&device_contexts_, p); #else - device_contexts_.emplace( - p, PtrType(new CPUDeviceContext(boost::get(p)))); + EmplaceDeviceContext(&device_contexts_, p); #endif } else if (platform::is_gpu_place(p)) { #ifdef PADDLE_WITH_CUDA - device_contexts_.emplace( - p, PtrType(new CUDADeviceContext(boost::get(p)))); + EmplaceDeviceContext(&device_contexts_, p); #else PADDLE_THROW( "'CUDAPlace' is not supported, Please re-compile with WITH_GPU " @@ -64,9 +73,8 @@ DeviceContextPool::DeviceContextPool( #endif } else if (platform::is_cuda_pinned_place(p)) { #ifdef PADDLE_WITH_CUDA - device_contexts_.emplace( - p, - PtrType(new CUDAPinnedDeviceContext(boost::get(p)))); + EmplaceDeviceContext( + &device_contexts_, p); #else PADDLE_THROW( "'CUDAPlace' is not supported, Please re-compile with WITH_GPU " @@ -145,62 +153,38 @@ class EigenCudaStreamDevice : public Eigen::StreamInterface { mutable unsigned int* semaphore_; }; -class CudnnHolder { - public: - CudnnHolder(const cudaStream_t* stream, const CUDAPlace& place) - : workspace_(nullptr), workspace_len_(0), stream_(stream), place_(place) { - PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_)); - PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, *stream_)); - } - - cudnnHandle_t cudnn_handle() const { return cudnn_handle_; } +CudnnHolder::CudnnHolder(const cudaStream_t* stream, const CUDAPlace& place) + : workspace_(nullptr), workspace_len_(0), stream_(stream), place_(place) { + PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_)); + PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, *stream_)); +} - void RunFunc(const std::function& cudnn_func, - size_t required_workspace_len) { - std::lock_guard lock(mtx_); - if (required_workspace_len > workspace_len_) { - ReallocateWorkspace(required_workspace_len); - } - cudnn_func(workspace_); +CudnnHolder::~CudnnHolder() { + PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_)); + if (workspace_ != nullptr) { + paddle::memory::Free(place_, workspace_); } +} - ~CudnnHolder() { - PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_)); - if (workspace_ != nullptr) { - paddle::memory::Free(place_, workspace_); - } +void CudnnHolder::ReallocateWorkspace(size_t required_workspace_len) { + if (required_workspace_len <= workspace_len_) { + return; } - - private: - void ReallocateWorkspace(size_t required_workspace_len) { - if (required_workspace_len <= workspace_len_) { - return; - } - if (workspace_ != nullptr) { - // Maybe someone is using the current workspace - PADDLE_ENFORCE(cudaStreamSynchronize(*stream_)); - paddle::memory::Free(place_, workspace_); - } - workspace_ = paddle::memory::Alloc(place_, required_workspace_len); - workspace_len_ = required_workspace_len; + if (workspace_ != nullptr) { + // Maybe someone is using the current workspace + PADDLE_ENFORCE(cudaStreamSynchronize(*stream_)); + paddle::memory::Free(place_, workspace_); } - - cudnnHandle_t cudnn_handle_; - void* workspace_; - size_t workspace_len_; - - const cudaStream_t* stream_; // not owned; - const CUDAPlace place_; - - std::mutex mtx_; -}; + workspace_ = paddle::memory::Alloc(place_, required_workspace_len); + workspace_len_ = required_workspace_len; +} CUDADeviceContext::CUDADeviceContext(CUDAPlace place) : place_(place), cudnn_holder_(nullptr) { SetDeviceId(place_.device); - compute_capability = GetCUDAComputeCapability(place_.device); - multi_process = GetCUDAMultiProcessors(place_.device); - max_threads_per_mp = GetCUDAMaxThreadsPerMultiProcessor(place_.device); + compute_capability_ = GetCUDAComputeCapability(place_.device); + multi_process_ = GetCUDAMultiProcessors(place_.device); + max_threads_per_mp_ = GetCUDAMaxThreadsPerMultiProcessor(place_.device); PADDLE_ENFORCE(cudaStreamCreate(&stream_)); eigen_stream_.reset(new EigenCudaStreamDevice()); eigen_stream_->Reinitialize(&stream_, place); @@ -211,6 +195,19 @@ CUDADeviceContext::CUDADeviceContext(CUDAPlace place) cudnn_holder_.reset(new CudnnHolder(&stream_, place)); } + driver_version_ = GetCUDADriverVersion(place_.device); + runtime_version_ = GetCUDARuntimeVersion(place_.device); + + LOG_FIRST_N(WARNING, 1) << "Please NOTE: device: " << place_.device + << ", CUDA Capability: " << compute_capability_ + << ", Driver Version: " << driver_version_ / 1000 + << "." << (driver_version_ % 100) / 10 + << ", Runtime Version: " << runtime_version_ / 1000 + << "." << (runtime_version_ % 100) / 10; + size_t cudnn_dso_ver = dynload::cudnnGetVersion(); + LOG_FIRST_N(WARNING, 1) << "device: " << place_.device + << ", cuDNN Version: " << cudnn_dso_ver / 1000 << "." + << (cudnn_dso_ver % 100) / 10 << "."; callback_manager_.reset(new StreamCallbackManager(stream_)); } @@ -232,11 +229,11 @@ void CUDADeviceContext::Wait() const { } int CUDADeviceContext::GetComputeCapability() const { - return compute_capability; + return compute_capability_; } int CUDADeviceContext::GetMaxPhysicalThreadCount() const { - return multi_process * max_threads_per_mp; + return multi_process_ * max_threads_per_mp_; } Eigen::GpuDevice* CUDADeviceContext::eigen_device() const { @@ -251,9 +248,8 @@ cudnnHandle_t CUDADeviceContext::cudnn_handle() const { return cudnn_holder_->cudnn_handle(); } -void CUDADeviceContext::RunCudnnFuncWithWorkspace( - const std::function& cudnn_func, size_t workspace_len) const { - cudnn_holder_->RunFunc(cudnn_func, workspace_len); +CudnnWorkspaceHandle CUDADeviceContext::cudnn_workspace_handle() const { + return CudnnWorkspaceHandle(cudnn_holder_.get()); } cudaStream_t CUDADeviceContext::stream() const { return stream_; } @@ -276,38 +272,73 @@ Place CUDAPinnedDeviceContext::GetPlace() const { return place_; } #ifdef PADDLE_WITH_MKLDNN MKLDNNDeviceContext::MKLDNNDeviceContext(CPUPlace place) - : CPUDeviceContext(place), engine_(mkldnn::engine::cpu, 0), p_blobs_() { - p_blobs_.reset(new std::unordered_map>()); + : CPUDeviceContext(place), engine_(mkldnn::engine::cpu, 0), p_blobmap_() { + p_blobmap_.reset(new BlobMap()); + p_mutex_.reset(new std::mutex()); +} + +namespace { +// Current thread's id. +thread_local int cur_thread_id = 0; } +void set_cur_thread_id(int tid) { cur_thread_id = tid; } +int get_cur_thread_id(void) { return cur_thread_id; } + void MKLDNNDeviceContext::SetBlob(const std::string& name, std::shared_ptr data) const { - std::unordered_map>* p; - p = p_blobs_.get(); + BlobMap* pMap = p_blobmap_.get(); + std::shared_ptr pBlob = nullptr; + + int tid = platform::get_cur_thread_id(); + + std::lock_guard lock(*p_mutex_.get()); - auto it = p->find(name); + // Find KeyBlob for current thread + auto map_it = pMap->find(tid); - if (it == p->end()) { - (*p)[name] = data; // create new blob + if (map_it == pMap->end()) { + // 1st time to set blob in current thread + pBlob = std::shared_ptr(new KeyBlob()); + (*pMap)[tid] = pBlob; } else { - it->second = data; // set data to existing blob + pBlob = map_it->second; } + // Find Key in found (or newly created) KeyBlob + auto key_it = pBlob->find(name); + + if (key_it == pBlob->end()) { + (*pBlob)[name] = data; // create new blob + } else { + key_it->second = data; // set data to existing blob + } + + // lock will be automatically released when out of scope return; } std::shared_ptr MKLDNNDeviceContext::GetBlob( const std::string& name) const { - std::unordered_map>* p; - p = p_blobs_.get(); + BlobMap* pMap = p_blobmap_.get(); + std::shared_ptr pBlob = nullptr; - auto it = p->find(name); + int tid = platform::get_cur_thread_id(); - if (it != p->end()) { - return it->second; - } + std::lock_guard lock(*p_mutex_.get()); + + // Find KeyBlob for current thread firstly + auto map_it = pMap->find(tid); + if (map_it == pMap->end()) return nullptr; + pBlob = map_it->second; + + // Find Blob via name + auto key_it = pBlob->find(name); + + if (key_it == pBlob->end()) return nullptr; - return nullptr; + // lock will be automatically released when out of scope + return key_it->second; } #endif diff --git a/paddle/fluid/platform/device_context.h b/paddle/fluid/platform/device_context.h index 79539195157d74d4d757edee5e008cbb76c93ee2..df248f9bb15591d5015ad01278797ec7e31ef9d1 100644 --- a/paddle/fluid/platform/device_context.h +++ b/paddle/fluid/platform/device_context.h @@ -10,6 +10,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include // NOLINT #include #include // NOLINT #include @@ -72,7 +73,60 @@ struct DefaultDeviceContextType { #ifdef PADDLE_WITH_CUDA class EigenCudaStreamDevice; -class CudnnHolder; +class CudnnHolder { + public: + CudnnHolder(const cudaStream_t* stream, const CUDAPlace& place); + ~CudnnHolder(); + cudnnHandle_t cudnn_handle() const { return cudnn_handle_; } + + private: + friend class CudnnWorkspaceHandle; + void ReallocateWorkspace(size_t required_workspace_len); + + template + void RunFuncImpl(Callback&& cudnn_func, size_t required_workspace_len) { + if (required_workspace_len > workspace_len_) { + ReallocateWorkspace(required_workspace_len); + } + cudnn_func(workspace_); + } + + std::mutex& Mutex() { return mtx_; } + + cudnnHandle_t cudnn_handle_; + void* workspace_; + size_t workspace_len_; + + const cudaStream_t* stream_; // not owned; + const CUDAPlace place_; + + std::mutex mtx_; +}; + +class CudnnWorkspaceHandle { + public: + /*! \brief The lock would not be acquired when constructor calls. + * The lock would be acquired when RunFunc() is called first time. */ + inline explicit CudnnWorkspaceHandle(CudnnHolder* holder) : holder_(holder) {} + + /*! \brief Thread which call RunFunc() would acquire the lock first + * before invoking cudnn functions. */ + template + inline void RunFunc(Callback&& cudnn_func, size_t required_workspace_len) { + if (!guard_) { + guard_.reset(new std::lock_guard(holder_->Mutex())); + } + holder_->RunFuncImpl(std::forward(cudnn_func), + required_workspace_len); + } + + CudnnWorkspaceHandle(CudnnWorkspaceHandle&&) = default; + CudnnWorkspaceHandle& operator=(CudnnWorkspaceHandle&&) = delete; + + private: + CudnnHolder* holder_; // not own + std::unique_ptr> guard_; +}; class CUDADeviceContext : public DeviceContext { public: @@ -100,10 +154,14 @@ class CUDADeviceContext : public DeviceContext { /*! \brief Return cudnn handle in the device context. */ cudnnHandle_t cudnn_handle() const; - /*! \brief Run a cudnn function with the workspace provided by - * CUDADeviceContext */ - void RunCudnnFuncWithWorkspace(const std::function& cudnn_func, - size_t workspace_len) const; + /*! \brief Return a cudnn workspace handle to call multiple cudnn + * functions without interrupting by other threads. + * Once the first cudnn function is called by the handle, a lock + * would be acquired to prevent other threads from accessing the + * workspace. Once the handle is destructed, the lock would be released. + * CudnnWorkspaceHandle is an RAII object to implement thread-safe + * sequential cudnn function calls. */ + CudnnWorkspaceHandle cudnn_workspace_handle() const; /*! \brief Return cuda stream in the device context. */ cudaStream_t stream() const; @@ -135,9 +193,11 @@ class CUDADeviceContext : public DeviceContext { cudaStream_t stream_; cublasHandle_t cublas_handle_; - int compute_capability; - int multi_process; - int max_threads_per_mp; + int compute_capability_; + int runtime_version_; + int driver_version_; + int multi_process_; + int max_threads_per_mp_; mutable std::mutex mtx_; @@ -174,6 +234,12 @@ struct DefaultDeviceContextType { #endif #ifdef PADDLE_WITH_MKLDNN +using KeyBlob = std::unordered_map>; +using BlobMap = std::unordered_map>; + +void set_cur_thread_id(int); +int get_cur_thread_id(void); + class MKLDNNDeviceContext : public CPUDeviceContext { public: explicit MKLDNNDeviceContext(CPUPlace place); @@ -189,8 +255,8 @@ class MKLDNNDeviceContext : public CPUDeviceContext { private: mkldnn::engine engine_; - std::shared_ptr>> - p_blobs_; + std::shared_ptr p_blobmap_; + std::shared_ptr p_mutex_; }; #endif @@ -226,7 +292,8 @@ class DeviceContextPool { private: static DeviceContextPool* pool; - std::map> device_contexts_; + std::map>> + device_contexts_; DISABLE_COPY_AND_ASSIGN(DeviceContextPool); }; diff --git a/paddle/fluid/platform/device_tracer.cc b/paddle/fluid/platform/device_tracer.cc index dc1d751141187edb7738e42c41514614d4d399b0..ea4564058d602a9abe43bd063f1ed73f88a2de08 100644 --- a/paddle/fluid/platform/device_tracer.cc +++ b/paddle/fluid/platform/device_tracer.cc @@ -203,7 +203,7 @@ class DeviceTracerImpl : public DeviceTracer { void AddCPURecords(const std::string &anno, uint64_t start_ns, uint64_t end_ns, int64_t device_id, int64_t thread_id) { if (anno.empty()) { - VLOG(1) << "Empty timeline annotation."; + VLOG(10) << "Empty timeline annotation."; return; } std::lock_guard l(trace_mu_); @@ -216,7 +216,7 @@ class DeviceTracerImpl : public DeviceTracer { 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) { - VLOG(3) << name << " cannot be traced"; + VLOG(30) << name << " cannot be traced"; return; } std::lock_guard l(trace_mu_); @@ -228,7 +228,7 @@ class DeviceTracerImpl : public DeviceTracer { int64_t stream_id, uint32_t correlation_id) { // 0 means timestamp information could not be collected for the kernel. if (start == 0 || end == 0) { - VLOG(3) << correlation_id << " cannot be traced"; + VLOG(30) << correlation_id << " cannot be traced"; return; } std::lock_guard l(trace_mu_); @@ -347,7 +347,7 @@ class DeviceTracerImpl : public DeviceTracer { tracer->AddAnnotation(cbInfo->correlationId, anno); } } else { - VLOG(1) << "Unhandled API Callback for " << domain << " " << cbid; + VLOG(10) << "Unhandled API Callback for " << domain << " " << cbid; } } CUpti_SubscriberHandle subscriber_; diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h index e6353f67ef118072a2d8e49111e8ecc486589998..c26143d2f2780f3042f66b99808c6b85866f9dc4 100644 --- a/paddle/fluid/platform/dynload/cudnn.h +++ b/paddle/fluid/platform/dynload/cudnn.h @@ -65,44 +65,54 @@ extern void EnforceCUDNNLoaded(const char* fn_name); * include all needed cudnn functions in HPPL * different cudnn version has different interfaces **/ -#define CUDNN_DNN_ROUTINE_EACH(__macro) \ - __macro(cudnnSetTensor4dDescriptor); \ - __macro(cudnnSetTensor4dDescriptorEx); \ - __macro(cudnnSetTensorNdDescriptor); \ - __macro(cudnnGetTensorNdDescriptor); \ - __macro(cudnnGetConvolutionNdForwardOutputDim); \ - __macro(cudnnGetConvolutionForwardAlgorithm); \ - __macro(cudnnCreateTensorDescriptor); \ - __macro(cudnnDestroyTensorDescriptor); \ - __macro(cudnnCreateFilterDescriptor); \ - __macro(cudnnSetFilter4dDescriptor); \ - __macro(cudnnSetFilterNdDescriptor); \ - __macro(cudnnGetFilterNdDescriptor); \ - __macro(cudnnSetPooling2dDescriptor); \ - __macro(cudnnSetPoolingNdDescriptor); \ - __macro(cudnnGetPoolingNdDescriptor); \ - __macro(cudnnDestroyFilterDescriptor); \ - __macro(cudnnCreateConvolutionDescriptor); \ - __macro(cudnnCreatePoolingDescriptor); \ - __macro(cudnnDestroyPoolingDescriptor); \ - __macro(cudnnSetConvolution2dDescriptor); \ - __macro(cudnnDestroyConvolutionDescriptor); \ - __macro(cudnnSetConvolutionNdDescriptor); \ - __macro(cudnnGetConvolutionNdDescriptor); \ - __macro(cudnnDeriveBNTensorDescriptor); \ - __macro(cudnnCreate); \ - __macro(cudnnDestroy); \ - __macro(cudnnSetStream); \ - __macro(cudnnActivationForward); \ - __macro(cudnnConvolutionForward); \ - __macro(cudnnConvolutionBackwardBias); \ - __macro(cudnnGetConvolutionForwardWorkspaceSize); \ - __macro(cudnnTransformTensor); \ - __macro(cudnnPoolingForward); \ - __macro(cudnnPoolingBackward); \ - __macro(cudnnSoftmaxBackward); \ - __macro(cudnnSoftmaxForward); \ - __macro(cudnnGetVersion); \ +#define CUDNN_DNN_ROUTINE_EACH(__macro) \ + __macro(cudnnSetTensor4dDescriptor); \ + __macro(cudnnSetTensor4dDescriptorEx); \ + __macro(cudnnSetTensorNdDescriptor); \ + __macro(cudnnGetTensorNdDescriptor); \ + __macro(cudnnGetConvolutionNdForwardOutputDim); \ + __macro(cudnnGetConvolutionForwardAlgorithm); \ + __macro(cudnnCreateTensorDescriptor); \ + __macro(cudnnDestroyTensorDescriptor); \ + __macro(cudnnCreateFilterDescriptor); \ + __macro(cudnnSetFilter4dDescriptor); \ + __macro(cudnnSetFilterNdDescriptor); \ + __macro(cudnnGetFilterNdDescriptor); \ + __macro(cudnnSetPooling2dDescriptor); \ + __macro(cudnnSetPoolingNdDescriptor); \ + __macro(cudnnGetPoolingNdDescriptor); \ + __macro(cudnnDestroyFilterDescriptor); \ + __macro(cudnnCreateConvolutionDescriptor); \ + __macro(cudnnCreatePoolingDescriptor); \ + __macro(cudnnDestroyPoolingDescriptor); \ + __macro(cudnnSetConvolution2dDescriptor); \ + __macro(cudnnDestroyConvolutionDescriptor); \ + __macro(cudnnSetConvolutionNdDescriptor); \ + __macro(cudnnGetConvolutionNdDescriptor); \ + __macro(cudnnDeriveBNTensorDescriptor); \ + __macro(cudnnCreateSpatialTransformerDescriptor); \ + __macro(cudnnSetSpatialTransformerNdDescriptor); \ + __macro(cudnnDestroySpatialTransformerDescriptor); \ + __macro(cudnnSpatialTfGridGeneratorForward); \ + __macro(cudnnSpatialTfGridGeneratorBackward); \ + __macro(cudnnSpatialTfSamplerForward); \ + __macro(cudnnSpatialTfSamplerBackward); \ + __macro(cudnnCreate); \ + __macro(cudnnDestroy); \ + __macro(cudnnSetStream); \ + __macro(cudnnActivationForward); \ + __macro(cudnnConvolutionForward); \ + __macro(cudnnConvolutionBackwardBias); \ + __macro(cudnnGetConvolutionForwardWorkspaceSize); \ + __macro(cudnnTransformTensor); \ + __macro(cudnnPoolingForward); \ + __macro(cudnnPoolingBackward); \ + __macro(cudnnSoftmaxBackward); \ + __macro(cudnnSoftmaxForward); \ + __macro(cudnnGetVersion); \ + __macro(cudnnFindConvolutionForwardAlgorithmEx); \ + __macro(cudnnFindConvolutionBackwardFilterAlgorithmEx); \ + __macro(cudnnFindConvolutionBackwardDataAlgorithmEx); \ __macro(cudnnGetErrorString); CUDNN_DNN_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP) diff --git a/paddle/fluid/platform/dynload/dynamic_loader.cc b/paddle/fluid/platform/dynload/dynamic_loader.cc index cc5cda6106c188f3156d33480b5d3641eed32556..d53907b749805d9c16737da3105d6c66cacb12fb 100644 --- a/paddle/fluid/platform/dynload/dynamic_loader.cc +++ b/paddle/fluid/platform/dynload/dynamic_loader.cc @@ -72,8 +72,8 @@ static inline std::string join(const std::string& part1, static inline void* GetDsoHandleFromDefaultPath(const std::string& dso_path, int dynload_flags) { - VLOG(3) << "Try to find library: " << dso_path - << " from default system path."; + VLOG(30) << "Try to find library: " << dso_path + << " from default system path."; // default search from LD_LIBRARY_PATH/DYLD_LIBRARY_PATH // and /usr/local/lib path void* dso_handle = dlopen(dso_path.c_str(), dynload_flags); diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h index f04395a8ac00f33501008aa12f22773ddda9b138..a251bfcd9914422cb6300adbbcdef3dfa79f441c 100644 --- a/paddle/fluid/platform/enforce.h +++ b/paddle/fluid/platform/enforce.h @@ -130,6 +130,13 @@ struct EOFException : public std::exception { #define UNLIKELY(condition) (condition == 0) #endif +#if !defined(_WIN32) +#define LIKELY(condition) __builtin_expect(static_cast(condition), 1) +#else +// there is no equivalent intrinsics in msvc. +#define LIKELY(condition) (condition != 0) +#endif + template inline typename std::enable_if::type throw_on_error( bool stat, const Args&... args) { diff --git a/paddle/fluid/platform/gpu_info.cc b/paddle/fluid/platform/gpu_info.cc index f599e7fbc886a60394ae4690e4160275b55b8596..c78f159ad25a17b38333a57a0650d9843c4c5632 100644 --- a/paddle/fluid/platform/gpu_info.cc +++ b/paddle/fluid/platform/gpu_info.cc @@ -46,6 +46,24 @@ int GetCUDAComputeCapability(int id) { return device_prop.major * 10 + device_prop.minor; } +int GetCUDARuntimeVersion(int id) { + PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count"); + int runtime_version = 0; + PADDLE_ENFORCE(cudaRuntimeGetVersion(&runtime_version), + "cudaRuntimeGetVersion failed in " + "paddle::platform::cudaRuntimeGetVersion"); + return runtime_version; +} + +int GetCUDADriverVersion(int id) { + PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count"); + int driver_version = 0; + PADDLE_ENFORCE(cudaDriverGetVersion(&driver_version), + "cudaDriverGetVersion failed in " + "paddle::platform::GetCUDADriverVersion"); + return driver_version; +} + int GetCUDAMultiProcessors(int id) { PADDLE_ENFORCE_LT(id, GetCUDADeviceCount(), "id must less than GPU count"); int count; @@ -106,8 +124,8 @@ size_t GpuMaxChunkSize() { size_t available = 0; GpuMemoryUsage(&available, &total); - VLOG(10) << "GPU Usage " << available / 1024 / 1024 << "M/" - << total / 1024 / 1024 << "M"; + VLOG(100) << "GPU Usage " << available / 1024 / 1024 << "M/" + << total / 1024 / 1024 << "M"; size_t reserving = static_cast(0.05 * total); // If available less than minimum chunk size, no usable memory exists. available = diff --git a/paddle/fluid/platform/gpu_info.h b/paddle/fluid/platform/gpu_info.h index f4640d3eaa2165c35e8e14690d83e9e7e7168c0b..be44158431ff80a41f7fdf4dfd4d070667f2ac63 100644 --- a/paddle/fluid/platform/gpu_info.h +++ b/paddle/fluid/platform/gpu_info.h @@ -29,6 +29,12 @@ int GetCUDADeviceCount(); //! Get the compute capability of the ith GPU (format: major * 10 + minor) int GetCUDAComputeCapability(int i); +//! Get the runtime version of the ith GPU +int GetCUDARuntimeVersion(int id); + +//! Get the driver version of the ith GPU +int GetCUDADriverVersion(int id); + //! Get the MultiProcessors of the ith GPU. int GetCUDAMultiProcessors(int i); diff --git a/paddle/fluid/platform/init.cc b/paddle/fluid/platform/init.cc index 4c99f4be321160caf0ee2f89a655bdfb933408e3..4cbfe0a69c06cb6793c877263b2feaafa7c3dc60 100644 --- a/paddle/fluid/platform/init.cc +++ b/paddle/fluid/platform/init.cc @@ -45,7 +45,7 @@ void InitGflags(std::vector argv) { line += ' '; } google::ParseCommandLineFlags(&argc, &arr, true); - VLOG(1) << "Init commandline: " << line; + VLOG(10) << "Init commandline: " << line; }); } @@ -116,21 +116,51 @@ void InitDevices(bool init_p2p, const std::vector devices) { platform::SetNumThreads(FLAGS_paddle_num_threads); #endif - if (platform::jit::MayIUse(platform::jit::avx512_common)) { -#ifndef __AVX512F__ - LOG(WARNING) << "AVX512F is available, Please re-compile on local machine"; +#if !defined(_WIN32) && !defined(__APPLE__) && !defined(__OSX__) + if (platform::jit::MayIUse(platform::jit::avx)) { +#ifndef __AVX__ + LOG(WARNING) << "AVX is available, Please re-compile on local machine"; #endif } - if (platform::jit::MayIUse(platform::jit::avx2)) { -#ifndef __AVX2__ - LOG(WARNING) << "AVX2 is available, Please re-compile on local machine"; + +// Throw some informations when CPU instructions mismatch. +#define AVX_GUIDE(compiletime, runtime) \ + LOG(FATAL) \ + << "This version is compiled on higher instruction(" #compiletime \ + ") system, you may encounter illegal instruction error running on" \ + " your local CPU machine. Please reinstall the " #runtime \ + " version or compile from source code." + +#ifdef __AVX512F__ + if (!platform::jit::MayIUse(platform::jit::avx512f)) { + if (platform::jit::MayIUse(platform::jit::avx2)) { + AVX_GUIDE(AVX512, AVX2); + } else if (platform::jit::MayIUse(platform::jit::avx)) { + AVX_GUIDE(AVX512, AVX); + } else { + AVX_GUIDE(AVX512, NonAVX); + } + } #endif + +#ifdef __AVX2__ + if (!platform::jit::MayIUse(platform::jit::avx2)) { + if (platform::jit::MayIUse(platform::jit::avx)) { + AVX_GUIDE(AVX2, AVX); + } else { + AVX_GUIDE(AVX2, NonAVX); + } } - if (platform::jit::MayIUse(platform::jit::avx)) { -#ifndef __AVX__ - LOG(WARNING) << "AVX is available, Please re-compile on local machine"; #endif + +#ifdef __AVX__ + if (!platform::jit::MayIUse(platform::jit::avx)) { + AVX_GUIDE(AVX, NonAVX); } +#endif +#undef AVX_GUIDE + +#endif } void InitGLOG(const std::string &prog_name) { diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index c0a2543ba5d8ff8f34cb6231c51cb5053a6a9481..814012e6c1fad414d10f5a64af283bed57e11fe3 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -187,6 +187,29 @@ class MKLDNNHandler { return mem_p; } + std::shared_ptr AcquireMemory( + const std::shared_ptr& user_memory_p, + const std::shared_ptr& target_memory_p, + const std::string& suffix, + std::vector& pipeline) { // NOLINT + auto local_key = key_ + suffix; + auto key_reorder_p = key_ + suffix + "reorder_p"; + + auto stored_reorder_p = std::static_pointer_cast( + dev_ctx_.GetBlob(key_reorder_p)); + + if (stored_reorder_p) { + pipeline.push_back(*stored_reorder_p); + } else { + auto reorder_p = + std::make_shared(*user_memory_p, *target_memory_p); + dev_ctx_.SetBlob(key_reorder_p, reorder_p); + pipeline.push_back(*reorder_p); + } + + return target_memory_p; + } + std::shared_ptr AcquireMemory( mkldnn::memory::primitive_desc& mpd, // NOLINT mkldnn::memory::primitive_desc& user_mpd, // NOLINT diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h index 115abb98d56e633c938695c8127c832eab602110..40af1f95208905231b933e5184a807b061164799 100644 --- a/paddle/fluid/platform/nccl_helper.h +++ b/paddle/fluid/platform/nccl_helper.h @@ -112,7 +112,7 @@ struct NCCLContextMap { NCCLGroupGuard gurad; for (auto &gpu_id : order_) { int rank = trainer_id * order_.size() + gpu_id; - VLOG(3) << "init nccl rank: " << rank << " nranks: " << nranks; + VLOG(30) << "init nccl rank: " << rank << " nranks: " << nranks; PADDLE_ENFORCE(cudaSetDevice(gpu_id)); PADDLE_ENFORCE(platform::dynload::ncclCommInitRank( comms.get() + gpu_id, nranks, *nccl_id, rank)); diff --git a/paddle/fluid/platform/profiler.cc b/paddle/fluid/platform/profiler.cc index 652a6ec7a4e2e823b28f39b449570cd375e88e18..56bf9e31a35fdec5b7f04849068ff96ac9776c0e 100644 --- a/paddle/fluid/platform/profiler.cc +++ b/paddle/fluid/platform/profiler.cc @@ -30,6 +30,8 @@ limitations under the License. */ #include "paddle/fluid/platform/device_tracer.h" #include "paddle/fluid/string/printf.h" +DEFINE_bool(enable_rpc_profiler, false, "Enable rpc profiler or not."); + namespace paddle { namespace platform { @@ -193,6 +195,13 @@ RecordEvent::~RecordEvent() { PopEvent(name_, dev_ctx_); } +RecordRPCEvent::RecordRPCEvent(const std::string& name, + const DeviceContext* dev_ctx) { + if (FLAGS_enable_rpc_profiler) { + event_.reset(new platform::RecordEvent(name, dev_ctx)); + } +} + RecordBlock::RecordBlock(int block_id) : is_enabled_(false), start_ns_(PosixInNsec()) { std::lock_guard l(profiler_mu); @@ -217,7 +226,7 @@ RecordBlock::~RecordBlock() { void EnableProfiler(ProfilerState state) { PADDLE_ENFORCE(state != ProfilerState::kDisabled, - "Can't enbale profling, since the input state is ", + "Can't enable profiling, since the input state is ", "ProfilerState::kDisabled"); std::lock_guard l(profiler_mu); @@ -276,7 +285,7 @@ struct EventItem { // Print results void PrintProfiler(const std::vector>& events_table, const std::string& sorted_domain, const size_t name_width, - const size_t data_width, double total) { + const size_t data_width, bool merge_thread) { // Output header information std::cout << "\n------------------------->" << " Profiling Report " @@ -292,6 +301,10 @@ void PrintProfiler(const std::vector>& events_table, PADDLE_THROW("Invalid profiler state", g_state); } + if (merge_thread) { + std::cout << "Note! This Report merge all thread info into one." + << std::endl; + } std::cout << "Place: " << place << std::endl; std::cout << "Time unit: ms" << std::endl; std::cout << "Sorted by " << sorted_domain @@ -312,8 +325,7 @@ void PrintProfiler(const std::vector>& events_table, << std::setw(data_width) << event_item.min_time << std::setw(data_width) << event_item.max_time << std::setw(data_width) << event_item.ave_time - << std::setw(data_width) << event_item.total_time / total - << std::endl; + << std::setw(data_width) << event_item.ratio << std::endl; } } std::cout << std::endl; @@ -321,8 +333,10 @@ void PrintProfiler(const std::vector>& events_table, // Parse the event list and output the profiling report void ParseEvents(const std::vector>& events, + bool merge_thread, EventSortingKey sorted_by = EventSortingKey::kDefault) { if (g_state == ProfilerState::kDisabled) return; + if (merge_thread && events.size() < 2) return; std::string sorted_domain; std::function sorted_func; @@ -361,34 +375,55 @@ void ParseEvents(const std::vector>& events, sorted_domain = "event first end time"; } + const std::vector>* analyze_events; + std::vector> merged_events_list; + if (merge_thread) { + std::vector merged_events; + for (size_t i = 0; i < events.size(); ++i) { + for (size_t j = 0; j < events[i].size(); ++j) { + merged_events.push_back(events[i][j]); + } + } + merged_events_list.push_back(merged_events); + analyze_events = &merged_events_list; + } else { + analyze_events = &events; + } + std::vector> events_table; size_t max_name_width = 0; - double total = 0.; // the total time - for (size_t i = 0; i < events.size(); i++) { + for (size_t i = 0; i < (*analyze_events).size(); i++) { + double total = 0.; // the total time in one thread std::list pushed_events; std::vector event_items; std::unordered_map event_idx; - for (size_t j = 0; j < events[i].size(); j++) { - if (events[i][j].type() == EventType::kPushRange) { - pushed_events.push_back(events[i][j]); - } else if (events[i][j].type() == EventType::kPopRange) { + for (size_t j = 0; j < (*analyze_events)[i].size(); j++) { + if ((*analyze_events)[i][j].type() == EventType::kPushRange) { + pushed_events.push_back((*analyze_events)[i][j]); + } else if ((*analyze_events)[i][j].type() == EventType::kPopRange) { std::list::reverse_iterator rit = pushed_events.rbegin(); while (rit != pushed_events.rend() && - rit->name() != events[i][j].name()) { + rit->name() != (*analyze_events)[i][j].name()) { ++rit; } if (rit != pushed_events.rend()) { double event_time = (g_state == ProfilerState::kCUDA || g_state == ProfilerState::kAll) - ? rit->CudaElapsedMs(events[i][j]) - : rit->CpuElapsedMs(events[i][j]); + ? rit->CudaElapsedMs((*analyze_events)[i][j]) + : rit->CpuElapsedMs((*analyze_events)[i][j]); total += event_time; - std::string event_name = - "thread" + std::to_string(rit->thread_id()) + "::" + rit->name(); - max_name_width = std::max(max_name_width, event_name.size()); + std::string event_name; + if (merge_thread) { + event_name = rit->name(); + max_name_width = std::max(max_name_width, event_name.size()); + } else { + event_name = "thread" + std::to_string(rit->thread_id()) + "::" + + rit->name(); + max_name_width = std::max(max_name_width, event_name.size()); + } if (event_idx.find(event_name) == event_idx.end()) { event_idx[event_name] = event_items.size(); @@ -413,7 +448,7 @@ void ParseEvents(const std::vector>& events, pushed_events.erase((++rit).base()); } else { LOG(WARNING) << "Cannot find the push marker of event \'" - << events[i][j].name() + << (*analyze_events)[i][j].name() << "\', which will be ignored in profiling report."; } } @@ -421,6 +456,7 @@ void ParseEvents(const std::vector>& events, // average time for (auto& item : event_items) { item.ave_time = item.total_time / item.calls; + item.ratio = item.total_time / total; } // sort if (sorted_by != EventSortingKey::kDefault) { @@ -438,7 +474,8 @@ void ParseEvents(const std::vector>& events, } // Print report - PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12, total); + PrintProfiler(events_table, sorted_domain, max_name_width + 4, 12, + merge_thread); } void DisableProfiler(EventSortingKey sorted_key, @@ -449,7 +486,8 @@ void DisableProfiler(EventSortingKey sorted_key, Mark("_stop_profiler_", nullptr); std::vector> all_events = GetAllEvents(); - ParseEvents(all_events, sorted_key); + ParseEvents(all_events, true, sorted_key); + ParseEvents(all_events, false, sorted_key); ResetProfiler(); DeviceTracer* tracer = GetDeviceTracer(); if (tracer->IsEnabled()) { diff --git a/paddle/fluid/platform/profiler.h b/paddle/fluid/platform/profiler.h index 38630686f7cf3c669373f941d989adf11ba6cfe6..e8eae874afa3d17f0d3374eef457cdbacb3f8424 100644 --- a/paddle/fluid/platform/profiler.h +++ b/paddle/fluid/platform/profiler.h @@ -71,6 +71,7 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx); #if !defined(_WIN32) struct RecordEvent { + // dev_ctx can be set to nullptr if device is cpu. RecordEvent(const std::string& name, const DeviceContext* dev_ctx); ~RecordEvent(); @@ -86,6 +87,16 @@ struct RecordEvent { std::string full_name_; }; +class RecordRPCEvent { + public: + // dev_ctx can be set to nullptr if device is cpu. + RecordRPCEvent(const std::string& name, const DeviceContext* dev_ctx); + ~RecordRPCEvent() {} + + private: + std::unique_ptr event_; +}; + struct RecordBlock { explicit RecordBlock(int block_id); ~RecordBlock(); diff --git a/paddle/fluid/platform/stream_callback_manager.h b/paddle/fluid/platform/stream_callback_manager.h index 6c984065aa5fa1a8875aebe84051ab396bc417ec..0e88a439cf6ca83e3d98725f58875adeeea86be0 100644 --- a/paddle/fluid/platform/stream_callback_manager.h +++ b/paddle/fluid/platform/stream_callback_manager.h @@ -24,8 +24,6 @@ namespace paddle { namespace platform { -using StreamCallback = std::function; - class StreamCallbackManager; struct StreamCallbackContext { @@ -35,7 +33,7 @@ struct StreamCallbackContext { : manager_(manager), callback_(callback) {} const StreamCallbackManager *manager_; // do not own - StreamCallback callback_; + std::function callback_; }; class StreamCallbackManager { @@ -45,16 +43,18 @@ class StreamCallbackManager { template inline void AddCallback(Callback &&callback) const { - AddCallbackWithStreamAndErrorInfo( - [=](cudaStream_t, cudaError_t) { callback(); }); - } - - template - inline void AddCallbackWithStreamAndErrorInfo(Callback &&callback) const { - auto *stream_callback_context = new StreamCallbackContext(this, callback); - PADDLE_ENFORCE(cudaStreamAddCallback( - stream_, StreamCallbackManager::StreamCallbackFunc, - stream_callback_context, 0)); + auto *stream_callback_context = + new StreamCallbackContext(this, std::forward(callback)); + PADDLE_ENFORCE( +#if CUDA_VERSION >= 10000 + cudaLaunchHostFunc(stream_, StreamCallbackManager::StreamCallbackFunc, + stream_callback_context) +#else + cudaStreamAddCallback(stream_, + StreamCallbackManager::StreamCallbackFunc, + stream_callback_context, 0) +#endif + ); // NOLINT } void Wait() const { thread_pool_.reset(new ThreadPool(1)); } @@ -63,17 +63,21 @@ class StreamCallbackManager { const cudaStream_t stream_; mutable std::unique_ptr thread_pool_; - // cudaStreamCallback cannot call CUDA API inside, so we have to use - // thread_pool here +// cudaStreamCallback cannot call CUDA API inside, so we have to use +// thread_pool here +#if CUDA_VERSION >= 10000 + static void CUDART_CB StreamCallbackFunc(void *user_data) +#else static void CUDART_CB StreamCallbackFunc(cudaStream_t stream, - cudaError_t status, - void *user_data) { + cudaError_t status, void *user_data) +#endif + { auto *callback_context_ptr = reinterpret_cast(user_data); callback_context_ptr->manager_->thread_pool_->enqueue([=]() { std::unique_ptr callback_context( callback_context_ptr); - callback_context->callback_(stream, status); + callback_context->callback_(); }); } }; diff --git a/paddle/fluid/pybind/const_value.cc b/paddle/fluid/pybind/const_value.cc index 1f61a0e289f32196ead04d71d07b513cbe4655b1..06d8b65fb1480d9f621ca937c1d66ab7e910f010 100644 --- a/paddle/fluid/pybind/const_value.cc +++ b/paddle/fluid/pybind/const_value.cc @@ -27,6 +27,7 @@ void BindConstValue(pybind11::module* m) { m->def("kZeroVarSuffix", [] { return framework::kZeroVarSuffix; }); m->def("kControlDepVarName", [] { return framework::ir::Node::kControlDepVarName; }); + m->def("kNewGradSuffix", [] { return framework::kNewGradSuffix; }); auto op_proto_and_checker_maker = m->def_submodule("op_proto_and_checker_maker"); diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index 3b22718a8c6f994dbc2dc3e7aaa19a7163f716ba..586e92c2b3146d75a673d1fe326dbee7297a3bfb 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -57,6 +57,18 @@ struct variant_caster> { auto caster = make_caster(); if (!load_success_ && caster.load(src, convert)) { load_success_ = true; + + if (std::is_same>::value) { + auto caster_ints = make_caster>(); + if (caster_ints.load(src, convert)) { + VLOG(40) << "This value are floats and int64_ts satisfy " + "simultaneously, will set it's type to " + "std::vector"; + value = cast_op>(caster_ints); + return true; + } + } + value = cast_op(caster); return true; } @@ -259,6 +271,8 @@ void BindOpDesc(pybind11::module *m) { pybind11::enum_(*m, "AttrType", "") .value("INT", pd::proto::AttrType::INT) .value("INTS", pd::proto::AttrType::INTS) + .value("LONG", pd::proto::AttrType::LONG) + .value("LONGS", pd::proto::AttrType::LONGS) .value("FLOAT", pd::proto::AttrType::FLOAT) .value("FLOATS", pd::proto::AttrType::FLOATS) .value("STRING", pd::proto::AttrType::STRING) diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 311cd944603e9bdfefef4daa3a9c690df5b30235..238cc19189cfd74afa38bdcb5f5c802f9521dfea 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -57,6 +57,10 @@ limitations under the License. */ #include "pybind11/stl.h" +DEFINE_bool(reader_queue_speed_test_mode, false, + "If set true, the queue.pop will only get data from queue but not " + "remove the data from queue for speed testing"); + // disable auto conversion to list in Python PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray); @@ -157,7 +161,50 @@ PYBIND11_PLUGIN(core) { .def("_get_double_element", TensorGetElement) .def("_dtype", [](Tensor &self) { return ToDataType(self.type()); }); - py::class_(m, "LoDTensor") + py::class_(m, "LoDTensor", R"DOC( + LoDTensor is a Tensor with optional LoD information. + + np.array(lod_tensor) can convert LoDTensor to numpy array. + lod_tensor.lod() can retrieve the LoD information. + + LoD is short for Level of Details and is usually used for varied sequence + length. You can skip the following comment if you don't need optional LoD. + + For example: + A LoDTensor X can look like the example below. It contains 2 sequences. + The first has length 2 and the second has length 3, as described by x.lod. + + The first tensor dimension 5=2+3 is calculated from LoD if it's available. + It means the total number of sequence element. In X, each element has 2 + columns, hence [5, 2]. + + x.lod = [[2, 3]] + x.data = [[1, 2], [3, 4], + [5, 6], [7, 8], [9, 10]] + x.shape = [5, 2] + + LoD can have multiple levels (for example, a paragraph can have multiple + sentences and a sentence can have multiple words). In the following + LodTensor Y, the lod_level is 2. It means there are 2 sequence, the + first sequence length is 2 (has 2 sub-sequences), the second one's + length is 1. The first sequence's 2 sub-sequences have length 2 and 2, + respectively. And the second sequence's 1 sub-sequence has length 3. + + y.lod = [[2 1], [2 2 3]] + y.shape = [2+2+3, ...] + + Note: + In above description, LoD is length-based. In Paddle internal + implementation, lod is offset-based. Hence, internally, + y.lod is represented as [[0, 2, 3], [0, 2, 4, 7]] (length-based + equivlent would be [[2-0, 3-2], [2-0, 4-2, 7-4]]). + + Sometimes LoD is called recursive_sequence_length to be more + self-explanatory. In this case, it must be length-based. Due to history + reasons. when LoD is called lod in public API, it might be offset-based. + Users should be careful about it. + + )DOC") .def_buffer( [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); }) .def("__init__", @@ -337,7 +384,8 @@ All parameter, weight, gradient are variables in Paddle. return make_ddim(shape); }); auto *holder = var.GetMutable(); - holder->InitOnce(capacity, dims); + holder->InitOnce(capacity, dims, + FLAGS_reader_queue_speed_test_mode); return holder->GetQueue(); }, py::return_value_policy::copy); @@ -597,9 +645,13 @@ All parameter, weight, gradient are variables in Paddle. py::class_> pass(m, "Pass"); pass.def(py::init()) - .def("set_str", [](ir::Pass &self, const std::string &name, - const std::string &attr) { - self.Set(name, new std::string(attr)); + .def( + "set_str", + [](ir::Pass &self, const std::string &name, const std::string &attr) { + self.Set(name, new std::string(attr)); + }) + .def("set_int", [](ir::Pass &self, const std::string &name, int val) { + self.Set(name, new int(val)); }); py::class_> pb( @@ -624,16 +676,17 @@ All parameter, weight, gradient are variables in Paddle. ExecutionStrategy allows the user to more preciously control how to run the program in ParallelExecutor by setting the property. - The available properties include: - use_cuda (bool): Whether to use CUDA or not. Default True. - num_threads (int): The number of threads that used to run the - operators in ParallelExecutor. If it is not set, it will be - set in ParallelExecutor according to the device count. - Default 0. - allow_op_delay (bool): Whether to delay the communication operators - to run. Default False. - num_iteration_per_drop_scope (int): how many iterations between - the two dropping local scopes. Default 100. + Examples: + .. code-block:: python + + exec_strategy = fluid.ExecutionStrategy() + exec_strategy.num_threads = 4 + + train_exe = fluid.ParallelExecutor(use_cuda=True, + loss_name=loss.name, + exec_strategy=exec_strategy) + + train_loss, = train_exe.run([loss.name], feed=feed_dict) )DOC"); @@ -643,19 +696,34 @@ All parameter, weight, gradient are variables in Paddle. [](const ExecutionStrategy &self) { return self.num_threads_; }, [](ExecutionStrategy &self, size_t num_threads) { self.num_threads_ = num_threads; - }) + }, + R"DOC(The type is INT, num_threads represents the size of thread pool that + used to run the operators of the current program in ParallelExecutor. + If :math:`num\_threads=1`, all the operators will execute one by one, + but the order maybe difference between iterations. + If it is not set, it will be set in ParallelExecutor according to the + device type and device count, for GPU, :math:`num\_threads=device\_count*4`, for CPU, + :math:`num\_threads=CPU\_NUM*4`, the explanation of:math:`CPU\_NUM` is in ParallelExecutor. + if it is not set, ParallelExecutor will get the cpu count by calling + `multiprocessing.cpu_count()`. Default 0.)DOC") .def_property( "use_cuda", [](const ExecutionStrategy &self) { return self.use_cuda_; }, [](ExecutionStrategy &self, bool use_cuda) { self.use_cuda_ = use_cuda; - }) + }) // FIXME(chengduo): Doesn't add doc for 'use_cuda', use_cuda may + // make user confuse, because ParallelExecutor has a parameter named + // 'use_cuda' too, in current implementation, ParallelExecutor's + // 'use_cuda' will rewrite ExecutionStrategy's 'use_cuda'. .def_property( "allow_op_delay", [](const ExecutionStrategy &self) { return self.allow_op_delay_; }, [](ExecutionStrategy &self, bool allow_op_delay) { self.allow_op_delay_ = allow_op_delay; - }) + }, + R"DOC(The type is BOOL, allow_op_delay represents whether to delay the + communication operators to run, it may make the execution faster. + Note that in some models, allow_op_delay may cause program hang. Default False.)DOC") .def_property( "num_iteration_per_drop_scope", [](const ExecutionStrategy &self) { @@ -663,7 +731,24 @@ All parameter, weight, gradient are variables in Paddle. }, [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) { self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope; - }); + }, + R"DOC(The type is INT, num_iteration_per_drop_scope indicates how + many iterations to clean up the temp variables which + is generated during execution. It may make the execution faster, + because the temp variable's shape maybe the same between two iterations. Default 100. + + NOTES: + 1. If you fetch data when calling the 'run', the ParallelExecutor + will clean up the temp variables at the end of the current iteration. + 2. In some NLP model, it may cause the GPU memory is insufficient, + in this case, you should reduce `num_iteration_per_drop_scope`. + )DOC") + .def_property("_dry_run", + [](const ExecutionStrategy &self) { return self.dry_run_; }, + [](ExecutionStrategy &self, bool dry_run) { + self.dry_run_ = dry_run; + }); + exec_strategy.def_property( "use_experimental_executor", [](const ExecutionStrategy &self) { @@ -678,20 +763,17 @@ All parameter, weight, gradient are variables in Paddle. BuildStrategy allows the user to more preciously control how to build the SSA Graph in ParallelExecutor by setting the property. - The available properties include: - reduce_strategy (str): There are two reduce strategies, 'AllReduce' - and 'Reduce'. If you want that all parameters will be optimized - on all devices, you can choose 'AllReduce'; if you choose - 'Reduce', all parameters will be evenly allocated to different - devices for optimization, and then broadcast the optimized - parameter to other devices. Default 'AllReduce'. - gradient_scale_strategy (str): There are two ways of defining loss@grad, - 'CoeffNumDevice' and 'Customized'. By default, ParallelExecutor - sets the loss@grad according to the number of devices. If you want - to customize loss@grad, you can choose 'Customized'. - Default 'CoeffNumDevice'. - debug_graphviz_path (str): Whether to write the SSA Graph to file in the - form of graphviz. It is useful for debugging. Default "". + Examples: + .. code-block:: python + + build_strategy = fluid.BuildStrategy() + build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + + train_exe = fluid.ParallelExecutor(use_cuda=True, + loss_name=loss.name, + build_strategy=build_strategy) + + train_loss, = train_exe.run([loss.name], feed=feed_dict) )DOC"); py::enum_(build_strategy, "ReduceStrategy") @@ -710,31 +792,69 @@ All parameter, weight, gradient are variables in Paddle. [](const BuildStrategy &self) { return self.reduce_; }, [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) { self.reduce_ = strategy; - }) + }, + R"DOC(The type is STR, there are two reduce strategies in ParallelExecutor, + 'AllReduce' and 'Reduce'. If you want that all the parameters' + optimization are done on all devices independently, you should choose 'AllReduce'; + if you choose 'Reduce', all the parameters' optimization will be evenly distributed + to different devices, and then broadcast the optimized parameter to other devices. + In some models, `Reduce` is faster. Default 'AllReduce'. )DOC") .def_property( "gradient_scale_strategy", [](const BuildStrategy &self) { return self.gradient_scale_; }, [](BuildStrategy &self, BuildStrategy::GradientScaleStrategy strategy) { self.gradient_scale_ = strategy; - }) + }, + R"DOC(The type is STR, there are three ways of defining :math:`loss@grad` in + ParallelExecutor, 'CoeffNumDevice', 'One' and 'Customized'. By default, + ParallelExecutor sets the :math:`loss@grad` according to the number of devices. + If you want to customize :math:`loss@grad`, you can choose 'Customized'. + Default 'CoeffNumDevice'.)DOC") .def_property( "debug_graphviz_path", [](const BuildStrategy &self) { return self.debug_graphviz_path_; }, [](BuildStrategy &self, const std::string &path) { self.debug_graphviz_path_ = path; - }) + }, + R"DOC(The type is STR, debug_graphviz_path indicate the path that + writing the SSA Graph to file in the form of graphviz, you. + It is useful for debugging. Default "")DOC") .def_property( "enable_data_balance", [](const BuildStrategy &self) { return self.enable_data_balance_; }, - [](BuildStrategy &self, bool b) { self.enable_data_balance_ = b; }) - .def_property("fuse_elewise_add_act_ops", - [](const BuildStrategy &self) { - return self.fuse_elewise_add_act_ops_; - }, - [](BuildStrategy &self, bool b) { - self.fuse_elewise_add_act_ops_ = b; - }) + [](BuildStrategy &self, bool b) { + self.enable_data_balance_ = b; + }) // FIXME(chengudo): enable_data_balance seems not important + .def_property( + "enable_sequential_execution", + [](const BuildStrategy &self) { + return self.enable_sequential_execution_; + }, + [](BuildStrategy &self, bool b) { + self.enable_sequential_execution_ = b; + }, + R"DOC(The type is BOOL. If set True, the execution order of ops would be the same as what is in the program. Default False.)DOC") + .def_property( + "remove_unnecessary_lock", + [](const BuildStrategy &self) { + return self.remove_unnecessary_lock_; + }, + [](BuildStrategy &self, bool b) { + self.remove_unnecessary_lock_ = b; + }, + R"DOC(The type is BOOL. If set True, some locks in GPU ops would be released and ParallelExecutor would run faster. Default False.)DOC") + .def_property( + "fuse_elewise_add_act_ops", + [](const BuildStrategy &self) { + return self.fuse_elewise_add_act_ops_; + }, + [](BuildStrategy &self, bool b) { + self.fuse_elewise_add_act_ops_ = b; + }, + R"DOC(The type is BOOL, fuse_elewise_add_act_ops indicate whether + to fuse elementwise_add_op and activation_op, + it may make the execution faster. Default False)DOC") .def("_create_passes_from_strategy", [](BuildStrategy &self) -> std::shared_ptr { return self.CreatePassesFromStrategy(); diff --git a/paddle/fluid/train/demo/CMakeLists.txt b/paddle/fluid/train/demo/CMakeLists.txt index 78d6e5ff554b9cd9facae85be166a697e0b75337..eabb51d370aff709e289e1fc727aa2dbb551d82e 100644 --- a/paddle/fluid/train/demo/CMakeLists.txt +++ b/paddle/fluid/train/demo/CMakeLists.txt @@ -15,6 +15,7 @@ include_directories("${PADDLE_LIB}") include_directories("${PADDLE_LIB}/third_party/install/protobuf/include") include_directories("${PADDLE_LIB}/third_party/install/glog/include") include_directories("${PADDLE_LIB}/third_party/install/gflags/include") +include_directories("${PADDLE_LIB}/third_party/install/xxhash/include") include_directories("${PADDLE_LIB}/third_party/install/snappy/include") include_directories("${PADDLE_LIB}/third_party/install/snappystream/include") include_directories("${PADDLE_LIB}/third_party/install/zlib/include") @@ -27,6 +28,7 @@ link_directories("${PADDLE_LIB}/third_party/install/snappystream/lib") link_directories("${PADDLE_LIB}/third_party/install/protobuf/lib") link_directories("${PADDLE_LIB}/third_party/install/glog/lib") link_directories("${PADDLE_LIB}/third_party/install/gflags/lib") +link_directories("${PADDLE_LIB}/third_party/install/xxhash/lib") link_directories("${PADDLE_LIB}/third_party/install/zlib/lib") add_executable(demo_trainer demo_trainer.cc) @@ -62,5 +64,5 @@ target_link_libraries(demo_trainer ${ARCHIVE_END} ${MATH_LIB} ${MKLDNN_LIB} - glog gflags protobuf snappystream snappy z + glog gflags protobuf snappystream snappy z xxhash ${EXTERNAL_LIB}) diff --git a/paddle/fluid/train/demo/README.md b/paddle/fluid/train/demo/README.md index 41b01d33828f750f67bba5f82cb7ed6fe4d4ea0a..191da20669e185d819ec5eed55427461cc0b10e4 100644 --- a/paddle/fluid/train/demo/README.md +++ b/paddle/fluid/train/demo/README.md @@ -15,7 +15,7 @@ cmake .. -DFLUID_INSTALL_DIR=$PADDLE_LIB \ -DWITH_MKL=OFF \ -DWITH_MKLDNN=OFF make -j8 -make -j8 inference_lib_dist +make -j8 fluid_lib_dist ``` ### step 2. generate program desc diff --git a/paddle/fluid/train/demo/demo_trainer.cc b/paddle/fluid/train/demo/demo_trainer.cc index a0757b53f37b29de0b3802c345b1ad9db69f16e9..ac1ac8e7c2348289516240b6eddf454d02828e2f 100644 --- a/paddle/fluid/train/demo/demo_trainer.cc +++ b/paddle/fluid/train/demo/demo_trainer.cc @@ -40,7 +40,7 @@ void ReadBinaryFile(const std::string& filename, std::string* contents) { std::unique_ptr Load( paddle::framework::Executor* executor, const std::string& model_filename) { - VLOG(3) << "loading model from " << model_filename; + VLOG(30) << "loading model from " << model_filename; std::string program_desc_str; ReadBinaryFile(model_filename, &program_desc_str); diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh index e133323ae420ba68d90215767ab940aed744acd6..a51c9becd416af243cb473c8856141db8d9f3bf0 100755 --- a/paddle/scripts/paddle_build.sh +++ b/paddle/scripts/paddle_build.sh @@ -95,9 +95,9 @@ function cmake_gen() { exit 1 fi fi - else + else if [ "$1" != "" ]; then - echo "using python abi: $1" + echo "using python abi: $1" if [ "$1" == "cp27-cp27m" ]; then export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:} export PATH=/opt/python/cp27-cp27m/bin/:${PATH} @@ -119,7 +119,7 @@ function cmake_gen() { fi fi fi - + if [ "$SYSTEM" == "Darwin" ]; then WITH_DISTRIBUTE=${WITH_DISTRIBUTE:-ON} WITH_AVX=${WITH_AVX:-ON} @@ -127,7 +127,7 @@ function cmake_gen() { else INFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR:-/root/.cache/inference_demo} fi - + cat <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <> ${PADDLE_ROOT}/build/Dockerfile <>> with program._optimized_guard([p,g]): >>> p = p - 0.001 * g """ + tmp_role = self._current_role + tmp_var = self._op_role_var + OpRole = core.op_proto_and_checker_maker.OpRole self._current_role = OpRole.Optimize self._op_role_var = [ @@ -1503,11 +1506,11 @@ class Program(object): for var in param_and_grads ] yield - self._op_role_var = [] - self._current_role = OpRole.Forward + self._op_role_var = tmp_var + self._current_role = tmp_role @contextlib.contextmanager - def _lr_schedule_guard(self): + def _lr_schedule_guard(self, is_with_opt=False): """ A with guard to set :code:`LRSched` :code:`OpRole` and :code:`OpRoleVar` automatically. The :code:`OpRoleVar` is @@ -1515,6 +1518,10 @@ class Program(object): Notes: This is a very low level API. Users should not use it directly. + Args: + is_with_opt: Only set to true if these ops a in the middle + of a bunch of optimize ops so that it can be treated + correctly. For example, sgd->lr_op->sgd->lr_op->sgd. Examples: @@ -1522,13 +1529,19 @@ class Program(object): >>> with program.lr_schedule_guard(): >>> lr = lr * decay """ + + tmp_role = self._current_role + tmp_var = self._op_role_var + OpRole = core.op_proto_and_checker_maker.OpRole self._current_role = OpRole.LRSched + if is_with_opt: + self._current_role = int(OpRole.LRSched) | int(OpRole.Optimize) # TODO(typhoonzero): how to set target learning rate var self._op_role_var = [] yield - self._op_role_var = [] - self._current_role = OpRole.Forward + self._op_role_var = tmp_var + self._current_role = tmp_role def __str__(self): """ diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index 604f3eacd75beff306915b224b30c369dd3a486f..8936d884dd9e1ebbe5f688c11430b64e51ad8bd5 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -65,7 +65,7 @@ def is_persistable(var): Examples: .. code-block:: python - param = fluid.default_main_program().global_block().var('fc.w') + param = fluid.default_main_program().global_block().var('fc.b') res = fluid.io.is_persistable(param) """ if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ @@ -625,8 +625,13 @@ def save_inference_model(dirname, main_program._distributed_lookup_table, main_program._endpoints) - if not os.path.isdir(dirname): + # when a pserver and a trainer running on the same machine, mkdir may conflict + try: os.makedirs(dirname) + except OSError as e: + if e.errno != errno.EEXIST: + raise + if model_filename is not None: model_basename = os.path.basename(model_filename) else: @@ -884,12 +889,13 @@ def _load_slice_up_vars(executor, dirname, slice_vars_and_attrs): load_prog = Program() load_block = load_prog.global_block() + need_delete_vars = [] for var_tuple in slice_vars_and_attrs: orig_var = var_tuple[0] start = var_tuple[1] slice_var = var_tuple[2] - end = start + reduce(lambda x, y: x * y, slice_var.shape) + end = start + slice_var.shape[0] clone_orig_var = load_block.create_var( name=orig_var.name, @@ -917,5 +923,8 @@ def _load_slice_up_vars(executor, dirname, slice_vars_and_attrs): attrs={'axes': [0], 'starts': [start], 'ends': [end]}) - + need_delete_vars.append(clone_orig_var) + load_block.append_op( + type='delete_var', + inputs={'X': need_delete_vars}, ) executor.run(load_prog) diff --git a/python/paddle/fluid/layer_helper.py b/python/paddle/fluid/layer_helper.py index bd9727b6ac0208b199091db00bd0fd5fae74d53b..dc317de9abbd06f4021e64b87ea88ba6af8809c9 100644 --- a/python/paddle/fluid/layer_helper.py +++ b/python/paddle/fluid/layer_helper.py @@ -324,10 +324,19 @@ class LayerHelper(object): raise ValueError("no Parameter name %s found" % name) return param - def create_tmp_variable(self, dtype, stop_gradient=False): + def create_variable_for_type_inference(self, dtype, stop_gradient=False): + """Create a temporary variable that should be type inferred layer. + + Note: + The default type will be set to LOD_TENSOR. However, when + the var is used as operator output, its type will be updated + based on operator's `VarTypeInference` implementation in + infer_var_type. + """ return self.main_program.current_block().create_var( name=unique_name.generate(".".join([self.name, 'tmp'])), dtype=dtype, + type=core.VarDesc.VarType.LOD_TENSOR, persistable=False, stop_gradient=stop_gradient) @@ -388,7 +397,7 @@ class LayerHelper(object): b = self.create_parameter( attr=bias_attr, shape=size, dtype=input_var.dtype, is_bias=True) - tmp = self.create_tmp_variable(dtype=input_var.dtype) + tmp = self.create_variable_for_type_inference(dtype=input_var.dtype) self.append_op( type='elementwise_add', inputs={'X': [input_var], @@ -414,7 +423,7 @@ class LayerHelper(object): tmp = input_var # NOTE(dzhwinter): some activation support inplace compution. if not core.IsInplace(act_type): - tmp = self.create_tmp_variable(dtype=input_var.dtype) + tmp = self.create_variable_for_type_inference(dtype=input_var.dtype) self.append_op( type=act_type, inputs={"X": [input_var]}, diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 4af97e8632a47fbd981362dc8249a3f6b7269ecd..9730fbf510cbe8c323b761b29821710f2c14a81d 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -80,8 +80,8 @@ def split_lod_tensor(input, mask, level=0): """ helper = LayerHelper('split_lod_tensor', **locals()) - out_true = helper.create_tmp_variable(dtype=input.dtype) - out_false = helper.create_tmp_variable(dtype=input.dtype) + out_true = helper.create_variable_for_type_inference(dtype=input.dtype) + out_false = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type='split_lod_tensor', inputs={ @@ -131,7 +131,7 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0): in_true=out_true, in_false=out_false, mask=y, x=x, level=level) """ helper = LayerHelper('merge_lod_tensor', **locals()) - out = helper.create_tmp_variable(dtype=in_true.dtype) + out = helper.create_variable_for_type_inference(dtype=in_true.dtype) helper.append_op( type='merge_lod_tensor', inputs={'X': x, @@ -524,7 +524,7 @@ class StaticRNN(object): if not isinstance(o, Variable): raise TypeError("step output takes a Variable") - tmp_o = self.helper.create_tmp_variable(dtype=o.dtype) + tmp_o = self.helper.create_variable_for_type_inference(dtype=o.dtype) self.helper.append_op( type='rnn_memory_helper', inputs={'X': [o]}, @@ -606,7 +606,8 @@ class StaticRNN(object): pre_memories.append(mem.pre_mem.name) mem_var = rnn_block.var(mem.mem.name) assert isinstance(mem_var, Variable) - new_mem = self.helper.create_tmp_variable(dtype=mem_var.dtype) + new_mem = self.helper.create_variable_for_type_inference( + dtype=mem_var.dtype) rnn_block.append_op( type='rnn_memory_helper', @@ -813,7 +814,7 @@ def max_sequence_len(rank_table): ${out_comment}. """ helper = LayerHelper("max_seqence_len", **locals()) - res = helper.create_tmp_variable(dtype="int64") + res = helper.create_variable_for_type_inference(dtype="int64") helper.append_op( type="max_sequence_len", inputs={"RankTable": rank_table}, @@ -884,7 +885,7 @@ def array_to_lod_tensor(x, table): lod_tensor = fluid.layers.array_to_lod_tensor(array, table) """ helper = LayerHelper("array_to_lod_tensor", **locals()) - tmp = helper.create_tmp_variable(dtype=x.dtype) + tmp = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type="array_to_lod_tensor", inputs={'X': x, @@ -915,7 +916,7 @@ def increment(x, value=1.0, in_place=True): """ helper = LayerHelper("increment", **locals()) if not in_place: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = x helper.append_op( @@ -1012,7 +1013,7 @@ def less_than(x, y, force_cpu=None, cond=None, **ignored): """ helper = LayerHelper("less_than", **locals()) if cond is None: - cond = helper.create_tmp_variable(dtype='bool') + cond = helper.create_variable_for_type_inference(dtype='bool') cond.stop_gradient = True attrs = dict() @@ -1051,7 +1052,7 @@ def equal(x, y, cond=None, **ignored): """ helper = LayerHelper("equal", **locals()) if cond is None: - cond = helper.create_tmp_variable(dtype='bool') + cond = helper.create_variable_for_type_inference(dtype='bool') cond.stop_gradient = True helper.append_op( @@ -1098,7 +1099,7 @@ def array_read(array, i): array, Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY: raise TypeError("array should be tensor array vairable") - out = helper.create_tmp_variable(dtype=array.dtype) + out = helper.create_variable_for_type_inference(dtype=array.dtype) helper.append_op( type='read_from_array', inputs={'X': [array], @@ -1133,7 +1134,7 @@ def shrink_memory(x, i, table): usage. """ helper = LayerHelper('shrink_memory', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='shrink_rnn_memory', inputs={'X': [x], @@ -1170,7 +1171,7 @@ def array_length(array): """ helper = LayerHelper('array_length', **locals()) - tmp = helper.create_tmp_variable(dtype='int64') + tmp = helper.create_variable_for_type_inference(dtype='int64') tmp.stop_gradient = True helper.append_op( type='lod_array_length', inputs={'X': [array]}, outputs={'Out': [tmp]}) @@ -1585,12 +1586,11 @@ class DynamicRNN(object): self.lod_rank_table = None self.max_seq_len = None self.step_idx = None - self.zero_idx = fill_constant( - shape=[1], value=0, dtype='int64', force_cpu=True) + self.zero_idx = None self.mem_dict = dict() self.output_array = [] self.outputs = [] - self.cond = self.helper.create_tmp_variable(dtype='bool') + self.cond = self.helper.create_variable_for_type_inference(dtype='bool') self.cond.stop_gradient = False self.while_op = While(self.cond) self.input_array = [] @@ -1791,6 +1791,7 @@ class DynamicRNN(object): """ self._assert_in_rnn_block_('memory') + self._init_zero_idx_() if init is not None: if not isinstance(init, Variable): raise TypeError( @@ -1904,6 +1905,22 @@ class DynamicRNN(object): array_write(x=each, i=self.step_idx, array=outside_array) self.output_array.append(outside_array) + def _init_zero_idx_(self): + if self.zero_idx is None: + parent_block = self._parent_block_() + self.zero_idx = parent_block.create_var( + name=unique_name.generate('zero_idx'), dtype='int64') + parent_block.append_op( + type='fill_constant', + inputs={}, + outputs={'Out': [self.zero_idx]}, + attrs={ + 'shape': [1], + 'dtype': self.zero_idx.dtype, + 'value': float(0), + 'force_cpu': True + }) + def _parent_block_(self): prog = self.helper.main_program parent_idx = prog.current_block().parent_idx @@ -1924,7 +1941,7 @@ def reorder_lod_tensor_by_rank(x, rank_table): helper.is_instance('x', Variable) helper.is_instance('rank_table', Variable) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='reorder_lod_tensor_by_rank', inputs={'X': [x], @@ -1958,7 +1975,7 @@ def is_empty(x, cond=None, **ignored): """ helper = LayerHelper("is_empty", **locals()) if cond is None: - cond = helper.create_tmp_variable(dtype='bool') + cond = helper.create_variable_for_type_inference(dtype='bool') cond.stop_gradient = True elif not isinstance(cond, Variable): raise TypeError("cond takes a variable") diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index 1cfcbbb9c1614f21848e62cce79befc673e1739c..96b6705e26c0f8d8d223e9020192a8f330c2c727 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -31,6 +31,7 @@ from functools import reduce __all__ = [ 'prior_box', + 'density_prior_box', 'multi_box_head', 'bipartite_match', 'target_assign', @@ -116,8 +117,8 @@ def rpn_target_assign(bbox_pred, Returns: tuple: A tuple(predicted_scores, predicted_location, target_label, - target_bbox) is returned. The predicted_scores and - predicted_location is the predicted result of the RPN. + target_bbox, bbox_inside_weight) is returned. The predicted_scores + and predicted_location is the predicted result of the RPN. The target_label and target_bbox is the ground truth, respectively. The predicted_location is a 2D Tensor with shape [F, 4], and the shape of target_bbox is same as the shape of @@ -126,6 +127,8 @@ def rpn_target_assign(bbox_pred, [F + B, 1], and the shape of target_label is same as the shape of the predicted_scores, B is the number of the background anchors, the F and B is depends on the input of this operator. + Bbox_inside_weight represents whether the predicted loc is fake_fg + or not and the shape is [F, 4]. Examples: .. code-block:: python @@ -138,7 +141,7 @@ def rpn_target_assign(bbox_pred, append_batch_size=False, dtype='float32') gt_boxes = layers.data(name='gt_boxes', shape=[10, 4], append_batch_size=False, dtype='float32') - loc_pred, score_pred, loc_target, score_target = + loc_pred, score_pred, loc_target, score_target, bbox_inside_weight = fluid.layers.rpn_target_assign(bbox_pred=bbox_pred, cls_logits=cls_logits, anchor_box=anchor_box, @@ -147,10 +150,13 @@ def rpn_target_assign(bbox_pred, helper = LayerHelper('rpn_target_assign', **locals()) # Assign target label to anchors - loc_index = helper.create_tmp_variable(dtype='int32') - score_index = helper.create_tmp_variable(dtype='int32') - target_label = helper.create_tmp_variable(dtype='int32') - target_bbox = helper.create_tmp_variable(dtype=anchor_box.dtype) + loc_index = helper.create_variable_for_type_inference(dtype='int32') + score_index = helper.create_variable_for_type_inference(dtype='int32') + target_label = helper.create_variable_for_type_inference(dtype='int32') + target_bbox = helper.create_variable_for_type_inference( + dtype=anchor_box.dtype) + bbox_inside_weight = helper.create_variable_for_type_inference( + dtype=anchor_box.dtype) helper.append_op( type="rpn_target_assign", inputs={ @@ -163,7 +169,8 @@ def rpn_target_assign(bbox_pred, 'LocationIndex': loc_index, 'ScoreIndex': score_index, 'TargetLabel': target_label, - 'TargetBBox': target_bbox + 'TargetBBox': target_bbox, + 'BBoxInsideWeight': bbox_inside_weight }, attrs={ 'rpn_batch_size_per_im': rpn_batch_size_per_im, @@ -178,13 +185,14 @@ def rpn_target_assign(bbox_pred, score_index.stop_gradient = True target_label.stop_gradient = True target_bbox.stop_gradient = True + bbox_inside_weight.stop_gradient = True cls_logits = nn.reshape(x=cls_logits, shape=(-1, 1)) bbox_pred = nn.reshape(x=bbox_pred, shape=(-1, 4)) predicted_cls_logits = nn.gather(cls_logits, score_index) predicted_bbox_pred = nn.gather(bbox_pred, loc_index) - return predicted_cls_logits, predicted_bbox_pred, target_label, target_bbox + return predicted_cls_logits, predicted_bbox_pred, target_label, target_bbox, bbox_inside_weight def detection_output(loc, @@ -282,7 +290,8 @@ def detection_output(loc, scores = nn.reshape(x=scores, shape=compile_shape, actual_shape=run_shape) scores = nn.transpose(scores, perm=[0, 2, 1]) scores.stop_gradient = True - nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype) + nmsed_outs = helper.create_variable_for_type_inference( + dtype=decoded_box.dtype) helper.append_op( type="multiclass_nms", inputs={'Scores': scores, @@ -314,7 +323,7 @@ def iou_similarity(x, y, name=None): """ helper = LayerHelper("iou_similarity", **locals()) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = helper.create_variable( name=name, dtype=x.dtype, persistable=False) @@ -351,7 +360,8 @@ def box_coder(prior_box, helper = LayerHelper("box_coder", **locals()) if name is None: - output_box = helper.create_tmp_variable(dtype=prior_box.dtype) + output_box = helper.create_variable_for_type_inference( + dtype=prior_box.dtype) else: output_box = helper.create_variable( name=name, dtype=prior_box.dtype, persistable=False) @@ -382,7 +392,7 @@ def polygon_box_transform(input, name=None): """ helper = LayerHelper("polygon_box_transform", **locals()) if name is None: - output = helper.create_tmp_variable(dtype=input.dtype) + output = helper.create_variable_for_type_inference(dtype=input.dtype) else: output = helper.create_variable( name=name, dtype=prior_box.input, persistable=False) @@ -450,7 +460,7 @@ def detection_map(detect_res, helper = LayerHelper("detection_map", **locals()) def __create_var(type): - return helper.create_tmp_variable(dtype=type) + return helper.create_variable_for_type_inference(dtype=type) map_out = __create_var('float32') accum_pos_count_out = out_states[0] if out_states else __create_var('int32') @@ -557,8 +567,9 @@ def bipartite_match(dist_matrix, >>> matched_indices, matched_dist = fluid.layers.bipartite_match(iou) """ helper = LayerHelper('bipartite_match', **locals()) - match_indices = helper.create_tmp_variable(dtype='int32') - match_distance = helper.create_tmp_variable(dtype=dist_matrix.dtype) + match_indices = helper.create_variable_for_type_inference(dtype='int32') + match_distance = helper.create_variable_for_type_inference( + dtype=dist_matrix.dtype) helper.append_op( type='bipartite_match', inputs={'DistMat': dist_matrix}, @@ -644,8 +655,8 @@ def target_assign(input, gt, matched_indices, mismatch_value=0) """ helper = LayerHelper('target_assign', **locals()) - out = helper.create_tmp_variable(dtype=input.dtype) - out_weight = helper.create_tmp_variable(dtype='float32') + out = helper.create_variable_for_type_inference(dtype=input.dtype) + out_weight = helper.create_variable_for_type_inference(dtype='float32') helper.append_op( type='target_assign', inputs={ @@ -816,9 +827,10 @@ def ssd_loss(location, conf_loss = nn.reshape( x=conf_loss, shape=(num, num_prior), actual_shape=actual_shape) conf_loss.stop_gradient = True - neg_indices = helper.create_tmp_variable(dtype='int32') + neg_indices = helper.create_variable_for_type_inference(dtype='int32') dtype = matched_indices.dtype - updated_matched_indices = helper.create_tmp_variable(dtype=dtype) + updated_matched_indices = helper.create_variable_for_type_inference( + dtype=dtype) helper.append_op( type='mine_hard_examples', inputs={ @@ -998,8 +1010,8 @@ def prior_box(input, max_sizes = [max_sizes] attrs['max_sizes'] = max_sizes - box = helper.create_tmp_variable(dtype) - var = helper.create_tmp_variable(dtype) + box = helper.create_variable_for_type_inference(dtype) + var = helper.create_variable_for_type_inference(dtype) helper.append_op( type="prior_box", inputs={"Input": input, @@ -1012,6 +1024,135 @@ def prior_box(input, return box, var +def density_prior_box(input, + image, + densities=None, + fixed_sizes=None, + fixed_ratios=None, + variance=[0.1, 0.1, 0.2, 0.2], + clip=False, + steps=[0.0, 0.0], + offset=0.5, + name=None): + """ + **Density Prior Box Operator** + + Generate density prior boxes for SSD(Single Shot MultiBox Detector) + algorithm. Each position of the input produce N prior boxes, N is + determined by the count of densities, fixed_sizes and fixed_ratios. + Boxes center at grid points around each input position is generated by + this operator, and the grid points is determined by densities and + the count of density prior box is determined by fixed_sizes and fixed_ratios. + Obviously, the number of fixed_sizes is equal to the number of densities. + For densities_i in densities: + N_density_prior_box =sum(N_fixed_ratios * densities_i^2), + + Args: + input(Variable): The Input Variables, the format is NCHW. + image(Variable): The input image data of PriorBoxOp, + the layout is NCHW. + densities(list|tuple|None): the densities of generated density prior + boxes, this attribute should be a list or tuple of integers. + Default: None. + fixed_sizes(list|tuple|None): the fixed sizes of generated density + prior boxes, this attribute should a list or tuple of same + length with :attr:`densities`. Default: None. + fixed_ratios(list|tuple|None): the fixed ratios of generated density + prior boxes, if this attribute is not set and :attr:`densities` + and :attr:`fix_sizes` is set, :attr:`aspect_ratios` will be used + to generate density prior boxes. + variance(list|tuple): the variances to be encoded in density prior boxes. + Default:[0.1, 0.1, 0.2, 0.2]. + clip(bool): Whether to clip out-of-boundary boxes. Default: False. + step(list|turple): Prior boxes step across width and height, If + step[0] == 0.0/step[1] == 0.0, the density prior boxes step across + height/weight of the input will be automatically calculated. + Default: [0., 0.] + offset(float): Prior boxes center offset. Default: 0.5 + name(str): Name of the density prior box op. Default: None. + + Returns: + tuple: A tuple with two Variable (boxes, variances) + + boxes: the output density prior boxes of PriorBox. + The layout is [H, W, num_priors, 4]. + H is the height of input, W is the width of input, + num_priors is the total + box count of each position of input. + + variances: the expanded variances of PriorBox. + The layout is [H, W, num_priors, 4]. + H is the height of input, W is the width of input + num_priors is the total + box count of each position of input + + + Examples: + .. code-block:: python + + box, var = fluid.layers.density_prior_box( + input=conv1, + image=images, + min_sizes=[100.], + max_sizes=[200.], + aspect_ratios=[1.0, 1.0 / 2.0, 2.0], + densities=[3, 4], + fixed_sizes=[50., 60.], + fixed_ratios=[1.0, 3.0, 1.0 / 3.0], + flip=True, + clip=True) + """ + helper = LayerHelper("density_prior_box", **locals()) + dtype = helper.input_dtype() + + def _is_list_or_tuple_(data): + return (isinstance(data, list) or isinstance(data, tuple)) + + if not _is_list_or_tuple_(densities): + raise TypeError('densities should be a list or a tuple or None.') + if not _is_list_or_tuple_(fixed_sizes): + raise TypeError('fixed_sizes should be a list or a tuple or None.') + if not _is_list_or_tuple_(fixed_ratios): + raise TypeError('fixed_ratios should be a list or a tuple or None.') + if len(densities) != len(fixed_sizes): + raise ValueError('densities and fixed_sizes length should be euqal.') + if not (_is_list_or_tuple_(steps) and len(steps) == 2): + raise ValueError('steps should be a list or tuple ', + 'with length 2, (step_width, step_height).') + + densities = list(map(int, densities)) + fixed_sizes = list(map(float, fixed_sizes)) + fixed_ratios = list(map(float, fixed_ratios)) + steps = list(map(float, steps)) + + attrs = { + 'variances': variance, + 'clip': clip, + 'step_w': steps[0], + 'step_h': steps[1], + 'offset': offset, + } + if densities is not None and len(densities) > 0: + attrs['densities'] = densities + if fixed_sizes is not None and len(fixed_sizes) > 0: + attrs['fixed_sizes'] = fixed_sizes + if fixed_ratios is not None and len(fixed_ratios) > 0: + attrs['fixed_ratios'] = fixed_ratios + + box = helper.create_variable_for_type_inference(dtype) + var = helper.create_variable_for_type_inference(dtype) + helper.append_op( + type="density_prior_box", + inputs={"Input": input, + "Image": image}, + outputs={"Boxes": box, + "Variances": var}, + attrs=attrs, ) + box.stop_gradient = True + var.stop_gradient = True + return box, var + + def multi_box_head(inputs, image, base_size, @@ -1337,8 +1478,8 @@ def anchor_generator(input, 'offset': offset } - anchor = helper.create_tmp_variable(dtype) - var = helper.create_tmp_variable(dtype) + anchor = helper.create_variable_for_type_inference(dtype) + var = helper.create_variable_for_type_inference(dtype) helper.append_op( type="anchor_generator", inputs={"Input": input}, @@ -1384,7 +1525,7 @@ def roi_perspective_transform(input, """ helper = LayerHelper('roi_perspective_transform', **locals()) dtype = helper.input_dtype() - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="roi_perspective_transform", inputs={"X": input, @@ -1413,16 +1554,49 @@ def generate_proposal_labels(rpn_rois, use_random=True): """ ** Generate proposal labels Faster-RCNN ** - TODO(buxingyuan): Add Document + This operator can be, for given the GenerateProposalOp output bounding boxes and groundtruth, + to sample foreground boxes and background boxes, and compute loss target. + + RpnRois is the output boxes of RPN and was processed by generate_proposal_op, these boxes + were combined with groundtruth boxes and sampled according to batch_size_per_im and fg_fraction, + If an instance with a groundtruth overlap greater than fg_thresh, then it was considered as a foreground sample. + If an instance with a groundtruth overlap greater than bg_thresh_lo and lower than bg_thresh_hi, + then it was considered as a background sample. + After all foreground and background boxes are chosen (so called Rois), + then we apply random sampling to make sure + the number of foreground boxes is no more than batch_size_per_im * fg_fraction. + + For each box in Rois, we assign the classification (class label) and regression targets (box label) to it. + Finally BboxInsideWeights and BboxOutsideWeights are used to specify whether it would contribute to training loss. + + Args: + rpn_rois(Variable): A 2-D LoDTensor with shape [N, 4]. N is the number of the GenerateProposalOp's output, each element is a bounding box with [xmin, ymin, xmax, ymax] format. + gt_classes(Variable): A 2-D LoDTensor with shape [M, 1]. M is the number of groundtruth, each element is a class label of groundtruth. + is_crowd(Variable): A 2-D LoDTensor with shape [M, 1]. M is the number of groundtruth, each element is a flag indicates whether a groundtruth is crowd. + gt_boxes(Variable): A 2-D LoDTensor with shape [M, 4]. M is the number of groundtruth, each element is a bounding box with [xmin, ymin, xmax, ymax] format. + im_info(Variable): A 2-D LoDTensor with shape [B, 3]. B is the number of input images, each element consists of im_height, im_width, im_scale. + + batch_size_per_im(int): Batch size of rois per images. + fg_fraction(float): Foreground fraction in total batch_size_per_im. + fg_thresh(float): Overlap threshold which is used to chose foreground sample. + bg_thresh_hi(float): Overlap threshold upper bound which is used to chose background sample. + bg_thresh_lo(float): Overlap threshold lower bound which is used to chose background sample. + bbox_reg_weights(list|tuple): Box regression weights. + class_nums(int): Class number. + use_random(bool): Use random sampling to choose foreground and background boxes. """ helper = LayerHelper('generate_proposal_labels', **locals()) - rois = helper.create_tmp_variable(dtype=rpn_rois.dtype) - labels_int32 = helper.create_tmp_variable(dtype=gt_classes.dtype) - bbox_targets = helper.create_tmp_variable(dtype=rpn_rois.dtype) - bbox_inside_weights = helper.create_tmp_variable(dtype=rpn_rois.dtype) - bbox_outside_weights = helper.create_tmp_variable(dtype=rpn_rois.dtype) + rois = helper.create_variable_for_type_inference(dtype=rpn_rois.dtype) + labels_int32 = helper.create_variable_for_type_inference( + dtype=gt_classes.dtype) + bbox_targets = helper.create_variable_for_type_inference( + dtype=rpn_rois.dtype) + bbox_inside_weights = helper.create_variable_for_type_inference( + dtype=rpn_rois.dtype) + bbox_outside_weights = helper.create_variable_for_type_inference( + dtype=rpn_rois.dtype) helper.append_op( type="generate_proposal_labels", @@ -1472,7 +1646,7 @@ def generate_proposals(scores, eta=1.0, name=None): """ - ** Generate proposal labels Faster-RCNN ** + ** Generate proposal Faster-RCNN ** This operation proposes RoIs according to each box with their probability to be a foreground object and the box can be calculated by anchors. Bbox_deltais and scores to be an object are the output of RPN. Final proposals @@ -1504,8 +1678,10 @@ def generate_proposals(scores, """ helper = LayerHelper('generate_proposals', **locals()) - rpn_rois = helper.create_tmp_variable(dtype=bbox_deltas.dtype) - rpn_roi_probs = helper.create_tmp_variable(dtype=scores.dtype) + rpn_rois = helper.create_variable_for_type_inference( + dtype=bbox_deltas.dtype) + rpn_roi_probs = helper.create_variable_for_type_inference( + dtype=scores.dtype) helper.append_op( type="generate_proposals", inputs={ diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 81c78cba219007a9348af961e4b0dc227edba747..1ab48c00548b58f4b3e411d8e46e8cf496d6b891 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -30,7 +30,8 @@ from ..unique_name import generate as unique_name __all__ = [ 'data', 'open_files', 'read_file', 'shuffle', 'batch', 'double_buffer', - 'random_data_generator', 'py_reader', 'Preprocessor', 'load' + 'random_data_generator', 'py_reader', 'create_py_reader_by_data', + 'Preprocessor', 'load' ] @@ -55,8 +56,12 @@ def data(name, Args: name(str): The name/alias of the function shape(list): Tuple declaring the shape. - append_batch_size(bool): Whether or not to append the data as a batch. - dtype(int|float): The type of data : float32, float_16, int etc + append_batch_size(bool): + 1. If true, it prepends -1 to the shape. + For example if shape=[1], the resulting shape is [-1, 1]. + 2. If shape contains -1, such as shape=[1, -1], + append_batch_size will be enforced to be be False (ineffective). + dtype(basestring): The type of data : float32, float_16, int etc type(VarType): The output type. By default it is LOD_TENSOR. lod_level(int): The LoD Level. 0 means the input data is not a sequence. stop_gradient(bool): A boolean that mentions whether gradient should flow. @@ -311,6 +316,7 @@ def _copy_reader_var_(block, var): new_var = block.create_var(name=var.name, type=core.VarDesc.VarType.READER) new_var.desc.set_shapes(var.desc.shapes()) new_var.desc.set_dtypes(var.desc.dtypes()) + new_var.desc.set_lod_levels(var.desc.lod_levels()) new_var.persistable = True return new_var @@ -470,6 +476,159 @@ def random_data_generator(low, high, shapes, lod_levels, for_parallel=True): return monkey_patch_reader_methods(main_prog_var) +def _py_reader(capacity, + shapes, + dtypes, + lod_levels=None, + name=None, + use_double_buffer=True, + feed_list=None): + + if feed_list is not None: + if not isinstance(feed_list, list): + raise TypeError("feed_list should be a list of Variable" + " instead of " + str(type(feed_list))) + lod_levels = [] + dtypes = [] + shape_concat = [] + ranks = [] + shapes = [] + + for feed_data in feed_list: + dtypes.append(feed_data.dtype) + shape_concat.extend(feed_data.shape) + ranks.append(len(feed_data.shape)) + shapes.append(feed_data.shape) + lod_levels.append(feed_data.lod_level) + else: + dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] + shape_concat = [] + ranks = [] + + for shape in shapes: + shape_concat.extend(shape) + ranks.append(len(shape)) + + if lod_levels is None: + lod_levels = [0] * len(shapes) + + if name is None: + queue_name = unique_name('lod_tensor_blocking_queue') + reader_name = unique_name('create_py_reader') + double_buffer_name = unique_name('double_buffer') + else: + queue_name = "_".join([name, "queue"]) + reader_name = "_".join([name, "reader"]) + double_buffer_name = "_".join([name, "double_buffer"]) + + var = global_scope().var(queue_name) + feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes) + + startup_blk = default_startup_program().current_block() + startup_var = startup_blk.create_var(name=reader_name) + startup_blk.append_op( + type='create_py_reader', + inputs={'blocking_queue': [queue_name]}, + outputs={'Out': [startup_var]}, + attrs={ + 'shape_concat': shape_concat, + 'lod_levels': lod_levels, + 'ranks': ranks + }) + + startup_var.desc.set_dtypes(dtypes) + startup_var.persistable = True + + main_prog_var = _copy_reader_var_(default_main_program().current_block(), + startup_var) + + reader = monkey_patch_reader_methods(main_prog_var) + if use_double_buffer: + double_buffer_reader = double_buffer(reader, name=double_buffer_name) + # we return a double buffer reader. However, the reset method comes from + # py_reader. + double_buffer_reader.reset = reader.reset + reader = double_buffer_reader + + # monkey patch py_reader special methods + reader.queue = feed_queue + current_reset_method = reader.reset + reader.thread = None + reader.tensor_provider = None + reader.exited = False + + def start_provide_thread(func): + def __provider_thread__(): + for tensors in func(): + array = core.LoDTensorArray() + for item in tensors: + if not isinstance(item, core.LoDTensor): + tmp = core.LoDTensor() + tmp.set(item, core.CPUPlace()) + item = tmp + + array.append(item) + + if reader.exited: + break + feed_queue.push(array) + if reader.exited: + break + feed_queue.close() + + reader.thread = threading.Thread(target=__provider_thread__) + reader.thread.daemon = True + reader.thread.start() + + def __set_tensor_provider__(func): + reader.tensor_provider = func + + def __set_paddle_reader__(paddle_reader): + with program_guard(Program(), Program()): + actual_feed_list = feed_list + if actual_feed_list is None: + actual_feed_list = [] + counter = 0 + for dtype, shape, lod_level in zip(dtypes, shapes, lod_levels): + name = str(counter) + actual_feed_list.append( + data( + name=name, + dtype=dtype, + shape=shape, + lod_level=lod_level)) + counter += 1 + + data_names = [feed_data.name for feed_data in actual_feed_list] + feeder = DataFeeder( + feed_list=actual_feed_list, place=core.CPUPlace()) + paddle_reader = feeder.decorate_reader( + paddle_reader, multi_devices=False) + + def __tensor_provider__(): + for slots in paddle_reader(): + yield [slots[data_name] for data_name in data_names] + + __set_tensor_provider__(__tensor_provider__) + + def __reset__(): + current_reset_method() + if reader.thread is not None and reader.tensor_provider is not None: + reader.exited = True + reader.thread.join() + reader.exited = False + + def __start__(): + start_provide_thread(reader.tensor_provider) + + reader.reset = __reset__ + reader.decorate_tensor_provider = __set_tensor_provider__ + reader.decorate_paddle_reader = __set_paddle_reader__ + reader.start = __start__ + + return reader + + def py_reader(capacity, shapes, dtypes, @@ -594,128 +753,72 @@ def py_reader(capacity, >>> except fluid.core.EOFException: >>> test_reader.reset() """ - dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] - shape_concat = [] - ranks = [] - - for shape in shapes: - shape_concat.extend(shape) - ranks.append(len(shape)) - - if lod_levels is None: - lod_levels = [0] * len(shapes) - - if name is None: - queue_name = unique_name('lod_tensor_blocking_queue') - reader_name = unique_name('create_py_reader') - double_buffer_name = unique_name('double_buffer') - else: - queue_name = "_".join([name, "queue"]) - reader_name = "_".join([name, "reader"]) - double_buffer_name = "_".join([name, "double_buffer"]) - - var = global_scope().var(queue_name) - feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes) - - startup_blk = default_startup_program().current_block() - startup_var = startup_blk.create_var(name=reader_name) - startup_blk.append_op( - type='create_py_reader', - inputs={'blocking_queue': [queue_name]}, - outputs={'Out': [startup_var]}, - attrs={ - 'shape_concat': shape_concat, - 'lod_levels': lod_levels, - 'ranks': ranks - }) - - startup_var.desc.set_dtypes(dtypes) - startup_var.persistable = True - - main_prog_var = _copy_reader_var_(default_main_program().current_block(), - startup_var) - - reader = monkey_patch_reader_methods(main_prog_var) - if use_double_buffer: - double_buffer_reader = double_buffer(reader, name=double_buffer_name) - # we return a double buffer reader. However, the reset method comes from - # py_reader. - double_buffer_reader.reset = reader.reset - reader = double_buffer_reader - - # monkey patch py_reader special methods - reader.queue = feed_queue - current_reset_method = reader.reset - reader.thread = None - reader.tensor_provider = None - reader.exited = False - - def start_provide_thread(func): - def __provider_thread__(): - for tensors in func(): - array = core.LoDTensorArray() - for item in tensors: - if not isinstance(item, core.LoDTensor): - tmp = core.LoDTensor() - tmp.set(item, core.CPUPlace()) - item = tmp - - array.append(item) - - if reader.exited: - break - feed_queue.push(array) - if reader.exited: - break - feed_queue.close() + return _py_reader( + capacity=capacity, + shapes=shapes, + dtypes=dtypes, + lod_levels=lod_levels, + name=name, + use_double_buffer=use_double_buffer) - reader.thread = threading.Thread(target=__provider_thread__) - reader.thread.daemon = True - reader.thread.start() - def __set_tensor_provider__(func): - reader.tensor_provider = func +def create_py_reader_by_data(capacity, + feed_list, + name=None, + use_double_buffer=True): + """ + Create a Python reader for data feeding in Python - def __set_paddle_reader__(paddle_reader): - with program_guard(Program(), Program()): - feed_list = [] - counter = 0 - for dtype, shape, lod_level in zip(dtypes, shapes, lod_levels): - name = str(counter) - feed_list.append( - data( - name=name, - dtype=dtype, - shape=shape, - lod_level=lod_level)) - counter += 1 - - feeder = DataFeeder(feed_list=feed_list, place=core.CPUPlace()) - paddle_reader = feeder.decorate_reader( - paddle_reader, multi_devices=False) + This layer returns a Reader Variable. - def __tensor_provider__(): - for slots in paddle_reader(): - yield [slots[str(idx)] for idx in six.moves.xrange(counter)] + Works much like py_reader except that it's input is feed_list + instead of shapes, dtypes and lod_levels - __set_tensor_provider__(__tensor_provider__) + Args: + capacity(int): The buffer capacity maintained by :code:`py_reader`. + feed_list(list(Variable)): The data feed list. + name(basestring): The prefix Python queue name and Reader name. None will + be generated automatically. + use_double_buffer(bool): Whether use double buffer or not. - def __reset__(): - current_reset_method() - if reader.thread is not None and reader.tensor_provider is not None: - reader.exited = True - reader.thread.join() - reader.exited = False + Returns: + Variable: A Reader from which we can get feeding data. - def __start__(): - start_provide_thread(reader.tensor_provider) + Examples: - reader.reset = __reset__ - reader.decorate_tensor_provider = __set_tensor_provider__ - reader.decorate_paddle_reader = __set_paddle_reader__ - reader.start = __start__ + 1. The basic usage of :code:`py_reader` is as follows: - return reader + >>> import paddle.fluid as fluid + >>> import paddle.dataset.mnist as mnist + >>> + >>> image = fluid.layers.data(name='image', shape=[3,224,224], dtypes='float32') + >>> label = fluid.layers.data(name='label', shape=[1], dtypes='int64') + >>> reader = fluid.layers.create_py_reader_by_data(capacity=64, feed_list=[image, label]) + >>> reader.decorate_paddle_reader( + >>> paddle.reader.shuffle(paddle.batch(mnist.train()) + >>> + >>> img, label = fluid.layers.read_file(reader) + >>> loss = network(img, label) # some network definition + >>> + >>> fluid.Executor(fluid.CUDAPlace(0)).run(fluid.default_startup_program()) + >>> + >>> exe = fluid.ParallelExecutor(use_cuda=True, loss_name=loss.name) + >>> for epoch_id in range(10): + >>> reader.start() + >>> try: + >>> while True: + >>> exe.run(fetch_list=[loss.name]) + >>> except fluid.core.EOFException: + >>> reader.reset() + """ + return _py_reader( + capacity=capacity, + shapes=None, + dtypes=None, + lod_levels=None, + name=name, + use_double_buffer=use_double_buffer, + feed_list=feed_list) def open_files(filenames, @@ -950,7 +1053,7 @@ def read_file(reader): """ helper = LayerHelper('read_file') out = [ - helper.create_tmp_variable( + helper.create_variable_for_type_inference( stop_gradient=True, dtype='float32') for _ in range(len(reader.desc.shapes())) ] diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index 8c11921d9bde0920f33368837302d39f36f45556..eea0a362a0c31083f304a2167d0fdadfb30fb640 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -202,10 +202,12 @@ def generate_layer_fn(op_type): out_var = out[0] if (isinstance(out, list) or isinstance(out, tuple)) else out else: - out_var = helper.create_tmp_variable(dtype=dtype) + out_var = helper.create_variable_for_type_inference(dtype=dtype) outputs[o_name] = [out_var] for name in intermediate_output_names: - outputs[name] = [helper.create_tmp_variable(dtype=dtype)] + outputs[name] = [ + helper.create_variable_for_type_inference(dtype=dtype) + ] helper.append_op( type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs) return helper.append_activation(out_var) @@ -229,7 +231,7 @@ def generate_layer_fn_noattr(op_type): def func(x, name=None): helper = LayerHelper(op_type, **locals()) - output = helper.create_tmp_variable(dtype=x.dtype) + output = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op(type=op_type, inputs={"X": x}, outputs={"Out": output}) return output diff --git a/python/paddle/fluid/layers/learning_rate_scheduler.py b/python/paddle/fluid/layers/learning_rate_scheduler.py index dfd801a098d6451dbdb20d9ba44187d1e3f8a91a..149224bb68ac869dec14ac9f953f0072bd24c7e2 100644 --- a/python/paddle/fluid/layers/learning_rate_scheduler.py +++ b/python/paddle/fluid/layers/learning_rate_scheduler.py @@ -27,7 +27,7 @@ from . import nn from . import ops from . import tensor from ..initializer import init_on_cpu -from ..framework import default_main_program, Parameter, unique_name +from ..framework import default_main_program, Parameter, unique_name, name_scope __all__ = [ 'exponential_decay', 'natural_exp_decay', 'inverse_time_decay', @@ -332,14 +332,16 @@ def append_LARS(params_grads, learning_rate, weight_decay): return grad_norm + weight_decay * param_norm for param, grad in params_grads: - param_lr = param.optimize_attr['learning_rate'] - param_norm = ops.sqrt(nn.reduce_sum(input=ops.square(param))) - grad_norm = ops.sqrt(nn.reduce_sum(input=ops.square(grad))) - if type(param_lr) == float and param_lr == 1.0: - decayed_lr = learning_rate * param_norm \ - / _balanced_weight(param_norm, grad_norm) - else: - decayed_lr = learning_rate * param_lr * param_norm \ - / _balanced_weight(param_norm, grad_norm) - # set back param local learning rate - param.optimize_attr['learning_rate'] = decayed_lr + with param.block.program.optimized_guard( + [param, grad]), name_scope("optimizer"): + param_lr = param.optimize_attr['learning_rate'] + param_norm = ops.sqrt(nn.reduce_sum(input=ops.square(param))) + grad_norm = ops.sqrt(nn.reduce_sum(input=ops.square(grad))) + if type(param_lr) == float and param_lr == 1.0: + decayed_lr = learning_rate * param_norm \ + / _balanced_weight(param_norm, grad_norm) + else: + decayed_lr = learning_rate * param_lr * param_norm \ + / _balanced_weight(param_norm, grad_norm) + # set back param local learning rate + param.optimize_attr['learning_rate'] = decayed_lr diff --git a/python/paddle/fluid/layers/metric_op.py b/python/paddle/fluid/layers/metric_op.py index a3064b565d096f7feda18379c66ffc8bf2f4a55c..b2d2c93ead80d781d0a55ca541a1b0bb4232ad81 100644 --- a/python/paddle/fluid/layers/metric_op.py +++ b/python/paddle/fluid/layers/metric_op.py @@ -58,11 +58,11 @@ def accuracy(input, label, k=1, correct=None, total=None): """ helper = LayerHelper("accuracy", **locals()) topk_out, topk_indices = nn.topk(input, k=k) - acc_out = helper.create_tmp_variable(dtype="float32") + acc_out = helper.create_variable_for_type_inference(dtype="float32") if correct is None: - correct = helper.create_tmp_variable(dtype="int64") + correct = helper.create_variable_for_type_inference(dtype="int64") if total is None: - total = helper.create_tmp_variable(dtype="int64") + total = helper.create_variable_for_type_inference(dtype="int64") helper.append_op( type="accuracy", inputs={ @@ -124,8 +124,8 @@ def auc(input, auc_out=fluid.layers.auc(input=prediction, label=label) """ helper = LayerHelper("auc", **locals()) - auc_out = helper.create_tmp_variable(dtype="float64") - batch_auc_out = helper.create_tmp_variable(dtype="float64") + auc_out = helper.create_variable_for_type_inference(dtype="float64") + batch_auc_out = helper.create_variable_for_type_inference(dtype="float64") # make tp, tn, fp, fn persistable, so that can accumulate all batches. # for batch auc diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 8c0ef7a82421ffc04bf669e6850e075226c09d27..625dee474abe27eb167d3e36eb5b5dce5e46d351 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -27,6 +27,7 @@ from .tensor import concat from . import utils from .. import unique_name from functools import reduce +from .. import core __all__ = [ 'fc', @@ -56,6 +57,7 @@ __all__ = [ 'sequence_expand', 'sequence_expand_as', 'sequence_pad', + 'sequence_unpad', 'lstm_unit', 'reduce_sum', 'reduce_mean', @@ -64,6 +66,7 @@ __all__ = [ 'reduce_prod', 'sequence_first_step', 'sequence_last_step', + 'sequence_slice', 'dropout', 'split', 'ctc_greedy_decoder', @@ -94,10 +97,12 @@ __all__ = [ 'pad_constant_like', 'label_smooth', 'roi_pool', + 'roi_align', 'dice_loss', 'image_resize', 'image_resize_short', 'resize_bilinear', + 'resize_nearest', 'gather', 'scatter', 'sequence_scatter', @@ -107,6 +112,7 @@ __all__ = [ 'log', 'crop', 'rank_loss', + 'margin_rank_loss', 'elu', 'relu6', 'pow', @@ -150,6 +156,16 @@ __all__ = [ 'mul', 'sigmoid_cross_entropy_with_logits', 'maxout', + 'space_to_depth', + 'affine_grid', + 'sequence_reverse', + 'affine_channel', + 'similarity_focus', + 'hash', + 'grid_sampler', + 'log_loss', + 'add_position_encoding', + 'bilinear_tensor_product', ] @@ -237,7 +253,7 @@ def fc(input, w = helper.create_parameter( attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False) - tmp = helper.create_tmp_variable(dtype) + tmp = helper.create_variable_for_type_inference(dtype) helper.append_op( type="mul", inputs={"X": input_var, @@ -250,7 +266,7 @@ def fc(input, if len(mul_results) == 1: pre_bias = mul_results[0] else: - pre_bias = helper.create_tmp_variable(dtype) + pre_bias = helper.create_variable_for_type_inference(dtype) helper.append_op( type="sum", inputs={"X": mul_results}, @@ -309,7 +325,7 @@ def embedding(input, helper = LayerHelper('embedding', **locals()) w = helper.create_parameter( attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False) - tmp = helper.create_tmp_variable(dtype) + tmp = helper.create_variable_for_type_inference(dtype) padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else ( size[0] + padding_idx) helper.append_op( @@ -351,7 +367,6 @@ def dynamic_lstm(input, c_0(Variable): The initial cell state is an optional input, default is zero. This is a tensor with shape (N x D), where N is the batch size. `h_0` and `c_0` can be NULL but only at the same time. - param_attr(ParamAttr|None): The parameter attribute for the learnable hidden-hidden weights. @@ -359,6 +374,11 @@ def dynamic_lstm(input, W_{fh}, W_{oh}`} - The shape is (D x 4D), where D is the hidden size. + + If it is set to None or one attribute of ParamAttr, + dynamic_lstm will create ParamAttr as param_attr. + If the Initializer of the param_attr is not set, the + parameter is initialized with Xavier. Default: None. bias_attr (ParamAttr|None): The bias attribute for the learnable bias weights, which contains two parts, input-hidden bias weights and peephole connections weights if @@ -371,6 +391,11 @@ def dynamic_lstm(input, - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \ W_{fc}, W_{oc}`}. - The shape is (1 x 7D). + + If it is set to None or one attribute of ParamAttr, + dynamic_lstm will create ParamAttr as bias_attr. + If the Initializer of the bias_attr is not set, + the bias is initialized zero. Default: None. use_peepholes (bool): ${use_peepholes_comment} is_reverse (bool): ${is_reverse_comment} gate_activation (str): ${gate_activation_comment} @@ -389,11 +414,11 @@ def dynamic_lstm(input, hidden_dim = 512 forward_proj = fluid.layers.fc(input=input_seq, size=hidden_dim * 4, - act=None, bias_attr=None) + bias_attr=False) forward, _ = fluid.layers.dynamic_lstm( input=forward_proj, size=hidden_dim * 4, use_peepholes=False) """ - + assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp." helper = LayerHelper('lstm', **locals()) size = size // 4 weight = helper.create_parameter( @@ -404,10 +429,10 @@ def dynamic_lstm(input, bias = helper.create_parameter( attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True) - hidden = helper.create_tmp_variable(dtype) - cell = helper.create_tmp_variable(dtype) - batch_gate = helper.create_tmp_variable(dtype) - batch_cell_pre_act = helper.create_tmp_variable(dtype) + hidden = helper.create_variable_for_type_inference(dtype) + cell = helper.create_variable_for_type_inference(dtype) + batch_gate = helper.create_variable_for_type_inference(dtype) + batch_cell_pre_act = helper.create_variable_for_type_inference(dtype) inputs = {'Input': input, 'Weight': weight, 'Bias': bias} batch_size = input.shape[0] if h_0: @@ -528,6 +553,11 @@ def dynamic_lstmp(input, size. - Projection weight = {:math:`W_{rh}`}. - The shape of projection weight is (D x P). + + If it is set to None or one attribute of ParamAttr, + dynamic_lstm will create ParamAttr as param_attr. + If the Initializer of the param_attr is not set, the + parameter is initialized with Xavier. Default: None. bias_attr(ParamAttr|None): The bias attribute for the learnable bias weights, which contains two parts, input-hidden bias weights and peephole connections weights if @@ -540,6 +570,11 @@ def dynamic_lstmp(input, - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \ W_{fc}, W_{oc}`}. - The shape is (1 x 7D). + + If it is set to None or one attribute of ParamAttr, + dynamic_lstm will create ParamAttr as bias_attr. + If the Initializer of the bias_attr is not set, + the bias is initialized zero. Default: None. use_peepholes(bool): Whether to enable diagonal/peephole connections, default `True`. is_reverse(bool): Whether to compute reversed LSTM, default `False`. @@ -584,6 +619,7 @@ def dynamic_lstmp(input, proj_activation="tanh") """ + assert bias_attr is not False, "bias_attr should not be False in dynamic_lstmp." helper = LayerHelper('lstmp', **locals()) size = size // 4 weight = helper.create_parameter( @@ -596,12 +632,12 @@ def dynamic_lstmp(input, bias = helper.create_parameter( attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True) - projection = helper.create_tmp_variable(dtype) - cell = helper.create_tmp_variable(dtype) - ordered_proj0 = helper.create_tmp_variable(dtype) - batch_hidden = helper.create_tmp_variable(dtype) - batch_gate = helper.create_tmp_variable(dtype) - batch_cell_pre_act = helper.create_tmp_variable(dtype) + projection = helper.create_variable_for_type_inference(dtype) + cell = helper.create_variable_for_type_inference(dtype) + ordered_proj0 = helper.create_variable_for_type_inference(dtype) + batch_hidden = helper.create_variable_for_type_inference(dtype) + batch_gate = helper.create_variable_for_type_inference(dtype) + batch_cell_pre_act = helper.create_variable_for_type_inference(dtype) helper.append_op( type='lstmp', @@ -681,8 +717,18 @@ def dynamic_gru(input, The first part are weights of the update gate and reset gate with shape :math:`(D \\times 2D)`, and the second part are weights for candidate hidden state with shape :math:`(D \\times D)`. - bias_attr(ParamAttr): The parameter attribute for learnable the - hidden-hidden bias. + + If it is set to None or one attribute of ParamAttr, dynamic_gru will + create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias + of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates + the bias in the update gate, reset gate and candidate calculations. + If it is set to False, no bias will be applied to the update gate, + reset gate and candidate calculations. If it is set to None or one + attribute of ParamAttr, dynamic_gru will create ParamAttr as + bias_attr. If the Initializer of the bias_attr is not set, the bias + is initialized zero. Default: None. is_reverse(bool): Whether to compute reversed GRU, default :attr:`False`. gate_activation(str): The activation for update gate and reset gate. @@ -720,16 +766,16 @@ def dynamic_gru(input, attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True) batch_size = input.shape[0] inputs = {'Input': input, 'Weight': weight, 'Bias': bias} - if h_0 != None: + if h_0: assert h_0.shape == ( batch_size, size ), 'The shape of h0 should be(batch_size, %d)' % size inputs['H0'] = h_0 - hidden = helper.create_tmp_variable(dtype) - batch_gate = helper.create_tmp_variable(dtype) - batch_reset_hidden_prev = helper.create_tmp_variable(dtype) - batch_hidden = helper.create_tmp_variable(dtype) + hidden = helper.create_variable_for_type_inference(dtype) + batch_gate = helper.create_variable_for_type_inference(dtype) + batch_reset_hidden_prev = helper.create_variable_for_type_inference(dtype) + batch_hidden = helper.create_variable_for_type_inference(dtype) helper.append_op( type='gru', @@ -781,10 +827,29 @@ def gru_unit(input, Args: input (Variable): The fc transformed input value of current step. - hidden (Variable): The hidden value of lstm unit from previous step. + hidden (Variable): The hidden value of gru unit from previous step. size (integer): The input dimension value. - param_attr (ParamAttr): The weight parameters for gru unit. Default: None - bias_attr (ParamAttr): The bias parameters for gru unit. Default: None + param_attr(ParamAttr|None): The parameter attribute for the learnable + hidden-hidden weight matrix. Note: + + - The shape of the weight matrix is :math:`(T \\times 3D)`, where + :math:`D` is the hidden size. + - All elements in the weight matrix can be divided into two parts. + The first part are weights of the update gate and reset gate with + shape :math:`(D \\times 2D)`, and the second part are weights for + candidate hidden state with shape :math:`(D \\times D)`. + + If it is set to None or one attribute of ParamAttr, gru_unit will + create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias + of GRU. Note that the bias with :math:`(1 \\times 3D)` concatenates + the bias in the update gate, reset gate and candidate calculations. + If it is set to False, no bias will be applied to the update gate, + reset gate and candidate calculations. If it is set to None or one + attribute of ParamAttr, gru_unit will create ParamAttr as + bias_attr. If the Initializer of the bias_attr is not set, the bias + is initialized zero. Default: None. activation (string): The activation type for cell (actNode). Default: 'tanh' gate_activation (string): The activation type for gates (actGate). @@ -819,9 +884,9 @@ def gru_unit(input, weight = helper.create_parameter( attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype) - gate = helper.create_tmp_variable(dtype) - reset_hidden_pre = helper.create_tmp_variable(dtype) - updated_hidden = helper.create_tmp_variable(dtype) + gate = helper.create_variable_for_type_inference(dtype) + reset_hidden_pre = helper.create_variable_for_type_inference(dtype) + updated_hidden = helper.create_variable_for_type_inference(dtype) inputs = {'Input': input, 'HiddenPrev': hidden, 'Weight': weight} # create bias if helper.bias_attr: @@ -871,10 +936,14 @@ def linear_chain_crf(input, label, param_attr=None): attr=helper.param_attr, shape=[size + 2, size], dtype=helper.input_dtype()) - alpha = helper.create_tmp_variable(dtype=helper.input_dtype()) - emission_exps = helper.create_tmp_variable(dtype=helper.input_dtype()) - transition_exps = helper.create_tmp_variable(dtype=helper.input_dtype()) - log_likelihood = helper.create_tmp_variable(dtype=helper.input_dtype()) + alpha = helper.create_variable_for_type_inference( + dtype=helper.input_dtype()) + emission_exps = helper.create_variable_for_type_inference( + dtype=helper.input_dtype()) + transition_exps = helper.create_variable_for_type_inference( + dtype=helper.input_dtype()) + log_likelihood = helper.create_variable_for_type_inference( + dtype=helper.input_dtype()) helper.append_op( type='linear_chain_crf', inputs={"Emission": [input], @@ -913,7 +982,8 @@ def crf_decoding(input, param_attr, label=None): """ helper = LayerHelper('crf_decoding', **locals()) transition = helper.get_parameter(param_attr.name) - viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype()) + viterbi_path = helper.create_variable_for_type_inference( + dtype=helper.input_dtype()) helper.append_op( type='crf_decoding', inputs={"Emission": [input], @@ -937,9 +1007,9 @@ def cos_sim(X, Y): Variable: the output of cosine(X, Y). """ helper = LayerHelper('cos_sim', **locals()) - out = helper.create_tmp_variable(dtype=X.dtype) - xnorm = helper.create_tmp_variable(dtype=X.dtype) - ynorm = helper.create_tmp_variable(dtype=X.dtype) + out = helper.create_variable_for_type_inference(dtype=X.dtype) + xnorm = helper.create_variable_for_type_inference(dtype=X.dtype) + ynorm = helper.create_variable_for_type_inference(dtype=X.dtype) helper.append_op( type='cos_sim', inputs={'X': [X], @@ -950,7 +1020,12 @@ def cos_sim(X, Y): return out -def dropout(x, dropout_prob, is_test=False, seed=None, name=None): +def dropout(x, + dropout_prob, + is_test=False, + seed=None, + name=None, + dropout_implementation="downgrade_in_infer"): """ Computes dropout. @@ -970,6 +1045,21 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None): units will be dropped. DO NOT use a fixed seed in training. name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. + dropout_implementation(string): ['downgrade_in_infer'(defauld)|'upscale_in_train'] + 1. downgrade_in_infer(default), downgrade the outcome at inference + train: out = input * mask + inference: out = input * dropout_prob + (make is a tensor same shape with input, value is 0 or 1 + ratio of 0 is dropout_prob) + 2. upscale_in_train, upscale the outcome at training time + train: out = input * mask / ( 1.0 - dropout_prob ) + inference: out = input + (make is a tensor same shape with input, value is 0 or 1 + ratio of 0 is dropout_prob) + dropout op can be removed from the program. + the program will be efficient + + Returns: Variable: A tensor variable is the shape with `x`. @@ -983,8 +1073,9 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None): """ helper = LayerHelper('dropout', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) - mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True) + out = helper.create_variable_for_type_inference(dtype=x.dtype) + mask = helper.create_variable_for_type_inference( + dtype=x.dtype, stop_gradient=True) if (seed is None or seed == 0) and helper.main_program.random_seed != 0: seed = helper.main_program.random_seed @@ -998,7 +1089,8 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None): 'dropout_prob': dropout_prob, 'is_test': is_test, 'fix_seed': seed is not None, - 'seed': seed if seed is not None else 0 + 'seed': seed if seed is not None else 0, + 'dropout_implementation': dropout_implementation, }) return out @@ -1069,7 +1161,7 @@ def cross_entropy(input, label, soft_label=False, ignore_index=-100): cost = fluid.layers.cross_entropy(input=predict, label=label) """ helper = LayerHelper('cross_entropy', **locals()) - out = helper.create_tmp_variable(dtype=input.dtype) + out = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type='cross_entropy', inputs={'X': [input], @@ -1116,14 +1208,14 @@ def square_error_cost(input, label): """ helper = LayerHelper('square_error_cost', **locals()) - minus_out = helper.create_tmp_variable(dtype=input.dtype) + minus_out = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type='elementwise_sub', inputs={'X': [input], 'Y': [label]}, outputs={'Out': [minus_out]}) - square_out = helper.create_tmp_variable(dtype=input.dtype) + square_out = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type='square', inputs={'X': [minus_out]}, outputs={'Out': [square_out]}) @@ -1229,12 +1321,13 @@ def chunk_eval(input, helper = LayerHelper("chunk_eval", **locals()) # prepare output - precision = helper.create_tmp_variable(dtype="float32") - recall = helper.create_tmp_variable(dtype="float32") - f1_score = helper.create_tmp_variable(dtype="float32") - num_infer_chunks = helper.create_tmp_variable(dtype="int64") - num_label_chunks = helper.create_tmp_variable(dtype="int64") - num_correct_chunks = helper.create_tmp_variable(dtype="int64") + precision = helper.create_variable_for_type_inference(dtype="float32") + recall = helper.create_variable_for_type_inference(dtype="float32") + f1_score = helper.create_variable_for_type_inference(dtype="float32") + num_infer_chunks = helper.create_variable_for_type_inference(dtype="int64") + num_label_chunks = helper.create_variable_for_type_inference(dtype="int64") + num_correct_chunks = helper.create_variable_for_type_inference( + dtype="int64") helper.append_op( type="chunk_eval", @@ -1265,7 +1358,8 @@ def sequence_conv(input, padding=None, bias_attr=None, param_attr=None, - act=None): + act=None, + name=None): """ This function creates the op for sequence_conv, using the inputs and other convolutional configurations for the filters and stride as given @@ -1277,9 +1371,19 @@ def sequence_conv(input, filter_size (int): the filter size (H and W). filter_stride (int): stride of the filter. padding (bool): if True, add paddings. - bias_attr (ParamAttr|None): attributes for bias - param_attr (ParamAttr|None): attributes for parameter - act (str): the activation type + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of sequence_conv. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, sequence_conv + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of sequence_conv. If it is set to None or one attribute of ParamAttr, sequence_conv + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + act (str): Activation type, if it is set to None, activation is not appended. + Default: None. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Variable: output of sequence_conv @@ -1290,7 +1394,7 @@ def sequence_conv(input, filter_shape = [filter_size * input.shape[1], num_filters] filter_param = helper.create_parameter( attr=helper.param_attr, shape=filter_shape, dtype=dtype) - pre_bias = helper.create_tmp_variable(dtype) + pre_bias = helper.create_variable_for_type_inference(dtype) helper.append_op( type='sequence_conv', @@ -1308,7 +1412,7 @@ def sequence_conv(input, return helper.append_activation(pre_act) -def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False): +def sequence_softmax(input, use_cudnn=False, name=None): """ This function computes the softmax activation among all time-steps for each sequence. The dimension of each time-step should be 1. Thus, the shape of @@ -1328,10 +1432,10 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False): Args: input (Variable): The input variable which is a LoDTensor. - bias_attr (ParamAttr|None): attributes for bias - param_attr (ParamAttr|None): attributes for parameter use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \ - library is installed. Default: False + library is installed. Default: False. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Variable: output of sequence_softmax @@ -1346,7 +1450,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False): """ helper = LayerHelper('sequence_softmax', **locals()) dtype = helper.input_dtype() - softmax_out = helper.create_tmp_variable(dtype) + softmax_out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="sequence_softmax", inputs={"X": input}, @@ -1355,7 +1459,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=False): return softmax_out -def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None): +def softmax(input, use_cudnn=True, name=None): """ The input of the softmax operator is a tensor of any rank. The output tensor has the same shape as the input. @@ -1382,10 +1486,10 @@ def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None): Args: input (Variable): The input variable. - bias_attr (ParamAttr): attributes for bias - param_attr (ParamAttr): attributes for parameter use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \ - library is installed. + library is installed. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Variable: output of softmax @@ -1400,7 +1504,7 @@ def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None): """ helper = LayerHelper('softmax', **locals()) dtype = helper.input_dtype() - softmax_out = helper.create_tmp_variable(dtype) + softmax_out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="softmax", inputs={"X": input}, @@ -1491,14 +1595,23 @@ def conv2d(input, convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only - connected to the second half of the input channels. Default: groups=1 - param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None - bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None + connected to the second half of the input channels. Default: groups=1. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv2d. If it is set to None or one attribute of ParamAttr, conv2d + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with :math:`Normal(0.0, std)`, + and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv2d + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True - act (str): Activation type. Default: None + act (str): Activation type, if it is set to None, activation is not appended. + Default: None name (str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + will be named automatically. Default: None Returns: Variable: The tensor variable storing the convolution and \ @@ -1516,7 +1629,7 @@ def conv2d(input, """ num_channels = input.shape[1] - + assert param_attr is not False, "param_attr should not be False here." l_type = 'conv2d' if (num_channels == groups and num_filters % num_channels == 0 and not use_cudnn): @@ -1544,7 +1657,8 @@ def conv2d(input, filter_shape = [num_filters, int(num_filter_channels)] + filter_size def _get_default_param_initializer(): - std = (2.0 / (filter_size[0]**2 * num_channels))**0.5 + filter_elem_num = filter_size[0] * filter_size[1] * num_channels + std = (2.0 / filter_elem_num)**0.5 return Normal(0.0, std, 0) filter_param = helper.create_parameter( @@ -1553,7 +1667,21 @@ def conv2d(input, dtype=dtype, default_initializer=_get_default_param_initializer()) - pre_bias = helper.create_tmp_variable(dtype) + pre_bias = helper.create_variable_for_type_inference(dtype) + + if use_cudnn: + helper.create_variable( + name="kCUDNNFwdAlgoCache", + persistable=True, + type=core.VarDesc.VarType.RAW) + helper.create_variable( + name="kCUDNNBwdDataAlgoCache", + persistable=True, + type=core.VarDesc.VarType.RAW) + helper.create_variable( + name="kCUDNNBwdFilterAlgoCache", + persistable=True, + type=core.VarDesc.VarType.RAW) helper.append_op( type=l_type, @@ -1568,7 +1696,7 @@ def conv2d(input, 'dilations': dilation, 'groups': groups, 'use_cudnn': use_cudnn, - 'use_mkldnn': False + 'use_mkldnn': False, }) pre_act = helper.append_bias_op(pre_bias, dim_start=1, dim_end=2) @@ -1655,13 +1783,22 @@ def conv3d(input, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels. Default: groups=1 - param_attr (ParamAttr): The parameters to the Conv3d Layer. Default: None - bias_attr (ParamAttr): Bias parameter for the Conv3d layer. Default: None + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv3d. If it is set to None or one attribute of ParamAttr, conv3d + will create ParamAttr as param_attr. If it is set to None, the parameter + is initialized with :math:`Normal(0.0, std)`, and the :math:`std` is + :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv3d + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True - act (str): Activation type. Default: None + act (str): Activation type, if it is set to None, activation is not appended. + Default: None. name (str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + will be named automatically. Default: None. Returns: Variable: The tensor variable storing the convolution and \ @@ -1679,7 +1816,7 @@ def conv3d(input, """ l_type = 'conv3d' - + assert param_attr is not False, "param_attr should not be False here." helper = LayerHelper(l_type, **locals()) dtype = helper.input_dtype() @@ -1704,7 +1841,9 @@ def conv3d(input, filter_shape = [num_filters, num_filter_channels] + filter_size def _get_default_param_initializer(): - std = (2.0 / (filter_size[0]**3 * num_channels))**0.5 + filter_elem_num = filter_size[0] * filter_size[1] * filter_size[ + 2] * num_channels + std = (2.0 / filter_elem_num)**0.5 return Normal(0.0, std, 0) filter_param = helper.create_parameter( @@ -1713,7 +1852,7 @@ def conv3d(input, dtype=dtype, default_initializer=_get_default_param_initializer()) - pre_bias = helper.create_tmp_variable(dtype) + pre_bias = helper.create_variable_for_type_inference(dtype) helper.append_op( type=l_type, @@ -1736,7 +1875,7 @@ def conv3d(input, return helper.append_activation(pre_act) -def sequence_pool(input, pool_type): +def sequence_pool(input, pool_type, is_test=False): """ This function add the operator for sequence pooling. It pools features of all time-steps of each instance, and is applied @@ -1773,6 +1912,7 @@ def sequence_pool(input, pool_type): input(variable): The input variable which is a LoDTensor. pool_type (string): The pooling type of sequence_pool. It supports average, sum, sqrt and max. + is_test(bool, Default False): Used distinguish training from scoring mode. Returns: The sequence pooling variable which is a Tensor. @@ -1792,15 +1932,16 @@ def sequence_pool(input, pool_type): """ helper = LayerHelper('sequence_pool', **locals()) dtype = helper.input_dtype() - pool_out = helper.create_tmp_variable(dtype) - max_index = helper.create_tmp_variable(dtype) + pool_out = helper.create_variable_for_type_inference(dtype) + max_index = helper.create_variable_for_type_inference(dtype) helper.append_op( type="sequence_pool", inputs={"X": input}, outputs={"Out": pool_out, "MaxIndex": max_index}, - attrs={"pooltype": pool_type.upper()}) + attrs={"pooltype": pool_type.upper(), + "is_test": is_test}) # when pool_type is max, variable max_index is initialized, # so we stop the gradient explicitly here @@ -1829,7 +1970,7 @@ def sequence_concat(input, name=None): out = fluid.layers.sequence_concat(input=[seq1, seq2, seq3]) """ helper = LayerHelper('sequence_concat', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) helper.append_op( type='sequence_concat', inputs={'X': input}, outputs={'Out': [out]}) return out @@ -1901,6 +2042,76 @@ def sequence_last_step(input): return sequence_pool(input=input, pool_type="last") +def sequence_slice(input, offset, length, name=None): + """ + **Sequence Slice Layer** + + The layer crops a subsequence from given sequence with given start + offset and subsequence length. + + It only supports sequence data (LoDTensor with lod_level equal to 1). + + .. code-block:: text + + - Case: + + Given the input Variable **input**: + + input.data = [[a1, a2], [b1, b2], [c1, c2], [d1, d2], [e1, e2]], + input.lod = [[3, 2]], + input.dims = (5, 2), + + with offset.data = [[0], [1]] and length.data = [[2], [1]], + + the output Variable will be + + out.data = [[a1, a2], [b1, b2], [e1, e2]], + out.lod = [[2, 1]], + out.dims = (3, 2). + + NOTE: The first dimension size of **input**, **offset** and **length** + should be equal. The **offset** should start from 0. + + Args: + input(Variable): The input Variable which consists of the complete + sequences. + offset(Variable): The offset to slice each sequence. + length(Variable): The length of each subsequence. + name(str|None): A name for this layer(optional). If set None, the + layer will be named automatically. + + Returns: + Variable: The output subsequences. + + Examples: + + .. code-block:: python + + import numpy as np + seqs = fluid.layers.data(name='x', shape=[10, 5], + dtype='float32', lod_level=1) + offset = fluid.layers.assign(input=np.array([[0, 1]]).astype("int32")) + length = fluid.layers.assign(input=np.array([[2, 1]]).astype("int32")) + subseqs = fluid.layers.sequence_slice(input=seqs, offset=offset, + length=length) + """ + helper = LayerHelper("sequence_slice", **locals()) + dtype = helper.input_dtype() + out = helper.create_variable_for_type_inference(dtype) + + offset.stop_gradient = True + length.stop_gradient = True + + helper.append_op( + type="sequence_slice", + inputs={"X": input, + "Offset": offset, + "Length": length}, + outputs={"Out": out}) + + return out + + @templatedoc() def pool2d(input, pool_size=-1, @@ -1910,7 +2121,8 @@ def pool2d(input, global_pooling=False, use_cudnn=True, ceil_mode=False, - name=None): + name=None, + exclusive=True): """ ${comment} @@ -1924,11 +2136,13 @@ def pool2d(input, pool_type: ${pooling_type_comment} pool_stride (int): stride of the pooling layer. pool_padding (int): padding size. - global_pooling: ${global_pooling_comment} - use_cudnn: ${use_cudnn_comment} - ceil_mode: ${ceil_mode_comment} + global_pooling (bool): ${global_pooling_comment} + use_cudnn (bool): ${use_cudnn_comment} + ceil_mode (bool): ${ceil_mode_comment} name (str|None): A name for this layer(optional). If set None, the layer will be named automatically. + exclusive (bool): Whether to exclude padding points in average pooling + mode, default is true Returns: Variable: The pooling result. @@ -1972,7 +2186,7 @@ def pool2d(input, helper = LayerHelper(l_type, **locals()) dtype = helper.input_dtype() - pool_out = helper.create_tmp_variable(dtype) + pool_out = helper.create_variable_for_type_inference(dtype) helper.append_op( type=l_type, @@ -1986,7 +2200,8 @@ def pool2d(input, "paddings": pool_padding, "use_cudnn": use_cudnn, "ceil_mode": ceil_mode, - "use_mkldnn": False + "use_mkldnn": False, + "exclusive": exclusive, }) return pool_out @@ -2000,7 +2215,8 @@ def pool3d(input, global_pooling=False, use_cudnn=True, ceil_mode=False, - name=None): + name=None, + exclusive=True): """ This function adds the operator for pooling in 3-dimensions, using the pooling configurations mentioned in input parameters. @@ -2016,6 +2232,8 @@ def pool3d(input, ceil_mode (bool): ${ceil_mode_comment} name (str): A name for this layer(optional). If set None, the layer will be named automatically. + exclusive (bool): Whether to exclude padding points in average pooling + mode, default is true Returns: Variable: output of pool3d layer. @@ -2040,7 +2258,7 @@ def pool3d(input, l_type = "pool3d" helper = LayerHelper(l_type, **locals()) dtype = helper.input_dtype() - pool_out = helper.create_tmp_variable(dtype) + pool_out = helper.create_variable_for_type_inference(dtype) helper.append_op( type=l_type, @@ -2054,7 +2272,8 @@ def pool3d(input, "paddings": pool_padding, "use_cudnn": use_cudnn, "ceil_mode": ceil_mode, - "use_mkldnn": False + "use_mkldnn": False, + "exclusive": exclusive, }) return pool_out @@ -2106,8 +2325,14 @@ def batch_norm(input, is_test(bool, Default False): Used for training or training. momentum(float, Default 0.9): epsilon(float, Default 1e-05): - param_attr(ParamAttr): The parameter attribute for Parameter `scale`. - bias_attr(ParamAttr): The parameter attribute for Parameter `bias`. + param_attr(ParamAttr|None): The parameter attribute for Parameter `scale` + of batch_norm. If it is set to None or one attribute of ParamAttr, batch_norm + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr(ParamAttr|None): The parameter attribute for the bias of batch_norm. + If it is set to None or one attribute of ParamAttr, batch_norm + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. data_layout(string, default NCHW): NCHW|NHWC in_place(bool, Default False): Make the input and output of batch norm reuse memory. name(string, Default None): A name for this layer(optional). If set None, the layer @@ -2127,6 +2352,7 @@ def batch_norm(input, hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w') hidden2 = fluid.layers.batch_norm(input=hidden1) """ + assert bias_attr is not False, "bias_attr should not be False in batch_norm." helper = LayerHelper('batch_norm', **locals()) dtype = helper.input_dtype() @@ -2176,10 +2402,13 @@ def batch_norm(input, mean_out = mean # variance and variance out share the same memory variance_out = variance - saved_mean = helper.create_tmp_variable(dtype=dtype, stop_gradient=True) - saved_variance = helper.create_tmp_variable(dtype=dtype, stop_gradient=True) + saved_mean = helper.create_variable_for_type_inference( + dtype=dtype, stop_gradient=True) + saved_variance = helper.create_variable_for_type_inference( + dtype=dtype, stop_gradient=True) - batch_norm_out = input if in_place else helper.create_tmp_variable(dtype) + batch_norm_out = input if in_place else helper.create_variable_for_type_inference( + dtype) helper.append_op( type="batch_norm", @@ -2243,19 +2472,28 @@ def layer_norm(input, Args: input(Variable): The input tensor variable. scale(bool): Whether to learn the adaptive gain :math:`g` after - normalization. + normalization. Default True. shift(bool): Whether to learn the adaptive bias :math:`b` after - normalization. - begin_norm_axis(bool): The normalization will be performed along + normalization. Default True. + begin_norm_axis(int): The normalization will be performed along dimensions from :attr:`begin_norm_axis` to :attr:`rank(input)`. + Default 1. epsilon(float): The small value added to the variance to prevent - division by zero. + division by zero. Default 1e-05. param_attr(ParamAttr|None): The parameter attribute for the learnable - gain :math:`g`. + gain :math:`g`. If :attr:`scale` is False, :attr:`param_attr` is + omitted. If :attr:`scale` is True and :attr:`param_attr` is None, + a default :code:`ParamAttr` would be added as scale. The + :attr:`param_attr` is initialized as 1 if it is added. Default None. bias_attr(ParamAttr|None): The parameter attribute for the learnable - bias :math:`b`. + bias :math:`b`. If :attr:`shift` is False, :attr:`bias_attr` is + omitted. If :attr:`shift` is True and :attr:`param_attr` is None, + a default :code:`ParamAttr` would be added as bias. The + :attr:`bias_attr` is initialized as 0 if it is added. Default None. act(str): Activation to be applied to the output of layer normalizaiton. - name (str): The name of this layer. It is optional. + Default None. + name(str): The name of this layer. It is optional. Default None, and a + unique name would be generated automatically. Returns: ${y_comment} @@ -2287,9 +2525,11 @@ def layer_norm(input, inputs['Bias'] = bias # create output - mean_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True) - variance_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True) - layer_norm_out = helper.create_tmp_variable(dtype) + mean_out = helper.create_variable_for_type_inference( + dtype=dtype, stop_gradient=True) + variance_out = helper.create_variable_for_type_inference( + dtype=dtype, stop_gradient=True) + layer_norm_out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="layer_norm", @@ -2396,15 +2636,22 @@ def conv2d_transpose(input, when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only connected to the second half of the input channels. - Default: groups=1 - param_attr(ParamAttr): The parameters to the Conv2d_transpose Layer. - Default: None - bias_attr(ParamAttr): Bias parameter for the Conv2d layer. Default: None + Default: groups = 1. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv2d_transpose. If it is set to None or one attribute of ParamAttr, conv2d_transpose + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d_transpose. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv2d_transpose + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn - library is installed. Default: True - act(str): Activation type. Default: None + library is installed. Default: True. + act (str): Activation type, if it is set to None, activation is not appended. + Default: None. name(str|None): A name for this layer(optional). If set None, the layer - will be named automatically. + will be named automatically. Default: True. Returns: Variable: The tensor variable storing the convolution transpose result. @@ -2419,7 +2666,7 @@ def conv2d_transpose(input, data = fluid.layers.data(name='data', shape=[3, 32, 32], dtype='float32') conv2d_transpose = fluid.layers.conv2d_transpose(input=data, num_filters=2, filter_size=3) """ - + assert param_attr is not False, "param_attr should not be False in conv2d_transpose." input_channel = input.shape[1] op_type = 'conv2d_transpose' @@ -2455,6 +2702,7 @@ def conv2d_transpose(input, else: filter_size = utils.convert_to_list(filter_size, 2, 'conv2d_transpose.filter_size') + if output_size is None: output_size = [] elif isinstance(output_size, list) or isinstance(output_size, int): @@ -2464,10 +2712,11 @@ def conv2d_transpose(input, padding = utils.convert_to_list(padding, 2, 'padding') groups = 1 if groups is None else groups filter_shape = [input_channel, num_filters // groups] + filter_size + img_filter = helper.create_parameter( dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) - pre_bias = helper.create_tmp_variable(dtype=input.dtype) + pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type=op_type, inputs={'Input': [input], @@ -2576,12 +2825,19 @@ def conv3d_transpose(input, first half of the input channels, while the second half of the filters is only connected to the second half of the input channels. Default: groups=1 - param_attr(ParamAttr): The parameters to the Conv3d_transpose Layer. - Default: None - bias_attr(ParamAttr): Bias parameter for the Conv3d layer. Default: None + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv3d_transpose. If it is set to None or one attribute of ParamAttr, conv3d_transpose + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv3d_transpose. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv3d_transpose + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. use_cudnn(bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True - act(str): Activation type. Default: None + act (str): Activation type, if it is set to None, activation is not appended. + Default: None. name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. @@ -2598,6 +2854,7 @@ def conv3d_transpose(input, data = fluid.layers.data(name='data', shape=[3, 12, 32, 32], dtype='float32') conv3d_transpose = fluid.layers.conv3d_transpose(input=data, num_filters=2, filter_size=3) """ + assert param_attr is not False, "param_attr should not be False in conv3d_transpose." l_type = "conv3d_transpose" helper = LayerHelper(l_type, **locals()) if not isinstance(input, Variable): @@ -2637,7 +2894,7 @@ def conv3d_transpose(input, img_filter = helper.create_parameter( dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) - pre_bias = helper.create_tmp_variable(dtype=input.dtype) + pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type=l_type, inputs={'Input': [input], @@ -2716,7 +2973,7 @@ def sequence_expand(x, y, ref_level=-1, name=None): """ helper = LayerHelper('sequence_expand', input=x, **locals()) dtype = helper.input_dtype() - tmp = helper.create_tmp_variable(dtype) + tmp = helper.create_variable_for_type_inference(dtype) helper.append_op( type='sequence_expand', inputs={'X': x, @@ -2782,7 +3039,7 @@ def sequence_expand_as(x, y, name=None): """ helper = LayerHelper('sequence_expand_as', input=x, **locals()) dtype = helper.input_dtype() - tmp = helper.create_tmp_variable(dtype) + tmp = helper.create_variable_for_type_inference(dtype) helper.append_op( type='sequence_expand_as', inputs={'X': x, @@ -2792,7 +3049,7 @@ def sequence_expand_as(x, y, name=None): @templatedoc() -def sequence_pad(x, pad_value, maxlen=None): +def sequence_pad(x, pad_value, maxlen=None, name=None): """ ${comment} @@ -2806,7 +3063,9 @@ def sequence_pad(x, pad_value, maxlen=None): None or any positive int. When it is None, all sequences will be padded up to the length of the longest one among them; when it a certain positive value, it must be greater than the length of the - longest original sequence." + longest original sequence. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Returns: Variable: The padded sequence batch and the original lengths before @@ -2819,14 +3078,15 @@ def sequence_pad(x, pad_value, maxlen=None): x = fluid.layers.data(name='y', shape=[10, 5], dtype='float32', lod_level=1) - pad_value = fluid.layers.assign(input=numpy.array([0])) + pad_value = fluid.layers.assign( + input=numpy.array([0.0], dtype=numpy.float32)) out = fluid.layers.sequence_pad(x=x, pad_value=pad_value) """ helper = LayerHelper('sequence_pad', input=x, **locals()) dtype = helper.input_dtype() - out = helper.create_tmp_variable(dtype) - length = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) + length = helper.create_variable_for_type_inference(dtype) pad_value.stop_gradient = True length.stop_gradient = True @@ -2843,6 +3103,66 @@ def sequence_pad(x, pad_value, maxlen=None): return out, length +def sequence_unpad(x, length, name=None): + """ + **Sequence Unpad Layer** + + This layer removes the padding data in the input sequences and convert + them into sequences with actual length as output, identitied by lod + information. + + .. code-block:: text + + Example: + + Given input Variable **x**: + x.data = [[ 1.0, 2.0, 3.0, 4.0, 5.0], + [ 6.0, 7.0, 8.0, 9.0, 10.0], + [11.0, 12.0, 13.0, 14.0, 15.0]], + + in which there are 3 sequences padded to length 5, and the acutal length + specified by input Variable **length**: + + length.data = [[2], [3], [4]], + + after unpadding, the output Variable will be: + + out.data = [[1.0, 2.0, 6.0, 7.0, 8.0, 11.0, 12.0, 13.0, 14.0]] + out.lod = [[2, 3, 4]] + + Args: + x(Variable): Input Variable which contains the padded sequences with + equal length. + length(Variable): The Variable that specifies the actual ength of + sequences after unpadding. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The Variable contains the unpadded sequences. + + Examples: + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[10, 5], dtype='float32') + len = fluid.layers.data(name='length', shape=[1], dtype='int64') + out = fluid.layers.sequence_unpad(x=x, length=len) + """ + + helper = LayerHelper('sequence_unpad', input=x, **locals()) + dtype = helper.input_dtype() + out = helper.create_variable_for_type_inference(dtype) + + length.stop_gradient = True + + helper.append_op( + type='sequence_unpad', + inputs={'X': x, + 'Length': length}, + outputs={'Out': out}) + return out + + def beam_search(pre_ids, pre_scores, ids, @@ -2930,8 +3250,9 @@ def beam_search(pre_ids, score_type = scores.dtype id_type = ids.dtype - selected_scores = helper.create_tmp_variable(dtype=score_type) - selected_ids = helper.create_tmp_variable(dtype=id_type) + selected_scores = helper.create_variable_for_type_inference( + dtype=score_type) + selected_ids = helper.create_variable_for_type_inference(dtype=id_type) helper.append_op( type='beam_search', @@ -2988,8 +3309,8 @@ def beam_search_decode(ids, scores, beam_size, end_id, name=None): ids, scores, beam_size=5, end_id=0) """ helper = LayerHelper('beam_search_decode', **locals()) - sentence_ids = helper.create_tmp_variable(dtype=ids.dtype) - sentence_scores = helper.create_tmp_variable(dtype=ids.dtype) + sentence_ids = helper.create_variable_for_type_inference(dtype=ids.dtype) + sentence_scores = helper.create_variable_for_type_inference(dtype=ids.dtype) helper.append_op( type="beam_search_decode", @@ -3054,10 +3375,18 @@ def lstm_unit(x_t, cell_t_prev (Variable): The cell value of lstm unit, a 2-D tensor with shape M x S, M for batch size and S for size of lstm unit. forget_bias (float): The forget bias of lstm unit. - param_attr (ParamAttr): The attributes of parameter weights, used to set - initializer, name etc. - bias_attr (ParamAttr): The attributes of bias weights, if not False, - bias weights will be created and be set to default value. + param_attr(ParamAttr|None): The parameter attribute for the learnable + hidden-hidden weights. + If it is set to None or one attribute of ParamAttr, + lstm_unit will create ParamAttr as param_attr. + If the Initializer of the param_attr is not set, the + parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|None): The bias attribute for the learnable bias + weights. If it is set to False, no bias will be added + to the output units. If it is set to None or one attribute of ParamAttr, + lstm_unit will create ParamAttr as bias_attr. + If the Initializer of the bias_attr is not set, + the bias is initialized zero. Default: None. name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. @@ -3111,8 +3440,8 @@ def lstm_unit(x_t, param_attr=param_attr, bias_attr=bias_attr) dtype = x_t.dtype - c = helper.create_tmp_variable(dtype) - h = helper.create_tmp_variable(dtype) + c = helper.create_variable_for_type_inference(dtype) + h = helper.create_variable_for_type_inference(dtype) helper.append_op( type='lstm_unit', @@ -3166,7 +3495,7 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None): """ helper = LayerHelper('reduce_sum', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) if dim is not None and not isinstance(dim, list): dim = [dim] helper.append_op( @@ -3223,7 +3552,7 @@ def reduce_mean(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_mean(x, dim=[0, 1]) # [4.0, 5.0] """ helper = LayerHelper('reduce_mean', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) if dim is not None and not isinstance(dim, list): dim = [dim] helper.append_op( @@ -3278,7 +3607,7 @@ def reduce_max(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_max(x, dim=[0, 1]) # [7.0, 8.0] """ helper = LayerHelper('reduce_max', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) if dim is not None and not isinstance(dim, list): dim = [dim] helper.append_op( @@ -3333,7 +3662,7 @@ def reduce_min(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_min(x, dim=[0, 1]) # [1.0, 2.0] """ helper = LayerHelper('reduce_min', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) if dim is not None and not isinstance(dim, list): dim = [dim] helper.append_op( @@ -3389,7 +3718,7 @@ def reduce_prod(input, dim=None, keep_dim=False, name=None): fluid.layers.reduce_prod(x, dim=[0, 1]) # [105.0, 384.0] """ helper = LayerHelper('reduce_prod', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) if dim is not None and not isinstance(dim, list): dim = [dim] helper.append_op( @@ -3449,7 +3778,7 @@ def split(input, num_or_sections, dim=-1, name=None): dim], 'len(num_or_sections) must not be more than input.shape[dim].' num = len(num_or_sections) outs = [ - helper.create_tmp_variable(dtype=helper.input_dtype()) + helper.create_variable_for_type_inference(dtype=helper.input_dtype()) for i in range(num) ] helper.append_op( @@ -3506,8 +3835,8 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None): axis = 0 helper = LayerHelper("l2_normalize", **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) - norm = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) + norm = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type="norm", inputs={"X": x}, @@ -3616,7 +3945,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None): __check_input(x, y) helper = LayerHelper('matmul', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='matmul', inputs={'X': x, @@ -3687,8 +4016,8 @@ def topk(input, k, name=None): top5_values, top5_indices = layers.topk(input, k=5) """ helper = LayerHelper("top_k", **locals()) - values = helper.create_tmp_variable(dtype=input.dtype) - indices = helper.create_tmp_variable(dtype="int64") + values = helper.create_variable_for_type_inference(dtype=input.dtype) + indices = helper.create_variable_for_type_inference(dtype="int64") helper.append_op( type="top_k", inputs={"X": [input]}, @@ -3738,16 +4067,16 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None): Examples: .. code-block:: python - x = fluid.layers.data(name='x', shape=[8], dtype='float32') - y = fluid.layers.data(name='y', shape=[7], dtype='float32') + x = fluid.layers.data(name='x', shape=[1], dtype='float32') + y = fluid.layers.data(name='y', shape=[1], dtype='float32') cost = fluid.layers.edit_distance(input=x,label=y) """ helper = LayerHelper("edit_distance", **locals()) # remove some tokens from input and labels if ignored_tokens is not None and len(ignored_tokens) > 0: - erased_input = helper.create_tmp_variable(dtype="int64") - erased_label = helper.create_tmp_variable(dtype="int64") + erased_input = helper.create_variable_for_type_inference(dtype="int64") + erased_label = helper.create_variable_for_type_inference(dtype="int64") helper.append_op( type="sequence_erase", @@ -3764,8 +4093,8 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None): label = erased_label # edit distance op - edit_distance_out = helper.create_tmp_variable(dtype="int64") - sequence_num = helper.create_tmp_variable(dtype="int64") + edit_distance_out = helper.create_variable_for_type_inference(dtype="int64") + sequence_num = helper.create_variable_for_type_inference(dtype="int64") helper.append_op( type="edit_distance", inputs={"Hyps": [input], @@ -3840,7 +4169,7 @@ def ctc_greedy_decoder(input, blank, name=None): _, topk_indices = topk(input, k=1) # ctc align op - ctc_out = helper.create_tmp_variable(dtype="int64") + ctc_out = helper.create_variable_for_type_inference(dtype="int64") helper.append_op( type="ctc_align", inputs={"Input": [topk_indices]}, @@ -3890,8 +4219,8 @@ def warpctc(input, label, blank=0, norm_by_times=False): """ helper = LayerHelper('warpctc', **locals()) - loss_out = helper.create_tmp_variable(dtype=input.dtype) - grad_out = helper.create_tmp_variable(dtype=input.dtype) + loss_out = helper.create_variable_for_type_inference(dtype=input.dtype) + grad_out = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type='warpctc', inputs={'Logits': [input], @@ -3952,7 +4281,7 @@ def sequence_reshape(input, new_dim): x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10) """ helper = LayerHelper('sequence_reshape', **locals()) - out = helper.create_tmp_variable(helper.input_dtype()) + out = helper.create_variable_for_type_inference(helper.input_dtype()) helper.append_op( type='sequence_reshape', inputs={'X': [input]}, @@ -3971,7 +4300,8 @@ def nce(input, sample_weight=None, param_attr=None, bias_attr=None, - num_neg_samples=None): + num_neg_samples=None, + name=None): """ ${comment} @@ -3982,9 +4312,18 @@ def nce(input, sample_weight (Variable|None): A Variable of shape [batch_size, 1] storing a weight for each sample. The default weight for each sample is 1.0. - param_attr (ParamAttr|None): attributes for parameter - bias_attr (ParamAttr|None): attributes for bias + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of nce. If it is set to None or one attribute of ParamAttr, nce + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of nce. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, nce + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. num_neg_samples (int): ${num_neg_samples_comment} + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Variable: The output nce loss. @@ -4017,22 +4356,31 @@ def nce(input, """ helper = LayerHelper('nce', **locals()) assert isinstance(input, Variable) - dim = input.shape[1] assert isinstance(label, Variable) + + dim = input.shape[1] num_true_class = label.shape[1] w = helper.create_parameter( attr=helper.param_attr, shape=[num_total_classes, dim], is_bias=False, dtype=input.dtype) - b = helper.create_parameter( - attr=helper.bias_attr, - shape=[num_total_classes, 1], - is_bias=True, - dtype=input.dtype) - cost = helper.create_tmp_variable(dtype=input.dtype) - sample_logits = helper.create_tmp_variable(dtype=input.dtype) - sample_labels = helper.create_tmp_variable(dtype=label.dtype) + inputs = { + 'Input': input, + 'Label': label, + 'Weight': w, + 'SampleWeight': sample_weight if sample_weight is not None else [] + } + if helper.bias_attr: + b = helper.create_parameter( + attr=helper.bias_attr, + shape=[num_total_classes, 1], + is_bias=True, + dtype=input.dtype) + inputs['Bias'] = b + cost = helper.create_variable_for_type_inference(dtype=input.dtype) + sample_logits = helper.create_variable_for_type_inference(dtype=input.dtype) + sample_labels = helper.create_variable_for_type_inference(dtype=label.dtype) if num_neg_samples is None: num_neg_samples = 10 @@ -4046,13 +4394,7 @@ def nce(input, helper.append_op( type='nce', - inputs={ - 'Input': input, - 'Label': label, - 'Weight': w, - 'Bias': b, - 'SampleWeight': sample_weight if sample_weight is not None else [] - }, + inputs=inputs, outputs={ 'Cost': cost, 'SampleLogits': sample_logits, @@ -4062,7 +4404,12 @@ def nce(input, return cost / (num_neg_samples + 1) -def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): +def hsigmoid(input, + label, + num_classes, + param_attr=None, + bias_attr=None, + name=None): """ The hierarchical sigmoid operator is used to accelerate the training process of language model. This operator organizes the classes into a @@ -4083,11 +4430,17 @@ def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): label (Variable): The tensor variable contains labels of training data. It's a tensor with shape is :math:`[N \\times 1]`. num_classes: (int), The number of classes, must not be less than 2. - param_attr (ParamAttr|list of ParamAttr, default None): The parameter - attribute for learnable parameters/weights of this layer. - bias_attr (ParamAttr|list of ParamAttr, default None): The parameter - attribute for the bias of this layer. If it is set to False, no - bias will be applied. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of hsigmoid. If it is set to None or one attribute of ParamAttr, hsigmoid + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with Xavier. Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of hsigmoid. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, hsigmoid + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. + name (str|None): A name for this layer(optional). If set None, the layer + will be named automatically. Default: None. Returns: Out: (Tensor) The cost of hierarchical sigmoid operator. the shape is [N, 1] @@ -4103,8 +4456,8 @@ def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): helper = LayerHelper('hierarchical_sigmoid', **locals()) dtype = helper.input_dtype() - out = helper.create_tmp_variable(dtype) - pre_out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) + pre_out = helper.create_variable_for_type_inference(dtype) dim = input.shape[1] if num_classes < 2: raise ValueError("num_classes must not be less than 2.") @@ -4148,7 +4501,10 @@ def transpose(x, perm, name=None): Examples: .. code-block:: python - x = fluid.layers.data(name='x', shape=[5, 10, 15], dtype='float32') + # use append_batch_size=False to avoid prepending extra + # batch size in shape + x = fluid.layers.data(name='x', shape=[5, 10, 15], + dtype='float32', append_batch_size=False) x_transposed = layers.transpose(x, perm=[1, 0, 2]) """ @@ -4164,8 +4520,8 @@ def transpose(x, perm, name=None): (idx, perm[idx], len(x.shape))) helper = LayerHelper('transpose', **locals()) - out = helper.create_tmp_variable(x.dtype) - x_shape = helper.create_tmp_variable(x.dtype) + out = helper.create_variable_for_type_inference(x.dtype) + x_shape = helper.create_variable_for_type_inference(x.dtype) helper.append_op( type='transpose2', inputs={'X': [x]}, @@ -4307,7 +4663,7 @@ def im2sequence(input, inputs["Y"] = input_image_size attrs["out_stride"] = out_stride helper = LayerHelper('im2sequence', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) helper.append_op( type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs) return out @@ -4340,7 +4696,7 @@ def row_conv(input, future_context_size, param_attr=None, act=None): filter_shape = [future_context_size + 1, input.shape[1]] filter_param = helper.create_parameter( attr=helper.param_attr, shape=filter_shape, dtype=dtype) - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type='row_conv', inputs={'X': [input], @@ -4373,7 +4729,7 @@ def multiplex(inputs, index): raise ValueError("inputs should be a list object and contains at least " "2 elements.") - out = helper.create_tmp_variable(inputs[0].dtype) + out = helper.create_variable_for_type_inference(inputs[0].dtype) helper.append_op( type='multiplex', inputs={'X': inputs, @@ -4385,7 +4741,9 @@ def multiplex(inputs, index): def softmax_with_cross_entropy(logits, label, soft_label=False, - ignore_index=-100): + ignore_index=-100, + numeric_stable_mode=False, + return_softmax=False): """ **Softmax With Cross Entropy Operator.** @@ -4419,6 +4777,18 @@ def softmax_with_cross_entropy(logits, \\left(\\text{logit}_i - \\log\\left(\\sum_{i=0}^{K} \\exp(\\text{logit}_i)\\right)\\right), j = 1,...,K + 3) If numeric_stable_mode is True, softmax is calculated first by: + + .. math:: + + max_j = \\max_{i=0}^{K}{\\text{logit}_i} + + log\\_max\\_sum_j = \\log\\sum_{i=0}^{K}\\exp(logit_i - max_j) + + softmax_j = \\exp(logit_j - max_j - {log\\_max\\_sum}_j) + + and then cross entropy loss is calculated by softmax and label. + Args: logits (Variable): The unscaled log probabilities, which is a 2-D tensor with shape [N x K]. N is the batch_size, and K is the class number. @@ -4430,9 +4800,22 @@ def softmax_with_cross_entropy(logits, ignore_index (int): Specifies a target value that is ignored and does not contribute to the input gradient. Only valid if soft_label is set to False. Default: -100 + numeric_stable_mode (bool): A flag to indicate whether to use a more + numerically stable algorithm. Only valid + when soft_label is False and GPU is used. + When soft_label is True or CPU is used, + the algorithm is always numerically stable. + Note that the speed may be slower when use + stable algorithm. Default: False + return_softmax (bool): A flag indicating whether to return the softmax + along with the cross entropy loss. Default: False Returns: - Variable: The cross entropy loss is a 2-D tensor with shape [N x 1]. + Variable or Tuple of two Variables: Return the cross entropy loss if + `return_softmax` is False, otherwise the tuple + (loss, softmax), where the cross entropy loss is + a 2-D tensor with shape [N x 1], and softmax is a + 2-D tensor with shape [N x K]. Examples: .. code-block:: python @@ -4444,16 +4827,23 @@ def softmax_with_cross_entropy(logits, logits=fc, label=label) """ helper = LayerHelper('softmax_with_cross_entropy', **locals()) - softmax = helper.create_tmp_variable(dtype=logits.dtype) - loss = helper.create_tmp_variable(dtype=logits.dtype) + softmax = helper.create_variable_for_type_inference(dtype=logits.dtype) + loss = helper.create_variable_for_type_inference(dtype=logits.dtype) helper.append_op( type='softmax_with_cross_entropy', inputs={'Logits': logits, 'Label': label}, outputs={'Softmax': softmax, 'Loss': loss}, - attrs={'soft_label': soft_label, - 'ignore_index': ignore_index}) + attrs={ + 'soft_label': soft_label, + 'ignore_index': ignore_index, + 'numeric_stable_mode': numeric_stable_mode + }) + + if return_softmax: + return loss, softmax + return loss @@ -4495,8 +4885,8 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None): """ helper = LayerHelper('smooth_l1_loss', **locals()) - diff = helper.create_tmp_variable(dtype=x.dtype) - loss = helper.create_tmp_variable(dtype=x.dtype) + diff = helper.create_variable_for_type_inference(dtype=x.dtype) + loss = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='smooth_l1_loss', inputs={ @@ -4529,7 +4919,7 @@ def one_hot(input, depth): one_hot_label = layers.one_hot(input=label, depth=10) """ helper = LayerHelper("one_hot", **locals()) - one_hot_out = helper.create_tmp_variable(dtype='float32') + one_hot_out = helper.create_variable_for_type_inference(dtype='float32') helper.append_op( type="one_hot", inputs={'X': input}, @@ -4577,7 +4967,7 @@ def autoincreased_step_counter(counter_name=None, begin=1, step=1): return counter -def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): +def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None): """ Gives a new shape to the input Tensor without changing its data. @@ -4625,15 +5015,22 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): :attr:`shape` specifying shape. That is to say :attr:`actual_shape` has a higher priority than :attr:`shape`. - act (str): The non-linear activation to be applied to output variable. - inplace(bool): If this flag is set true, the output - shares data with input without copying, otherwise - a new output tensor is created - whose data is copied from input x. + act (str): The non-linear activation to be applied to the reshaped tensor + variable. + inplace(bool): Must use :attr:`False` if :attr:`x` is used in multiple + operators. If this flag is set :attr:`True`, reuse input + :attr:`x` to reshape, which will change the shape of + tensor variable :attr:`x` and might cause errors when + :attr:`x` is used in multiple operators. If :attr:`False`, + preserve the shape :attr:`x` and create a new output tensor + variable whose data is copied from input x but reshaped. name (str): The name of this layer. It is optional. Returns: - Variable: The output tensor. + Variable: The reshaped tensor variable if :attr:`act` is None. It is a \ + new tensor variable if :attr:`inplace` is :attr:`False`, \ + otherwise it is :attr:`x`. If :attr:`act` is not None, return \ + the activated tensor variable. Raises: TypeError: if actual_shape is neither Variable nor None. @@ -4644,7 +5041,7 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): data = fluid.layers.data( name='data', shape=[2, 4, 6], dtype='float32') reshaped = fluid.layers.reshape( - x=data, shape=[-1, 0, 3, 2], act='tanh', inplace=True) + x=data, shape=[-1, 0, 3, 2], inplace=True) """ if not (isinstance(shape, list) or isinstance(shape, tuple)): @@ -4671,8 +5068,9 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None): "except one unknown dimension.") helper = LayerHelper("reshape2", **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) - x_shape = helper.create_tmp_variable(dtype=x.dtype) + out = x if inplace else helper.create_variable_for_type_inference( + dtype=x.dtype) + x_shape = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type="reshape2", inputs=inputs, @@ -4721,8 +5119,8 @@ def squeeze(input, axes, name=None): y = layers.sequeeze(input=x, axes=[1]) """ helper = LayerHelper("squeeze", **locals()) - out = helper.create_tmp_variable(dtype=input.dtype) - x_shape = helper.create_tmp_variable(dtype=input.dtype) + out = helper.create_variable_for_type_inference(dtype=input.dtype) + x_shape = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type="squeeze2", inputs={"X": input}, @@ -4758,8 +5156,8 @@ def unsqueeze(input, axes, name=None): y = layers.unsequeeze(input=x, axes=[1]) """ helper = LayerHelper("unsqueeze", **locals()) - out = helper.create_tmp_variable(dtype=input.dtype) - x_shape = helper.create_tmp_variable(dtype=input.dtype) + out = helper.create_variable_for_type_inference(dtype=input.dtype) + x_shape = helper.create_variable_for_type_inference(dtype=input.dtype) helper.append_op( type="unsqueeze2", inputs={"X": input}, @@ -4849,7 +5247,7 @@ def lod_reset(x, y=None, target_lod=None): out = layers.lod_reset(x=x, y=y) """ helper = LayerHelper("lod_reset", **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) if y is not None: helper.append_op( type="lod_reset", inputs={'X': x, @@ -4918,8 +5316,9 @@ def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None): "dims of input must be 4(not %d), and it's order must be NCHW" % (dims)) - mid_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True) - lrn_out = helper.create_tmp_variable(dtype) + mid_out = helper.create_variable_for_type_inference( + dtype=dtype, stop_gradient=True) + lrn_out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="lrn", inputs={"X": input}, @@ -4984,7 +5383,7 @@ def pad(x, paddings, pad_value=0., name=None): """ helper = LayerHelper('pad', input=x, **locals()) dtype = helper.input_dtype() - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type='pad', inputs={'X': x}, @@ -5064,7 +5463,7 @@ def pad_constant_like(x, y, pad_value=0., name=None): """ helper = LayerHelper('pad_constant_like', input=x, **locals()) dtype = helper.input_dtype() - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type='pad_constant_like', inputs={'X': x, @@ -5129,7 +5528,7 @@ def label_smooth(label, raise ValueError("The value of epsilon must be between 0 and 1.") helper = LayerHelper("label_smooth", **locals()) label.stop_gradient = True - smooth_label = helper.create_tmp_variable(dtype) + smooth_label = helper.create_variable_for_type_inference(dtype) helper.append_op( type="label_smooth", inputs={"X": label, @@ -5161,8 +5560,8 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): """ helper = LayerHelper('roi_pool', **locals()) dtype = helper.input_dtype() - pool_out = helper.create_tmp_variable(dtype) - argmaxes = helper.create_tmp_variable(dtype='int32') + pool_out = helper.create_variable_for_type_inference(dtype) + argmaxes = helper.create_variable_for_type_inference(dtype='int32') helper.append_op( type="roi_pool", inputs={"X": input, @@ -5177,6 +5576,54 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0): return pool_out +@templatedoc() +def roi_align(input, + rois, + pooled_height=1, + pooled_width=1, + spatial_scale=1.0, + sampling_ratio=-1, + name=None): + """ + ${comment} + + Args: + input (Variable): ${x_comment} + rois (Variable): ROIs (Regions of Interest) to pool over. + pooled_height (integer): ${pooled_height_comment} Default: 1 + pooled_width (integer): ${pooled_width_comment} Default: 1 + spatial_scale (float): ${spatial_scale_comment} Default: 1.0 + sampling_ratio(intger): ${sampling_ratio_comment} Default: -1 + + Returns: + Variable: ${out_comment}. + Examples: + .. code-block:: python + + align_out = fluid.layers.roi_align(input=x, + rois=rois, + pooled_height=7, + pooled_width=7, + spatial_scale=0.5, + sampling_ratio=-1) + """ + helper = LayerHelper('roi_align', **locals()) + dtype = helper.input_dtype() + align_out = helper.create_variable_for_type_inference(dtype) + helper.append_op( + type="roi_align", + inputs={"X": input, + "ROIs": rois}, + outputs={"Out": align_out}, + attrs={ + "pooled_height": pooled_height, + "pooled_width": pooled_width, + "spatial_scale": spatial_scale, + "sampling_ratio": sampling_ratio + }) + return align_out + + def dice_loss(input, label, epsilon=0.00001): """ Dice loss for comparing the similarity of two batch of data, @@ -5222,7 +5669,8 @@ def image_resize(input, out_shape=None, scale=None, name=None, - resample='BILINEAR'): + resample='BILINEAR', + actual_shape=None): """ **Resize a Batch of Images** @@ -5232,6 +5680,7 @@ def image_resize(input, Supporting resample methods: 'BILINEAR' : Bilinear interpolation + 'NEAREST' : Nearest neighbor interpolation Args: input (Variable): The input tensor of image resize layer, @@ -5246,25 +5695,51 @@ def image_resize(input, Default: None name(str|None): A name for this layer(optional). If set None, the layer will be named automatically. - resample(str): The resample method. It can only be 'BILINEAR' currently. + resample(str): The resample method. It supports 'BILINEAR' and 'NEAREST' + currently. Default: 'BILINEAR' + actual_shape(Variable): An optional input to specify output shape + dynamically. If provided, image resize + according to this given shape rather than + :attr:`out_shape` and :attr:`scale` specifying + shape. That is to say actual_shape has the + highest priority. It is recommended to use + actual_shape instead of :attr:`out_shape` if you + want to specify output shape dynamically. When + using actual_shape to specify output shape, one of + :attr:`out_shape` and :attr:`scale` should also be + set, otherwise errors would be occured in graph + constructing stage. + Default: None Returns: Variable: The output is a 4-D tensor of the shape (num_batches, channls, out_h, out_w). + Raises: + TypeError: out_shape should be a list or tuple or Variable. + TypeError: actual_shape should either be Variable or None. + ValueError: The 'resample' of image_resize can only be 'BILINEAR' + or 'NEAREST' currently. + ValueError: One of out_shape and scale must not be None. + ValueError: out_shape length should be 2. + Examples: .. code-block:: python out = fluid.layers.image_resize(input, out_shape=[12, 12]) """ - resample_methods = {'BILINEAR': 'bilinear_interp'} + resample_methods = { + 'BILINEAR': 'bilinear', + 'NEAREST': 'nearest', + } if resample not in resample_methods: raise ValueError( - "The 'resample' of image_resize can only be 'BILINEAR' currently.") + "The 'resample' of image_resize can only be 'BILINEAR' or 'NEAREST' currently." + ) if out_shape is None and scale is None: - raise ValueError("One of out_shape and scale must not be None") - helper = LayerHelper('bilinear_interp', **locals()) + raise ValueError("One of out_shape and scale must not be None.") + helper = LayerHelper('interpolate', **locals()) dtype = helper.input_dtype() def _is_list_or_turple_(data): @@ -5274,33 +5749,60 @@ def image_resize(input, out_w = 0 inputs = {"X": input} if out_shape is not None: - if not (_is_list_or_turple_(out_shape) and - len(out_shape) == 2) and not isinstance(out_shape, Variable): - raise ValueError('out_shape should be a list or tuple or variable') - if _is_list_or_turple_(out_shape): - out_shape = list(map(int, out_shape)) - out_h = out_shape[0] - out_w = out_shape[1] - else: + if isinstance(out_shape, Variable): + warnings.warn("out_shape as Variable type is deprecated, \ + it is recommended to use actual_shape instead of \ + out_shape to specify output shape dynamically.") inputs['OutSize'] = out_shape + elif not (_is_list_or_turple_(out_shape)): + raise TypeError("out_shape should be a list or tuple or Variable.") + elif len(out_shape) != 2: + raise ValueError("out_shape length should be 2.") + + out_shape = list(map(int, out_shape)) + out_h = out_shape[0] + out_w = out_shape[1] else: out_h = int(input.shape[2] * scale) out_w = int(input.shape[3] * scale) - out = helper.create_tmp_variable(dtype) + if isinstance(actual_shape, Variable): + inputs["OutSize"] = actual_shape + elif actual_shape is not None: + raise TypeError("actual_shape should either be Variable or None.") + + out = helper.create_variable_for_type_inference(dtype) helper.append_op( - type=resample_methods[resample], + type='interpolate', inputs=inputs, outputs={"Out": out}, - attrs={"out_h": out_h, - "out_w": out_w}) + attrs={ + "out_h": out_h, + "out_w": out_w, + "interp_method": resample_methods[resample] + }) return out -@templatedoc(op_type="bilinear_interp") -def resize_bilinear(input, out_shape=None, scale=None, name=None): +@templatedoc(op_type="interpolate") +def resize_bilinear(input, + out_shape=None, + scale=None, + name=None, + actual_shape=None): """ - ${comment} + Resize input by performing bilinear interpolation based on given + output shape which specified by actual_shape, out_shape and scale + in priority order. + + Bilinear interpolation is an extension of linear interpolation for + interpolating functions of two variables (e.g. H-direction and + W-direction in this op) on a rectilinear 2D grid. The key idea is + to perform linear interpolation first in one direction, and then + again in the other direction. + + For details of bilinear interpolation, please refer to Wikipedia: + https://en.wikipedia.org/wiki/Bilinear_interpolation Args: input(${x_type}): ${x_comment}. @@ -5312,12 +5814,71 @@ def resize_bilinear(input, out_shape=None, scale=None, name=None): a higher priority than scale. Default: None. name(str|None): The output variable name. + actual_shape(Variable): An optional input to specify output shape + dynamically. If provided, image resize + according to this given shape rather than + :attr:`out_shape` and :attr:`scale` specifying + shape. That is to say actual_shape has the + highest priority. It is recommended to use + actual_shape instead of :attr:`out_shape` if you + want to specify output shape dynamically. When + using actual_shape to specify output shape, one of + :attr:`out_shape` and :attr:`scale` should also be + set, otherwise errors would be occured in graph + constructing stage. + Default: None Returns: ${out_comment}. """ - return image_resize(input, out_shape, scale, name, 'BILINEAR') + return image_resize(input, out_shape, scale, name, 'BILINEAR', actual_shape) + + +@templatedoc(op_type="interpolate") +def resize_nearest(input, + out_shape=None, + scale=None, + name=None, + actual_shape=None): + """ + Resize input by performing nearest neighbor interpolation in both the + 3rd dimention(in height direction) and the 4th dimention(in width + direction) based on given output shape which specified by actual_shape, + out_shape and scale in priority order. + + For details of nearest neighbor interpolation, please refer to Wikipedia: + https://en.wikipedia.org/wiki/Nearest-neighbor_interpolation + + Args: + input(${x_type}): ${x_comment}. + + out_shape(${out_size_type}): ${out_size_comment}. + + scale(float|None): The multiplier for the input height or width. At + least one of out_shape or scale must be set. And out_shape has + a higher priority than scale. Default: None. + + name(str|None): The output variable name. + actual_shape(Variable): An optional input to specify output shape + dynamically. If provided, image resize + according to this given shape rather than + :attr:`out_shape` and :attr:`scale` specifying + shape. That is to say actual_shape has the + highest priority. It is recommended to use + actual_shape instead of :attr:`out_shape` if you + want to specify output shape dynamically. When + using actual_shape to specify output shape, one of + :attr:`out_shape` and :attr:`scale` should also be + set, otherwise errors would be occured in graph + constructing stage. + Default: None + + Returns: + ${out_comment}. + """ + + return image_resize(input, out_shape, scale, name, 'NEAREST', actual_shape) def image_resize_short(input, out_short_len, resample='BILINEAR'): @@ -5396,7 +5957,7 @@ def gather(input, index): """ helper = LayerHelper('gather', **locals()) dtype = helper.input_dtype() - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="gather", inputs={"X": input, @@ -5436,7 +5997,7 @@ def scatter(input, index, updates, name=None): """ helper = LayerHelper('scatter', **locals()) dtype = helper.input_dtype() - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="scatter", inputs={"X": input, @@ -5496,7 +6057,7 @@ def sequence_scatter(input, index, updates, name=None): """ helper = LayerHelper('sequence_scatter', **locals()) dtype = helper.input_dtype() - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="sequence_scatter", inputs={"X": input, @@ -5526,7 +6087,7 @@ def random_crop(x, shape, seed=None): """ helper = LayerHelper("random_crop", **locals()) dtype = x.dtype - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) if seed is None: seed = np.random.randint(-65536, 65536) op_attrs = {"shape": shape} @@ -5572,7 +6133,7 @@ def log(x, name=None): """ helper = LayerHelper('log', **locals()) dtype = helper.input_dtype(input_param_name='x') - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out}) return out @@ -5603,7 +6164,7 @@ def relu(x, name=None): """ helper = LayerHelper('relu', **locals()) dtype = helper.input_dtype(input_param_name='x') - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out}) return out @@ -5642,9 +6203,9 @@ def mean_iou(input, label, num_classes): """ helper = LayerHelper('mean_iou', **locals()) dtype = helper.input_dtype() - out_mean_iou = helper.create_tmp_variable(dtype='float32') - out_wrong = helper.create_tmp_variable(dtype='int32') - out_correct = helper.create_tmp_variable(dtype='int32') + out_mean_iou = helper.create_variable_for_type_inference(dtype='float32') + out_wrong = helper.create_variable_for_type_inference(dtype='int32') + out_correct = helper.create_variable_for_type_inference(dtype='int32') helper.append_op( type="mean_iou", inputs={"Predictions": input, @@ -5736,7 +6297,7 @@ def crop(x, shape=None, offsets=None, name=None): if offsets is None: offsets = [0] * len(x.shape) - out = helper.create_tmp_variable(x.dtype) + out = helper.create_variable_for_type_inference(x.dtype) ipts = {'X': x} attrs = {} if isinstance(shape, Variable): @@ -5756,6 +6317,124 @@ def crop(x, shape=None, offsets=None, name=None): return out +def affine_grid(theta, out_shape, name=None): + """ + It generates a grid of (x,y) coordinates using the parameters of + the affine transformation that correspond to a set of points where + the input feature map should be sampled to produce the transformed + output feature map. + + .. code-block:: text + + * Case 1: + + Given: + + theta = [[[x_11, x_12, x_13] + [x_14, x_15, x_16]] + [[x_21, x_22, x_23] + [x_24, x_25, x_26]]] + + out_shape = [2, 3, 5, 5] + + Step 1: + + Generate normalized coordinates according to out_shape. + The values of the normalized coordinates are in the interval between -1 and 1. + The shape of the normalized coordinates is [2, H, W] as below: + + C = [[[-1. -1. -1. -1. -1. ] + [-0.5 -0.5 -0.5 -0.5 -0.5] + [ 0. 0. 0. 0. 0. ] + [ 0.5 0.5 0.5 0.5 0.5] + [ 1. 1. 1. 1. 1. ]] + [[-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ] + [-1. -0.5 0. 0.5 1. ]]] + C[0] is the coordinates in height axis and C[1] is the coordinates in width axis. + + Step2: + + Tanspose and reshape C to shape [H * W, 2] and append ones to last dimension. The we get: + C_ = [[-1. -1. 1. ] + [-0.5 -1. 1. ] + [ 0. -1. 1. ] + [ 0.5 -1. 1. ] + [ 1. -1. 1. ] + [-1. -0.5 1. ] + [-0.5 -0.5 1. ] + [ 0. -0.5 1. ] + [ 0.5 -0.5 1. ] + [ 1. -0.5 1. ] + [-1. 0. 1. ] + [-0.5 0. 1. ] + [ 0. 0. 1. ] + [ 0.5 0. 1. ] + [ 1. 0. 1. ] + [-1. 0.5 1. ] + [-0.5 0.5 1. ] + [ 0. 0.5 1. ] + [ 0.5 0.5 1. ] + [ 1. 0.5 1. ] + [-1. 1. 1. ] + [-0.5 1. 1. ] + [ 0. 1. 1. ] + [ 0.5 1. 1. ] + [ 1. 1. 1. ]] + Step3: + Compute output by equation $$Output[i] = C_ * Theta[i]^T$$ + + Args: + theta (Variable): A batch of affine transform parameters with shape [N, 2, 3]. + out_shape (Variable | list | tuple): The shape of target output with format [N, C, H, W]. + out_shape can be a Variable or a list or tuple. + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: The output with shape [N, H, W, 2]. + + Raises: + ValueError: If the type of arguments is not supported. + + Examples: + + .. code-block:: python + theta = fluid.layers.data(name="x", shape=[2, 3], dtype="float32") + out_shape = fluid.layers.data(name="y", shape=[-1], dtype="float32") + data = fluid.layers.affine_grid(theta, out_shape) + + # or + data = fluid.layers.affine_grid(theta, [5, 3, 28, 28]) + + """ + helper = LayerHelper('affine_grid') + + if not (isinstance(out_shape, list) or isinstance(out_shape, tuple) or \ + isinstance(out_shape, Variable)): + raise ValueError("The out_shape should be a list, tuple or Variable.") + + if not isinstance(theta, Variable): + raise ValueError("The theta should be a Variable.") + + out = helper.create_variable_for_type_inference(theta.dtype) + ipts = {'Theta': theta} + attrs = {} + if isinstance(out_shape, Variable): + ipts['OutputShape'] = out_shape + else: + attrs['output_shape'] = out_shape + + helper.append_op( + type='affine_grid', + inputs=ipts, + outputs={'Output': out}, + attrs=None if len(attrs) == 0 else attrs) + return out + + def rank_loss(label, left, right, name=None): """ **Rank loss layer for RankNet** @@ -5816,7 +6495,7 @@ def rank_loss(label, left, right, name=None): if not (isinstance(right, Variable)): raise ValueError("The right should be a Variable") - out = helper.create_tmp_variable("float32") + out = helper.create_variable_for_type_inference("float32") helper.append_op( type='rank_loss', @@ -5827,6 +6506,54 @@ def rank_loss(label, left, right, name=None): return out +def margin_rank_loss(label, left, right, margin=0.1, name=None): + """ + Margin Ranking Loss Layer for ranking problem, + which compares left score and right score passed in. + The ranking loss can be defined as following equation: + + .. math:: + + rank\_loss &= max(0, -label * (left - right) + margin) + + Args: + label (Variable): Indicates whether the left is ranked higher than the right or not. + left (Variable): Ranking score for left. + right (Variable): Ranking score for right. + margin (float): Indicates the given margin. + name (str|None): A name for this layer (optional). If set None, the layer + will be named automatically. + Returns: + Variable: The ranking loss. + Raises: + ValueError: Any of label, left, and right is not a Variable. + Examples: + .. code-block:: python + label = fluid.layers.data(name="label", shape=[4, 1], dtype="float32") + left = fluid.layers.data(name="left", shape=[4, 1], dtype="float32") + right = fluid.layers.data(name="right", shape=[4, 1], dtype="float32") + out = fluid.layers.margin_rank_loss(label, left, right) + """ + helper = LayerHelper('margin_rank_loss', **locals()) + if not isinstance(label, Variable): + raise ValueError("The label should be a Variable.") + if not isinstance(left, Variable): + raise ValueError("The left should be a Variable.") + if not isinstance(right, Variable): + raise ValueError("The right should be a Variable.") + out = helper.create_variable_for_type_inference(left.dtype) + act = helper.create_variable_for_type_inference(left.dtype) + helper.append_op( + type='margin_rank_loss', + inputs={"Label": label, + "X1": left, + "X2": right}, + outputs={'Out': out, + 'Activated': act}, + attrs={'margin': margin}) + return out + + def pad2d(input, paddings=[0, 0, 0, 0], mode='constant', @@ -5900,7 +6627,7 @@ def pad2d(input, helper = LayerHelper('pad2d', **locals()) dtype = helper.input_dtype(input_param_name='input') - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type='pad2d', inputs={'X': input}, @@ -5929,7 +6656,7 @@ def elu(x, alpha=1.0, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('elu', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='elu', inputs={'X': x}, @@ -5952,7 +6679,7 @@ def relu6(x, threshold=6.0, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('relu6', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='relu6', inputs={'X': x}, @@ -5975,7 +6702,7 @@ def pow(x, factor=1.0, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('pow', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='pow', inputs={'X': x}, @@ -5999,7 +6726,7 @@ def stanh(x, scale_a=2.0 / 3.0, scale_b=1.7159, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('stanh', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='stanh', inputs={'X': x}, @@ -6024,7 +6751,7 @@ def hard_sigmoid(x, slope=0.2, offset=0.5, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('hard_sigmoid', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='hard_sigmoid', inputs={'X': x}, @@ -6048,7 +6775,7 @@ def swish(x, beta=1.0, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('swish', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='swish', inputs={'X': x}, @@ -6100,7 +6827,7 @@ def prelu(x, mode, param_attr=None, name=None): dtype='float32', is_bias=False, default_initializer=Constant(1.0)) - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type="prelu", inputs={"X": x, @@ -6124,7 +6851,7 @@ def brelu(x, t_min=0.0, t_max=24.0, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('brelu', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='brelu', inputs={'X': x}, @@ -6147,7 +6874,7 @@ def leaky_relu(x, alpha=0.02, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('leaky_relu', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='leaky_relu', inputs={'X': x}, @@ -6169,7 +6896,7 @@ def soft_relu(x, threshold=40.0, name=None): output(${out_type}): ${out_comment} """ helper = LayerHelper('soft_relu', **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='soft_relu', inputs={'X': x}, @@ -6236,8 +6963,8 @@ def flatten(x, axis=1, name=None): if not (isinstance(axis, int)) or axis > len(x.shape) or axis < 0: raise ValueError("The axis should be a int, and in range [0, rank(x)]") - out = helper.create_tmp_variable(x.dtype) - x_shape = helper.create_tmp_variable(x.dtype) + out = helper.create_variable_for_type_inference(x.dtype) + x_shape = helper.create_variable_for_type_inference(x.dtype) helper.append_op( type='flatten2', inputs={"X": x}, @@ -6283,13 +7010,15 @@ def sequence_enumerate(input, win_size, pad_value=0, name=None): out = fluid.layers.sequence_enumerate(input=x, win_size=3, pad_value=0) """ helper = LayerHelper('sequence_enumerate', **locals()) - out = helper.create_tmp_variable(helper.input_dtype(), stop_gradient=True) + out = helper.create_variable_for_type_inference( + helper.input_dtype(), stop_gradient=True) helper.append_op( type='sequence_enumerate', inputs={'X': input}, outputs={'Out': out}, attrs={'win_size': win_size, 'pad_value': pad_value}) + return out def sequence_mask(x, maxlen=None, dtype='int64', name=None): @@ -6322,9 +7051,9 @@ def sequence_mask(x, maxlen=None, dtype='int64', name=None): helper = LayerHelper('sequence_mask', **locals()) if name is None: - out = helper.create_tmp_variable(dtype=dtype) + out = helper.create_variable_for_type_inference(dtype=dtype) else: - out = helper.create_tmp_variable(dtype=dtype, name=name) + out = helper.create_variable_for_type_inference(dtype=dtype, name=name) helper.append_op( type='sequence_mask', @@ -6367,7 +7096,7 @@ def stack(x, axis=0): if not isinstance(x, list) and not isinstance(x, tuple): x = [x] - out = helper.create_tmp_variable(x[0].dtype) + out = helper.create_variable_for_type_inference(x[0].dtype) helper.append_op( type='stack', inputs={'X': x}, outputs={'Y': out}, attrs={'axis': axis}) @@ -6405,7 +7134,7 @@ def unstack(x, axis=0, num=None): outs = [] for _ in num: - outs.append(helper.create_tmp_variable(x.dtype)) + outs.append(helper.create_variable_for_type_inference(x.dtype)) helper.append_op( type='unstack', @@ -6457,7 +7186,7 @@ def expand(x, expand_times, name=None): """ helper = LayerHelper('expand', input=x, **locals()) dtype = helper.input_dtype(input_param_name='x') - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type='expand', inputs={'X': x}, @@ -6496,7 +7225,7 @@ def uniform_random_batch_size_like(input, """ helper = LayerHelper('uniform_random_batch_size_like', **locals()) - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) c_dtype = convert_np_dtype_to_dtype_(dtype) helper.append_op( type='uniform_random_batch_size_like', @@ -6533,7 +7262,7 @@ def gaussian_random(shape, mean=0.0, std=1.0, seed=0, dtype='float32'): """ helper = LayerHelper('gaussian_random', **locals()) - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) c_dtype = convert_np_dtype_to_dtype_(dtype) helper.append_op( type='gaussian_random', @@ -6568,7 +7297,7 @@ def sampling_id(x, min=0.0, max=1.0, seed=0, dtype='float32'): """ helper = LayerHelper('sampling_id', **locals()) - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) helper.append_op( type='sampling_id', inputs={'X': x}, @@ -6607,7 +7336,7 @@ def gaussian_random_batch_size_like(input, """ helper = LayerHelper('gaussian_random_batch_size_like', **locals()) - out = helper.create_tmp_variable(dtype) + out = helper.create_variable_for_type_inference(dtype) c_dtype = convert_np_dtype_to_dtype_(dtype) helper.append_op( type='gaussian_random_batch_size_like', @@ -6639,7 +7368,8 @@ def sum(x): """ helper = LayerHelper('sum', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype('x')) + out = helper.create_variable_for_type_inference( + dtype=helper.input_dtype('x')) helper.append_op( type='sum', inputs={'X': x}, @@ -6666,7 +7396,8 @@ def slice(input, axes, starts, ends): """ helper = LayerHelper('slice', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype('input')) + out = helper.create_variable_for_type_inference( + dtype=helper.input_dtype('input')) helper.append_op( type='slice', inputs={'Input': input}, @@ -6692,7 +7423,8 @@ def shape(input): """ helper = LayerHelper('shape', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype('input')) + out = helper.create_variable_for_type_inference( + dtype=helper.input_dtype('input')) helper.append_op( type='shape', inputs={'Input': input}, outputs={'Out': out}) @@ -6709,7 +7441,7 @@ def _elementwise_op(helper): use_mkldnn = helper.kwargs.get('use_mkldnn', False) name = helper.kwargs.get('name', None) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = helper.create_variable( name=name, dtype=x.dtype, persistable=False) @@ -6743,7 +7475,7 @@ def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None): helper = LayerHelper('scale', **locals()) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = helper.create_variable( name=name, dtype=x.dtype, persistable=False) @@ -6809,7 +7541,7 @@ def _logical_op(op_name, x, y, out=None, name=None, binary_op=True): if out is None: if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = helper.create_variable( name=name, dtype=x.dtype, persistable=False) @@ -6917,10 +7649,10 @@ def clip(x, min, max, name=None): helper = LayerHelper("clip", **locals()) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) - else: - out = helper.create_variable( - name=name, dtype=x.dtype, persistable=False) + name = unique_name.generate(".".join([helper.name, 'tmp'])) + + out = helper.create_variable( + type=x.type, name=name, dtype=x.dtype, persistable=False) helper.append_op( type="clip", @@ -6949,10 +7681,10 @@ def clip_by_norm(x, max_norm, name=None): helper = LayerHelper("clip_by_norm", **locals()) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) - else: - out = helper.create_variable( - name=name, dtype=x.dtype, persistable=False) + name = unique_name.generate(".".join([helper.name, 'tmp'])) + + out = helper.create_variable( + type=x.type, name=name, dtype=x.dtype, persistable=False) helper.append_op( type="clip_by_norm", @@ -6979,7 +7711,7 @@ def mean(x, name=None): helper = LayerHelper("mean", **locals()) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = helper.create_variable( name=name, dtype=x.dtype, persistable=False) @@ -7009,7 +7741,7 @@ def mul(x, y, x_num_col_dims=1, y_num_col_dims=1, name=None): helper = LayerHelper("mul", **locals()) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = helper.create_variable( name=name, dtype=x.dtype, persistable=False) @@ -7043,7 +7775,7 @@ def sigmoid_cross_entropy_with_logits(x, label, name=None): helper = LayerHelper("sigmoid_cross_entropy_with_logits", **locals()) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = helper.create_variable( name=name, dtype=x.dtype, persistable=False) @@ -7073,7 +7805,7 @@ def maxout(x, groups, name=None): helper = LayerHelper("maxout", **locals()) if name is None: - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) else: out = helper.create_variable( name=name, dtype=x.dtype, persistable=False) @@ -7084,3 +7816,557 @@ def maxout(x, groups, name=None): attrs={"groups": groups}, outputs={"Out": out}) return out + + +def space_to_depth(x, blocksize, name=None): + """ + Gives a blocksize to space_to_depth the input LoDtensor with Layout: [batch, channel, height, width] + + This op rearranges blocks of spatial data, into depth. More specifically, this op outputs a copy of the + input LoDtensor where values from the height and width dimensions are moved to the channel dimension. + The attr blocksize indicates the input block size. + + space_to_depth will reorgnize the elements of input with shape[batch, channel, height, width] according + to blocksize to construct output with shape [batch, channel * blocksize * blocksize, height/blocksize, width/blocksize]: + + space_to_depth is used to This operation is useful for resizing the activations between convolutions + (but keeping all data) + + - Non-overlapping blocks of size block_size x block size are rearranged into depth at each location. + - The depth of the output tensor is block_size * block_size * input channel + - The Y, X coordinates within each block of the input become the high order component of the output channel index + - channel should be divisible by square of blocksize + - height, width should be divsible by blocksize + + + Args: + x(variable): The input LoDtensor. + blocksize(variable): The blocksize to select the element on each feature map should be > 2 + + Returns: + Variable: The output LoDtensor. + + Raises: + TypeError: blocksize type must be a long. + + Examples: + .. code-block:: python + + data = fluid.layers.data( + name='data', shape=[1, 4, 2, 2], dtype='float32') + space_to_depthed = fluid.layers.space_to_depth( + x=data, blocksize=2) + """ + + helper = LayerHelper("space_to_depth", **locals()) + + if not (isinstance(blocksize, int)): + raise ValueError("blocksize must be a python Int") + + if name is None: + out = helper.create_variable_for_type_inference( + dtype=x.dtype) #fix create + else: + out = helper.create_variable( + name=name, dtype=x.dtype, persistable=False) + + helper.append_op( + type="space_to_depth", + inputs={"X": x}, + attrs={"blocksize": blocksize}, + outputs={"Out": out}) + return out + + +@templatedoc() +def sequence_reverse(x, name=None): + """ + ${comment} + + Args: + x(${x_type}): ${x_comment} + name(basestring|None): Name of the output. + + Returns: + out(${y_type}): ${y_comment} + """ + helper = LayerHelper("sequence_reverse", **locals()) + if name is None: + out = helper.create_variable_for_type_inference(dtype=x.dtype) + else: + out = helper.create_variable( + name=name, dtype=x.dtype, persistable=False) + + helper.append_op( + type="sequence_reverse", + inputs={"X": x}, + outputs={"Y": out}, + attrs=dict()) + return out + + +def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None): + """ + Applies a separate affine transformation to each channel of the input. + Useful for replacing spatial batch norm with its equivalent fixed + transformation. The input also can be 2D tensor and applies a affine + transformation in second dimension. + + Args: + x (Variable): Feature map input can be a 4D tensor with order NCHW + or NHWC. It also can be a 2D tensor and the affine transformation + is applied in the second dimension. + scale (Variable): 1D input of shape (C), the c-th element is the scale + factor of the affine transformation for the c-th channel of + the input. + bias (Variable): 1D input of shape (C), the c-th element is the bias + of the affine transformation for the c-th channel of the input. + data_layout (string, default NCHW): NCHW or NHWC. If input is 2D + tensor, you can ignore data_layout. + name (str, default None): The name of this layer. + + Returns: + out (Variable): A tensor of the same shape and data layout with x. + """ + helper = LayerHelper("affine_channel", **locals()) + + if name is None: + out = helper.create_variable_for_type_inference(dtype=x.dtype) + else: + out = helper.create_variable( + name=name, dtype=x.dtype, persistable=False) + + helper.append_op( + type="affine_channel", + inputs={"X": x, + 'Scale': scale, + 'Bias': bias}, + attrs={"data_layout": data_layout}, + outputs={"Out": out}) + return out + + +def similarity_focus(input, axis, indexes, name=None): + """ + SimilarityFocus Operator + + Generate a similarity focus mask with the same shape of input using the following method: + 1. Extract the 3-D tensor(here the first dimension is BatchSize) corresponding + to the axis according to the indexes. For example, if axis=1 and indexes=[a], + it will get the matrix T=X[:, a, :, :]. In this case, if the shape of input X + is (BatchSize, A, B, C), the shape of tensor T is (BatchSize, B, C). + 2. For each index, find the largest numbers in the tensor T, so that the same + row and same column has at most one number(what it means is that if the + largest number has been found in the i-th row and the j-th column, then + the numbers in the i-th row or j-th column will be skipped. And then the + next largest number will be selected from the remaining numbers. Obviously + there will be min(B, C) numbers), and mark the corresponding position of the + 3-D similarity focus mask as 1, otherwise as 0. Do elementwise-or for + each index. + 3. Broadcast the 3-D similarity focus mask to the same shape of input X. + + Refer to `Similarity Focus Layer `_ + + .. code-block:: text + + * Example : + + Given a 4-D tensor x with the shape (BatchSize, C, A, B), where C is + the number of channels and the shape of feature map is (A, B): + x.shape = (2, 3, 2, 2) + x.data = [[[[0.8, 0.1], + [0.4, 0.5]], + + [[0.9, 0.7], + [0.9, 0.9]], + + [[0.8, 0.9], + [0.1, 0.2]]], + + + [[[0.2, 0.5], + [0.3, 0.4]], + + [[0.9, 0.7], + [0.8, 0.4]], + + [[0.0, 0.2], + [0.4, 0.7]]]] + + Given axis: 1 (the axis of the channel) + Given indexes: [0] + + then we get a 4-D tensor out with the same shape of input x: + out.shape = (2, 3, 2, 2) + out.data = [[[[1.0, 0.0], + [0.0, 1.0]], + + [[1.0, 0.0], + [0.0, 1.0]], + + [[1.0, 0.0], + [0.0, 1.0]]], + + [[[0.0, 1.0], + [1.0, 0.0]], + + [[0.0, 1.0], + [1.0, 0.0]], + + [[0.0, 1.0], + [1.0, 0.0]]]] + + Args: + input(Variable): The input tensor variable(default float). It should + be a 4-D tensor with shape [BatchSize, A, B, C]. + axis(int): Indicating the dimension to be selected. It can only be + 1, 2 or 3. + indexes(list): Indicating the indexes of the selected dimension. + + Returns: + Variable: A tensor variable with the same shape and same type + as the input. + + Examples: + .. code-block:: python + data = fluid.layers.data( + name='data', shape=[2, 3, 2, 2], dtype='float32') + x = fluid.layers.layer_norm(input=data, axis=1, indexes=[0]) + """ + helper = LayerHelper('similarity_focus', **locals()) + # check attrs + if isinstance(axis, int) is False: + raise TypeError("axis must be int type.") + if isinstance(indexes, list) is False: + raise TypeError("indexes must be list type.") + if axis != 1 and axis != 2 and axis != 3: + raise ValueError("axis must be 1, 2 or 3.") + if len(indexes) == 0: + raise ValueError("indexes can not be empty.") + + if name is None: + out = helper.create_variable_for_type_inference(dtype=input.dtype) + else: + out = helper.create_variable( + name=name, dtype=input.dtype, persistable=False) + helper.append_op( + type='similarity_focus', + inputs={'X': input}, + outputs={'Out': out}, + attrs={"axis": axis, + "indexes": indexes}) + return out + + +def hash(input, hash_size, num_hash=1, name=None): + """ + Hash the input to an integer whose value is less than the given hash size. + + The hash algorithm we used was xxHash - Extremely fast hash algorithm + (https://github.com/Cyan4973/xxHash/tree/v0.6.5) + + A simple example as below: + + .. code-block:: text + + Given: + + # shape [2, 2] + input.data = [ + [[1], [2]], + [[3], [4]], + ] + + input.lod = [[0, 2]] + + hash_size = 10000 + + num_hash = 4 + + Then: + + Hash op will take all number in input's 2nd dimension as hash algorithm's + input for each time. Each input will be hashed for 4 times, and get an + array whose length is 4. Each value in the array ranges from 0 to 9999. + + # shape [2, 4] + output.data = [ + [[9662], [9217], [1129], [8487]], + [[8310], [1327], [1654], [4567]], + ] + + output.lod = [[0, 2]] + + Args: + input (Variable): The input variable which is a one-hot word. The + dimensions of the input variable must be 2. + hash_size (int): The space size for hash algorithm. The output value + will keep in the range:math:`[0, hash_size - 1]`. + num_hash (int): The times of hash, default 1. + name (str, default None): The name of this layer. + + Returns: + Variable: The hash result variable which is a LoDTensor. + + Examples: + .. code-block:: python + word_dict = paddle.dataset.imdb.word_dict() + x = fluid.layers.data(shape[1], dtype='int32', lod_level=1) + out = fluid.layers.hash(input=x, num_hash=4, hash_size=1000) + """ + helper = LayerHelper('hash', **locals()) + out = helper.create_variable_for_type_inference( + helper.input_dtype(), stop_gradient=True) + helper.append_op( + type='hash', + inputs={'X': input}, + outputs={'Out': out}, + attrs={'num_hash': num_hash, + 'mod_by': hash_size}) + return out + + +@templatedoc() +def grid_sampler(x, grid, name=None): + """ + This operation samples input X by using bilinear interpolation based on + flow field grid, which is usually gennerated by affine_grid. The grid of + shape [N, H, W, 2] is the concatenation of (grid_x, grid_y) coordinates + with shape [N, H, W] each, where grid_x is indexing the 4th dimension + (in width dimension) of input data x and grid_y is indexng the 3rd + dimention (in height dimension), finally results is the bilinear + interpolation value of 4 nearest corner points. + + Step 1: + Get (x, y) grid coordinates and scale to [0, H-1/W-1]. + + grid_x = 0.5 * (grid[:, :, :, 0] + 1) * (W - 1) + grid_y = 0.5 * (grid[:, :, :, 1] + 1) * (H - 1) + + Step 2: + Indices input data X with grid (x, y) in each [H, W] area, and bilinear + interpolate point value by 4 nearest points. + + wn ------- y_n ------- en + | | | + | d_n | + | | | + x_w --d_w-- grid--d_e-- x_e + | | | + | d_s | + | | | + ws ------- y_s ------- wn + + x_w = floor(x) // west side x coord + x_e = x_w + 1 // east side x coord + y_n = floor(y) // north side y coord + y_s = y_s + 1 // south side y coord + + d_w = grid_x - x_w // distance to west side + d_e = x_e - grid_x // distance to east side + d_n = grid_y - y_n // distance to north side + d_s = y_s - grid_y // distance to south side + + wn = X[:, :, y_n, x_w] // north-west point value + en = X[:, :, y_n, x_e] // north-east point value + ws = X[:, :, y_s, x_w] // south-east point value + es = X[:, :, y_s, x_w] // north-east point value + + output = wn * d_e * d_s + en * d_w * d_s + + ws * d_e * d_n + es * d_w * d_n + + Args: + x(Variable): Input data of shape [N, C, H, W]. + grid(Variable): Input grid tensor of shape [N, H, W, 2]. + name (str, default None): The name of this layer. + + Returns: + out(Variable): Output of shape [N, C, H, W] data samples input X + using bilnear interpolation based on input grid. + + Exmples: + .. code-block:: python + + x = fluid.layers.data(name='x', shape=[3, 10, 32, 32], dtype='float32') + theta = fluid.layers.data(name='theta', shape=[3, 2, 3], dtype='float32') + grid = fluid.layers.affine_grid(input=theta, size=[3, 10, 32, 32]}) + out = fluid.layers.grid_sampler(x=x, grid=grid) + """ + helper = LayerHelper("grid_sampler", **locals()) + + if not isinstance(x, Variable): + return ValueError("The x should be a Variable") + + if not isinstance(grid, Variable): + return ValueError("The grid should be a Variable") + + out = helper.create_variable_for_type_inference(x.dtype) + ipts = {'X': x, 'Grid': grid} + + helper.append_op(type='grid_sampler', inputs=ipts, outputs={'Output': out}) + return out + + +def log_loss(input, label, epsilon=1e-4, name=None): + """ + **Negative Log Loss Layer** + + This layer accepts input predictions and target label and returns the + negative log loss. + + .. math:: + + Out = -label * \\log{(input + \\epsilon)} + - (1 - label) * \\log{(1 - input + \\epsilon)} + + Args: + input (Variable|list): a 2-D tensor with shape [N x 1], where N is the + batch size. This input is a probability computed + by the previous operator. + label (Variable|list): the ground truth which is a 2-D tensor with + shape [N x 1], where N is the batch size. + epsilon (float): epsilon + name (string): the name of log_loss + + Returns: + Variable: A 2-D tensor with shape [N x 1], the negative log loss. + + Examples: + .. code-block:: python + + prob = fluid.layers.sigmoid(net) + cost = fluid.layers.log_loss(input=prob, label=label) + """ + helper = LayerHelper('log_loss', **locals()) + + if name is None: + loss = helper.create_variable_for_type_inference(dtype=input.dtype) + else: + loss = helper.create_variable( + name=name, dtype=input.dtype, persistable=False) + + helper.append_op( + type='log_loss', + inputs={'Predicted': [input], + 'Labels': [label]}, + outputs={'Loss': [loss]}, + attrs={'epsilon': epsilon}) + return loss + + +def add_position_encoding(input, alpha, beta, name=None): + """ + **Add Position Encoding Layer** + + This layer accepts an input 3D-Tensor of shape [N x M x P], and return an + output Tensor of shape [N x M x P] with positional encoding value. + + Refer to `Attention Is All You Need`_ . + + .. math:: + PE(pos, 2i) = \\sin{(pos / 10000^{2i / P})} \\\\ + PE(pos, 2i + 1) = \\cos{(pos / 10000^{2i / P})} \\\\ + Out(:, pos, i) = \\alpha * input(:, pos, i) + \\beta * PE(pos, i) + + Where: + * PE(pos, 2i): the increment for the number at even position + * PE(pos, 2i + 1): the increment for the number at odd position + + Args: + input (Variable): 3-D input tensor with shape [N x M x P] + alpha (float): multiple of Input Tensor + beta (float): multiple of Positional Encoding Tensor + name (string): the name of position encoding layer + + Returns: + Variable: A 3-D Tensor of shape [N x M x P] with positional encoding. + + Examples: + .. code-block:: python + + position_tensor = fluid.layers.add_position_encoding(input=tensor) + """ + helper = LayerHelper('add_position_encoding', **locals()) + dtype = helper.input_dtype() + + if name is None: + out = helper.create_variable_for_type_inference(dtype=dtype) + else: + out = helper.create_variable(name=name, dtype=dtype, persistable=False) + + helper.append_op( + type="add_position_encoding", + inputs={"X": input}, + outputs={"Out": out}, + attrs={"alpha": alpha, + "beta": beta}) + return out + + +def bilinear_tensor_product(x, + y, + size, + act=None, + name=None, + param_attr=None, + bias_attr=None): + """ + **Add Bilinear Tensor Product Layer** + + This layer performs bilinear tensor product on two inputs. + For example: + + .. math:: + out{i} = x * W_{i} * {y^\mathrm{T}}, i=0,1,...,size-1 + + In this formula: + - :math:`x`: the first input contains M elements, shape is [batch_size, M]. + - :math:`y`: the second input contains N elements, shape is [batch_size, N]. + - :math:`W_{i}`: the i-th learned weight, shape is [M, N] + - :math:`out{i}`: the i-th element of out, shape is [batch_size, size]. + - :math:`y^\mathrm{T}`: the transpose of :math:`y_{2}`. + + Args: + x (Variable): 2-D input tensor with shape [batch_size, M] + y (Variable): 2-D input tensor with shape [batch_size, N] + size (int): The dimension of this layer. + act (str, default None): Activation to be applied to the output of this layer. + name (str, default None): The name of this layer. + param_attr (ParamAttr, default None): The parameter attribute for the learnable w. + parameters/weights of this layer. + bias_attr (ParamAttr, default None): The parameter attribute for the bias + of this layer. If it is set to False, no bias will be added to the output units. + If it is set to None, the bias is initialized zero. Default: None. + + Returns: + Variable: A 2-D Tensor of shape [batch_size, size]. + + Examples: + .. code-block:: python + + tensor = bilinear_tensor_product(x=layer1, y=layer2, size=1000) + """ + helper = LayerHelper('bilinear_tensor_product', **locals()) + dtype = helper.input_dtype('x') + + param_shape = [size, x.shape[1], y.shape[1]] + + w = helper.create_parameter( + attr=helper.param_attr, shape=param_shape, dtype=dtype, is_bias=False) + + if name is None: + out = helper.create_variable_for_type_inference(dtype=dtype) + else: + out = helper.create_variable(name=name, dtype=dtype, persistable=False) + + inputs = {"X": x, "Y": y, "Weight": w} + if helper.bias_attr: + bias_size = [1, size] + bias = helper.create_parameter( + attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True) + inputs["Bias"] = bias + helper.append_op( + type="bilinear_tensor_product", inputs=inputs, outputs={"Out": out}) + + # add activation + return helper.append_activation(out) diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index 9a8300524d8784fae598635796888382b1adbccf..1ff40a26f2f24e2ff06719972489b0c1e5d140c3 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -14,6 +14,8 @@ from __future__ import print_function from .layer_function_generator import generate_layer_fn, generate_layer_fn_noattr +from .. import core +from ..framework import convert_np_dtype_to_dtype_ __activations_noattr__ = [ 'sigmoid', @@ -58,8 +60,11 @@ _uniform_random_ = generate_layer_fn('uniform_random') def uniform_random(shape, dtype=None, min=None, max=None, seed=None): + locals_var = locals().keys() + if not isinstance(dtype, core.VarDesc.VarType): + dtype = convert_np_dtype_to_dtype_(dtype) kwargs = dict() - for name in locals(): + for name in locals_var: val = locals()[name] if val is not None: kwargs[name] = val @@ -78,8 +83,9 @@ _hard_shrink_ = generate_layer_fn('hard_shrink') def hard_shrink(x, threshold=None): + locals_var = locals().keys() kwargs = dict() - for name in locals(): + for name in locals_var: val = locals()[name] if val is not None: kwargs[name] = val @@ -99,12 +105,12 @@ _cum_sum_ = generate_layer_fn('cumsum') def cumsum(x, axis=None, exclusive=None, reverse=None): + locals_var = locals().keys() kwargs = dict() - for name in locals(): + for name in locals_var: val = locals()[name] if val is not None: kwargs[name] = val - return _cum_sum_(**kwargs) @@ -121,8 +127,9 @@ _thresholded_relu_ = generate_layer_fn('thresholded_relu') def thresholded_relu(x, threshold=None): + locals_var = locals().keys() kwargs = dict() - for name in locals(): + for name in locals_var: val = locals()[name] if val is not None: kwargs[name] = val diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index 44b92af7acc012f89b271c74b026d18e1a4075f8..ff32c00104171bf42c00be33f05758a4387228e1 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -24,10 +24,10 @@ from .layer_function_generator import templatedoc import numpy __all__ = [ - 'create_tensor', 'create_parameter', 'create_global_var', 'cast', 'concat', - 'sums', 'assign', 'fill_constant_batch_size_like', 'fill_constant', - 'argmin', 'argmax', 'argsort', 'ones', 'zeros', 'reverse', 'has_inf', - 'has_nan', 'isfinite' + 'create_tensor', 'create_parameter', 'create_global_var', 'cast', + 'tensor_array_to_tensor', 'concat', 'sums', 'assign', + 'fill_constant_batch_size_like', 'fill_constant', 'argmin', 'argmax', + 'argsort', 'ones', 'zeros', 'reverse', 'has_inf', 'has_nan', 'isfinite' ] @@ -100,7 +100,7 @@ def create_global_var(shape, force_cpu=False, name=None): """ - Create a new variable in the global block(block 0). + Create a new tensor variable with value in the global block(block 0). Args: shape(list[int]): shape of the variable @@ -152,7 +152,7 @@ def cast(x, dtype): result = fluid.layers.cast(x=data, dtype='float64') """ helper = LayerHelper('cast', **locals()) - out = helper.create_tmp_variable(dtype=dtype) + out = helper.create_variable_for_type_inference(dtype=dtype) helper.append_op( type='cast', inputs={'X': [x]}, @@ -184,7 +184,7 @@ def concat(input, axis=0, name=None): out = fluid.layers.concat(input=[Efirst, Esecond, Ethird, Efourth]) """ helper = LayerHelper('concat', **locals()) - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) helper.append_op( type='concat', inputs={'X': input}, @@ -193,6 +193,60 @@ def concat(input, axis=0, name=None): return out +def tensor_array_to_tensor(input, axis=1, name=None): + """ + This function concatenates the input LodTensorArray along the axis mentioned + and returns that as the output. + + A simple example as below: + + .. code-block:: text + + Given: + + input.data = {[[0.6, 0.1, 0.3], + [0.5, 0.3, 0.2]], + [[1.3], + [1.8]], + [[2.3, 2.1], + [2.5, 2.4]]} + + axis = 1 + + Then: + + output.data = [[0.6, 0.1, 0.3, 1.3, 2.3, 2.1], + [0.5, 0.3, 0.2, 1.8, 2.5, 2.4]] + + output_index.data = [3, 1, 2] + + Args: + input(list): Input LodTensorArray + axis(int): Integer axis along which the tensors will be concatenated + name(str|None): A name for this layer(optional). If set None, the layer + will be named automatically. + + Returns: + Variable: Output variable of the concatenation + Variable: The input LodTensorArray items' dims along the axis + + Examples: + .. code-block:: python + + output, output_index = fluid.layers.tensor_array_to_tensor(input=tensor_array) + """ + helper = LayerHelper('tensor_array_to_tensor', **locals()) + out = helper.create_variable_for_type_inference(dtype=helper.input_dtype()) + out_index = helper.create_variable_for_type_inference(dtype="int32") + helper.append_op( + type='tensor_array_to_tensor', + inputs={'X': input}, + outputs={'Out': [out], + 'OutIndex': [out_index]}, + attrs={'axis': axis}) + return out, out_index + + def sums(input, out=None): """ This function performs the sum operation on the input and returns the @@ -221,7 +275,8 @@ def sums(input, out=None): """ helper = LayerHelper('sum', **locals()) if out is None: - out = helper.create_tmp_variable(dtype=helper.input_dtype()) + out = helper.create_variable_for_type_inference( + dtype=helper.input_dtype()) helper.append_op( type='sum', inputs={'X': input}, @@ -252,7 +307,7 @@ def assign(input, output=None): """ helper = LayerHelper('assign', **locals()) if output is None: - output = helper.create_tmp_variable(dtype=input.dtype) + output = helper.create_variable_for_type_inference(dtype=input.dtype) if isinstance(input, Variable): helper.append_op( type='assign', inputs={'X': [input]}, outputs={'Out': [output]}) @@ -311,7 +366,7 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None): helper = LayerHelper("fill_constant", **locals()) if out is None: - out = helper.create_tmp_variable(dtype=dtype) + out = helper.create_variable_for_type_inference(dtype=dtype) helper.append_op( type='fill_constant', inputs={}, @@ -358,7 +413,7 @@ def fill_constant_batch_size_like(input, ${out_comment}. """ helper = LayerHelper("fill_constant_batch_size_like", **locals()) - out = helper.create_tmp_variable(dtype=dtype) + out = helper.create_variable_for_type_inference(dtype=dtype) helper.append_op( type='fill_constant_batch_size_like', inputs={'Input': input}, @@ -396,7 +451,7 @@ def argmin(x, axis=0): out = fluid.layers.argmin(x=in, axis=-1) """ helper = LayerHelper("arg_min", **locals()) - out = helper.create_tmp_variable(VarDesc.VarType.INT64) + out = helper.create_variable_for_type_inference(VarDesc.VarType.INT64) helper.append_op( type='arg_min', inputs={'X': x}, @@ -427,7 +482,7 @@ def argmax(x, axis=0): out = fluid.layers.argmax(x=in, axis=-1) """ helper = LayerHelper("arg_max", **locals()) - out = helper.create_tmp_variable(VarDesc.VarType.INT64) + out = helper.create_variable_for_type_inference(VarDesc.VarType.INT64) helper.append_op( type='arg_max', inputs={'X': x}, @@ -477,8 +532,10 @@ def argsort(input, axis=-1, name=None): out, indices = fluid.layers.argsort(input, axis=0) """ helper = LayerHelper("argsort", **locals()) - out = helper.create_tmp_variable(dtype=input.dtype, stop_gradient=True) - ids = helper.create_tmp_variable(VarDesc.VarType.INT64, stop_gradient=True) + out = helper.create_variable_for_type_inference( + dtype=input.dtype, stop_gradient=True) + ids = helper.create_variable_for_type_inference( + VarDesc.VarType.INT64, stop_gradient=True) helper.append_op( type='argsort', inputs={'X': input}, @@ -562,7 +619,7 @@ def reverse(x, axis): if isinstance(axis, int): axis = [axis] helper = LayerHelper("reverse", **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op( type='reverse', inputs={'Input': x}, @@ -654,7 +711,7 @@ def has_inf(x): Variable: The tensor variable storing the output, only a bool value. """ helper = LayerHelper("isinf", **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op(type="isinf", inputs={"X": x}, outputs={"Out": out}) return out @@ -670,7 +727,7 @@ def has_nan(x): Variable: The tensor variable storing the output, only a bool value. """ helper = LayerHelper("isnan", **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op(type="isnan", inputs={"X": x}, outputs={"Out": out}) return out @@ -687,6 +744,6 @@ def isfinite(x): Variable: The tensor variable storing the output, contains a bool value. """ helper = LayerHelper("isfinite", **locals()) - out = helper.create_tmp_variable(dtype=x.dtype) + out = helper.create_variable_for_type_inference(dtype=x.dtype) helper.append_op(type="isfinite", inputs={"X": x}, outputs={"Out": out}) return out diff --git a/python/paddle/fluid/metrics.py b/python/paddle/fluid/metrics.py index 0c2800dcf35ed156b71625babea2724f520575e5..f65b37903a35fa2bf9f2c2b2f169ce6fd4c478db 100644 --- a/python/paddle/fluid/metrics.py +++ b/python/paddle/fluid/metrics.py @@ -13,8 +13,6 @@ # limitations under the License. """ Fluid Metrics - -The metrics are accomplished via Python natively. """ from __future__ import print_function @@ -24,6 +22,12 @@ import copy import warnings import six +from .layer_helper import LayerHelper +from .initializer import Constant +from . import unique_name +from .framework import Program, Variable, program_guard +from . import layers + __all__ = [ 'MetricBase', 'CompositeMetric', @@ -190,7 +194,7 @@ class CompositeMetric(MetricBase): or soft-label, should custom the corresponding update rule. """ for m in self._metrics: - ans.append(m.update(preds, labels)) + m.update(preds, labels) def eval(self): """ @@ -474,71 +478,10 @@ class EditDistance(MetricBase): "There is no data in EditDistance Metric. Please check layers.edit_distance output has been added to EditDistance." ) avg_distance = self.total_distance / self.seq_num - avg_instance_error = self.instance_error / self.seq_num + avg_instance_error = self.instance_error / float(self.seq_num) return avg_distance, avg_instance_error -class DetectionMAP(MetricBase): - """ - Calculate the detection mean average precision (mAP). - mAP is the metric to measure the accuracy of object detectors - like Faster R-CNN, SSD, etc. - It is the average of the maximum precisions at different recall values. - Please get more information from the following articles: - https://sanchom.wordpress.com/tag/average-precision/ - - https://arxiv.org/abs/1512.02325 - - The general steps are as follows: - - 1. calculate the true positive and false positive according to the input - of detection and labels. - 2. calculate mAP value, support two versions: '11 point' and 'integral'. - - Examples: - .. code-block:: python - - pred = fluid.layers.fc(input=data, size=1000, act="tanh") - batch_map = layers.detection_map( - input, - label, - class_num, - background_label, - overlap_threshold=overlap_threshold, - evaluate_difficult=evaluate_difficult, - ap_version=ap_version) - metric = fluid.metrics.DetectionMAP() - for data in train_reader(): - loss, preds, labels = exe.run(fetch_list=[cost, batch_map]) - batch_size = data[0] - metric.update(value=batch_map, weight=batch_size) - numpy_map = metric.eval() - """ - - def __init__(self, name=None): - super(DetectionMAP, self).__init__(name) - # the current map value - self.value = .0 - self.weight = .0 - - def update(self, value, weight): - if not _is_number_or_matrix_(value): - raise ValueError( - "The 'value' must be a number(int, float) or a numpy ndarray.") - if not _is_number_(weight): - raise ValueError("The 'weight' must be a number(int, float).") - self.value += value - self.weight += weight - - def eval(self): - if self.weight == 0: - raise ValueError( - "There is no data in DetectionMAP Metrics. " - "Please check layers.detection_map output has added to DetectionMAP." - ) - return self.value / self.weight - - class Auc(MetricBase): """ Auc metric adapts to the binary classification. @@ -616,3 +559,179 @@ class Auc(MetricBase): idx -= 1 return auc / tot_pos / tot_neg if tot_pos > 0.0 and tot_neg > 0.0 else 0.0 + + +class DetectionMAP(object): + """ + Calculate the detection mean average precision (mAP). + + The general steps are as follows: + 1. calculate the true positive and false positive according to the input + of detection and labels. + 2. calculate mAP value, support two versions: '11 point' and 'integral'. + + Please get more information from the following articles: + https://sanchom.wordpress.com/tag/average-precision/ + https://arxiv.org/abs/1512.02325 + + Args: + input (Variable): The detection results, which is a LoDTensor with shape + [M, 6]. The layout is [label, confidence, xmin, ymin, xmax, ymax]. + gt_label (Variable): The ground truth label index, which is a LoDTensor + with shape [N, 1]. + gt_box (Variable): The ground truth bounding box (bbox), which is a + LoDTensor with shape [N, 4]. The layout is [xmin, ymin, xmax, ymax]. + gt_difficult (Variable|None): Whether this ground truth is a difficult + bounding bbox, which can be a LoDTensor [N, 1] or not set. If None, + it means all the ground truth labels are not difficult bbox. + class_num (int): The class number. + background_label (int): The index of background label, the background + label will be ignored. If set to -1, then all categories will be + considered, 0 by defalut. + overlap_threshold (float): The threshold for deciding true/false + positive, 0.5 by defalut. + evaluate_difficult (bool): Whether to consider difficult ground truth + for evaluation, True by defalut. This argument does not work when + gt_difficult is None. + ap_version (string): The average precision calculation ways, it must be + 'integral' or '11point'. Please check + https://sanchom.wordpress.com/tag/average-precision/ for details. + - 11point: the 11-point interpolated average precision. + - integral: the natural integral of the precision-recall curve. + + Examples: + .. code-block:: python + + exe = fluid.Executor(place) + map_evaluator = fluid.Evaluator.DetectionMAP(input, + gt_label, gt_box, gt_difficult) + cur_map, accum_map = map_evaluator.get_map_var() + fetch = [cost, cur_map, accum_map] + for epoch in PASS_NUM: + map_evaluator.reset(exe) + for data in batches: + loss, cur_map_v, accum_map_v = exe.run(fetch_list=fetch) + + In the above example: + + 'cur_map_v' is the mAP of current mini-batch. + 'accum_map_v' is the accumulative mAP of one pass. + """ + + def __init__(self, + input, + gt_label, + gt_box, + gt_difficult=None, + class_num=None, + background_label=0, + overlap_threshold=0.5, + evaluate_difficult=True, + ap_version='integral'): + + self.helper = LayerHelper('map_eval') + gt_label = layers.cast(x=gt_label, dtype=gt_box.dtype) + if gt_difficult: + gt_difficult = layers.cast(x=gt_difficult, dtype=gt_box.dtype) + label = layers.concat([gt_label, gt_difficult, gt_box], axis=1) + else: + label = layers.concat([gt_label, gt_box], axis=1) + + # calculate mean average precision (mAP) of current mini-batch + map = layers.detection_map( + input, + label, + class_num, + background_label, + overlap_threshold=overlap_threshold, + evaluate_difficult=evaluate_difficult, + ap_version=ap_version) + + states = [] + states.append( + self._create_state( + dtype='int32', shape=None, suffix='accum_pos_count')) + states.append( + self._create_state( + dtype='float32', shape=None, suffix='accum_true_pos')) + states.append( + self._create_state( + dtype='float32', shape=None, suffix='accum_false_pos')) + var = self._create_state(dtype='int32', shape=[1], suffix='has_state') + self.helper.set_variable_initializer( + var, initializer=Constant(value=int(0))) + self.has_state = var + + # calculate accumulative mAP + accum_map = layers.detection_map( + input, + label, + class_num, + background_label, + overlap_threshold=overlap_threshold, + evaluate_difficult=evaluate_difficult, + has_state=self.has_state, + input_states=states, + out_states=states, + ap_version=ap_version) + + layers.fill_constant( + shape=self.has_state.shape, + value=1, + dtype=self.has_state.dtype, + out=self.has_state) + + self.cur_map = map + self.accum_map = accum_map + + def _create_state(self, suffix, dtype, shape): + """ + Create state variable. + Args: + suffix(str): the state suffix. + dtype(str|core.VarDesc.VarType): the state data type + shape(tuple|list): the shape of state + Returns: State variable + """ + state = self.helper.create_variable( + name="_".join([unique_name.generate(self.helper.name), suffix]), + persistable=True, + dtype=dtype, + shape=shape) + return state + + def get_map_var(self): + """ + Returns: mAP variable of current mini-batch and + accumulative mAP variable cross mini-batches. + """ + return self.cur_map, self.accum_map + + def reset(self, executor, reset_program=None): + """ + Reset metric states at the begin of each pass/user specified batch. + + Args: + executor(Executor): a executor for executing + the reset_program. + reset_program(Program|None): a single Program for reset process. + If None, will create a Program. + """ + + def _clone_var_(block, var): + assert isinstance(var, Variable) + return block.create_var( + name=var.name, + shape=var.shape, + dtype=var.dtype, + type=var.type, + lod_level=var.lod_level, + persistable=var.persistable) + + if reset_program is None: + reset_program = Program() + with program_guard(main_program=reset_program): + var = _clone_var_(reset_program.current_block(), self.has_state) + layers.fill_constant( + shape=var.shape, value=0, dtype=var.dtype, out=var) + executor.run(reset_program) diff --git a/python/paddle/fluid/nets.py b/python/paddle/fluid/nets.py index 1dabad54f5b976e0fcabf6918d3bc6ece4eed384..00d33b36fcc3266bf7f08020052d28172665e53e 100644 --- a/python/paddle/fluid/nets.py +++ b/python/paddle/fluid/nets.py @@ -64,23 +64,33 @@ def simple_img_conv_pool(input, average-pooling. Default :math:`max`. global_pooling (bool): Whether to use the global pooling. If global_pooling = true, pool_size and pool_padding while be ignored. Default False - conv_stride (int|list|tuple): The stride size of the Conv2d Layer. If stride is a + conv_stride (int|list|tuple): The stride size of the conv2d Layer. If stride is a list or tuple, it must contain two integers, (conv_stride_H, conv_stride_W). Otherwise, the conv_stride_H = conv_stride_W = conv_stride. Default: conv_stride = 1. - conv_padding (int|list|tuple): The padding size of the Conv2d Layer. If padding is + conv_padding (int|list|tuple): The padding size of the conv2d Layer. If padding is a list or tuple, it must contain two integers, (conv_padding_H, conv_padding_W). Otherwise, the conv_padding_H = conv_padding_W = conv_padding. Default: conv_padding = 0. - conv_dilation (int|list|tuple): The dilation size of the Conv2d Layer. If dilation is + conv_dilation (int|list|tuple): The dilation size of the conv2d Layer. If dilation is a list or tuple, it must contain two integers, (conv_dilation_H, conv_dilation_W). Otherwise, the conv_dilation_H = conv_dilation_W = conv_dilation. Default: conv_dilation = 1. - conv_groups (int): The groups number of the Conv2d Layer. According to grouped + conv_groups (int): The groups number of the conv2d Layer. According to grouped convolution in Alex Krizhevsky's Deep CNN paper: when group=2, the first half of the filters is only connected to the first half of the input channels, while the second half of the filters is only - connected to the second half of the input channels. Default: groups=1 - param_attr (ParamAttr): The parameters to the Conv2d Layer. Default: None - bias_attr (ParamAttr): Bias parameter for the Conv2d layer. Default: None - act (str): Activation type for Conv2d. Default: None + connected to the second half of the input channels. Default: groups=1. + param_attr (ParamAttr|None): The parameter attribute for learnable parameters/weights + of conv2d. If it is set to None or one attribute of ParamAttr, conv2d + will create ParamAttr as param_attr. If the Initializer of the param_attr + is not set, the parameter is initialized with :math:`Normal(0.0, std)`, + and the :math:`std` is :math:`(\\frac{2.0 }{filter\_elem\_num})^{0.5}`. + Default: None. + bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of conv2d. + If it is set to False, no bias will be added to the output units. + If it is set to None or one attribute of ParamAttr, conv2d + will create ParamAttr as bias_attr. If the Initializer of the bias_attr + is not set, the bias is initialized zero. Default: None. + act (str): Activation type for conv2d, if it is set to None, activation is not + appended. Default: None. use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn library is installed. Default: True diff --git a/python/paddle/fluid/op.py b/python/paddle/fluid/op.py index 667db10d3ebdd24ddd9efbe2310ebb331e268ee2..b8bb3db1eedcf25c9b6a02ad3b4f261e8be8efce 100644 --- a/python/paddle/fluid/op.py +++ b/python/paddle/fluid/op.py @@ -108,6 +108,8 @@ class OpDescCreationMethod(object): new_attr.i = user_defined_attr elif attr.type == framework_pb2.FLOAT: new_attr.f = user_defined_attr + elif attr.type == framework_pb2.LONG: + new_attr.l = user_defined_attr elif attr.type == framework_pb2.STRING: new_attr.s = user_defined_attr elif attr.type == framework_pb2.BOOLEAN: @@ -120,6 +122,8 @@ class OpDescCreationMethod(object): new_attr.strings.extend(user_defined_attr) elif attr.type == framework_pb2.BOOLEANS: new_attr.bools.extend(user_defined_attr) + elif attr.type == framework_pb2.LONGS: + new_attr.longs.extend(user_defined_attr) elif attr.type == framework_pb2.INT_PAIRS: for p in user_defined_attr: pair = new_attr.int_pairs.add() diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index 1b9571f6d3a6a69d1ac35f6be74b80eaa2ce6251..da92826d410505c9a80820f655162dd22e6b5966 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -13,26 +13,30 @@ # limitations under the License. from __future__ import print_function -import re + from collections import defaultdict -from paddle.fluid.framework import Program, Variable, name_scope +from contextlib import contextmanager + +from paddle.fluid.framework import Program, Variable, name_scope, default_main_program +from paddle.fluid.distribute_lookup_table import find_distributed_lookup_table + from . import framework from . import layers +from . import unique_name from .backward import append_backward +from .clip import append_gradient_clip_ops, error_clip_callback from .framework import program_guard -from . import unique_name from .initializer import Constant from .layer_helper import LayerHelper -from .regularizer import append_regularization_ops -from .clip import append_gradient_clip_ops, error_clip_callback -from contextlib import contextmanager from .layers import ops +from .regularizer import append_regularization_ops __all__ = [ 'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl', 'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer', 'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer', - 'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'RMSPropOptimizer' + 'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'LarsMomentum', + 'LarsMomentumOptimizer' ] @@ -83,7 +87,7 @@ class Optimizer(object): name=unique_name.generate("learning_rate"), shape=[1], value=float(self._learning_rate), - dtype='float32' if self._dtype == None else self._dtype, + dtype='float32' if self._dtype is None else self._dtype, persistable=True) def _global_learning_rate(self, program=None): @@ -105,13 +109,15 @@ class Optimizer(object): param = param_and_grad[0] param_lr = param.optimize_attr['learning_rate'] if type(param_lr) == Variable: - print("returns updated param lr ", param_lr) return param_lr else: if param_lr == 1.0: return self._global_learning_rate() else: - return self._global_learning_rate() * param_lr + with default_main_program()._lr_schedule_guard( + is_with_opt=True), framework.name_scope( + 'scale_with_param_lr'): + return self._global_learning_rate() * param_lr def _create_accumulators(self, block, parameters): """Create all accumulators needed by the parameters @@ -241,6 +247,50 @@ class Optimizer(object): end = len(global_block.ops) return global_block._slice_ops(start, end) + def _process_distribute_lookuptable(self, param_grads, loss, + startup_program): + """ + Because distribute lookup table only support SGD optimizer for now, not support + other optimizer and regularization, so we should find the table parameter out, + and avoid to add regularization and other op for it, and add sgd optimize op + for it independently. + :param param_grads(list((Var, Var))): list of (param, grad) pair. + :param loss: the loss variable. + :param startup_program: the startup program + """ + program = loss.block.program + table_name = find_distributed_lookup_table(program) + table_param = None + table_grad = None + new_param_grads = [] + for p, g in param_grads: + if p.name == table_name: + if table_param is not None: + raise RuntimeError( + "multi dist table var found, only support one now!") + table_param = p + table_grad = g + else: + new_param_grads.append((p, g)) + sgd_op = None + if table_param is not None: + with program_guard(program, startup_program): + param_and_grad = [table_param, table_grad] + with table_param.block.program._optimized_guard(param_and_grad), \ + framework.name_scope("optimizer"): + self._create_global_learning_rate() + # create the optimize op + sgd_op = loss.block.append_op( + type='sgd', + inputs={ + "Param": table_param, + "Grad": table_grad, + "LearningRate": + self._create_param_lr(param_and_grad) + }, + outputs={"ParamOut": param_and_grad[0]}) + return new_param_grads, (table_param, table_grad), sgd_op + def minimize(self, loss, startup_program=None, @@ -256,6 +306,9 @@ class Optimizer(object): params_grads = sorted(params_grads, key=lambda x: x[0].name) + params_grads, table_param_and_grad, table_optimize_op = \ + self._process_distribute_lookuptable(params_grads, loss, startup_program) + params_grads = append_gradient_clip_ops(params_grads) # Add regularization if any @@ -264,6 +317,9 @@ class Optimizer(object): optimize_ops = self._create_optimization_pass(params_grads, loss, startup_program) + if table_optimize_op is not None: + optimize_ops.append(table_optimize_op) + params_grads.append(table_param_and_grad) return optimize_ops, params_grads @@ -397,6 +453,91 @@ class MomentumOptimizer(Optimizer): return momentum_op +class LarsMomentumOptimizer(Optimizer): + """ + Momentum optimizer with LARS support + + The update equations are as follows: + + .. math:: + + & local\_learning\_rate = learning\_rate * lars\_coeff * \\ + \\frac{||param||}{||gradient|| + lars\_weight\_decay * ||param||} + + & velocity = mu * velocity + local\_learning\_rate * (gradient + lars\_weight\_decay * param) + + & param = param - velocity + + Args: + learning_rate (float|Variable): the learning rate used to update parameters. \ + Can be a float value or a Variable with one float value as data element. + momentum (float): momentum factor + lars_coeff (float): defines how much we trust the layer to change its weights. + lars_weight_decay (float): weight decay coefficient for decaying using LARS. + regularization: A Regularizer, such as + fluid.regularizer.L2DecayRegularizer. + name: A optional name prefix. + + + Examples: + .. code-block:: python + + optimizer = fluid.optimizer.LarsMomentum(learning_rate=0.2, momentum=0.1, lars_weight_decay=0.001) + optimizer.minimize(cost) + """ + _velocity_acc_str = "velocity" + + def __init__(self, + learning_rate, + momentum, + lars_coeff=0.001, + lars_weight_decay=0.0005, + regularization=None, + name=None): + assert learning_rate is not None + assert momentum is not None + super(LarsMomentumOptimizer, self).__init__( + learning_rate=learning_rate, + regularization=regularization, + name=name) + self.type = "lars_momentum" + self._momentum = momentum + self._lars_coeff = float(lars_coeff) + self._lars_weight_decay = float(lars_weight_decay) + + def _create_accumulators(self, block, parameters): + assert isinstance(block, framework.Block) + + for p in parameters: + self._add_accumulator(self._velocity_acc_str, p) + + def _append_optimize_op(self, block, param_and_grad): + assert isinstance(block, framework.Block) + + velocity_acc = self._get_accumulator(self._velocity_acc_str, + param_and_grad[0]) + # create the momentum optimize op + momentum_op = block.append_op( + type=self.type, + inputs={ + "Param": param_and_grad[0], + "Grad": param_and_grad[1], + "Velocity": velocity_acc, + "LearningRate": self._create_param_lr(param_and_grad) + }, + outputs={ + "ParamOut": param_and_grad[0], + "VelocityOut": velocity_acc + }, + attrs={ + "mu": self._momentum, + "lars_coeff": self._lars_coeff, + "lars_weight_decay": self._lars_weight_decay + }) + + return momentum_op + + class AdagradOptimizer(Optimizer): """ **Adaptive Gradient Algorithm (Adagrad)** @@ -601,7 +742,8 @@ class AdamOptimizer(Optimizer): for param, grad in param_and_grads: if grad is None: continue - with param.block.program._optimized_guard([param, grad]): + with param.block.program._optimized_guard( + [param, grad]), name_scope("optimizer"): beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str, param) beta2_pow_acc = self._get_accumulator(self._beta2_pow_acc_str, @@ -659,6 +801,9 @@ class AdamaxOptimizer(Optimizer): optimizer = fluid.optimizer.Adamax(learning_rate=0.2) optimizer.minimize(cost) + + Notes: + Currently, AdamaxOptimizer doesn't support sparse parameter optimization. """ _moment_acc_str = "moment" _inf_norm_acc_str = "inf_norm" @@ -736,7 +881,8 @@ class AdamaxOptimizer(Optimizer): for param, grad in parameters_and_grads: if grad is None: continue - with param.block.program._optimized_guard([param, grad]): + with param.block.program._optimized_guard( + [param, grad]), name_scope('adamx'): beta1_pow_acc = self._get_accumulator(self._beta1_pow_acc_str, param) main_block.append_op( @@ -778,6 +924,9 @@ class DecayedAdagradOptimizer(Optimizer): optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2) optimizer.minimize(cost) + + Notes: + Currently, DecayedAdagradOptimizer doesn't support sparse parameter optimization. """ _moment_acc_str = "moment" @@ -858,6 +1007,9 @@ class AdadeltaOptimizer(Optimizer): optimizer = fluid.optimizer.Adadelta( learning_rate=0.0003, epsilon=1.0e-6, rho=0.95) _, params_grads = optimizer.minimize(cost) + + Notes: + Currently, AdadeltaOptimizer doesn't support sparse parameter optimization. """ _avg_squared_grad_acc_str = "_avg_squared_grad" @@ -1126,6 +1278,9 @@ class FtrlOptimizer(Optimizer): optimizer = fluid.optimizer.Ftrl(0.0001) _, params_grads = optimizer.minimize(cost) + + Notes: + Currently, FtrlOptimizer doesn't support sparse parameter optimization. """ _squared_acc_str = "squared" @@ -1204,6 +1359,7 @@ DecayedAdagrad = DecayedAdagradOptimizer Adadelta = AdadeltaOptimizer RMSProp = RMSPropOptimizer Ftrl = FtrlOptimizer +LarsMomentum = LarsMomentumOptimizer class ModelAverage(Optimizer): @@ -1266,7 +1422,8 @@ class ModelAverage(Optimizer): for param, grad in self.params_grads: if grad is None: continue - with param.block.program._optimized_guard([param, grad]): + with param.block.program._optimized_guard( + [param, grad]), name_scope('move_average'): self._append_average_accumulate_op(param) self.apply_program = Program() diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 57d272cbfb948840679e80e8db40379c57603113..3f4dd5eb712e738bbee8f93c062375033b8ab2f6 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -31,15 +31,32 @@ BuildStrategy = core.ParallelExecutor.BuildStrategy class ParallelExecutor(object): """ - ParallelExecutor can run program in parallel. + ParallelExecutor is designed for data parallelism, which focuses on distributing + the data across different nodes and every node operates on the data in parallel. + If you use ParallelExecutor to run the current program on GPU, the node means GPU + device, and ParallelExecutor will get the available GPU device automatically on + the current machine. If you use ParallelExecutor to run the current program on CPU, + the node means the CPU device, and you can specify the CPU device number by adding + 'CPU_NUM' environment variable, for example 'CPU_NUM=4', if the environment variable + is not found, ParallelExecutor will call `multiprocessing.cpu_count` to get the number + of CPUs in the system. Args: use_cuda (bool): Whether to use CUDA or not. loss_name (str): The loss name must set in training. Default None. main_program (Program): The program that need to run, if not provided, then default_main_program will be used. Default None. - share_vars_from(ParallelExecutor): If provied, it will share variables + share_vars_from(ParallelExecutor): If provide, it will share variables from the specified ParallelExecutor. Default None. + exec_strategy(ExecutionStrategy): exec_strategy is used to control how to run + the program in ParallelExecutor, for example how many threads are used to + execute the program, how many iterations to clean up the temp variables + which is generated during execution. For more information, please refer + to fluid.ExecutionStrategy. Default None. + build_strategy(BuildStrategy): build_strategy is used to control how to + build the SSA Graph in ParallelExecutor by setting the property, + for example reduce_strategy, gradient_scale_strategy. For more information, + please refer to fluid.BuildStrategy. Default None. num_trainers(int): If greater than 1, NCCL will be initialized with multiple rank of nodes, each node should have same number of GPUs. Distributed training will be enabled then. Default 1. diff --git a/python/paddle/fluid/recordio_writer.py b/python/paddle/fluid/recordio_writer.py index a69c0c29d4675d3e6b9b2a2d766b8be9935092cf..076a942cdde5623faa570bf98f889e8145b60f8b 100644 --- a/python/paddle/fluid/recordio_writer.py +++ b/python/paddle/fluid/recordio_writer.py @@ -41,9 +41,6 @@ def convert_reader_to_recordio_file( """ Convert a Python Reader to a recordio file. - Please see :ref:`api_guide_python_reader` and :ref:`api_guide_reader_op` for - details. - Examples: >>> import paddle.fluid as fluid diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py index a4336e955f21b0b09bf3dadbd437855c06745860..d8aace9fdfa601413bb4d4b1b2a309ba6a8e4ece 100644 --- a/python/paddle/fluid/regularizer.py +++ b/python/paddle/fluid/regularizer.py @@ -47,7 +47,8 @@ def append_regularization_ops(parameters_and_grads, regularization=None): if grad is None: params_and_grads.append((param, grad)) continue - with param.block.program._optimized_guard([param, grad]): + with param.block.program._optimized_guard( + [param, grad]), framework.name_scope('regularization'): regularization_term = None if param.regularizer is not None: # Add variable for regularization term in grad block @@ -60,14 +61,25 @@ def append_regularization_ops(parameters_and_grads, regularization=None): params_and_grads.append((param, grad)) continue - assert grad.shape == regularization_term.shape + new_grad = grad + if grad.type == core.VarDesc.VarType.SELECTED_ROWS: + # FIXME(zcd): If the grad is SELECTED_ROWS, after regularization, + # the grad's type and name will be changed. But the gradient's name + # is used in ParallelExecutor Reduce mode, so I add a flag for + # the new_grad here. + new_grad = grad.block.create_var( + name=grad.name + core.kNewGradSuffix(), + dtype=param.dtype, + shape=param.shape, + lod_level=param.lod_level, + type=core.VarDesc.VarType.LOD_TENSOR) grad.block.append_op( - type='elementwise_add', - inputs={"X": grad, - "Y": regularization_term}, - outputs={"Out": grad}) - params_and_grads.append((param, grad)) + type='sum', + inputs={"X": [grad, regularization_term]}, + outputs={"Out": new_grad}) + + params_and_grads.append((param, new_grad)) return params_and_grads @@ -141,26 +153,7 @@ class L2DecayRegularizer(WeightDecayRegularizer): assert isinstance(block, framework.Block) decay = block.create_var( - dtype="float32", shape=param.shape, lod_level=param.lod_level) - - if grad.type == core.VarDesc.VarType.SELECTED_ROWS: - idx = block.create_var( - dtype="int64", - shape=param.shape, - type=core.VarDesc.VarType.LOD_TENSOR) - decay = block.create_var( - dtype="float32", - shape=param.shape, - type=core.VarDesc.VarType.SELECTED_ROWS) - block.append_op( - type='extract_rows', inputs={'X': grad}, outputs={'Out': idx}) - block.append_op( - type='lookup_table', - inputs={'W': param, - 'Ids': idx}, - outputs={'Out': decay}, - attrs={'is_sparse': True}) - param = decay + dtype=param.dtype, shape=param.shape, lod_level=param.lod_level) # Append Op to calculate decay block.append_op( @@ -217,26 +210,9 @@ class L1DecayRegularizer(WeightDecayRegularizer): """ assert isinstance(param, framework.Parameter) assert isinstance(block, framework.Block) + decay = block.create_var( - dtype="float32", shape=param.shape, lod_level=param.lod_level) - - if grad.type == core.VarDesc.VarType.SELECTED_ROWS: - idx = block.create_var( - dtype="int64", - shape=param.shape, - type=core.VarDesc.VarType.LOD_TENSOR) - decay = block.create_var( - dtype="float32", - shape=param.shape, - type=core.VarDesc.VarType.SELECTED_ROWS) - block.append_op( - type='extract_rows', inputs={'X': grad}, outputs={'Out': idx}) - block.append_op( - type='lookup_table', - inputs={'W': param, - 'Ids': idx}, - outputs={'Out': decay}, - attrs={'is_sparse': True}) + dtype=param.dtype, shape=param.shape, lod_level=param.lod_level) # Append sign op block.append_op( diff --git a/python/paddle/fluid/tests/CMakeLists.txt b/python/paddle/fluid/tests/CMakeLists.txt index 1885dda44ab5eaeca6a4f54e4b84379c71ec3167..d24417bbacb503d9ea70e68e7e0edb59e7dddbde 100644 --- a/python/paddle/fluid/tests/CMakeLists.txt +++ b/python/paddle/fluid/tests/CMakeLists.txt @@ -1,4 +1,3 @@ -set(PYTHON_TESTS_DIR ${CMAKE_CURRENT_BINARY_DIR} CACHE PATH "python tests directory") file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt b/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt index 673c965b662a022739f8d489c331f4de9455a926..91c1d17eb5391ea37a41a886594cc71c6e6c56bd 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt @@ -1,7 +1,19 @@ file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py") string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") -# default test -foreach(src ${TEST_OPS}) - py_test(${src} SRCS ${src}.py) -endforeach() +if(NOT APPLE) + # default test + foreach(src ${TEST_OPS}) + py_test(${src} SRCS ${src}.py) + endforeach() +else() + foreach(src ${TEST_OPS}) + if(${src} STREQUAL "test_image_classification_vgg") + message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src}) + elseif(${src} STREQUAL "test_image_classification_resnet") + message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src}) + elseif() + py_test(${src} SRCS ${src}.py) + endif() + endforeach() +endif() diff --git a/python/paddle/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/fluid/tests/book/test_label_semantic_roles.py index f63387a90617dc4e9b7c9ee7caa2d01595237a03..42ab9b231153f7ede7b8f8dd4e754f8cc92f65fe 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -38,7 +38,7 @@ depth = 8 mix_hidden_lr = 1e-3 IS_SPARSE = True -PASS_NUM = 10 +PASS_NUM = 1 BATCH_SIZE = 10 embedding_name = 'emb' diff --git a/python/paddle/fluid/tests/test_detection.py b/python/paddle/fluid/tests/test_detection.py index 56129641ce5900d82aedf243d2fa1eadfd6b8d86..982d29180141d052e25ea3dcba6e3e7ce4181c48 100644 --- a/python/paddle/fluid/tests/test_detection.py +++ b/python/paddle/fluid/tests/test_detection.py @@ -128,6 +128,24 @@ class TestPriorBox(unittest.TestCase): assert box.shape[3] == 4 +class TestDensityPriorBox(unittest.TestCase): + def test_density_prior_box(self): + data_shape = [3, 224, 224] + images = fluid.layers.data( + name='pixel', shape=data_shape, dtype='float32') + conv1 = fluid.layers.conv2d(images, 3, 3, 2) + box, var = layers.density_prior_box( + input=conv1, + image=images, + densities=[3, 4], + fixed_sizes=[50., 60.], + fixed_ratios=[1.0], + clip=True) + assert len(box.shape) == 4 + assert box.shape == var.shape + assert box.shape[3] == 4 + + class TestAnchorGenerator(unittest.TestCase): def test_anchor_generator(self): data_shape = [3, 224, 224] @@ -301,7 +319,7 @@ class TestRpnTargetAssign(unittest.TestCase): dtype='float32', lod_level=1, append_batch_size=False) - pred_scores, pred_loc, tgt_lbl, tgt_bbox = layers.rpn_target_assign( + pred_scores, pred_loc, tgt_lbl, tgt_bbox, bbox_inside_weight = layers.rpn_target_assign( bbox_pred=bbox_pred, cls_logits=cls_logits, anchor_box=anchor_box, @@ -313,15 +331,18 @@ class TestRpnTargetAssign(unittest.TestCase): rpn_straddle_thresh=0.0, rpn_fg_fraction=0.5, rpn_positive_overlap=0.7, - rpn_negative_overlap=0.3) + rpn_negative_overlap=0.3, + use_random=False) self.assertIsNotNone(pred_scores) self.assertIsNotNone(pred_loc) self.assertIsNotNone(tgt_lbl) self.assertIsNotNone(tgt_bbox) + self.assertIsNotNone(bbox_inside_weight) assert pred_scores.shape[1] == 1 assert pred_loc.shape[1] == 4 assert pred_loc.shape[1] == tgt_bbox.shape[1] + print(str(program)) class TestGenerateProposals(unittest.TestCase): diff --git a/python/paddle/fluid/tests/unittests/CMakeLists.txt b/python/paddle/fluid/tests/unittests/CMakeLists.txt index 7de0ebce06e9de439d3570bee9ac7dbce33ee868..1513eca51439288acac35729300bcbe4e71e4205 100644 --- a/python/paddle/fluid/tests/unittests/CMakeLists.txt +++ b/python/paddle/fluid/tests/unittests/CMakeLists.txt @@ -17,6 +17,10 @@ if(NOT WITH_DISTRIBUTE) list(REMOVE_ITEM TEST_OPS test_listen_and_serv_op) LIST(REMOVE_ITEM TEST_OPS test_dist_mnist) LIST(REMOVE_ITEM TEST_OPS test_dist_word2vec) + LIST(REMOVE_ITEM TEST_OPS test_dist_ctr) + LIST(REMOVE_ITEM TEST_OPS test_dist_simnet_bow) + LIST(REMOVE_ITEM TEST_OPS test_dist_mnist_batch_merge) + LIST(REMOVE_ITEM TEST_OPS test_dist_text_classification) endif(NOT WITH_DISTRIBUTE) list(REMOVE_ITEM TEST_OPS test_seq_concat_op) # FIXME(helin): https://github.com/PaddlePaddle/Paddle/issues/8290 @@ -55,6 +59,7 @@ function(py_test_modules TARGET_NAME) if (py_test_modules_SERIAL) set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1) endif() + set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600) endif() endfunction() list(REMOVE_ITEM TEST_OPS test_warpctc_op) @@ -78,9 +83,11 @@ if(WITH_DISTRIBUTE) set_tests_properties(test_dist_word2vec PROPERTIES TIMEOUT 200) py_test_modules(test_dist_se_resnext MODULES test_dist_se_resnext) set_tests_properties(test_dist_se_resnext PROPERTIES TIMEOUT 1000) - # TODO: fix this test - #py_test_modules(test_dist_transformer MODULES test_dist_transformer) - #set_tests_properties(test_dist_transformer PROPERTIES TIMEOUT 1000) + # FIXME(typhoonzero): add this back + #py_test_modules(test_dist_transformer MODULES test_dist_transformer) + #set_tests_properties(test_dist_transformer PROPERTIES TIMEOUT 1000) + # TODO(typhoonzero): make dist test parallel when fix port management issue + set_tests_properties(test_dist_mnist test_dist_word2vec test_dist_se_resnext test_dist_ctr test_dist_simnet_bow test_dist_save_load test_dist_text_classification test_dist_mnist_batch_merge PROPERTIES RUN_SERIAL TRUE) endif(NOT APPLE) py_test_modules(test_dist_transpiler MODULES test_dist_transpiler) endif() @@ -88,4 +95,6 @@ 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 150) py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL) -py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL) +if(NOT APPLE) + py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL) +endif() diff --git a/python/paddle/fluid/tests/unittests/dist_mnist.py b/python/paddle/fluid/tests/unittests/dist_mnist.py index 877d21ae882ab4efb49beb6a846ab71a22c2aab7..1cda2711f765622b0bda6f4c688f69352bbd2a6f 100644 --- a/python/paddle/fluid/tests/unittests/dist_mnist.py +++ b/python/paddle/fluid/tests/unittests/dist_mnist.py @@ -90,12 +90,14 @@ class TestDistMnist2x2(TestDistRunnerBase): inference_program = fluid.default_main_program().clone() # Optimization - opt = fluid.optimizer.AdamOptimizer( - learning_rate=0.001, beta1=0.9, beta2=0.999) + # TODO(typhoonzero): fix distributed adam optimizer + # opt = fluid.optimizer.AdamOptimizer( + # learning_rate=0.001, beta1=0.9, beta2=0.999) + opt = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9) # Reader train_reader = paddle.batch( - paddle.dataset.mnist.train(), batch_size=batch_size) + paddle.dataset.mnist.test(), batch_size=batch_size) test_reader = paddle.batch( paddle.dataset.mnist.test(), batch_size=batch_size) opt.minimize(avg_cost) diff --git a/python/paddle/fluid/tests/unittests/dist_mnist_batch_merge.py b/python/paddle/fluid/tests/unittests/dist_mnist_batch_merge.py new file mode 100644 index 0000000000000000000000000000000000000000..d386e75fd887a898f5a13e48e378e08ff6c99ea0 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/dist_mnist_batch_merge.py @@ -0,0 +1,80 @@ +# 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 numpy as np +import argparse +import time +import math + +import paddle +import paddle.fluid as fluid +import paddle.fluid.profiler as profiler +from paddle.fluid import core +import unittest +from multiprocessing import Process +import os +import signal +from functools import reduce +from test_dist_base import TestDistRunnerBase, runtime_main +from dist_mnist import cnn_model + +DTYPE = "float32" + + +def test_merge_reader(repeat_batch_size=8): + orig_reader = paddle.dataset.mnist.test() + record_batch = [] + b = 0 + for d in orig_reader(): + if b >= repeat_batch_size: + break + record_batch.append(d) + b += 1 + while True: + for d in record_batch: + yield d + + +class TestDistMnist2x2(TestDistRunnerBase): + def get_model(self, batch_size=2): + # Input data + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + # Train program + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_size_tensor = fluid.layers.create_tensor(dtype='int64') + batch_acc = fluid.layers.accuracy( + input=predict, label=label, total=batch_size_tensor) + + inference_program = fluid.default_main_program().clone() + # Optimization + opt = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9) + + # Reader + train_reader = paddle.batch(test_merge_reader, batch_size=batch_size) + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + opt.minimize(avg_cost) + return inference_program, avg_cost, train_reader, test_reader, batch_acc, predict + + +if __name__ == "__main__": + runtime_main(TestDistMnist2x2) diff --git a/python/paddle/fluid/tests/unittests/dist_mnist_lars.py b/python/paddle/fluid/tests/unittests/dist_mnist_lars.py new file mode 100644 index 0000000000000000000000000000000000000000..977e17c37f7676ae81d9ab29b6b36089ccbeeacf --- /dev/null +++ b/python/paddle/fluid/tests/unittests/dist_mnist_lars.py @@ -0,0 +1,73 @@ +# 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 numpy as np +import argparse +import time +import math + +import paddle +import paddle.fluid as fluid +import paddle.fluid.profiler as profiler +from paddle.fluid import core +import unittest +from multiprocessing import Process +import os +import signal +from functools import reduce +from test_dist_base import TestDistRunnerBase, runtime_main +from dist_mnist import cnn_model + +DTYPE = "float32" +paddle.dataset.mnist.fetch() + +# Fix seed for test +fluid.default_startup_program().random_seed = 1 +fluid.default_main_program().random_seed = 1 + + +class TestDistMnist2x2(TestDistRunnerBase): + def get_model(self, batch_size=2): + # Input data + images = fluid.layers.data(name='pixel', shape=[1, 28, 28], dtype=DTYPE) + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + + # Train program + predict = cnn_model(images) + cost = fluid.layers.cross_entropy(input=predict, label=label) + avg_cost = fluid.layers.mean(x=cost) + + # Evaluator + batch_size_tensor = fluid.layers.create_tensor(dtype='int64') + batch_acc = fluid.layers.accuracy( + input=predict, label=label, total=batch_size_tensor) + + inference_program = fluid.default_main_program().clone() + # Optimization + opt = fluid.optimizer.LarsMomentumOptimizer( + learning_rate=0.001, momentum=0.9) + + # Reader + train_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + test_reader = paddle.batch( + paddle.dataset.mnist.test(), batch_size=batch_size) + opt.minimize(avg_cost) + return inference_program, avg_cost, train_reader, test_reader, batch_acc, predict + + +if __name__ == "__main__": + runtime_main(TestDistMnist2x2) diff --git a/python/paddle/fluid/tests/unittests/dist_save_load.py b/python/paddle/fluid/tests/unittests/dist_save_load.py new file mode 100644 index 0000000000000000000000000000000000000000..edc60550058f53da456c21de4b41142b907743df --- /dev/null +++ b/python/paddle/fluid/tests/unittests/dist_save_load.py @@ -0,0 +1,174 @@ +# 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 os +import sys +import signal +import subprocess +import argparse +import time +import math +import random +from multiprocessing import Process +from functools import reduce + +import numpy as np +import unittest +import six + +import paddle +import paddle.fluid as fluid +from paddle.fluid import core +from paddle.fluid import io + +from test_dist_base import TestDistRunnerBase, runtime_main, RUN_STEP +from dist_simnet_bow import TestDistSimnetBow2x2, DATA_URL, DATA_MD5 + + +class TestDistSaveLoad2x2(TestDistSimnetBow2x2): + def _load_persistable_vars(self, executor, dirname, program): + def _is_checkpoint_var(var): + """ + the checkpoint will not save or load all the variables. + var type is FEED_MINIBATCH/FETCH_LIST/RAW or var name ends with @GRAD are discarded. + + : param var(Variable) + """ + if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \ + var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \ + var.desc.type() == core.VarDesc.VarType.RAW: + return False + # @GRAD are named for gradient variables, checkpoint will not save it. + if "@GRAD" in var.name: + return False + # .trainer_ are named for distribute train variables, checkpoint will not save it. + if ".trainer_" in var.name: + return False + + # .block is named for distribute train variables, checkpoint will not save it. + if ".block" in var.name: + return False + + if "tmp_" in var.name: + return False + + return var.persistable + + io.load_vars( + executor, + dirname=dirname, + main_program=program, + predicate=_is_checkpoint_var, + filename=None) + + def run_pserver(self, args): + self.get_model(batch_size=2) + # NOTE: pserver should not call memory optimize + t = self.get_transpiler(args.trainer_id, + fluid.default_main_program(), args.endpoints, + args.trainers, args.sync_mode) + pserver_prog = t.get_pserver_program(args.current_endpoint) + startup_prog = t.get_startup_program(args.current_endpoint, + pserver_prog) + + need_load = bool(int(os.getenv("LOAD", "0"))) + model_dir = os.getenv("MODEL_DIR", "") + + place = fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup_prog) + + if need_load and model_dir: + self._load_persistable_vars(exe, model_dir, startup_prog) + exe.run(pserver_prog) + + def run_trainer(self, args): + test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \ + self.get_model(batch_size=2) + + if args.mem_opt: + fluid.memory_optimize(fluid.default_main_program(), skip_grads=True) + if args.is_dist: + t = self.get_transpiler(args.trainer_id, + fluid.default_main_program(), + args.endpoints, args.trainers, + args.sync_mode) + + trainer_prog = t.get_trainer_program() + else: + trainer_prog = fluid.default_main_program() + + if args.use_cuda: + place = fluid.CUDAPlace(0) + else: + place = fluid.CPUPlace() + + startup_exe = fluid.Executor(place) + startup_exe.run(fluid.default_startup_program()) + + strategy = fluid.ExecutionStrategy() + strategy.num_threads = 1 + strategy.allow_op_delay = False + + build_stra = fluid.BuildStrategy() + + if args.use_reduce: + build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce + else: + build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + + exe = fluid.ParallelExecutor( + args.use_cuda, + loss_name=avg_cost.name, + exec_strategy=strategy, + build_strategy=build_stra) + + feed_var_list = [ + var for var in trainer_prog.global_block().vars.values() + if var.is_data + ] + + feeder = fluid.DataFeeder(feed_var_list, place) + reader_generator = train_reader() + + def get_data(): + origin_batch = next(reader_generator) + if args.is_dist and args.use_reader_alloc: + new_batch = [] + for offset, item in enumerate(origin_batch): + if offset % 2 == args.trainer_id: + new_batch.append(item) + return new_batch + else: + return origin_batch + + need_save = bool(int(os.getenv("SAVE", "0"))) + model_dir = os.getenv("MODEL_DIR", "") + + if need_save: + for _ in six.moves.xrange(RUN_STEP): + loss, = exe.run(fetch_list=[avg_cost.name], + feed=feeder.feed(get_data())) + if need_save and model_dir: + io.save_persistables(startup_exe, model_dir, trainer_prog) + + var = np.array(fluid.global_scope().find_var('__fc_b__').get_tensor()) + print(np.ravel(var).tolist()) + + +if __name__ == "__main__": + paddle.dataset.common.download(DATA_URL, 'simnet', DATA_MD5, "train") + runtime_main(TestDistSaveLoad2x2) diff --git a/python/paddle/fluid/tests/unittests/dist_simnet_bow.py b/python/paddle/fluid/tests/unittests/dist_simnet_bow.py index 6456d1b53a129db04ace7ff4413a3d76e922ccde..fac5e037a46715d146e354825f09ee8ccc4f3d70 100644 --- a/python/paddle/fluid/tests/unittests/dist_simnet_bow.py +++ b/python/paddle/fluid/tests/unittests/dist_simnet_bow.py @@ -81,7 +81,10 @@ def get_optimizer(): return optimizer -def train_network(batch_size, is_distributed=False, is_sparse=False): +def train_network(batch_size, + is_distributed=False, + is_sparse=False, + is_self_contained_lr=False): # query q = fluid.layers.data( name="query_ids", shape=[1], dtype="int64", lod_level=1) @@ -93,7 +96,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False): param_attr=fluid.ParamAttr( initializer=fluid.initializer.Constant(value=0.01), name="__emb__", - learning_rate=emb_lr), + learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.01), + name="__emb__"), is_sparse=is_sparse) ## vsum q_sum = fluid.layers.sequence_pool(input=q_emb, pool_type='sum') @@ -119,7 +124,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False): param_attr=fluid.ParamAttr( initializer=fluid.initializer.Constant(value=0.01), name="__emb__", - learning_rate=emb_lr), + learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.01), + name="__emb__"), is_sparse=is_sparse) ## vsum pt_sum = fluid.layers.sequence_pool(input=pt_emb, pool_type='sum') @@ -144,7 +151,9 @@ def train_network(batch_size, is_distributed=False, is_sparse=False): param_attr=fluid.ParamAttr( initializer=fluid.initializer.Constant(value=0.01), name="__emb__", - learning_rate=emb_lr), + learning_rate=emb_lr) if is_self_contained_lr else fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.01), + name="__emb__"), is_sparse=is_sparse) ## vsum nt_sum = fluid.layers.sequence_pool(input=nt_emb, pool_type='sum') @@ -220,7 +229,10 @@ class TestDistSimnetBow2x2(TestDistRunnerBase): def get_model(self, batch_size=2): # Train program avg_cost, acc, predict = \ - train_network(batch_size, bool(int(os.environ["IS_DISTRIBUTED"])), bool(int(os.environ["IS_SPARSE"]))) + train_network(batch_size, + bool(int(os.environ["IS_DISTRIBUTED"])), + bool(int(os.environ["IS_SPARSE"])), + bool(int(os.environ["IS_SELF_CONTAINED_LR"]))) inference_program = fluid.default_main_program().clone() diff --git a/python/paddle/fluid/tests/unittests/dist_transformer.py b/python/paddle/fluid/tests/unittests/dist_transformer.py index a2cc57425841100a2b61279d1b447b88ed4b9a54..27c67edf4f62dd3c5d396826348f8da4513667ba 100644 --- a/python/paddle/fluid/tests/unittests/dist_transformer.py +++ b/python/paddle/fluid/tests/unittests/dist_transformer.py @@ -35,7 +35,7 @@ import paddle import paddle.fluid as fluid import paddle.fluid.layers as layers from paddle.fluid import core -from test_dist_base import TestDistRunnerBase, runtime_main +from test_dist_base import TestDistRunnerBase, runtime_main, RUN_STEP import paddle.compat as cpt from paddle.compat import long_type @@ -562,18 +562,12 @@ def train_loop(exe, train_progm, dev_count, sum_cost, avg_cost, lr_scheduler, for pass_id in six.moves.xrange(TrainTaskConfig.pass_num): pass_start_time = time.time() for batch_id, data in enumerate(train_data()): - if batch_id >= 5: + if batch_id >= RUN_STEP: break feed_list = [] total_num_token = 0 - #if TrainTaskConfig.local: - # lr_rate = lr_scheduler.update_learning_rate() - #for place_id, data_buffer in enumerate( - # split_data( - # data, num_part=dev_count)): - if TrainTaskConfig.local: lr_rate = lr_scheduler.update_learning_rate() @@ -619,12 +613,11 @@ def train_loop(exe, train_progm, dev_count, sum_cost, avg_cost, lr_scheduler, init = True # Validate and save the model for inference. - if batch_id == 0 or batch_id == 4: - if TrainTaskConfig.val_file_pattern is not None: - val_avg_cost, val_ppl = test() - print("[%f]" % val_avg_cost) - else: - assert (False) + if TrainTaskConfig.val_file_pattern is not None: + val_avg_cost, val_ppl = test() + print("[%f]" % val_avg_cost) + else: + assert (False) #import transformer_reader as reader @@ -1166,6 +1159,7 @@ def prepare_encoder(src_word, name=pos_enc_param_name, trainable=False, initializer=fluid.initializer.ConstantInitializer(0.001))) + src_pos_enc.stop_gradient = True enc_input = src_word_emb + src_pos_enc return layers.dropout( enc_input, @@ -1701,7 +1695,7 @@ class DistTransformer2x2(TestDistRunnerBase): def run_trainer(self, args): TrainTaskConfig.use_gpu = args.use_cuda - sum_cost, avg_cost, predict, token_num, local_lr_scheduler = get_model( + sum_cost, avg_cost, predict, token_num, local_lr_scheduler, test_program = get_model( args.is_dist, not args.sync_mode) if args.is_dist: diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index e97643cddef22465436051a41ef4b825e9634d23..690c4cf0ad6b2c741689e419223cfa6b6e1e5cf3 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -54,14 +54,6 @@ def get_numeric_gradient(place, def product(dim): return six.moves.reduce(lambda a, b: a * b, dim, 1) - def get_output(): - sum = [] - op.run(scope, place) - for output_name in output_names: - sum.append( - np.array(scope.find_var(output_name).get_tensor()).mean()) - return np.array(sum).sum() / len(output_names) - tensor_to_check = scope.find_var(input_to_check).get_tensor() tensor_size = product(tensor_to_check.shape()) tensor_to_check_dtype = tensor_to_check._dtype() @@ -77,6 +69,15 @@ def get_numeric_gradient(place, raise ValueError("Not supported data type " + str( tensor_to_check_dtype)) + def get_output(): + sum = [] + op.run(scope, place) + for output_name in output_names: + sum.append( + np.array(scope.find_var(output_name).get_tensor()).astype( + tensor_to_check_dtype).mean()) + return tensor_to_check_dtype(np.array(sum).sum() / len(output_names)) + gradient_flat = np.zeros(shape=(tensor_size, ), dtype=tensor_to_check_dtype) def __get_elem__(tensor, i): 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 ee291fe746f3a1b6ce18df9fb6aa174a89e2eadd..86f861674c26fe61e624103c2a0d70f816a1aebc 100644 --- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -18,6 +18,7 @@ import multiprocessing import os import unittest import paddle.fluid as fluid +import paddle.fluid.core as core import time import numpy as np import math @@ -40,7 +41,8 @@ class TestParallelExecutorBase(unittest.TestCase): use_reduce=False, fuse_elewise_add_act_ops=False, optimizer=fluid.optimizer.Adam, - use_fast_executor=False): + use_fast_executor=False, + enable_sequential_execution=False): def run_executor(exe, feed, fetch_list, program=None): if isinstance(exe, fluid.ParallelExecutor): res = exe.run(fetch_list=fetch_list, feed=feed) @@ -80,6 +82,9 @@ class TestParallelExecutorBase(unittest.TestCase): build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce \ if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops + build_strategy.enable_sequential_execution = enable_sequential_execution + if use_cuda and core.is_compiled_with_cuda(): + build_strategy.remove_unnecessary_lock = True if use_parallel_executor: exe = fluid.ParallelExecutor( diff --git a/python/paddle/fluid/tests/unittests/test_activation_op.py b/python/paddle/fluid/tests/unittests/test_activation_op.py index 30651c1326328180592520447e597aa722146a42..ad7591417ec116a2232bfb7cd94be37a32edfc2e 100644 --- a/python/paddle/fluid/tests/unittests/test_activation_op.py +++ b/python/paddle/fluid/tests/unittests/test_activation_op.py @@ -21,7 +21,7 @@ from op_test import OpTest from scipy.special import expit -class TestExp(OpTest): +class TestActivation(OpTest): def setUp(self): self.op_type = "exp" self.dtype = np.float32 @@ -42,24 +42,12 @@ class TestExp(OpTest): self.check_grad(['X'], 'Out', max_relative_error=0.007) def init_dtype(self): - pass - - -class TestFP16Exp(TestExp): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) + self.dtype = np.float32 -class TestSigmoid(OpTest): +class TestSigmoid(TestActivation): def setUp(self): self.op_type = "sigmoid" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype) @@ -68,33 +56,15 @@ class TestSigmoid(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.01) - def init_dtype(self): - pass - - -class TestFP16Sigmoid(TestSigmoid): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestLogSigmoid(OpTest): +class TestLogSigmoid(TestActivation): def setUp(self): self.op_type = "logsigmoid" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype) @@ -103,33 +73,15 @@ class TestLogSigmoid(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.008) - def init_dtype(self): - pass - - -class TestFP16LogSigmoid(TestLogSigmoid): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestTanh(OpTest): +class TestTanh(TestActivation): def setUp(self): self.op_type = "tanh" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype) @@ -138,33 +90,15 @@ class TestTanh(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - - -class TestFP16Tanh(TestTanh): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestTanhShrink(OpTest): +class TestTanhShrink(TestActivation): def setUp(self): self.op_type = "tanh_shrink" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(0.1, 1, [10, 17]).astype(self.dtype) @@ -173,33 +107,15 @@ class TestTanhShrink(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.008) - def init_dtype(self): - pass - - -class TestFP16TanhShrink(TestTanhShrink): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestHardShrink(OpTest): +class TestHardShrink(TestActivation): def setUp(self): self.op_type = "hard_shrink" - self.dtype = np.float32 self.init_dtype() threshold = 0.5 @@ -211,33 +127,15 @@ class TestHardShrink(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.005) - def init_dtype(self): - pass - - -class TestFP16HardShrink(TestHardShrink): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestSoftShrink(OpTest): +class TestSoftShrink(TestActivation): def setUp(self): self.op_type = "softshrink" - self.dtype = np.float32 self.init_dtype() lambda_val = 0.1 @@ -250,33 +148,15 @@ class TestSoftShrink(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - - -class TestFP16SoftShrink(TestSoftShrink): - def init_dtype(self): - self.dtype = np.float16 - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestSqrt(OpTest): +class TestSqrt(TestActivation): def setUp(self): self.op_type = "sqrt" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype) @@ -285,33 +165,15 @@ class TestSqrt(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - -class TestFP16Sqrt(TestSqrt): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestAbs(OpTest): +class TestAbs(TestActivation): def setUp(self): self.op_type = "abs" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype) @@ -325,33 +187,15 @@ class TestAbs(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - -class TestFP16Abs(TestAbs): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestCeil(OpTest): +class TestCeil(TestActivation): def setUp(self): self.op_type = "ceil" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype) @@ -360,30 +204,14 @@ class TestCeil(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - # The same reason with TestFloor - - def init_dtype(self): + def test_check_grad(self): pass -class TestFP16Ceil(TestCeil): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestFloor(OpTest): +class TestFloor(TestActivation): def setUp(self): self.op_type = "floor" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype) @@ -392,31 +220,16 @@ class TestFloor(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - # the gradient on floor, ceil, round is undefined. # we return zero as gradient, but the numpy return nan - - def init_dtype(self): + # The same reason with TestFloor + def test_check_grad(self): pass -class TestFP16Floor(TestFloor): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestCos(OpTest): +class TestCos(TestActivation): def setUp(self): self.op_type = "cos" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype) @@ -425,33 +238,15 @@ class TestCos(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - - -class TestFP16Cos(TestCos): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestSin(OpTest): +class TestSin(TestActivation): def setUp(self): self.op_type = "sin" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype) @@ -460,33 +255,15 @@ class TestSin(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - - -class TestFP16Sin(TestSin): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestRound(OpTest): +class TestRound(TestActivation): def setUp(self): self.op_type = "round" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype) @@ -495,28 +272,13 @@ class TestRound(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - - def init_dtype(self): + def test_check_grad(self): pass -class TestFP16Round(TestRound): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestRelu(OpTest): +class TestRelu(TestActivation): def setUp(self): self.op_type = "relu" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype) @@ -527,33 +289,15 @@ class TestRelu(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - - -class TestFP16Relu(TestRelu): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestBRelu(OpTest): +class TestBRelu(TestActivation): def setUp(self): self.op_type = "brelu" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [4, 4]).astype(self.dtype) @@ -570,33 +314,15 @@ class TestBRelu(OpTest): self.attrs = {'t_min': t_min, 't_max': t_max} self.outputs = {'Out': t} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.02) - def init_dtype(self): - pass - -class TestFP16BRelu(TestBRelu): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestRelu6(OpTest): +class TestRelu6(TestActivation): def setUp(self): self.op_type = "relu6" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [4, 10]).astype(self.dtype) @@ -610,33 +336,15 @@ class TestRelu6(OpTest): self.attrs = {'threshold': threshold} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.02) - def init_dtype(self): - pass - -class TestFP16Relu6(TestRelu6): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestSoftRelu(OpTest): +class TestSoftRelu(TestActivation): def setUp(self): self.op_type = "soft_relu" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-3, 3, [4, 4]).astype(self.dtype) @@ -653,33 +361,15 @@ class TestSoftRelu(OpTest): self.attrs = {'threshold': threshold} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.02) - def init_dtype(self): - pass - - -class TestFP16SoftRelu(TestSoftRelu): - def init_dtype(self): - self.dtype = np.float16 - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestELU(OpTest): +class TestELU(TestActivation): def setUp(self): self.op_type = "elu" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-3, 3, [4, 4]).astype(self.dtype) @@ -691,33 +381,15 @@ class TestELU(OpTest): self.attrs = {'alpha': alpha} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.02) - def init_dtype(self): - pass - - -class TestFP16ELU(TestELU): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestReciprocal(OpTest): +class TestReciprocal(TestActivation): def setUp(self): self.op_type = "reciprocal" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype) @@ -726,33 +398,15 @@ class TestReciprocal(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.01) - def init_dtype(self): - pass - - -class TestFP16Reciprocal(TestReciprocal): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestLog(OpTest): +class TestLog(TestActivation): def setUp(self): self.op_type = "log" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype) @@ -761,33 +415,15 @@ class TestLog(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - - -class TestFP16Log(TestLog): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestSquare(OpTest): +class TestSquare(TestActivation): def setUp(self): self.op_type = "square" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype) @@ -796,33 +432,15 @@ class TestSquare(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - -class TestFP16Square(TestSquare): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestPow(OpTest): +class TestPow(TestActivation): def setUp(self): self.op_type = "pow" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(1, 2, [11, 17]).astype(self.dtype) @@ -832,33 +450,15 @@ class TestPow(OpTest): self.attrs = {'factor': 3.0} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.02) - def init_dtype(self): - pass - - -class TestFP16Pow(TestPow): - def init_dtype(self): - self.dtype = np.float16 - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=5e-2) - - -class TestSTanh(OpTest): +class TestSTanh(TestActivation): def setUp(self): self.op_type = "stanh" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype) @@ -870,34 +470,17 @@ class TestSTanh(OpTest): self.attrs = {'scale_a': scale_a, 'scale_b': scale_b} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - -class TestFP16STanh(TestSTanh): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestSoftplus(OpTest): +class TestSoftplus(TestActivation): def setUp(self): self.op_type = "softplus" - self.dtype = np.float64 self.init_dtype() + self.dtype = np.float64 x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype) out = np.log(1 + np.exp(x)) @@ -905,33 +488,15 @@ class TestSoftplus(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - - -class TestFP16Softplus(TestSoftplus): - def init_dtype(self): - self.dtype = np.float16 - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestSoftsign(OpTest): +class TestSoftsign(TestActivation): def setUp(self): self.op_type = "softsign" - self.dtype = np.float32 self.init_dtype() x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype) @@ -940,33 +505,15 @@ class TestSoftsign(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.007) - def init_dtype(self): - pass - - -class TestFP16Softsign(TestSoftsign): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestThresholdedRelu(OpTest): +class TestThresholdedRelu(TestActivation): def setUp(self): self.op_type = "thresholded_relu" - self.dtype = np.float32 self.init_dtype() threshold = 0.25 @@ -981,33 +528,15 @@ class TestThresholdedRelu(OpTest): self.attrs = {'threshold': threshold} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=self.relative_error) - def init_dtype(self): - pass - - -class TestFP16ThresholdedRelu(TestThresholdedRelu): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestHardSigmoid(OpTest): +class TestHardSigmoid(TestActivation): def setUp(self): self.op_type = "hard_sigmoid" - self.dtype = np.float32 self.init_dtype() self.relative_error = 0.002 @@ -1030,33 +559,15 @@ class TestHardSigmoid(OpTest): self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(X)} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.002) - def init_dtype(self): - pass - - -class TestFP16HardSigmoid(TestHardSigmoid): - def init_dtype(self): - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - -class TestSwish(OpTest): +class TestSwish(TestActivation): def setUp(self): self.op_type = "swish" - self.dtype = np.float32 self.init_dtype() X = np.random.uniform(0.1, 1, [11, 17]).astype(self.dtype) @@ -1067,28 +578,70 @@ class TestSwish(OpTest): self.attrs = {'beta': beta} self.outputs = {'Out': out} - def test_check_output(self): - self.check_output() - def test_check_grad(self): if self.dtype == np.float16: return self.check_grad(['X'], 'Out', max_relative_error=0.008) - def init_dtype(self): - pass - -class TestFP16Swish(TestSwish): - def init_dtype(self): - self.dtype = np.float16 +#------------------ Test Fp16 ---------------------- +def create_test_act_fp16_class(parent, + atol=1e-3, + grad_check=True, + grad_atol=0.80): + @unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") + class TestActFp16(parent): + def init_dtype(self): + self.dtype = np.float16 - def test_check_output(self): - if core.is_compiled_with_cuda(): + def test_check_output(self): place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) + support_fp16 = core.is_float16_supported(place) + if support_fp16: + self.check_output_with_place(place, atol=atol) + def test_check_grad(self): + place = core.CUDAPlace(0) + support_fp16 = core.is_float16_supported(place) + if support_fp16 and grad_check: + self.check_grad_with_place( + place, ['X'], 'Out', max_relative_error=grad_atol) + + cls_name = "{0}_{1}".format(parent.__name__, "fp16") + TestActFp16.__name__ = cls_name + globals()[cls_name] = TestActFp16 + + +create_test_act_fp16_class(TestActivation) +create_test_act_fp16_class(TestSigmoid) +create_test_act_fp16_class(TestLogSigmoid) +create_test_act_fp16_class(TestTanh) +create_test_act_fp16_class(TestTanhShrink) +create_test_act_fp16_class(TestHardShrink) +create_test_act_fp16_class(TestSoftShrink) +create_test_act_fp16_class(TestSqrt) +create_test_act_fp16_class(TestAbs) +create_test_act_fp16_class(TestCeil, grad_check=False) +create_test_act_fp16_class(TestFloor, grad_check=False) +create_test_act_fp16_class(TestCos, grad_atol=0.85) +create_test_act_fp16_class(TestSin) +create_test_act_fp16_class(TestRound, grad_check=False) +create_test_act_fp16_class(TestRelu) +create_test_act_fp16_class(TestBRelu) +create_test_act_fp16_class(TestRelu6) +create_test_act_fp16_class(TestSoftRelu) +create_test_act_fp16_class(TestELU) +create_test_act_fp16_class(TestReciprocal) +create_test_act_fp16_class(TestLog) +create_test_act_fp16_class(TestSquare) +create_test_act_fp16_class(TestPow, atol=5e-2) +create_test_act_fp16_class(TestSTanh, grad_atol=0.9) +create_test_act_fp16_class(TestSoftplus) +create_test_act_fp16_class(TestSoftsign) +create_test_act_fp16_class(TestThresholdedRelu) +create_test_act_fp16_class(TestHardSigmoid) +create_test_act_fp16_class(TestSwish) if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_add_position_encoding_op.py b/python/paddle/fluid/tests/unittests/test_add_position_encoding_op.py new file mode 100644 index 0000000000000000000000000000000000000000..3f2a33793028f0883ffe94dd8a32626ad5c0351c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_add_position_encoding_op.py @@ -0,0 +1,134 @@ +# 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 unittest +import numpy as np +import math +import paddle.fluid.core as core +from op_test import OpTest + + +class TestAddPositionEncodingTensorOp(OpTest): + """ + This class is to test the AddPositionEncodingOp + """ + + def setUp(self): + """ + the prepared section for add position encoding op + """ + self.op_type = "add_position_encoding" + self.dtype = np.float32 + self.init_input_output() + + self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(self.x), } + self.outputs = {'Out': self.out} + self.attrs = {'alpha': self.alpha, 'beta': self.beta} + + def test_check_output(self): + """ + check the correctness of output + """ + self.check_output() + + def test_check_grad(self): + """ + check the correctness of grad + """ + self.check_grad(['X'], 'Out', max_relative_error=0.005) + + def init_input_output(self): + """ + init the input and output for test cases + """ + self.alpha = 0.6 + self.beta = 0.5 + self.x = np.random.uniform(0.1, 1, [2, 4, 4]).astype(self.dtype) + self.out = np.copy(self.x) + + batch_size = self.x.shape[0] + max_length = self.x.shape[1] + enc_size = self.x.shape[2] + + half_shape = int(enc_size / 2) + for i in range(batch_size): + for j in range(max_length): + for k in range(half_shape): + val = j / pow(10000.0, k / ( + half_shape - 1)) if half_shape > 1 else j / 10000.0 + self.out[i, j, k] = \ + self.x[i, j, k] * self.alpha + math.sin(val) * self.beta + self.out[i, j, half_shape + k] = \ + self.x[i, j, half_shape + k] * self.alpha + math.cos(val) * self.beta + + +class TestAddPositionEncodingLoDTensorOp(OpTest): + """ + This class is to test the AddPositionEncodingLoDTensorOp + """ + + def setUp(self): + """ + the prepared section for add position encoding LoDTensor op + """ + self.op_type = "add_position_encoding" + self.dtype = np.float32 + self.init_input_output() + + self.inputs = {'X': (self.x, self.lod), } + self.outputs = {'Out': (self.out, self.lod)} + self.attrs = {'alpha': self.alpha, 'beta': self.beta} + + def test_check_output(self): + """ + check the correctness of output + """ + self.check_output() + + def test_check_grad(self): + """ + check the correctness of grad + """ + self.check_grad(['X'], 'Out', max_relative_error=0.005) + + def init_input_output(self): + """ + init the input and output for test cases + """ + self.alpha = 0.6 + self.beta = 0.5 + self.x = np.random.uniform(0.1, 1, [10, 4]).astype(self.dtype) + self.lod = [[3, 7]] + self.out = np.copy(self.x) + + batch_size = len(self.lod[0]) + enc_size = self.x.shape[1] + + start = 0 + half_shape = int(enc_size / 2) + for i in range(batch_size): + max_length = self.lod[0][i] + for j in range(max_length): + for k in range(half_shape): + val = j / pow(10000.0, k / ( + half_shape - 1)) if half_shape > 1 else j / 10000.0 + pos = start + j + self.out[pos, k] = \ + self.x[pos, k] * self.alpha + math.sin(val) * self.beta + self.out[pos, half_shape + k] = \ + self.x[pos, half_shape + k] * self.alpha + math.cos(val) * self.beta + start += max_length + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_affine_channel_op.py b/python/paddle/fluid/tests/unittests/test_affine_channel_op.py new file mode 100644 index 0000000000000000000000000000000000000000..2c9a063e6ee75371e0d05e1ff6964753017881a1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_affine_channel_op.py @@ -0,0 +1,106 @@ +# 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 +from op_test import OpTest +import paddle.fluid.core as core + + +def affine_channel(x, scale, bias, layout): + C = x.shape[1] if layout == 'NCHW' else x.shape[-1] + if len(x.shape) == 4: + new_shape = (1, C, 1, 1) if layout == 'NCHW' else (1, 1, 1, C) + else: + new_shape = (1, C) + scale = scale.reshape(new_shape) + bias = bias.reshape(new_shape) + return x * scale + bias + + +class TestAffineChannelOp(OpTest): + def setUp(self): + self.op_type = "affine_channel" + self.init_test_case() + + x = np.random.random(self.shape).astype("float32") + scale = np.random.random(self.C).astype("float32") + bias = np.random.random(self.C).astype("float32") + + y = affine_channel(x, scale, bias, self.layout) + + self.inputs = {'X': x, 'Scale': scale, 'Bias': bias} + self.attrs = {'data_layout': self.layout} + self.outputs = {'Out': y} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X', 'Scale', 'Bias'], 'Out') + + def test_check_grad_stopgrad_dx(self): + self.check_grad(['Scale', 'Bias'], 'Out', no_grad_set=set('X')) + + def test_check_grad_stopgrad_dscale_dbias(self): + self.check_grad(['X'], 'Out', no_grad_set=set(['Scale', 'Bias'])) + + def init_test_case(self): + self.shape = [2, 32, 14, 14] + self.C = 32 + self.layout = 'NCHW' + + +class TestAffineChannelNHWC(TestAffineChannelOp): + def init_test_case(self): + self.shape = [2, 14, 14, 32] + self.C = 32 + self.layout = 'NHWC' + + +class TestAffineChannel2D(TestAffineChannelOp): + def init_test_case(self): + self.shape = [16, 64] + self.C = 64 + self.layout = 'NCHW' + + +class TestAffineChannelNCHWLargeShape(TestAffineChannelOp): + def init_test_case(self): + self.shape = [64, 128, 112, 112] + self.C = 128 + self.layout = 'NCHW' + + # since the gradient check is very slow in large shape, so skip check_grad + def test_check_grad(self): + pass + + def test_check_grad_stopgrad_dx(self): + pass + + def test_check_grad_stopgrad_dscale_dbias(self): + pass + + +class TestAffineChannelNCHWLargeShape(TestAffineChannelNCHWLargeShape): + def init_test_case(self): + self.shape = [64, 112, 112, 512] + self.C = 512 + self.layout = 'NHWC' + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_affine_grid_op.py b/python/paddle/fluid/tests/unittests/test_affine_grid_op.py new file mode 100644 index 0000000000000000000000000000000000000000..576d00940c4c7a5e30af5550e14b674a73e7df11 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_affine_grid_op.py @@ -0,0 +1,79 @@ +# 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 unittest +import numpy as np +from op_test import OpTest + + +def AffineGrid(theta, size): + n = size[0] + w = size[3] + h = size[2] + h_idx = np.repeat( + np.linspace(-1, 1, h)[np.newaxis, :], w, axis=0).T[:, :, np.newaxis] + w_idx = np.repeat( + np.linspace(-1, 1, w)[np.newaxis, :], h, axis=0)[:, :, np.newaxis] + grid = np.concatenate( + [w_idx, h_idx, np.ones([h, w, 1])], axis=2) # h * w * 3 + grid = np.repeat(grid[np.newaxis, :], size[0], axis=0) # n * h * w *3 + + ret = np.zeros([n, h * w, 2]) + theta = theta.transpose([0, 2, 1]) + for i in range(len(theta)): + ret[i] = np.dot(grid[i].reshape([h * w, 3]), theta[i]) + +# print ret.reshape([h * w, 2]).astype("float32") + return ret.reshape([n, h, w, 2]).astype("float32") + + +class TestAffineGridOp(OpTest): + def setUp(self): + self.initTestCase() + self.op_type = "affine_grid" + theta = np.random.randint(1, 3, self.theta_shape).astype("float32") + theta = np.ones(self.theta_shape).astype("float32") + self.inputs = {'Theta': theta} + self.attrs = {"use_cudnn": True} + if self.dynamic_shape: + self.inputs['OutputShape'] = self.output_shape + else: + self.attrs['output_shape'] = self.output_shape + self.outputs = {'Output': AffineGrid(theta, self.output_shape)} + + def test_check_output(self): + self.check_output() + + def test_check_grad_normal(self): + self.check_grad( + ['Theta'], + 'Output', + no_grad_set=['OutputShape'], + max_relative_error=0.006) + + def initTestCase(self): + self.theta_shape = (3, 2, 3) + self.output_shape = np.array([3, 2, 5, 7]).astype("int32") + self.dynamic_shape = False + + +class TestAffineGridOpCase1(TestAffineGridOp): + def initTestCase(self): + self.theta_shape = (3, 2, 3) + self.output_shape = np.array([3, 2, 5, 7]).astype("int32") + self.dynamic_shape = True + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py deleted file mode 100644 index bed847c3c168c906a89c32631b2a8f0ba2e6e7be..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py +++ /dev/null @@ -1,168 +0,0 @@ -# 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 -from op_test import OpTest -import paddle.fluid.core as core - - -def bilinear_interp_np(input, out_h, out_w, out_size): - if out_size is not None: - out_h = out_size[0] - out_w = out_size[1] - batch_size, channel, in_h, in_w = input.shape - if out_h > 1: - ratio_h = (in_h - 1.0) / (out_h - 1.0) - else: - ratio_h = 0.0 - if out_w > 1: - ratio_w = (in_w - 1.0) / (out_w - 1.0) - else: - ratio_w = 0.0 - - out = np.zeros((batch_size, channel, out_h, out_w)) - for i in range(out_h): - h = int(ratio_h * i) - hid = 1 if h < in_h - 1 else 0 - h1lambda = ratio_h * i - h - h2lambda = 1.0 - h1lambda - for j in range(out_w): - w = int(ratio_w * j) - wid = 1 if w < in_w - 1 else 0 - w1lambda = ratio_w * j - w - w2lambda = 1.0 - w1lambda - - out[:, :, i, j] = h2lambda*(w2lambda*input[:, :, h, w] + - w1lambda*input[:, :, h, w+wid]) + \ - h1lambda*(w2lambda*input[:, :, h+hid, w] + - w1lambda*input[:, :, h+hid, w+wid]) - return out.astype(input.dtype) - - -class TestBilinearInterpOp(OpTest): - def setUp(self): - self.out_size = None - self.init_test_case() - self.op_type = "bilinear_interp" - input_np = np.random.random(self.input_shape).astype("float32") - output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, - self.out_size) - self.inputs = {'X': input_np} - if self.out_size is not None: - self.inputs['OutSize'] = self.out_size - self.attrs = {'out_h': self.out_h, 'out_w': self.out_w} - self.outputs = {'Out': output_np} - - def test_check_output(self): - self.check_output() - - def test_check_grad(self): - self.check_grad(['X'], 'Out', in_place=True) - - def init_test_case(self): - self.input_shape = [2, 3, 4, 4] - self.out_h = 2 - self.out_w = 2 - self.out_size = np.array([3, 3]).astype("int32") - - -class TestCase1(TestBilinearInterpOp): - def init_test_case(self): - self.input_shape = [4, 1, 7, 8] - self.out_h = 1 - self.out_w = 1 - - -class TestCase2(TestBilinearInterpOp): - def init_test_case(self): - self.input_shape = [3, 3, 9, 6] - self.out_h = 12 - self.out_w = 12 - - -class TestCase3(TestBilinearInterpOp): - def init_test_case(self): - self.input_shape = [1, 1, 128, 64] - self.out_h = 64 - self.out_w = 128 - - -class TestCase4(TestBilinearInterpOp): - def init_test_case(self): - self.input_shape = [4, 1, 7, 8] - self.out_h = 1 - self.out_w = 1 - self.out_size = np.array([2, 2]).astype("int32") - - -class TestCase5(TestBilinearInterpOp): - def init_test_case(self): - self.input_shape = [3, 3, 9, 6] - self.out_h = 12 - self.out_w = 12 - self.out_size = np.array([11, 11]).astype("int32") - - -class TestCase6(TestBilinearInterpOp): - def init_test_case(self): - self.input_shape = [1, 1, 128, 64] - self.out_h = 64 - self.out_w = 128 - self.out_size = np.array([65, 129]).astype("int32") - - -class TestBilinearInterpOpUint8(OpTest): - def setUp(self): - self.out_size = None - self.init_test_case() - self.op_type = "bilinear_interp" - input_np = np.random.randint( - low=0, high=256, size=self.input_shape).astype("uint8") - output_np = bilinear_interp_np(input_np, self.out_h, self.out_w, - self.out_size) - self.inputs = {'X': input_np} - if self.out_size is not None: - self.inputs['OutSize'] = self.out_size - self.attrs = {'out_h': self.out_h, 'out_w': self.out_w} - self.outputs = {'Out': output_np} - - def test_check_output(self): - self.check_output_with_place(place=core.CPUPlace(), atol=1) - - def init_test_case(self): - self.input_shape = [1, 3, 9, 6] - self.out_h = 10 - self.out_w = 9 - - -class TestCase1Uint8(TestBilinearInterpOpUint8): - def init_test_case(self): - self.input_shape = [2, 3, 128, 64] - self.out_h = 120 - self.out_w = 50 - - -class TestCase2Uint8(TestBilinearInterpOpUint8): - def init_test_case(self): - self.input_shape = [4, 1, 7, 8] - self.out_h = 5 - self.out_w = 13 - self.out_size = np.array([6, 15]).astype("int32") - - -if __name__ == "__main__": - unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py b/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py index 6103c3aafc0bb154194314830c5c8c5d89460cfe..46433d78252219fe02c3c4b5ddfc525bda177f18 100644 --- a/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py @@ -18,6 +18,9 @@ import unittest import numpy as np from op_test import OpTest +import paddle.fluid as fluid +import paddle.fluid.core as core + class TestClipByNormOp(OpTest): def setUp(self): @@ -62,5 +65,59 @@ class TestCase3(TestClipByNormOp): self.max_norm = 1.0 +class TestClipByNormOpWithSelectedRows(OpTest): + def check_with_place(self, place): + self.config_test_case() + scope = core.Scope() + + # set input + x_selected_rows = scope.var('X').get_selected_rows() + x_selected_rows.set_rows(self.grad_rows) + x_tensor = x_selected_rows.get_tensor() + x_np = np.random.random(self.grad_shape).astype("float32") + x_np[np.abs(x_np) < self.max_relative_error] = 0.5 + x_tensor.set(x_np, place) + + # set output + out_selected_rows = scope.var('Out').get_selected_rows() + + # run clip_by_norm_op + clip_by_norm_op = fluid.op.Operator( + "clip_by_norm", max_norm=self.max_norm, X='X', Out='Out') + clip_by_norm_op.run(scope, place) + + # check output + self.assertEqual(out_selected_rows.rows(), self.grad_clipped_rows) + out_tensor = out_selected_rows.get_tensor() + y_np = np.zeros(self.grad_clipped_shape) + y_np[0] = np.sum(x_np[0:2]) + y_np[1] = x_np[2] + y_np[2] = x_np[3] + norm = np.sqrt(np.sum(np.square(y_np))) + if norm > self.max_norm: + output = self.max_norm * y_np / norm + else: + output = y_np + self.assertTrue( + np.allclose( + np.array(out_tensor), output, atol=1e-5, equal_nan=False)) + + def test_clip_by_norm_with_selected_ros(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + + for place in places: + self.check_with_place(place) + + def config_test_case(self): + self.max_norm = 1.0 + self.max_relative_error = 0.006 + self.grad_shape = (4, 1) + self.grad_clipped_shape = (3, 1) + self.grad_rows = [0, 0, 1, 2] + self.grad_clipped_rows = [0, 1, 2] + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_op.py index 2ecc2504a8c9c5ecfc32cee96df9e368ff219cbb..ebbbf3ab8b00ff49d55ea5d472a2f7c4eae0da52 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_op.py @@ -67,6 +67,7 @@ class TestConv2dOp(OpTest): def setUp(self): self.op_type = "conv2d" self.use_cudnn = False + self.exhaustive_search = False self.use_cuda = False self.use_mkldnn = False self.data_format = "AnyLayout" @@ -98,7 +99,8 @@ class TestConv2dOp(OpTest): 'dilations': self.dilations, 'use_cudnn': self.use_cudnn, 'use_mkldnn': self.use_mkldnn, - 'data_format': self.data_format + 'data_format': self.data_format, + 'exhaustive_search': self.exhaustive_search } self.outputs = {'Output': output} @@ -223,106 +225,75 @@ class TestWithInput1x1Filter1x1(TestConv2dOp): #----------------Conv2dCUDNN---------------- -class TestCUDNN(TestConv2dOp): - def init_kernel_type(self): - self.use_cudnn = True -class TestFP16CUDNN(TestConv2dOp): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=2e-2) - +def create_test_cudnn_class(parent): + @unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") + class TestCUDNNCase(parent): + def init_kernel_type(self): + self.use_cudnn = True -class TestCUDNNWithPad(TestWithPad): - def init_kernel_type(self): - self.use_cudnn = True + cls_name = "{0}_{1}".format(parent.__name__, "CUDNN") + TestCUDNNCase.__name__ = cls_name + globals()[cls_name] = TestCUDNNCase -class TestFP16CUDNNWithPad(TestWithPad): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 +create_test_cudnn_class(TestConv2dOp) +create_test_cudnn_class(TestWithPad) +create_test_cudnn_class(TestWithStride) +create_test_cudnn_class(TestWithGroup) +create_test_cudnn_class(TestWith1x1) +create_test_cudnn_class(TestWithInput1x1Filter1x1) - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=2e-2) +#----------------Conv2dCUDNN---------------- -class TestCUDNNWithStride(TestWithStride): - def init_kernel_type(self): - self.use_cudnn = True +def create_test_cudnn_fp16_class(parent, grad_check=True): + @unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") + class TestConv2DCUDNNFp16(parent): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) -class TestFP16CUDNNWithStride(TestWithStride): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): + def test_check_grad_no_filter(self): place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=2e-2) - - -class TestCUDNNWithGroup(TestWithGroup): - def init_kernel_type(self): - self.use_cudnn = True - - -class TestFP16CUDNNWithGroup(TestWithGroup): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=2e-2) - - -class TestCUDNNWith1x1(TestWith1x1): - def init_kernel_type(self): - self.use_cudnn = True - - -class TestFP16CUDNNWith1x1(TestWith1x1): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): + if core.is_float16_supported(place) and grad_check: + self.check_grad_with_place( + place, ['Input'], + 'Output', + max_relative_error=0.02, + no_grad_set=set(['Filter'])) + + def test_check_grad_no_input(self): place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=2e-2) + if core.is_float16_supported(place) and grad_check: + self.check_grad_with_place( + place, ['Filter'], + 'Output', + max_relative_error=0.02, + no_grad_set=set(['Input'])) + cls_name = "{0}_{1}".format(parent.__name__, "CUDNNFp16") + TestConv2DCUDNNFp16.__name__ = cls_name + globals()[cls_name] = TestConv2DCUDNNFp16 -class TestCUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1): - def init_kernel_type(self): - self.use_cudnn = True +create_test_cudnn_fp16_class(TestConv2dOp, grad_check=False) +create_test_cudnn_fp16_class(TestWithPad, grad_check=False) +create_test_cudnn_fp16_class(TestWithStride, grad_check=False) +create_test_cudnn_fp16_class(TestWithGroup, grad_check=False) +create_test_cudnn_fp16_class(TestWith1x1, grad_check=False) +create_test_cudnn_fp16_class(TestWithInput1x1Filter1x1, grad_check=False) -class TestFP16CUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=2e-2) +# -------TestDepthwiseConv class TestDepthwiseConv(TestConv2dOp): @@ -392,6 +363,12 @@ class TestDepthwiseConvWithDilation2(TestConv2dOp): self.op_type = "depthwise_conv2d" +class TestCUDNNExhaustiveSearch(TestConv2dOp): + def init_kernel_type(self): + self.use_cudnn = True + self.exhaustive_search = True + + # Please Don't remove the following code. # Currently, CI use cudnn V5.0 which not support dilation conv. # class TestCUDNNWithDilation(TestWithDilation): diff --git a/python/paddle/fluid/tests/unittests/test_conv3d_op.py b/python/paddle/fluid/tests/unittests/test_conv3d_op.py index ddaf99fe061205f0f2e4c592c9e28e27e657c16a..69c5ab7a4a4cbd552d27dcb07052d46752eeb54a 100644 --- a/python/paddle/fluid/tests/unittests/test_conv3d_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv3d_op.py @@ -335,6 +335,12 @@ class TestFP16WithInput1x1Filter1x1CUDNN(TestWithInput1x1Filter1x1): self.check_output_with_place(place, atol=2e-2) +class TestCUDNNExhaustiveSearch(TestCUDNN): + def init_kernel_type(self): + self.use_cudnn = True + self.exhaustive_search = True + + # FIXME(typhoonzero): find a way to determine if # using cudnn > 6 in python # class TestWithDilationCUDNN(TestWithDilation): diff --git a/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py b/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py index f22badbea0c67b210f7ac4e14e5d647f1cffa6cc..4bdc6403cb4fde2b1f4efd957e922b7ea5cd8f38 100644 --- a/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py +++ b/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py @@ -16,28 +16,58 @@ from __future__ import print_function import unittest import numpy as np +import paddle.fluid.core as core from op_test import OpTest, randomize_probability -class TestCrossEntropyOp1(OpTest): +class TestCrossEntropyOp(OpTest): """Test cross-entropy with discrete one-hot labels. """ def setUp(self): self.op_type = "cross_entropy" - batch_size = 30 - class_num = 10 + self.soft_label = False + self.ignore_index = -100 + self.dtype = np.float64 + self.batch_size = 30 + self.class_num = 10 + + self.init_dtype_type() + self.init_attr_type() + self.init_bs_class_num() + self.init_x() + self.init_label() + self.get_cross_entropy() + + self.inputs = {"X": self.x, "Label": self.label} + self.outputs = {"Y": self.cross_entropy} + self.attrs = { + "soft_label": self.soft_label, + "ignore_index": self.ignore_index + } + + def init_x(self): + self.x = randomize_probability( + self.batch_size, self.class_num, dtype=self.dtype) + + def init_label(self): + self.label = np.random.randint( + 0, self.class_num, (self.batch_size, 1), dtype="int64") + + def get_cross_entropy(self): + self.cross_entropy = np.asmatrix( + [[-np.log(self.x[i][self.label[i][0]])] + for i in range(self.x.shape[0])], + dtype="float64") - X = randomize_probability(batch_size, class_num, dtype='float64') + def init_attr_type(self): + pass - label = np.random.randint(0, class_num, (batch_size, 1), dtype="int64") - cross_entropy = np.asmatrix( - [[-np.log(X[i][label[i][0]])] for i in range(X.shape[0])], - dtype="float64") + def init_dtype_type(self): + pass - self.inputs = {"X": X, "Label": label} - self.outputs = {"Y": cross_entropy} - self.attrs = {"soft_label": False} + def init_bs_class_num(self): + pass def test_check_output(self): self.check_output() @@ -46,197 +76,231 @@ class TestCrossEntropyOp1(OpTest): self.check_grad(["X"], "Y", numeric_grad_delta=0.001) -class TestCrossEntropyOp2(OpTest): +class TestCrossEntropyOp2(TestCrossEntropyOp): """Test cross-entropy with vectorized soft labels. """ - def setUp(self): - self.op_type = "cross_entropy" - batch_size = 5 - class_num = 37 + def init_label(self): + self.label = np.random.uniform( + 0.1, 1.0, [self.batch_size, self.class_num]).astype(self.dtype) + self.label /= self.label.sum(axis=1, keepdims=True) - X = randomize_probability(batch_size, class_num) - label = np.random.uniform(0.1, 1.0, - [batch_size, class_num]).astype("float32") - label /= label.sum(axis=1, keepdims=True) - cross_entropy = (-label * np.log(X)).sum( - axis=1, keepdims=True).astype("float32") + def get_cross_entropy(self): + self.cross_entropy = (-self.label * np.log(self.x)).sum( + axis=1, keepdims=True).astype(self.dtype) - self.inputs = {"X": X, "Label": label} - self.outputs = {"Y": cross_entropy} - self.attrs = {"soft_label": True} + def init_attr_type(self): + self.soft_label = True - def test_check_output(self): - self.check_output() + def init_dtype_type(self): + self.dtype = np.float32 + + def init_bs_class_num(self): + self.batch_size = 5 + self.class_num = 37 def test_check_grad(self): self.check_grad( ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001) -class TestCrossEntropyOp3(OpTest): +class TestCrossEntropyOp3(TestCrossEntropyOp): """Test cross-entropy with vectorized one-hot representation of labels. """ - def setUp(self): - self.op_type = "cross_entropy" - batch_size = 5 - class_num = 17 + def init_label(self): + self.label_index = np.random.randint(0, self.class_num, + (self.batch_size)) + self.label = np.zeros(self.x.shape).astype(self.dtype) + self.label[np.arange(self.batch_size), self.label_index] = 1 - X = randomize_probability(batch_size, class_num) - label_index = np.random.randint( - 0, class_num, (batch_size), dtype="int32") - label = np.zeros(X.shape) - label[np.arange(batch_size), label_index] = 1 + def get_cross_entropy(self): + self.cross_entropy = np.asmatrix( + [[-np.log(self.x[i][self.label_index[i]])] + for i in range(self.x.shape[0])]).astype(self.dtype) - cross_entropy = np.asmatrix( - [[-np.log(X[i][label_index[i]])] for i in range(X.shape[0])], - dtype="float32") - cross_entropy2 = (-label * np.log(X)).sum( - axis=1, keepdims=True).astype("float32") + def init_attr_type(self): + self.soft_label = True - self.inputs = {"X": X, "Label": label.astype(np.float32)} - self.outputs = {"Y": cross_entropy} - self.attrs = {"soft_label": True} + def init_dtype_type(self): + self.dtype = np.float32 - def test_check_output(self): - self.check_output() + def init_bs_class_num(self): + self.batch_size = 5 + self.class_num = 17 def test_check_grad(self): self.check_grad( ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001) -class TestCrossEntropyOp4(OpTest): +class TestCrossEntropyOp4(TestCrossEntropyOp): """Test high rank tensor cross-entropy with discrete one-hot labels. """ - def setUp(self): - self.op_type = "cross_entropy" - shape = [10, 2, 4] - ins_num = np.prod(np.array(shape)) - class_num = 10 + def init_x(self): + self.shape = [10, 2, 4] + self.ins_num = np.prod(np.array(self.shape)) + self.X_2d = randomize_probability(self.ins_num, + self.class_num).astype(self.dtype) + self.x = self.X_2d.reshape(self.shape + [self.class_num]) - X_2d = randomize_probability(ins_num, class_num, dtype='float64') + def init_label(self): + self.label_2d = np.random.randint( + 0, self.class_num, (self.ins_num, 1), dtype="int64") + self.label = self.label_2d.reshape(self.shape + [1]) - label_2d = np.random.randint(0, class_num, (ins_num, 1), dtype="int64") + def get_cross_entropy(self): cross_entropy_2d = np.asmatrix( - [[-np.log(X_2d[i][label_2d[i][0]])] for i in range(X_2d.shape[0])], - dtype="float64") + [[-np.log(self.X_2d[i][self.label_2d[i][0]])] + for i in range(self.X_2d.shape[0])]).astype(self.dtype) + self.cross_entropy = np.array(cross_entropy_2d).reshape(self.shape + + [1]) - X = X_2d.reshape(shape + [class_num]) - label = label_2d.reshape(shape + [1]) - cross_entropy = np.array(cross_entropy_2d).reshape(shape + [1]) + def init_attr_type(self): + self.soft_label = False - self.inputs = {"X": X, "Label": label} - self.outputs = {"Y": cross_entropy} - self.attrs = {"soft_label": False} - - def test_check_output(self): - self.check_output() + def init_dtype_type(self): + self.dtype = np.float64 - def test_check_grad(self): - self.check_grad(["X"], "Y", numeric_grad_delta=0.001) + def init_bs_class_num(self): + self.class_num = 10 -class TestCrossEntropyOp5(OpTest): +class TestCrossEntropyOp5(TestCrossEntropyOp): """Test high rank tensor cross-entropy with vectorized soft labels. """ - def setUp(self): - self.op_type = "cross_entropy" - shape = [4, 3] - ins_num = np.prod(np.array(shape)) - class_num = 37 + def init_x(self): + self.shape = [4, 3] + self.ins_num = np.prod(np.array(self.shape)) + self.X_2d = randomize_probability(self.ins_num, + self.class_num).astype(self.dtype) + self.x = self.X_2d.reshape(self.shape + [self.class_num]) - X_2d = randomize_probability(ins_num, class_num) - label_2d = np.random.uniform(0.1, 1.0, - [ins_num, class_num]).astype("float32") - label_2d /= label_2d.sum(axis=1, keepdims=True) - cross_entropy_2d = (-label_2d * np.log(X_2d)).sum( - axis=1, keepdims=True).astype("float32") + def init_label(self): + self.label_2d = np.random.uniform( + 0.1, 1.0, [self.ins_num, self.class_num]).astype(self.dtype) + self.label_2d /= self.label_2d.sum(axis=1, keepdims=True) + self.label = self.label_2d.reshape(self.shape + [self.class_num]) - X = X_2d.reshape(shape + [class_num]) - label = label_2d.reshape(shape + [class_num]) - cross_entropy = np.array(cross_entropy_2d).reshape(shape + [1]) + def get_cross_entropy(self): + cross_entropy_2d = (-self.label_2d * np.log(self.X_2d)).sum( + axis=1, keepdims=True).astype(self.dtype) + self.cross_entropy = np.array(cross_entropy_2d).reshape(self.shape + + [1]) - self.inputs = {"X": X, "Label": label} - self.outputs = {"Y": cross_entropy} - self.attrs = {"soft_label": True} + def init_attr_type(self): + self.soft_label = True - def test_check_output(self): - self.check_output() + def init_dtype_type(self): + self.dtype = np.float32 + + def init_bs_class_num(self): + self.class_num = 37 def test_check_grad(self): self.check_grad( ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001) -class TestCrossEntropyOp6(OpTest): +class TestCrossEntropyOp6(TestCrossEntropyOp): """Test high rank tensor cross-entropy with vectorized one-hot representation of labels. """ - def setUp(self): - self.op_type = "cross_entropy" - shape = [4, 3, 2] - ins_num = np.prod(np.array(shape)) - class_num = 17 - - X_2d = randomize_probability(ins_num, class_num) - label_index_2d = np.random.randint( - 0, class_num, (ins_num), dtype="int32") - label_2d = np.zeros(X_2d.shape) - label_2d[np.arange(ins_num), label_index_2d] = 1 - + def init_x(self): + self.shape = [4, 3, 2] + self.ins_num = np.prod(np.array(self.shape)) + self.X_2d = randomize_probability(self.ins_num, + self.class_num).astype(self.dtype) + self.x = self.X_2d.reshape(self.shape + [self.class_num]) + + def init_label(self): + self.label_index_2d = np.random.randint( + 0, self.class_num, (self.ins_num), dtype="int64") + label_2d = np.zeros(self.X_2d.shape) + label_2d[np.arange(self.ins_num), self.label_index_2d] = 1 + self.label = label_2d.reshape(self.shape + [self.class_num]).astype( + self.dtype) + + def get_cross_entropy(self): cross_entropy_2d = np.asmatrix( - [[-np.log(X_2d[i][label_index_2d[i]])] - for i in range(X_2d.shape[0])], - dtype="float32") + [[-np.log(self.X_2d[i][self.label_index_2d[i]])] + for i in range(self.X_2d.shape[0])]) + self.cross_entropy = np.array(cross_entropy_2d).reshape( + self.shape + [1]).astype(self.dtype) - X = X_2d.reshape(shape + [class_num]) - label = label_2d.reshape(shape + [class_num]) - cross_entropy = np.array(cross_entropy_2d).reshape(shape + [1]) + def init_attr_type(self): + self.soft_label = True - self.inputs = {"X": X, "Label": label.astype(np.float32)} - self.outputs = {"Y": cross_entropy} - self.attrs = {"soft_label": True} + def init_dtype_type(self): + self.dtype = np.float32 - def test_check_output(self): - self.check_output() + def init_bs_class_num(self): + self.class_num = 17 def test_check_grad(self): self.check_grad( ["X"], "Y", max_relative_error=0.05, numeric_grad_delta=0.001) -class TestCrossEntropyOp7(OpTest): +class TestCrossEntropyOp7(TestCrossEntropyOp): """Test cross-entropy with ignore index. """ - def setUp(self): - self.op_type = "cross_entropy" - batch_size = 30 - class_num = 10 - ignore_index = 3 - - X = randomize_probability(batch_size, class_num, dtype='float64') - - label = np.random.randint(0, class_num, (batch_size, 1), dtype="int64") - cross_entropy = np.asmatrix( - [[-np.log(X[i][label[i][0]])] - if label[i][0] != ignore_index else [0] - for i in range(X.shape[0])], - dtype="float64") - self.inputs = {"X": X, "Label": label} - self.outputs = {"Y": cross_entropy} - self.attrs = {"soft_label": False, "ignore_index": ignore_index} - - def test_check_output(self): - self.check_output() - - def test_check_grad(self): - self.check_grad(["X"], "Y", numeric_grad_delta=0.001) - + def init_label(self): + self.label = np.random.randint( + 0, self.class_num, (self.batch_size, 1), dtype="int64") + + def get_cross_entropy(self): + self.cross_entropy = np.asmatrix( + [[-np.log(self.x[i][self.label[i][0]])] + if self.label[i][0] != self.ignore_index else [0] + for i in range(self.x.shape[0])]).astype(self.dtype) + + def init_attr_type(self): + self.soft_label = False + self.ignore_index = 3 + + def init_dtype_type(self): + self.dtype = np.float64 + + def init_bs_class_num(self): + self.batch_size = 30 + self.class_num = 10 + + +# Add Fp16 test +def create_test_class(parent, cls_name): + @unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") + class TestCrossEntropyFP16Op(parent): + def init_dtype_type(self): + return np.float16 + + def test_check_output(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-1) + + def test_check_grad(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['X'], 'Y', max_relative_error=0.9) + + cls_name = "{0}".format(cls_name) + TestCrossEntropyFP16Op.__name__ = cls_name + globals()[cls_name] = TestCrossEntropyFP16Op + + +create_test_class(TestCrossEntropyOp, "TestCrossEntropyF16Op") +#create_test_class(TestCrossEntropyOp2, "TestCrossEntropyF16Op2") +create_test_class(TestCrossEntropyOp3, "TestCrossEntropyF16Op3") +create_test_class(TestCrossEntropyOp4, "TestCrossEntropyF16Op4") +#create_test_class(TestCrossEntropyOp5, "TestCrossEntropyF16Op5") +create_test_class(TestCrossEntropyOp6, "TestCrossEntropyF16Op6") +create_test_class(TestCrossEntropyOp7, "TestCrossEntropyF16Op7") if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_density_prior_box_op.py b/python/paddle/fluid/tests/unittests/test_density_prior_box_op.py new file mode 100644 index 0000000000000000000000000000000000000000..79d1fd3d7171e06a88a75cf50b6a51ef4da51f07 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_density_prior_box_op.py @@ -0,0 +1,142 @@ +# 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 sys +import math +from op_test import OpTest + + +class TestDensityPriorBoxOp(OpTest): + def set_data(self): + self.init_test_params() + self.init_test_input() + self.init_test_output() + self.inputs = {'Input': self.input, 'Image': self.image} + + self.attrs = { + 'variances': self.variances, + 'clip': self.clip, + 'step_w': self.step_w, + 'step_h': self.step_h, + 'offset': self.offset, + 'densities': self.densities, + 'fixed_sizes': self.fixed_sizes, + 'fixed_ratios': self.fixed_ratios + } + self.outputs = {'Boxes': self.out_boxes, 'Variances': self.out_var} + + def test_check_output(self): + self.check_output() + + def setUp(self): + self.op_type = "density_prior_box" + self.set_data() + + def set_density(self): + self.densities = [] + self.fixed_sizes = [] + self.fixed_ratios = [] + + def init_test_params(self): + self.layer_w = 32 + self.layer_h = 32 + + self.image_w = 40 + self.image_h = 40 + + self.step_w = float(self.image_w) / float(self.layer_w) + self.step_h = float(self.image_h) / float(self.layer_h) + + self.input_channels = 2 + self.image_channels = 3 + self.batch_size = 10 + + self.variances = [0.1, 0.1, 0.2, 0.2] + self.variances = np.array(self.variances, dtype=np.float).flatten() + + self.set_density() + + self.clip = True + self.num_priors = 0 + if len(self.fixed_sizes) > 0 and len(self.densities) > 0: + for density in self.densities: + if len(self.fixed_ratios) > 0: + self.num_priors += len(self.fixed_ratios) * (pow(density, + 2)) + self.offset = 0.5 + + def init_test_input(self): + self.image = np.random.random( + (self.batch_size, self.image_channels, self.image_w, + self.image_h)).astype('float32') + + self.input = np.random.random( + (self.batch_size, self.input_channels, self.layer_w, + self.layer_h)).astype('float32') + + def init_test_output(self): + out_dim = (self.layer_h, self.layer_w, self.num_priors, 4) + out_boxes = np.zeros(out_dim).astype('float32') + out_var = np.zeros(out_dim).astype('float32') + + step_average = int((self.step_w + self.step_h) * 0.5) + for h in range(self.layer_h): + for w in range(self.layer_w): + idx = 0 + c_x = (w + self.offset) * self.step_w + c_y = (h + self.offset) * self.step_h + # Generate density prior boxes with fixed size + for density, fixed_size in zip(self.densities, + self.fixed_sizes): + if (len(self.fixed_ratios) > 0): + for ar in self.fixed_ratios: + shift = int(step_average / density) + box_width_ratio = fixed_size * math.sqrt(ar) + box_height_ratio = fixed_size / math.sqrt(ar) + for di in range(density): + for dj in range(density): + c_x_temp = c_x - step_average / 2.0 + shift / 2.0 + dj * shift + c_y_temp = c_y - step_average / 2.0 + shift / 2.0 + di * shift + out_boxes[h, w, idx, :] = [ + max((c_x_temp - box_width_ratio / 2.0) / + self.image_w, 0), + max((c_y_temp - box_height_ratio / 2.0) + / self.image_h, 0), + min((c_x_temp + box_width_ratio / 2.0) / + self.image_w, 1), + min((c_y_temp + box_height_ratio / 2.0) + / self.image_h, 1) + ] + idx += 1 + if self.clip: + out_boxes = np.clip(out_boxes, 0.0, 1.0) + out_var = np.tile(self.variances, + (self.layer_h, self.layer_w, self.num_priors, 1)) + self.out_boxes = out_boxes.astype('float32') + self.out_var = out_var.astype('float32') + + +class TestDensityPriorBox(TestDensityPriorBoxOp): + def set_density(self): + self.densities = [3, 4] + self.fixed_sizes = [1.0, 2.0] + self.fixed_ratios = [1.0] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py index 04924bec057e301bfb342a62bb4c1e0b3c3aff4c..4b8a215190a90c974a9ecc8658d044c59b80c989 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_base.py +++ b/python/paddle/fluid/tests/unittests/test_dist_base.py @@ -22,22 +22,30 @@ import signal import subprocess import six import argparse +import pickle +import numpy as np import paddle.fluid as fluid RUN_STEP = 10 +DEFAULT_BATCH_SIZE = 2 class TestDistRunnerBase(object): - def get_model(self, batch_size=2): + def get_model(self, batch_size=DEFAULT_BATCH_SIZE): raise NotImplementedError( "get_model should be implemented by child classes.") @staticmethod - def get_transpiler(trainer_id, main_program, pserver_endpoints, trainers, - sync_mode): + def get_transpiler(trainer_id, + main_program, + pserver_endpoints, + trainers, + sync_mode, + dc_asgd=False): # NOTE: import fluid until runtime, or else forking processes will cause error. config = fluid.DistributeTranspilerConfig() + config.enable_dc_asgd = dc_asgd t = fluid.DistributeTranspiler(config=config) t.transpile( trainer_id=trainer_id, @@ -48,12 +56,11 @@ class TestDistRunnerBase(object): return t def run_pserver(self, args): - - self.get_model(batch_size=2) + self.get_model(batch_size=args.batch_size) # NOTE: pserver should not call memory optimize t = self.get_transpiler(args.trainer_id, fluid.default_main_program(), args.endpoints, - args.trainers, args.sync_mode) + args.trainers, args.sync_mode, args.dc_asgd) pserver_prog = t.get_pserver_program(args.current_endpoint) startup_prog = t.get_startup_program(args.current_endpoint, pserver_prog) @@ -65,7 +72,7 @@ class TestDistRunnerBase(object): def run_trainer(self, args): test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \ - self.get_model(batch_size=2) + self.get_model(batch_size=args.batch_size) if args.mem_opt: fluid.memory_optimize(fluid.default_main_program(), skip_grads=True) @@ -73,8 +80,7 @@ class TestDistRunnerBase(object): t = self.get_transpiler(args.trainer_id, fluid.default_main_program(), args.endpoints, args.trainers, - args.sync_mode) - + args.sync_mode, args.dc_asgd) trainer_prog = t.get_trainer_program() else: trainer_prog = fluid.default_main_program() @@ -98,6 +104,12 @@ class TestDistRunnerBase(object): else: build_stra.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce + if args.batch_merge_repeat > 1: + pass_builder = build_stra._create_passes_from_strategy() + mypass = pass_builder.insert_pass( + len(pass_builder.all_passes()) - 2, "multi_batch_merge_pass") + mypass.set_int("num_repeats", args.batch_merge_repeat) + exe = fluid.ParallelExecutor( args.use_cuda, loss_name=avg_cost.name, @@ -123,10 +135,15 @@ class TestDistRunnerBase(object): else: return origin_batch + out_losses = [] for _ in six.moves.xrange(RUN_STEP): loss, = exe.run(fetch_list=[avg_cost.name], feed=feeder.feed(get_data())) - print(loss) + out_losses.append(loss[0]) + if six.PY2: + print(pickle.dumps(out_losses)) + else: + sys.stdout.buffer.write(pickle.dumps(out_losses)) def runtime_main(test_class): @@ -143,8 +160,12 @@ def runtime_main(test_class): parser.add_argument('--mem_opt', action='store_true') parser.add_argument('--use_cuda', action='store_true') parser.add_argument('--use_reduce', action='store_true') + parser.add_argument('--dc_asgd', action='store_true') parser.add_argument( - '--use_reader_alloc', action='store_true', required=False, default=True) + '--use_reader_alloc', action='store_true', required=False) + parser.add_argument('--batch_size', required=False, type=int, default=2) + parser.add_argument( + '--batch_merge_repeat', required=False, type=int, default=1) args = parser.parse_args() @@ -180,11 +201,12 @@ class TestDistBase(unittest.TestCase): self._pservers = 2 self._ps_endpoints = "127.0.0.1:%s,127.0.0.1:%s" % ( self._find_free_port(), self._find_free_port()) - self._python_interp = "python" + self._python_interp = sys.executable self._sync_mode = True self._enforce_place = None self._mem_opt = False self._use_reduce = False + self._dc_asgd = False # must use with async mode self._use_reader_alloc = True self._setup_config() self._after_setup_config() @@ -229,24 +251,18 @@ class TestDistBase(unittest.TestCase): return ps0_proc, ps1_proc, ps0_pipe, ps1_pipe - def _wait_ps_ready(self, pid): - retry_times = 50 - while True: - assert retry_times >= 0, "wait ps ready failed" - time.sleep(3) - try: - # the listen_and_serv_op would touch a file which contains the listen port - # on the /tmp directory until it was ready to process all the RPC call. - os.stat("/tmp/paddle.%d.port" % pid) - return - except os.error as e: - sys.stderr.write('waiting for pserver: %s, left retry %d\n' % - (e, retry_times)) - retry_times -= 1 - - def _run_local(self, model, envs, check_error_log): + def _run_local(self, + model, + envs, + check_error_log=False, + batch_size=DEFAULT_BATCH_SIZE, + batch_merge_repeat=1): cmd = "%s %s --role trainer" % (self._python_interp, model) + if batch_size != DEFAULT_BATCH_SIZE: + cmd += " --batch_size %d" % batch_size + if batch_merge_repeat > 1: + cmd += " --batch_merge_repeat %d" % batch_merge_repeat if self.__use_cuda: cmd += " --use_cuda" @@ -271,23 +287,20 @@ class TestDistBase(unittest.TestCase): env=envs) local_out, local_err = local_proc.communicate() - local_ret = cpt.to_text(local_out) if check_error_log: err_log.close() - sys.stderr.write('local_stdout: %s\n' % local_ret) + sys.stderr.write('local_stdout: %s\n' % pickle.loads(local_out)) sys.stderr.write('local_stderr: %s\n' % local_err) - local_losses = local_ret.split("\n") - return local_losses + return pickle.loads(local_out) def _run_cluster(self, model, envs, check_error_log): # Run dist train to compare with local results ps0, ps1, ps0_pipe, ps1_pipe = self.start_pserver(model, check_error_log, envs) - self._wait_ps_ready(ps0.pid) - self._wait_ps_ready(ps1.pid) + ps0_ep, ps1_ep = self._ps_endpoints.split(",") tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --is_dist" @@ -322,8 +335,8 @@ class TestDistBase(unittest.TestCase): env0.update(envs) env1.update(envs) - print("tr0_cmd:{}, env0: {}".format(tr0_cmd, env0)) - print("tr1_cmd:{}, env1: {}".format(tr1_cmd, env1)) + print("tr0_cmd:{}".format(tr0_cmd)) + print("tr1_cmd:{}".format(tr1_cmd)) tr0_pipe = open("/tmp/tr0_err.log", "wb") tr1_pipe = open("/tmp/tr1_err.log", "wb") @@ -339,9 +352,7 @@ class TestDistBase(unittest.TestCase): env=env1) tr0_out, tr0_err = tr0_proc.communicate() - tr0_loss_text = cpt.to_text(tr0_out) tr1_out, tr1_err = tr1_proc.communicate() - tr1_loss_text = cpt.to_text(tr1_out) # close trainer file tr0_pipe.close() @@ -356,15 +367,13 @@ class TestDistBase(unittest.TestCase): ps1.terminate() # print log - sys.stderr.write('trainer 0 stdout:\n %s\n' % tr0_loss_text) - sys.stderr.write('trainer 0 stderr:\n %s\n' % tr0_err) - sys.stderr.write('trainer 1 stdout: %s\n' % tr1_loss_text) + sys.stderr.write('trainer 0 stdout: %s\n' % pickle.loads(tr0_out)) + sys.stderr.write('trainer 0 stderr: %s\n' % tr0_err) + sys.stderr.write('trainer 1 stdout: %s\n' % pickle.loads(tr1_out)) sys.stderr.write('trainer 1 stderr: %s\n' % tr1_err) - tr0_losses = tr0_loss_text.split("\n") - tr1_losses = tr1_loss_text.split("\n") - - return tr0_losses, tr1_losses + # return tr0_losses, tr1_losses + return pickle.loads(tr0_out), pickle.loads(tr1_out) def check_with_place(self, model_file, @@ -394,9 +403,9 @@ class TestDistBase(unittest.TestCase): check_error_log) for step_id in range(RUN_STEP): - local_loss = eval(local_losses[step_id])[0] - tr0_loss = eval(tr0_losses[step_id])[0] - tr1_loss = eval(tr1_losses[step_id])[0] - dist_loss = (tr0_loss + tr1_loss) / 2 - print(str(local_loss) + ":" + str(dist_loss)) - self.assertAlmostEqual(local_loss, dist_loss, delta=delta) + local_loss = local_losses[step_id] + tr0_loss = tr0_losses[step_id] + tr1_loss = tr1_losses[step_id] + dist_loss = (np.array([tr0_loss]) + np.array([tr1_loss])) / 2 + print("=======", local_loss, ":", dist_loss[0], "=======") + self.assertAlmostEqual(local_loss, dist_loss[0], delta=delta) diff --git a/python/paddle/fluid/tests/unittests/test_dist_ctr.py b/python/paddle/fluid/tests/unittests/test_dist_ctr.py index 3575fd07fc727bd6c6b07a19a60b1df6656ae9e2..b2d979729bc9b2546375cb657f78abe0d8c2dcc7 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_ctr.py +++ b/python/paddle/fluid/tests/unittests/test_dist_ctr.py @@ -18,14 +18,14 @@ import unittest from test_dist_base import TestDistBase +# FIXME(tangwei): sum op can not handle when inputs is empty. class TestDistCTR2x2(TestDistBase): def _setup_config(self): self._sync_mode = True self._enforce_place = "CPU" - -def test_dist_ctr(self): - self.check_with_place("dist_ctr.py", delta=1e-7, check_error_log=False) + def test_dist_ctr(self): + self.check_with_place("dist_ctr.py", delta=1e-7, check_error_log=False) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_dist_mnist.py b/python/paddle/fluid/tests/unittests/test_dist_mnist.py index f65dd7e2a28c4ace3988c0cc1267ebe981fbd9cb..81eb651878209164b3f339cc5030dbac847942d1 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_mnist.py +++ b/python/paddle/fluid/tests/unittests/test_dist_mnist.py @@ -26,6 +26,15 @@ class TestDistMnist2x2(TestDistBase): self.check_with_place("dist_mnist.py", delta=1e-5) +class TestDistMnist2x2Lars(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._use_reduce = False + + def test_se_resnext(self): + self.check_with_place("dist_mnist_lars.py", delta=1e-5) + + class TestDistMnist2x2WithMemopt(TestDistBase): def _setup_config(self): self._sync_mode = True @@ -44,6 +53,15 @@ class TestDistMnistAsync(TestDistBase): self.check_with_place("dist_mnist.py", delta=200) +class TestDistMnistDcAsgd(TestDistBase): + def _setup_config(self): + self._sync_mode = False + self._dc_asgd = True + + def test_se_resnext(self): + self.check_with_place("dist_mnist.py", delta=200) + + # FIXME(typhoonzero): enable these tests once we have 4 # 4 GPUs on CI machine, and the base class should be updated. # diff --git a/python/paddle/fluid/tests/unittests/test_dist_mnist_batch_merge.py b/python/paddle/fluid/tests/unittests/test_dist_mnist_batch_merge.py new file mode 100644 index 0000000000000000000000000000000000000000..22d4b7929033529c5cea60064e6d9de57eddeb8e --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_mnist_batch_merge.py @@ -0,0 +1,67 @@ +# 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 +from test_dist_base import TestDistBase +import os + + +class TestDistMnist2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._use_reduce = False + + def test_dist_train(self): + self.check_with_place("dist_mnist_batch_merge.py", delta=1e-5) + + def check_with_place(self, + model_file, + delta=1e-3, + check_error_log=False, + need_envs={}): + # TODO(typhoonzero): should auto adapt GPU count on the machine. + required_envs = { + "PATH": os.getenv("PATH", ""), + "PYTHONPATH": os.getenv("PYTHONPATH", ""), + "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH", ""), + "FLAGS_fraction_of_gpu_memory_to_use": "0.15", + "FLAGS_cudnn_deterministic": "1", + } + + required_envs.update(need_envs) + + if check_error_log: + required_envs["GLOG_v"] = "7" + required_envs["GLOG_logtostderr"] = "1" + + no_merge_losses = self._run_local( + model_file, + required_envs, + check_error_log=check_error_log, + batch_size=4) + + batch_merge_losses = self._run_local( + model_file, + required_envs, + check_error_log=check_error_log, + batch_size=2, + batch_merge_repeat=2) + # Ensure both result have values. + self.assertGreater(len(no_merge_losses), 1) + self.assertEqual(len(no_merge_losses), len(batch_merge_losses)) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_save_load.py b/python/paddle/fluid/tests/unittests/test_dist_save_load.py new file mode 100644 index 0000000000000000000000000000000000000000..03066fee48b703f8b55bd4ae6a9c4bb8deecab1e --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_dist_save_load.py @@ -0,0 +1,90 @@ +# 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 os +import shutil +import unittest +import tempfile + +import numpy as np + +from test_dist_base import TestDistBase, RUN_STEP + + +class TestDistSaveLoadDense2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._enforce_place = "CPU" + + def check_with_place(self, + model_file, + delta=1e-3, + check_error_log=False, + need_envs={}): + + required_envs = { + "PATH": os.getenv("PATH", ""), + "PYTHONPATH": os.getenv("PYTHONPATH", ""), + "LD_LIBRARY_PATH": os.getenv("LD_LIBRARY_PATH", ""), + "http_proxy": "" + } + + required_envs.update(need_envs) + + if check_error_log: + required_envs["GLOG_v"] = "7" + required_envs["GLOG_logtostderr"] = "1" + + model_dir = tempfile.mkdtemp() + + local_env = {} + local_env["SAVE"] = "1" + local_env["MODEL_DIR"] = model_dir + local_env.update(required_envs) + + cluster_env = {} + cluster_env["LOAD"] = "1" + cluster_env["MODEL_DIR"] = model_dir + cluster_env.update(required_envs) + + local_var = self._run_local(model_file, local_env, check_error_log) + tr0_var, tr1_var = self._run_cluster(model_file, cluster_env, + check_error_log) + + shutil.rmtree(model_dir) + + local_np = np.array(eval(local_var[0])) + train0_np = np.array(eval(tr0_var[0])) + train1_np = np.array(eval(tr1_var[0])) + self.assertAlmostEqual(local_np.all(), train0_np.all(), delta=delta) + self.assertAlmostEqual(local_np.all(), train1_np.all(), delta=delta) + self.assertAlmostEqual(train0_np.all(), train1_np.all(), delta=delta) + + @unittest.skip(reason="CI fail") + def test_dist(self): + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '0', + 'IS_SELF_CONTAINED_LR': '1' + } + self.check_with_place( + "dist_save_load.py", + delta=0, + check_error_log=False, + need_envs=need_envs) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py index c0989ca709e100d8f147a08970b0e858c81ce09b..c2a4e5ca0c050813785f602c5d2088466e616971 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py +++ b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py @@ -23,16 +23,17 @@ class TestDistSeResneXt2x2(TestDistBase): self._use_reader_alloc = False def test_dist_train(self): - self.check_with_place("dist_se_resnext.py", delta=100) + self.check_with_place("dist_se_resnext.py", delta=1e-7) class TestDistseResnXt2x2WithMemopt(TestDistBase): def _setup_config(self): self._sync_mode = True self._mem_opt = True + self._use_reader_alloc = False def test_dist_train(self): - self.check_with_place("dist_se_resnext.py", delta=100) + self.check_with_place("dist_se_resnext.py", delta=1e-7) class TestDistSeResneXt2x2Async(TestDistBase): diff --git a/python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py b/python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py index e971f29db42a7c1a2394505a8ece3d2fd6b347e9..102a4dab05fe1adc6a503920714f50415b29dc19 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py +++ b/python/paddle/fluid/tests/unittests/test_dist_simnet_bow.py @@ -25,7 +25,11 @@ class TestDistSimnetBowDense2x2(TestDistBase): self._enforce_place = "CPU" def test_simnet_bow(self): - need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '0'} + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '0', + 'IS_SELF_CONTAINED_LR': '1' + } self.check_with_place( "dist_simnet_bow.py", delta=1e-5, @@ -38,8 +42,12 @@ class TestDistSimnetBow2x2DenseAsync(TestDistBase): self._sync_mode = False self._enforce_place = "CPU" - def test_simnet_bow(self): - need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '0'} + def no_test_simnet_bow(self): + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '0', + 'IS_SELF_CONTAINED_LR': '1' + } self.check_with_place( "dist_simnet_bow.py", delta=100, @@ -53,7 +61,11 @@ class TestDistSimnetBowSparse2x2(TestDistBase): self._enforce_place = "CPU" def test_simnet_bow(self): - need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '1'} + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '1' + } self.check_with_place( "dist_simnet_bow.py", delta=1e-5, @@ -67,7 +79,48 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase): self._enforce_place = "CPU" def test_simnet_bow(self): - need_envs = {"IS_DISTRIBUTED": '0', "IS_SPARSE": '1'} + need_envs = { + "IS_DISTRIBUTED": '0', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '1' + } + self.check_with_place( + "dist_simnet_bow.py", + delta=100, + check_error_log=False, + need_envs=need_envs) + + +# FIXME(tangwei): Learningrate variable is not created on pserver. +class TestDistSimnetBow2x2LookupTableSync(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._enforce_place = "CPU" + + def test_simnet_bow(self): + need_envs = { + "IS_DISTRIBUTED": '1', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '1' + } + self.check_with_place( + "dist_simnet_bow.py", + delta=1e-5, + check_error_log=True, + need_envs=need_envs) + + +class TestDistSimnetBow2x2LookupTableAsync(TestDistBase): + def _setup_config(self): + self._sync_mode = False + self._enforce_place = "CPU" + + def test_simnet_bow(self): + need_envs = { + "IS_DISTRIBUTED": '1', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '1' + } self.check_with_place( "dist_simnet_bow.py", delta=100, @@ -75,5 +128,23 @@ class TestDistSimnetBow2x2SparseAsync(TestDistBase): need_envs=need_envs) +class TestDistSimnetBow2x2LookupTableNotContainLRSync(TestDistBase): + def _setup_config(self): + self._sync_mode = True + self._enforce_place = "CPU" + + def test_simnet_bow(self): + need_envs = { + "IS_DISTRIBUTED": '1', + "IS_SPARSE": '1', + 'IS_SELF_CONTAINED_LR': '0' + } + self.check_with_place( + "dist_simnet_bow.py", + delta=1e-5, + check_error_log=False, + need_envs=need_envs) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_transformer.py b/python/paddle/fluid/tests/unittests/test_dist_transformer.py index 47e8dfaf03ceb27a74f5e48d662d2b534d2d152b..25dcccc28d710695d4c5e08c17816669d0fae5d8 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transformer.py @@ -61,7 +61,8 @@ class TestDistTransformer2x2Sync(TestDistBase): def test_dist_train(self): download_files() - self.check_with_place("dist_transformer.py", delta=1e-5) + self.check_with_place( + "dist_transformer.py", delta=1e-5, check_error_log=False) class TestDistTransformer2x2Async(TestDistBase): @@ -70,7 +71,8 @@ class TestDistTransformer2x2Async(TestDistBase): def test_dist_train(self): download_files() - self.check_with_place("dist_transformer.py", delta=1.0) + self.check_with_place( + "dist_transformer.py", delta=1.0, check_error_log=False) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py index 54a1c68a37f6929890aab697b48d621e6effb7d8..d132dd3c48f55c07725515e40faeb5076398adeb 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -283,6 +283,25 @@ class TestDecayedAdagrad(TranspilerTest): trainer, _ = self.get_trainer() +class TestFtrl(TranspilerTest): + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=fluid.ParamAttr(name='fc_b')) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + opt = fluid.optimizer.Ftrl(learning_rate=0.1) + opt.minimize(avg_cost) + + def transpiler_test_impl(self): + pserver, startup = self.get_pserver(self.pserver1_ep) + trainer, _ = self.get_trainer() + + class TestLRDecayConditional(TranspilerTest): def net_conf(self): x = fluid.layers.data(name='x', shape=[1000], dtype='float32') @@ -354,9 +373,8 @@ class TestL2Decay(TranspilerTest): self.assertEqual(len(pserver.blocks), 3) self.assertEqual([op.type for op in pserver.blocks[1].ops], ["sum", "scale", "clip", "sgd"]) - self.assertEqual( - [op.type for op in pserver.blocks[2].ops], - ["sum", "scale", "clip", "scale", "elementwise_add", "sgd"]) + self.assertEqual([op.type for op in pserver.blocks[2].ops], + ["sum", "scale", "clip", "scale", "sum", "sgd"]) # TODO(typhoonzero): test clipping and L2Decay ops are removed from trainer @@ -397,12 +415,35 @@ class TestL2DecayWithPiecewise(TranspilerTest): "logical_and", "conditional_block", "fill_constant", "conditional_block" ]) - self.assertEqual( - [op.type for op in pserver.blocks[7].ops], - ["sum", "scale", "scale", "elementwise_add", "momentum"]) - self.assertEqual( - [op.type for op in pserver.blocks[8].ops], - ["sum", "scale", "scale", "elementwise_add", "momentum"]) + self.assertEqual([op.type for op in pserver.blocks[7].ops], + ["sum", "scale", "scale", "sum", "momentum"]) + self.assertEqual([op.type for op in pserver.blocks[8].ops], + ["sum", "scale", "scale", "sum", "momentum"]) + + +class TestEmptyPserverOptimizeBlocks(TranspilerTest): + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + # only one parameter + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=False) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + sgd_optimizer = fluid.optimizer.SGD(learning_rate=1.0) + sgd_optimizer.minimize(avg_cost) + + def transpiler_test_impl(self): + config = fluid.DistributeTranspilerConfig() + config.slice_var_up = False + + pserver, startup = self.get_pserver(ep=self.pserver2_ep, config=config) + + self.assertEqual(len(pserver.blocks), 2) + self.assertEqual(len(pserver.blocks[1].ops), 0) class TestDistLookupTableBase(TranspilerTest): @@ -411,12 +452,12 @@ class TestDistLookupTableBase(TranspilerTest): self.emb_size = 64 self.lookup_table_name = 'shared_w' - def emb_pool(ids): + def emb_pool(ids, table_name, is_distributed): emb = fluid.layers.embedding( input=ids, size=[self.table_size, self.emb_size], dtype='float32', - param_attr=self.lookup_table_name, # share parameter + param_attr=table_name, is_sparse=is_sparse, is_distributed=is_distributed) pool = fluid.layers.sequence_pool(input=emb, pool_type='average') @@ -426,9 +467,13 @@ class TestDistLookupTableBase(TranspilerTest): name='title_ids', shape=[1], dtype='int64', lod_level=1) brand_ids = fluid.layers.data( name='brand_ids', shape=[1], dtype='int64', lod_level=1) - title_emb = emb_pool(title_ids) - brand_emb = emb_pool(brand_ids) - fc0 = fluid.layers.concat(input=[title_emb, brand_emb], axis=1) + profile_ids = fluid.layers.data( + name='brand_ids', shape=[1], dtype='int64', lod_level=1) + title_emb = emb_pool(title_ids, self.lookup_table_name, is_distributed) + brand_emb = emb_pool(brand_ids, self.lookup_table_name, is_distributed) + profile_emb = emb_pool(profile_ids, "profile_emb", False) + fc0 = fluid.layers.concat( + input=[title_emb, brand_emb, profile_emb], axis=1) predict = fluid.layers.fc(input=fc0, size=2, act=None, @@ -449,7 +494,7 @@ class TestLocalLookupTable(TestDistLookupTableBase): def transpiler_test_impl(self): pserver1, startup1 = self.get_pserver(self.pserver1_ep) - self.assertEqual(len(pserver1.blocks), 3) + self.assertEqual(len(pserver1.blocks), 4) # 0 listen_and_serv # 1 optimize for fc_w or fc_b adam self.assertEqual([op.type for op in pserver1.blocks[1].ops], @@ -459,16 +504,23 @@ class TestLocalLookupTable(TestDistLookupTableBase): self.assertEqual([op.type for op in pserver1.blocks[2].ops], ["sum", "scale", "adam", "scale", "scale"]) + # 3 optimize for table 2 adam + # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num + self.assertEqual([op.type for op in pserver1.blocks[3].ops], + ["sum", "scale", "adam", "scale", "scale"]) + trainer, _ = self.get_trainer() self.assertEqual(len(trainer.blocks), 1) ops = [ 'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool', - 'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean', - 'fill_constant', 'mean_grad', 'cross_entropy_grad', - 'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad', - 'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad', - 'lookup_table_grad', 'sum', 'split_selected_rows', 'send', - 'send_barrier', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat' + 'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add', + 'cross_entropy', 'mean', 'fill_constant', 'mean_grad', + 'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad', + 'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad', + 'split_selected_rows', 'send', 'sequence_pool_grad', + 'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad', + 'sum', 'split_selected_rows', 'send', 'send_barrier', 'recv', + 'recv', 'recv', 'recv', 'fetch_barrier', 'concat', 'concat' ] self.assertEqual([op.type for op in trainer.blocks[0].ops], ops) @@ -485,31 +537,42 @@ class TestDistLookupTable(TestDistLookupTableBase): # 1 optimize for fc_w or fc_b adam self.assertEqual([op.type for op in pserver1.blocks[1].ops], ["sum", "scale", "adam", "scale", "scale"]) - # 2 optimize for table sgd + # 4 prefetch -> lookup_sparse_table for data0 self.assertEqual([op.type for op in pserver1.blocks[2].ops], + ["sum", "scale", "adam", "scale", "scale"]) + # 2 optimize for table sgd + self.assertEqual([op.type for op in pserver1.blocks[3].ops], ["sum", "sgd"]) # 3 prefetch -> lookup_sparse_table for data0 - self.assertEqual([op.type for op in pserver1.blocks[3].ops], - ["lookup_sparse_table"]) - # 4 prefetch -> lookup_sparse_table for data1 self.assertEqual([op.type for op in pserver1.blocks[4].ops], ["lookup_sparse_table"]) # 5 save table self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"]) - trainer, _ = self.get_trainer() + trainer, trainer_startup = self.get_trainer() self.assertEqual(len(trainer.blocks), 1) ops = [ - 'split_ids', 'prefetch', 'merge_ids', 'sequence_pool', 'split_ids', - 'prefetch', 'merge_ids', 'sequence_pool', 'concat', 'mul', + 'split_ids', 'prefetch', 'merge_ids', 'sequence_pool', + 'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean', 'fill_constant', 'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad', - 'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad', - 'sum', 'split_ids', 'send', 'send_barrier', 'recv', 'recv', - 'fetch_barrier' + 'lookup_table_grad', 'split_selected_rows', 'send', + 'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad', + 'lookup_table_grad', 'sum', 'split_ids', 'send', 'send_barrier', + 'recv', 'recv', 'recv', 'fetch_barrier', 'concat' ] self.assertEqual([op.type for op in trainer.blocks[0].ops], ops) + startup_ops = [ + 'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant', + 'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant', + 'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant', + 'fill_constant', 'fill_constant', 'uniform_random', + 'uniform_random', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat', + 'fake_init' + ] + self.assertEqual([op.type for op in trainer_startup.blocks[0].ops], + startup_ops) class TestAsyncLocalLookupTable(TestDistLookupTableBase): @@ -520,7 +583,7 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase): config = fluid.DistributeTranspilerConfig() pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False) - self.assertEqual(len(pserver1.blocks), 3) + self.assertEqual(len(pserver1.blocks), 4) # 0 listen_and_serv # 1 optimize for fc_w or fc_b adam self.assertEqual([op.type for op in pserver1.blocks[1].ops], @@ -529,17 +592,23 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase): # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num self.assertEqual([op.type for op in pserver1.blocks[2].ops], ["adam", "scale", "scale"]) + # 3 optimize for table adam + # NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num + self.assertEqual([op.type for op in pserver1.blocks[3].ops], + ["adam", "scale", "scale"]) trainer, _ = self.get_trainer(config) self.assertEqual(len(trainer.blocks), 1) ops = [ 'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool', - 'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean', - 'fill_constant', 'mean_grad', 'cross_entropy_grad', - 'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad', - 'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad', - 'lookup_table_grad', 'sum', 'split_selected_rows', 'send', 'recv', - 'recv', 'recv', 'concat' + 'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add', + 'cross_entropy', 'mean', 'fill_constant', 'mean_grad', + 'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad', + 'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad', + 'split_selected_rows', 'send', 'sequence_pool_grad', + 'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad', + 'sum', 'split_selected_rows', 'send', 'recv', 'recv', 'recv', + 'recv', 'concat', 'concat' ] self.assertEqual([op.type for op in trainer.blocks[0].ops], ops) @@ -558,29 +627,41 @@ class TestAsyncDistLookupTable(TestDistLookupTableBase): # 1 optimize for fc_w or fc_b adam self.assertEqual([op.type for op in pserver1.blocks[1].ops], ["adam", "scale", "scale"]) - # 2 optimize for table sgd - self.assertEqual([op.type for op in pserver1.blocks[2].ops], ["sgd"]) - # 3 prefetch -> lookup_sparse_table for data0 - self.assertEqual([op.type for op in pserver1.blocks[3].ops], - ["lookup_sparse_table"]) - # 4 prefetch -> lookup_sparse_table for data1 + # 2 optimize for table adam + self.assertEqual([op.type for op in pserver1.blocks[2].ops], + ["adam", "scale", "scale"]) + # 3 optimize for table sgd + self.assertEqual([op.type for op in pserver1.blocks[3].ops], ["sgd"]) + # 4 prefetch -> lookup_sparse_table for data0 self.assertEqual([op.type for op in pserver1.blocks[4].ops], ["lookup_sparse_table"]) # 5 save table self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"]) - trainer, _ = self.get_trainer(config) + trainer, trainer_startup = self.get_trainer(config) self.assertEqual(len(trainer.blocks), 1) ops = [ - 'split_ids', 'prefetch', 'merge_ids', 'sequence_pool', 'split_ids', - 'prefetch', 'merge_ids', 'sequence_pool', 'concat', 'mul', + 'split_ids', 'prefetch', 'merge_ids', 'sequence_pool', + 'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean', 'fill_constant', 'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad', - 'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad', - 'sum', 'split_ids', 'send', 'recv', 'recv' + 'lookup_table_grad', 'split_selected_rows', 'send', + 'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad', + 'lookup_table_grad', 'sum', 'split_ids', 'send', 'recv', 'recv', + 'recv', 'concat' ] self.assertEqual([op.type for op in trainer.blocks[0].ops], ops) + startup_ops = [ + 'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant', + 'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant', + 'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant', + 'fill_constant', 'fill_constant', 'uniform_random', + 'uniform_random', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat', + 'fake_init' + ] + self.assertEqual([op.type for op in trainer_startup.blocks[0].ops], + startup_ops) class TestDistLookupTableSliceSize(TestDistLookupTableBase): diff --git a/python/paddle/fluid/tests/unittests/test_dropout_op.py b/python/paddle/fluid/tests/unittests/test_dropout_op.py index 0296bc2af4e0b79478c34b4cceab32b5a8a50f2f..be3c5f3b9558ec522803ed9a5acedea75cda6ccc 100644 --- a/python/paddle/fluid/tests/unittests/test_dropout_op.py +++ b/python/paddle/fluid/tests/unittests/test_dropout_op.py @@ -85,6 +85,69 @@ class TestDropoutOp5(OpTest): self.check_output() +class TestDropoutOp6(TestDropoutOp): + def setUp(self): + self.op_type = "dropout" + self.inputs = {'X': np.random.random((32, 64)).astype("float32")} + self.attrs = { + 'dropout_prob': 1.0, + 'fix_seed': True, + 'is_test': False, + 'dropout_implementation': 'upscale_in_train' + } + self.outputs = { + 'Out': np.zeros((32, 64)).astype('float32'), + 'Mask': np.zeros((32, 64)).astype('float32') + } + + +class TestDropoutOp7(TestDropoutOp): + def setUp(self): + self.op_type = "dropout" + self.inputs = {'X': np.random.random((32, 64, 2)).astype("float32")} + self.attrs = { + 'dropout_prob': 0.0, + 'fix_seed': True, + 'is_test': False, + 'dropout_implementation': 'upscale_in_train' + } + self.outputs = { + 'Out': self.inputs['X'], + 'Mask': np.ones((32, 64, 2)).astype('float32') + } + + +class TestDropoutOp8(OpTest): + def setUp(self): + self.op_type = "dropout" + self.inputs = {'X': np.random.random((32, 64)).astype("float32")} + self.attrs = { + 'dropout_prob': 0.35, + 'fix_seed': True, + 'is_test': True, + 'dropout_implementation': 'upscale_in_train' + } + self.outputs = {'Out': self.inputs['X']} + + def test_check_output(self): + self.check_output() + + +class TestDropoutOp9(OpTest): + def setUp(self): + self.op_type = "dropout" + self.inputs = {'X': np.random.random((32, 64, 3)).astype("float32")} + self.attrs = { + 'dropout_prob': 0.75, + 'is_test': True, + 'dropout_implementation': 'upscale_in_train' + } + self.outputs = {'Out': self.inputs['X']} + + def test_check_output(self): + self.check_output() + + class TestFP16DropoutOp(OpTest): def setUp(self): self.op_type = "dropout" diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py index 6a129b6df9bf1830fdf5eb5cb9ae0c5e4f7bb4ec..53409e436c0739bce63a3a8f90591e0ca6836859 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py @@ -117,56 +117,5 @@ class TestElementwiseMulOp_broadcast_3(ElementwiseMulOp): } -class TestElementWiseMulSelectedRows(OpTest): - def setUp(self): - self.rows = [0, 1, 2, 3, 4, 5, 6] - self.feature = 12 - self.height = 100 - self.input_shape = (len(self.rows), self.feature) - - def prepare_input(self, scope, place): - self.input = { - "X": np.random.random(self.input_shape).astype("float32"), - "Y": np.random.random(self.input_shape).astype("float32") - } - - def init_input(in_name): - x_selected_rows = scope.var(in_name).get_selected_rows() - x_selected_rows.set_height(self.height) - x_selected_rows.set_rows(self.rows) - x_array = self.input[in_name] - x_tensor = x_selected_rows.get_tensor() - x_tensor.set(x_array, place) - - init_input("X") - init_input("Y") - - def create_out_selected_row(self, scope): - return scope.var('Out').get_selected_rows() - - def check_result(self, out_selected_rows): - assert out_selected_rows.height() == self.height - assert out_selected_rows.rows() == self.rows - out_tensor = np.array(out_selected_rows.get_tensor()) - assert out_tensor.shape == self.input_shape - - def check_with_place(self, place): - scope = core.Scope() - self.prepare_input(scope, place) - - out_selected_rows = self.create_out_selected_row(scope) - out_selected_rows.set_height(0) - out_selected_rows.set_rows([]) - - elementwise_mul = Operator("elementwise_mul", X='X', Y='Y', Out='Out') - elementwise_mul.run(scope, place) - self.check_result(out_selected_rows) - - def test_elewisemul_with_selected_rows_input(self): - places = [core.CPUPlace()] - for place in places: - self.check_with_place(place) - - if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_extract_rows_op.py b/python/paddle/fluid/tests/unittests/test_extract_rows_op.py deleted file mode 100644 index 8629bcf0f2e3c37aefdbf79b203176a43e0c3a7e..0000000000000000000000000000000000000000 --- a/python/paddle/fluid/tests/unittests/test_extract_rows_op.py +++ /dev/null @@ -1,60 +0,0 @@ -# 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 paddle.fluid.core as core -from paddle.fluid.op import Operator -from op_test import OpTest - - -class TestExtractRows(OpTest): - def check_with_place(self, place): - scope = core.Scope() - - # create and initialize Variable - feature_len = 12 - rows = [0, 4, 4, 7] - np_array = np.ones((len(rows), feature_len)).astype("float32") - - in_x = scope.var('X').get_selected_rows() - in_x.set_height(len(rows)) - in_x.set_rows(rows) - in_x_tensor = in_x.get_tensor() - in_x_tensor.set(np_array, place) - - # create Out Variable - out_tensor = scope.var('Out').get_tensor() - - # create and run lookup_table operator - extract_rows_op = Operator("extract_rows", X='X', Out='Out') - extract_rows_op.run(scope, place) - - # get result from Out - result_array = np.array(out_tensor) - result_array = [ele[0] for ele in result_array] - assert result_array == rows - - def test_concat_rows(self): - places = [core.CPUPlace()] - if core.is_compiled_with_cuda(): - places.append(core.CUDAPlace(0)) - for place in places: - self.check_with_place(place) - - -if __name__ == '__main__': - unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fake_init_op.py b/python/paddle/fluid/tests/unittests/test_fake_init_op.py new file mode 100644 index 0000000000000000000000000000000000000000..a62b7aed66b59940b4ba654d98479e3e35c7b78b --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fake_init_op.py @@ -0,0 +1,52 @@ +# 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 paddle.fluid.core as core +from paddle.fluid.op import Operator + + +class TestFakeInitOpSelectedRows(unittest.TestCase): + def check_with_place(self, place, is_selected_rows): + scope = core.Scope() + + out_var_name = 'Out' + if is_selected_rows: + out_tensor = scope.var(out_var_name).get_selected_rows().get_tensor( + ) + else: + out_tensor = scope.var(out_var_name).get_tensor() + + var_shape = [4, 784] + + # create and run fake_init_op + fake_init_op = Operator("fake_init", Out=out_var_name, shape=var_shape) + fake_init_op.run(scope, place) + + self.assertEqual(var_shape, out_tensor._get_dims()) + + def test_fake_init_selected_rows(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + for place in places: + for is_selected_rows in [True, False]: + self.check_with_place(place, is_selected_rows) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fused_embedding_fc_lstm_op.py b/python/paddle/fluid/tests/unittests/test_fused_embedding_fc_lstm_op.py new file mode 100644 index 0000000000000000000000000000000000000000..70ca521d3387ac11cd41d8496b4d094667232d4c --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fused_embedding_fc_lstm_op.py @@ -0,0 +1,218 @@ +# 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 +from op_test import OpTest +from test_lstm_op import lstm, ACTIVATION + + +def fc(x, w, b): + return np.dot(x, w) + b + + +def fused_embedded_fc_lstm( + ids, # T x 1 + lod, # 1 x N + embeddings=None, # Dict_size x M + wx=None, # M x 4D + bx=None, # 1 x 4D + h0=None, # N x D + c0=None, # N x D + w_h=None, # D x 4D + w_b=None, # 1 x 4D + w_c=None, # 1 x 3D + is_reverse=False, + act_gate=None, + act_cell=None, + act_cand=None): + # Make a lookup for embeddings and pass result into lstm reference + T = ids.shape[0] + M = embeddings.shape[1] + x = embeddings[ids].reshape([T, M]) + return lstm( + fc(x, wx, bx), lod, h0, c0, w_h, w_b, w_c, is_reverse, act_gate, + act_cell, act_cand) + + +class TestFusionLSTMOp(OpTest): + def set_conf(self): + pass + + def setUp(self): + self.op_type = 'fused_embedding_fc_lstm' + self.lod = [[2, 3, 5, 4]] + self.M = 8 # Embedding size + self.D = 16 # Hidden size + self.dict_size = 18 + self.has_initial_state = False + self.use_peepholes = False + self.is_reverse = False + self.act_gate = 'sigmoid' + self.act_cell = 'tanh' + self.act_cand = 'tanh' + self.set_conf() + + T = sum(self.lod[0]) + bs = len(self.lod[0]) + + # this is the weight of fc + wx = np.random.normal(size=(self.M, 4 * self.D)).astype('float32') + # this is the bias of fc + bx = np.random.normal(size=(1, 4 * self.D)).astype('float32') + + if self.use_peepholes: + b = np.random.normal(size=(1, 7 * self.D)).astype('float32') + else: + b = np.random.normal(size=(1, 4 * self.D)).astype('float32') + w_b = np.copy(b[:, 0:4 * self.D]) + w_c = b[:, 4 * self.D:] if self.use_peepholes else None + + # low is 0 , high is voc_size - 1 + ids = np.random.randint( + low=0, high=self.dict_size - 1, size=(T, 1)).astype("int64") + # embeddings as they were trained , so each entry is of M size + embeddings = np.random.random( + (self.dict_size, self.M)).astype("float32") + + # multiply embeddings via Weights + fc_embeddings = np.dot(embeddings, wx) + + # bias should be manually added into the bias of this fused embedding fc LSTM + b[0, 0:4 * self.D] += bx[0, :] + combined_biases = b[:, 0:4 * self.D] + # So let broadcast it , so they can be added + ones = np.ones([self.dict_size, 1]) + broadcasted_biases = np.dot(ones, combined_biases) + # Sum biases with Wx*embeddings + fc_embeddings += broadcasted_biases + + if self.has_initial_state: + h0 = np.random.normal(size=(bs, self.D)).astype('float32') + c0 = np.random.normal(size=(bs, self.D)).astype('float32') + else: + h0 = np.zeros((bs, self.D)).astype('float32') + c0 = np.zeros((bs, self.D)).astype('float32') + + wh = np.random.normal(size=(self.D, 4 * self.D)).astype('float32') + + h, c = fused_embedded_fc_lstm( + ids, self.lod, embeddings, wx, bx, h0, c0, wh, w_b, w_c, + self.is_reverse, ACTIVATION[self.act_gate], + ACTIVATION[self.act_cell], ACTIVATION[self.act_cand]) + + self.inputs = { + 'Ids': (ids, self.lod), + 'Embeddings': fc_embeddings, + 'WeightH': wh, + 'Bias': b + } + + if self.has_initial_state: + self.inputs['H0'] = h0 + self.inputs['C0'] = c0 + + self.outputs = { + 'Hidden': (h, self.lod), + 'Cell': (c, self.lod), + } + self.attrs = { + 'use_peepholes': self.use_peepholes, + 'is_reverse': self.is_reverse, + 'gate_activation': self.act_gate, + 'cell_activation': self.act_cell, + 'candidate_activation': self.act_cand + } + + def test_check_output(self): + for use_seq in {True, False}: + self.attrs['use_seq'] = use_seq + self.check_output() + + +class TestFusionLSTMOpInit(TestFusionLSTMOp): + def set_conf(self): + self.has_initial_state = True + + +class TestFusionLSTMOpReverse(TestFusionLSTMOp): + def set_conf(self): + self.is_reverse = True + + +class TestFusionLSTMOpInitReverse(TestFusionLSTMOp): + def set_conf(self): + self.has_initial_state = True + self.is_reverse = True + + +class TestFusionLSTMOpMD1(TestFusionLSTMOp): + def set_conf(self): + self.M = 36 + self.D = 8 + + +class TestFusionLSTMOpMD2(TestFusionLSTMOp): + def set_conf(self): + self.M = 8 + self.D = 8 + + +class TestFusionLSTMOpMD3(TestFusionLSTMOp): + def set_conf(self): + self.M = 15 + self.D = 3 + + +class TestFusionLSTMOpBS1(TestFusionLSTMOp): + def set_conf(self): + self.lod = [[3]] + self.D = 16 + + +class TestFusionLSTMOpPeepholes(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + + +class TestFusionLSTMOpPeepholesInit(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + self.has_initial_state = True + + +class TestFusionLSTMOpPeepholesReverse(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + self.is_reverse = True + + +class TestFusionLSTMOpPeepholesInitReverse(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + self.has_initial_state = True + self.is_reverse = True + + +class TestFusionLSTMOpPeepholesBS1(TestFusionLSTMOp): + def set_conf(self): + self.use_peepholes = True + self.lod = [[2]] + self.D = 8 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_fusion_gru_op.py b/python/paddle/fluid/tests/unittests/test_fusion_gru_op.py index 36ebc8fb6ea9efdcd1807f5c8917ab1428b3381e..377454e7802e40f90c371987adfe50cce922c764 100644 --- a/python/paddle/fluid/tests/unittests/test_fusion_gru_op.py +++ b/python/paddle/fluid/tests/unittests/test_fusion_gru_op.py @@ -125,6 +125,12 @@ class TestFusionGRUOpMD2(TestFusionGRUOp): self.D = 8 +class TestFusionGRUOpMD3(TestFusionGRUOp): + def set_confs(self): + self.M = 17 + self.D = 15 + + class TestFusionGRUOpBS1(TestFusionGRUOp): def set_confs(self): self.lod = [[3]] diff --git a/python/paddle/fluid/tests/unittests/test_fusion_seqconv_eltadd_relu_op.py b/python/paddle/fluid/tests/unittests/test_fusion_seqconv_eltadd_relu_op.py new file mode 100644 index 0000000000000000000000000000000000000000..ba6f1415b1c832eb688443953866652e3458b172 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fusion_seqconv_eltadd_relu_op.py @@ -0,0 +1,94 @@ +# 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 random +from op_test import OpTest +from test_seq_conv import seqconv + + +class TestSeqConvEltAddRelu(OpTest): + def set_conf(self): + pass + + def setUp(self): + self.op_type = 'fusion_seqconv_eltadd_relu' + self.lod = [[6, 4]] + self.in_fea_size = 16 + self.out_fea_size = 8 + self.context_length = 4 + self.context_stride = 1 + self.context_start = 0 + self.set_conf() + + assert self.context_stride == 1 + + T = sum(self.lod[0]) + x = np.random.uniform(-1, 1, [T, self.in_fea_size]).astype('float32') + w = np.random.uniform( + -1, 1, [self.in_fea_size * self.context_length, + self.out_fea_size]).astype('float32') + b = np.random.uniform(-2, 1, [1, self.out_fea_size]).astype('float32') + out = seqconv(x, self.lod, w, self.context_length, self.context_start) + out = np.maximum(out + b, 0) + + self.inputs = {'X': (x, self.lod), 'Filter': w, 'Bias': b} + self.attrs = { + 'contextStart': self.context_start, + 'contextLength': self.context_length, + 'contextStride': self.context_stride + } + self.outputs = {'Out': out} + + def test_check_output(self): + self.check_output() + + +class TestSeqConvEltAddReluBS1(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[10]] + + +class TestSeqConvEltAddReluBS1Case2(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[2]] + + +class TestSeqConvEltAddReluCase1(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[3, 5, 1, 6]] + self.context_length = 3 + self.context_start = -2 + + +class TestSeqConvEltAddReluCase2(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[10, 1, 2, 4, 1, 5, 6]] + self.in_fea_size = 2 + self.context_length = 4 + self.context_start = -1 + + +class TestSeqConvEltAddReluCase3(TestSeqConvEltAddRelu): + def set_conf(self): + self.lod = [[10, 1, 2, 4, 1, 5, 6]] + self.context_length = 5 + self.context_start = -4 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_grid_sampler_op.py b/python/paddle/fluid/tests/unittests/test_grid_sampler_op.py new file mode 100644 index 0000000000000000000000000000000000000000..c2529e0d70c9a359d2a44c671769d50a92650a73 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_grid_sampler_op.py @@ -0,0 +1,123 @@ +# 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 unittest +import numpy as np +from op_test import OpTest + + +def AffineGrid(theta, size): + n = size[0] + h = size[2] + w = size[3] + h_idx = np.repeat( + np.linspace(-1, 1, h)[np.newaxis, :], w, axis=0).T[:, :, np.newaxis] + w_idx = np.repeat( + np.linspace(-1, 1, w)[np.newaxis, :], h, axis=0)[:, :, np.newaxis] + grid = np.concatenate( + [w_idx, h_idx, np.ones([h, w, 1])], axis=2) # h * w * 3 + grid = np.repeat(grid[np.newaxis, :], size[0], axis=0) # n * h * w *3 + + ret = np.zeros([n, h * w, 2]) + theta = theta.transpose([0, 2, 1]) + for i in range(len(theta)): + ret[i] = np.dot(grid[i].reshape([h * w, 3]), theta[i]) + + return ret.reshape([n, h, w, 2]).astype("float32") + + +def getGridPointValue(data, x, y): + data_shape = data.shape + N = data_shape[0] + H = data_shape[2] + W = data_shape[3] + + out = np.zeros(data_shape, dtype='float') + for i in range(N): + for j in range(H): + for k in range(W): + if y[i, j, k] < 0 or y[i, j, k] > H - 1 or x[i, j, k] < 0 or x[ + i, j, k] > W - 1: + out[i, :, j, k] = 0 + else: + out[i, :, j, k] = data[i, :, y[i, j, k], x[i, j, k]] + + return out + + +def GridSampler(data, grid): + dims = data.shape + N = dims[0] + C = dims[1] + H = dims[2] + W = dims[3] + + x = grid[:, :, :, 0] + y = grid[:, :, :, 1] + y_max = H - 1 + x_max = W - 1 + + x = 0.5 * ((x.astype('float32') + 1.0) * x_max) + y = 0.5 * ((y.astype('float32') + 1.0) * y_max) + + x0 = np.floor(x).astype('int32') + x1 = x0 + 1 + y0 = np.floor(y).astype('int32') + y1 = y0 + 1 + + wa = np.tile(((x1 - x) * (y1 - y)).reshape((N, 1, H, W)), (1, C, 1, 1)) + wb = np.tile(((x1 - x) * (y - y0)).reshape((N, 1, H, W)), (1, C, 1, 1)) + wc = np.tile(((x - x0) * (y1 - y)).reshape((N, 1, H, W)), (1, C, 1, 1)) + wd = np.tile(((x - x0) * (y - y0)).reshape((N, 1, H, W)), (1, C, 1, 1)) + + va = getGridPointValue(data, x0, y0) + vb = getGridPointValue(data, x0, y1) + vc = getGridPointValue(data, x1, y0) + vd = getGridPointValue(data, x1, y1) + + out = (wa * va + wb * vb + wc * vc + wd * vd).astype('float32') + return out + + +class TestGridSamplerOp(OpTest): + def setUp(self): + self.initTestCase() + self.op_type = 'grid_sampler' + x = np.random.randint(0, 255, self.x_shape).astype('float32') + + theta = np.zeros(self.theta_shape).astype('float32') + for i in range(self.theta_shape[0]): + for j in range(2): + for k in range(3): + theta[i, j, k] = np.random.rand(1)[0] + grid = AffineGrid(theta, self.x_shape) + + self.inputs = {'X': x, 'Grid': grid} + self.attrs = {'use_cudnn': True} + self.outputs = {'Output': GridSampler(x, grid)} + + def test_check_output(self): + self.check_output(atol=1e-3) + + def test_check_grad_normal(self): + self.check_grad(['X', 'Grid'], 'Output', max_relative_error=0.61) + + def initTestCase(self): + self.x_shape = (2, 5, 7, 3) + self.grid_shape = (2, 7, 3, 2) + self.theta_shape = (2, 2, 3) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_hash_op.py b/python/paddle/fluid/tests/unittests/test_hash_op.py new file mode 100644 index 0000000000000000000000000000000000000000..1130ea39c42204283885ab1072a52db8c22f8b2e --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_hash_op.py @@ -0,0 +1,57 @@ +# 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 unittest +import numpy as np +from op_test import OpTest + + +class TestScaleOp(OpTest): + def setUp(self): + self.op_type = "hash" + self.init_test_case() + self.inputs = {'X': (self.in_seq, self.lod)} + self.attrs = {'num_hash': 4, 'mod_by': 10000} + self.outputs = {'Out': (self.out_seq, self.lod)} + + def init_test_case(self): + np.random.seed = 1 + self.in_seq = np.random.randint(0, 10, (30, 1)).astype("int32") + self.lod = [[9, 4, 11, 6]] + # self.out_seq = np.ones([30, 4, 1], dtype=np.int32) + self.out_seq = [ + [[9662], [9217], [1129], [8487]], [[9662], [9217], [1129], [8487]], + [[8310], [1327], [1654], [4567]], [[6897], [3218], [2013], [1241]], + [[9407], [6715], [6949], [8094]], [[8473], [694], [5142], [2479]], + [[8310], [1327], [1654], [4567]], [[6897], [3218], [2013], [1241]], + [[4372], [9456], [8204], [6695]], [[6897], [3218], [2013], [1241]], + [[8473], [694], [5142], [2479]], [[4372], [9456], [8204], [6695]], + [[4372], [9456], [8204], [6695]], [[8473], [694], [5142], [2479]], + [[9407], [6715], [6949], [8094]], [[9369], [4525], [8935], [9210]], + [[4372], [9456], [8204], [6695]], [[4372], [9456], [8204], [6695]], + [[9369], [4525], [8935], [9210]], [[6897], [3218], [2013], [1241]], + [[9038], [7951], [5953], [8657]], [[9407], [6715], [6949], [8094]], + [[9662], [9217], [1129], [8487]], [[9369], [4525], [8935], [9210]], + [[9038], [7951], [5953], [8657]], [[9662], [9217], [1129], [8487]], + [[9369], [4525], [8935], [9210]], [[1719], [5986], [9919], [3421]], + [[4372], [9456], [8204], [6695]], [[9038], [7951], [5953], [8657]] + ] + self.out_seq = np.array(self.out_seq) + + def test_check_output(self): + self.check_output() + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_interpolate_op.py b/python/paddle/fluid/tests/unittests/test_interpolate_op.py new file mode 100644 index 0000000000000000000000000000000000000000..9748d094cda6ee9dc649d95d1ca7f1c4b55d1031 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_interpolate_op.py @@ -0,0 +1,335 @@ +# 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 +from op_test import OpTest +import paddle.fluid.core as core + + +def nearest_neighbor_interp_np(X, + out_h, + out_w, + out_size=None, + actual_shape=None): + """nearest neighbor interpolation implement in shape [N, C, H, W]""" + if out_size is not None: + out_h = out_size[0] + out_w = out_size[1] + if actual_shape is not None: + out_h = actual_shape[0] + out_w = actual_shape[1] + n, c, in_h, in_w = X.shape + + ratio_h = ratio_w = 0.0 + if out_h > 1: + ratio_h = (in_h - 1.0) / (out_h - 1.0) + if out_w > 1: + ratio_w = (in_w - 1.0) / (out_w - 1.0) + + out = np.zeros((n, c, out_h, out_w)) + for i in range(out_h): + in_i = int(ratio_h * i + 0.5) + for j in range(out_w): + in_j = int(ratio_w * j + 0.5) + out[:, :, i, j] = X[:, :, in_i, in_j] + + return out.astype(X.dtype) + + +def bilinear_interp_np(input, out_h, out_w, out_size=None, actual_shape=None): + """bilinear interpolation implement in shape [N, C, H, W]""" + if out_size is not None: + out_h = out_size[0] + out_w = out_size[1] + if actual_shape is not None: + out_h = actual_shape[0] + out_w = actual_shape[1] + batch_size, channel, in_h, in_w = input.shape + if out_h > 1: + ratio_h = (in_h - 1.0) / (out_h - 1.0) + else: + ratio_h = 0.0 + if out_w > 1: + ratio_w = (in_w - 1.0) / (out_w - 1.0) + else: + ratio_w = 0.0 + + out = np.zeros((batch_size, channel, out_h, out_w)) + for i in range(out_h): + h = int(ratio_h * i) + hid = 1 if h < in_h - 1 else 0 + h1lambda = ratio_h * i - h + h2lambda = 1.0 - h1lambda + for j in range(out_w): + w = int(ratio_w * j) + wid = 1 if w < in_w - 1 else 0 + w1lambda = ratio_w * j - w + w2lambda = 1.0 - w1lambda + + out[:, :, i, j] = h2lambda*(w2lambda*input[:, :, h, w] + + w1lambda*input[:, :, h, w+wid]) + \ + h1lambda*(w2lambda*input[:, :, h+hid, w] + + w1lambda*input[:, :, h+hid, w+wid]) + return out.astype(input.dtype) + + +INTERPOLATE_FUNCS = { + 'bilinear': bilinear_interp_np, + 'nearest': nearest_neighbor_interp_np, +} + + +class TestInterpolateOp(OpTest): + def setUp(self): + self.out_size = None + self.actual_shape = None + self.init_test_case() + self.op_type = "interpolate" + input_np = np.random.random(self.input_shape).astype("float32") + + output_np = INTERPOLATE_FUNCS[self.interp_method]( + input_np, self.out_h, self.out_w, self.out_size, self.actual_shape) + self.inputs = {'X': input_np} + if self.out_size is not None: + self.inputs['OutSize'] = self.out_size + if self.actual_shape is not None: + self.inputs['OutSize'] = self.actual_shape + self.attrs = { + 'out_h': self.out_h, + 'out_w': self.out_w, + 'interp_method': self.interp_method + } + self.outputs = {'Out': output_np} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out', in_place=True) + + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [2, 3, 4, 4] + self.out_h = 2 + self.out_w = 2 + self.out_size = np.array([3, 3]).astype("int32") + + +class TestBilinearInterpCase1(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [4, 1, 7, 8] + self.out_h = 1 + self.out_w = 1 + + +class TestBilinearInterpCase2(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [3, 3, 9, 6] + self.out_h = 12 + self.out_w = 12 + + +class TestBilinearInterpCase3(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [1, 1, 128, 64] + self.out_h = 64 + self.out_w = 128 + + +class TestBilinearInterpCase4(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [4, 1, 7, 8] + self.out_h = 1 + self.out_w = 1 + self.out_size = np.array([2, 2]).astype("int32") + + +class TestBilinearInterpCase5(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [3, 3, 9, 6] + self.out_h = 12 + self.out_w = 12 + self.out_size = np.array([11, 11]).astype("int32") + + +class TestBilinearInterpCase6(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [1, 1, 128, 64] + self.out_h = 64 + self.out_w = 128 + self.out_size = np.array([65, 129]).astype("int32") + + +class TestBilinearInterpActualShape(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [3, 2, 32, 16] + self.out_h = 64 + self.out_w = 32 + self.out_size = np.array([66, 40]).astype("int32") + + +class TestBilinearInterpBigScale(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [4, 4, 64, 32] + self.out_h = 100 + self.out_w = 50 + self.out_size = np.array([101, 51]).astype('int32') + + +class TestInterpolateOpUint8(OpTest): + def setUp(self): + self.out_size = None + self.actual_shape = None + self.init_test_case() + self.op_type = "interpolate" + input_np = np.random.randint( + low=0, high=256, size=self.input_shape).astype("uint8") + output_np = INTERPOLATE_FUNCS[self.interp_method]( + input_np, self.out_h, self.out_w, self.out_size, self.actual_shape) + self.inputs = {'X': input_np} + if self.out_size is not None: + self.inputs['OutSize'] = self.out_size + self.attrs = { + 'out_h': self.out_h, + 'out_w': self.out_w, + 'interp_method': self.interp_method + } + self.outputs = {'Out': output_np} + + def test_check_output(self): + self.check_output_with_place(place=core.CPUPlace(), atol=1) + + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [1, 3, 9, 6] + self.out_h = 10 + self.out_w = 9 + + +class TestBilinearInterpCase1Uint8(TestInterpolateOpUint8): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [2, 3, 128, 64] + self.out_h = 120 + self.out_w = 50 + + +class TestBilinearInterpCase2Uint8(TestInterpolateOpUint8): + def init_test_case(self): + self.interp_method = 'bilinear' + self.input_shape = [4, 1, 7, 8] + self.out_h = 5 + self.out_w = 13 + self.out_size = np.array([6, 15]).astype("int32") + + +class TestNearestNeighborInterpCase1(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [4, 1, 7, 8] + self.out_h = 1 + self.out_w = 1 + + +class TestNearestNeighborInterpCase2(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [3, 3, 9, 6] + self.out_h = 12 + self.out_w = 12 + + +class TestNearestNeighborInterpCase3(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [1, 1, 128, 64] + self.out_h = 64 + self.out_w = 128 + + +class TestNearestNeighborInterpCase4(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [4, 1, 7, 8] + self.out_h = 1 + self.out_w = 1 + self.out_size = np.array([2, 2]).astype("int32") + + +class TestNearestNeighborInterpCase5(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [3, 3, 9, 6] + self.out_h = 12 + self.out_w = 12 + self.out_size = np.array([11, 11]).astype("int32") + + +class TestNearestNeighborInterpCase6(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [1, 1, 128, 64] + self.out_h = 64 + self.out_w = 128 + self.out_size = np.array([65, 129]).astype("int32") + + +class TestNearestNeighborInterpActualShape(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [3, 2, 32, 16] + self.out_h = 64 + self.out_w = 32 + self.out_size = np.array([66, 40]).astype("int32") + + +class TestNearestNeighborInterpBigScale(TestInterpolateOp): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [4, 4, 64, 32] + self.out_h = 100 + self.out_w = 50 + self.out_size = np.array([101, 51]).astype('int32') + + +class TestNearestNeighborInterpCase1Uint8(TestInterpolateOpUint8): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [2, 3, 128, 64] + self.out_h = 120 + self.out_w = 50 + + +class TestNearestNeighborInterpCase2Uint8(TestInterpolateOpUint8): + def init_test_case(self): + self.interp_method = 'nearest' + self.input_shape = [4, 1, 7, 8] + self.out_h = 5 + self.out_w = 13 + self.out_size = np.array([6, 15]).astype("int32") + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 1d8d0b55f0c5d7cffa01a100847bdf48b6d7023d..a8fa5436c43d2f05f632b920f67d43d837d28da9 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -194,6 +194,14 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(layers.sequence_expand(x=x, y=y, ref_level=1)) print(str(program)) + def test_sequence_unpad(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[10, 5], dtype='float32') + length = layers.data(name='length', shape=[1], dtype='int64') + self.assertIsNotNone(layers.sequence_unpad(x=x, length=length)) + print(str(program)) + def test_lstm_unit(self): program = Program() with program_guard(program): @@ -240,6 +248,17 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(layers.softmax(hid)) print(str(program)) + def test_space_to_depth(self): + program = Program() + with program_guard(program): + data = layers.data( + name='data', + shape=[32, 9, 6, 6], + append_batch_size=False, + dtype='float32') + self.assertIsNotNone(layers.space_to_depth(data, 3)) + print(str(program)) + def test_sequence_unsqueeze(self): program = Program() with program_guard(program): @@ -350,6 +369,10 @@ class TestBook(unittest.TestCase): with program_guard(program): x = layers.data(name='x', shape=[16], dtype='float32') y = layers.data(name='label', shape=[1], dtype='int64') + loss, softmax = layers.softmax_with_cross_entropy( + x, y, return_softmax=True) + self.assertIsNotNone(loss) + self.assertIsNotNone(softmax) loss = layers.softmax_with_cross_entropy(x, y) self.assertIsNotNone(loss) print(str(program)) @@ -406,6 +429,19 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(out) print(str(program)) + def test_sequence_slice(self): + program = Program() + with program_guard(program): + import numpy as np + seqs = layers.data( + name='x', shape=[10, 5], dtype='float32', lod_level=1) + offset = layers.assign(input=np.array([[0, 1]]).astype('int32')) + length = layers.assign(input=np.array([[2, 1]]).astype('int32')) + out = layers.sequence_slice( + input=seqs, offset=offset, length=length) + self.assertIsNotNone(out) + print(str(program)) + def test_lod_reset(self): program = Program() with program_guard(program): @@ -444,6 +480,16 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(output) print(str(program)) + def test_roi_align(self): + program = Program() + with program_guard(program): + x = layers.data(name="x", shape=[256, 30, 30], dtype="float32") + rois = layers.data( + name="rois", shape=[4], dtype="float32", lod_level=1) + output = layers.roi_align(x, rois, 14, 14, 0.5, 2) + self.assertIsNotNone(output) + print(str(program)) + def test_resize_bilinear(self): program = Program() with program_guard(program): @@ -454,6 +500,16 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(output) print(str(program)) + def test_resize_nearest(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[3, 9, 6], dtype="float32") + output = layers.resize_nearest(x, out_shape=[12, 12]) + self.assertIsNotNone(output) + output = layers.resize_nearest(x, scale=3) + self.assertIsNotNone(output) + print(str(program)) + def test_polygon_box_transform(self): program = Program() with program_guard(program): @@ -834,6 +890,41 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(out) print(str(program)) + def test_grid_sampler(self): + program = Program() + with program_guard(program): + x = layers.data(name='x', shape=[3, 5, 7], dtype='float32') + grid = layers.data(name='grid', shape=[5, 7, 2], dtype='float32') + out = layers.grid_sampler(x, grid) + self.assertIsNotNone(out) + print(str(program)) + + def test_affine_grid(self): + program = Program() + with program_guard(program): + data = layers.data(name='data', shape=[2, 3, 3], dtype="float32") + out, ids = layers.argsort(input=data, axis=1) + + theta = layers.data(name="theta", shape=[2, 3], dtype="float32") + out_shape = layers.data( + name="out_shape", shape=[-1], dtype="float32") + data_0 = layers.affine_grid(theta, out_shape) + data_1 = layers.affine_grid(theta, [5, 3, 28, 28]) + + self.assertIsNotNone(data_0) + self.assertIsNotNone(data_1) + print(str(program)) + + def test_bilinear_tensor_product_layer(self): + program = Program() + with program_guard(program): + data = layers.data(name='data', shape=[4], dtype="float32") + + theta = layers.data(name="theta", shape=[5], dtype="float32") + out = layers.bilinear_tensor_product(data, theta, 6) + + print(str(program)) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py index 48b52a5412eb99fbc7a5c8534a766ede4954e849..a0358f8b401e301312b5b9c0b18733d4275045e3 100644 --- a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py +++ b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py @@ -55,6 +55,46 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainers, trainer_id): exe.run(pserver_prog) +def run_pserver_with_empty_block(use_cuda, sync_mode, ip, port, trainers, + trainer_id): + x = fluid.layers.data(name='x', shape=[1], dtype='float32') + y_predict = fluid.layers.fc(input=x, size=1, act=None, bias_attr=False) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + + # loss function + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + + # optimizer + sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) + sgd_optimizer.minimize(avg_cost) + + place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() + exe = fluid.Executor(place) + + ps1 = ip + ":" + str(int(port) + 1) + ps2 = ip + ":" + port + pserver_endpoints = ps1 + "," + ps2 + + config = fluid.DistributeTranspilerConfig() + config.slice_var_up = False + t = fluid.DistributeTranspiler(config=config) + t.transpile( + trainer_id, + pservers=pserver_endpoints, + trainers=trainers, + sync_mode=sync_mode) + pserver_prog = t.get_pserver_program(ps2) + + # pserver2 have no parameter + assert (len(pserver_prog.blocks) == 2) + assert (len(pserver_prog.blocks[1].ops) == 0) + + pserver_startup = t.get_startup_program(ps2, pserver_prog) + exe.run(pserver_startup) + exe.run(pserver_prog) + + class TestListenAndServOp(OpTest): def setUp(self): self.ps_timeout = 5 @@ -63,9 +103,9 @@ class TestListenAndServOp(OpTest): self.trainers = 1 self.trainer_id = 0 - def _start_pserver(self, use_cuda, sync_mode): + def _start_pserver(self, use_cuda, sync_mode, pserver_func): p = Process( - target=run_pserver, + target=pserver_func, args=(use_cuda, sync_mode, self.ip, self.port, self.trainers, self.trainer_id)) p.daemon = True @@ -92,7 +132,24 @@ class TestListenAndServOp(OpTest): def test_handle_signal_in_serv_op(self): # run pserver on CPU in sync mode - p1 = self._start_pserver(False, True) + p1 = self._start_pserver(False, True, run_pserver) + self._wait_ps_ready(p1.pid) + + # raise SIGTERM to pserver + os.kill(p1.pid, signal.SIGINT) + p1.join() + + # run pserver on CPU in async mode + p2 = self._start_pserver(False, False, run_pserver) + self._wait_ps_ready(p2.pid) + + # raise SIGTERM to pserver + os.kill(p2.pid, signal.SIGTERM) + p2.join() + + def test_list_and_serv_run_empty_optimize_block(self): + # run pserver on CPU in sync mode + p1 = self._start_pserver(False, True, run_pserver_with_empty_block) self._wait_ps_ready(p1.pid) # raise SIGTERM to pserver @@ -100,7 +157,7 @@ class TestListenAndServOp(OpTest): p1.join() # run pserver on CPU in async mode - p2 = self._start_pserver(False, False) + p2 = self._start_pserver(False, False, run_pserver_with_empty_block) self._wait_ps_ready(p2.pid) # raise SIGTERM to pserver diff --git a/python/paddle/fluid/tests/unittests/test_mean_op.py b/python/paddle/fluid/tests/unittests/test_mean_op.py index ff338f0e0037307e81a92eed804096c9a2a87361..beae909e9b4c88eb7ddfbbe4e5ad2cf583a953ef 100644 --- a/python/paddle/fluid/tests/unittests/test_mean_op.py +++ b/python/paddle/fluid/tests/unittests/test_mean_op.py @@ -17,14 +17,20 @@ from __future__ import print_function import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core class TestMeanOp(OpTest): def setUp(self): self.op_type = "mean" - self.inputs = {'X': np.random.random((10, 10)).astype("float32")} + self.dtype = np.float32 + self.init_dtype_type() + self.inputs = {'X': np.random.random((10, 10)).astype(self.dtype)} self.outputs = {'Out': np.mean(self.inputs["X"])} + def init_dtype_type(self): + pass + def test_check_output(self): self.check_output() @@ -32,5 +38,23 @@ class TestMeanOp(OpTest): self.check_grad(['X'], 'Out') +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestFP16MeanOp(TestMeanOp): + def init_dtype_type(self): + self.dtype = np.float16 + + def test_check_output(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-3) + + def test_checkout_grad(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['X'], 'Out', max_relative_error=0.8) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_merge_ids_op.py b/python/paddle/fluid/tests/unittests/test_merge_ids_op.py index 26ce7024117162e8bad403a9d8b8518c27578c83..b109e4ea62669c735128f4824eb9d02ad43900e0 100644 --- a/python/paddle/fluid/tests/unittests/test_merge_ids_op.py +++ b/python/paddle/fluid/tests/unittests/test_merge_ids_op.py @@ -22,15 +22,28 @@ from op_test import OpTest class TestMergeIdsOp(OpTest): def setUp(self): self.op_type = "merge_ids" - ids = np.array([[0], [2], [2], [3], [5], [5], [6]]).astype('int64') - x0 = np.array([[0.1, 0.2], [0.2, 0.3], [0.3, 0.4]]).astype('float32') - x1 = np.array([]).astype('float32') - x2 = np.array([[0.4, 0.5], [0.4, 0.5], [0.5, 0.6], - [0.5, 0.6]]).astype('float32') - out = np.array([[0.1, 0.2], [0.4, 0.5], [0.4, 0.5], [0.2, 0.3], - [0.5, 0.6], [0.5, 0.6], [0.3, 0.4]]).astype('float32') - self.inputs = {'Ids': ids, "X": [('x0', x0), ('x1', x1), ('x2', x2)]} - self.outputs = {'Out': out} + ids1 = np.array([[0], [2], [5], [6]]).astype('int64') + ids2 = np.array([[0], [2], [2], [3]]).astype('int64') + + rows1 = np.array([[0], [2]]).astype('int64') + rows2 = np.array([[3], [5]]).astype('int64') + rows3 = np.array([[6]]).astype('int64') + + x0 = np.array([[0.1, 0.2], [0.2, 0.3]]).astype('float32') + x1 = np.array([[0.3, 0.4], [0.4, 0.5]]).astype('float32') + x2 = np.array([[0.5, 0.6]]).astype('float32') + + out1 = np.array( + [[0.1, 0.2], [0.2, 0.3], [0.4, 0.5], [0.5, 0.6]]).astype('float32') + out2 = np.array( + [[0.1, 0.2], [0.2, 0.3], [0.2, 0.3], [0.3, 0.4]]).astype('float32') + + self.inputs = { + 'Ids': [('ids1', ids1), ('ids2', ids2)], + "Rows": [('rows1', rows1), ('rows2', rows2), ('rows3', rows3)], + "X": [('x0', x0), ('x1', x1), ('x2', x2)] + } + self.outputs = {'Out': [('out1', out1), ('out2', out2)]} def test_check_output(self): self.check_output() diff --git a/python/paddle/fluid/tests/unittests/test_metrics.py b/python/paddle/fluid/tests/unittests/test_metrics.py new file mode 100644 index 0000000000000000000000000000000000000000..ec27884cae2b0462951f6597b1b83e58d1c8af5d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_metrics.py @@ -0,0 +1,49 @@ +# 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 unittest + +import paddle.fluid as fluid +from paddle.fluid.framework import Program, program_guard + + +class TestMetricsDetectionMap(unittest.TestCase): + def test_detection_map(self): + program = fluid.Program() + with program_guard(program): + detect_res = fluid.layers.data( + name='detect_res', + shape=[10, 6], + append_batch_size=False, + dtype='float32') + label = fluid.layers.data( + name='label', + shape=[10, 1], + append_batch_size=False, + dtype='float32') + box = fluid.layers.data( + name='bbox', + shape=[10, 4], + append_batch_size=False, + dtype='float32') + map_eval = fluid.metrics.DetectionMAP( + detect_res, label, box, class_num=21) + cur_map, accm_map = map_eval.get_map_var() + self.assertIsNotNone(cur_map) + self.assertIsNotNone(accm_map) + print(str(program)) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_momentum_op.py b/python/paddle/fluid/tests/unittests/test_momentum_op.py index 7137fd0fdb7c503492107da684b95989037eb872..cf4346cf2e7a099334ec273546901a91d0ad925d 100644 --- a/python/paddle/fluid/tests/unittests/test_momentum_op.py +++ b/python/paddle/fluid/tests/unittests/test_momentum_op.py @@ -16,6 +16,8 @@ from __future__ import print_function import unittest import numpy as np +import paddle.fluid.core as core +from paddle.fluid.op import Operator from op_test import OpTest @@ -88,5 +90,136 @@ class TestMomentumOp2(OpTest): self.check_output() +class TestLarsMomentumOp(OpTest): + def setUp(self): + self.op_type = "lars_momentum" + + param = np.random.random((123, 321)).astype("float32") + grad = np.random.random((123, 321)).astype("float32") + velocity = np.zeros((123, 321)).astype("float32") + learning_rate = np.array([0.001]).astype("float32") + mu = 0.0001 + lars_coeff = 0.001 + lars_weight_decay = 0.0005 + + self.inputs = { + 'Param': param, + 'Grad': grad, + 'Velocity': velocity, + 'LearningRate': learning_rate + } + + self.attrs = { + 'mu': mu, + 'lars_coeff': lars_coeff, + 'lars_weight_decay': lars_weight_decay + } + + pnorm = np.sqrt(np.square(param).sum()) + gnorm = np.sqrt(np.square(grad).sum()) + local_lr = learning_rate * lars_coeff * pnorm / ( + gnorm + lars_weight_decay * param) + velocity_out = mu * velocity + local_lr * (grad + lars_weight_decay * + param) + param_out = param - velocity_out + + self.outputs = {'ParamOut': param_out, 'VelocityOut': velocity_out} + + def test_check_output(self): + self.check_output() + + +class TestSparseMomentumOp(unittest.TestCase): + def setUp(self): + self.use_nesterov = False + + def check_with_place(self, place): + self.init_kernel() + scope = core.Scope() + # create and initialize Grad Variable + height = 10 + rows = [0, 4, 7] + row_numel = 12 + mu = 1.0 + use_nesterov = self.use_nesterov + + # create and initialize Param Variable + param = scope.var('Param').get_tensor() + param_array = np.full((height, row_numel), 5.0).astype("float32") + param.set(param_array, place) + param_out = scope.var("ParamOut").get_tensor() + param_out_array = np.full((height, row_numel), 0.0).astype("float32") + param_out.set(param_out_array, place) + + grad_selected_rows = scope.var('Grad').get_selected_rows() + grad_selected_rows.set_height(height) + grad_selected_rows.set_rows(rows) + grad_np_array = np.ones((len(rows), row_numel)).astype("float32") + grad_np_array[0, 0] = 2.0 + grad_np_array[2, 8] = 4.0 + grad_tensor = grad_selected_rows.get_tensor() + grad_tensor.set(grad_np_array, place) + + velocity = scope.var('Velocity').get_tensor() + velocity_np_array = np.ones((height, row_numel)).astype("float32") + velocity.set(velocity_np_array, place) + velocity_out = scope.var('VelocityOut').get_tensor() + velocity_out_np_array = np.full((height, row_numel), + 0.0).astype("float32") + velocity_out.set(velocity_out_np_array, place) + + # create and initialize LeraningRate Variable + lr = scope.var('LearningRate').get_tensor() + lr_array = np.full((1), 2.0).astype("float32") + lr.set(lr_array, place) + + # create and run operator + op = Operator( + "momentum", + Param='Param', + Grad='Grad', + Velocity='Velocity', + ParamOut='ParamOut', + VelocityOut='VelocityOut', + LearningRate='LearningRate', + mu=mu, + use_nesterov=use_nesterov) + op.run(scope, place) + + # get and compare result + param_out_np_array = np.array(param_out) + velocity_out_np_array = np.array(velocity_out) + + # TODO(dzh): add a more suitable general numpy interface + # for sparse update. + _grad_np_array = np.full((height, row_numel), 0.0).astype("float32") + for i in range(len(rows)): + _grad_np_array[rows[i]] = grad_np_array[i] + _velocity_out = mu * velocity_np_array + _grad_np_array + _param = param_array + if use_nesterov: + _param_out = _param - (_grad_np_array + _velocity_out * mu + ) * lr_array + else: + _param_out = _param - lr_array * _velocity_out + self.assertTrue((_velocity_out == velocity_out_np_array).all()) + self.assertTrue((_param_out == param_out_np_array).all()) + + def init_kernel(self): + pass + + def test_sparse_momentum(self): + places = [core.CPUPlace()] + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) + for place in places: + self.check_with_place(place) + + +class TestSparseMomentumOp2(TestSparseMomentumOp): + def init_kernel(self): + self.use_nesterov = True + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_mul_op.py b/python/paddle/fluid/tests/unittests/test_mul_op.py index fca4ffa88b79ebfad009b436d440e86ddceaaed7..d54326714acf47bd5d2abd2d919b0e3b0cab3546 100644 --- a/python/paddle/fluid/tests/unittests/test_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_mul_op.py @@ -23,12 +23,17 @@ from op_test import OpTest class TestMulOp(OpTest): def setUp(self): self.op_type = "mul" + self.dtype = np.float32 + self.init_dtype_type() self.inputs = { - 'X': np.random.random((2, 5)).astype("float32"), - 'Y': np.random.random((5, 3)).astype("float32") + 'X': np.random.random((2, 5)).astype(self.dtype), + 'Y': np.random.random((5, 3)).astype(self.dtype) } self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])} + def init_dtype_type(self): + pass + def test_check_output(self): self.check_output() @@ -47,9 +52,11 @@ class TestMulOp(OpTest): class TestMulOp2(OpTest): def setUp(self): self.op_type = "mul" + self.dtype = np.float32 + self.init_dtype_type() self.inputs = { - 'X': np.random.random((3, 4, 4, 3)).astype("float32"), - 'Y': np.random.random((2, 6, 1, 2, 3)).astype("float32") + 'X': np.random.random((3, 4, 4, 3)).astype(self.dtype), + 'Y': np.random.random((2, 6, 1, 2, 3)).astype(self.dtype) } self.attrs = { 'x_num_col_dims': 2, @@ -60,6 +67,9 @@ class TestMulOp2(OpTest): result = result.reshape(3, 4, 1, 2, 3) self.outputs = {'Out': result} + def init_dtype_type(self): + pass + def test_check_output(self): self.check_output() @@ -75,40 +85,76 @@ class TestMulOp2(OpTest): ['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y')) -class TestFP16MulOp1(OpTest): - def setUp(self): - self.op_type = "mul" - x = np.random.random((3, 5)).astype("float16") - y = np.random.random((5, 4)).astype("float16") - self.inputs = {'X': x.view(np.float16), 'Y': y.view(np.float16)} - self.outputs = {'Out': np.dot(x, y)} +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestFP16MulOp1(TestMulOp): + def init_dtype_type(self): + self.dtype = np.float16 def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-1) + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-1) + def test_check_grad_normal(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['X', 'Y'], 'Out', max_relative_error=0.5) -class TestFP16MulOp2(OpTest): - def setUp(self): - self.op_type = "mul" - x = np.random.random((3, 4, 4, 3)).astype("float16") - y = np.random.random((2, 6, 1, 2, 3)).astype("float16") - self.inputs = {'X': x.view(np.float16), 'Y': y.view(np.float16)} - self.attrs = { - 'x_num_col_dims': 2, - 'y_num_col_dims': 2, - } - result = np.dot(x.reshape(3 * 4, 4 * 3), y.reshape(2 * 6, 1 * 2 * 3)) - result = result.reshape(3, 4, 1, 2, 3) - self.outputs = {'Out': result} + def test_check_grad_ingore_x(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['Y'], + 'Out', + max_relative_error=0.5, + no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['X'], + 'Out', + max_relative_error=0.5, + no_grad_set=set('Y')) + + +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestFP16MulOp2(TestMulOp2): + def init_dtype_type(self): + self.dtype = np.float16 def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=2e-1) + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-1) + + def test_check_grad_normal(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['X', 'Y'], 'Out', max_relative_error=0.9) + + def test_check_grad_ingore_x(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['Y'], + 'Out', + max_relative_error=0.5, + no_grad_set=set("X")) + + def test_check_grad_ingore_y(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ['X'], + 'Out', + max_relative_error=0.9, + no_grad_set=set('Y')) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py index 6d6917300cb66afcc8a0c509986a0f26be8b1f09..84b0aad8acb096a32f625e32fb640599f2882d97 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py @@ -16,6 +16,7 @@ from __future__ import print_function import paddle.dataset.conll05 as conll05 import paddle.fluid as fluid +import paddle.fluid.core as core import unittest import paddle import numpy as np @@ -174,39 +175,39 @@ class TestCRFModel(unittest.TestCase): print(pe.run(feed=feeder.feed(cur_batch), fetch_list=[avg_cost.name])[0]) - @unittest.skip(reason="CI hangs") def test_update_sparse_parameter_all_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce - self.check_network_convergence( - is_sparse=True, build_strategy=build_strategy, use_cuda=True) + if core.is_compiled_with_cuda(): + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy, use_cuda=True) self.check_network_convergence( is_sparse=True, build_strategy=build_strategy, use_cuda=False) - @unittest.skip(reason="CI hangs") def test_update_dense_parameter_all_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce - self.check_network_convergence( - is_sparse=False, build_strategy=build_strategy, use_cuda=True) + if core.is_compiled_with_cuda(): + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy, use_cuda=True) self.check_network_convergence( is_sparse=False, build_strategy=build_strategy, use_cuda=False) - @unittest.skip(reason="CI hangs") def test_update_sparse_parameter_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce - self.check_network_convergence( - is_sparse=True, build_strategy=build_strategy, use_cuda=True) + if core.is_compiled_with_cuda(): + self.check_network_convergence( + is_sparse=True, build_strategy=build_strategy, use_cuda=True) self.check_network_convergence( is_sparse=True, build_strategy=build_strategy, use_cuda=False) - @unittest.skip(reason="CI hangs") def test_update_dense_parameter_reduce(self): build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce - self.check_network_convergence( - is_sparse=False, build_strategy=build_strategy, use_cuda=True) + if core.is_compiled_with_cuda(): + self.check_network_convergence( + is_sparse=False, build_strategy=build_strategy, use_cuda=True) self.check_network_convergence( is_sparse=False, build_strategy=build_strategy, use_cuda=False) diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_dry_run.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_dry_run.py new file mode 100644 index 0000000000000000000000000000000000000000..18d95c94ad36316b7149eb5412260b40a57ac002 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_dry_run.py @@ -0,0 +1,80 @@ +# 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 paddle.fluid as fluid +import unittest +import logging +import six + + +class TestBase(unittest.TestCase): + def main(self, + network_func, + iter=10, + iter_per_pe=10, + use_gpu=True, + use_experimental_executor=False): + if use_gpu and not fluid.core.is_compiled_with_cuda(): + logging.warning( + "Paddle is not compiled with CUDA, skip GPU unittests") + return + + main_prog = fluid.Program() + startup_prog = fluid.Program() + scope = fluid.Scope() + with fluid.program_guard(main_prog, startup_prog): + with fluid.scope_guard(scope): + loss = network_func() + fluid.Executor( + fluid.CUDAPlace(0) + if use_gpu else fluid.CPUPlace()).run(startup_prog) + + for _ in six.moves.xrange(iter): + exe_strategy = fluid.ExecutionStrategy() + exe_strategy._dry_run = True + exe_strategy.use_experimental_executor = use_experimental_executor + pe = fluid.ParallelExecutor( + use_cuda=use_gpu, + loss_name=loss.name, + main_program=main_prog, + exec_strategy=exe_strategy) + for _ in six.moves.xrange(iter_per_pe): + pe.run([]) + + +class TestMNISTDryRun(TestBase): + def test_mnist_dry_run(self): + for use_gpu in (False, True): + for use_experimental_executor in (False, True): + self.main( + network_func=TestMNISTDryRun.network_func, + use_gpu=use_gpu, + use_experimental_executor=use_experimental_executor) + + @staticmethod + def network_func(): + img = fluid.layers.data(name='img', shape=[784], dtype='float32') + label = fluid.layers.data(name='label', shape=[1], dtype='int64') + hidden = img + for _ in six.moves.xrange(10): + hidden = fluid.layers.fc(input=img, size=200, act='tanh') + prediction = fluid.layers.fc(input=hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + avg_loss = fluid.layers.mean(loss) + fluid.optimizer.Adam().minimize(avg_loss) + return avg_loss + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py index af3745987aa3eae96968bdc6b5c9cd951e9ca6fa..3eecc4670152e72443f731c71d7db67ca8e02e72 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py @@ -14,30 +14,18 @@ from __future__ import print_function -from parallel_executor_test_base import TestParallelExecutorBase -import paddle.fluid as fluid -import paddle.fluid.core as core -import numpy as np -import paddle -import paddle.dataset.mnist as mnist import unittest -import os -MNIST_RECORDIO_FILE = "./mnist_test_pe.recordio" +import numpy as np +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): - if use_feed: - img = fluid.layers.data(name='image', shape=[784], dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - else: - reader = fluid.layers.open_files( - filenames=[MNIST_RECORDIO_FILE], - shapes=[[-1, 784], [-1, 1]], - lod_levels=[0, 0], - dtypes=['float32', 'int64']) - reader = fluid.layers.io.double_buffer(reader) - img, label = fluid.layers.read_file(reader) + 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( @@ -53,17 +41,8 @@ def simple_fc_net(use_feed): def fc_with_batchnorm(use_feed): - if use_feed: - img = fluid.layers.data(name='image', shape=[784], dtype='float32') - label = fluid.layers.data(name='label', shape=[1], dtype='int64') - else: - reader = fluid.layers.open_files( - filenames=[MNIST_RECORDIO_FILE], - shapes=[[-1, 784], [-1, 1]], - lod_levels=[0, 0], - dtypes=['float32', 'int64']) - reader = fluid.layers.io.double_buffer(reader) - img, label = fluid.layers.read_file(reader) + 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(1): @@ -88,19 +67,6 @@ class TestMNIST(TestParallelExecutorBase): @classmethod def setUpClass(cls): os.environ['CPU_NUM'] = str(4) - # Convert mnist to recordio file - with fluid.program_guard(fluid.Program(), fluid.Program()): - reader = paddle.batch(mnist.train(), batch_size=4) - feeder = fluid.DataFeeder( - feed_list=[ # order is image and label - fluid.layers.data( - name='image', shape=[784]), - fluid.layers.data( - name='label', shape=[1], dtype='int64'), - ], - place=fluid.CPUPlace()) - fluid.recordio_writer.convert_reader_to_recordio_file( - MNIST_RECORDIO_FILE, reader, feeder) def _init_data(self): np.random.seed(5) @@ -111,10 +77,6 @@ class TestMNIST(TestParallelExecutorBase): def _compare_reduce_and_allreduce(self, model, use_cuda): if use_cuda and not core.is_compiled_with_cuda(): return - self.check_network_convergence( - model, use_cuda=use_cuda, use_reduce=True) - self.check_network_convergence( - model, use_cuda=use_cuda, allow_op_delay=True, use_reduce=True) img, label = self._init_data() @@ -140,9 +102,6 @@ class TestMNIST(TestParallelExecutorBase): def check_simple_fc_convergence(self, use_cuda, use_reduce=False): if use_cuda and not core.is_compiled_with_cuda(): return - self.check_network_convergence(simple_fc_net, use_cuda=use_cuda) - self.check_network_convergence( - simple_fc_net, use_cuda=use_cuda, allow_op_delay=True) img, label = self._init_data() @@ -199,8 +158,6 @@ class TestMNIST(TestParallelExecutorBase): if use_cuda and not core.is_compiled_with_cuda(): return - self.check_network_convergence(fc_with_batchnorm, use_cuda=use_cuda) - img, label = self._init_data() self.check_network_convergence( diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py index cc2d692e18430eb48e6e800106eab0c3739d3f53..e7a56bb6386a812e43e5c1b5c08cd0682aa9223a 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py @@ -232,6 +232,46 @@ class TestResnet(TestParallelExecutorBase): for loss in zip(all_reduce_last_loss, reduce_last_loss): self.assertAlmostEquals(loss[0], loss[1], delta=delta2) + if not use_cuda: + return + + all_reduce_first_loss_seq, all_reduce_last_loss_seq = self.check_network_convergence( + model, + feed_dict={"image": img, + "label": label}, + iter=iter, + batch_size=batch_size, + use_cuda=use_cuda, + use_reduce=False, + optimizer=optimizer, + enable_sequential_execution=True) + + reduce_first_loss_seq, reduce_last_loss_seq = self.check_network_convergence( + model, + feed_dict={"image": img, + "label": label}, + iter=iter, + batch_size=batch_size, + use_cuda=use_cuda, + use_reduce=True, + optimizer=optimizer, + enable_sequential_execution=True) + + for loss in zip(all_reduce_first_loss, all_reduce_first_loss_seq): + self.assertAlmostEquals(loss[0], loss[1], delta=1e-6) + for loss in zip(all_reduce_last_loss, all_reduce_last_loss_seq): + self.assertAlmostEquals(loss[0], loss[1], delta=delta2) + + for loss in zip(reduce_first_loss, reduce_first_loss_seq): + self.assertAlmostEquals(loss[0], loss[1], delta=1e-6) + for loss in zip(reduce_last_loss, reduce_last_loss_seq): + self.assertAlmostEquals(loss[0], loss[1], delta=delta2) + + for loss in zip(all_reduce_first_loss_seq, reduce_first_loss_seq): + self.assertAlmostEquals(loss[0], loss[1], delta=1e-6) + for loss in zip(all_reduce_last_loss_seq, reduce_last_loss_seq): + self.assertAlmostEquals(loss[0], loss[1], delta=delta2) + def _check_resnet_convergence(self, model, use_cuda=True, diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py index a55b2002ed989d4588716202a37aa6f4139825ea..3827743908c1d76931572277323d1dd5ddd05523 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py @@ -173,6 +173,8 @@ class TestTransformer(TestParallelExecutorBase): def test_main(self): if core.is_compiled_with_cuda(): self.check_network_convergence(transformer, use_cuda=True) + self.check_network_convergence( + transformer, use_cuda=True, enable_sequential_execution=True) self.check_network_convergence(transformer, use_cuda=False, iter=5) diff --git a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py index dfedf8190f75ec26532f281338f076ca0c7d83af..7f266056a9d98be1a6f67473be65a74957f943e9 100644 --- a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py +++ b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py @@ -37,7 +37,7 @@ def PolygonBoxRestore(input): indexes = indexes.repeat( [batch_size], axis=0) # [batch_size, geo_channels/2, 2, h, w] return indexes.reshape( - input.shape) - input # [batch_size, geo_channels, h, w] + input.shape) * 4 - input # [batch_size, geo_channels, h, w] class TestPolygonBoxRestoreOp(OpTest): diff --git a/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py index 14d7ed9057d622b136056e1b5bbbe57f9a04d5d7..19f29c78269eef6414342a200b719e498c2037d2 100644 --- a/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py @@ -15,10 +15,10 @@ from __future__ import print_function import unittest -from test_pool2d_op import TestPool2d_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5 +from test_pool2d_op import TestPool2D_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5 -class TestMKLDNNCase1(TestPool2d_Op): +class TestMKLDNNCase1(TestPool2D_Op): def init_kernel_type(self): self.use_mkldnn = True diff --git a/python/paddle/fluid/tests/unittests/test_pool2d_op.py b/python/paddle/fluid/tests/unittests/test_pool2d_op.py index 26969bd5230afdac83a943d2dc21094a0972d60a..47b2e71a4e52a327831fde7494bd7a2306b6f2ea 100644 --- a/python/paddle/fluid/tests/unittests/test_pool2d_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool2d_op.py @@ -26,7 +26,8 @@ def max_pool2D_forward_naive(x, strides, paddings, global_pool=0, - ceil_mode=False): + ceil_mode=False, + exclusive=True): N, C, H, W = x.shape if global_pool == 1: ksize = [H, W] @@ -54,7 +55,8 @@ def avg_pool2D_forward_naive(x, strides, paddings, global_pool=0, - ceil_mode=False): + ceil_mode=False, + exclusive=True): N, C, H, W = x.shape if global_pool == 1: ksize = [H, W] @@ -73,12 +75,13 @@ def avg_pool2D_forward_naive(x, c_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) x_masked = x[:, :, r_start:r_end, c_start:c_end] - out[:, :, i, j] = np.sum(x_masked, axis=(2, 3)) / ( - (r_end - r_start) * (c_end - c_start)) + field_size = ((r_end - r_start) * (c_end - c_start)) if exclusive \ + else (ksize[0] * ksize[1]) + out[:, :, i, j] = np.sum(x_masked, axis=(2, 3)) / field_size return out -class TestPool2d_Op(OpTest): +class TestPool2D_Op(OpTest): def setUp(self): self.op_type = "pool2d" self.use_cudnn = False @@ -89,12 +92,13 @@ class TestPool2d_Op(OpTest): self.init_kernel_type() self.init_pool_type() self.init_ceil_mode() + self.init_exclusive() if self.global_pool: self.paddings = [0 for _ in range(len(self.paddings))] input = np.random.random(self.shape).astype(self.dtype) - output = self.pool2D_forward_naive(input, self.ksize, self.strides, - self.paddings, self.global_pool, - self.ceil_mode).astype(self.dtype) + output = self.pool2D_forward_naive( + input, self.ksize, self.strides, self.paddings, self.global_pool, + self.ceil_mode, self.exclusive).astype(self.dtype) self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(input)} self.attrs = { @@ -106,7 +110,9 @@ class TestPool2d_Op(OpTest): 'use_cudnn': self.use_cudnn, 'use_mkldnn': self.use_mkldnn, 'ceil_mode': self.ceil_mode, - 'data_format': 'AnyLayout' # TODO(dzhwinter) : should be fix latter + 'data_format': + 'AnyLayout', # TODO(dzhwinter) : should be fix latter + 'exclusive': self.exclusive } self.outputs = {'Out': output} @@ -150,8 +156,11 @@ class TestPool2d_Op(OpTest): def init_ceil_mode(self): self.ceil_mode = False + def init_exclusive(self): + self.exclusive = True -class TestCase1(TestPool2d_Op): + +class TestCase1(TestPool2D_Op): def init_test_case(self): self.shape = [2, 3, 7, 7] self.ksize = [3, 3] @@ -166,7 +175,7 @@ class TestCase1(TestPool2d_Op): self.global_pool = False -class TestCase2(TestPool2d_Op): +class TestCase2(TestPool2D_Op): def init_test_case(self): self.shape = [2, 3, 7, 7] self.ksize = [3, 3] @@ -181,7 +190,7 @@ class TestCase2(TestPool2d_Op): self.global_pool = False -class TestCase3(TestPool2d_Op): +class TestCase3(TestPool2D_Op): def init_pool_type(self): self.pool_type = "max" self.pool2D_forward_naive = max_pool2D_forward_naive @@ -199,127 +208,111 @@ class TestCase5(TestCase2): self.pool2D_forward_naive = max_pool2D_forward_naive -#--------------------test pool2d-------------------- -class TestCUDNNCase1(TestPool2d_Op): - def init_kernel_type(self): - self.use_cudnn = True - - -class TestFP16CUDNNCase1(TestPool2d_Op): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 - - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) +#--------------------test pool2d cudnn-------------------- -class TestCUDNNCase2(TestCase1): - def init_kernel_type(self): - self.use_cudnn = True +def create_test_cudnn_class(parent): + @unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") + class TestCUDNNCase(parent): + def init_kernel_type(self): + self.use_cudnn = True + cls_name = "{0}_{1}".format(parent.__name__, "CUDNNOp") + TestCUDNNCase.__name__ = cls_name + globals()[cls_name] = TestCUDNNCase -class TestFP16CUDNNCase2(TestCase1): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) +create_test_cudnn_class(TestPool2D_Op) +create_test_cudnn_class(TestCase1) +create_test_cudnn_class(TestCase2) +create_test_cudnn_class(TestCase3) +create_test_cudnn_class(TestCase4) +create_test_cudnn_class(TestCase5) +#--------------------test pool2d cudnn_fp16-------------------- -class TestCUDNNCase3(TestCase2): - def init_kernel_type(self): - self.use_cudnn = True +def create_test_cudnn_fp16_class(parent, check_grad=True): + @unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") + class TestCUDNNFp16Case(parent): + def init_kernel_type(self): + self.use_cudnn = True + self.dtype = np.float16 -class TestFP16CUDNNCase3(TestCase2): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) - def test_check_output(self): - if core.is_compiled_with_cuda(): + def test_check_grad(self): place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) - - -class TestCUDNNCase4(TestCase3): - def init_kernel_type(self): - self.use_cudnn = True - + if core.is_float16_supported( + place) and self.pool_type != "max" and check_grad: + self.check_grad_with_place( + place, set(['X']), 'Out', max_relative_error=0.07) -class TestFP16CUDNNCase4(TestCase3): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 + cls_name = "{0}_{1}".format(parent.__name__, "CUDNNFp16Op") + TestCUDNNFp16Case.__name__ = cls_name + globals()[cls_name] = TestCUDNNFp16Case - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) +create_test_cudnn_fp16_class(TestPool2D_Op) +create_test_cudnn_fp16_class(TestCase1, check_grad=False) +create_test_cudnn_fp16_class(TestCase2) +create_test_cudnn_fp16_class(TestCase3) +create_test_cudnn_fp16_class(TestCase4) +create_test_cudnn_fp16_class(TestCase5) -class TestCUDNNCase5(TestCase4): - def init_kernel_type(self): - self.use_cudnn = True +#--------------------test pool2d use ceil mode-------------------- -class TestFP16CUDNNCase5(TestCase4): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 +def create_test_cudnn_use_ceil_class(parent): + @unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") + class TestPool2DUseCeilCase(parent): + def init_kernel_type(self): + self.use_cudnn = True - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) + def init_ceil_mode(self): + self.ceil_mode = True + cls_name = "{0}_{1}".format(parent.__name__, "CUDNNOpCeilMode") + TestPool2DUseCeilCase.__name__ = cls_name + globals()[cls_name] = TestPool2DUseCeilCase -class TestCUDNNCase6(TestCase5): - def init_kernel_type(self): - self.use_cudnn = True +create_test_cudnn_use_ceil_class(TestPool2D_Op) +create_test_cudnn_use_ceil_class(TestCase1) -class TestFP16CUDNNCase6(TestCase5): - def init_kernel_type(self): - self.use_cudnn = True - self.dtype = np.float16 - def test_check_output(self): - if core.is_compiled_with_cuda(): - place = core.CUDAPlace(0) - if core.is_float16_supported(place): - self.check_output_with_place(place, atol=1e-3) +def create_test_use_ceil_class(parent): + class TestPool2DUseCeilCase(parent): + def init_ceil_mode(self): + self.ceil_mode = True + cls_name = "{0}_{1}".format(parent.__name__, "CeilModeCast") + TestPool2DUseCeilCase.__name__ = cls_name + globals()[cls_name] = TestPool2DUseCeilCase -class TestCeilModeCase1(TestCUDNNCase1): - def init_ceil_mode(self): - self.ceil_mode = True +create_test_use_ceil_class(TestCase1) +create_test_use_ceil_class(TestCase2) -class TestCeilModeCase2(TestCUDNNCase2): - def init_ceil_mode(self): - self.ceil_mode = True +class TestAvgInclude(TestCase2): + def init_exclusive(self): + self.exclusive = False -class TestCeilModeCase3(TestCase1): - def init_ceil_mode(self): - self.ceil_mode = True +class TestCUDNNAvgInclude(TestCase2): + def init_kernel_type(self): + self.use_cudnn = True -class TestCeilModeCase4(TestCase2): - def init_ceil_mode(self): - self.ceil_mode = True + def init_exclusive(self): + self.exclusive = False if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_pool3d_op.py b/python/paddle/fluid/tests/unittests/test_pool3d_op.py index 77045c1307baead3711d58ed368dfa5f2acc3699..f05f8ccb3985be162d89da099496d5b2baf4afdc 100644 --- a/python/paddle/fluid/tests/unittests/test_pool3d_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool3d_op.py @@ -26,7 +26,8 @@ def max_pool3D_forward_naive(x, strides, paddings, global_pool=0, - ceil_mode=False): + ceil_mode=False, + exclusive=True): N, C, D, H, W = x.shape if global_pool == 1: ksize = [D, H, W] @@ -60,7 +61,8 @@ def avg_pool3D_forward_naive(x, strides, paddings, global_pool=0, - ceil_mode=False): + ceil_mode=False, + exclusive=True): N, C, D, H, W = x.shape if global_pool == 1: ksize = [D, H, W] @@ -85,8 +87,10 @@ def avg_pool3D_forward_naive(x, w_end = np.min((j * strides[1] + ksize[1] - paddings[1], W)) x_masked = x[:, :, d_start:d_end, h_start:h_end, w_start:w_end] - out[:, :, k, i, j] = np.sum(x_masked, axis=(2, 3, 4)) / ( - (d_end - d_start) * (h_end - h_start) * (w_end - w_start)) + field_size = (d_end - d_start) * (h_end - h_start) * (w_end - w_start) \ + if exclusive else ksize[0] * ksize[1] * ksize[2] + out[:, :, k, i, j] = np.sum(x_masked, axis=(2, 3, + 4)) / field_size return out @@ -100,13 +104,14 @@ class TestPool3d_Op(OpTest): self.init_kernel_type() self.init_pool_type() self.init_ceil_mode() + self.init_exclusive() if self.global_pool: self.paddings = [0 for _ in range(len(self.paddings))] input = np.random.random(self.shape).astype(self.dtype) - output = self.pool3D_forward_naive(input, self.ksize, self.strides, - self.paddings, self.global_pool, - self.ceil_mode).astype(self.dtype) + output = self.pool3D_forward_naive( + input, self.ksize, self.strides, self.paddings, self.global_pool, + self.ceil_mode, self.exclusive).astype(self.dtype) self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(input)} self.attrs = { @@ -117,7 +122,9 @@ class TestPool3d_Op(OpTest): 'global_pooling': self.global_pool, 'use_cudnn': self.use_cudnn, 'ceil_mode': self.ceil_mode, - 'data_format': 'AnyLayout' # TODO(dzhwinter) : should be fix latter + 'data_format': + 'AnyLayout', # TODO(dzhwinter) : should be fix latter + 'exclusive': self.exclusive } self.outputs = {'Out': output} @@ -161,6 +168,9 @@ class TestPool3d_Op(OpTest): def init_ceil_mode(self): self.ceil_mode = False + def init_exclusive(self): + self.exclusive = True + class TestCase1(TestPool3d_Op): def init_test_case(self): @@ -333,5 +343,15 @@ class TestCeilModeCase4(TestCase2): self.ceil_mode = True +class TestAvgInclude(TestCase2): + def init_exclusive(self): + self.exclusive = False + + +class TestCUDNNAvgInclude(TestCUDNNCase3): + def init_exclusive(self): + self.exclusive = False + + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_py_reader_lod_level_share.py b/python/paddle/fluid/tests/unittests/test_py_reader_lod_level_share.py new file mode 100644 index 0000000000000000000000000000000000000000..55dc3a7aa341ff09eb3d7d219cd1c23427e25da1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_py_reader_lod_level_share.py @@ -0,0 +1,43 @@ +# 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 paddle.fluid as fluid +import unittest + + +class TestLoDLevelShare(unittest.TestCase): + def setUp(self): + self.use_double_buffer = False + + def test_lod_level_share(self): + reader = fluid.layers.py_reader( + capacity=16, + shapes=([-1, 256], [-1, 512], [-1, 100]), + dtypes=('float32', 'int64', 'double'), + lod_levels=(1, 2, 0), + use_double_buffer=self.use_double_buffer) + + x, y, z = fluid.layers.read_file(reader) + self.assertEqual(x.lod_level, 1) + self.assertEqual(y.lod_level, 2) + self.assertEqual(z.lod_level, 0) + + +class TestLoDLevelShare2(TestLoDLevelShare): + def setUp(self): + self.use_double_buffer = True + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_py_reader_pin_memory.py b/python/paddle/fluid/tests/unittests/test_py_reader_pin_memory.py new file mode 100644 index 0000000000000000000000000000000000000000..b913127ad625eb25de3ec36edd2161019ed09749 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_py_reader_pin_memory.py @@ -0,0 +1,130 @@ +# 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 paddle +import paddle.fluid as fluid +import paddle.fluid.core as core +import numpy as np +from threading import Thread + + +def user_reader(inputs): + def _reader(): + for d in inputs: + yield d + + return _reader + + +def batch_feeder(batch_reader, pin_memory=False, img_dtype="float32"): + def _feeder(): + for batch_data in batch_reader(): + sample_batch = [] + label_batch = [] + for sample, label in batch_data: + sample_batch.append(sample) + label_batch.append([label]) + tensor = core.LoDTensor() + label = core.LoDTensor() + place = core.CUDAPinnedPlace() if pin_memory else core.CPUPlace() + tensor.set(np.array(sample_batch, dtype=img_dtype), place) + label.set(np.array(label_batch, dtype="int64"), place) + yield [tensor, label] + + return _feeder + + +class TestPyReader(unittest.TestCase): + def setUp(self): + self.capacity = 10 + self.shapes = [(-1, 3, 2, 1), (-1, 1)] + self.lod_levels = [0, 0] + self.dtypes = ['float32', 'int64'] + + def test_pin_memory_pyreader(self): + with fluid.program_guard(fluid.Program(), fluid.Program()): + place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda( + ) else fluid.CPUPlace() + executor = fluid.Executor(place) + + data_file = fluid.layers.py_reader( + capacity=self.capacity, + dtypes=self.dtypes, + lod_levels=self.lod_levels, + shapes=self.shapes) + # feed_queue = data_file.queue + read_out_data = fluid.layers.read_file(data_file) + + self.inputs = [] + for _ in range(10): + sample = np.random.uniform( + low=0, high=1, size=[3, 2, 1]).astype("float32") + label = np.random.uniform( + low=0, high=10, size=[1]).astype("int64") + self.inputs.append((sample, label)) + + self.input_tensors = [] + for d, l in batch_feeder( + paddle.batch( + user_reader(self.inputs), batch_size=2), + pin_memory=True + if fluid.core.is_compiled_with_cuda() else False)(): + ta = fluid.LoDTensorArray() + ta.append(d) + ta.append(l) + self.input_tensors.append(ta) + + self.batched_inputs = [] + for batch in paddle.batch(user_reader(self.inputs), batch_size=2)(): + feed_d = [] + feed_l = [] + for d, l in batch: + feed_d.append(d) + feed_l.append([l]) + self.batched_inputs.append([feed_d, feed_l]) + + data_file.decorate_tensor_provider( + batch_feeder( + paddle.batch( + user_reader(self.inputs), batch_size=2), + pin_memory=True + if fluid.core.is_compiled_with_cuda() else False)) + + executor.run(fluid.default_startup_program()) + self.outputs = [] + + data_file.start() + for _ in self.input_tensors: + self.outputs.append( + executor.run(fetch_list=list(read_out_data))) + data_file.reset() + self.validate() + + def validate(self): + self.assertEqual(len(self.batched_inputs), len(self.outputs)) + for in_data_list, out_data_list in zip(self.batched_inputs, + self.outputs): + self.assertEqual(len(in_data_list), len(out_data_list)) + in_data_list_np = [ + np.array(in_lod_tensor) for in_lod_tensor in in_data_list + ] + for in_data, out_data in zip(in_data_list_np, out_data_list): + self.assertTrue((in_data == out_data).all()) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py b/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py index b7fad9b3a60632adb564e1d155a3d935706b467f..d94494e219c5f348a08b4c3c2d111674ea6badf3 100644 --- a/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py +++ b/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py @@ -53,15 +53,24 @@ def simple_fc_net(in_size, hidden_sizes, batch_size, queue_capacity, - use_double_buffer=False): - reader = fluid.layers.py_reader( - capacity=queue_capacity, - shapes=[[-1, in_size], [-1, 1]], - lod_levels=[0, 0], - dtypes=['float32', 'int64'], - use_double_buffer=False) - feed_queue = reader.queue - reader = fluid.layers.batch(reader, batch_size=batch_size) + use_double_buffer=False, + use_feed_list=True): + if use_feed_list: + data = fluid.layers.data(name="data", dtype='float32', shape=[in_size]) + label = fluid.layers.data(name='label', dtype='int64', shape=[1]) + py_reader = fluid.layers.create_py_reader_by_data( + capacity=queue_capacity, + use_double_buffer=False, + feed_list=[data, label]) + else: + py_reader = fluid.layers.py_reader( + capacity=queue_capacity, + shapes=[[-1, in_size], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64'], + use_double_buffer=False) + feed_queue = py_reader.queue + reader = fluid.layers.batch(py_reader, batch_size=batch_size) if use_double_buffer: reader = fluid.layers.double_buffer(reader) @@ -83,7 +92,7 @@ def simple_fc_net(in_size, optimizer = fluid.optimizer.Adam() optimizer.minimize(loss) - return in_data, label, loss, optimizer, feed_queue + return in_data, label, loss, optimizer, feed_queue, py_reader class TestPyReaderUsingExecutor(unittest.TestCase): @@ -100,16 +109,22 @@ class TestPyReaderUsingExecutor(unittest.TestCase): if core.is_compiled_with_cuda() else [False]): for use_parallel_executor in [False, True]: for use_double_buffer in [False, True]: - print('Test Parameters:'), - print({ - 'use_cuda': use_cuda, - 'use_parallel_executor': use_parallel_executor, - 'use_double_buffer': use_double_buffer - }) - self.main(use_cuda, use_parallel_executor, - use_double_buffer) - - def random_reader(self): + for use_feed_list in [False, True]: + for use_decorate_paddle_reader in [False, True]: + print('Test Parameters:'), + print({ + 'use_cuda': use_cuda, + 'use_parallel_executor': use_parallel_executor, + 'use_double_buffer': use_double_buffer, + 'use_feed_list': use_feed_list, + 'use_decorate_paddle_reader': + use_decorate_paddle_reader + }) + self.main(use_cuda, use_parallel_executor, + use_double_buffer, use_feed_list, + use_decorate_paddle_reader) + + def tensor_reader(self, use_decorate_paddle_reader): def reader(): self.inputs = [] cnt = 0 @@ -133,34 +148,43 @@ class TestPyReaderUsingExecutor(unittest.TestCase): elif not self.use_double_buffer: break - yield tensors + if use_decorate_paddle_reader: + yield [(in_data, label)] + else: + yield tensors cnt += 1 - yield None + if not use_decorate_paddle_reader: + yield None return reader def main(self, use_cuda=True, use_parallel_executor=False, - use_double_buffer=False): + use_double_buffer=False, + use_feed_list=False, + use_decorate_paddle_reader=False): assert not use_cuda or use_cuda and core.is_compiled_with_cuda() self.use_cuda = use_cuda self.use_parallel_executor = use_parallel_executor self.use_double_buffer = use_double_buffer + self.use_feed_list = use_feed_list + self.use_decorate_paddle_reader = use_decorate_paddle_reader startup_program = fluid.Program() main_program = fluid.Program() with fluid.program_guard(main_program, startup_program): - in_data, label, loss, optimizer, feed_queue = simple_fc_net( + in_data, label, loss, optimizer, feed_queue, py_reader = simple_fc_net( in_size=self.in_size, class_num=self.class_num, hidden_sizes=self.hidden_sizes, batch_size=self.batch_size, queue_capacity=self.queue_capacity, - use_double_buffer=self.use_double_buffer) + use_double_buffer=self.use_double_buffer, + use_feed_list=self.use_feed_list) place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace() @@ -178,10 +202,14 @@ class TestPyReaderUsingExecutor(unittest.TestCase): main_exe = startup_exe self.batch_size_times = 1 - reader = self.random_reader() - thread = threading.Thread( - target=feed_data, args=(feed_queue, reader)) - thread.start() + reader = self.tensor_reader(use_decorate_paddle_reader) + if use_decorate_paddle_reader: + py_reader.decorate_paddle_reader(reader) + py_reader.start() + else: + thread = threading.Thread( + target=feed_data, args=(feed_queue, reader)) + thread.start() self.outputs = [] for _ in range(self.iterations): diff --git a/python/paddle/fluid/tests/unittests/test_ref_by_trainer_id_op.py b/python/paddle/fluid/tests/unittests/test_ref_by_trainer_id_op.py new file mode 100644 index 0000000000000000000000000000000000000000..e4872829edb325edcadbd4e1aefaf5014b800d3a --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_ref_by_trainer_id_op.py @@ -0,0 +1,36 @@ +# 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 unittest +import numpy as np +from op_test import OpTest + + +class TestRefByTrainerIdOp(OpTest): + def setUp(self): + self.op_type = "ref_by_trainer_id" + param_baks = [("x%d" % x, np.random.random((10, 10)).astype("float32")) + for x in range(10)] + self.inputs = { + 'X': param_baks, + 'TrainerId': np.array([8]).astype("int64") + } + self.outputs = {'Out': param_baks[8][1]} + + def test_check_output(self): + self.check_output() + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_regularizer.py b/python/paddle/fluid/tests/unittests/test_regularizer.py index 6727335c6059161d235a64a1b90d36b84004f9b3..20f91cf4485f2e79c20fe90143c8b7deebb9fc49 100644 --- a/python/paddle/fluid/tests/unittests/test_regularizer.py +++ b/python/paddle/fluid/tests/unittests/test_regularizer.py @@ -55,7 +55,7 @@ class TestL2DecayRegularizer(unittest.TestCase): params_grads = optimizer.append_regularization_ops(params_grads) self.assertEqual(len(params_grads), 1) self.assertEqual(len(block.ops), count_ops + 2) - self.assertEqual(block.ops[-1].type, 'elementwise_add') + self.assertEqual(block.ops[-1].type, 'sum') self.assertEqual(block.ops[-2].type, 'scale') @@ -92,7 +92,7 @@ class TestL1DecayRegularizer(unittest.TestCase): params_grads = optimizer.append_regularization_ops(params_grads) self.assertEqual(len(params_grads), 1) self.assertEqual(len(block.ops), count_ops + 3) - self.assertEqual(block.ops[-1].type, 'elementwise_add') + self.assertEqual(block.ops[-1].type, 'sum') self.assertEqual(block.ops[-2].type, 'scale') self.assertEqual(block.ops[-3].type, 'sign') diff --git a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py index 70848e4e2239e2be160bb0c1a28a5aecd01a87dc..eb12bc741767340a3e7e3580a8b95065d4267693 100644 --- a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py +++ b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py @@ -19,33 +19,76 @@ import unittest import numpy as np import paddle.fluid.core as core from paddle.fluid.op import Operator +import paddle.fluid as fluid + + +def create_selected_rows_and_tensor(scope, place, height, row_num, + embedding_size): + sr = scope.var("@selected_rows@").get_selected_rows() + tensor = scope.var("grad").get_tensor() + + rows = np.random.random_integers( + low=0, high=height - 1, size=[row_num, ]).astype('int64') + sr_val = np.random.random(size=[row_num, embedding_size]).astype('float32') + + sr.set_height(height) + sr.set_rows(rows) + sr.get_tensor().set(sr_val, place) + + tensor_val = np.zeros(shape=[height, embedding_size], dtype='float32') + for i in range(row_num): + row = rows[i] + tensor_val[row, :] = tensor_val[row, :] + sr_val[i, :] + + tensor.set(tensor_val, place) + return tensor_val, sr_val class TestBase(unittest.TestCase): - def setup(self, centered, epsilon=1e-6): + def setup(self, + place, + is_sparse, + centered, + size, + row_num=None, + epsilon=1e-6): np.random.seed(5) # fix seed + self.scope = fluid.global_scope() + self.place = place + self.param_name = "param" - self.param = np.random.random((123, 321)).astype("float32") + self.param = np.random.random(size).astype("float32") self.mean_square_name = "mean_square" - self.mean_square = np.random.random((123, 321)).astype("float32") + self.mean_square = np.random.uniform( + low=1, high=2, size=size).astype("float32") self.mean_grad_name = "mean_grad" - self.mean_grad = np.random.random((123, 321)).astype("float32") + self.mean_grad = np.random.random(size).astype("float32") self.lr_name = "lr" self.learning_rate = np.array([0.01]).astype("float32") self.grad_name = "grad" - self.grad = np.random.random((123, 321)).astype("float32") + + self.is_sparse = is_sparse + if self.is_sparse: + self.grad_sr_name = "@selected_rows@" + self.grad, self.grad_sr = create_selected_rows_and_tensor( + self.scope, place, size[0], row_num, size[1]) + else: + self.grad = np.random.random(size).astype("float32") + grad_tensor = self.scope.var(self.grad_name).get_tensor() + grad_tensor.set(self.grad, place) self.moment_name = "moment" - self.moment = np.zeros((123, 321)).astype("float32") + self.moment = np.random.uniform( + low=0, high=1, size=size).astype("float32") self.epsilon = epsilon self.decay = 0.9 - self.momentum = 0.0 + self.momentum = 0.1 self.centered = centered self.ms_out = self.decay * self.mean_square + (1 - self.decay @@ -61,118 +104,122 @@ class TestBase(unittest.TestCase): self.param_out = self.param - self.moment_out - def check(self, - actual_t, - expect_t, - place, - out_name, - atol=1e-5, - equal_nan=False): - self.assertTrue( - np.allclose( - actual_t, expect_t, atol=atol, equal_nan=equal_nan), - "Output (" + out_name + ") has diff at " + str(place) + "\nExpect " - + str(expect_t) + "\n" + "But Got" + str(actual_t)) - - -class TestRmspropOp(TestBase): - def check_with_place(self, place, centered, epsilon): - self.setup(centered, epsilon) - scope = core.Scope() - # create and initialize Param Variable - param = scope.var(self.param_name).get_tensor() - param.set(self.param, place) + self.param_tensor = self.scope.var(self.param_name).get_tensor() + self.param_tensor.set(self.param, place) - mean_square = scope.var(self.mean_square_name).get_tensor() - mean_square.set(self.mean_square, place) + self.mean_square_tensor = self.scope.var( + self.mean_square_name).get_tensor() + self.mean_square_tensor.set(self.mean_square, place) - lr = scope.var(self.lr_name).get_tensor() + lr = self.scope.var(self.lr_name).get_tensor() lr.set(self.learning_rate, place) - grad = scope.var(self.grad_name).get_tensor() - grad.set(self.grad, place) + self.moment_tensor = self.scope.var(self.moment_name).get_tensor() + self.moment_tensor.set(self.moment, place) - moment = scope.var(self.moment_name).get_tensor() - moment.set(self.moment, place) + if self.centered: + self.mean_grad_tensor = self.scope.var( + self.mean_grad_name).get_tensor() + self.mean_grad_tensor.set(self.mean_grad, place) - # create and run sgd operator + def check(self, actual_t, expect_t, place, out_name, atol=1e-5): + self.assertTrue( + np.allclose( + actual_t, expect_t, atol=atol), + "Output (" + out_name + ") has diff at " + str(place) + "\nExpect " + + str(expect_t) + "\n" + "But Got" + str(actual_t)) - if self.centered: - mean_grad = scope.var(self.mean_grad_name).get_tensor() - mean_grad.set(self.mean_grad, place) - - rmsprop_op = Operator( - "rmsprop", - Param=self.param_name, - Grad=self.grad_name, - MeanSquare=self.mean_square_name, - MeanGrad=self.mean_grad_name, - Moment=self.moment_name, - LearningRate=self.lr_name, - ParamOut=self.param_name, - MeanSquareOut=self.mean_square_name, - MomentOut=self.moment_name, - MeanGradOut=self.mean_grad_name, - epsilon=self.epsilon, - decay=self.decay, - momentum=self.momentum, - centered=True) - else: - rmsprop_op = Operator( - "rmsprop", - Param=self.param_name, - Grad=self.grad_name, - MeanSquare=self.mean_square_name, - Moment=self.moment_name, - LearningRate=self.lr_name, - ParamOut=self.param_name, - MeanSquareOut=self.mean_square_name, - MomentOut=self.moment_name, - epsilon=self.epsilon, - decay=self.decay, - momentum=self.momentum, - centered=False) - - rmsprop_op.run(scope, place) - - atol = 1e-5 - equal_nan = False + +class TestRmspropOp(TestBase): + def check_with_place(self, + place, + is_sparse, + centered, + size, + row_num=None, + epsilon=1e-6): + self.setup(place, is_sparse, centered, size, row_num, epsilon) + self.run_and_check() + + def run_and_check(self): + grad_name = self.grad_sr_name if self.is_sparse else self.grad_name + + kwargs = { + 'Param': self.param_name, + 'Grad': grad_name, + 'MeanSquare': self.mean_square_name, + 'Moment': self.moment_name, + 'LearningRate': self.lr_name, + 'ParamOut': self.param_name, + 'MeanSquareOut': self.mean_square_name, + 'MomentOut': self.moment_name, + 'epsilon': self.epsilon, + 'decay': self.decay, + 'momentum': self.momentum, + 'centered': self.centered + } if self.centered: - atol = 1e-3 - equal_nan = True + kwargs['MeanGrad'] = self.mean_grad_name + kwargs['MeanGradOut'] = self.mean_grad_name + + rmsprop_op = Operator('rmsprop', **kwargs) + atol = 1e-6 + + rmsprop_op.run(self.scope, self.place) self.check( - np.array(mean_square), self.ms_out, place, self.mean_square_name) + np.array(self.mean_square_tensor), + self.ms_out, + self.place, + self.mean_square_name, + atol=atol) self.check( - np.array(moment), + np.array(self.moment_tensor), self.moment_out, - place, + self.place, self.moment_name, - atol=atol, - equal_nan=equal_nan) + atol=atol) self.check( - np.array(param), + np.array(self.param_tensor), self.param_out, - place, + self.place, self.param_name, - atol=atol, - equal_nan=equal_nan) + atol=atol) if self.centered: self.check( - np.array(mean_grad), self.mg_out, place, self.mean_grad_name) + np.array(self.mean_grad_tensor), self.mg_out, self.place, + self.mean_grad_name) def test_rmsprop(self): places = [core.CPUPlace()] if core.is_compiled_with_cuda(): places.append(core.CUDAPlace(0)) + + size = (128, 320) for place in places: - self.check_with_place(place, False, 1e-6) - self.check_with_place(place, False, 1e-10) - self.check_with_place(place, True, 1e-6) - self.check_with_place(place, True, 1e-10) + for centered in [False, True]: + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, is_sparse=False, centered=centered, size=size) + + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, + is_sparse=True, + centered=centered, + row_num=512, + size=size) + + with fluid.scope_guard(core.Scope()): + self.check_with_place( + place, + is_sparse=True, + centered=centered, + row_num=60, + size=size) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_roi_align_op.py b/python/paddle/fluid/tests/unittests/test_roi_align_op.py new file mode 100644 index 0000000000000000000000000000000000000000..1a252ea547e4d93d83f64fa9cdb3605eeef0a3cf --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_roi_align_op.py @@ -0,0 +1,170 @@ +# 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 math +import sys +from op_test import OpTest + + +class TestROIAlignOp(OpTest): + def set_data(self): + self.init_test_case() + self.make_rois() + self.calc_roi_align() + self.inputs = {'X': self.x, 'ROIs': (self.rois[:, 1:5], self.rois_lod)} + self.attrs = { + 'spatial_scale': self.spatial_scale, + 'pooled_height': self.pooled_height, + 'pooled_width': self.pooled_width, + 'sampling_ratio': self.sampling_ratio + } + + self.outputs = {'Out': self.out_data} + + def init_test_case(self): + self.batch_size = 3 + self.channels = 3 + self.height = 8 + self.width = 6 + + # n, c, h, w + self.x_dim = (self.batch_size, self.channels, self.height, self.width) + + self.spatial_scale = 1.0 / 2.0 + self.pooled_height = 2 + self.pooled_width = 2 + self.sampling_ratio = -1 + + self.x = np.random.random(self.x_dim).astype('float32') + + def pre_calc(self, x_i, roi_xmin, roi_ymin, roi_bin_grid_h, roi_bin_grid_w, + bin_size_h, bin_size_w): + count = roi_bin_grid_h * roi_bin_grid_w + bilinear_pos = np.zeros( + [self.channels, self.pooled_height, self.pooled_width, count, 4], + np.float32) + bilinear_w = np.zeros( + [self.pooled_height, self.pooled_width, count, 4], np.float32) + for ph in range(self.pooled_width): + for pw in range(self.pooled_height): + c = 0 + for iy in range(roi_bin_grid_h): + y = roi_ymin + ph * bin_size_h + (iy + 0.5) * \ + bin_size_h / roi_bin_grid_h + for ix in range(roi_bin_grid_w): + x = roi_xmin + pw * bin_size_w + (ix + 0.5) * \ + bin_size_w / roi_bin_grid_w + if y < -1.0 or y > self.height or \ + x < -1.0 or x > self.width: + continue + if y <= 0: + y = 0 + if x <= 0: + x = 0 + y_low = int(y) + x_low = int(x) + if y_low >= self.height - 1: + y = y_high = y_low = self.height - 1 + else: + y_high = y_low + 1 + if x_low >= self.width - 1: + x = x_high = x_low = self.width - 1 + else: + x_high = x_low + 1 + ly = y - y_low + lx = x - x_low + hy = 1 - ly + hx = 1 - lx + for ch in range(self.channels): + bilinear_pos[ch, ph, pw, c, 0] = x_i[ch, y_low, + x_low] + bilinear_pos[ch, ph, pw, c, 1] = x_i[ch, y_low, + x_high] + bilinear_pos[ch, ph, pw, c, 2] = x_i[ch, y_high, + x_low] + bilinear_pos[ch, ph, pw, c, 3] = x_i[ch, y_high, + x_high] + bilinear_w[ph, pw, c, 0] = hy * hx + bilinear_w[ph, pw, c, 1] = hy * lx + bilinear_w[ph, pw, c, 2] = ly * hx + bilinear_w[ph, pw, c, 3] = ly * lx + c = c + 1 + return bilinear_pos, bilinear_w + + def calc_roi_align(self): + self.out_data = np.zeros( + (self.rois_num, self.channels, self.pooled_height, + self.pooled_width)).astype('float32') + + for i in range(self.rois_num): + roi = self.rois[i] + roi_batch_id = int(roi[0]) + x_i = self.x[roi_batch_id] + roi_xmin = roi[1] * self.spatial_scale + roi_ymin = roi[2] * self.spatial_scale + roi_xmax = roi[3] * self.spatial_scale + roi_ymax = roi[4] * self.spatial_scale + roi_width = max(roi_xmax - roi_xmin, 1) + roi_height = max(roi_ymax - roi_ymin, 1) + bin_size_h = float(roi_height) / float(self.pooled_height) + bin_size_w = float(roi_width) / float(self.pooled_width) + roi_bin_grid_h = self.sampling_ratio if self.sampling_ratio > 0 else \ + math.ceil(roi_height / self.pooled_height) + roi_bin_grid_w = self.sampling_ratio if self.sampling_ratio > 0 else \ + math.ceil(roi_width / self.pooled_width) + count = int(roi_bin_grid_h * roi_bin_grid_w) + pre_size = count * self.pooled_width * self.pooled_height + bilinear_pos, bilinear_w = self.pre_calc(x_i, roi_xmin, roi_ymin, + int(roi_bin_grid_h), + int(roi_bin_grid_w), + bin_size_h, bin_size_w) + for ch in range(self.channels): + align_per_bin = (bilinear_pos[ch] * bilinear_w).sum(axis=-1) + output_val = align_per_bin.mean(axis=-1) + self.out_data[i, ch, :, :] = output_val + + def make_rois(self): + rois = [] + self.rois_lod = [[]] + for bno in range(self.batch_size): + self.rois_lod[0].append(bno + 1) + for i in range(bno + 1): + x1 = np.random.random_integers( + 0, self.width // self.spatial_scale - self.pooled_width) + y1 = np.random.random_integers( + 0, self.height // self.spatial_scale - self.pooled_height) + + x2 = np.random.random_integers(x1 + self.pooled_width, + self.width // self.spatial_scale) + y2 = np.random.random_integers( + y1 + self.pooled_height, self.height // self.spatial_scale) + + roi = [bno, x1, y1, x2, y2] + rois.append(roi) + self.rois_num = len(rois) + self.rois = np.array(rois).astype("float32") + + def setUp(self): + self.op_type = "roi_align" + self.set_data() + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out') diff --git a/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py b/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py index f63dbcd3d7f6bfce3ccc1c42ae41afe42bfad003..1a2c9bb5f43d55d8e6183de0d55bfcc2b9ac3f08 100644 --- a/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py +++ b/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py @@ -50,8 +50,10 @@ def rpn_target_assign(anchor_by_gt_overlap, fg_inds, size=(len(fg_inds) - num_fg), replace=False) else: disable_inds = fg_inds[num_fg:] + labels[disable_inds] = -1 fg_inds = np.where(labels == 1)[0] + bbox_inside_weight = np.zeros((len(fg_inds), 4), dtype=np.float32) num_bg = rpn_batch_size_per_im - np.sum(labels == 1) bg_inds = np.where(anchor_to_gt_max < rpn_negative_overlap)[0] @@ -59,18 +61,27 @@ def rpn_target_assign(anchor_by_gt_overlap, enable_inds = bg_inds[np.random.randint(len(bg_inds), size=num_bg)] else: enable_inds = bg_inds[:num_bg] + + fg_fake_inds = np.array([], np.int32) + fg_value = np.array([fg_inds[0]], np.int32) + fake_num = 0 + for bg_id in enable_inds: + if bg_id in fg_inds: + fake_num += 1 + fg_fake_inds = np.hstack([fg_fake_inds, fg_value]) labels[enable_inds] = 0 + + bbox_inside_weight[fake_num:, :] = 1 fg_inds = np.where(labels == 1)[0] bg_inds = np.where(labels == 0)[0] - - loc_index = fg_inds - score_index = np.hstack((fg_inds, bg_inds)) + loc_index = np.hstack([fg_fake_inds, fg_inds]) + score_index = np.hstack([fg_inds, bg_inds]) labels = labels[score_index] assert not np.any(labels == -1), "Wrong labels with -1" - gt_inds = anchor_to_gt_argmax[fg_inds] + gt_inds = anchor_to_gt_argmax[loc_index] - return loc_index, score_index, labels, gt_inds + return loc_index, score_index, labels, gt_inds, bbox_inside_weight def get_anchor(n, c, h, w): @@ -123,9 +134,12 @@ def rpn_target_assign_in_python(all_anchors, gt_boxes_slice = gt_boxes_slice[not_crowd_inds] iou = _bbox_overlaps(inside_anchors, gt_boxes_slice) - loc_inds, score_inds, labels, gt_inds = rpn_target_assign( - iou, rpn_batch_size_per_im, rpn_positive_overlap, - rpn_negative_overlap, rpn_fg_fraction, use_random) + loc_inds, score_inds, labels, gt_inds, bbox_inside_weight = \ + rpn_target_assign(iou, rpn_batch_size_per_im, + rpn_positive_overlap, + rpn_negative_overlap, + rpn_fg_fraction, + use_random) # unmap to all anchor loc_inds = inds_inside[loc_inds] score_inds = inds_inside[score_inds] @@ -139,6 +153,7 @@ def rpn_target_assign_in_python(all_anchors, score_indexes = score_inds tgt_labels = labels tgt_bboxes = box_deltas + bbox_inside_weights = bbox_inside_weight else: loc_indexes = np.concatenate( [loc_indexes, loc_inds + i * anchor_num]) @@ -146,8 +161,10 @@ def rpn_target_assign_in_python(all_anchors, [score_indexes, score_inds + i * anchor_num]) tgt_labels = np.concatenate([tgt_labels, labels]) tgt_bboxes = np.vstack([tgt_bboxes, box_deltas]) + bbox_inside_weights = np.vstack([bbox_inside_weights, \ + bbox_inside_weight]) - return loc_indexes, score_indexes, tgt_bboxes, tgt_labels + return loc_indexes, score_indexes, tgt_bboxes, tgt_labels, bbox_inside_weights class TestRpnTargetAssignOp(OpTest): @@ -182,10 +199,12 @@ class TestRpnTargetAssignOp(OpTest): rpn_fg_fraction = 0.5 use_random = False - loc_index, score_index, tgt_bbox, labels = rpn_target_assign_in_python( - all_anchors, gt_boxes, is_crowd, im_info, lod, rpn_straddle_thresh, - rpn_batch_size_per_im, rpn_positive_overlap, rpn_negative_overlap, - rpn_fg_fraction, use_random) + loc_index, score_index, tgt_bbox, labels, bbox_inside_weights = \ + rpn_target_assign_in_python(all_anchors, gt_boxes, is_crowd, + im_info, lod, rpn_straddle_thresh, + rpn_batch_size_per_im, rpn_positive_overlap, + rpn_negative_overlap, + rpn_fg_fraction, use_random) labels = labels[:, np.newaxis] self.op_type = "rpn_target_assign" @@ -207,7 +226,8 @@ class TestRpnTargetAssignOp(OpTest): 'LocationIndex': loc_index.astype('int32'), 'ScoreIndex': score_index.astype('int32'), 'TargetBBox': tgt_bbox.astype('float32'), - 'TargetLabel': labels.astype('int32') + 'TargetLabel': labels.astype('int32'), + 'BBoxInsideWeight': bbox_inside_weights.astype('float32') } def test_check_output(self): diff --git a/python/paddle/fluid/tests/unittests/test_scale_op.py b/python/paddle/fluid/tests/unittests/test_scale_op.py index 032af6ed5ce9e1007d6775306ef4c0aefb9dcc41..9893c92ad68f4d460c4bb428bb44a30df25fd6e0 100644 --- a/python/paddle/fluid/tests/unittests/test_scale_op.py +++ b/python/paddle/fluid/tests/unittests/test_scale_op.py @@ -24,9 +24,16 @@ from paddle.fluid.op import Operator class TestScaleOp(OpTest): def setUp(self): self.op_type = "scale" - self.inputs = {'X': np.random.random((10, 10)).astype("float32")} + self.dtype = np.float32 + self.init_dtype_type() + self.inputs = {'X': np.random.random((10, 10)).astype(self.dtype)} self.attrs = {'scale': -2.3} - self.outputs = {'Out': self.inputs['X'] * self.attrs['scale']} + self.outputs = { + 'Out': self.inputs['X'] * self.dtype(self.attrs['scale']) + } + + def init_dtype_type(self): + pass def test_check_output(self): self.check_output() @@ -36,9 +43,15 @@ class TestScaleOp(OpTest): class TestScaleOpSelectedRows(unittest.TestCase): + def init_dtype_type(self): + pass + def check_with_place(self, place, in_name, out_name): scope = core.Scope() + self.dtype = np.float32 + self.init_dtype_type() + # create and initialize Grad Variable in_height = 10 in_rows = [0, 4, 7] @@ -49,7 +62,7 @@ class TestScaleOpSelectedRows(unittest.TestCase): in_selected_rows.set_height(in_height) in_selected_rows.set_rows(in_rows) in_array = np.random.random( - (len(in_rows), in_row_numel)).astype("float32") + (len(in_rows), in_row_numel)).astype(self.dtype) in_tensor = in_selected_rows.get_tensor() in_tensor.set(in_array, place) @@ -87,5 +100,41 @@ class TestScaleOpSelectedRows(unittest.TestCase): self.check_with_place(place, 'in', 'in') +# Add FP16 test +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestScaleFp16Op(TestScaleOp): + def init_dtype_type(self): + self.dtype = np.float16 + + def test_check_output(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=0.002) + + def test_check_grad(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ["X"], "Out", max_relative_error=0.05) + + +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestScaleFp16OpSelectedRows(TestScaleOpSelectedRows): + def init_dtype_type(self): + self.dtype = np.float16 + + def test_scale_selected_rows(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_with_place(place, 'in', 'out') + + def test_scale_selected_rows_inplace(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_with_place(place, 'in', 'in') + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_seq_conv.py b/python/paddle/fluid/tests/unittests/test_seq_conv.py index dcc86382e5286f354c4f2e81ead598f12c75b2c1..2285e9496768aea6f48fb7796536e8344839d862 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_conv.py +++ b/python/paddle/fluid/tests/unittests/test_seq_conv.py @@ -20,6 +20,53 @@ import random from op_test import OpTest +def seqconv(x, + lod, + filter, + context_length, + context_start, + padding_trainable=False, + padding_data=None): + [T, M] = x.shape + col = np.zeros((T, context_length * M)).astype('float32') + offset = [0] + for seq_len in lod[0]: + offset.append(offset[-1] + seq_len) + begin_pad = np.max([0, -context_start]) + for i in range(len(offset) - 1): + for j in range(context_length): + in_begin = offset[i] + context_start + j + in_end = offset[i + 1] + context_start + j + out_begin = offset[i] + out_end = offset[i + 1] + if in_begin < offset[i]: + pad_size = np.min( + [offset[i] - in_begin, offset[i + 1] - offset[i]]) + if padding_trainable: + sub_w = padding_data[j:j + pad_size, :] + col[offset[i]:offset[i] + pad_size, j * M:(j + 1) * + M] = sub_w + out_begin = offset[i] + pad_size + in_begin = offset[i] + + if in_end > offset[i + 1]: + pad_size = np.min( + [in_end - offset[i + 1], offset[i + 1] - offset[i]]) + if padding_trainable: + sub_w = padding_data[begin_pad + context_start + j - + pad_size:begin_pad + context_start + + j, :] + col[offset[i + 1] - pad_size:offset[i + 1], j * M:(j + 1) * + M] = sub_w + in_end = offset[i + 1] + out_end = offset[i + 1] - pad_size + if in_end <= in_begin: + continue + in_sub = x[in_begin:in_end, :] + col[out_begin:out_end, j * M:(j + 1) * M] += in_sub + return np.dot(col, filter) + + class TestSeqProject(OpTest): def setUp(self): self.init_test_case() @@ -66,57 +113,9 @@ class TestSeqProject(OpTest): 'paddingTrainable': self.padding_trainable, 'contextStride': self.context_stride } - out = np.zeros( - (self.input_size[0], self.output_represention)).astype('float32') + out = seqconv(x, self.lod, w, self.context_length, self.context_start, + self.padding_trainable, self.pad_data) self.outputs = {'Out': out} - self.compute() - - def compute(self): - x, lod = self.inputs['X'] - filter = self.inputs['Filter'] - pading_data = self.pad_data - out = np.zeros((self.input_size[0], self.context_length * - self.input_size[1])).astype('float32') - offset = [0] - for seq_len in lod[0]: - offset.append(offset[-1] + seq_len) - begin_pad = np.max([0, -self.context_start]) - - for i in range(len(offset) - 1): - for j in range(self.context_length): - in_begin = offset[i] + self.context_start + j - in_end = offset[i + 1] + self.context_start + j - out_begin = offset[i] - out_end = offset[i + 1] - if in_begin < offset[i]: - pad_size = np.min( - [offset[i] - in_begin, offset[i + 1] - offset[i]]) - if self.padding_trainable: - sub_w = pading_data[j:j + pad_size, :] - out[offset[i]:offset[i] + pad_size, j * self.input_size[ - 1]:(j + 1) * self.input_size[1]] = sub_w - out_begin = offset[i] + pad_size - in_begin = offset[i] - - if in_end > offset[i + 1]: - pad_size = np.min( - [in_end - offset[i + 1], offset[i + 1] - offset[i]]) - if self.padding_trainable: - sub_w = pading_data[begin_pad + self.context_start + j - - pad_size:begin_pad + - self.context_start + j, :] - out[offset[i + 1] - pad_size:offset[i + 1], j * self. - input_size[1]:(j + 1) * self.input_size[1]] = sub_w - in_end = offset[i + 1] - out_end = offset[i + 1] - pad_size - if in_end <= in_begin: - continue - - in_sub = x[in_begin:in_end, :] - out[out_begin:out_end, j * self.input_size[1]:(j + 1) * - self.input_size[1]] += in_sub - - np.dot(out, filter, out=self.outputs['Out']) def test_check_output(self): self.check_output() diff --git a/python/paddle/fluid/tests/unittests/test_seq_pool.py b/python/paddle/fluid/tests/unittests/test_seq_pool.py index 641eb03a5fbf1bb140b20cc3518cea83386fa577..a80ad5b079891efe1b0e1222b3c2455d4891d5f5 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_pool.py +++ b/python/paddle/fluid/tests/unittests/test_seq_pool.py @@ -184,6 +184,20 @@ class TestSeqMaxPool2D(TestSeqAvgPool2D): out[i] = np.reshape(np.amax(sub_x, axis=0), (3, 11)) +class TestSeqMaxPool2DInference(TestSeqMaxPool2D): + def compute(self, x, offset, out): + self.attrs = {'pooltype': "MAX", 'is_test': True} + for i in range(len(offset[0]) - 1): + sub_x = np.reshape(x[offset[0][i]:offset[0][i + 1], :], + (-1, 3 * 11)) + out[i] = np.reshape(np.amax(sub_x, axis=0), (3, 11)) + + def test_check_grad(self): + """Grad computation does not apply to Sequence MAX + Pool executed when is_test is true """ + return + + class TestSeqLastPool2D(TestSeqAvgPool2D): def compute(self, x, offset, out): self.attrs = {'pooltype': "LAST"} diff --git a/python/paddle/fluid/tests/unittests/test_sequence_reverse.py b/python/paddle/fluid/tests/unittests/test_sequence_reverse.py new file mode 100644 index 0000000000000000000000000000000000000000..eebd25e0975f1711ea86093f007212cadc6334f5 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_sequence_reverse.py @@ -0,0 +1,69 @@ +# 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 unittest +import paddle.fluid as fluid +import paddle.fluid.core as core +from op_test import OpTest +import numpy as np + + +class TestSequenceReverseBase(OpTest): + def initParameters(self): + pass + + def setUp(self): + self.size = (10, 3, 4) + self.lod = [2, 3, 5] + self.dtype = 'float32' + self.initParameters() + self.op_type = 'sequence_reverse' + self.x = np.random.random(self.size).astype(self.dtype) + self.y = self.get_output() + + self.inputs = {'X': (self.x, [self.lod, ]), } + self.outputs = {'Y': (self.y, [self.lod, ]), } + + def get_output(self): + tmp_x = np.reshape(self.x, newshape=[self.x.shape[0], -1]) + tmp_y = np.ndarray(tmp_x.shape).astype(self.dtype) + prev_idx = 0 + for cur_len in self.lod: + idx_range = range(prev_idx, prev_idx + cur_len) + tmp_y[idx_range, :] = np.flip(tmp_x[idx_range, :], 0) + prev_idx += cur_len + + return np.reshape(tmp_y, newshape=self.x.shape).astype(self.dtype) + + def test_output(self): + self.check_output(0) + + def test_grad(self): + self.check_grad(['X'], 'Y') + + +class TestSequenceReserve1(TestSequenceReverseBase): + def initParameters(self): + self.size = (12, 10) + self.lod = [4, 5, 3] + + +class TestSequenceReverse2(TestSequenceReverseBase): + def initParameters(self): + self.size = (12, 10) + self.lod = [12] + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_sequence_unpad_op.py b/python/paddle/fluid/tests/unittests/test_sequence_unpad_op.py new file mode 100644 index 0000000000000000000000000000000000000000..673b0ea180464b8b8f6f5c6e76d5c5c80f347d25 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_sequence_unpad_op.py @@ -0,0 +1,75 @@ +# 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 unittest +import six +import numpy as np +from op_test import OpTest + + +class TestSequenceUnpadOp(OpTest): + def init(self): + self.length = [2, 3, 4] + self.x_shape = (3, 5) + self.dtype = "float32" + + def compute(self): + assert len(self.length) == self.x_shape[0] + x = np.random.random(self.x_shape).astype(self.dtype) + out_lod = [self.length] + + out = x[0, 0:self.length[0]] + for i in six.moves.xrange(1, x.shape[0]): + out = np.append(out, x[i, 0:self.length[i]], axis=0) + + out_shape = (sum(self.length), ) + if len(self.x_shape) == 2: + out_shape = out_shape + (1, ) + else: + out_shape = out_shape + self.x_shape[2:] + + self.inputs = { + 'X': x, + 'Length': np.array(self.length).astype('int64').reshape(-1, 1) + } + self.outputs = {'Out': (out.reshape(out_shape), out_lod)} + + def setUp(self): + self.op_type = 'sequence_unpad' + self.init() + self.compute() + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + +class TestSequenceUnpadOp2(TestSequenceUnpadOp): + def init(self): + self.length = [2, 3, 4] + self.x_shape = (3, 5, 4, 3) + self.dtype = "float32" + + +class TestSequenceUnpadOp3(TestSequenceUnpadOp): + def init(self): + self.length = [5, 2, 3, 4] + self.x_shape = (4, 5, 3, 3, 6) + self.dtype = "float64" + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_similarity_focus_op.py b/python/paddle/fluid/tests/unittests/test_similarity_focus_op.py new file mode 100755 index 0000000000000000000000000000000000000000..b3833f05f1aa3aac7b5bcc5b6fdc138870cc8844 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_similarity_focus_op.py @@ -0,0 +1,217 @@ +# 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 paddle.fluid.core as core +from op_test import OpTest + + +class TestSimilarityFocusOp(OpTest): + def setUp(self): + self.op_type = "similarity_focus" + batch_size = 2 + x_dim, y_dim, z_dim = 3, 2, 2 + self.inputs = { + 'X': np.array([[[[0.8, 0.1], [0.4, 0.5]], [[0.9, 0.7], [0.9, 0.9]], + [[0.8, 0.9], [0.1, 0.2]]], + [[[0.2, 0.5], [0.3, 0.4]], [[0.9, 0.7], [0.8, 0.4]], + [[0.0, 0.2], [0.4, 0.7]]]]), + } + self.attrs = { + 'axis': 1, + 'indexes': [0], + } + + output = None + for batch in range(batch_size): + res = np.zeros((1, y_dim, z_dim)).astype("float32").reshape(-1) + for index in self.attrs['indexes']: + channel = self.inputs['X'][batch, index, :, :].reshape(-1).copy( + ) + tag1 = [0 for i in range(y_dim)] + tag2 = [0 for i in range(z_dim)] + cnt = 0 + for i in range(channel.size): + index = channel.argmax() + idx1 = index // z_dim + idx2 = index % z_dim + if tag1[idx1] + tag2[idx2] == 0: + tag1[idx1] = 1 + tag2[idx2] = 1 + res[index] = 1 + cnt += 1 + if cnt == min(y_dim, z_dim): + break + channel[index] = -1 + res = res.reshape(1, y_dim, z_dim).repeat([x_dim], axis=0) + res = res.reshape(1, x_dim, y_dim, z_dim) + if output is not None: + output = np.concatenate((output, res), axis=0) + else: + output = res + self.outputs = {'Out': output} + + def test_check_output(self): + self.check_output() + + +class TestSimilarityFocusOp_axis1(OpTest): + def setUp(self): + self.op_type = "similarity_focus" + batch_size = 3 + x_dim, y_dim, z_dim = 4, 5, 6 + self.inputs = { + 'X': np.random.random( + (batch_size, x_dim, y_dim, z_dim)).astype("float32"), + } + self.attrs = { + 'axis': 1, + 'indexes': [0, 3], + } + + output = None + for batch in range(batch_size): + res = np.zeros((1, y_dim, z_dim)).astype("float32").reshape(-1) + for index in self.attrs['indexes']: + channel = self.inputs['X'][batch, index, :, :].reshape(-1).copy( + ) + tag1 = [0 for i in range(y_dim)] + tag2 = [0 for i in range(z_dim)] + cnt = 0 + for i in range(channel.size): + index = channel.argmax() + idx1 = index // z_dim + idx2 = index % z_dim + if tag1[idx1] + tag2[idx2] == 0: + tag1[idx1] = 1 + tag2[idx2] = 1 + res[index] = 1 + cnt += 1 + if cnt == min(y_dim, z_dim): + break + channel[index] = -1 + res = res.reshape(1, y_dim, z_dim) + res = res.repeat([x_dim], axis=0) + res = res.reshape(1, x_dim, y_dim, z_dim) + if output is not None: + output = np.concatenate((output, res), axis=0) + else: + output = res + self.outputs = {'Out': output} + + def test_check_output(self): + self.check_output() + + +class TestSimilarityFocusOp_axis2(OpTest): + def setUp(self): + self.op_type = "similarity_focus" + batch_size = 6 + x_dim, y_dim, z_dim = 7, 8, 9 + self.inputs = { + 'X': np.random.random( + (batch_size, x_dim, y_dim, z_dim)).astype("float32"), + } + self.attrs = { + 'axis': 2, + 'indexes': [0, 3, 5], + } + + output = None + for batch in range(batch_size): + res = np.zeros((x_dim, 1, z_dim)).astype("float32").reshape(-1) + for index in self.attrs['indexes']: + channel = self.inputs['X'][batch, :, index, :].reshape(-1).copy( + ) + tag1 = [0 for i in range(x_dim)] + tag2 = [0 for i in range(z_dim)] + cnt = 0 + for i in range(channel.size): + index = channel.argmax() + idx1 = index // z_dim + idx2 = index % z_dim + if tag1[idx1] + tag2[idx2] == 0: + tag1[idx1] = 1 + tag2[idx2] = 1 + res[index] = 1 + cnt += 1 + if cnt == min(x_dim, z_dim): + break + channel[index] = -1 + res = res.reshape(x_dim, 1, z_dim) + res = res.repeat([y_dim], axis=1) + res = res.reshape(1, x_dim, y_dim, z_dim) + if output is not None: + output = np.concatenate((output, res), axis=0) + else: + output = res + self.outputs = {'Out': output} + + def test_check_output(self): + self.check_output() + + +class TestSimilarityFocusOp_axis3(OpTest): + def setUp(self): + self.op_type = "similarity_focus" + batch_size = 64 + x_dim, y_dim, z_dim = 48, 48, 13 + self.inputs = { + 'X': np.random.random( + (batch_size, x_dim, y_dim, z_dim)).astype("float32"), + } + self.attrs = { + 'axis': 3, + 'indexes': [0, 2, 7, 9], + } + + output = None + for batch in range(batch_size): + res = np.zeros((x_dim, y_dim, 1)).astype("float32").reshape(-1) + for index in self.attrs['indexes']: + channel = self.inputs['X'][batch, :, :, index].reshape(-1).copy( + ) + tag1 = [0 for i in range(x_dim)] + tag2 = [0 for i in range(y_dim)] + cnt = 0 + for i in range(channel.size): + index = channel.argmax() + idx1 = index // y_dim + idx2 = index % y_dim + if tag1[idx1] + tag2[idx2] == 0: + tag1[idx1] = 1 + tag2[idx2] = 1 + res[index] = 1 + cnt += 1 + if cnt == min(x_dim, y_dim): + break + channel[index] = -1 + res = res.reshape(x_dim, y_dim, 1) + res = res.repeat([z_dim], axis=2) + res = res.reshape(1, x_dim, y_dim, z_dim) + if output is not None: + output = np.concatenate((output, res), axis=0) + else: + output = res + self.outputs = {'Out': output} + + def test_check_output(self): + self.check_output() + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_slice_var.py b/python/paddle/fluid/tests/unittests/test_slice_var.py index fab63b7d5631829feffd26fc1dce2bd338d2036b..b16c744603534ad07f9318a5a195f1d7550b1f10 100644 --- a/python/paddle/fluid/tests/unittests/test_slice_var.py +++ b/python/paddle/fluid/tests/unittests/test_slice_var.py @@ -30,7 +30,6 @@ class TestSliceVar(unittest.TestCase): var = program.global_block().create_var( name=str(random.randint(10000, 99999)), persistable=True, - # dtype=core.VarDesc.VarType.LOD_TENSOR, shape=shape) var_list.append(var) blocks = slice_variable(var_list, 10, min_size) diff --git a/python/paddle/fluid/tests/unittests/test_softmax_op.py b/python/paddle/fluid/tests/unittests/test_softmax_op.py index d88aa1ae1c9d848eba7a2224d22b5201fc27b857..40c3135183a128cd9b7324ce27da798fa2d93afd 100644 --- a/python/paddle/fluid/tests/unittests/test_softmax_op.py +++ b/python/paddle/fluid/tests/unittests/test_softmax_op.py @@ -62,12 +62,11 @@ class TestSoftmaxOp(OpTest): self.check_output() def test_check_grad(self): - if self.dtype == np.float16: - return - if self.use_cudnn: + if self.use_cudnn or self.dtype == np.float16: place = core.CUDAPlace(0) - self.check_grad_with_place( - place, ["X"], "Out", max_relative_error=0.01) + if core.is_float16_supported(place): + self.check_grad_with_place( + place, ["X"], "Out", max_relative_error=0.01) else: self.check_grad(["X"], "Out", max_relative_error=0.01) @@ -103,10 +102,23 @@ class TestSoftmaxFP16Op(TestSoftmaxOp): if core.is_float16_supported(place): self.check_output_with_place(place, atol=1e-3) + # FIXME: If the x_shape is [10, 10], gradient failed. + def test_check_grad(self): + pass + @unittest.skipIf(not core.is_compiled_with_cuda(), "core is not compiled with CUDA") -class TestSoftmaxFP16Op2(TestSoftmaxFP16Op): +class TestSoftmaxFP16Op2(TestSoftmaxOp): + def init_kernel_type(self): + self.dtype = np.float16 + + def test_check_output(self): + if core.is_compiled_with_cuda(): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=1e-3) + def get_x_shape(self): return [2, 3, 4, 5] diff --git a/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py b/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py index a18941dd3126ac027f022ddafbbaed8516166233..37ee880970cf7f6f235e7c43697b2b7872bed38b 100644 --- a/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py +++ b/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py @@ -26,7 +26,11 @@ class TestSoftmaxWithCrossEntropyOp(OpTest): Test softmax with cross entropy operator with discreate one-hot labels. """ + def initParams(self): + self.numeric_stable_mode = False + def setUp(self): + self.initParams() self.op_type = "softmax_with_cross_entropy" batch_size = 41 class_num = 37 @@ -46,6 +50,7 @@ class TestSoftmaxWithCrossEntropyOp(OpTest): "Softmax": softmax.astype("float64"), "Loss": cross_entropy.astype("float64") } + self.attrs = {"numeric_stable_mode": self.numeric_stable_mode} def test_check_output(self): self.check_output() @@ -54,6 +59,11 @@ class TestSoftmaxWithCrossEntropyOp(OpTest): self.check_grad(["Logits"], "Loss") +class TestSoftmaxWithCrossEntropyOpNoCudnn(TestSoftmaxWithCrossEntropyOp): + def initParams(self): + self.numeric_stable_mode = True + + class TestSoftmaxWithCrossEntropyOp2(OpTest): """ Test softmax with cross entropy operator with soft labels. @@ -93,7 +103,11 @@ class TestSoftmaxWithCrossEntropyOp3(OpTest): Test softmax with cross entropy operator with ignore_index. """ + def initParams(self): + self.numeric_stable_mode = False + def setUp(self): + self.initParams() self.op_type = "softmax_with_cross_entropy" batch_size = 41 class_num = 37 @@ -114,7 +128,10 @@ class TestSoftmaxWithCrossEntropyOp3(OpTest): "Softmax": softmax.astype("float64"), "Loss": cross_entropy.astype("float64") } - self.attrs = {"ignore_index": ignore_index} + self.attrs = { + "ignore_index": ignore_index, + "numeric_stable_mode": self.numeric_stable_mode + } def test_check_output(self): self.check_output() @@ -123,5 +140,10 @@ class TestSoftmaxWithCrossEntropyOp3(OpTest): self.check_grad(["Logits"], "Loss") +class TestSoftmaxWithCrossEntropyOp3NoCudnn(TestSoftmaxWithCrossEntropyOp3): + def initParams(self): + self.numeric_stable_mode = True + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_space_to_depth_op.py b/python/paddle/fluid/tests/unittests/test_space_to_depth_op.py new file mode 100644 index 0000000000000000000000000000000000000000..5fdad44f1242b9ee99040b43d7ce2cf84664eed1 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_space_to_depth_op.py @@ -0,0 +1,135 @@ +# 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 paddle.fluid as fluid +from op_test import OpTest + + +class TestSpaceToDepthOp(OpTest): + @staticmethod + def helper(in_, width, height, channel, batch, blocksize, forward, out_): + channel_out = channel // (blocksize * blocksize) + for b in range(batch): + for k in range(channel): + for j in range(height): + for i in range(width): + in_index = i + width * (j + height * (k + channel * b)) + channel2 = k % channel_out + offset = k // channel_out + width2 = i * blocksize + offset % blocksize + height2 = j * blocksize + offset // blocksize + out_index = width2 + width * blocksize * ( + height2 + height * blocksize * + (channel2 + channel_out * b)) + if forward: + out_[out_index] = in_[in_index] + else: + out_[in_index] = in_[out_index] + + def setUp(self): + self.init_data() + + self.op_type = "space_to_depth" + self.inputs = {"X": self.x} + self.helper(self.x_1d, self.x.shape[3], self.x.shape[2], + self.x.shape[1], self.x.shape[0], self.blocksize, + self.forward, self.out_1d) + self.out = np.reshape(self.out_1d, self.infered_shape) + self.attrs = {"blocksize": self.blocksize} + self.outputs = {"Out": self.out} + + def init_data(self): + self.ori_shape = (32, 12, 6, 6) + self.infered_shape = (32, 48, 3, 3) + self.one_d_len = 32 * 48 * 3 * 3 + + self.blocksize = 2 + self.x = np.random.random(self.ori_shape).astype('float32') + self.x_1d = np.reshape(self.x, self.one_d_len) + self.out = np.zeros(self.infered_shape).astype('float32') + self.out_1d = np.reshape(self.out, self.one_d_len) + self.forward = 1 + + def test_check_output(self): + place = fluid.core.CUDAPlace(0) if fluid.core.is_compiled_with_cuda( + ) else fluid.core.CPUPlace() + self.check_output_with_place(place, 1e-5, None, False) + + def test_check_grad(self): + place = fluid.core.CUDAPlace(0) if fluid.core.is_compiled_with_cuda( + ) else fluid.core.CPUPlace() + self.check_grad_with_place(place, ['X'], 'Out') + + +class TestSpaceToDepthOpBasic(TestSpaceToDepthOp): + def init_data(self): + self.ori_shape = (32, 8, 6, 6) + self.infered_shape = (32, 32, 3, 3) + self.one_d_len = 32 * 32 * 3 * 3 + + self.blocksize = 2 + self.x = np.random.random(self.ori_shape).astype('float32') + self.x_1d = np.reshape(self.x, self.one_d_len) + self.out = np.zeros(self.infered_shape).astype('float32') + self.out_1d = np.reshape(self.out, self.one_d_len) + self.forward = 1 + + +class TestSpaceToDepthOpDoubleBasic(TestSpaceToDepthOp): + def init_data(self): + self.ori_shape = (32, 8, 6, 6) + self.infered_shape = (32, 32, 3, 3) + self.one_d_len = 32 * 32 * 3 * 3 + + self.blocksize = 2 + self.x = np.random.random(self.ori_shape).astype('float64') + self.x_1d = np.reshape(self.x, self.one_d_len) + self.out = np.zeros(self.infered_shape).astype('float64') + self.out_1d = np.reshape(self.out, self.one_d_len) + self.forward = 1 + + +class TestSpaceToDepthOpWithStride3(TestSpaceToDepthOp): + def init_data(self): + self.ori_shape = (32, 9, 6, 6) + self.infered_shape = (32, 81, 2, 2) + self.one_d_len = 32 * 81 * 2 * 2 + + self.blocksize = 3 + self.x = np.random.random(self.ori_shape).astype('float32') + self.x_1d = np.reshape(self.x, self.one_d_len) + self.out = np.zeros(self.infered_shape).astype('float32') + self.out_1d = np.reshape(self.out, self.one_d_len) + self.forward = 1 + + +class TestSpaceToDepthOpWithNotSquare(TestSpaceToDepthOp): + def init_data(self): + self.ori_shape = (32, 9, 9, 6) + self.infered_shape = (32, 81, 3, 2) + self.one_d_len = 32 * 81 * 3 * 2 + + self.blocksize = 3 + self.x = np.random.random(self.ori_shape).astype('float32') + self.x_1d = np.reshape(self.x, self.one_d_len) + self.out = np.zeros(self.infered_shape).astype('float32') + self.out_1d = np.reshape(self.out, self.one_d_len) + self.forward = 1 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_split_ids_op.py b/python/paddle/fluid/tests/unittests/test_split_ids_op.py index 4c3d0258980fd8595704a65219deb520b96e222e..d674dad2293921c06135b4ee528538d266cb2904 100644 --- a/python/paddle/fluid/tests/unittests/test_split_ids_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_ids_op.py @@ -25,18 +25,21 @@ from paddle.fluid.op import Operator class TestSplitIdsOp(OpTest): def setUp(self): self.op_type = "split_ids" - ids = np.array([[0], [2], [2], [3], [5], [5], [6]]).astype('int64') + ids1 = np.array([[0], [2], [2], [3], [5], [5], [6]]).astype('int64') + ids2 = np.array([[6], [2], [3], [3], [5], [2], [6]]).astype('int64') + ids3 = np.array([[2], [2], [2], [3], [5], [5], [6]]).astype('int64') + out0 = np.array([[0], [3], [6]]).astype('int64') out1 = np.array([[]]).astype('int64') - out2 = np.array([[2], [2], [5], [5]]).astype('int64') - self.inputs = {'Ids': ids} + out2 = np.array([[2], [5]]).astype('int64') + self.inputs = {'Ids': [('ids1', ids1), ('ids2', ids2), ('ids3', ids3)]} self.outputs = {'Out': [('out0', out0), ('out1', out1), ('out2', out2)]} def test_check_output(self): self.check_output() -class TestSpliteIds(unittest.TestCase): +class TestSplitSelectedRows(unittest.TestCase): def get_places(self): places = [core.CPUPlace()] return places diff --git a/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py b/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py index 41a5ee59ea523b1f6c5015974a12c526e883fa35..50204b8a77c187aa695da83860960566448d290f 100644 --- a/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py @@ -99,7 +99,6 @@ class TestSpliteSelectedRows(unittest.TestCase): out0_grad.set_height(height) out0_grad_tensor = out0_grad.get_tensor() np_array = np.ones((len(rows0), row_numel)).astype("float32") - np_array[0, 0] = 2.0 out0_grad_tensor.set(np_array, place) out1_grad = scope.var("out1@GRAD").get_selected_rows() @@ -108,7 +107,6 @@ class TestSpliteSelectedRows(unittest.TestCase): out1_grad.set_height(height) out1_grad_tensor = out1_grad.get_tensor() np_array = np.ones((len(rows1), row_numel)).astype("float32") - np_array[0, 1] = 4.0 out1_grad_tensor.set(np_array, place) x_grad = scope.var("X@GRAD").get_selected_rows() @@ -121,11 +119,13 @@ class TestSpliteSelectedRows(unittest.TestCase): grad_op.run(scope, place) - self.assertEqual(x_grad.rows(), rows0 + rows1) + merged_rows = set(rows0 + rows1) + self.assertEqual(set(x_grad.rows()), set(rows0 + rows1)) self.assertEqual(x_grad.height(), height) + print(np.array(x_grad.get_tensor())) self.assertAlmostEqual(2.0, np.array(x_grad.get_tensor())[0, 0]) - self.assertAlmostEqual(4.0, np.array(x_grad.get_tensor())[2, 1]) + self.assertAlmostEqual(1.0, np.array(x_grad.get_tensor())[2, 1]) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_sum_op.py b/python/paddle/fluid/tests/unittests/test_sum_op.py index 74797bb65678404b7b35d06eecc7f9a12b2a346e..0be5be6e97d26c6ec42471d078e8e5995727e594 100644 --- a/python/paddle/fluid/tests/unittests/test_sum_op.py +++ b/python/paddle/fluid/tests/unittests/test_sum_op.py @@ -24,16 +24,20 @@ from paddle.fluid.op import Operator class TestSumOp(OpTest): def setUp(self): self.op_type = "sum" + self.init_kernel_type() self.use_mkldnn = False self.init_kernel_type() - x0 = np.random.random((3, 4)).astype('float32') - x1 = np.random.random((3, 4)).astype('float32') - x2 = np.random.random((3, 4)).astype('float32') + x0 = np.random.random((3, 4)).astype(self.dtype) + x1 = np.random.random((3, 4)).astype(self.dtype) + x2 = np.random.random((3, 4)).astype(self.dtype) self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]} y = x0 + x1 + x2 self.outputs = {'Out': y} self.attrs = {'use_mkldnn': self.use_mkldnn} + def init_kernel_type(self): + self.dtype = np.float32 + def test_check_output(self): self.check_output() @@ -45,16 +49,36 @@ class TestSumOp(OpTest): class TestSelectedRowsSumOp(OpTest): - def check_with_place(self, place): - scope = core.Scope() - self.check_input_and_optput(scope, place, True, True, True) - self.check_input_and_optput(scope, place, False, True, True) - self.check_input_and_optput(scope, place, False, False, True) - self.check_input_and_optput(scope, place, False, False, False) + def setUp(self): + self.height = 10 + self.row_numel = 12 + self.rows = [0, 1, 2, 3, 4, 5, 6] + self.dtype = np.float32 + self.init_kernel_type() + + def check_with_place(self, place, inplace): + self.check_input_and_optput(core.Scope(), place, inplace, True, True, + True) + self.check_input_and_optput(core.Scope(), place, inplace, False, True, + True) + self.check_input_and_optput(core.Scope(), place, inplace, False, False, + True) + self.check_input_and_optput(core.Scope(), place, inplace, False, False, + False) + + def init_kernel_type(self): + pass + + def _get_array(self, rows, row_numel): + array = np.ones((len(rows), row_numel)).astype(self.dtype) + for i in range(len(rows)): + array[i] *= rows[i] + return array def check_input_and_optput(self, scope, place, + inplace, w1_has_data=False, w2_has_data=False, w3_has_data=False): @@ -64,35 +88,43 @@ class TestSelectedRowsSumOp(OpTest): self.create_selected_rows(scope, place, "W3", w3_has_data) # create Out Variable - out = scope.var('Out').get_selected_rows() + if inplace: + out_var_name = "W1" + else: + out_var_name = "Out" + out = scope.var(out_var_name).get_selected_rows() # create and run sum operator - sum_op = Operator("sum", X=["W1", "W2", "W3"], Out='Out') + sum_op = Operator("sum", X=["W1", "W2", "W3"], Out=out_var_name) sum_op.run(scope, place) has_data_w_num = 0 - for w in [w1_has_data, w2_has_data, w3_has_data]: - if not w: + for has_data in [w1_has_data, w2_has_data, w3_has_data]: + if has_data: has_data_w_num += 1 - self.assertEqual(7 * has_data_w_num, len(out.rows())) + if has_data_w_num > 0: + self.assertEqual(len(out.rows()), 7) + self.assertTrue( + np.array_equal( + np.array(out.get_tensor()), + self._get_array(self.rows, self.row_numel) * + has_data_w_num)) + else: + self.assertEqual(len(out.rows()), 0) - def create_selected_rows(self, scope, place, var_name, isEmpty): + def create_selected_rows(self, scope, place, var_name, has_data): # create and initialize W Variable - if not isEmpty: - rows = [0, 1, 2, 3, 4, 5, 6] - row_numel = 12 + if has_data: + rows = self.rows else: rows = [] - row_numel = 12 var = scope.var(var_name) w_selected_rows = var.get_selected_rows() - w_selected_rows.set_height(len(rows)) + w_selected_rows.set_height(self.height) w_selected_rows.set_rows(rows) - w_array = np.ones((len(rows), row_numel)).astype("float32") - for i in range(len(rows)): - w_array[i] *= i + w_array = self._get_array(self.rows, self.row_numel) w_tensor = w_selected_rows.get_tensor() w_tensor.set(w_array, place) @@ -100,10 +132,98 @@ class TestSelectedRowsSumOp(OpTest): def test_w_is_selected_rows(self): places = [core.CPUPlace()] - # currently only support CPU + if core.is_compiled_with_cuda(): + places.append(core.CUDAPlace(0)) for place in places: - self.check_with_place(place) + for inplace in [True, False]: + self.check_with_place(place, inplace) + + +class TestLoDTensorAndSelectedRowsOp(TestSelectedRowsSumOp): + def setUp(self): + self.height = 10 + self.row_numel = 12 + self.rows = [0, 1, 2, 2, 4, 5, 6] + + def check_with_place(self, place, inplace): + scope = core.Scope() + if inplace: + self.create_lod_tensor(scope, place, "x1") + self.create_selected_rows(scope, place, "x2", True) + out = scope.var("x1").get_tensor() + out_name = "x1" + else: + self.create_selected_rows(scope, place, "x1", True) + self.create_lod_tensor(scope, place, "x2") + out = scope.var("out").get_tensor() + out_name = "out" + + # create and run sum operator + sum_op = Operator("sum", X=["x1", "x2"], Out=out_name) + sum_op.run(scope, place) + + result = np.ones((1, self.height)).astype(np.int32).tolist()[0] + for ele in self.rows: + result[ele] += 1 + + out_t = np.array(out) + self.assertEqual(out_t.shape[0], self.height) + self.assertTrue( + np.array_equal(out_t, + self._get_array([i for i in range( + self.height)], self.row_numel) * np.tile( + np.array(result).reshape(self.height, 1), + self.row_numel))) + + def create_lod_tensor(self, scope, place, var_name): + var = scope.var(var_name) + w_tensor = var.get_tensor() + w_array = self._get_array([i for i in range(self.height)], + self.row_numel) + w_tensor.set(w_array, place) + return var + + +#----------- test fp16 ----------- +@unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") +class TestFP16SumOp(TestSumOp): + def init_kernel_type(self): + self.dtype = np.float16 + + def test_check_output(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_output_with_place(place, atol=2e-2) + + # FIXME: Because of the precision fp16, max_relative_error + # should be 0.15 here. + def test_check_grad(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + self.check_grad(['x0'], 'Out', max_relative_error=0.15) + + +def create_test_sum_fp16_class(parent): + @unittest.skipIf(not core.is_compiled_with_cuda(), + "core is not compiled with CUDA") + class TestSumFp16Case(parent): + def init_kernel_type(self): + self.dtype = np.float16 + + def test_w_is_selected_rows(self): + place = core.CUDAPlace(0) + if core.is_float16_supported(place): + for inplace in [True, False]: + self.check_with_place(place, inplace) + + cls_name = "{0}_{1}".format(parent.__name__, "SumFp16Test") + TestSumFp16Case.__name__ = cls_name + globals()[cls_name] = TestSumFp16Case + +create_test_sum_fp16_class(TestSelectedRowsSumOp) +create_test_sum_fp16_class(TestLoDTensorAndSelectedRowsOp) if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_tensor_array_to_tensor.py b/python/paddle/fluid/tests/unittests/test_tensor_array_to_tensor.py new file mode 100644 index 0000000000000000000000000000000000000000..78b95de7e07b1d1fcdeeae63498e740c2b474c6d --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_tensor_array_to_tensor.py @@ -0,0 +1,142 @@ +# 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 +import paddle.fluid as fluid +import paddle.fluid.core as core +from paddle.fluid.op import Operator +from paddle.fluid.executor import Executor + + +class TestLoDTensorArrayConcat(unittest.TestCase): + def setUp(self): + self.op_type = "tensor_array_to_tensor" + self.attrs = {"axis": 0} + self.outputs = ["Out"] + + def test_get_set(self): + scope = core.Scope() + program = fluid.Program() + block = program.global_block() + + input_arr = block.create_var( + name="tmp_lod_tensor_array", + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY) + input_arr.persistable = True + input_arr_var = scope.var('tmp_lod_tensor_array') + input_tensor_array = input_arr_var.get_lod_tensor_array() + self.assertEqual(0, len(input_tensor_array)) + + cpu = core.CPUPlace() + for i in range(10): + t = core.LoDTensor() + if i == 0: + t.set(numpy.array([[i], [i]], dtype='float32'), cpu) + else: + t.set(numpy.array([[i]], dtype='float32'), cpu) + input_tensor_array.append(t) + + self.assertEqual(10, len(input_tensor_array)) + + random_grad = numpy.random.random_sample([11]).astype(numpy.float32) + + y_out = block.create_var(name="Out") + y_out.persistable = True + y_out_index = block.create_var(name="OutIndex") + y_out_index.persistable = True + + y_grad_arr = block.create_var( + name='Out@GRAD', dtype='float32', shape=[11]) + y_grad_arr.persistable = True + y_grad = scope.var('Out@GRAD') + y_grad_tensor = y_grad.get_tensor() + y_grad_tensor.set(random_grad, cpu) + + op = block.append_op( + type=self.op_type, + inputs={"X": input_arr}, + outputs={"Out": y_out, + "OutIndex": y_out_index}, + attrs=self.attrs) + + out_grad = block.create_var( + name="tmp_lod_tensor_array@GRAD", + type=core.VarDesc.VarType.LOD_TENSOR_ARRAY) + out_grad.persistable = True + + grad_op_desc_list, op_grad_to_var = core.get_grad_op_desc(op.desc, + set(), []) + grad_op_desc = grad_op_desc_list[0] + new_op_desc = block.desc.append_op() + new_op_desc.copy_from(grad_op_desc) + for var_name in grad_op_desc.output_arg_names(): + block.desc.var(var_name.encode("ascii")) + + grad_op_desc.infer_var_type(block.desc) + grad_op_desc.infer_shape(block.desc) + for arg in grad_op_desc.output_arg_names(): + grad_var = block.desc.find_var(arg.encode("ascii")) + grad_var.set_dtype(core.VarDesc.VarType.FP32) + + fetch_list = [] + fetch_list.append(block.var('Out')) + fetch_list.append(block.var('OutIndex')) + + exe = fluid.Executor(fluid.CPUPlace()) + out = exe.run(program, fetch_list=fetch_list, scope=scope) + #print ("index: ", numpy.array(out[1])) + + # test forward + tensor_res = numpy.array(out[0]) + tensor_res_out_idx = numpy.array(out[1]) + tensor_gt = numpy.array( + [0] + [0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype='float32') + + self.assertEqual(len(tensor_res), len(tensor_gt)) + self.assertEqual(len(tensor_res_out_idx), 10) + + for i in range(len(tensor_res)): + self.assertEqual(tensor_res[i], tensor_gt[i]) + + for i in range(len(tensor_res_out_idx)): + if i == 0: + self.assertEqual(tensor_res_out_idx[i], 2) + else: + self.assertEqual(tensor_res_out_idx[i], 1) + + # test backward + grad_tensor = scope.var('tmp_lod_tensor_array@GRAD') + grad_tensor_array = grad_tensor.get_lod_tensor_array() + + self.assertEqual(10, len(grad_tensor_array)) + + for i in range(len(grad_tensor_array)): + if i == 0: + self.assertEqual( + numpy.array(grad_tensor_array[i])[0], + numpy.array(random_grad[i])) + self.assertEqual( + numpy.array(grad_tensor_array[i])[1], + numpy.array(random_grad[i + 1])) + if i == 1: + self.assertEqual( + numpy.array(grad_tensor_array[i]), + numpy.array(random_grad[i + 1])) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_top_k_op.py b/python/paddle/fluid/tests/unittests/test_top_k_op.py index e54e170f7f1e03db4b63db72edb7395d18130f68..69b29db83a43d18c0825b610642009a0377b9901 100644 --- a/python/paddle/fluid/tests/unittests/test_top_k_op.py +++ b/python/paddle/fluid/tests/unittests/test_top_k_op.py @@ -21,22 +21,27 @@ from op_test import OpTest class TestTopkOp(OpTest): def setUp(self): + self.set_args() self.op_type = "top_k" - k = 1 - input = np.random.random((32, 84)).astype("float32") - output = np.ndarray((32, k)) - indices = np.ndarray((32, k)).astype("int64") + k = self.top_k + input = np.random.random((self.row, k)).astype("float32") + output = np.ndarray((self.row, k)) + indices = np.ndarray((self.row, k)).astype("int64") self.inputs = {'X': input} self.attrs = {'k': k} - for rowid in range(32): + for rowid in range(self.row): row = input[rowid] - output[rowid] = np.sort(row)[-k:] - indices[rowid] = row.argsort()[-k:] + output[rowid] = np.sort(row)[::-1][:k] + indices[rowid] = row.argsort()[::-1][:k] self.outputs = {'Out': output, 'Indices': indices} + def set_args(self): + self.row = 32 + self.top_k = 1 + def test_check_output(self): self.check_output() @@ -50,14 +55,39 @@ class TestTopkOp3d(OpTest): output = np.ndarray((64, k)) indices = np.ndarray((64, k)).astype("int64") - # FIXME: should use 'X': input for a 3d input - self.inputs = {'X': input_flat_2d} + self.inputs = {'X': input} self.attrs = {'k': k} for rowid in range(64): row = input_flat_2d[rowid] - output[rowid] = np.sort(row)[-k:] - indices[rowid] = row.argsort()[-k:] + output[rowid] = np.sort(row)[::-1][:k] + indices[rowid] = row.argsort()[::-1][:k] + + self.outputs = { + 'Out': output.reshape((32, 2, k)), + 'Indices': indices.reshape((32, 2, k)) + } + + def test_check_output(self): + self.check_output() + + +class TestTopkOp2(OpTest): + def setUp(self): + self.op_type = "top_k" + k = 1 + m = 2056 + input = np.random.random((m, 84)).astype("float32") + output = np.ndarray((m, k)) + indices = np.ndarray((m, k)).astype("int64") + + self.inputs = {'X': input} + self.attrs = {'k': k} + + for rowid in range(m): + row = input[rowid] + output[rowid] = -np.sort(-row)[:k] + indices[rowid] = (-row).argsort()[:k] self.outputs = {'Out': output, 'Indices': indices} @@ -65,5 +95,17 @@ class TestTopkOp3d(OpTest): self.check_output() +class TestTopkOp3(TestTopkOp): + def set_args(self): + self.row = 2056 + self.top_k = 3 + + +class TestTopkOp4(TestTopkOp): + def set_args(self): + self.row = 40000 + self.top_k = 1 + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index ecdbe27f4d90268d755a712e25289cfaf4715f29..89bc24802751340b6d4657be8673d714f3d3dc2b 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -31,17 +31,17 @@ Steps to transpile pserver: """ import math -import sys import numpy as np import collections -import six +import logging -from .ps_dispatcher import RoundRobin, HashName, PSDispatcher -from .. import core, framework +from .ps_dispatcher import RoundRobin, PSDispatcher +from .. import core, framework, unique_name from ..framework import Program, default_main_program, \ default_startup_program, Block, \ Parameter, grad_var_name from .details import * +from ..distribute_lookup_table import find_distributed_lookup_table from functools import reduce LOOKUP_TABLE_TYPE = "lookup_table" @@ -49,6 +49,7 @@ LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad" OP_ROLE_VAR_ATTR_NAME = core.op_proto_and_checker_maker.kOpRoleVarAttrName() RPC_OP_ROLE_ATTR_NAME = op_role_attr_name = core.op_proto_and_checker_maker.kOpRoleAttrName( ) +OPT_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Optimize RPC_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.RPC DIST_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.Dist LR_SCHED_OP_ROLE_ATTR_VALUE = core.op_proto_and_checker_maker.OpRole.LRSched @@ -136,6 +137,7 @@ class DistributeTranspilerConfig(object): slice_var_up = True split_method = None min_block_size = 8192 + enable_dc_asgd = False # supported modes: pserver, nccl2 mode = "pserver" print_log = False @@ -250,6 +252,8 @@ class DistributeTranspiler(object): n workers, the id may range from 0 ~ n-1 program (Program|None): program to transpile, default is fluid.default_main_program(). + startup_program (Program|None): startup_program to transpile, + default is fluid.default_startup_program(). pservers (str): comma separated ip:port string for the pserver list. trainers (int|str): in pserver mode this is the number of @@ -287,7 +291,8 @@ class DistributeTranspiler(object): self.optimize_ops, self.params_grads = self._get_optimize_pass() ps_dispatcher = self.config.split_method(self.pserver_endpoints) - self.has_distributed_lookup_table = self._has_distributed_lookup_table() + self.table_name = find_distributed_lookup_table(self.origin_program) + self.has_distributed_lookup_table = self.table_name != None self.param_name_to_grad_name = dict() self.grad_name_to_param_name = dict() for param_var, grad_var in self.params_grads: @@ -381,6 +386,8 @@ class DistributeTranspiler(object): outputs={"Out": send_barrier_out}, attrs={ "endpoints": pserver_endpoints, + "sync_mode": self.sync_mode, + "trainer_id": self.trainer_id, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) @@ -424,6 +431,7 @@ class DistributeTranspiler(object): outputs={"Out": splited_var}, attrs={ "epmap": eps, + "trainer_id": self.trainer_id, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE, OP_ROLE_VAR_ATTR_NAME: [param_varname, recv_op_role_var_name], @@ -438,6 +446,7 @@ class DistributeTranspiler(object): outputs={"Out": all_recv_outputs}, attrs={ "endpoints": pserver_endpoints, + "trainer_id": self.trainer_id, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) @@ -474,6 +483,26 @@ class DistributeTranspiler(object): delete_ops(self.origin_program.global_block(), self.optimize_ops) delete_ops(self.origin_program.global_block(), lr_ops) + # delete table init op + if self.has_distributed_lookup_table: + table_var = self.startup_program.global_block().vars[ + self.table_name] + table_param_init_op = [] + for op in self.startup_program.global_block().ops: + if self.table_name in op.output_arg_names: + table_param_init_op.append(op) + init_op_num = len(table_param_init_op) + if init_op_num != 1: + raise ValueError("table init op num should be 1, now is " + str( + init_op_num)) + table_init_op = table_param_init_op[0] + self.startup_program.global_block().append_op( + type="fake_init", + inputs={}, + outputs={"Out": table_var}, + attrs={"shape": table_init_op.attr('shape')}) + delete_ops(self.startup_program.global_block(), table_param_init_op) + self.origin_program.__str__() if wait_port: @@ -629,6 +658,24 @@ in a single call.") endpoint, op): opt_op_on_pserver.append(op) # step 3.3 + # prepare if dc asgd is enabled + if self.config.enable_dc_asgd == True: + assert (self.sync_mode == False) + self.param_bak_list = [] + # add param_bak for each trainer + for p in self.param_grad_ep_mapping[endpoint]["params"]: + # each parameter should have w_bak for each trainer id + for i in range(self.trainer_num): + param_bak_name = "%s.trainer_%d_bak" % (p.name, i) + tmpvar = pserver_program.global_block().create_var( + # NOTE: this var name format is used in `request_get_handler` + name=param_bak_name, + type=p.type, + shape=p.shape, + dtype=p.dtype) + self.param_bak_list.append((p, tmpvar)) + + # step 3.4 # Iterate through the ops, and if an op and the optimize ops # which located on current pserver are in one set, then # append it into the sub program. @@ -712,14 +759,14 @@ in a single call.") for _, op in enumerate(self.optimize_ops): # optimizer is connected to itself if op.attr(OP_ROLE_VAR_ATTR_NAME)[0] == optimize_target_param_name and \ - op not in global_ops: + op not in global_ops: log("append opt op: ", op.type, op.input_arg_names, merged_var) __append_optimize_op__(op, per_opt_block, grad_to_block_id, merged_var, lr_ops) - # dedup grad to ids list +# dedup grad to ids list grad_to_block_id = list(set(grad_to_block_id)) # append global ops if global_ops: @@ -746,6 +793,15 @@ in a single call.") prefetch_var_name_to_block_id.extend( lookup_table_var_name_to_block_id) + if len(optimize_blocks) == 0: + logging.warn("pserver [" + str(endpoint) + + "] has no optimize block!!") + pre_block_idx = pserver_program.num_blocks - 1 + empty_block = pserver_program._create_block(pre_block_idx) + optimize_blocks.append(empty_block) + + # In some case, some parameter server will have no parameter to optimize + # So we give an empty optimize block to parameter server. attrs = { "optimize_blocks": optimize_blocks, "endpoint": endpoint, @@ -756,6 +812,8 @@ in a single call.") if self.has_distributed_lookup_table: attrs['checkpint_block_id'] = checkpoint_block_id + if self.config.enable_dc_asgd: + attrs['dc_asgd'] = True if len(prefetch_var_name_to_block_id) > 0: attrs[ @@ -788,7 +846,8 @@ in a single call.") tuple: (main_program, startup_program), of type "Program" """ pserver_prog = self.get_pserver_program(endpoint) - pserver_startup = self.get_startup_program(endpoint) + pserver_startup = self.get_startup_program( + endpoint, pserver_program=pserver_prog) return pserver_prog, pserver_startup def get_startup_program(self, @@ -871,6 +930,15 @@ to transpile() call.") inputs=new_inputs, outputs=new_outputs, attrs=op.all_attrs()) + if self.config.enable_dc_asgd: + for p, p_bak in self.param_bak_list: + startup_param_var = s_prog.global_block().vars[p.name] + startup_tmpvar = s_prog.global_block().vars[p_bak.name] + # copy init random value to param_bak + s_prog.global_block().append_op( + type="assign", + inputs={"X": startup_param_var}, + outputs={"Out": startup_tmpvar}) # add slice vars s_prog._slice_vars_and_attrs = self._get_slice_vars_and_attrs(endpoint) @@ -888,38 +956,16 @@ to transpile() call.") block_idx = int(block_name.split(block_suffix)[1]) orig_var = self.origin_program.global_block().vars[orig_var_name] - skip_numel = 0 + skip_dim0 = 0 slice_vars = self.param_var_mapping[orig_var_name] for slice_var in slice_vars[:block_idx]: - skip_numel += reduce(lambda x, y: x * y, slice_var.shape) - slice_vars_and_attrs.append([orig_var, skip_numel, param]) + skip_dim0 += slice_var.shape[0] + slice_vars_and_attrs.append([orig_var, skip_dim0, param]) return slice_vars_and_attrs # ====================== private transpiler functions ===================== - def _has_distributed_lookup_table(self): - # process lookup_table_op - # 1. check all lookup_table_op is distributed - # 2. check all lookup_table_op share the same table. - distributed_lookup_table_ops = [] - # support only one distributed_lookup_table now - self.table_name = None - for op in self.origin_program.global_block().ops: - if op.type == LOOKUP_TABLE_TYPE: - if op.attr('is_distributed') is True: - if self.table_name is None: - self.table_name = op.input("W")[0] - if self.table_name != op.input("W")[0]: - raise RuntimeError("all distributed lookup_table_ops" - " should have only one table") - distributed_lookup_table_ops.append(op) - else: - if self.table_name is not None: - assert op.input("W")[0] != self.table_name - - return len(distributed_lookup_table_ops) > 0 - def _update_dist_lookup_table_vars(self, param_list, grad_list, params_grads): # TODO(wuyi): put find a way to put dist lookup table stuff all together. @@ -1032,92 +1078,90 @@ to transpile() call.") def _replace_lookup_table_op_with_prefetch(self, program, pserver_endpoints): # 1. replace lookup_table_op with split_ids_op -> prefetch_op -> sum_op - # self.all_prefetch_input_vars = - # [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1] - # [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]] + self.all_in_ids_vars = [] self.all_prefetch_input_vars = [] - - # self.all_prefetch_input_vars = - # [[var0_prefetch_in_pserver0, var0_prefetch_in_pserver1] - # [var1_prefetch_in_pserver0, var1_prefetch_in_pserver1]] self.all_prefetch_output_vars = [] + self.all_out_emb_vars = [] + lookup_table_op_index = -1 continue_search_lookup_table_op = True while continue_search_lookup_table_op: continue_search_lookup_table_op = False all_ops = program.global_block().ops for op in all_ops: - if op.type == LOOKUP_TABLE_TYPE: + if op.type == LOOKUP_TABLE_TYPE and self.table_name == op.input( + "W")[0]: + if not op.attr('is_distributed'): + raise RuntimeError( + "lookup_table_op that lookup an distributed embedding table" + "should set is_distributed to true") continue_search_lookup_table_op = True - lookup_table_op_index = list(all_ops).index(op) + lookup_table_op_index = lookup_table_op_index if lookup_table_op_index != -1 else list( + all_ops).index(op) ids_name = op.input("Ids") out_name = op.output("Out") ids_var = program.global_block().vars[ids_name[0]] - prefetch_input_vars = self._create_splited_vars( - source_var=ids_var, - block=program.global_block(), - tag="_prefetch_in_") - self.all_prefetch_input_vars.append(prefetch_input_vars) + self.all_in_ids_vars.append(ids_var) out_var = program.global_block().vars[out_name[0]] - prefetch_output_vars = self._create_splited_vars( - source_var=out_var, - block=program.global_block(), - tag="_prefetch_out_") - self.all_prefetch_output_vars.append(prefetch_output_vars) - - # insert split_ids_op - program.global_block()._insert_op( - index=lookup_table_op_index, - type="split_ids", - inputs={ - 'Ids': [ - program.global_block().vars[varname] - for varname in ids_name - ] - }, - outputs={"Out": prefetch_input_vars}) - - # insert prefetch_op - program.global_block()._insert_op( - index=lookup_table_op_index + 1, - type="prefetch", - inputs={'X': prefetch_input_vars}, - outputs={"Out": prefetch_output_vars}, - attrs={ - "epmap": pserver_endpoints, - # FIXME(qiao) temporarily disable this config because prefetch - # is not act as other rpc op, it's more like a forward op - # RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE - }) - - # insert concat_op - program.global_block()._insert_op( - index=lookup_table_op_index + 2, - type="merge_ids", - inputs={ - 'Ids': [ - program.global_block().vars[varname] - for varname in ids_name - ], - 'X': prefetch_output_vars - }, - outputs={ - "Out": [ - program.global_block().vars[varname] - for varname in out_name - ] - }) + self.all_out_emb_vars.append(out_var) # delete lookup_table_op delete_ops(program.global_block(), [op]) # break for loop break + for index in range(len(self.pserver_endpoints)): + in_var = program.global_block().create_var( + name=str("prefetch_compress_in_tmp_" + str(index)), + type=self.all_in_ids_vars[0].type, + shape=self.all_in_ids_vars[0].shape, + dtype=self.all_in_ids_vars[0].dtype) + self.all_prefetch_input_vars.append(in_var) + + out_var = program.global_block().create_var( + name=str("prefetch_compress_out_tmp_" + str(index)), + type=self.all_out_emb_vars[0].type, + shape=self.all_out_emb_vars[0].shape, + dtype=self.all_out_emb_vars[0].dtype) + self.all_prefetch_output_vars.append(out_var) + + # insert split_ids_op + program.global_block()._insert_op( + index=lookup_table_op_index, + type="split_ids", + inputs={'Ids': self.all_in_ids_vars}, + outputs={"Out": self.all_prefetch_input_vars}) + + # insert prefetch_op + program.global_block()._insert_op( + index=lookup_table_op_index + 1, + type="prefetch", + inputs={'X': self.all_prefetch_input_vars}, + outputs={"Out": self.all_prefetch_output_vars}, + attrs={ + "epmap": pserver_endpoints, + # FIXME(qiao) temporarily disable this config because prefetch + # is not act as other rpc op, it's more like a forward op + # RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + }) + + # insert concat_op + program.global_block()._insert_op( + index=lookup_table_op_index + 2, + type="merge_ids", + inputs={ + 'Ids': self.all_in_ids_vars, + 'Rows': self.all_prefetch_input_vars, + 'X': self.all_prefetch_output_vars + }, + outputs={"Out": self.all_out_emb_vars}) + def _split_table_grad_and_add_send_vars(self, program, pserver_endpoints): # 2. add split_ids_op and send_op to send gradient to pservers + # there should only be one table_name all_ops = program.global_block().ops table_grad_name = grad_var_name(self.table_name) @@ -1131,7 +1175,8 @@ to transpile() call.") inputs={ 'Ids': [program.global_block().vars[table_grad_name]] }, - outputs={"Out": self.trainer_side_table_grad_list}) + outputs={"Out": self.trainer_side_table_grad_list}, + attrs={RPC_OP_ROLE_ATTR_NAME: DIST_OP_ROLE_ATTR_VALUE}) program.global_block()._insert_op( index=op_index + 2, type="send", @@ -1142,8 +1187,9 @@ to transpile() call.") if self.sync_mode else [] }, attrs={ - "sync_mode": self.sync_mode, + "sync_mode": not self.sync_mode, "epmap": pserver_endpoints, + "trainer_id": self.trainer_id, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE, OP_ROLE_VAR_ATTR_NAME: [ self.grad_name_to_param_name[table_grad_name], @@ -1157,38 +1203,45 @@ to transpile() call.") # STEP: create prefetch block table_var = pserver_program.global_block().vars[self.table_name] prefetch_var_name_to_block_id = [] - for index in range(len(self.all_prefetch_input_vars)): - prefetch_block = pserver_program._create_block(optimize_block.idx) - trainer_ids = self.all_prefetch_input_vars[index][pserver_index] - pserver_ids = pserver_program.global_block().create_var( - name=trainer_ids.name, - type=trainer_ids.type, - shape=trainer_ids.shape, - dtype=trainer_ids.dtype) - trainer_out = self.all_prefetch_output_vars[index][pserver_index] - pserver_out = pserver_program.global_block().create_var( - name=trainer_out.name, - type=trainer_out.type, - shape=trainer_out.shape, - dtype=trainer_out.dtype) - prefetch_block.append_op( - type="lookup_sparse_table", - inputs={'Ids': pserver_ids, - "W": table_var}, - outputs={"Out": pserver_out}, - attrs={ - "is_sparse": True, # has no effect on lookup_table op - "is_distributed": True, - "padding_idx": -1 - }) - prefetch_var_name_to_block_id.append(trainer_ids.name + ":" + str( - prefetch_block.idx)) + prefetch_block = pserver_program._create_block(optimize_block.idx) + trainer_ids = self.all_prefetch_input_vars[pserver_index] + pserver_ids = pserver_program.global_block().create_var( + name=trainer_ids.name, + type=trainer_ids.type, + shape=trainer_ids.shape, + dtype=trainer_ids.dtype) + trainer_out = self.all_prefetch_output_vars[pserver_index] + pserver_out = pserver_program.global_block().create_var( + name=trainer_out.name, + type=trainer_out.type, + shape=trainer_out.shape, + dtype=trainer_out.dtype) + prefetch_block.append_op( + type="lookup_sparse_table", + inputs={'Ids': pserver_ids, + "W": table_var}, + outputs={"Out": pserver_out}, + attrs={ + "is_sparse": True, # has no effect on lookup_table op + "is_distributed": True, + "padding_idx": -1 + }) + prefetch_var_name_to_block_id.append(trainer_ids.name + ":" + str( + prefetch_block.idx)) return prefetch_var_name_to_block_id def _create_table_optimize_block(self, pserver_index, pserver_program, pre_block_idx, grad_to_block_id): # STEP: create table optimize block + table_opt_block = pserver_program._create_block(pre_block_idx) # create table param and grad var in pserver program + # create table optimize block in pserver program + table_opt_op = [ + op for op in self.optimize_ops + if 'Param' in op.input_names and op.input("Param")[0] == + self.table_name + ][0] + origin_param_var = self.origin_program.global_block().vars[ self.table_name] @@ -1204,19 +1257,16 @@ to transpile() call.") dtype=origin_param_var.dtype, type=core.VarDesc.VarType.SELECTED_ROWS, persistable=True) + # parameter must be selected rows param_var.desc.set_type(core.VarDesc.VarType.SELECTED_ROWS) grad_var = pserver_program.global_block()._clone_variable( self.origin_program.global_block().vars[grad_var_name( self.table_name)]) - # create table optimize block in pserver program - table_opt_op = [ - op for op in self.optimize_ops - if 'Param' in op.input_names and op.input("Param")[0] == - self.table_name - ][0] - table_opt_block = pserver_program._create_block(pre_block_idx) + lr_var = pserver_program.global_block()._clone_variable( + self.origin_program.global_block().vars[table_opt_op.input( + "LearningRate")[0]]) if self.sync_mode: # create grad vars in pserver program @@ -1248,8 +1298,6 @@ to transpile() call.") grad_var = pserver_program.global_block()._rename_var( origin_grad_name, splited_grad_name) - lr_var = pserver_program.global_block().vars[table_opt_op.input( - "LearningRate")[0]] inputs = { "Param": [param_var], "Grad": [grad_var], @@ -1257,7 +1305,6 @@ to transpile() call.") } outputs = {"ParamOut": [param_var]} # only support sgd now - import logging logging.warn( "distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of " + table_opt_op.type) @@ -1272,7 +1319,6 @@ to transpile() call.") """ create a new block to handle save checkpoint. """ - import os pserver_program.global_block().create_var( name="kLookupTablePath", @@ -1358,16 +1404,6 @@ to transpile() call.") program.global_block()._sync_with_cpp() return var_mapping - def _create_splited_vars(self, source_var, block, tag): - return [ - block.create_var( - name=str(source_var.name + tag + str(index)), - type=source_var.type, - shape=source_var.shape, - dtype=source_var.dtype) - for index in range(len(self.pserver_endpoints)) - ] - def _clone_var(self, block, var, persistable=True): return block.create_var( name=var.name, @@ -1425,7 +1461,7 @@ to transpile() call.") elif op_type == "adamax": if varkey in ["Moment", "InfNorm"]: return param_shape - elif op_type == "momentum": + elif op_type in ["momentum", "lars_momentum"]: if varkey == "Velocity": return param_shape elif op_type == "rmsprop": @@ -1434,8 +1470,15 @@ to transpile() call.") elif op_type == "decayed_adagrad": if varkey == "Moment": return param_shape + elif op_type == "ftrl": + if varkey in ["SquaredAccumulator", "LinearAccumulator"]: + return param_shape elif op_type == "sgd": pass + else: + raise ValueError( + "Not supported optimizer for distributed training: %s" % + op_type) return orig_shape def _get_varname_parts(self, varname): @@ -1502,6 +1545,68 @@ to transpile() call.") attrs={"scale": 1.0 / float(self.trainer_num)}) return merged_var + def _append_dc_asgd_ops(self, block, param_var, grad_var): + # NOTE: can not use grammar candy here, should put ops in specific block + local_param_bak = block.create_var( + name="%s.local_bak" % param_var.name, + shape=param_var.shape, + type=param_var.type, + dtype=param_var.dtype, + persistable=False) + # trainer_id_var is block local + trainer_id_var = block.create_var( + name="@TRAINER_ID@", + type=core.VarDesc.VarType.LOD_TENSOR, + dtype=core.VarDesc.VarType.INT64, + shape=[1], + persistable=False) + + # ref_inputs = [x[1] for x in self.param_bak_list] + ref_inputs = [] + for p, p_bak in self.param_bak_list: + if p.name == param_var.name: + ref_inputs.append(p_bak) + block.append_op( + type="ref_by_trainer_id", + inputs={"X": ref_inputs, + "TrainerId": trainer_id_var}, + outputs={"Out": local_param_bak}) + + def __create_temp_var__(): + return block.create_var( + name=unique_name.generate("tmp_dc_output"), + shape=param_var.shape, + type=param_var.type, + dtype=param_var.dtype, + persistable=False) + + o1 = __create_temp_var__() + block.append_op( + type="elementwise_sub", + inputs={"X": param_var, + "Y": local_param_bak}, + outputs={"Out": o1}) + o2 = __create_temp_var__() + block.append_op( + type="elementwise_mul", + inputs={"X": o1, + "Y": grad_var}, + outputs={"Out": o2}) + o3 = __create_temp_var__() + block.append_op( + type="elementwise_mul", + inputs={"X": o2, + "Y": grad_var}, + outputs={"Out": o3}) + # TODO(typhoonzero): append scale + o4 = __create_temp_var__() + block.append_op( + type="elementwise_add", + inputs={"X": grad_var, + "Y": o3}, + outputs={"Out": o4}) + return o4 + def _append_pserver_ops(self, optimize_block, opt_op, endpoint, grad_to_block_id, origin_program, merged_var): program = optimize_block.program @@ -1517,9 +1622,16 @@ to transpile() call.") break return param_block + if self.config.enable_dc_asgd: + param_var = _get_param_block(opt_op) + dc = self._append_dc_asgd_ops(optimize_block, param_var, merged_var) + for key in opt_op.input_names: if key == "Grad": - new_inputs[key] = merged_var + if self.config.enable_dc_asgd: + new_inputs[key] = dc + else: + new_inputs[key] = merged_var elif key == "Param": param_block = _get_param_block(opt_op) if not param_block: @@ -1571,13 +1683,27 @@ to transpile() call.") outputs=outputs, attrs=opt_op.all_attrs()) - def _is_splited_grad_var(self, var, var_dict): + def _get_pserver_grad_param_var(self, var, var_dict): + """ + Return pserver side grad/param variable, return None + if the variable is not grad/param, e.g. + + a@GRAD -> a@GRAD.block0 + a@GRAD -> a@GRAD (a is not splited) + fc_0.w_0 -> fc_0.w_0.block_0 + fc_0.w_0 -> fc_0.w_0 (weight is not splited) + _generated_var_123 -> None + """ grad_block = None for _, g in six.iteritems(var_dict): if self._orig_varname(g.name) == self._orig_varname(var.name): + # skip per trainer vars if g.name.find(".trainer_") == -1: - grad_block = g - break + # only param or grads have splited blocks + if self._orig_varname(g.name) in self.grad_name_to_param_name or\ + self._orig_varname(g.name) in self.param_name_to_grad_name: + grad_block = g + break return grad_block def _clone_lr_op(self, program, block, op): @@ -1610,32 +1736,38 @@ to transpile() call.") for key, varlist in six.iteritems(inputs): if not isinstance(varlist, list): varlist = [varlist] - for var in varlist: - # for ops like clipping and weight decay, get the splited var + for i in range(len(varlist)): + var = varlist[i] + # for ops like clipping and weight decay, get the splited var (xxx.block0) # for inputs/outputs - grad_block = self._is_splited_grad_var( + grad_block = self._get_pserver_grad_param_var( var, program.global_block().vars) if grad_block: - inputs[key] = grad_block + varlist[i] = grad_block elif var.name not in program.global_block().vars: - program.global_block().create_var( - name=var.name, - persistable=var.persistable, - dtype=var.dtype, - shape=var.shape) + tmpvar = program.global_block()._clone_variable(var) + varlist[i] = tmpvar + else: + varlist[i] = program.global_block().vars[var.name] + inputs[key] = varlist outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, opt_op) for key, varlist in six.iteritems(outputs): if not isinstance(varlist, list): varlist = [varlist] - for var in varlist: - grad_block = self._is_splited_grad_var( + for i in range(len(varlist)): + var = varlist[i] + grad_block = self._get_pserver_grad_param_var( var, program.global_block().vars) if grad_block: - outputs[key] = grad_block + varlist[i] = grad_block elif var.name not in program.global_block().vars: - program.global_block()._clone_variable(var) + tmpvar = program.global_block()._clone_variable(var) + varlist[i] = tmpvar + else: + varlist[i] = program.global_block().vars[var.name] + outputs[key] = varlist return optimize_block.append_op( type=opt_op.type, @@ -1712,8 +1844,10 @@ to transpile() call.") lr_ops = [] block = self.origin_program.global_block() for op in block.ops: - if int(op.attr(RPC_OP_ROLE_ATTR_NAME)) == int( - LR_SCHED_OP_ROLE_ATTR_VALUE): + role_id = int(op.attr(RPC_OP_ROLE_ATTR_NAME)) + if role_id == int(LR_SCHED_OP_ROLE_ATTR_VALUE) or \ + role_id == int(LR_SCHED_OP_ROLE_ATTR_VALUE) | \ + int(OPT_OP_ROLE_ATTR_VALUE): lr_ops.append(op) log("append lr op: ", op.type) return lr_ops diff --git a/python/paddle/fluid/transpiler/inference_transpiler.py b/python/paddle/fluid/transpiler/inference_transpiler.py index c402535b27142e94af339a6c18401ba20bc6564d..9a13cecc646e8534a157fad882fd97836348deb4 100644 --- a/python/paddle/fluid/transpiler/inference_transpiler.py +++ b/python/paddle/fluid/transpiler/inference_transpiler.py @@ -61,6 +61,9 @@ class InferenceTranspiler(object): raise TypeError("scope should be as Scope type or None") use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False)) + if use_mkldnn: + self._depthwise_conv_mkldnn(program) + self._fuse_batch_norm(program, place, scope) if use_mkldnn: self._fuse_conv_bias_mkldnn(program) @@ -70,11 +73,36 @@ class InferenceTranspiler(object): program) # ResNet residual block merging self._fuse_bn_relu_mkldnn(program) + def _depthwise_conv_mkldnn(self, program): + ''' + Transpile the program by replacing depthwise_conv2d to conv2d for MKLDNN program. + The result is: + - before: + - any_other_op->depthwise_conv->any_other_op + - after: + - any_other_op->conv->any_other_op + :param program: program to transpile + :type program: Program + ''' + self.block = program.block(0) + + i = 0 + while i < len(self.block.ops): + current_op = self.block.ops[i] + if current_op.type == 'depthwise_conv2d': + current_op.desc.set_type("conv2d") + i = i + 1 + + # TODO(luotao): use clone() method to flush the program.desc in force, + # since some large program.desc will not be flushed immediately. + # And a better solution will be considered later. + program = program.clone() + def _fuse_conv_eltwise_mkldnn(self, program): ''' Transpile the program fusing elementwise_add into conv for MKLDNN program. Elementwise add following convolution OP can be fused by adding - 'fuse_eltwise' attribute to convolution OP and replacing its output + 'fuse_residual_connection' attribute to convolution OP and replacing its output Tensor with second parameter of elementwise_add. The result of fuse is: - before: @@ -92,7 +120,8 @@ class InferenceTranspiler(object): if current_op.type in ['conv2d']: next_op = self.block.ops[i + 1] if next_op.type == 'elementwise_add': - self._fuse_conv_eltwise(current_op, next_op) + self._fuse_conv_eltwise(i, current_op, next_op) + self.block._remove_op(i + 1) # Remove old conv self.block._remove_op(i + 1) # Remove elementwise_add i = i + 1 self._adjust_input() @@ -444,7 +473,7 @@ class InferenceTranspiler(object): outputs={"Output": out_var}, attrs=attrs) - def _fuse_conv_eltwise(self, conv_op, eltwise_op): + def _fuse_conv_eltwise(self, index, conv_op, eltwise_op): ''' fuse the conv op with elementwise_add @@ -454,9 +483,30 @@ class InferenceTranspiler(object): :type eltwise_op: Operator ''' - conv_op._set_attr("fuse_eltwise", True) - self.input_map[conv_op.output("Output")[0]] = eltwise_op.input("Y")[0] - self.input_map[eltwise_op.output("Out")[0]] = eltwise_op.input("Y")[0] + eltwise_input = "X" + if eltwise_op.input("X")[0] == conv_op.output("Output")[0]: + eltwise_input = "Y" + + residual_var = self.block.vars[eltwise_op.input(eltwise_input)[0]] + out_var = self.block.vars[eltwise_op.output("Out")[0]] + filter_var = self.block.vars[conv_op.input("Filter")[0]] + in_var = self.block.vars[conv_op.input("Input")[0]] + bias_var = self.block.vars[conv_op.input("Bias")[0]] + + conv_op._set_attr("fuse_residual_connection", True) + attrs = {name: conv_op.attr(name) for name in conv_op.attr_names} + + self.block._insert_op( + index, + type="conv2d", + inputs={ + "Input": in_var, + "Filter": filter_var, + "Bias": bias_var, + "ResidualData": residual_var + }, + outputs={"Output": out_var}, + attrs=attrs) def _adjust_input(self): for i in range(len(self.block.ops)): diff --git a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py index 861bb5fae5d7a8561ded1f547fbb86ae1e1a073e..c9f1be934773cc28f026f2b867b9e3a4f7aa8472 100755 --- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -171,7 +171,7 @@ class ControlFlowGraph(object): self._live_out[i] |= self._live_in[s] self._live_in[i] = self._uses[i] | ( self._live_out[i] - self._defs[i]) - if live_in[i] != self._live_in[i]: + if live_in[i] != set(self._live_in[i]): for d in self._presuccessors[i]: worklist.append(d) @@ -321,8 +321,7 @@ class ControlFlowGraph(object): if not compare_shape(x_shape, cache_shape, level): continue - # TODO(qijun): actually, we should compare - # dtype_to_size[x_dtype] and dtype_to_size[cache_dtype] + # TODO(qijun): dtype_to_size[x_dtype] and dtype_to_size[cache_dtype] if x_dtype != cache_dtype: continue @@ -487,7 +486,6 @@ def memory_optimize(input_program, skip_opt_set = grad_set else: skip_opt_set.update(grad_set) - cfgs = _get_cfgs(input_program) for cfg in cfgs: cfg.memory_optimize(skip_opt_set=skip_opt_set, level=level) diff --git a/python/paddle/utils/__init__.py b/python/paddle/utils/__init__.py index 15595d208583b567b8f768c8d7bd84986ca5a03f..db6fe2d5fff4ed1617d793faee23f01395841768 100644 --- a/python/paddle/utils/__init__.py +++ b/python/paddle/utils/__init__.py @@ -12,4 +12,5 @@ # See the License for the specific language governing permissions and # limitations under the License. -__all__ = ['dump_config'] +from .plot import Ploter +__all__ = ['dump_config', 'Ploter'] diff --git a/python/paddle/utils/plot.py b/python/paddle/utils/plot.py new file mode 100644 index 0000000000000000000000000000000000000000..08889c0313fc24151cde6ca7b662d81eb53c9d7b --- /dev/null +++ b/python/paddle/utils/plot.py @@ -0,0 +1,115 @@ +# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os + + +class PlotData(object): + def __init__(self): + self.step = [] + self.value = [] + + def append(self, step, value): + self.step.append(step) + self.value.append(value) + + def reset(self): + self.step = [] + self.value = [] + + +class Ploter(object): + """ + Plot input data in a 2D graph + + Args: + title: assign the title of input data. + step: x_axis of the data. + value: y_axis of the data. + """ + + def __init__(self, *args): + self.__args__ = args + self.__plot_data__ = {} + for title in args: + self.__plot_data__[title] = PlotData() + # demo in notebooks will use Ploter to plot figure, but when we convert + # the ipydb to py file for testing, the import of matplotlib will make the + # script crash. So we can use `export DISABLE_PLOT=True` to disable import + # these libs + self.__disable_plot__ = os.environ.get("DISABLE_PLOT") + if not self.__plot_is_disabled__(): + import matplotlib.pyplot as plt + from IPython import display + self.plt = plt + self.display = display + + def __plot_is_disabled__(self): + return self.__disable_plot__ == "True" + + def append(self, title, step, value): + """ + Feed data + + Args: + title: assign the group data to this subtitle. + step: the x_axis of data. + value: the y_axis of data. + + Examples: + .. code-block:: python + plot_curve = Ploter("Curve 1","Curve 2") + plot_curve.append(title="Curve 1",step=1,value=1) + """ + assert isinstance(title, basestring) + assert self.__plot_data__.has_key(title) + data = self.__plot_data__[title] + assert isinstance(data, PlotData) + data.append(step, value) + + def plot(self, path=None): + """ + Plot data in a 2D graph + + Args: + path: store the figure to this file path. Defaul None. + + Examples: + .. code-block:: python + plot_curve = Ploter() + plot_cure.plot() + """ + if self.__plot_is_disabled__(): + return + + titles = [] + for title in self.__args__: + data = self.__plot_data__[title] + assert isinstance(data, PlotData) + if len(data.step) > 0: + titles.append(title) + self.plt.plot(data.step, data.value) + self.plt.legend(titles, loc='upper left') + if path is None: + self.display.clear_output(wait=True) + self.display.display(self.plt.gcf()) + else: + self.plt.savefig(path) + self.plt.gcf().clear() + + def reset(self): + for key in self.__plot_data__: + data = self.__plot_data__[key] + assert isinstance(data, PlotData) + data.reset() diff --git a/python/setup.py.in b/python/setup.py.in index b376be0ea373f089ef17f27435d979712fbdff72..c623057d5081a6fedcd90eb5f5d53531a5d62bb8 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -14,7 +14,8 @@ RC = 0 def git_commit(): try: cmd = ['git', 'rev-parse', 'HEAD'] - git_commit = subprocess.Popen(cmd, stdout = subprocess.PIPE).communicate()[0].strip() + git_commit = subprocess.Popen(cmd, stdout = subprocess.PIPE, + cwd="@PADDLE_SOURCE_DIR@").communicate()[0].strip() except: git_commit = 'Unknown' git_commit = git_commit.decode() @@ -27,7 +28,7 @@ def _get_version_detail(idx): if re.match('@TAG_VERSION_REGEX@', '@PADDLE_VERSION@'): version_details = '@PADDLE_VERSION@'.split('.') - if len(version_details) == 3: + if len(version_details) >= 3: return version_details[idx] return 0 @@ -44,7 +45,7 @@ def get_patch(): def is_taged(): try: cmd = ['git', 'describe', '--exact-match', '--tags', 'HEAD', '2>/dev/null'] - git_tag = subprocess.Popen(cmd, stdout = subprocess.PIPE).communicate()[0].strip() + git_tag = subprocess.Popen(cmd, stdout = subprocess.PIPE, cwd="@PADDLE_SOURCE_DIR@").communicate()[0].strip() git_tag = git_tag.decode() except: return False @@ -55,8 +56,7 @@ def is_taged(): return False def write_version_py(filename='paddle/version.py'): - cnt = ''' -# THIS FILE IS GENERATED FROM PADDLEPADDLE SETUP.PY + cnt = '''# THIS FILE IS GENERATED FROM PADDLEPADDLE SETUP.PY # full_version = '%(major)d.%(minor)d.%(patch)s' major = '%(major)d' @@ -174,6 +174,18 @@ if '${CMAKE_BUILD_TYPE}' == 'Release': raise Exception("patch libmkldnn.so failed, command: %s" % command) package_data['paddle.libs']+=['libmkldnn.so.0'] shutil.copy('${MKLDNN_SHARED_LIB}', libs_path) +if '${WITH_NGRAPH}' == 'ON': + if '${CMAKE_BUILD_TYPE}' == 'Release': + # only change rpath in Release mode. + command = "patchelf --set-rpath '$ORIGIN/' ${NGRAPH_SHARED_LIB}" + if os.system(command) != 0: + raise Exception("patch ${NGRAPH_SHARED_LIB_NAME} failed, command: %s" % command) + shutil.copy('${NGRAPH_SHARED_LIB}', libs_path) + shutil.copy('${NGRAPH_CPU_LIB}', libs_path) + shutil.copy('${NGRAPH_TBB_LIB}', libs_path) + package_data['paddle.libs']+=['${NGRAPH_SHARED_LIB_NAME}', + '${NGRAPH_CPU_LIB_NAME}', + '${NGRAPH_TBB_LIB_NAME}'] # remove unused paddle/libs/__init__.py os.remove(libs_path+'/__init__.py') package_dir['paddle.libs']=libs_path