diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index 89c620bb2f7ef634fa80b64eec7037e8cb9a190c..6140340890c0e5025eb08209e8ea78df918b4dc0 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -1,3 +1,4 @@ +repos: - repo: https://github.com/Lucas-C/pre-commit-hooks.git sha: v1.0.1 hooks: @@ -25,6 +26,14 @@ entry: bash ./.clang_format.hook -i language: system files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$ +- repo: local + hooks: + - id: cpplint-cpp-source + name: cpplint + description: Check C++ code style using cpplint.py. + entry: bash ./tools/codestyle/cpplint_pre_commit.hook + language: system + files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx)$ - repo: https://github.com/PaddlePaddle/pre-commit-golang sha: 8337620115c25ff8333f1b1a493bd031049bd7c0 hooks: diff --git a/.travis.yml b/.travis.yml index bf6a41d13c4eabc2d8543ab821ce0ff747a061df..929c847bd36d64e79a199b2634ebf68c3225429b 100644 --- a/.travis.yml +++ b/.travis.yml @@ -34,7 +34,7 @@ addons: - automake - libtool - ccache - ssh_known_hosts: 52.76.173.135 + ssh_known_hosts: 13.229.163.131 before_install: - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python diff --git a/CMakeLists.txt b/CMakeLists.txt index 0ec65bac84b0b0d89123473a8941f80c90f1b339..c649aafeddaf9f28c213d086236c3779d3137d92 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -36,6 +36,7 @@ include(simd) ################################ Configurations ####################################### option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND}) +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_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON) @@ -52,8 +53,7 @@ option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF) option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF) option(ON_TRAVIS "Exclude special unit test on Travis CI" OFF) option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF) -# TODO: Only compile PaddlePaddle fluid version by WITH_FLUID option. -option(WITH_FLUID "Compile PaddlePaddle fluid only(TODO)" OFF) +option(WITH_FLUID_ONLY "Compile PaddlePaddle fluid only" OFF) option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) option(GLIDE_INSTALL "Download and install go dependencies " ON) option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) @@ -108,7 +108,7 @@ if (WITH_C_API AND WITH_PYTHON) endif() if (WITH_C_API) - set(WITH_FLUID OFF CACHE STRING "Disable install fluid when compile the C_API" FORCE) + set(WITH_FLUID_ONLY OFF CACHE STRING "Disable install fluid when compile the C_API" FORCE) endif() if(MOBILE_INFERENCE) @@ -146,6 +146,7 @@ include(external/cares) include(external/grpc) include(external/snappy) # download snappy include(external/snappystream) +include(external/threadpool) include(cudnn) # set cudnn libraries, must before configure include(cupti) @@ -180,6 +181,11 @@ if(WITH_GPU) include(cuda) endif(WITH_GPU) +if(WITH_AMD_GPU) + find_package(HIP) + include(hip) +endif(WITH_AMD_GPU) + if(WITH_MKLML) list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB}) endif() diff --git a/benchmark/cluster/vgg16/vgg16_fluid.py b/benchmark/cluster/vgg16/vgg16_fluid.py index 786f224608f7d41c438411de0e09fedbcf2264b8..8b29227cfab2a36d5b9f6d17b837b33da8d2a92e 100644 --- a/benchmark/cluster/vgg16/vgg16_fluid.py +++ b/benchmark/cluster/vgg16/vgg16_fluid.py @@ -18,12 +18,13 @@ import sys import time import numpy as np import paddle.v2 as paddle -import paddle.v2.fluid as fluid -import paddle.v2.fluid.core as core -import paddle.v2.fluid.profiler as profiler +import paddle.fluid as fluid +import paddle.fluid.core as core +import paddle.fluid.profiler as profiler import argparse import functools import os +from paddle.fluid import debuger def str2bool(v): @@ -182,28 +183,27 @@ def main(): start_time = time.time() num_samples = 0 train_pass_acc.reset() - with profiler.profiler("CPU", 'total') as prof: - for batch_id, data in enumerate(train_reader()): - ts = time.time() - img_data = np.array( - map(lambda x: x[0].reshape(data_shape), data)).astype( - "float32") - y_data = np.array(map(lambda x: x[1], data)).astype("int64") - y_data = y_data.reshape([-1, 1]) - - loss, acc, b_size = exe.run( - trainer_prog, - feed={"pixel": img_data, - "label": y_data}, - fetch_list=[avg_cost, batch_acc, batch_size]) - iters += 1 - num_samples += len(data) - train_pass_acc.add(value=acc, weight=b_size) - print( - "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed = %.2f img/s" - % (pass_id, iters, loss, acc, - len(data) / (time.time() - ts)) - ) # The accuracy is the accumulation of batches, but not the current batch. + for batch_id, data in enumerate(train_reader()): + ts = time.time() + img_data = np.array( + map(lambda x: x[0].reshape(data_shape), data)).astype( + "float32") + y_data = np.array(map(lambda x: x[1], data)).astype("int64") + y_data = y_data.reshape([-1, 1]) + + loss, acc, b_size = exe.run( + trainer_prog, + feed={"pixel": img_data, + "label": y_data}, + fetch_list=[avg_cost, batch_acc, batch_size]) + iters += 1 + num_samples += len(data) + train_pass_acc.add(value=acc, weight=b_size) + print( + "Pass = %d, Iters = %d, Loss = %f, Accuracy = %f, Speed = %.2f img/s" + % (pass_id, iters, loss, acc, + len(data) / (time.time() - ts)) + ) # The accuracy is the accumulation of batches, but not the current batch. pass_elapsed = time.time() - start_time pass_train_acc = train_pass_acc.eval() @@ -254,9 +254,7 @@ def main(): pserver_prog = t.get_pserver_program(current_endpoint) pserver_startup = t.get_startup_program(current_endpoint, pserver_prog) - print("starting server side startup") exe.run(pserver_startup) - print("starting parameter server...") exe.run(pserver_prog) elif training_role == "TRAINER": # Parameter initialization diff --git a/benchmark/cluster/vgg16/vgg16_tf.py b/benchmark/cluster/vgg16/vgg16_tf.py index 996df0e314b867ea8de618dfd3977f490fbe8372..2d220478acae46566760209dbc012cff316946aa 100644 --- a/benchmark/cluster/vgg16/vgg16_tf.py +++ b/benchmark/cluster/vgg16/vgg16_tf.py @@ -292,14 +292,18 @@ def run_benchmark(cluster_spec, server): return np.mean(test_accs) config = tf.ConfigProto( - intra_op_parallelism_threads=1, inter_op_parallelism_threads=1) + intra_op_parallelism_threads=1, + inter_op_parallelism_threads=1, + log_device_placement=True) config.gpu_options.allow_growth = True hooks = [tf.train.StopAtStepHook(last_step=1000000)] with tf.train.MonitoredTrainingSession( - master=server.target, is_chief=(args.task_index == 0), - hooks=hooks) as sess: + master=server.target, + is_chief=(args.task_index == 0), + hooks=hooks, + config=config) as sess: iters, num_samples, start_time = 0, 0, 0.0 for pass_id in range(args.num_passes): # train diff --git a/cmake/configure.cmake b/cmake/configure.cmake index 0f76f55270592c5625a9624b33f4c0f82efdc627..f726405c4773994f6ca6509e5218750805b03995 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -57,11 +57,7 @@ if(NOT WITH_GOLANG) add_definitions(-DPADDLE_WITHOUT_GOLANG) endif(NOT WITH_GOLANG) -if(NOT WITH_GPU) - add_definitions(-DHPPL_STUB_FUNC) - - list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu) -else() +if(WITH_GPU) add_definitions(-DPADDLE_WITH_CUDA) FIND_PACKAGE(CUDA REQUIRED) @@ -84,7 +80,14 @@ else() # Include cuda and cudnn include_directories(${CUDNN_INCLUDE_DIR}) include_directories(${CUDA_TOOLKIT_INCLUDE}) -endif(NOT WITH_GPU) +elseif(WITH_AMD_GPU) + add_definitions(-DPADDLE_WITH_HIP) + set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__") + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__HIP_PLATFORM_HCC__") +else() + add_definitions(-DHPPL_STUB_FUNC) + list(APPEND CMAKE_CXX_SOURCE_FILE_EXTENSIONS cu) +endif() if (WITH_MKLML AND MKLML_IOMP_LIB) message(STATUS "Enable Intel OpenMP with ${MKLML_IOMP_LIB}") diff --git a/cmake/external/boost.cmake b/cmake/external/boost.cmake index d9cd264b49d546c35a2c57a82ead83ea654b60ae..10662fc96704685f030a5d76c6857d4bc20a63d9 100644 --- a/cmake/external/boost.cmake +++ b/cmake/external/boost.cmake @@ -24,7 +24,7 @@ set(BOOST_PROJECT "extern_boost") # So we use 1.41.0 here. set(BOOST_VER "1.41.0") set(BOOST_TAR "boost_1_41_0") -set(BOOST_URL "http://paddlepaddledeps.s3-website-us-west-1.amazonaws.com/${BOOST_TAR}.tar.gz") +set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz") set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost) set(BOOST_DOWNLOAD_DIR "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}") set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE) diff --git a/cmake/external/eigen.cmake b/cmake/external/eigen.cmake index 6a701e076c95372f903a09d35d4208ee73bd584c..73d70c34dce8bedd9e62519c207e5be3dcf7dba3 100644 --- a/cmake/external/eigen.cmake +++ b/cmake/external/eigen.cmake @@ -4,18 +4,33 @@ SET(EIGEN_SOURCE_DIR ${THIRD_PARTY_PATH}/eigen3) SET(EIGEN_INCLUDE_DIR ${EIGEN_SOURCE_DIR}/src/extern_eigen3) INCLUDE_DIRECTORIES(${EIGEN_INCLUDE_DIR}) -ExternalProject_Add( - extern_eigen3 - ${EXTERNAL_PROJECT_LOG_ARGS} - GIT_REPOSITORY "https://github.com/RLovelett/eigen.git" - GIT_TAG 70661066beef694cadf6c304d0d07e0758825c10 - PREFIX ${EIGEN_SOURCE_DIR} - UPDATE_COMMAND "" - CONFIGURE_COMMAND "" - BUILD_COMMAND "" - INSTALL_COMMAND "" - TEST_COMMAND "" -) +if(WITH_AMD_GPU) + ExternalProject_Add( + extern_eigen3 + ${EXTERNAL_PROJECT_LOG_ARGS} + GIT_REPOSITORY "https://github.com/sabreshao/hipeigen.git" + GIT_TAG 0cba03ff9f8f9f70bbd92ac5857b031aa8fed6f9 + PREFIX ${EIGEN_SOURCE_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + INSTALL_COMMAND "" + TEST_COMMAND "" + ) +else() + ExternalProject_Add( + extern_eigen3 + ${EXTERNAL_PROJECT_LOG_ARGS} + GIT_REPOSITORY "https://github.com/RLovelett/eigen.git" + GIT_TAG 70661066beef694cadf6c304d0d07e0758825c10 + PREFIX ${EIGEN_SOURCE_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + INSTALL_COMMAND "" + TEST_COMMAND "" + ) +endif() if (${CMAKE_VERSION} VERSION_LESS "3.3.0") set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/eigen3_dummy.c) diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake index 739a910c7c670b7b9f89e543582a32a80546fb11..df3f0c7f0c31efaa127515bb98e5668b8f9df199 100644 --- a/cmake/external/mklml.cmake +++ b/cmake/external/mklml.cmake @@ -34,7 +34,7 @@ SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}") SET(MKLML_DST_DIR "mklml") SET(MKLML_INSTALL_ROOT "${THIRD_PARTY_PATH}/install") SET(MKLML_INSTALL_DIR ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR}) -SET(MKLML_ROOT ${MKLML_INSTALL_DIR}/${MKLML_VER}) +SET(MKLML_ROOT ${MKLML_INSTALL_DIR}) SET(MKLML_INC_DIR ${MKLML_ROOT}/include) SET(MKLML_LIB_DIR ${MKLML_ROOT}/lib) SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so) @@ -46,7 +46,7 @@ INCLUDE_DIRECTORIES(${MKLML_INC_DIR}) FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt "PROJECT(MKLML)\n" "cmake_minimum_required(VERSION 3.0)\n" - "install(DIRECTORY ${MKLML_VER}\n" + "install(DIRECTORY ${MKLML_VER}/include ${MKLML_VER}/lib \n" " DESTINATION ${MKLML_DST_DIR})\n") ExternalProject_Add( diff --git a/cmake/external/threadpool.cmake b/cmake/external/threadpool.cmake new file mode 100644 index 0000000000000000000000000000000000000000..0159815fed81bdff6de3e561af569e9edc75f947 --- /dev/null +++ b/cmake/external/threadpool.cmake @@ -0,0 +1,30 @@ +INCLUDE(ExternalProject) + +SET(THREADPOOL_SOURCE_DIR ${THIRD_PARTY_PATH}/threadpool) +SET(THREADPOOL_INCLUDE_DIR ${THREADPOOL_SOURCE_DIR}/src/extern_threadpool) +INCLUDE_DIRECTORIES(${THREADPOOL_INCLUDE_DIR}) + +ExternalProject_Add( + extern_threadpool + ${EXTERNAL_PROJECT_LOG_ARGS} + GIT_REPOSITORY "https://github.com/progschj/ThreadPool.git" + GIT_TAG 9a42ec1329f259a5f4881a291db1dcb8f2ad9040 + PREFIX ${THREADPOOL_SOURCE_DIR} + UPDATE_COMMAND "" + CONFIGURE_COMMAND "" + BUILD_COMMAND "" + INSTALL_COMMAND "" + TEST_COMMAND "" +) + +if (${CMAKE_VERSION} VERSION_LESS "3.3.0") + set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/threadpool_dummy.c) + file(WRITE ${dummyfile} "const char *dummy_threadpool = \"${dummyfile}\";") + add_library(simple_threadpool STATIC ${dummyfile}) +else() + add_library(simple_threadpool INTERFACE) +endif() + +add_dependencies(simple_threadpool extern_threadpool) + +LIST(APPEND external_project_dependencies simple_threadpool) diff --git a/cmake/generic.cmake b/cmake/generic.cmake index 471e3929069d0d28105404b4f0f6baa303faf0e0..3fe750f47efc149bb1af6086841bffd5dd8e85fd 100644 --- a/cmake/generic.cmake +++ b/cmake/generic.cmake @@ -317,6 +317,82 @@ function(nv_test TARGET_NAME) endif() endfunction(nv_test) +function(hip_library TARGET_NAME) + if (WITH_AMD_GPU) + set(options STATIC static SHARED shared) + set(oneValueArgs "") + set(multiValueArgs SRCS DEPS) + cmake_parse_arguments(hip_library "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + set(_sources ${hip_library_SRCS}) + HIP_PREPARE_TARGET_COMMANDS(${TARGET_NAME} OBJ _generated_files _source_files ${_sources} HIPCC_OPTIONS ${_hipcc_options} HCC_OPTIONS ${_hcc_options} NVCC_OPTIONS ${_nvcc_options}) + if(_source_files) + list(REMOVE_ITEM _sources ${_source_files}) + endif() + if(hip_library_SRCS) + if (hip_library_SHARED OR hip_library_shared) # build *.so + add_library(${TARGET_NAME} SHARED ${_cmake_options} ${_generated_files} ${_sources}) + set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP) + else() + add_library(${TARGET_NAME} STATIC ${_cmake_options} ${_generated_files} ${_sources}) + set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE CXX) + target_link_libraries(${TARGET_NAME} /opt/rocm/hip/lib/libhip_hcc.so /opt/rocm/hip/lib/libhip_device.a) + find_fluid_modules(${TARGET_NAME}) + endif() + if (hip_library_DEPS) + add_dependencies(${TARGET_NAME} ${hip_library_DEPS}) + target_link_libraries(${TARGET_NAME} ${hip_library_DEPS}) + endif() + # cpplint code style + foreach(source_file ${hip_library_SRCS}) + string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file}) + if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + list(APPEND hip_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h) + endif() + endforeach() + add_style_check_target(${TARGET_NAME} ${hip_library_SRCS} ${hip_library_HEADERS}) + else(hip_library_SRCS) + if (hip_library_DEPS) + merge_static_libs(${TARGET_NAME} ${hip_library_DEPS}) + else() + message(FATAL "Please specify source file or library in nv_library.") + endif() + endif(hip_library_SRCS) + endif() +endfunction(hip_library) + +function(hip_binary TARGET_NAME) + if (WITH_AMD_GPU) + set(options "") + set(oneValueArgs "") + set(multiValueArgs SRCS DEPS) + cmake_parse_arguments(hip_binary "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + hip_add_executable(${TARGET_NAME} ${hip_binary_SRCS}) + if(hip_binary_DEPS) + target_link_libraries(${TARGET_NAME} ${hip_binary_DEPS}) + add_dependencies(${TARGET_NAME} ${hip_binary_DEPS}) + endif() + endif() +endfunction(hip_binary) + +function(hip_test TARGET_NAME) + if (WITH_AMD_GPU AND WITH_TESTING) + set(options "") + set(oneValueArgs "") + set(multiValueArgs SRCS DEPS) + cmake_parse_arguments(hip_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) + set(_sources ${hip_test_SRCS}) + HIP_PREPARE_TARGET_COMMANDS(${TARGET_NAME} OBJ _generated_files _source_files ${_sources} HIPCC_OPTIONS ${_hipcc_options} HCC_OPTIONS ${_hcc_options} NVCC_OPTIONS ${_nvcc_options}) + if(_source_files) + list(REMOVE_ITEM _sources ${_source_files}) + endif() + add_executable(${TARGET_NAME} ${_cmake_options} ${_generated_files} ${_sources}) + set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP) + target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) + add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main paddle_memory gtest gflags) + add_test(${TARGET_NAME} ${TARGET_NAME}) + endif() +endfunction(hip_test) + function(go_library TARGET_NAME) set(options STATIC static SHARED shared) set(oneValueArgs "") @@ -511,6 +587,9 @@ function(grpc_library TARGET_NAME) get_filename_component(PROTO_WE ${grpc_library_PROTO} NAME_WE) get_filename_component(PROTO_PATH ${ABS_PROTO} PATH) + #FIXME(putcn): the follwoing line is supposed to generate *.pb.h and cc, but + # somehow it didn't. line 602 to 604 is to patching this. Leaving this here + # for now to enable dist CI. protobuf_generate_cpp(grpc_proto_srcs grpc_proto_hdrs "${ABS_PROTO}") set(grpc_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.cc") set(grpc_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.h") @@ -521,6 +600,9 @@ function(grpc_library TARGET_NAME) COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}" --plugin=protoc-gen-grpc="${GRPC_CPP_PLUGIN}" "${ABS_PROTO}" + COMMAND ${PROTOBUF_PROTOC_EXECUTABLE} + ARGS --cpp_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}" + "${ABS_PROTO}" DEPENDS "${ABS_PROTO}" ${PROTOBUF_PROTOC_EXECUTABLE} extern_grpc) # FIXME(typhoonzero): grpc generated code do not generate virtual-dtor, mark it diff --git a/cmake/hip.cmake b/cmake/hip.cmake new file mode 100644 index 0000000000000000000000000000000000000000..bfe491bd6b7602959d3dd60bd06c67993593cc9b --- /dev/null +++ b/cmake/hip.cmake @@ -0,0 +1,43 @@ +if(NOT WITH_AMD_GPU) + return() +endif() + +include_directories("/opt/rocm/include") +include_directories("/opt/rocm/hipblas/include") +include_directories("/opt/rocm/hiprand/include") +include_directories("/opt/rocm/rocrand/include") +include_directories("/opt/rocm/rccl/include") +include_directories("/opt/rocm/thrust") + +list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc") + +set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++14" ) + +if(WITH_DSO) + set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_USE_DSO") +endif(WITH_DSO) + +if(WITH_DOUBLE) + set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_TYPE_DOUBLE") +endif(WITH_DOUBLE) + +if(WITH_TESTING) + set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_TESTING") +endif(WITH_TESTING) + +if(CMAKE_BUILD_TYPE STREQUAL "Debug") + list(APPEND HIP_HCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG}) +elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo") + list(APPEND HIP_HCC_FLAGS ${CMAKE_CXX_FLAGS_RELWITHDEBINFO}) +elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel") + list(APPEND HIP_HCC_FLAGS ${CMAKE_CXX_FLAGS_MINSIZEREL}) +endif() + +if("x${HCC_HOME}" STREQUAL "x") + set(HCC_HOME "/opt/rocm/hcc") +endif() + +set(CMAKE_HIP_LINK_EXECUTABLE "${HIP_HIPCC_CMAKE_LINKER_HELPER} ${HCC_HOME} -o ") +set(CMAKE_HIP_CREATE_SHARED_LIBRARY "${HIP_HIPCC_CMAKE_LINKER_HELPER} ${HCC_HOME} -o -shared") +set(CMAKE_HIP_CREATE_SHARED_MODULE "${HIP_HIPCC_CMAKE_LINKER_HELPER} ${HCC_HOME} -o -shared") + diff --git a/cmake/inference_lib.cmake b/cmake/inference_lib.cmake index 6b2237b858380f384be0aa3c6ae24a4c83ad646d..0323cd9698cba916d2aa04403be97c0a6a463830 100644 --- a/cmake/inference_lib.cmake +++ b/cmake/inference_lib.cmake @@ -69,6 +69,12 @@ if(NOT CBLAS_FOUND) SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include DSTS ${dst_dir} ${dst_dir} ) +elseif (WITH_MKLML) + set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mklml") + copy(mklml_lib + SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR} + DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir} + ) endif() # paddle fluid module diff --git a/doc/CMakeLists.txt b/doc/CMakeLists.txt index da67701ec1af57df742dce105990cffa40f45d7c..7066637a7cb27b83724cb4030c29a1019981f52b 100644 --- a/doc/CMakeLists.txt +++ b/doc/CMakeLists.txt @@ -1 +1,9 @@ +add_custom_target(paddle_apis ALL + DEPENDS paddle_v2_apis paddle_fluid_apis) + +add_custom_target(paddle_docs ALL + DEPENDS paddle_v2_docs paddle_v2_docs_cn + paddle_fluid_docs paddle_fluid_docs_cn) + add_subdirectory(v2) +add_subdirectory(fluid) diff --git a/doc/design/images/parallel_executor_overview.dot b/doc/design/images/parallel_executor_overview.dot new file mode 100644 index 0000000000000000000000000000000000000000..40753cb140540c08d9d4c449b8d377e315280436 --- /dev/null +++ b/doc/design/images/parallel_executor_overview.dot @@ -0,0 +1,83 @@ +digraph G { + subgraph cluster_init { + label="Initialization" + startup_program [label="startup", shape=box] + node_w_g0 [label="W\nGPU0"] + startup_program -> node_w_g0 [label="Initialize"] + node_w_g1 [label="W\nGPU1"] + node_w_g0 -> node_w_g1 [label="broadcast"] + } + + subgraph cluster_train { + label="forward_backward" + + subgraph cluster_gpu0 { + label="GPU0" + fc_0 [label="fc\nGPU0", shape=box] + hidden_0 [label="hidden\nGPU0"] + node_w_g0 -> fc_0 + fc_0 -> hidden_0 + loss0 [label="loss\nGPU0"] + hidden_0 -> loss0 [label="many ops omitted"] + scale_loss_0 [label="scale_loss_gradient\nGPU0", shape=box] + loss_g0 [label="loss_grad\nGPU0"] + scale_loss_0->loss_g0 + + fc_g_0 [label="w_grad\nGPU0", shape=box] + loss0 -> fc_g_0 + loss_g0 -> fc_g_0 + hidden_0 -> fc_g_0 + } + + subgraph cluster_gpu1 { + label="GPU1" + fc_1 [label="fc\nGPU1", shape=box] + hidden_1 [label="hidden\nGPU1"] + node_w_g1 -> fc_1 + fc_1 -> hidden_1 + loss1 [label="loss\nGPU1"] + hidden_1 -> loss1 [label="many ops omitted"] + scale_loss_1 [label="scale_loss_gradient\nGPU1", shape=box] + loss_g1 [label="loss_grad\nGPU1"] + scale_loss_1->loss_g1 + + fc_g_1 [label="w_grad\nGPU1", shape=box] + loss1 -> fc_g_1 + loss_g1 -> fc_g_1 + hidden_1 -> fc_g_1 + } + } + + all_reduce_w [label="Merge Gradients(AllReduce)", shape=box] + fc_g_0 -> all_reduce_w + fc_g_1 -> all_reduce_w + + fc_g_0_merged [label="w_grad\nMerged\nGPU0"] + fc_g_1_merged [label="w_grad\nMerged\nGPU1"] + all_reduce_w -> fc_g_0_merged + all_reduce_w -> fc_g_1_merged + + subgraph cluster_optimization { + label="Optimization" + subgraph cluster_opt_gpu0 { + label="GPU0" + sgd_0 [label="SGD Op\nGPU0", shape=box] + + fc_g_0_merged -> sgd_0 + node_w_g0 -> sgd_0 + optimized_w_0 [label="Optimized W\nGPU0"] + sgd_0 -> optimized_w_0 + } + subgraph cluster_opt_gpu1 { + label="GPU1" + sgd_1 [label="SGD Op\nGPU1", shape=box] + + fc_g_1_merged -> sgd_1 + node_w_g1 -> sgd_1 + optimized_w_1 [label="Optimized W\nGPU0"] + sgd_1 -> optimized_w_1 + } + } + + +} diff --git a/doc/design/images/parallel_executor_overview.png b/doc/design/images/parallel_executor_overview.png new file mode 100644 index 0000000000000000000000000000000000000000..d890c0ffee3b38dc7cb74a2b56c2ab4831532211 Binary files /dev/null and b/doc/design/images/parallel_executor_overview.png differ diff --git a/doc/design/parallel_executor.md b/doc/design/parallel_executor.md new file mode 100644 index 0000000000000000000000000000000000000000..9aed3b059a1595ba3971d7d5acfc0d16a731584b --- /dev/null +++ b/doc/design/parallel_executor.md @@ -0,0 +1,104 @@ +# ParallelExecutor + +## Background + +Neural network models are defined as a `ProgramDesc` in Fluid. The `ProgramDesc` can be executed by an interpreter(i.e. the `executor` concept in Fluid). The instructions or operators in a `Program` will be executed, and the results will be fetched in Python side. + +The executor is a very naive interpreter. It runs operators one by one. We can use `Parallel.Do` to support data parallelism, however, lacking device information in `ProgramDesc`; it is not possible to optimize the performance of `Parallel.Do`. + +We want a `ProgramDesc` can be run on different nodes. It is better not to contain device information in `ProgramDesc`. However, we can write a high-performance interpreter, which can hold an alternative intermediate representation of `ProgramDesc`, to take full usage of Multi-GPUs. + +ParallelExecutor is an interpreter of `ProgramDesc` which will [out-of-order execute](https://en.wikipedia.org/wiki/Out-of-order_execution) `Program` in data parallelism mode and maximise the utility of Multi-GPUs. + + +## Overview of MultiGPUs logic + +The ParallelExecutor takes the startup program and main program as inputs. The parameters will be initialised on `GPU0` by startup program and will broadcast to multi-GPUs. The main program will be duplicated into multi-GPUs. The gradient will be merged during each iteration, and each device will optimize parameters independently. Since the gradients on each device will be merged before parameter optimization, the parameters will be the same on each device and it does not need to be broadcast the parameters. + +![alt](images/parallel_executor_overview.png) + +There are several optimizations for this logic. + +1. We use an alternate representation in ParallelExecutor. It because the device information is critical for performance optimization. +2. The execution is out-of-order, i.e., an operator will be executed whenever the inputs of the operator are ready. + * GPU is a high-performance device; only one CPU thread cannot fulfil one GPU. So there is a thread pool to execute operators. + * Out-of-order also helps transpilers to generate `ProgramDesc`. It is no need to concern about the best order of performance when implementing a transpiler. +3. The streams of computation, merge gradients and fetch data are different. + +The performance of `ResNeXt152` on `TitanX` which `batch_size=12` is shown below. + +| Number of GPUs | 1 | 2 | 3 | 4| +| --- | --- | --- | --- | --- | +| Image/Sec | 17.9906 | 25.771 | 36.911 | 48.8428 | +| Speed Up | N/A | 1.43247029 | 2.05168255 | 2.71490667 | + + +## Static single assignment Graph + +[Static single assignment form](https://en.wikipedia.org/wiki/Static_single_assignment_form)(`SSA` for short) is a common form for compiler optimization. To implement concurrent execution, we uses an `SSA` graph as an intermedia representation of `ProgramDesc`. + +The `Program` is a directed acyclic graph, since a variable can be assigned multiple times. We enforce a variable will be assigned once, by adding version number to varaibles. We parsing the `Program` into a `SSA` graph. Also, ProgramExecutor duplicate `Program` into multi-devices. We also add a device number to varaibles and insert `NCCLAllReduce` into Graph. + +The data structure of `SSA` graph is: + +```c++ +struct VarHandleBase { + OpHandleBase* generated_op_; + vector pending_ops_; + + string name; + Place place; + size_t version; +}; + +struct OpHandleBase { + vector inputs_; + vector outputs_; +}; + +struct SSAGraph { + // vars on each devices. + // * the vars in each map in vector is on different device. + // * the map is mapping a variable name to variable handles + // with different versions + vector>> vars_; + + // All ops + vector ops_; +}; +``` +The variable handles are the wrapper of `Variables`. The operator handles are the wrapper of `OperatorBase`. Some `OpHandle` is not an `OperatorBase`, such as `NCCLAllReduceOpHandle`, because `AllReduceOpHandle` will use new device contexts. + +When the `ProgramDesc` converted into an `SSA` Graph, the [data hazard](https://en.wikipedia.org/wiki/Hazard_(computer_architecture)) problem is also need to be taken care. The dummy variables, which represent the dependency between operators, will be manually inserted into SSA graph to resolve the [data hazard](https://en.wikipedia.org/wiki/Hazard_(computer_architecture)) problem. + +## Execute SSA Graph + +The SSA graph can be out-of-order executed by an approximate [topological sorting](https://en.wikipedia.org/wiki/Topological_sorting) algorithm. The algorithm is + +1. Maintaining a map of an operator and its needed input number. +2. If a variable is not generated by an operator, i.e., `var.generated_op == nullptr`, decrease the needed input number of its pending operators. +3. If there is an operator which needed input number is decreased to zero, just run this operator. +4. After run this operator, just mark the variables are generated and repeat step 2 until all variables are generated. + +Running an operator can be asynchronized. There is a thread pool to execute an `SSA` graph. + +## Synchronize GPU Kernels + +The GPU is a non-blocking device. The different streams need be synchronized when switing streams. In current implementation, the synchronization based on the following algorithm: + +1. `OpHandle` will record `DeviceContext` that it is used. +2. In `OpHandle::Run`, if the `DeviceContext` of current operator is different from `DeviceContext` of any input variable, just wait the generate operator of this input variable. + +The `wait` are implemented by two strategies: + +1. Invoke `DeviceContext->Wait()`, It will wait all operators on this device contexts complete. +2. Uses `cudaStreamWaitEvent` to sending a event to the stream. It is a non-blocking call. The wait operators will be executed in GPU. + +Generally, the `cudaStreamWaitEvent` will have a better perforamnce. However, `DeviceContext->Wait()` strategy is easier to debug. The strategy can be changed in runtime. + +## What's next? + +* Merging gradient of dense parameters has been done. However, the merging of sparse parameters has not been done. +* The CPU version of Parallel Executor has not been implemented. The out-of-order logic will make CPU compuatation faster, too. +* A better strategy to merge gradients can be introduced. We can shrink the gradients from `float32` to `int8` or `int4` while merging. It will significantly speed up multi-GPUs training without much loss of precision. +* Combine multi-Nodes implementation. By the benifit of out-of-order, sending and recving operator can be an blocking operator, and the transpiler does not need to concern about the best position of operator. diff --git a/doc/fluid/CMakeLists.txt b/doc/fluid/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..9fe79323ef9377a459d8405cfa74c88c52ce9346 --- /dev/null +++ b/doc/fluid/CMakeLists.txt @@ -0,0 +1,55 @@ +if(NOT DEFINED SPHINX_THEME) + set(SPHINX_THEME default) +endif() + +if(NOT DEFINED SPHINX_THEME_DIR) + set(SPHINX_THEME_DIR) +endif() + +# configured documentation tools and intermediate build results +set(BINARY_BUILD_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_build") + +# Sphinx cache with pickled ReST documents +set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees") + +# HTML output director +set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html") + +configure_file( + "${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.en.in" + "${BINARY_BUILD_DIR_EN}/conf.py" + @ONLY) + +sphinx_add_target(paddle_fluid_docs + html + ${BINARY_BUILD_DIR_EN} + ${SPHINX_CACHE_DIR_EN} + ${CMAKE_CURRENT_SOURCE_DIR} + ${SPHINX_HTML_DIR_EN}) + +add_dependencies(paddle_fluid_docs gen_proto_py) + +# configured documentation tools and intermediate build results +set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build") + +# Sphinx cache with pickled ReST documents +set(SPHINX_CACHE_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_doctrees") + +# HTML output directory +set(SPHINX_HTML_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/html") + +configure_file( + "${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.cn.in" + "${BINARY_BUILD_DIR_CN}/conf.py" + @ONLY) + +sphinx_add_target(paddle_fluid_docs_cn + html + ${BINARY_BUILD_DIR_CN} + ${SPHINX_CACHE_DIR_CN} + ${CMAKE_CURRENT_SOURCE_DIR} + ${SPHINX_HTML_DIR_CN}) + +add_dependencies(paddle_fluid_docs_cn gen_proto_py) + +add_subdirectory(api) diff --git a/doc/fluid/api/CMakeLists.txt b/doc/fluid/api/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..ca40dfb9644cea69329be0ec231378506c138bc0 --- /dev/null +++ b/doc/fluid/api/CMakeLists.txt @@ -0,0 +1,22 @@ +# configured documentation tools and intermediate build results +set(BINARY_BUILD_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_build") + +# Sphinx cache with pickled ReST documents +set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees") + +# HTML output director +set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html") + +configure_file( + "${CMAKE_CURRENT_SOURCE_DIR}/../../templates/conf.py.en.in" + "${BINARY_BUILD_DIR_EN}/conf.py" + @ONLY) + +sphinx_add_target(paddle_fluid_apis + html + ${BINARY_BUILD_DIR_EN} + ${SPHINX_CACHE_DIR_EN} + ${CMAKE_CURRENT_SOURCE_DIR} + ${SPHINX_HTML_DIR_EN}) + +add_dependencies(paddle_fluid_apis gen_proto_py framework_py_proto copy_paddle_pybind) diff --git a/doc/v2/api/fluid/data_feeder.rst b/doc/fluid/api/data_feeder.rst similarity index 100% rename from doc/v2/api/fluid/data_feeder.rst rename to doc/fluid/api/data_feeder.rst diff --git a/doc/v2/api/fluid/evaluator.rst b/doc/fluid/api/evaluator.rst similarity index 100% rename from doc/v2/api/fluid/evaluator.rst rename to doc/fluid/api/evaluator.rst diff --git a/doc/v2/api/fluid/executor.rst b/doc/fluid/api/executor.rst similarity index 100% rename from doc/v2/api/fluid/executor.rst rename to doc/fluid/api/executor.rst diff --git a/doc/v2/api/fluid/gen_doc.py b/doc/fluid/api/gen_doc.py similarity index 100% rename from doc/v2/api/fluid/gen_doc.py rename to doc/fluid/api/gen_doc.py diff --git a/doc/v2/api/fluid/gen_doc.sh b/doc/fluid/api/gen_doc.sh similarity index 100% rename from doc/v2/api/fluid/gen_doc.sh rename to doc/fluid/api/gen_doc.sh diff --git a/doc/v2/api/fluid/index.rst b/doc/fluid/api/index_en.rst similarity index 100% rename from doc/v2/api/fluid/index.rst rename to doc/fluid/api/index_en.rst diff --git a/doc/v2/api/fluid/initializer.rst b/doc/fluid/api/initializer.rst similarity index 100% rename from doc/v2/api/fluid/initializer.rst rename to doc/fluid/api/initializer.rst diff --git a/doc/v2/api/fluid/io.rst b/doc/fluid/api/io.rst similarity index 100% rename from doc/v2/api/fluid/io.rst rename to doc/fluid/api/io.rst diff --git a/doc/v2/api/fluid/layers.rst b/doc/fluid/api/layers.rst similarity index 100% rename from doc/v2/api/fluid/layers.rst rename to doc/fluid/api/layers.rst diff --git a/doc/v2/api/fluid/nets.rst b/doc/fluid/api/nets.rst similarity index 100% rename from doc/v2/api/fluid/nets.rst rename to doc/fluid/api/nets.rst diff --git a/doc/v2/api/fluid/optimizer.rst b/doc/fluid/api/optimizer.rst similarity index 100% rename from doc/v2/api/fluid/optimizer.rst rename to doc/fluid/api/optimizer.rst diff --git a/doc/v2/api/fluid/param_attr.rst b/doc/fluid/api/param_attr.rst similarity index 100% rename from doc/v2/api/fluid/param_attr.rst rename to doc/fluid/api/param_attr.rst diff --git a/doc/v2/api/fluid/profiler.rst b/doc/fluid/api/profiler.rst similarity index 100% rename from doc/v2/api/fluid/profiler.rst rename to doc/fluid/api/profiler.rst diff --git a/doc/v2/api/fluid/regularizer.rst b/doc/fluid/api/regularizer.rst similarity index 100% rename from doc/v2/api/fluid/regularizer.rst rename to doc/fluid/api/regularizer.rst diff --git a/doc/fluid/build_and_install/build_from_source_cn.rst b/doc/fluid/build_and_install/build_from_source_cn.rst new file mode 120000 index 0000000000000000000000000000000000000000..ae4e8c7c48e584ec16a7be5466f83dd154ffb5fb --- /dev/null +++ b/doc/fluid/build_and_install/build_from_source_cn.rst @@ -0,0 +1 @@ +../../v2/build_and_install/build_from_source_cn.rst \ No newline at end of file diff --git a/doc/fluid/build_and_install/build_from_source_en.rst b/doc/fluid/build_and_install/build_from_source_en.rst new file mode 120000 index 0000000000000000000000000000000000000000..1ac828c973826bb8374c4aa8e17fda3ea1bb939f --- /dev/null +++ b/doc/fluid/build_and_install/build_from_source_en.rst @@ -0,0 +1 @@ +../../v2/build_and_install/build_from_source_en.rst \ No newline at end of file diff --git a/doc/fluid/build_and_install/docker_install_cn.rst b/doc/fluid/build_and_install/docker_install_cn.rst new file mode 120000 index 0000000000000000000000000000000000000000..965b2e20559291989422938c418fadbac16941b9 --- /dev/null +++ b/doc/fluid/build_and_install/docker_install_cn.rst @@ -0,0 +1 @@ +../../v2/build_and_install/docker_install_cn.rst \ No newline at end of file diff --git a/doc/fluid/build_and_install/docker_install_en.rst b/doc/fluid/build_and_install/docker_install_en.rst new file mode 120000 index 0000000000000000000000000000000000000000..79d7341a7bbb9e477c773134f24983fd7607769a --- /dev/null +++ b/doc/fluid/build_and_install/docker_install_en.rst @@ -0,0 +1 @@ +../../v2/build_and_install/docker_install_en.rst \ No newline at end of file diff --git a/doc/fluid/build_and_install/index_cn.rst b/doc/fluid/build_and_install/index_cn.rst new file mode 120000 index 0000000000000000000000000000000000000000..f697fcd8fac9131862ae7f8f51c5ebe93737ad2d --- /dev/null +++ b/doc/fluid/build_and_install/index_cn.rst @@ -0,0 +1 @@ +../../v2/build_and_install/index_cn.rst \ No newline at end of file diff --git a/doc/fluid/build_and_install/index_en.rst b/doc/fluid/build_and_install/index_en.rst new file mode 120000 index 0000000000000000000000000000000000000000..502f66a41319d4f41ae1774628ca36da9dca76ce --- /dev/null +++ b/doc/fluid/build_and_install/index_en.rst @@ -0,0 +1 @@ +../../v2/build_and_install/index_en.rst \ No newline at end of file diff --git a/doc/fluid/build_and_install/pip_install_cn.rst b/doc/fluid/build_and_install/pip_install_cn.rst new file mode 120000 index 0000000000000000000000000000000000000000..07deca84b82ff553e0c19324695089dcfb6be90e --- /dev/null +++ b/doc/fluid/build_and_install/pip_install_cn.rst @@ -0,0 +1 @@ +../../v2/build_and_install/pip_install_cn.rst \ No newline at end of file diff --git a/doc/fluid/build_and_install/pip_install_en.rst b/doc/fluid/build_and_install/pip_install_en.rst new file mode 120000 index 0000000000000000000000000000000000000000..7f39c998195b719b05443e96f1c4a6a8d44b98c9 --- /dev/null +++ b/doc/fluid/build_and_install/pip_install_en.rst @@ -0,0 +1 @@ +../../v2/build_and_install/pip_install_en.rst \ No newline at end of file diff --git a/doc/fluid/design/algorithm/index_cn.rst b/doc/fluid/design/algorithm/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..0883a9dc9c457f393ac1bdc930cb47ebcb0a25d9 --- /dev/null +++ b/doc/fluid/design/algorithm/index_cn.rst @@ -0,0 +1,7 @@ +梯度更新算法 +------------ + +.. toctree:: + :maxdepth: 1 + + parameter_average.md diff --git a/doc/fluid/design/algorithm/index_en.rst b/doc/fluid/design/algorithm/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..59fe68dcf79ce2ef90b9adc829a0db45a4f0b3dc --- /dev/null +++ b/doc/fluid/design/algorithm/index_en.rst @@ -0,0 +1,7 @@ +Gradient Update Algorithm +-------------------------------------- + +.. toctree:: + :maxdepth: 1 + + parameter_average.md diff --git a/doc/fluid/design/concepts/README.md b/doc/fluid/design/concepts/README.md index bf0e4dddc1b640ecbce489f65820aaf8a4b3b1e7..ed3f5aab2882c16ca6ac1446b4c4d4d27a373af7 100644 --- a/doc/fluid/design/concepts/README.md +++ b/doc/fluid/design/concepts/README.md @@ -2,7 +2,7 @@ A few months ago when we were trying to replace CMake with Bazel, @emailweixu su Here are some initial thoughts. Your comments are welcome! -### Required CMake Function +# Required CMake Function I think we need only the following few CMake functions to make a project description mean and clean: @@ -25,7 +25,7 @@ Also, - to describe external dependencies, we need `external_library`. - to build shared libraries, we need `shared_library`. -### An Example Project +## An Example Project Suppose that we have aforementioned functions defined in our `/cmake` directory. The following example `CMakeLists.txt` describes a project including the following source files: @@ -102,11 +102,11 @@ shared_library(api ``` -### Implementation +## Implementation As above example CMakeLists.txt executes, each function invocation adds "nodes" to a dependency graph. It also use this graph to generate CMake commands including `add_executable`, `add_dependencies`, `target_link_libraries`, and `add_test`. -### Using Package Manager For Go +## Using Package Manager For Go Building Go binaries and libraries need to satisfy their dependencies, generally we can do `go get ./...` to download and compile all external dependencies. The @@ -122,7 +122,7 @@ problems are: at many cloud file hosting, so users what to compile paddle by themselves can download this "vendor" package from a mirror site. -#### Choose A Suitable Tool +### Choose A Suitable Tool As mentioned by @wangkuiyi, [Here](https://github.com/golang/go/wiki/PackageManagementTools) list dozens of Go package managers. We choose the tool using following principles: @@ -140,7 +140,7 @@ management tool has been started at: https://github.com/golang/dep to resolve such problems, but it's currently at Alpha stage. So the best choice now is glide obviously. -#### Manage Go Packages +### Manage Go Packages - Dependencies: `go/glide.yaml` will store the dependencies and their versions which is directly imported by paddle. `go/glide.lock` will store all dependencies recursively diff --git a/doc/fluid/design/concepts/cpp_data_feeding.md b/doc/fluid/design/concepts/cpp_data_feeding.md index 8607b40ccbbe01db77afed72c1efa780b520744c..aabc1ba75a67c5767d409bd6e7e6240dec86b16c 100644 --- a/doc/fluid/design/concepts/cpp_data_feeding.md +++ b/doc/fluid/design/concepts/cpp_data_feeding.md @@ -113,7 +113,7 @@ To solve this problem, we introduce `ReaderHolder` as a wrapper. It acts as an e To create and invoke readers, some new ops are introduced: -### CreateReaderOp +### Operators That Create Readers Each reader has its creation op. File readers' creation ops have no input and yield the created file reader as its output. Decorated readers' creation ops take the underlying readers as inputs and then yield new decorated readers. @@ -153,19 +153,52 @@ double_buffer_reader = create_double_buffer_op(batch_reader) The forwarding ops of the corresponding `main_program` would be like this: ``` -while_op { +not_completed = true +pass_count = 0 +while_op(not_completed) { has_next = has_next_op(double_buffer_reader) if_else_op(has_next) { batch_data = read_op(double_buffer_reader) ... (subsequent training ops) } else { reset_op(double_buffer_reader) + increase_op(pass_count) + not_completed = less_than_op(pass_count, reqiured_pass_num) } } ``` -Two important considerations for these programs are as follows: +A few important considerations for these programs are as follows: -1. The multiple\_reader is the batch\_reader's underlying reader, and the batch\_reader is the double\_buffer\_reader's underlying reader. `read_op`, `has_next_op` and other reader related ops will only invoke the top-most reader. In this case, it's the double\_buffer\_reader. +1. `not_completed`, `pass_count` and other variables shown above are all Fluid Variables. -2. All readers exist in both `startup_program` and `main_program`. And they are persistable. +2. The multiple\_reader is the batch\_reader's underlying reader, and the batch\_reader is the double\_buffer\_reader's underlying reader. `read_op`, `has_next_op` and other reader related ops will only invoke the top-most reader. In this case, it's the double\_buffer\_reader. + +3. All readers exist in both `startup_program` and `main_program`. And they are persistable. + +### Simplify Configuration by MultiPassReader + +The Program configuration mentioned above is complicated. Users need to be very familiar to concepts of Program and Block to prevent making mistakes in their code. To make the usage of C++ readers more friendly to new users, we introduce `MultiPassReader`. + +`MultiPassReader` is a decorated reader. A multi-pass reader is used to continuously yield data for several training passes. It takes the number of passes to run as one of its attributes('pass_num') and maintains a counter to record how many passes it has completed. Each time its underlying reader reaches the EOF, the multi-pass reader checks whether it has completed the training of given number of pass. If not, the underlying reader will be re-initialized and starts a new pass automatically. Before completing the whole training, the return of MultiPassReader's `HasNext()` will always be `true`. + +With `MultiPassReader`, the startup program would be like this: + +``` +multiple_reader = open_files_op(...) +batch_reader = create_batch_reader_op(multiple_reader) +multi_pass_reader = create_multi_pass_reader_op(batch_reader) +double_buffer_reader = create_double_buffer_op(multi_pass_reader) +... (other initializers) +``` + +The forwarding part of the corresponding `main_program` would be like this: + +``` +not_completed = true +while_op(not_completed) { + batch_data = read_op(double_buffer_reader) + ... (subsequent training ops) + not_completed = has_next_op(double_buffer_reader) +} +``` diff --git a/doc/fluid/design/concepts/index_cn.rst b/doc/fluid/design/concepts/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..eec8a2f14ca9e8b3bf0d0acbbb6004972790d795 --- /dev/null +++ b/doc/fluid/design/concepts/index_cn.rst @@ -0,0 +1,18 @@ +核心概念 +------------- + +.. toctree:: + :maxdepth: 1 + + README.md + cpp_data_feeding.md + functions_operators_layers.md + program.md + variable.md + var_desc.md + tensor.md + tensor_array.md + lod_tensor.md + block.md + scope.md + executor.md diff --git a/doc/fluid/design/concepts/index_en.rst b/doc/fluid/design/concepts/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..036e1da2550cf520f5c40ecd9657f71603755adc --- /dev/null +++ b/doc/fluid/design/concepts/index_en.rst @@ -0,0 +1,18 @@ +Core Concepts +-------------------------------------- + +.. toctree:: + :maxdepth: 1 + + README.md + cpp_data_feeding.md + functions_operators_layers.md + program.md + variable.md + var_desc.md + tensor.md + tensor_array.md + lod_tensor.md + block.md + scope.md + executor.md diff --git a/doc/fluid/design/concepts/scope.md b/doc/fluid/design/concepts/scope.md index 4da76eebb74abcd26ec2b8671399e6bc4fb58574..dcf76649357aaef80d6bc1a933ece8c4c1063547 100644 --- a/doc/fluid/design/concepts/scope.md +++ b/doc/fluid/design/concepts/scope.md @@ -30,7 +30,7 @@ Scope is an association of a name to variable. All variables belong to `Scope`. Variable can not belong to many scopes. If you want to use variables from parent scope, you can use `parent scope`. -1. Scope should destruct all Variables inside it when itself is destructed. User can never store `Variable` pointer somewhere else. +1. Scope should destruct all Variables inside it when itself is destructed. User can never store `Variable` pointer somewhere else. Because Variable can only be got from Scope. When destroying Scope, we also need to destroy all the Variables in it. If user store `Variable` pointer to private data member or some global variable, the pointer will be an invalid pointer when associated `Scope` is destroyed. @@ -78,7 +78,7 @@ In `Scope` class, there is a private data member called `parent_`. `parent_` is A local scope is very useful when we implement Recurrent Neural Network. Each timestep of an RNN should be a `Net`. Each `Net` of timestep (`StepNet` for short) should use an independent local scope. Just like variables in a while loop is inside a local scope in programming languages. By using a single `StepNet` and changing local scope, we can implement an RNN easily. -# Interface Design +## Interface Design ```cpp class Variable { diff --git a/doc/fluid/design/concepts/var_desc.md b/doc/fluid/design/concepts/var_desc.md index 6a45af1995463402ba9c65ddb51c6c8bb107f99e..fcba08c07f40177d54a91048cb616198402a9d5d 100644 --- a/doc/fluid/design/concepts/var_desc.md +++ b/doc/fluid/design/concepts/var_desc.md @@ -1,3 +1,5 @@ +# Design Doc: Var_desc + ## Background PaddlePaddle divides the description of neural network computation into two stages: compile time and runtime. At compile time, the neural network computation is described as a `ProgramDesc` whereas at runtime an `Executor` interprets the `ProgramDesc` to compute the operations. diff --git a/doc/fluid/design/concurrent/channel.md b/doc/fluid/design/concurrent/channel.md new file mode 100644 index 0000000000000000000000000000000000000000..a00a3325e7b49381f0f82ebbf32b74683f02de5f --- /dev/null +++ b/doc/fluid/design/concurrent/channel.md @@ -0,0 +1,139 @@ +# Channel Design + +## Introduction + +A Channel is a data structure that allows for synchronous interprocess +communication via message passing. It is a fundemental component of CSP +(communicating sequential processes), and allows for users to pass data +between threads without having to worry about synchronization. + +## How to use it + +Paddle offers python APIs to open and close channels, along with sending +and receiving data to/from a channel. + +### Create a channel + +Creates a new channel that takes in variables of a specific dtype. + +- **fluid.make_channel(dtype, capacity=0)** + - **dtype**: The data type of variables being sent/received through channel + - **capacity**: The capacity of the channel. A capacity of 0 represents + an unbuffered channel. Capacity > 0 represents a buffered channel + +``` +ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR, 10) +``` + +### Close a channel + +Closes a channel. Any pending senders and receivers will be awoken during +this time. Receivers can still receive from a closed channel, but senders +are not allowed to send any additional data to the channel (Paddle will +raise an exception if users try to send to a closed channel.) + +- **fluid.channel_close(channel)** + +``` +fluid.channel_close(ch) +``` + +### Send data to a channel + +Sends a variable to a channel. Currently, variables of dtype `LoDTensor`, +`LoDRankTable`, `LoDTensorArray`, `SelectedRows`, `ReaderHolder`, and +`ChannelHolder` are supported. + +By default, the data of the Variable is moved from the sender to the receiver, +however the user can optionally copy the data before performing the send. + +- **channel_send(channel, variable, is_copy=False)** + - **channel**: The channel to send the variable to + - **variable**: The variable to send to the channel + - **is_copy**: If set to True, channel_send will perform a variable assign + to copy the source variable to a new variable to be sent. + +``` +ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) +var = fill_constant(shape=[1],dtype=core.VarDesc.VarType.INT32, value=100) +fluid.channel_send(ch, var, True) +``` + +### Receive data from a channel + +Receives a variable from a channel. The data of the variable is moved to the +receiving variable. + +- **channel_recv(channel, return_variable)** + - **channel**: The channel to receive the variable from + - **return_variable**: The destination variable used to store the data of the + variable received from the channel + +``` +ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) +var = fill_constant(shape=[1],dtype=core.VarDesc.VarType.INT32, value=-1) +fluid.channel_recv(ch, var) +``` + +## How it Works + +Channels provides a simple interface for different threads to share data. +To support the synchronization requirements, channels utilizes a series of +internal queues, locks, and conditional variables. + +### QueueMessage + +QueueMessage encapsulates the state of the channel send/receive operation to be +put in the **sendq/recvq**. It contains a condition variable used to lock the +thread (when there are no available sends/receives). In addition, it contains +a callback function to notify a thread when the QueueMessage is being +processed by the channel. + +### Queues + +- **buff_**: This queue holds the data buffer in a buffered channel. The +capacity is set to the capacity of the channel. This data buffer is not +used in an unbuffered channel. + +- **sendq**: This queue holds the QueueMessage of any pending senders of a +channel. When a thread performs a channel_send operation on the channel, the +channel_send operation will put a new QueueMessage on the sendq and block the +current thread under two conditions: + 1. The channel is buffered and is full + 2. The channel is unbuffered and does not have a receiver + +- **recvq**: This queue holds the QueueMessage of any pending receivers of a +channel. When a thread performs a channel_recv operation on the channel, the +channel_recv operation will put a new QueueMessage on the recvq and block the +current thread under two conditions: + 1. The channel is buffered and there is no data on the buff_ + 2. The channel is unbuffered and does not have a sender + +### State diagram + +#### Channel Send + +

+
+

+ +#### Channel Receive + +

+
+

+ +## Limitations and Considerations + +### Variable Copy + +In golang, variables in channels are copied from the sender to the receiver. +In Paddle, the data from our variables are **moved** from sender to receiver. +As a result, these variables should not be used after they are sent. We +provide a flag in channel_send method to allow users to copy the variable to +be sent before it is sent. + +Please note that this is acheived by adding an **assign** operator and creating +a temporary variable that is sent in place of the original variable. Please +note that **assign** operator has limited support for only certain variables +datatypes. diff --git a/doc/fluid/design/concurrent/go_op.md b/doc/fluid/design/concurrent/go_op.md new file mode 100644 index 0000000000000000000000000000000000000000..c18b788e80f432ebb2f14b15229e7823c112001e --- /dev/null +++ b/doc/fluid/design/concurrent/go_op.md @@ -0,0 +1,231 @@ +# go_op Design + +## Introduction + +The **go_op** allows user's of PaddlePaddle to run program blocks on a detached +thread. It works in conjuction with CSP operators (channel_send, +channel_receive, channel_open, channel_close, and select) to allow users to +concurrently process data and communicate easily between different threads. + +## How to use it + +``` +channel = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR) + +with fluid.Go(): + # Send a tensor of value 99 to "channel" on a detached thread + tensor = fill_constant(shape=[1], dtype='int', value=99) + tensor.stop_gradient = True + fluid.channel_send(channel, tensor) + +# Receive sent tensor from "channel" on the main thread +result = fill_constant(shape=[1], dtype='int', value=-1) +fluid.channel_recv(ch, result) +``` + +The go operator can be accessed by using the fluid.Go() control flow. This +will create a new sub block, where the user can add additional operators +to be ran on the thread. + +**Note:** Since back propegation is currently not support in the go_op, users +should ensure that operators in the go block does not require gradient +calculations. + +## How it Works + +Similar to other control blocks, go_op will create a sub block and add it +as a child to the current block. Operators and variables defined in this +block will be added to the go sub_block. + +In addition, the go operator will create a new child scope whose parent is +the global scope. Please refer to [block captures](#block-captures) for more +information. + +When Paddle executor runs go_op, go_op will take the sub_block and pass it to +the executor.run method (along with a newly created local scope) on a detached +thread. + +An example of the generated program description is shown below. Take note of +the **go_op** in particular. It is added as an operator in the current +block (in this example, block0). The **go_op** contains a `sub_block` +attribute, which points to the id of the block that will be executed in a +detached thread. + +``` +blocks { + idx: 0 + parent_idx: -1 + vars { + name: "return_value" + type { + type: LOD_TENSOR + lod_tensor { + tensor { + data_type: INT64 + } + } + } + } + vars { + name: "status_recv" + type { + type: LOD_TENSOR + lod_tensor { + tensor { + data_type: BOOL + } + } + } + } + ... + ops { + outputs { + parameter: "Out" + arguments: "channel" + } + type: "channel_create" + attrs { + name: "data_type" + type: INT + i: 7 + } + attrs { + name: "capacity" + type: INT + i: 0 + } + } + ops { + inputs { + parameter: "X" + arguments: "channel" + } + type: "go" + attrs { + name: "sub_block" + type: BLOCK + block_idx: 1 + } + } + ops { + inputs { + parameter: "Channel" + arguments: "channel" + } + outputs { + parameter: "Out" + arguments: "return_value" + } + outputs { + parameter: "Status" + arguments: "status_recv" + } + type: "channel_recv" + } + ... +} + +blocks { + idx: 1 + parent_idx: 0 + vars { + name: "status" + type { + type: LOD_TENSOR + lod_tensor { + tensor { + data_type: BOOL + } + } + } + } + ... + + ops { + outputs { + parameter: "Out" + arguments: "fill_constant_1.tmp_0" + } + type: "fill_constant" + attrs { + name: "force_cpu" + type: BOOLEAN + b: false + } + attrs { + name: "value" + type: FLOAT + f: 99.0 + } + attrs { + name: "shape" + type: INTS + ints: 1 + } + attrs { + name: "dtype" + type: INT + i: 3 + } + } + ops { + inputs { + parameter: "Channel" + arguments: "channel" + } + inputs { + parameter: "X" + arguments: "fill_constant_1.tmp_0" + } + outputs { + parameter: "Status" + arguments: "status" + } + type: "channel_send" + attrs { + name: "copy" + type: BOOLEAN + b: false + } + } +``` + +## Current Limitations + +#### Scopes and block captures: + +Paddle utilizes [scopes](./../concepts/scope.md) to store variables used in a +block. When a block is executed, a new local scope is created from the parent +scope (ie: scope derived from the parent block) and associated with the new +child block. After the block finishes executing, then the local scope and +all associated variables in the scope is deleted. + +This works well in a single threaded scenario, however with introduction of +go_op, a child block may continue to execute even after the parent block has +exited. If the go_op tries to access variables located in the parent block's +scope, it may receive a segmentation fault because the parent scope may have +been deleted. + +We need to implement block closures in order to prevent access to parent +scope variables from causing a segmentation fault. As a temporary workaround, +please ensure that all variables accessed in the go block is not destructed +before it is being accessed. Currently, the go_op will explicitly enforce +this requirement and raise an exception if a variable could not be found in +the scope. + +Please refer to [Closure issue](https://github.com/PaddlePaddle/Paddle/issues/8502) +for more details. + +#### Green Threads + +Golang utilizes `green threads`, which is a mechnism for the runtime library to +manage multiple threads (instead of natively by the OS). Green threads usually +allows for faster thread creation and switching, as there is less overhead +when spawning these threads. For the first version of CSP, we only support +OS threads. + + +#### Backward Propegation: + +go_op currently does not support backwards propagation. Please use go_op with +non training operators. diff --git a/doc/fluid/design/concurrent/images/channel_recv.png b/doc/fluid/design/concurrent/images/channel_recv.png new file mode 100644 index 0000000000000000000000000000000000000000..c06cd15ae7b8a8c94d5742f6675e389081fcf789 Binary files /dev/null and b/doc/fluid/design/concurrent/images/channel_recv.png differ diff --git a/doc/fluid/design/concurrent/images/channel_send.png b/doc/fluid/design/concurrent/images/channel_send.png new file mode 100644 index 0000000000000000000000000000000000000000..006ebb4a5a4bcd32c97847e9fb7729a740255f7c Binary files /dev/null and b/doc/fluid/design/concurrent/images/channel_send.png differ diff --git a/doc/fluid/design/concurrent/index_cn.rst b/doc/fluid/design/concurrent/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..e47135e9fc42760898083710e0a6767252a0225b --- /dev/null +++ b/doc/fluid/design/concurrent/index_cn.rst @@ -0,0 +1,8 @@ +并发编程 +------------ + +.. toctree:: + :maxdepth: 1 + + concurrent_programming.md + parallel_do.md diff --git a/doc/fluid/design/concurrent/index_en.rst b/doc/fluid/design/concurrent/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..0727e75798b2a869588f80d3cce7a886554e4ffb --- /dev/null +++ b/doc/fluid/design/concurrent/index_en.rst @@ -0,0 +1,8 @@ +Concurrent Programming +------------------------- + +.. toctree:: + :maxdepth: 1 + + concurrent_programming.md + parallel_do.md diff --git a/doc/fluid/design/data_type/index_cn.rst b/doc/fluid/design/data_type/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..b60167b6b1599df69dfc5073ebf32bdbb0a316ec --- /dev/null +++ b/doc/fluid/design/data_type/index_cn.rst @@ -0,0 +1,7 @@ +数据类型 +------------ + +.. toctree:: + :maxdepth: 1 + + float16.md diff --git a/doc/fluid/design/data_type/index_en.rst b/doc/fluid/design/data_type/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..6a88d17943f49134a2d00363845e919537ff4545 --- /dev/null +++ b/doc/fluid/design/data_type/index_en.rst @@ -0,0 +1,7 @@ +Data Type +------------ + +.. toctree:: + :maxdepth: 1 + + float16.md diff --git a/doc/fluid/design/dist_train/distributed_lookup_table_design.md b/doc/fluid/design/dist_train/distributed_lookup_table_design.md index e543adf0f97cc6b47415b807d7a1ed1effec9b22..988729138926f035750b59eb245dde82502a3ad2 100644 --- a/doc/fluid/design/dist_train/distributed_lookup_table_design.md +++ b/doc/fluid/design/dist_train/distributed_lookup_table_design.md @@ -1,4 +1,4 @@ -## Design Doc: Distributed Lookup Table Operator +# Design Doc: Distributed Lookup Table Operator A lookup table operator in PaddlePaddle where the table could be out of the memory of a computer. diff --git a/doc/fluid/design/dist_train/index_cn.rst b/doc/fluid/design/dist_train/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..ed6f3dda271d2de58d92aa7ec804fa9e68dfc48a --- /dev/null +++ b/doc/fluid/design/dist_train/index_cn.rst @@ -0,0 +1,9 @@ +分布式训练 +------------ + +.. toctree:: + :maxdepth: 1 + + distributed_architecture.md + distributed_lookup_table_design.md + parameter_server.md diff --git a/doc/fluid/design/dist_train/index_en.rst b/doc/fluid/design/dist_train/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..f84688f168021113bd933802709bcd787b474bca --- /dev/null +++ b/doc/fluid/design/dist_train/index_en.rst @@ -0,0 +1,9 @@ +Distributed Training +--------------------- + +.. toctree:: + :maxdepth: 1 + + distributed_architecture.md + distributed_lookup_table_design.md + parameter_server.md diff --git a/doc/fluid/design/dynamic_rnn/index_cn.rst b/doc/fluid/design/dynamic_rnn/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..1d224d22cf7103616f44115db01f0ae55f1cb88a --- /dev/null +++ b/doc/fluid/design/dynamic_rnn/index_cn.rst @@ -0,0 +1,8 @@ +动态RNN +------------ + +.. toctree:: + :maxdepth: 1 + + rnn.md + rnn_design.md diff --git a/doc/fluid/design/dynamic_rnn/index_en.rst b/doc/fluid/design/dynamic_rnn/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..568f496e4ffe21a5e730488aef905f7e2d98839e --- /dev/null +++ b/doc/fluid/design/dynamic_rnn/index_en.rst @@ -0,0 +1,8 @@ +Dynamic RNN +------------ + +.. toctree:: + :maxdepth: 1 + + rnn.md + rnn_design.md diff --git a/doc/fluid/design/dynamic_rnn/rnn_design.md b/doc/fluid/design/dynamic_rnn/rnn_design.md index 3d38b9a0ad225fd8e0c1bb037474b292b1887f5b..cecfcd3307ae4c4fa603220a360e9e124069fa58 100644 --- a/doc/fluid/design/dynamic_rnn/rnn_design.md +++ b/doc/fluid/design/dynamic_rnn/rnn_design.md @@ -99,7 +99,7 @@ private: - 由于传递过程是以复制`shared_ptr`的方式实现，因此框架只需要传递一次 `lod_start_pos` 2. 对于不感知 `lod_start_pos` 的Op足够透明 -3. 需要修改 `lod_start_pos` 的producer Op可以在 `Run` 时更新自己的 `lod_start_pos` 数据 +3. 需要修改 `lod_start_pos` 的producer Op可以在 `Run` 时更新自己的 `lod_start_pos` 数据 具体的设计分为以下3小节 @@ -189,7 +189,7 @@ struct SortedSeqItem { std::vector sorted_seqs; ``` -来追踪序列排序后的位置，并添加一个新的接口 +来追踪序列排序后的位置，并添加一个新的接口 ```c++ std::vector SortBySeqLen(const LODTensor& tensor); @@ -233,7 +233,10 @@ x x - 将每个序列concat 为规则的mini-batch表示 ## 参考文献 -1. [Tensorflow Bucketing](https://www.tensorflow.org/versions/r0.12/api_docs/python/contrib.training/bucketing) -2. [mxnet Bucketing](http://mxnet.io/how_to/bucketing.html) -3. [variable length input in RNN scenario](https://discuss.pytorch.org/t/about-the-variable-length-input-in-rnn-scenario/345/5) -4. [Level of details](https://en.wikipedia.org/wiki/Level_of_detail) +[Tensorflow Bucketing](https://www.tensorflow.org/versions/r0.12/api_docs/python/contrib.training/bucketing) + +[mxnet Bucketing](http://mxnet.io/how_to/bucketing.html) + +[variable length input in RNN scenario](https://discuss.pytorch.org/t/about-the-variable-length-input-in-rnn-scenario/345/5) + +[Level of details](https://en.wikipedia.org/wiki/Level_of_detail) diff --git a/doc/fluid/design/execution/index_cn.rst b/doc/fluid/design/execution/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..ed31b017429d168b2466d8f6b423f48bd5d78d1f --- /dev/null +++ b/doc/fluid/design/execution/index_cn.rst @@ -0,0 +1,8 @@ +执行流程 +------------- + +.. toctree:: + :maxdepth: 1 + + switch.md + if_else_op.md diff --git a/doc/fluid/design/execution/index_en.rst b/doc/fluid/design/execution/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..fcf846da348ff0bed707c42718e08314998fbac0 --- /dev/null +++ b/doc/fluid/design/execution/index_en.rst @@ -0,0 +1,8 @@ +Execution Process +-------------------------------------- + +.. toctree:: + :maxdepth: 1 + + switch.md + if_else_op.md diff --git a/doc/fluid/design/execution/switch.md b/doc/fluid/design/execution/switch.md index 827d0601c621e4a230de28e2baad8e196e69625e..1c337bd7159b25e594c2f91f9a143b3f4bc3c8e8 100644 --- a/doc/fluid/design/execution/switch.md +++ b/doc/fluid/design/execution/switch.md @@ -1,6 +1,6 @@ -### Design Doc: Switch +# Design Doc: Switch -### Background +## Background Many programming languages provide `switch` as a generalization of `if-elif-else`. We want to add it to Fluid. @@ -19,7 +19,7 @@ with switch() as switch: fluid.print("Case 3") ``` -### The Semantics +## The Semantics 1. A `switch` control-flow checks cases one-by-one. 1. The condition of each case is a boolean value, which is a scalar, and differs from the `fluid.if_else` control-flow, which condition could be a vector of boolean values. diff --git a/doc/fluid/design/index_cn.rst b/doc/fluid/design/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..e9f55214f411abb11bef180d7af4716ad85a0b09 --- /dev/null +++ b/doc/fluid/design/index_cn.rst @@ -0,0 +1,19 @@ +设计思想 +------------ + +.. toctree:: + :maxdepth: 1 + + motivation/index_cn.rst + execution/index_cn.rst + concepts/index_cn.rst + data_type/index_cn.rst + memory/index_cn.rst + muti_devices/index_cn.rst + dynamic_rnn/index_cn.rst + concurrent/index_cn.rst + algorithm/index_cn.rst + network/index_cn.rst + modules/index_cn.rst + interface/index_cn.rst + dist_train/index_cn.rst diff --git a/doc/fluid/design/index_en.rst b/doc/fluid/design/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..2802dc3a31d540c5a19bf9042053496aad152f98 --- /dev/null +++ b/doc/fluid/design/index_en.rst @@ -0,0 +1,19 @@ +Design +------------ + +.. toctree:: + :maxdepth: 1 + + motivation/index_en.rst + execution/index_en.rst + concepts/index_en.rst + data_type/index_en.rst + memory/index_en.rst + muti_devices/index_en.rst + dynamic_rnn/index_en.rst + concurrent/index_en.rst + algorithm/index_en.rst + network/index_en.rst + modules/index_en.rst + interface/index_en.rst + dist_train/index_en.rst diff --git a/doc/fluid/design/interface/index_cn.rst b/doc/fluid/design/interface/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..69a8d9bad4fe88935b9fa87757abf0105ca8eb75 --- /dev/null +++ b/doc/fluid/design/interface/index_cn.rst @@ -0,0 +1,4 @@ +多语言接口 +------------ + +TBD diff --git a/doc/fluid/design/interface/index_en.rst b/doc/fluid/design/interface/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..22abc71f984aa5da7151d5ebf0c3bdbcc69a3624 --- /dev/null +++ b/doc/fluid/design/interface/index_en.rst @@ -0,0 +1,4 @@ +Multi-Language Interface +----------------------- + +TBD diff --git a/doc/fluid/design/memory/index_cn.rst b/doc/fluid/design/memory/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..c507c638bd1a6eb428175ed2756a6ecfc6cca198 --- /dev/null +++ b/doc/fluid/design/memory/index_cn.rst @@ -0,0 +1,7 @@ +内存管理 +------------ + +.. toctree:: + :maxdepth: 1 + + memory_optimization.md diff --git a/doc/fluid/design/memory/index_en.rst b/doc/fluid/design/memory/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..f7526437a73a09b300f05e138084755f5528b242 --- /dev/null +++ b/doc/fluid/design/memory/index_en.rst @@ -0,0 +1,7 @@ +Memory Management +------------------- + +.. toctree:: + :maxdepth: 1 + + memory_optimization.md diff --git a/doc/fluid/design/modules/evaluator.md b/doc/fluid/design/modules/evaluator.md index 11cc129d56905a9ee666da92fbe6f8559c6d325a..de9605b0e67a035ab1ef1e4cafbe838f83bc5807 100644 --- a/doc/fluid/design/modules/evaluator.md +++ b/doc/fluid/design/modules/evaluator.md @@ -1,10 +1,10 @@ -## Evaluator Design +# Evaluator Design -### Problem Statement +## Problem Statement During training or inference, we provide an evaluation function to measure the model performance, for example, accuracy, precision, etc. In the operator based framework design, the data passes through the network pipeline batch by batch. As a result, inside the operator, we only calculate the metrics for one minibatch. Thus, we need to provide a mechanism to calculate the metrics for each N pass/batch the user wants. -### Evaluator Design +## Evaluator Design Currently, every operation is expressed in the graph. We divide the evaluator process into three steps. 1. Initialize the metric state and add it into the block. @@ -14,11 +14,11 @@ Currently, every operation is expressed in the graph. We divide the evaluator pr 3. Merge the mini-batch statistics to form the evaluation result for multiple mini-batches. When it comes to distributed training/Multi-GPU training, aggregate the value from different devices. -### Implementation -This design is shown in the Python API. -Each metric operator needs to caculate the metric statistic and return the batch-aware states. Python side is responsible for accumulating the states for each pass. +## Implementation +This design is shown in the Python API. +Each metric operator needs to caculate the metric statistic and return the batch-aware states. Python side is responsible for accumulating the states for each pass. + - ```python class Evaluator(object): """ @@ -32,7 +32,7 @@ class Evaluator(object): The initialization of Evaluator should be responsible for: create metric states and append to the main_program - """ + """ pass def _update_ops(self, input, label, **kwargs) @@ -40,14 +40,14 @@ class Evaluator(object): Add mini-batch evaluator caculate operators to the main_program. Add increment operator to accumulate the metric states. """ - + def reset(self, executor, reset_program=None): """ Reset metric states at the begin of each pass/user specified batch number. Execute the reset_program to reset the states. """ - + def eval(self, executor, eval_program=None): """ diff --git a/doc/fluid/design/modules/index_cn.rst b/doc/fluid/design/modules/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..b25783f0f5120991c29ba31b7b512bd4c183eecf --- /dev/null +++ b/doc/fluid/design/modules/index_cn.rst @@ -0,0 +1,14 @@ +代码结构和重要模块 +----------------- + +.. toctree:: + :maxdepth: 1 + + backward.md + python_api.md + regularization.md + infer_var_type.md + optimizer.md + prune.md + register_grad_op.md + net_op_design.md diff --git a/doc/fluid/design/modules/index_en.rst b/doc/fluid/design/modules/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..2108156e080996916f2650448f0a56f998757204 --- /dev/null +++ b/doc/fluid/design/modules/index_en.rst @@ -0,0 +1,14 @@ +Code Structure and Important Modules +------------------------------------- + +.. toctree:: + :maxdepth: 1 + + backward.md + python_api.md + regularization.md + infer_var_type.md + optimizer.md + prune.md + register_grad_op.md + net_op_design.md diff --git a/doc/fluid/design/modules/net_op_design.md b/doc/fluid/design/modules/net_op_design.md index a5f0483081e8a03b2d001a551fcc02bbd392016d..e64ac2fb1c6898bfeb883250347da3d9a4757b97 100644 --- a/doc/fluid/design/modules/net_op_design.md +++ b/doc/fluid/design/modules/net_op_design.md @@ -1,16 +1,16 @@ # Network Design `Network` is the container and controller of a set of operators, -user can build a real network from a `NetDesc` which is a protobuf message +user can build a real network from a `NetDesc` which is a protobuf message and use `Network.Run()` to run all the operators in the network. -A network object knows all Operators belonging to this network. Variables, -which are inputs and outputs of these operators, +A network object knows all Operators belonging to this network. Variables, +which are inputs and outputs of these operators, are created and managed by a hierarchy of Scope objects. -# API +## API -## Net +### Net To make the `Network` extendable, a base class is defined like this ```c++ @@ -43,8 +43,8 @@ class Net { }; ``` -All network implementations should build networks from a protobuf message which -describes the structure of a real network; `Run` method should be implemented by +All network implementations should build networks from a protobuf message which +describes the structure of a real network; `Run` method should be implemented by all implementations to offer a universal method to forward or backward compute a network. `Net::Create` is a method of factory pattern and can be implemented like @@ -64,7 +64,7 @@ std::unique Net::Create(const NetDesc& def) { ``` Network is designed as the container of operators. to make it more extendable, -we decouple it from the related variable resources. +we decouple it from the related variable resources. `Run(Scope* scope)` takes the scope as a argument so that it can run in different scopes. @@ -80,7 +80,7 @@ if (net) { } ``` -## `PlainNet` as a simple implementation of `BaseNet` +### `PlainNet` as a simple implementation of `BaseNet` A very basic implementation is as follows. All it does is simply to run every operators in sequence. @@ -211,9 +211,9 @@ class NetBuilder final { } ``` -## Compatibility with RNN +### Compatibility with RNN -Benefitting from the decoupling of `PlainNet.Run` and `Scope`, `PlainNet` is compatible with future RNN design, +Benefitting from the decoupling of `PlainNet.Run` and `Scope`, `PlainNet` is compatible with future RNN design, for example we can implement a simple recurrent neural network as follows ```c++ diff --git a/doc/fluid/design/modules/optimizer.md b/doc/fluid/design/modules/optimizer.md index 691081c268b848811bf5ee6d6a41edfe0f47eec0..1c25fde9cafb322f789662077d3fc6cc1d64ce38 100644 --- a/doc/fluid/design/modules/optimizer.md +++ b/doc/fluid/design/modules/optimizer.md @@ -1,6 +1,6 @@ -## Optimizer Design +# Optimizer Design -### The Problem +## The Problem A PaddlePaddle program, or a block, is a sequence of operators operating variables. A training program needs to do three kinds of works: @@ -19,7 +19,7 @@ It's true that users should be able to create all these operators manually by ca In this design, we propose a high-level API that automatically derives the optimisation pass and operators from the forward pass. -### High-level Python API to describe the training process +## High-level Python API to describe the training process 1. User write code to describe the network: @@ -54,7 +54,7 @@ In this design, we propose a high-level API that automatically derives the optim sess.run(target= opt_op_list, ...) ``` -#### Optimizer Python interface: +### Optimizer Python interface: ```python class Optimizer(object): diff --git a/doc/fluid/design/motivation/index_cn.rst b/doc/fluid/design/motivation/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..7706e73eca644ed6db772fd77da947395313237f --- /dev/null +++ b/doc/fluid/design/motivation/index_cn.rst @@ -0,0 +1,10 @@ +设计动机和目标 +------------- + +.. toctree:: + :maxdepth: 1 + + api.md + refactorization.md + fluid.md + fluid_compiler.md diff --git a/doc/fluid/design/motivation/index_en.rst b/doc/fluid/design/motivation/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..10b64b257c604ced6b957d6d6018e8a363f00fac --- /dev/null +++ b/doc/fluid/design/motivation/index_en.rst @@ -0,0 +1,10 @@ +Design Motivations and Goals +-------------------------------------- + +.. toctree:: + :maxdepth: 1 + + api.md + refactorization.md + fluid.md + fluid_compiler.md diff --git a/doc/fluid/design/motivation/refactorization.md b/doc/fluid/design/motivation/refactorization.md index f93d6155e1764386b01d2f0df3f141ab75cd55d4..7c39fabcc6df76afdb6a77b4cbc2edf0bf3ef780 100644 --- a/doc/fluid/design/motivation/refactorization.md +++ b/doc/fluid/design/motivation/refactorization.md @@ -97,13 +97,13 @@ Compile Time -> IR -> Runtime --- -# Operator/OpWithKernel/OpKernel +## Operator/OpWithKernel/OpKernel ![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/49caf1fb70820fb4a6c217634317c9306f361f36/op_op_with_kern_class_diagram.dot) --- -# Operator +## Operator ![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/dd598e8f1976f5759f58af5e5ef94738a6b2e661/op.dot) * `Operator` is the fundamental building block of the user interface. @@ -113,7 +113,7 @@ Compile Time -> IR -> Runtime --- -# OpWithKernel/Kernel +## OpWithKernel/Kernel ![class_diagram](http://api.paddlepaddle.org/graphviz?dot=https://gist.githubusercontent.com/reyoung/53df507f6749762675dff3e7ce53372f/raw/9d7f4eba185cf41c8e2fbfb40ae21890dbddcd39/op_with_kernel.dot) @@ -124,7 +124,7 @@ Compile Time -> IR -> Runtime --- -# Why separate Kernel and Operator +## Why separate Kernel and Operator * Separate GPU and CPU code. * Make Paddle capable of running without GPU. @@ -132,7 +132,7 @@ Compile Time -> IR -> Runtime * For example, same multiplication op can have different implementations kernels such as FP16 kernel, FP32 kernel, MKL, eigen kernel. --- -# Libraries for Kernel development +## Libraries for Kernel development * `Eigen::Tensor` contains basic math and element-wise functions. * Note that `Eigen::Tensor` has broadcast implementation. @@ -143,16 +143,16 @@ Compile Time -> IR -> Runtime * Hand-writing `GPUKernel` and `CPU` code * Do not write in header (`.h`) files. CPU Kernel should be in cpp source (`.cc`) and GPU kernels should be in cuda (`.cu`) files. (GCC cannot compile GPU code.) --- -# Operator Registration +## Operator Registration -## Why is registration necessary? +### Why is registration necessary? We need a method to build mappings between Op type names and Op classes. -## How is registration implemented? +### How is registration implemented? Maintaining a map, whose key is the type name and the value is the corresponding Op constructor. --- -# The Registry Map +## The Registry Map ### `OpInfoMap` @@ -166,7 +166,7 @@ Maintaining a map, whose key is the type name and the value is the corresponding - **`checker`**: Used to check attributes. --- -# Related Concepts +## Related Concepts ### Op_Maker It's constructor takes `proto` and `checker`. They are completed during Op_Maker's construction. ([ScaleOpMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)) @@ -178,7 +178,7 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) ``` --- -# Registration Process +## Registration Process 1. Write an Op class and its gradient Op class, if required. 2. Write an Op maker class. In the constructor of this class, describe the inputs, outputs and attributes of the operator. 3. Invoke the macro `REGISTER_OP`. This macro will @@ -186,13 +186,13 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) 2. Using the completed `proto` and `checker`, it will add a new key-value pair to the `OpInfoMap` --- -# Backward Module (1/2) +## Backward Module (1/2) ### Create Backward Operator - Mapping from forward Op to backward Op ![backward](https://gist.githubusercontent.com/dzhwinter/a6fbd4623ee76c459f7f94591fd1abf0/raw/61026ab6e518e66bde66a889bc42557a1fccff33/backward.png) --- -# Backward Module (2/2) +## Backward Module (2/2) ### Build Backward Network - **Input**: a graph of forward operators - **Output**: a graph of backward operators @@ -205,7 +205,7 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) --- -# Scope, Variable, Tensor +## Scope, Variable, Tensor * `Tensor` is an n-dimension array with type. * Only dims and data pointers are stored in `Tensor`. @@ -218,8 +218,8 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) * `Scope` has a hierarchical structure. The local scope can get variables from its parent scope. --- -# Block (in design) -## the difference between original RNNOp and Block +## Block (in design) +### the difference between original RNNOp and Block - As an operator is more intuitive than `RNNOp`, - Offers a new interface `Eval(targets)` to deduce the minimal block to `Run`, - Fits the compile-time/ runtime separation design paradigm. @@ -227,7 +227,7 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) - When graph executes, a Block with `BlockDesc` is passed. It then creates `Op` and `Var` instances and then invokes `Run`. --- -# Milestone +## Milestone - Take Paddle/books as the main line, the requirement of the models motivates framework refactoring, - Model migration - Framework development gives **priority support** to model migration, for example, @@ -240,7 +240,7 @@ REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) - Accept imperfection, concentrate on solving the specific problem at the right price. --- -# Control the migration quality +## Control the migration quality - Compare the performance of migrated models with old ones. - Follow the google C++ style guide. - Build the automatic workflow of generating Python/C++ documentations. diff --git a/doc/fluid/design/muti_devices/index_cn.rst b/doc/fluid/design/muti_devices/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..1f8439e8623e1c1ae9a12c24d08079f0ec3d761f --- /dev/null +++ b/doc/fluid/design/muti_devices/index_cn.rst @@ -0,0 +1,9 @@ +多设备支持 +------------ + +.. toctree:: + :maxdepth: 1 + + operator_kernel_type.md + kernel_selection.md + kernel_hint_design.md diff --git a/doc/fluid/design/muti_devices/index_en.rst b/doc/fluid/design/muti_devices/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..819e9c5d77b2abf8da0e2ce6f494ea5174c1d0a2 --- /dev/null +++ b/doc/fluid/design/muti_devices/index_en.rst @@ -0,0 +1,9 @@ +Multi-Device Support +---------------------- + +.. toctree:: + :maxdepth: 1 + + operator_kernel_type.md + kernel_selection.md + kernel_hint_design.md diff --git a/doc/fluid/design/muti_devices/kernel_hint_design.md b/doc/fluid/design/muti_devices/kernel_hint_design.md index a54b7da045e1a362626ef066f9ebb56af2c3181a..728c8f0b964c02c1efa019945f7427fa879d3aa1 100644 --- a/doc/fluid/design/muti_devices/kernel_hint_design.md +++ b/doc/fluid/design/muti_devices/kernel_hint_design.md @@ -1,4 +1,4 @@ -## Problem +# Problem In PaddlePaddle's [Design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md), one Operator may have multiple kernels. Users may have some personal preference to choose a certain type of kernel for an operator, such as `force_cpu` to choose a CPU kernel, `use_cudnn` to choose a CUDNN kernel, we need to provide a way for users to do this. In the current design, we use KernelType to describe one kernel. diff --git a/doc/fluid/design/muti_devices/kernel_selection.md b/doc/fluid/design/muti_devices/kernel_selection.md index 9719e031c70979cd95400701efd30879662e19bc..39ea2b00090a864f95610d6d2846ca5e5c904e78 100644 --- a/doc/fluid/design/muti_devices/kernel_selection.md +++ b/doc/fluid/design/muti_devices/kernel_selection.md @@ -1,4 +1,4 @@ -## Background +# Background Every operator has many kernels because there are multiple data types, places, data layout, library type that Fluid supports. We use the `OpKernelType ` to describe kernel types that operators can hold. The `OpKernelType ` is as follows: diff --git a/doc/fluid/design/network/index_cn.rst b/doc/fluid/design/network/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..3557d55fe4dbae1f712e0760ca15111ec6f6792d --- /dev/null +++ b/doc/fluid/design/network/index_cn.rst @@ -0,0 +1,7 @@ +复杂网络设计 +------------ + +.. toctree:: + :maxdepth: 1 + + sequence_decoder.md diff --git a/doc/fluid/design/network/index_en.rst b/doc/fluid/design/network/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..73a7137236bdf0548d35721609351d6deca3013b --- /dev/null +++ b/doc/fluid/design/network/index_en.rst @@ -0,0 +1,7 @@ +Complex Network Design +------------------------ + +.. toctree:: + :maxdepth: 1 + + sequence_decoder.md diff --git a/doc/fluid/dev/api_doc_std_cn.md b/doc/fluid/dev/api_doc_std_cn.md index 5596b2653ae6ed9917f77dad08f926bcb1fb3419..b50f18f21df0787b9761bf0935ed7f4384ff0f98 100644 --- a/doc/fluid/dev/api_doc_std_cn.md +++ b/doc/fluid/dev/api_doc_std_cn.md @@ -45,11 +45,11 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接 - Python API Definition - 格式： - + [Python API Definition] - + - 示例 - + ``` fc(input, size, @@ -63,19 +63,19 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接 ``` - Function Description - + - 格式 本模块应包含以下内容（排列顺序为文档撰写顺序）： [Function Description] - + [Formula] - + [Symbols' Descriptions if necessary] - + [References if necessary] - + - 示例 [Function Description] @@ -119,18 +119,18 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接 [References if necessary] 因fc没有必要列出的参考文献，故该内容省略。其他情况下需明确给出对应的参考文献和对应连接，以 layer_norm 为例： - + ``` Refer to `Layer Normalization `_ for more details. ``` - + - Args Description - + - 格式 - + \[Arg's Name\][(Data Type, Default Value)][Description] - + - 示例 fc的部分参数注释如下： @@ -145,35 +145,35 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接 ``` - Returns - + - 格式 - + [Name][Shape] - + - 示例 - + ``` Returns: A tensor variable storing the transformation result. ``` - + 当返回值为包含多个参数的tuple时，应按顺序逐个介绍各参数，以dynamic_lstm为例： - + ``` Returns: A tuple containing: The hidden state of LSTM whose shape is (T X D). The cell state of LSTM whose shape is (T X D). ``` - + - Raises - 格式 - + [Exception Type][Condition] - 示例 - + ``` Raises: ValueError: If the rank of the input is less than 2. @@ -182,7 +182,7 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接 - Note - 格式 - + [Note] - 示例 @@ -198,15 +198,15 @@ API文档须使用reStructuredText格式撰写，该格式详情请参考[链接 2. When num_heads == 1, scaled_dot_product_attention has no learnable parameters. ``` - + - Examples - 格式 \[Python Code Snipper] - + - 示例 - + ``` Examples: .. code-block:: python diff --git a/doc/fluid/dev/api_doc_std_en.md b/doc/fluid/dev/api_doc_std_en.md new file mode 100644 index 0000000000000000000000000000000000000000..e57072d52fd162e92a3482aef33f99ab9394c532 --- /dev/null +++ b/doc/fluid/dev/api_doc_std_en.md @@ -0,0 +1,226 @@ +# API Doc Standard + +- [API Doc Structure](#API Doc Structure) +- [Format and Examples](#Format and Examples) +- [Complete Example](#Complete Example) + + +## API Doc Structure + +API Doc should contain the following parts(please write them in order): + +- Python API Definition + + The definition of API + +- Function Description + + Description of API's function. + The description includes: meaning, purpose and operation on input of API, reference and corresponding link(if any), formula(if necessary) and explanations of key variables in the formula. + +- Args Description + + Description of API parameters. + Introduce parameters one by one according to the order in API definition. + The introduction includes: data type, default value(if any), meaning, etc. + +- Returns + + Introduction of API returned value. + Introduce meaning of returned value, provide correspoding format if necessary. + If returned value is a tuple containing multiple parameters, then introduce parameters one by one in order. + +- Raises（if any） + + Abnormality, error that may occur, and possible reasons. If there are more than one possible abnormity or error, they should be listed in order. + +- Note（if any） + + Matters needing attention. If there are more than one matters, they should be listed in order. + +- Examples + + Examples of how to use API. + + +## Format and Examples + +API documentation must obey reStructuredText format, please refer to [here](http://sphinx-doc-zh.readthedocs.io/en/latest/rest.html). +Format and examples of each part of API documantation are as follows: (take fc for example) + +- Python API Definition + + - Format + + [Python API Definition] + + - Example + + ``` + fc(input, + size, + num_flatten_dims=1, + param_attr=None, + bias_attr=None, + act=None, + name=None, + main_program=None, + startup_program=None) + ``` + +- Function Description + + - Format + + This part contains (please write them in order): + + [Function Description] + + [Formula] + + [Symbols' Descriptions if necessary] + + [References if necessary] + + - Example + + [Function Description] + + ``` + **Fully Connected Layer** + + The fully connected layer can take multiple tensors as its inputs. It + creates a variable called weights for each input tensor, which represents + a fully connected weight matrix from each input unit to each output unit. + The fully connected layer multiplies each input tensor with its coresponding + weight to produce an output Tensor. If multiple input tensors are given, + the results of multiple multiplications will be sumed up. If bias_attr is + not None, a bias variable will be created and added to the output. Finally, + if activation is not None, it will be applied to the output as well. + ``` + + [Formula] + + ``` + This process can be formulated as follows: + + .. math:: + + Out = Act({\sum_{i=0}^{N-1}X_iW_i + b}) + ``` + + [Symbols' Descriptions if necessary] + + ``` + In the above equation: + + * :math:`N`: Number of the input. + * :math:`X_i`: The input tensor. + * :math:`W`: The weights created by this layer. + * :math:`b`: The bias parameter created by this layer (if needed). + * :math:`Act`: The activation function. + * :math:`Out`: The output tensor. + ``` + + [References if necessary] + + Since there is no need for reference of fc, we omit them here. Under other circumstances, please provide explicit reference and link, take layer_norm for example: + + ``` + Refer to `Layer Normalization `_ for more details. + ``` + + +- Args Description + + - Format + + \[Arg's Name\][(Data Type, Default Value)][Description] + + - Example + + part of fc parameters are as follows: + + ``` + Args: + input (Variable|list of Variable): The input tensor(s) of this layer, and the dimension of + the input tensor(s) is at least 2. + param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable + parameters/weights of this layer. + name (str, default None): The name of this layer. + ``` + +- Returns + + - Format + + [Name][Shape] + + - Example + + ``` + Returns: + A tensor variable storing the transformation result. + ``` + + when returned value contain more than one tuple, please introduce every parameter in order, take dynamic_lstm for example: + + ``` + Returns: + A tuple containing: + The hidden state of LSTM whose shape is (T X D). + The cell state of LSTM whose shape is (T X D). + ``` + +- Raises + + - Format + + [Exception Type][Condition] + + - Example + + ``` + Raises: + ValueError: If the rank of the input is less than 2. + ``` + +- Note + + - Format + + [Note] + + - Example + + there is no Note in fc, so we omit this part. If there is any note, please write clearly. If there are more than one notes, please list them in order. Take scaled\_dot\_product\_attention for example: + + ``` + Note: + 1. When num_heads > 1, three linear projections are learned respectively + to map input queries, keys and values into queries', keys' and values'. + queries', keys' and values' have the same shapes with queries, keys + and values. + 2. When num_heads == 1, scaled_dot_product_attention has no learnable + parameters. + ``` + +- Examples + + - Format + + \[Python Code Snipper] + + - Example + + ``` + Examples: + .. code-block:: python + + data = fluid.layers.data(name="data", shape=[32, 32], dtype="float32") + fc = fluid.layers.fc(input=data, size=1000, act="tanh") + ``` + +## Complete Example + +Complete Example of fc please see [here](src/fc.py)。 diff --git a/doc/fluid/dev/index_cn.rst b/doc/fluid/dev/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..e70bf5dff3849f2ff82315f7eba4a92c93539843 --- /dev/null +++ b/doc/fluid/dev/index_cn.rst @@ -0,0 +1,13 @@ +开发标准 +------------ + +.. toctree:: + :maxdepth: 1 + + new_op_en.md + new_op_kernel_en.md + use_eigen_en.md + name_convention.md + support_new_device.md + releasing_process.md + op_markdown_format.md diff --git a/doc/fluid/dev/index_en.rst b/doc/fluid/dev/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..f0e9afcfcc9edfb9a91f58375cd415ea414f8f82 --- /dev/null +++ b/doc/fluid/dev/index_en.rst @@ -0,0 +1,13 @@ +Development +------------ + +.. toctree:: + :maxdepth: 1 + + new_op_en.md + new_op_kernel_en.md + use_eigen_en.md + name_convention.md + support_new_device.md + releasing_process.md + op_markdown_format.md diff --git a/doc/fluid/dev/name_convention.md b/doc/fluid/dev/name_convention.md index a02b356f058da68442516c2705d0bac140f8ef18..75830ef28c67dc4694d899efe503084b7b5852e1 100644 --- a/doc/fluid/dev/name_convention.md +++ b/doc/fluid/dev/name_convention.md @@ -1,8 +1,8 @@ -## Operator's Parameter Name Convention +# Operator's Parameter Name Convention To make the operator document itself more clear, we recommend operator names obey the listing conventions. -### OpProtoMaker names +## OpProtoMaker names When defining an operator in Paddle, a corresponding [OpProtoMaker](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L170) (TODO: OpProtoMaker Doc)need to be defined. All the Input/Output and Attributes will write into the [OpProto](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L61) , and will be used in client language to create operator. @@ -20,7 +20,7 @@ When defining an operator in Paddle, a corresponding [OpProtoMaker](https://gith - Order. - Follow the order of Input/Output, then Attribute, then Comments. See the example in best practice. -### Best Practice +## Best Practice Here we give some examples to show how these rules will be used. diff --git a/doc/fluid/dev/new_op_kernel_en.md b/doc/fluid/dev/new_op_kernel_en.md index 123df0a7ee4943c0b789ef9cfa6e0804d0fdd564..55dea8d0a39232ede59d4663d6e1a47fbfc60853 100644 --- a/doc/fluid/dev/new_op_kernel_en.md +++ b/doc/fluid/dev/new_op_kernel_en.md @@ -1,14 +1,14 @@ -## Add Kernels for a New Device +# Add Kernels for a New Device -### Background +## Background PaddlePaddle Fluid have hundreds of operators. Each operator could have one or more kernels. A kernel is an implementation of the operator for a certain device, which could be a hardware device, e.g., the CUDA GPU, or a library that utilizes a device, e.g., Intel MKL that makes full use of the Xeon CPU. [This document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md) explains how to add an operator, and its kernels. The kernels of an operator are indexed by a C++ type [`OpKernelType`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/operator_kernel_type.md). An operator chooses the right kernel at runtime. This choosing mechanism is described [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md). -### Write Kernels for A New Device +## Write Kernels for A New Device -#### Add A New Device +### Add A New Device For some historical reaons, we misuse the word *library* for *device*. For example, we call the deivce type by *library type*. An example is the header file [`library_type.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/library_type.h#L24). We will correct this ASAP. @@ -23,7 +23,7 @@ enum class LibraryType { ``` -#### Add A New [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L53) +### Add A New [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L53) If you have a new kind of Device, firstly you need to add a new kind of [`Place`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L53). For example `CUDAPlace`: @@ -45,7 +45,7 @@ struct CUDAPlace { typedef boost::variant Place; ``` -#### Add [device context]((https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L37)) +### Add [device context]((https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L37)) After a new kind of Device is added, you should add a corresponding [DeviceContext](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h#L37) for it. ```cpp @@ -58,7 +58,7 @@ class DeviceContext { }; ``` -#### Implement new [OpKernel](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L351) for your Device. +### Implement new [OpKernel](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L351) for your Device. A detailed documentation can be found in [`new_op_and_kernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/new_op_en.md) @@ -85,7 +85,7 @@ class OpKernel : public OpKernelBase { ``` -#### Register the OpKernel to framework +### Register the OpKernel to framework After writing the components described above, we should register the kernel to the framework. @@ -107,7 +107,7 @@ take [`conv2d`]((https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/oper REGISTER_OP_KERNEL(conv2d, CPU, paddle::platform::CPUPlace, paddle::operators::GemmConvKernel, paddle::operators::GemmConvKernel); - + REGISTER_OP_KERNEL(conv2d, CUDNN, ::paddle::platform::CUDAPlace, paddle::operators::CUDNNConvOpKernel, paddle::operators::CUDNNConvOpKernel); diff --git a/doc/fluid/dev/op_markdown_format.md b/doc/fluid/dev/op_markdown_format.md index 0ee804d592252c727622cbe59b0644813db3c4fd..4e539d7992e5f67ee7b07193b59b6b425b73c9e5 100644 --- a/doc/fluid/dev/op_markdown_format.md +++ b/doc/fluid/dev/op_markdown_format.md @@ -15,26 +15,26 @@ The signature of the operator. Each section mentioned above has been covered in further detail in the rest of the document. -# PaddlePaddle Operator Name +## PaddlePaddle Operator Name This should be in all small letters, in case of multiple words, we separate them with an underscore. For example: `array to lod tensor` should be written as `array_to_lod_tensor`. This naming convention should be standard across all PaddlePaddle operators. -# Standard Operator Name +## Standard Operator Name This is the standard name of the operator as used in the community. The general standard is usually: - Standard abbreviations like `SGD` are written in all capital letters. - Operator names that have multiple words inside a single word use `camelCase` (capitalize word boundaries inside of a word). - Keep numbers inside a word as is, with no boundary delimiters. - Follow the name of the operator with the keyword: `Activation Operator.` -# Operator description +## Operator description This section should contain the description of what the operator does, including the operation performed, the literature from where it comes and was introduced first, and other important details. The relevant paper/article including the hyperlink should be cited in this section. -# LaTeX equation +## LaTeX equation This section should contain an overall equation of the update or operation that the operator performs. The variables used in the equation should follow the naming convention of operators as described [here](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/name_convention.md). Two words in the same word should be separated by an underscore (`_`). -# The signature +## The signature This section describes the signature of the operator. A list of Inputs and Outputs, each of which have a small description of what the variable represents and the type of variable. The variable names follow the `CamelCase` naming convention. The proposed format for this is: `Section : VariableName : (VariableType) VariableDescription diff --git a/doc/fluid/dev/use_eigen_cn.md b/doc/fluid/dev/use_eigen_cn.md index f36843b4408c21bdca1fa83853e5b0a40116791c..75922e7d85a13e53ce94619a48d8da8b960e6c9a 100644 --- a/doc/fluid/dev/use_eigen_cn.md +++ b/doc/fluid/dev/use_eigen_cn.md @@ -1,16 +1,16 @@ -## 在Paddle中如何使用Eigen +# 在Paddle中如何使用Eigen 神经网络本质上是一个计算图，计算需要的数据存放在`Tensor`中，而计算过程是由`Operartor`来描述的。在执行时，`Operator`调用对应`OpKernel`中的`Compute`接口，实现对`Tensor`的操作。 -### Eigen Tensor模块 +## Eigen Tensor模块 Eigen Tensor模块对element-wise计算提供了强大的支持，并且书写一份代码，可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块，因此可能测试不够完备，文档较少。 关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md) -### paddle::framework::Tensor +## paddle::framework::Tensor Paddle Tensor定义在framework目录下，其主要接口如下： @@ -20,14 +20,14 @@ class Tensor { /*! Return a pointer to mutable memory block. */ template inline T* data(); - + /** * @brief Return a pointer to mutable memory block. * @note If not exist, then allocation. */ template inline T* mutable_data(platform::Place place); - + /** * @brief Return a pointer to mutable memory block. * @@ -38,17 +38,17 @@ class Tensor { */ template inline T* mutable_data(DDim dims, platform::Place place); - + /*! Resize the dimensions of the memory block. */ inline Tensor& Resize(const DDim& dims); - + /*! Return the dimensions of the memory block. */ inline const DDim& dims() const; private: /*! holds the memory block if allocated. */ std::shared_ptr holder_; - + /*! points to dimensions of memory block. */ DDim dim_; }; @@ -129,7 +129,7 @@ From是EigenTensor模板提供的一个接口，可以实现从paddle::framework -### 实现计算 +## 实现计算 当需要完成计算时，我们需要等式左边的EigenTensor调用device接口。在这里需要注意的是，这里的EigenTensor之间的运算只是改变了原有Tensor中的数据，而不会改变原有Tensor的shape信息。 diff --git a/doc/fluid/dev/use_eigen_en.md b/doc/fluid/dev/use_eigen_en.md index 3a466f73d1f9b94a29b171015279c782ca50bd02..3313d097cb21e40c23aa13187b6a50562f12403a 100644 --- a/doc/fluid/dev/use_eigen_en.md +++ b/doc/fluid/dev/use_eigen_en.md @@ -1,9 +1,9 @@ -## How to use Eigen in Paddle +# How to use Eigen in Paddle Essentially, a neural network is a compute graph. T data needed for the computation is stored in `Tensor`s and its computation procedure is described by `Operator`s. An `Operator` calls the `Compute` interface in its corresponding `OpKernel` and operates on the `Tensor`. -### Eigen Tensor Module +## Eigen Tensor Module The Eigen Tensor module supports powerful element-wise computation. In addition, a piece of code written using it can be run on both the CPU and the GPU. @@ -12,7 +12,7 @@ Note that Eigen Tensor is still being actively developed, so its tests are not c For details on Eigen Tensor module, please see [doc 1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) and [doc 2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md). -### paddle::framework::Tensor +## paddle::framework::Tensor Paddle Tensor's is defined in the framework directory with the following interface: @@ -105,7 +105,7 @@ void Compute(const framework::ExecutionContext& context) const override { ``` -### paddle::framework::Tensor到EigenTensor的转换 +## paddle::framework::Tensor到EigenTensor的转换 As shown above, in actual computation, we need to transform the input and output `Tensor`s into formats Eigen supports. We show some functions in [eigen.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/eigen.h) to implement the transformation from `paddle::framework::Tensor`to `EigenTensor/EigenMatrix/EigenVector/EigenScalar`. @@ -129,7 +129,7 @@ For more transformations, see the [unit tests](https://github.com/PaddlePaddle/P -### Implementing Computation +## Implementing Computation While computing, the device interface is needed from the EigenTensors on the left hand side of the assignments. Note that the computation between EigenTensors only changes the data originally inthe Tensor and does not change all the shape information associated with the Tensor. diff --git a/doc/fluid/faq/index_cn.rst b/doc/fluid/faq/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..395c1109891b5a00eab6f0b44d855658def7fdd6 --- /dev/null +++ b/doc/fluid/faq/index_cn.rst @@ -0,0 +1,2 @@ +FAQ +------------ diff --git a/doc/fluid/faq/index_en.rst b/doc/fluid/faq/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..395c1109891b5a00eab6f0b44d855658def7fdd6 --- /dev/null +++ b/doc/fluid/faq/index_en.rst @@ -0,0 +1,2 @@ +FAQ +------------ diff --git a/doc/fluid/getstarted/concepts/index_cn.rst b/doc/fluid/getstarted/concepts/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..2e7f70fc4cb871a80ffaffec6c06797973cd2f85 --- /dev/null +++ b/doc/fluid/getstarted/concepts/index_cn.rst @@ -0,0 +1,4 @@ +基本使用概念 +============ + +TBD diff --git a/doc/fluid/getstarted/concepts/index_en.rst b/doc/fluid/getstarted/concepts/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..78cca1e2a3443c2949ca0655190b0f05502f519a --- /dev/null +++ b/doc/fluid/getstarted/concepts/index_en.rst @@ -0,0 +1,4 @@ +Concepts +============ + +TBD diff --git a/doc/fluid/getstarted/index_cn.rst b/doc/fluid/getstarted/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..75af7354be93a6eeabfa9ccf86903505402a7ca6 --- /dev/null +++ b/doc/fluid/getstarted/index_cn.rst @@ -0,0 +1,19 @@ +新手入门 +============ + + +如果需要快速了解PaddlePaddle的使用，可以参考以下指南。 + +.. toctree:: + :maxdepth: 1 + + quickstart_cn.rst + + +在使用PaddlePaddle构建应用时，需要了解一些基本概念。 +这里以一个线性回归为例子，详细介绍了PaddlePaddle的使用流程，包括数据格式，模型配置与训练等。 + +.. toctree:: + :maxdepth: 1 + + concepts/use_concepts_cn.rst diff --git a/doc/fluid/getstarted/index_en.rst b/doc/fluid/getstarted/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..75a43f4af87c34830ec940068196e6ca72640501 --- /dev/null +++ b/doc/fluid/getstarted/index_en.rst @@ -0,0 +1,18 @@ +GET STARTED +============ + +If you want to quickly know how to use PaddlePaddle, please refer to the following guide: + +.. toctree:: + :maxdepth: 1 + + quickstart_en.rst + +While using PaddlePaddle to build applications, please understand some basic concepts. + +Here is an example of linear regression. It introduces workflow of PaddlePaddle, including data format, model configuration and training, etc. + +.. toctree:: + :maxdepth: 1 + + concepts/index_en.rst diff --git a/doc/fluid/getstarted/quickstart_cn.rst b/doc/fluid/getstarted/quickstart_cn.rst new file mode 120000 index 0000000000000000000000000000000000000000..93a9e4e37a8495c553cec257c27363ca8d062d39 --- /dev/null +++ b/doc/fluid/getstarted/quickstart_cn.rst @@ -0,0 +1 @@ +../../v2/getstarted/quickstart_cn.rst \ No newline at end of file diff --git a/doc/fluid/getstarted/quickstart_en.rst b/doc/fluid/getstarted/quickstart_en.rst new file mode 120000 index 0000000000000000000000000000000000000000..6e1894faa1176bb9e77f616e07df36191e54b782 --- /dev/null +++ b/doc/fluid/getstarted/quickstart_en.rst @@ -0,0 +1 @@ +../../v2/getstarted/quickstart_en.rst \ No newline at end of file diff --git a/doc/fluid/howto/index_cn.rst b/doc/fluid/howto/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..97aeaf167d329529f2b120b5a3d4085e0510fe16 --- /dev/null +++ b/doc/fluid/howto/index_cn.rst @@ -0,0 +1,7 @@ +进阶使用 +------------ + +.. toctree:: + :maxdepth: 1 + + optimization/index_cn.rst diff --git a/doc/fluid/howto/index_en.rst b/doc/fluid/howto/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..fd21e167ce3a46da167db1e9d7013804f730e047 --- /dev/null +++ b/doc/fluid/howto/index_en.rst @@ -0,0 +1,7 @@ +HOW TO +------------ + +.. toctree:: + :maxdepth: 1 + + optimization/index_en.rst diff --git a/doc/fluid/howto/optimization/benchmark/README.md b/doc/fluid/howto/optimization/benchmark/README.md new file mode 120000 index 0000000000000000000000000000000000000000..db30af7f53231c687f9ad61ad961a685733cbad0 --- /dev/null +++ b/doc/fluid/howto/optimization/benchmark/README.md @@ -0,0 +1 @@ +../../../../../benchmark/cluster/README.md \ No newline at end of file diff --git a/doc/fluid/howto/optimization/benchmark/index_cn.rst b/doc/fluid/howto/optimization/benchmark/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..9404800eb86ca6d27886258b67393028c76954dc --- /dev/null +++ b/doc/fluid/howto/optimization/benchmark/index_cn.rst @@ -0,0 +1,8 @@ +基准 +------------ + +.. toctree:: + :maxdepth: 1 + + vgg16/README.md + README.md diff --git a/doc/fluid/howto/optimization/benchmark/index_en.rst b/doc/fluid/howto/optimization/benchmark/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..1e200b660cc7f6aeaf8b3d94fd7a14999a52bccd --- /dev/null +++ b/doc/fluid/howto/optimization/benchmark/index_en.rst @@ -0,0 +1,8 @@ +Benchmark +------------ + +.. toctree:: + :maxdepth: 1 + + vgg16/README.md + README.md diff --git a/doc/fluid/howto/optimization/benchmark/vgg16/README.md b/doc/fluid/howto/optimization/benchmark/vgg16/README.md new file mode 120000 index 0000000000000000000000000000000000000000..ca963ef5f06aa0c2fe507ba7548dca8017358120 --- /dev/null +++ b/doc/fluid/howto/optimization/benchmark/vgg16/README.md @@ -0,0 +1 @@ +../../../../../../benchmark/cluster/vgg16/README.md \ No newline at end of file diff --git a/doc/fluid/howto/optimization/cpu_profiling_cn.md b/doc/fluid/howto/optimization/cpu_profiling_cn.md index d59be670c2b33b64d9b6f96b53f50e5bf9f0613b..17f895573a65731db34b2addddaa22e7f32157ec 100644 --- a/doc/fluid/howto/optimization/cpu_profiling_cn.md +++ b/doc/fluid/howto/optimization/cpu_profiling_cn.md @@ -8,7 +8,7 @@ PaddlePaddle 用户一般通过调用 Python API 编写深度学习程序。大 * Python 与 C++ 混合代码的性能分析 -## Python代码的性能分析 +# Python代码的性能分析 ### 生成性能分析文件 diff --git a/doc/fluid/howto/optimization/cpu_profiling_en.md b/doc/fluid/howto/optimization/cpu_profiling_en.md index 01e5fddf61547f9fc86ef18a6f2e2ac508d22dbb..abe4493c175fb4ee57f1acf45931e2890620d9c1 100644 --- a/doc/fluid/howto/optimization/cpu_profiling_en.md +++ b/doc/fluid/howto/optimization/cpu_profiling_en.md @@ -14,7 +14,7 @@ the profiling and tuning of 1. the Python code and 1. the mixture of Python and C++ code. -## Profiling the Python Code +# Profiling the Python Code ### Generate the Performance Profiling File @@ -81,7 +81,7 @@ focus on. We can sort above profiling file by tottime: We can see that the most time-consuming function is the `built-in method run`, which is a C++ function in `libpaddle.so`. We will -explain how to profile C++ code in the next section. At this +explain how to profile C++ code in the next section. At this moment, let's look into the third function `sync_with_cpp`, which is a Python function. We can click it to understand more about it: diff --git a/doc/fluid/howto/optimization/index_cn.rst b/doc/fluid/howto/optimization/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..27cc96702356703b339db845dc81913bdcc9f23b --- /dev/null +++ b/doc/fluid/howto/optimization/index_cn.rst @@ -0,0 +1,9 @@ +性能优化 +------------ + +.. toctree:: + :maxdepth: 1 + + timeline.md + cpu_profiling_cn.md + benchmark/index_cn.rst diff --git a/doc/fluid/howto/optimization/index_en.rst b/doc/fluid/howto/optimization/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..4ce624fe8f108a6afc7cd08a1542332755d22e04 --- /dev/null +++ b/doc/fluid/howto/optimization/index_en.rst @@ -0,0 +1,9 @@ +Performance Optimization +--------------------------- + +.. toctree:: + :maxdepth: 1 + + timeline.md + cpu_profiling_en.md + benchmark/index_en.rst diff --git a/doc/fluid/howto/optimization/timeline.md b/doc/fluid/howto/optimization/timeline.md index 9d9565a3e698a83ca465c5da83ff892360c33b8f..96481ae2a6e4442d40803f8d5361e5f942502df3 100644 --- a/doc/fluid/howto/optimization/timeline.md +++ b/doc/fluid/howto/optimization/timeline.md @@ -1,4 +1,4 @@ -## how to use timeline tool to do profile +# how to use timeline tool to do profile 1. Add `with profiler.profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number. diff --git a/doc/fluid/index_cn.rst b/doc/fluid/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..d878d192cae7ee9e8b8fdb4f615839c186fdf334 --- /dev/null +++ b/doc/fluid/index_cn.rst @@ -0,0 +1,12 @@ + PaddlePaddle Fluid +========================== + +.. toctree:: + :maxdepth: 1 + + getstarted/index_cn.rst + build_and_install/index_cn.rst + design/index_cn.rst + howto/index_cn.rst + dev/index_cn.rst + faq/index_cn.rst diff --git a/doc/fluid/index_en.rst b/doc/fluid/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..2bc76b58982cf50e637d15cca0c5d78166aa73a9 --- /dev/null +++ b/doc/fluid/index_en.rst @@ -0,0 +1,12 @@ + PaddlePaddle Fluid +========================== + +.. toctree:: + :maxdepth: 1 + + getstarted/index_en.rst + build_and_install/index_en.rst + design/index_en.rst + howto/index_en.rst + dev/index_en.rst + faq/index_en.rst diff --git a/doc/v2/CMakeLists.txt b/doc/v2/CMakeLists.txt index 286fe8845cd7a909d4030540e72362864b536063..82de7a3a3e1ca7724e1eda877d53454a4fa4129a 100644 --- a/doc/v2/CMakeLists.txt +++ b/doc/v2/CMakeLists.txt @@ -20,13 +20,15 @@ configure_file( "${BINARY_BUILD_DIR_EN}/conf.py" @ONLY) -sphinx_add_target(paddle_docs +sphinx_add_target(paddle_v2_docs html ${BINARY_BUILD_DIR_EN} ${SPHINX_CACHE_DIR_EN} ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_EN}) +add_dependencies(paddle_v2_docs gen_proto_py) + # configured documentation tools and intermediate build results set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build") @@ -41,11 +43,13 @@ configure_file( "${BINARY_BUILD_DIR_CN}/conf.py" @ONLY) -sphinx_add_target(paddle_docs_cn +sphinx_add_target(paddle_v2_docs_cn html ${BINARY_BUILD_DIR_CN} ${SPHINX_CACHE_DIR_CN} ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_CN}) +add_dependencies(paddle_v2_docs_cn gen_proto_py) + add_subdirectory(api) diff --git a/doc/v2/api/CMakeLists.txt b/doc/v2/api/CMakeLists.txt index 2ad589e8a260e48d46cba2300d6e2bcd4bdd8019..da1eafc02ed8cd155d4f0f1fbadcb7b237b6fcc1 100644 --- a/doc/v2/api/CMakeLists.txt +++ b/doc/v2/api/CMakeLists.txt @@ -12,9 +12,11 @@ configure_file( "${BINARY_BUILD_DIR_EN}/conf.py" @ONLY) -sphinx_add_target(paddle_api_docs +sphinx_add_target(paddle_v2_apis html ${BINARY_BUILD_DIR_EN} ${SPHINX_CACHE_DIR_EN} ${CMAKE_CURRENT_SOURCE_DIR} ${SPHINX_HTML_DIR_EN}) + +add_dependencies(paddle_v2_apis gen_proto_py framework_py_proto copy_paddle_pybind) diff --git a/doc/fluid/design/interface/00.why_plain_c.md b/doc/v2/design/interface/00.why_plain_c.md similarity index 100% rename from doc/fluid/design/interface/00.why_plain_c.md rename to doc/v2/design/interface/00.why_plain_c.md diff --git a/doc/fluid/design/interface/01.inference_implementation.md b/doc/v2/design/interface/01.inference_implementation.md similarity index 100% rename from doc/fluid/design/interface/01.inference_implementation.md rename to doc/v2/design/interface/01.inference_implementation.md diff --git a/doc/v2/design/interface/index_cn.rst b/doc/v2/design/interface/index_cn.rst new file mode 100644 index 0000000000000000000000000000000000000000..2509a5c5f4182d8ce3a16a3b7bd92c0d7bf5b056 --- /dev/null +++ b/doc/v2/design/interface/index_cn.rst @@ -0,0 +1,7 @@ +多语言接口 +------------ + +.. toctree:: + :maxdepth: 1 + + 00.why_plain_c.md diff --git a/doc/v2/design/interface/index_en.rst b/doc/v2/design/interface/index_en.rst new file mode 100644 index 0000000000000000000000000000000000000000..356e58c39c5ef6ee5ee50ab999b85f88628bfb85 --- /dev/null +++ b/doc/v2/design/interface/index_en.rst @@ -0,0 +1,7 @@ +Multilingual Interface +----------------------- + +.. toctree:: + :maxdepth: 1 + + 00.why_plain_c.md diff --git a/doc/v2/design/mkl/mkldnn.md b/doc/v2/design/mkl/mkldnn.md index e2fe1e6b26ffa73fda81863abfadf697c0acbfcf..1bd2e7bc34ee79eb753b3520d97e5e7beca89b0b 100644 --- a/doc/v2/design/mkl/mkldnn.md +++ b/doc/v2/design/mkl/mkldnn.md @@ -44,7 +44,7 @@ MKL，MKLML以及MKL-DNN三者关系如下表： | Name | Open Source | License | Descriptions | | :---------- | :--------------- | :---------- | :------------ | -| MKL | No | Proprietary | Accelerate math processing routines | +| MKL | No | Proprietary | Accelerate math processing routines | | MKLML | No | Proprietary | Small package of MKL, especially for Machine Learning | | MKL-DNN | Yes | Apache 2.0 | Accelerate primitives processing routines especially for Deep Neural Networks | @@ -89,7 +89,7 @@ PaddlePaddle/Paddle ### CMake 在`CMakeLists.txt`中提供一个与MKL有关的总开关：`WITH_MKL`，它负责决定编译时是否使用MKLML和MKL-DNN -- `WITH_MKLML` 控制是否使用MKLML库。 +- `WITH_MKLML` 控制是否使用MKLML库。 当打开`WITH_MKL`时，会自动使用MKLML库作为PaddlePaddle的CBLAS和LAPACK库，同时会开启Intel OpenMP用于提高MKLML的性能。 编译时会把对应的头文件和库放在`build/third_party/install/mklml/*`目录下对应的地方。 MKLML的库目前都是动态库，主要包括`libiomp5.so`和`libmklml_intel.so`。 @@ -172,7 +172,7 @@ if use_mkldnn self.layer_type = mkldnn_* ``` -所有MKL-DNN的`layer_type`会以*mkldnn_*开头，这些会在`MKLDNN*Layer`注册layer的时候保证，以示区分。 +所有MKL-DNN的`layer_type`会以*mkldnn_*开头，这些会在`MKLDNN*Layer`注册layer的时候保证，以示区分。 同时,会在`paddle/utils.Flags`中添加一个`use_mkldnn`的flag，用于选择是否使用MKL-DNN的相关功能。 diff --git a/doc/v2/dev/index_en.rst b/doc/v2/dev/index_en.rst index 549f5fa9aace7eb699d229e5f61fe10ae4ed4d66..36516b7953224e799e1065fd7930509eec0aa650 100644 --- a/doc/v2/dev/index_en.rst +++ b/doc/v2/dev/index_en.rst @@ -1,9 +1,27 @@ Development ------------ + +PaddlePaddle adheres to the following three sections of code and document specifications. + + +PaddlePaddle uses git for version control and Docker is used for building and testing environment. The code includes Cuda, C++, Python, Shell and other programming languages，which comply with Google C++ Style, Pep-8, and the code base includes style checking by an automatic inspection tool. Code comments need to follow the Doxygen specification. The code that does not meet the style requirements will fail to compile. We provide the following guidelines for the use of Git, build tests and code development. .. toctree:: :maxdepth: 1 contribute_to_paddle_en.md + + +PaddlePaddle is well documented in English and Chinese. We recommend using the English version of the documents and problem description. The design documents focus on problem descriptions, backgrounds, and are followed by solutions. As documents are generated by Sphinx, code comments should comply with the Sphinx documentation standard. We recommend to use the paddlepaddle.org tool to compile and generate and preview documents locally. Please refer to: + +.. toctree:: + :maxdepth: 1 + write_docs_en.rst + +PaddlePaddle V2 defines new operations by adding new Layers. You can implement various complex layers by combining basic APIs to satisfy most applications. If you want to customize layer, please refer to the following, and welcome to propose patch. + +.. toctree:: + :maxdepth: 1 + new_layer_en.rst diff --git a/doc/v2/dev/write_docs_cn.rst b/doc/v2/dev/write_docs_cn.rst index a055bb04c0c093c9159290067e5ccbd2525cd519..23615f8830e99633676c83ec5d28139a732c623c 100644 --- a/doc/v2/dev/write_docs_cn.rst +++ b/doc/v2/dev/write_docs_cn.rst @@ -2,13 +2,14 @@ 如何贡献文档 ############# -PaddlePaddle的文档包括中英文两个部分。文档都是通过 ``cmake`` 驱动 ``sphinx`` 编译生成，也可以利用paddlepaddle.org工具来编译和预览文档。 +PaddlePaddle的文档包括中英文两个部分。文档都是通过 ``cmake`` 驱动 ``sphinx`` 编译生成的，PaddlePaddle.org工具可以帮助我们实现这一编译过程，并提供更好的预览效果。 如何构建文档 ============ PaddlePaddle的文档构建有两种方式，分别为使用paddlepaddle.org工具和不使用paddlepaddle.org工具，两种方式都有各自的优点，前者方便预览，后者方便开发者进行调试。这两种方式中又分别有使用docker和不使用docker的两种构建方法。 +我们建议使用PaddlePaddle.org工具来构建文档。 使用PaddlePaddle.org工具 ------------------------ @@ -31,7 +32,7 @@ PaddlePaddle.org工具可以配合Docker使用，需要在系统里先安装好D docker run -it -p 8000:8000 -v `pwd`:/var/content paddlepaddle/paddlepaddle.org:latest 注意: PaddlePaddle.org 会在 -v (volume) 指定的内容存储库运行命令 -之后再用网页连到http://localhost:8000就可以在网页上生成需要的文档 +之后再用网页连到 http://localhost:8000 就可以在网页上生成需要的文档 编译后的文件将被存储在工作目录 /.ppo_workspace/content。 如果不想使用Docker，你还可以通过运行Django框架直接激活工具的服务器。使用下面的命令来运行它。 @@ -56,7 +57,7 @@ PaddlePaddle.org工具可以配合Docker使用，需要在系统里先安装好D python manage.py runserver 工具服务器将读取环境变量 CONTENT_DIR 搜索代码库。请指定的PaddlePaddle工作目录给环境变量 CONTENT_DIR。 -之后再用网页连到http://localhost:8000就可以在网页上生成需要的文档。 +之后再用网页连到 http://localhost:8000 就可以在网页上生成需要的文档。 编译后的文件将被存储在工作目录 /.ppo_workspace/content。 想了解更多PaddlePaddle.org工具的详细信息，可以 `点击这里 `_ 。 @@ -96,7 +97,7 @@ PaddlePaddle.org工具可以配合Docker使用，需要在系统里先安装好D python -m SimpleHTTPServer 8088 -在浏览器中输入http://localhost:8088就可以看到编译生成的中/英文的文档页面和英文的API页面,下图为生成的英文文档首页示例。注意，示例中由于使用了sphinx的原始主题，所以页面的风格与官网并不一致，但这并不影响开发者进行调试。 +在浏览器中输入 http://localhost:8088 就可以看到编译生成的中/英文的文档页面和英文的API页面,下图为生成的英文文档首页示例。注意，示例中由于使用了sphinx的原始主题，所以页面的风格与官网并不一致，但这并不影响开发者进行调试。 .. image:: src/doc_en.png :align: center diff --git a/doc/v2/dev/write_docs_en.rst b/doc/v2/dev/write_docs_en.rst index f3408a84269aaeef19986c220454555fbbe30e23..15ff0d34ad622f100fe98d8738b830e47c35b41b 100644 --- a/doc/v2/dev/write_docs_en.rst +++ b/doc/v2/dev/write_docs_en.rst @@ -2,21 +2,20 @@ Contribute Documentation ######################## -PaddlePaddle supports English documentation ``doc`` and Chinese documentation ``doc_cn``. -Both are compiled by `cmake`_ and `sphinx`_ , the compiled documentations will be stored under ``doc`` and ``doc_cn`` directories. -When using the PaddlePaddle.org to compile documentations, the compiled documentations will be stored under a consolidated directory: .ppo_workspace/content +PaddlePaddle's documentation includes both Chinese and English versions. The documentation is built using the ``cmake`` command to drive the ``sphinx`` compiler. The PaddlePaddle.org tool helps us to implement this compilation process and provides better preview results. -How to Build Documentations -============ +How to build Documentation +=========================== -We recommend using PaddlePaddle.org tool to build documentation +PaddlePaddle's documentation is built in two ways: using the PaddlePaddle.org tool and without using it. Both methods have their own advantages. The former facilitates previewing, while the latter facilitates debugging by the developer. We could choose to build the documentation with Docker or without it in each of the above ways. +We recommend using PaddlePaddle.org tool to build documentation. -Use PaddlePaddle.org tool --------------- -This is the recommended method to build documentation. It can compile documentation and preview the documentation in a web browser. +Using PaddlePaddle.org tool +----------------------------- +This is the recommended method to build documentation, because it can automatically compile the documentation and preview the documentation directly in a web page. Note that, although you can preview the documentation in other ways, its style may not be consistent with the official website. Compiling with the PaddlePaddle.org tool produces a preview that will be consistent with the official website documentation style. -The tool uses Docker, please install it on your system. Please check Docker official website on how to install Docker. You may use the following commands to activate the tool +The PaddlePaddle.org tool can be used with Docker and Docker needs to be installed first. Please refer to `Docker's official website `_ on how to install Docker. After installing Docker, you may use the following commands to activate the tool .. code-block:: bash @@ -32,8 +31,8 @@ The tool uses Docker, please install it on your system. Please check Docker offi # Please specify the working directory through -v docker run -it -p 8000:8000 -v `pwd`:/var/content paddlepaddle/paddlepaddle.org:latest -Note: PaddlePaddle.org will read the content repos specified in the -v (volume) flag of the docker run command -Use a web browser and navigate to http://localhost:8000, click the buttons to compile the documentation +Note: PaddlePaddle.org will read the content repos specified in the -v (volume) flag of the docker run commands +Use a web browser and navigate to http://localhost:8000. Click the buttons to compile the documentation. The compiled documentations will be stored in /.ppo_workspace/content @@ -58,19 +57,62 @@ If you don't wish to use Docker, you can also activate the tool through Django. pip install -r requirements.txt python manage.py runserver -Use a web browser and navigate to http://localhost:8000, click the buttons to compile the documentation +Specify the PaddlePaddle working directory for the environment variable CONTENT_DIR so that the tool could find where the working directory is. + +Use a web browser and navigate to http://localhost:8000. Click the buttons to compile the documentation The compiled documentations will be stored in /.ppo_workspace/content -If you want to learn more on the PaddlePaddle.org, please `click here `_ 。 +Please `click here `_ for more information about the PaddlePaddle.org tool. + + +Manually Building the Documentation +------------------------------------- + +Build PaddlePaddle's documentation with Docker，you need to install Docker first. Please refer to `Docker's official website `_ on how to install Docker. After Docker is installed, you could use the scripts in the source directory to build the documentation. + +[TBD] + +If you do not wish to use Docker, you can also use the following commands to directly build the PaddlePaddle documentation. + +.. code-block:: bash + + mkdir paddle + cd paddle + git clone https://github.com/PaddlePaddle/Paddle.git + mkdir -p build + cd build + cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON + + # If you only need to build documents, use the following commands + make -j $processors gen_proto_py + make -j $processors paddle_docs paddle_docs_cn + + # If you only need to build APIs, use the following commands + make -j $processors gen_proto_py framework_py_proto + make -j $processors copy_paddle_pybind + make -j $processors paddle_api_docs + +$processors indicates that as many processes as the CPU cores are started to compile in parallel. It should be set according to the number of CPU cores of your machine. + +After the compilation is complete, enter the ``doc/v2`` directory. If you chose to build documents, it will generate ``cn/html/`` and ``en/html`` subdirectories under this directory. If you chose to build APIs，it will generate``api/en/html`` subdirectory. Please enter these directories respectively and execute the following commands: + +.. code-block:: bash + + python -m SimpleHTTPServer 8088 + +Use a web browser and navigate to http://localhost:8000, you could see the compiled Chinese/English documents page and the English APIs page. The following figure is an example of the built English documents home page. Note that due to the sphinx's original theme used in the example, the style of the page is not consistent with the official website, but this does not affect the developer's debugging. -How to write Documentations -============ +.. image:: src/doc_en.png + :align: center + :scale: 60 % -PaddlePaddle uses `sphinx`_ to compile documentations，Please check sphinx official website for more detail. +How to write Documentation +=========================== +PaddlePaddle uses `sphinx`_ to compile documentation，Please check sphinx official website for more detail. How to update www.paddlepaddle.org -============================ +=================================== Please create PRs and submit them to github, please check `Contribute Code `_ 。 PaddlePaddle develop branch will update the documentation once the PR is merged. User may check latest `Chinese Docs `_ and diff --git a/doc/v2/faq/build_and_install/index_cn.rst b/doc/v2/faq/build_and_install/index_cn.rst index 7c7e896d187e4fe1544d7ec933fa4fa9f24df3cd..f292684fb5fe2df06db5239e7f43fdfa1dd2f2bd 100644 --- a/doc/v2/faq/build_and_install/index_cn.rst +++ b/doc/v2/faq/build_and_install/index_cn.rst @@ -139,3 +139,77 @@ PaddlePaddle使用avx SIMD指令提高cpu执行效率，因此错误的使用二 touch ../extern_mklml-stamp/extern_mklml-download // 4. 接着编译即可 + +9. 在Mac上无法安装numpy等Python包，权限错误 +------------------ + +Mac上对自带的Python和包有严格的权限保护，最好不要在自带的Python上安装。建议用virtualenv建立一个新的Python环境来操作。 + +virtualenv的基本原理是将机器上的Python运行所需的运行环境完整地拷贝一份。我们可以在一台机器上制造多份拷贝，并在这多个拷贝之间自由切换，这样就相当于在一台机器上拥有了多个相互隔离、互不干扰的Python环境。 + +下面简单介绍下如何用virtualenv为Paddle生成一个专用的Python环境： + +安装virtualenv： +:::::::::::::::: + +virtualenv本身也是Python的一个包，可以用pip进行安装： + +.. code-block:: bash + + sudo -H pip install virtualenv + +由于virtualenv需要安装给系统自带的Python，因此需要使用sudo权限。 + +创建一个新的Python运行环境： +::::::::::::::::::: + +.. code-block:: bash + + virtualenv --no-site-packages paddle + +--no-site-packages 参数表示不拷贝已有的任何第三方包，创造一个完全干净的新Python环境。后面的paddle是我们为这个新创建的环境取的名字。 + +执行完这一步后，当前目录下应该会出现一个名为paddle（或者你取的其他名字）的目录。这个目录里保存了运行一个Python环境所需要的各种文件。 + +启动运行环境： +:::::::::::::::: + +.. code-block:: bash + + source paddle/bin/activate + +执行后会发现命令提示符前面增加了(paddle)字样，说明已经成功启动了名为‘paddle’的Python环境。执行which python，可以发现使用的已经是刚刚创建的paddle目录下的Python。 + +在这个环境中，我们可以自由地进行Paddle的安装、使用和开发工作，无需担心对系统自带Python的影响。 + +退出运行环境： +::::::::::::::: + +直接执行： + +.. code-block:: bash + + deactivate + +可以看到命令提示符前面的(paddle)字样消失。 + +自动启动某一Python环境： +:::::::::::::::: + +如果我们经常使用Paddle，我们每次打开终端后都需要执行一下source paddle/bin/activate来启动环境，比较繁琐。为了简便，可以修改终端的配置文件，来让终端每次启动后自动启动特定的Python环境。 + +执行: + +.. code-block:: bash + + vi ~/.bash_profile + +打开终端配置文件，并在文件的最后添加一行： + +.. code-block:: bash + + source paddle/bin/activate + +保存并关闭文件。 + +这样，每次打开终端时就会自动启动名为‘paddle’的Python环境了。 diff --git a/doc/v2/faq/cluster/index_en.rst b/doc/v2/faq/cluster/index_en.rst index 855b7e8e53307b82a72c156be4ef509e27edf822..fa942a09625bef78b28456beeb735272b686e061 100644 --- a/doc/v2/faq/cluster/index_en.rst +++ b/doc/v2/faq/cluster/index_en.rst @@ -2,4 +2,15 @@ Cluster Training and Prediction ############################### -TBD +.. contents:: + +1. Network connection errors in the log during multi-node cluster training +------------------------------------------------ +There are maybe some errors in the log belonging to network connection problem during multi-node cluster training, for example, :code:`Connection reset by peer`. +This kind of error is usually caused by the abnormal exit of a training process in some node, and the other nodes cannot connect with this node any longer. Steps to troubleshoot the problem are as follows: + +* Find the first error in the :code:`train.log`, :code:`server.log`, check whether other fault casued the problem, such as FPE, lacking of memory or disk. + +* If the first error in server.log says "Address already used", this may be caused by the port conflict of the non-exclusive execution. Connect the sys-admin to check if the current MPI cluster supports jobs submitted with parameter :code:`resource=full`. If the current MPI cluster does not support this parameter, change the server port and try agian. + +* If the current MPI cluster does not support exclusive pattern which allows a process to occupy the whole node, ask the administrator to replace or update the this cluster. diff --git a/doc/v2/faq/model/index_en.rst b/doc/v2/faq/model/index_en.rst index cb26f59655f97dc28a2047994643ae16b8857964..67a33e08e192e5627ac3b0abd76e979f21ed2079 100644 --- a/doc/v2/faq/model/index_en.rst +++ b/doc/v2/faq/model/index_en.rst @@ -2,4 +2,80 @@ Model Configuration ################### -TBD +.. contents:: + +1. How to deal with error :code:`Duplicated layer name` +---------------------------------------------------------- + +The general reason for this error is that users may have set the same value for the attribute :code:`name` in different layers. Try to find out the :code:`name` attribute with the same value in diffrent layers and set them differently. + +2. How to use :code:`paddle.layer.memory`'s attribute :code:`name` +---------------------------------------------------------------------- + +* :code:`paddle.layer.memory` is used to get the output of a layer's last timestep and the layer is specified by the attribute :code:`name` . Thus, :code:`paddle.layer.memory` will associate with the layer that has the same value of attribute :code:`name` , and uses the output of the layer's last timestep as the input of its current timestep. + +* All the PaddlePaddle's layers have a unique name, which is set by the attribute :code:`name` . PaddlePaddle will automatically set it for the user when it is not explicitly set. :code:`paddle.layer.memory` is not a real layer, its name is set by the attribute :code:`memory_name` and PaddlePaddle will also automatically set it when the user does not explicitly set. The :code:`paddle.layer.memory` attribute :code:`name` is used to specify the layer it is associated with, and needs to be explicitly set by the user. + + +3. What is the difference between the two ways of using dropout +----------------------------------------------------------------- + +* There are two ways to use dropout in PaddlePaddle + + * Set the :code:`drop_rate` parameter in the layer's :code:`layer_atter` attribute. Take :code:`paddle.layer.fc` as an example: + + .. code-block:: python + + fc = paddle.layer.fc(input=input, layer_attr=paddle.attr.ExtraLayerAttribute(drop_rate=0.5)) + + * Use :code:`paddle.layer.dropout` layer. Take :code:`paddle.layer.fc` as an example: + + .. code-block:: python + + fc = paddle.layer.fc(input=input) + drop_fc = paddle.layer.dropout(input=fc, dropout_rate=0.5) + +* :code:`paddle.layer.dropout` actually uses the :code:`paddle.layer.add_to` layer and sets :code:`drop_rate` as the previous method. This method is very memory intensive. + +* PaddlePaddle implements dropout in the activation function rather than in the layer. + +* :code:`paddle.layer.lstmemory`, :code:`paddle.layer.grumemory`, :code:`paddle.layer.recurrent` implement activation of output in an unusual way, so we cannot use dropout by setting :code:`drop_rate` . To use dropout for these layers, we could use the second method, which is to use :code:`paddle.layer.dropout`. + +4. The differences between different recurrent layers +-------------------------------------------------------- +Take LSTM as an example. There are several kinds of recurrent layers in PaddlePaddle: + +* :code:`paddle.layer.lstmemory` +* :code:`paddle.networks.simple_lstm` +* :code:`paddle.networks.lstmemory_group` +* :code:`paddle.networks.bidirectional_lstm` + +According to implementations, recurrent layer can be classified into 2 types: + +1. Recurrent layer implemented by recurrent_group: + + * Using this type of recurrent layers, users can access the intermediate value calculated by the recurrent unit within a timestep (eg: hidden states, memory cells, etc.) + * :code:`paddle.networks.lstmemory_group` belongs to this type of recurrent layers. + +2. Recurrent layer implemented as a complete operation： + + * Users can only access output values when using this type of recurrent layers. + * :code:`paddle.networks.lstmemory_group` , :code:`paddle.networks.simple_lstm` and :code:`paddle.networks.bidirectional_lstm` belong to this type of recurrent layer； + +By implementing recurrent layer as a complete operation, CPU and GPU calculations can be optimized. Therefore, the second type of recurrent layer is more efficient than the first one. In practical applications, we propose to use the second type of recurrent layers if there is no need to access the intermediate variable of LSTM. + +In addition, PaddlePaddle also contains a kind of LSTM calculation unit: :code:`paddle.networks.lstmemory_unit`: + + * Unlike the recurrent layer described above, :code:`paddle.networks.lstmemory_unit` defines the computational process of an LSTM unit in a timestep. It is not a complete recurrent layer, nor can it receive sequence data as input. + * :code:`paddle.networks.lstmemory_unit` can only be used as a step function in recurrent_group. + +5. Can Softmax's calculation dimension be specified？ +-------------------------------------------------------------------- + +We can't specify calculation dimension for PaddlePaddle's softmax. It can only be calculated by rows. +In image tasks, for NCHW, if you need to calculate softmax in C dimension, you could use :code:`paddle.layer.switch_order` to change the dimension order, that is, convert NCHW to NHWC, then do the reshape operation and calculate softmax. + +6. Does PaddlePaddle support variable-dimensional data inputs +---------------------------------------------------------------- + +PaddlePaddle provides :code:`paddle.data_type.dense_array` to support variable-dimensional data input. Simply set the dimension of the data layer to a value larger than the dimension of the input data for occupancy. diff --git a/doc/v2/howto/capi/index_en.rst b/doc/v2/howto/capi/index_en.rst index 2cbbe362fd8e06abe9866d998f60fbb3458a80b5..4ec39c9d5223442cf6872edaf7befeb5053b538e 100644 --- a/doc/v2/howto/capi/index_en.rst +++ b/doc/v2/howto/capi/index_en.rst @@ -1,6 +1,23 @@ -C-API Prediction Library +C-API Inference Library ======================== +After we train a neural network, we use it to do inference. Inference is the process of preparing input data and propagating it through the model to produce the result. + +Compared with model training, prediction has the following features: + +#. Inference does not require backpropagation and parameter updates, as required during training. +#. Labels are not needed in prediction. +#. Most of the time, predictions need to be integrated with the user system. + +Therefore, the model prediction SDK needs to be designed separately and has the following features: + +#. The predictive SDK does not include backpropagation and parameter updates to reduce the size of the SDK. +#. The predictive SDK needs a simple user interface for ease of use. +#. Since the input data may have a variety of structures, the format of the input data is clearly and compactly packaged. +#. In order to be compatible with user's system, the SDK's interface must conform to the C-standard interface. + +PaddlePaddle provides C-API to solve the above problem. Following are the guidelines to use the C-API: + .. toctree:: :maxdepth: 1 diff --git a/doc/v2/howto/cmd_parameter/index_en.rst b/doc/v2/howto/cmd_parameter/index_en.rst index 0e3c72d27aca063f1b6f1c23e55718dba373c40a..f49683948ef78f363e2439cc25332431830eeb24 100644 --- a/doc/v2/howto/cmd_parameter/index_en.rst +++ b/doc/v2/howto/cmd_parameter/index_en.rst @@ -2,10 +2,25 @@ Set Command-line Parameters =========================== +The implementation of deep learning algorithms has a variety of characteristics, such as running environment, running stage, structure of the model and the traning strategy. PaddlePaddle supports the user to set various command-line parameters flexibly, which helps to achieve control of the model training or prediction process. + +In this part, we take several actual scenarios as an example, and the use of some command-line parameters is displayed: .. toctree:: :maxdepth: 1 use_case_en.md + +Then, we summarize and classify the use of all command-line parameters: + +.. toctree:: + :maxdepth: 1 + arguments_en.md + +Finally, the detailed descriptions are given, and we try to explain the propeties and significance of these command-line parameters in detail: + +.. toctree:: + :maxdepth: 1 + detail_introduction_en.md diff --git a/doc/v2/howto/index_en.rst b/doc/v2/howto/index_en.rst index 2079be766f2d8e6d63ca11dccd98f80613309ceb..35ef197f58f1f865e2cdbdebb567d5637284637a 100644 --- a/doc/v2/howto/index_en.rst +++ b/doc/v2/howto/index_en.rst @@ -1,11 +1,37 @@ HOW TO -======= +======== + +PaddlePaddle provides the users the ability to flexibly set various command line parameters to control the model training and inference process. Please refer to the following instructions on using PaddlePaddle: .. toctree:: :maxdepth: 1 cmd_parameter/index_en.rst + +PaddlePaddle supports distributed training tasks on fabric clusters, MPI clusters, and Kubernetes clusters. For detailed configuration and usage instructions, refer to: + +.. toctree:: + :maxdepth: 1 + cluster/index_en.rst + +PaddlePaddle provides a C-API for inference. We provide the following guidelines for using the C-API: + +.. toctree:: + :maxdepth: 1 + capi/index_en.rst + +PaddlePaddle supports a variety of flexible and efficient recurrent neural networks. For details, please refer to： + +.. toctree:: + :maxdepth: 1 + rnn/index_en.rst + +How to use the built-in timing tool, nvprof, or nvvp to run performance analysis and tuning, please refer to： + +.. toctree:: + :maxdepth: 1 + optimization/gpu_profiling_en.rst diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt index d2a4b1335464f553a361728e64ed5ca177ca53da..c44f8a8a8ecc1ba1f886fc41aec863b4ca3458a6 100644 --- a/paddle/CMakeLists.txt +++ b/paddle/CMakeLists.txt @@ -1,4 +1,4 @@ -if(NOT WITH_FLUID) +if(NOT WITH_FLUID_ONLY) add_subdirectory(cuda) add_subdirectory(function) add_subdirectory(utils) diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index a4ea74a6d2fbc29dc33a6b57ee453f49ed36c7fa..c425c71160a8fa3830a5fbdae1baaed850710877 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -1,3 +1,4 @@ +add_subdirectory(details) # ddim lib proto_library(framework_proto SRCS framework.proto) @@ -87,6 +88,9 @@ cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glo cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto backward glog lod_rank_table feed_fetch_method) + +cc_library(parallel_executor SRCS parallel_executor.cc DEPS multi_devices_graph_builder threaded_ssa_graph_executor) + cc_library(prune SRCS prune.cc DEPS framework_proto) cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context) cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry diff --git a/paddle/fluid/framework/block_desc.cc b/paddle/fluid/framework/block_desc.cc index 3693bc25d81a8309df1a6ddf3d9b08d484596ea9..fbe08349c37c4fde115ceea954ba2b84880088d7 100644 --- a/paddle/fluid/framework/block_desc.cc +++ b/paddle/fluid/framework/block_desc.cc @@ -147,15 +147,52 @@ void BlockDesc::RemoveOp(size_t s, size_t e) { if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) { return; } + auto get_vars = [](std::deque>::iterator &op, + std::vector &v) { + auto in_names = (*op)->InputArgumentNames(); + v.insert(v.end(), in_names.begin(), in_names.end()); + auto out_names = (*op)->OutputArgumentNames(); + v.insert(v.end(), out_names.begin(), out_names.end()); + std::sort(v.begin(), v.end()); + auto last = std::unique(v.begin(), v.end()); + v.erase(last, v.end()); + }; need_update_ = true; - for (auto it = ops_.begin() + s; it != ops_.begin() + e; it++) { - auto names = (*it)->InputArgumentNames(); - for (auto n : names) { - // TODO(typhoonzero): delete vars if no other op use it. - VLOG(3) << "deleting var " << n; + + for (size_t i = s; i < e; i++) { + // since remove op one by one, every time remove the first op. + auto op = ops_.begin() + s; + + // collect input and output variables from current delete op + std::vector cur_vars; + get_vars(op, cur_vars); + + // remove current op + ops_.erase(ops_.begin() + s); + + // collect input and output variables from other ops + std::vector other_vars; + for (auto it = ops_.begin(); it != ops_.end(); it++) { + get_vars(it, other_vars); + } + + // variables should be deleted + std::vector delete_vars; + // delete_vars = cur_vars - cur_vars ^ other_input_vars + std::set_difference(cur_vars.begin(), cur_vars.end(), other_vars.begin(), + other_vars.end(), + std::inserter(delete_vars, delete_vars.end())); + // remove variables + for (size_t i = 0; i < delete_vars.size(); i++) { + auto name = delete_vars[i]; + auto it = vars_.find(name); + PADDLE_ENFORCE(it != vars_.end(), + "%s is not in variable list, it should not be deleted", + name); + vars_.erase(it); + VLOG(3) << "deleting variable " << name; } } - ops_.erase(ops_.begin() + s, ops_.begin() + e); } std::vector BlockDesc::AllOps() const { diff --git a/paddle/fluid/framework/block_desc.h b/paddle/fluid/framework/block_desc.h index 185f018ac1b5863e0ee86fdaa17df1ccbc6e030e..468423e0e8e7b8c9ebc14b7568c9c3bd21645ea7 100644 --- a/paddle/fluid/framework/block_desc.h +++ b/paddle/fluid/framework/block_desc.h @@ -89,6 +89,11 @@ class BlockDesc { OpDesc *InsertOp(size_t index); + /* + * Remove Op and its input/output variables. + * Note that for either input or ouput variable, if it is also an input or + * output variable of other ops, we should remain it. + */ void RemoveOp(size_t s, size_t e); std::vector AllOps() const; diff --git a/paddle/fluid/framework/channel.h b/paddle/fluid/framework/channel.h index adfaba26ace78f547161ad4029a741f3ca8a6764..019bea600f496a6b58579ad0aa8af836cd6134a9 100644 --- a/paddle/fluid/framework/channel.h +++ b/paddle/fluid/framework/channel.h @@ -34,7 +34,7 @@ class Channel { public: virtual bool CanSend() = 0; virtual bool CanReceive() = 0; - virtual bool Send(T*) = 0; + virtual void Send(T*) = 0; virtual bool Receive(T*) = 0; virtual size_t Cap() = 0; virtual void Lock() = 0; @@ -84,69 +84,81 @@ class ChannelHolder { } template - bool Send(T* data) { - if (!IsInitialized()) return false; - PADDLE_ENFORCE_EQ(holder_->Type(), std::type_index(typeid(T))); + void Send(T* data) { + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + PADDLE_ENFORCE_EQ( + holder_->Type(), std::type_index(typeid(T)), + "Channel type is not same as the type of the data being sent"); // Static cast should be safe because we have ensured that types are same Channel* channel = static_cast*>(holder_->Ptr()); - return channel != nullptr ? channel->Send(data) : false; + PADDLE_ENFORCE_EQ(channel != nullptr, true, "Channel should not be null."); + channel->Send(data); } template bool Receive(T* data) { - if (!IsInitialized()) return false; - PADDLE_ENFORCE_EQ(holder_->Type(), std::type_index(typeid(T))); + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + PADDLE_ENFORCE_EQ( + holder_->Type(), std::type_index(typeid(T)), + "Channel type is not same as the type of the data being sent"); Channel* channel = static_cast*>(holder_->Ptr()); - return channel != nullptr ? channel->Receive(data) : false; + PADDLE_ENFORCE_EQ(channel != nullptr, true, "Channel should not be null."); + return channel->Receive(data); } bool IsClosed() { - if (IsInitialized()) { - return holder_->IsClosed(); - } - return false; + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + return holder_->IsClosed(); } bool CanSend() { - if (IsInitialized()) { - return holder_->CanSend(); - } - return false; + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + return holder_->CanSend(); } bool CanReceive() { - if (IsInitialized()) { - return holder_->CanReceive(); - } - return false; + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + return holder_->CanReceive(); } void close() { - if (IsInitialized()) holder_->Close(); + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + holder_->Close(); } size_t Cap() { - if (IsInitialized()) return holder_->Cap(); - return -1; + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + return holder_->Cap(); } void Lock() { - if (IsInitialized()) holder_->Lock(); + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + holder_->Lock(); } void Unlock() { - if (IsInitialized()) holder_->Unlock(); + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + holder_->Unlock(); } template void AddToSendQ(const void* referrer, T* data, std::shared_ptr cond, std::function cb) { - if (IsInitialized()) { - Channel* channel = static_cast*>(holder_->Ptr()); - if (channel != nullptr) { - channel->AddToSendQ(referrer, data, cond, cb); - } + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + Channel* channel = static_cast*>(holder_->Ptr()); + if (channel != nullptr) { + channel->AddToSendQ(referrer, data, cond, cb); } } @@ -154,26 +166,31 @@ class ChannelHolder { void AddToReceiveQ(const void* referrer, T* data, std::shared_ptr cond, std::function cb) { - if (IsInitialized()) { - Channel* channel = static_cast*>(holder_->Ptr()); - if (channel != nullptr) { - channel->AddToReceiveQ(referrer, data, cond, cb); - } + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + Channel* channel = static_cast*>(holder_->Ptr()); + if (channel != nullptr) { + channel->AddToReceiveQ(referrer, data, cond, cb); } } void RemoveFromSendQ(const void* referrer) { - if (IsInitialized()) holder_->RemoveFromSendQ(referrer); + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + holder_->RemoveFromSendQ(referrer); } void RemoveFromReceiveQ(const void* referrer) { - if (IsInitialized()) holder_->RemoveFromReceiveQ(referrer); + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); + holder_->RemoveFromReceiveQ(referrer); } inline bool IsInitialized() const { return holder_ != nullptr; } inline const std::type_index Type() { - PADDLE_ENFORCE_EQ(IsInitialized(), true); + PADDLE_ENFORCE_EQ(IsInitialized(), true, + "The Channel hasn't been initialized"); return holder_->Type(); } diff --git a/paddle/fluid/framework/channel_impl.h b/paddle/fluid/framework/channel_impl.h index 457abbf373d4549229e8fd8bd6b2087cc6b8f5c8..e056779ea0dd0a31191b628f82724298efaf50ff 100644 --- a/paddle/fluid/framework/channel_impl.h +++ b/paddle/fluid/framework/channel_impl.h @@ -31,7 +31,7 @@ class ChannelImpl : public paddle::framework::Channel { public: virtual bool CanSend(); virtual bool CanReceive(); - virtual bool Send(T *); + virtual void Send(T *); virtual bool Receive(T *); virtual size_t Cap() { return cap_; } virtual void Lock(); @@ -76,10 +76,9 @@ class ChannelImpl : public paddle::framework::Channel { } }; - bool send_return(bool value) { + void send_return() { send_ctr--; destructor_cond_.notify_all(); - return value; } bool recv_return(bool value) { @@ -88,6 +87,21 @@ class ChannelImpl : public paddle::framework::Channel { return value; } + std::shared_ptr get_first_message( + std::deque> &queue, ChannelAction action) { + while (!queue.empty()) { + // Check whether this message was added by Select + // If this was added by Select then execute the callback + // to check if you can execute this message. The callback + // can return false if some other case was executed in Select. + // In that case just discard this QueueMessage and process next. + std::shared_ptr m = queue.front(); + queue.pop_front(); + if (m->callback == nullptr || m->callback(action)) return m; + } + return nullptr; + } + size_t cap_; std::recursive_mutex mu_; bool closed_; @@ -118,45 +132,33 @@ bool ChannelImpl::CanReceive() { } template -bool ChannelImpl::Send(T *item) { +void ChannelImpl::Send(T *item) { send_ctr++; std::unique_lock lock{mu_}; - // If channel is closed, do nothing + // If channel is closed, throw exception if (closed_) { + send_return(); lock.unlock(); - // TODO(abhinavarora) Should panic on closed channel - return send_return(false); + PADDLE_THROW("Cannot send on closed channel"); } // If there is a receiver, directly pass the value we want // to send to the receiver, bypassing the channel buffer if any if (!recvq.empty()) { - std::shared_ptr m = recvq.front(); - recvq.pop_front(); - // Do the data transfer - // We will do this data transfer if either of the following - // cases are true - // 1. callback == nullptr // This means it was a regular channel send - // 2. callback returns true - bool do_send = true; - if (m->callback != nullptr) do_send = m->callback(ChannelAction::SEND); - if (do_send) + std::shared_ptr m = + get_first_message(recvq, ChannelAction::SEND); + + if (m != nullptr) { *(m->data) = std::move(*item); - else - // We cannot do the data transfer because - // this QueueMessage was added by Select - // and some other case was executed. - // So call the Send function again. - // We do not care about notifying other - // because they would have been notified - // by the executed select case. - return send_return(Send(item)); - - // Wake up the blocked process and unlock - m->Notify(); - lock.unlock(); - return send_return(true); + m->Notify(); + send_return(); + return; + } else { + Send(item); + send_return(); + return; + } } // Unbuffered channel will always bypass this @@ -165,9 +167,8 @@ bool ChannelImpl::Send(T *item) { if (buf_.size() < cap_) { // Copy to buffer buf_.push_back(std::move(*item)); - // Release lock and return true - lock.unlock(); - return send_return(true); + send_return(); + return; } // Block on channel, because some receiver will complete @@ -175,8 +176,12 @@ bool ChannelImpl::Send(T *item) { auto m = std::make_shared(item); sendq.push_back(m); m->Wait(lock); - // TODO(abhinavarora) Should panic on closed channel - return send_return(!m->chan_closed); + if (m->chan_closed) { + send_return(); + lock.unlock(); + PADDLE_THROW("Cannot send on closed channel"); + } + send_return(); } template @@ -186,39 +191,37 @@ bool ChannelImpl::Receive(T *item) { // If channel is closed and buffer is empty or // channel is unbuffered - if (closed_ && buf_.empty()) { - lock.unlock(); - return recv_return(false); - } + if (closed_ && buf_.empty()) return recv_return(false); // If there is a sender, directly receive the value we want - // from the sender, bypassing the channel buffer if any + // from the sender. In case of a buffered channel, read from + // buffer and move front of send queue to the buffer if (!sendq.empty()) { - std::shared_ptr m = sendq.front(); - sendq.pop_front(); - // Do the data transfer - // We will do this data transfer if either of the following - // cases are true - // 1. callback == nullptr // This means it was a regular channel send - // 2. callback returns true - bool do_receive = true; - if (m->callback != nullptr) - do_receive = m->callback(ChannelAction::RECEIVE); - if (do_receive) - *item = std::move(*(m->data)); - else - // We cannot do the data transfer because - // this QueueMessage was added by Select - // and some other case was executed. - // So call the Receive function again. - // We do not care about notifying other - // because they would have been notified - // by the executed select case. - return recv_return(Receive(item)); - - // Wake up the blocked process and unlock - m->Notify(); - lock.unlock(); + std::shared_ptr m = + get_first_message(sendq, ChannelAction::RECEIVE); + if (buf_.size() > 0) { + // Case 1 : Channel is Buffered + // Do Data transfer from front of buffer + // and add a QueueMessage to the buffer + *item = std::move(buf_.front()); + buf_.pop_front(); + // If first message from sendq is not null + // add it to the buffer and notify it + if (m != nullptr) { + // Copy to buffer + buf_.push_back(std::move(*(m->data))); + m->Notify(); + } // Ignore if there is no first message + } else { + // Case 2: Channel is Unbuffered + // Do data transfer from front of SendQ + // If front is nullptr, then recursively call itself + if (m != nullptr) { + *item = std::move(*(m->data)); + m->Notify(); + } else + return recv_return(Receive(item)); + } return recv_return(true); } @@ -227,8 +230,7 @@ bool ChannelImpl::Receive(T *item) { // Directly read from buffer *item = std::move(buf_.front()); buf_.pop_front(); - // Release lock and return true - lock.unlock(); + // return true return recv_return(true); } diff --git a/paddle/fluid/framework/channel_test.cc b/paddle/fluid/framework/channel_test.cc index 73be5cdbe2a1f5994ecee4c415e83962f50532fe..1184bfdae1940286fb72d9091ae4f23ff7f84a54 100644 --- a/paddle/fluid/framework/channel_test.cc +++ b/paddle/fluid/framework/channel_test.cc @@ -16,7 +16,6 @@ limitations under the License. */ #include #include - #include "gtest/gtest.h" using paddle::framework::Channel; @@ -37,23 +36,25 @@ TEST(Channel, ChannelCapacityTest) { delete ch; } -void RecevingOrderEqualToSendingOrder(Channel *ch) { +void RecevingOrderEqualToSendingOrder(Channel *ch, int num_items) { unsigned sum_send = 0; std::thread t([&]() { - for (int i = 0; i < 5; i++) { - EXPECT_EQ(ch->Send(&i), true); + for (int i = 0; i < num_items; i++) { + ch->Send(&i); sum_send += i; } }); - for (int i = 0; i < 5; i++) { - int recv = 999; + std::this_thread::sleep_for(std::chrono::milliseconds(200)); + for (int i = 0; i < num_items; i++) { + int recv = -1; EXPECT_EQ(ch->Receive(&recv), true); EXPECT_EQ(recv, i); } std::this_thread::sleep_for(std::chrono::milliseconds(200)); CloseChannel(ch); t.join(); - EXPECT_EQ(sum_send, 10U); + unsigned expected_sum = (num_items * (num_items - 1)) / 2; + EXPECT_EQ(sum_send, expected_sum); delete ch; } @@ -61,7 +62,7 @@ TEST(Channel, SufficientBufferSizeDoesntBlock) { const size_t buffer_size = 10; auto ch = MakeChannel(buffer_size); for (size_t i = 0; i < buffer_size; ++i) { - EXPECT_EQ(ch->Send(&i), true); // should not block + ch->Send(&i); } size_t out; @@ -82,7 +83,7 @@ void SendReceiveWithACloseChannelShouldPanic(Channel *ch) { const size_t data = 5; std::thread send_thread{[&]() { size_t i = data; - EXPECT_EQ(ch->Send(&i), true); // should not block + ch->Send(&i); // should not block }}; std::thread recv_thread{[&]() { @@ -94,12 +95,18 @@ void SendReceiveWithACloseChannelShouldPanic(Channel *ch) { send_thread.join(); recv_thread.join(); - // After closing send should return false. Receive should - // also return false as there is no data in queue. + // After closing send should panic. Receive should + // also false as there is no data in queue. CloseChannel(ch); send_thread = std::thread{[&]() { size_t i = data; - EXPECT_EQ(ch->Send(&i), false); // should return false + bool is_exception = false; + try { + ch->Send(&i); + } catch (paddle::platform::EnforceNotMet e) { + is_exception = true; + } + EXPECT_EQ(is_exception, true); }}; recv_thread = std::thread{[&]() { size_t i; @@ -129,7 +136,7 @@ TEST(Channel, ReceiveFromBufferedChannelReturnResidualValuesTest) { auto ch = MakeChannel(buffer_size); for (size_t i = 0; i < buffer_size; ++i) { - EXPECT_EQ(ch->Send(&i), true); // sending should not block + ch->Send(&i); // sending should not block } size_t out; @@ -160,9 +167,16 @@ TEST(Channel, ConcurrentSendNonConcurrentReceiveWithSufficientBufferSize) { // Try to write more than buffer size. for (size_t i = 0; i < 2 * buffer_size; ++i) { if (i < buffer_size) - EXPECT_EQ(ch->Send(&i), true); // should block after 10 iterations - else - EXPECT_EQ(ch->Send(&i), false); + ch->Send(&i); // should block after 10 iterations + else { + bool is_exception = false; + try { + ch->Send(&i); + } catch (paddle::platform::EnforceNotMet e) { + is_exception = true; + } + EXPECT_EQ(is_exception, true); + } } }); std::this_thread::sleep_for(std::chrono::milliseconds(200)); // wait 0.2 sec @@ -173,12 +187,28 @@ TEST(Channel, ConcurrentSendNonConcurrentReceiveWithSufficientBufferSize) { TEST(Channel, RecevingOrderEqualToSendingOrderWithUnBufferedChannel) { auto ch = MakeChannel(0); - RecevingOrderEqualToSendingOrder(ch); + RecevingOrderEqualToSendingOrder(ch, 20); +} + +TEST(Channel, RecevingOrderEqualToSendingOrderWithBufferedChannel1) { + // Test that Receive Order is same as Send Order when number of items + // sent is less than size of buffer + auto ch = MakeChannel(10); + RecevingOrderEqualToSendingOrder(ch, 5); } -TEST(Channel, RecevingOrderEqualToSendingOrderWithBufferedChannel) { +TEST(Channel, RecevingOrderEqualToSendingOrderWithBufferedChannel2) { + // Test that Receive Order is same as Send Order when number of items + // sent is equal to size of buffer auto ch = MakeChannel(10); - RecevingOrderEqualToSendingOrder(ch); + RecevingOrderEqualToSendingOrder(ch, 10); +} + +TEST(Channel, RecevingOrderEqualToSendingOrderWithBufferedChannel3) { + // Test that Receive Order is same as Send Order when number of items + // sent is greater than the size of buffer + auto ch = MakeChannel(10); + RecevingOrderEqualToSendingOrder(ch, 20); } void ChannelCloseUnblocksReceiversTest(Channel *ch) { @@ -231,7 +261,13 @@ void ChannelCloseUnblocksSendersTest(Channel *ch, bool isBuffered) { t[i] = std::thread( [&](bool *ended, bool *success) { int data = 10; - *success = ch->Send(&data); + bool is_exception = false; + try { + ch->Send(&data); + } catch (paddle::platform::EnforceNotMet e) { + is_exception = true; + } + *success = !is_exception; *ended = true; }, &thread_ended[i], &send_success[i]); @@ -316,8 +352,11 @@ TEST(Channel, UnbufferedLessReceiveMoreSendTest) { // Try to send more number of times // than receivers for (int i = 0; i < 4; i++) { - ch->Send(&i); - sum_send += i; + try { + ch->Send(&i); + sum_send += i; + } catch (paddle::platform::EnforceNotMet e) { + } } }); for (int i = 0; i < 3; i++) { @@ -382,7 +421,13 @@ void ChannelDestroyUnblockSenders(Channel *ch, bool isBuffered) { t[i] = std::thread( [&](bool *ended, bool *success) { int data = 10; - *success = ch->Send(&data); + bool is_exception = false; + try { + ch->Send(&data); + } catch (paddle::platform::EnforceNotMet e) { + is_exception = true; + } + *success = !is_exception; *ended = true; }, &thread_ended[i], &send_success[i]); @@ -508,7 +553,7 @@ void ChannelHolderSendReceive(ChannelHolder *ch) { unsigned sum_send = 0; std::thread t([&]() { for (int i = 0; i < 5; i++) { - EXPECT_EQ(ch->Send(&i), true); + ch->Send(&i); sum_send += i; } }); @@ -541,8 +586,22 @@ TEST(ChannelHolder, ChannelUninitializedTest) { ChannelHolder *ch = new ChannelHolder(); EXPECT_EQ(ch->IsInitialized(), false); int i = 10; - EXPECT_EQ(ch->Send(&i), false); - EXPECT_EQ(ch->Receive(&i), false); + bool send_exception = false; + try { + ch->Send(&i); + } catch (paddle::platform::EnforceNotMet e) { + send_exception = true; + } + EXPECT_EQ(send_exception, true); + + bool recv_exception = false; + try { + ch->Receive(&i); + } catch (paddle::platform::EnforceNotMet e) { + recv_exception = true; + } + EXPECT_EQ(recv_exception, true); + bool is_exception = false; try { ch->Type(); @@ -669,7 +728,13 @@ void ChannelHolderCloseUnblocksSendersTest(ChannelHolder *ch, bool isBuffered) { t[i] = std::thread( [&](bool *ended, bool *success) { int data = 10; - *success = ch->Send(&data); + bool is_exception = false; + try { + ch->Send(&data); + } catch (paddle::platform::EnforceNotMet e) { + is_exception = true; + } + *success = !is_exception; *ended = true; }, &thread_ended[i], &send_success[i]); @@ -760,7 +825,13 @@ void ChannelHolderDestroyUnblockSenders(ChannelHolder *ch, bool isBuffered) { t[i] = std::thread( [&](bool *ended, bool *success) { int data = 10; - *success = ch->Send(&data); + bool is_exception = false; + try { + ch->Send(&data); + } catch (paddle::platform::EnforceNotMet e) { + is_exception = true; + } + *success = !is_exception; *ended = true; }, &thread_ended[i], &send_success[i]); diff --git a/paddle/fluid/framework/concurrency_test.cc b/paddle/fluid/framework/concurrency_test.cc index 25152054eb8452a9667bd65b4441665476c1d46d..e98e9d94bf71fe9ac226ab3ad7f587b37a5c6e33 100644 --- a/paddle/fluid/framework/concurrency_test.cc +++ b/paddle/fluid/framework/concurrency_test.cc @@ -150,8 +150,9 @@ void AddFibonacciSelect(Scope *scope, p::CPUPlace *place, ProgramDesc *program, // Select block AddOp("select", {{"X", {dataChanName, quitChanName}}, {"case_to_execute", {"caseToExecute"}}}, - {}, {{"sub_block", casesBlock}, - {"cases", std::vector{case0Config, case1Config}}}, + {{"Out", {}}}, + {{"sub_block", casesBlock}, + {"cases", std::vector{case0Config, case1Config}}}, whileBlock); scope->Var("stepScopes"); @@ -209,9 +210,8 @@ TEST(Concurrency, Go_Op) { executor.Run(program, &scope, 0, true, true); - // After we call executor.run, the Go operator should do a channel_send to set - // the - // "result" variable to 99 + // After we call executor.run, the Go operator should do a channel_send to + // set the "result" variable to 99. auto *finalData = tensor.data(); EXPECT_EQ(finalData[0], 99); } diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..bf1a705ef50b663efa53393ead1f81fd6bcf8c48 --- /dev/null +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -0,0 +1,21 @@ +cc_library(var_handle SRCS var_handle.cc DEPS place) +cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context) +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) +nv_library(nccl_all_reduce_op_handle SRCS nccl_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory + dynload_cuda) +cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry) + +cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base) +cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph) + +if(WITH_GPU) + set(multi_devices_graph_builder_deps nccl_all_reduce_op_handle) +else() + set(multi_devices_graph_builder_deps) +endif() +cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle + scale_loss_grad_op_handle ${multi_devices_graph_builder_deps}) +cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph) +cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope + simple_threadpool device_context) diff --git a/paddle/fluid/framework/details/computation_op_handle.cc b/paddle/fluid/framework/details/computation_op_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..7a1b40c0b60a788b1f0a70e688f8fcbe427ad076 --- /dev/null +++ b/paddle/fluid/framework/details/computation_op_handle.cc @@ -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. + +#include "paddle/fluid/framework/details/computation_op_handle.h" + +namespace paddle { +namespace framework { +namespace details { +ComputationOpHandle::ComputationOpHandle(const OpDesc &op_desc, Scope *scope, + platform::Place place) + : op_(framework::OpRegistry::CreateOp(op_desc)), + scope_(scope), + place_(place) {} + +void ComputationOpHandle::RunImpl() { + auto *cur_ctx = dev_ctxes_[place_]; + for (auto *in : inputs_) { + bool need_wait = + in->generated_op_ && in->generated_op_->dev_ctxes_[place_] != cur_ctx; + if (need_wait) { + in->generated_op_->Wait(cur_ctx); + } + } + + op_->Run(*scope_->FindVar("@TMP_SCOPE@")->Get(), place_); +} + +std::string ComputationOpHandle::Name() const { return op_->Type(); } +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/computation_op_handle.h b/paddle/fluid/framework/details/computation_op_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..d6d2d731ca80a0fbc0a2a34027b5b7c3c1977c07 --- /dev/null +++ b/paddle/fluid/framework/details/computation_op_handle.h @@ -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. + +#pragma once + +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace framework { +namespace details { +struct ComputationOpHandle : public OpHandleBase { + std::unique_ptr op_; + Scope *scope_; + platform::Place place_; + + ComputationOpHandle(const OpDesc &op_desc, Scope *scope, + platform::Place place); + + std::string Name() const override; + + protected: + void RunImpl() override; +}; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fetch_op_handle.cc b/paddle/fluid/framework/details/fetch_op_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..9180903b864d03e59f55f41410b2240fa4199496 --- /dev/null +++ b/paddle/fluid/framework/details/fetch_op_handle.cc @@ -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. + +#include "paddle/fluid/framework/details/fetch_op_handle.h" + +namespace paddle { +namespace framework { +namespace details { + +FetchOpHandle::FetchOpHandle(FeedFetchList *data, size_t offset, + std::vector *local_scopes) + : data_(data), offset_(offset), local_scopes_(local_scopes) {} + +FetchOpHandle::~FetchOpHandle() { + for (auto *input_var : inputs_) { + input_var->pending_ops_.erase(this); + } +} + +void FetchOpHandle::Wait(platform::DeviceContext *waited_dev) { + PADDLE_THROW("Nobody should wait FetchOp. Unexpceted Error"); +} + +void FetchOpHandle::WaitAndMergeCPUTensors() const { + std::vector tensors_ptr; + tensors_ptr.reserve(tensors_.size()); + for (auto &t : tensors_) { + tensors_ptr.emplace_back(&t); + } + data_->at(offset_).MergeLoDTensor(tensors_ptr, platform::CPUPlace()); +} + +void FetchOpHandle::RunImpl() { + auto cpu_ctx = + platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); + for (auto *input : inputs_) { + auto *var = static_cast(input); + var->generated_op_->Wait(cpu_ctx); + } + + tensors_.resize(inputs_.size()); + auto *var = static_cast(inputs_[0]); + auto &var_name = var->name_; + platform::CPUPlace cpu; + auto &scopes = *local_scopes_; + + for (size_t i = 0; i < scopes.size(); ++i) { + auto &scope = scopes[i]; + auto &t = scope->FindVar(var_name)->Get(); + if (platform::is_gpu_place(var->place_)) { +#ifdef PADDLE_WITH_CUDA + TensorCopy(t, cpu, *dev_ctxes_[t.place()], &tensors_[i]); + dev_ctxes_[t.place()]->Wait(); +#endif + } else { + tensors_[i].ShareDataWith(t); + tensors_[i].set_lod(t.lod()); + } + } + + this->WaitAndMergeCPUTensors(); +} + +std::string FetchOpHandle::Name() const { return "Fetch"; } + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fetch_op_handle.h b/paddle/fluid/framework/details/fetch_op_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..904b2d669f8b156b99197afb0155380d1170a68b --- /dev/null +++ b/paddle/fluid/framework/details/fetch_op_handle.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 "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/feed_fetch_type.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace framework { +namespace details { + +struct FetchOpHandle : public OpHandleBase { + FeedFetchList *data_; + size_t offset_; + std::vector *local_scopes_; + std::vector tensors_; + + FetchOpHandle(FeedFetchList *data, size_t offset, + std::vector *local_scopes); + + ~FetchOpHandle(); + + void Wait(platform::DeviceContext *waited_dev) override; + + void WaitAndMergeCPUTensors() const; + + std::string Name() const override; + + protected: + void RunImpl() override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.cc b/paddle/fluid/framework/details/multi_devices_graph_builder.cc new file mode 100644 index 0000000000000000000000000000000000000000..a1b913a863cc1853ea3a786d22e6e8baa8c98a02 --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.cc @@ -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. + +#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" +#include "paddle/fluid/framework/details/computation_op_handle.h" +#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h" +#include "paddle/fluid/framework/scope.h" + +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h" +#endif + +namespace paddle { +namespace framework { +namespace details { + +#ifdef PADDLE_WITH_CUDA +MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( + const std::vector &places, + const std::string &loss_var_name, + const std::unordered_set ¶ms, + const std::vector &local_scopes, + platform::NCCLContextMap *nccl_ctxs) + : loss_var_name_(loss_var_name), + places_(places), + local_scopes_(local_scopes), + nccl_ctxs_(nccl_ctxs) { +#else +MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder( + const std::vector &places, + const std::string &loss_var_name, + const std::unordered_set ¶ms, + const std::vector &local_scopes) + : loss_var_name_(loss_var_name), + places_(places), + local_scopes_(local_scopes) { +#endif + for (auto &p : params) { + grad_names_.insert(GradVarName(p)); + } +} + +std::unique_ptr MultiDevSSAGraphBuilder::Build( + const ProgramDesc &program) const { + auto graph = new SSAGraph(); + SSAGraph &result = *graph; + result.vars_.resize(places_.size()); + + bool is_forwarding = true; + for (auto *op : program.Block(0).AllOps()) { + bool change_forward = false; + if (!is_forwarding) { + // FIXME(yy): Do not hard code like this + if (op->OutputArgumentNames().size() == 1 && + op->OutputArgumentNames()[0] == GradVarName(loss_var_name_)) { + continue; // Drop fill 1. for backward coeff; + } + } + + for (size_t i = 0; i < places_.size(); ++i) { + auto &p = places_[i]; + auto *s = local_scopes_[i]; + + result.ops_.emplace_back(new ComputationOpHandle(*op, s, p)); + auto *op_handle = result.ops_.back().get(); + op_handle->dev_ctxes_[p] = const_cast( + platform::DeviceContextPool::Instance().Get(p)); + + auto var_names = op->InputArgumentNames(); + + for (auto &each_var_name : var_names) { + VarHandle *var = + CreateOrGetLatestVarHandle(&result, each_var_name, p, i); + op_handle->AddInput(var); + } + var_names = op->OutputArgumentNames(); + + for (auto &each_var_name : var_names) { + CreateOpOutput(&result, op_handle, each_var_name, p, i); + } + + if (is_forwarding) { + if (var_names.size() == 1 && var_names[0] == loss_var_name_) { +// Insert ScaleCost OpHandle +#ifdef PADDLE_WITH_CUDA + auto *communication_dev_ctx = nccl_ctxs_->DevCtx(p); +#else + auto *communication_dev_ctx = + platform::DeviceContextPool::Instance().Get(platform::CPUPlace()); +#endif + + op_handle = new ScaleLossGradOpHandle(local_scopes_.size(), s, p, + communication_dev_ctx); + result.ops_.emplace_back(op_handle); + + // FIXME: Currently ScaleLossGradOp only use device_count as scale + // factor. So it does not depend on any other operators. + // VarHandle *loss = GetVarHandle(loss_var_name, place); + // loss->pending_ops_.emplace_back(op_handle); + // op_handle->inputs_.emplace_back(loss); + + CreateOpOutput(&result, op_handle, GradVarName(loss_var_name_), p, i); + change_forward = true; + } + } + } + + if (change_forward) { + is_forwarding = false; + } + + if (!is_forwarding) { + auto var_names = op->OutputArgumentNames(); + for (auto &og : var_names) { + if (grad_names_.count(og) != 0) { // is param grad + // Insert NCCL AllReduce Op +#ifdef PADDLE_WITH_CUDA + result.ops_.emplace_back( + new NCCLAllReduceOpHandle(local_scopes_, places_, *nccl_ctxs_)); + auto *op_handle = result.ops_.back().get(); + + for (size_t i = 0; i < places_.size(); ++i) { + auto &p = places_[i]; + auto &vars = result.vars_[i][og]; + + if (vars.empty()) { // This device has no data. continue. + continue; + } + auto *prev_grad = &vars[vars.size() - 1]; + op_handle->AddInput(prev_grad); + + auto &var = vars[vars.size()]; + var.place_ = p; + var.name_ = og; + var.version_ = vars.size() - 1; + + op_handle->AddOutput(&var); + } +#else + PADDLE_ENFORCE("Not implemented"); +#endif + } + } + } + } + + /* + Dependency graph has been constructed. However, there are still data + harzaeds need to be handled. + */ + PolishGraphToSupportDataHazards(&result); + + if (VLOG_IS_ON(10)) { + std::ostringstream sout; + PrintGraphviz(*graph, sout); + VLOG(10) << sout.str(); + } + + return std::unique_ptr(graph); +} // namespace details +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/multi_devices_graph_builder.h b/paddle/fluid/framework/details/multi_devices_graph_builder.h new file mode 100644 index 0000000000000000000000000000000000000000..d3c8e582cf2cdf26198822e4bd2602883622df21 --- /dev/null +++ b/paddle/fluid/framework/details/multi_devices_graph_builder.h @@ -0,0 +1,56 @@ +// 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/details/ssa_graph_builder.h" + +namespace paddle { +namespace platform { +class NCCLContextMap; +} + +namespace framework { +class Scope; +namespace details { +class MultiDevSSAGraphBuilder : public SSAGraphBuilder { + public: +#ifdef PADDLE_WITH_CUDA + MultiDevSSAGraphBuilder(const std::vector &places, + const std::string &loss_var_name, + const std::unordered_set ¶ms, + const std::vector &local_scopes, + platform::NCCLContextMap *nccl_ctxs); +#else + MultiDevSSAGraphBuilder(const std::vector &places, + const std::string &loss_var_name, + const std::unordered_set ¶ms, + const std::vector &local_scopes); +#endif + + std::unique_ptr Build(const ProgramDesc &program) const override; + + private: + std::string loss_var_name_; + const std::vector &places_; + const std::vector &local_scopes_; + std::unordered_set grad_names_; + +#ifdef PADDLE_WITH_CUDA + platform::NCCLContextMap *nccl_ctxs_; +#endif +}; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..5ddf331cfca39a4e81a42d9ff8efd5af7bcf6829 --- /dev/null +++ b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.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. + +#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h" + +namespace paddle { +namespace framework { +namespace details { +NCCLAllReduceOpHandle::NCCLAllReduceOpHandle( + const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap &ctxs) + : local_scopes_(local_scopes), places_(places), nccl_ctxs_(ctxs) { + for (auto &p : places_) { + this->dev_ctxes_[p] = nccl_ctxs_.DevCtx(p); + } +} + +void NCCLAllReduceOpHandle::RunImpl() { + if (inputs_.size() == 1) { + return; // No need to all reduce when GPU count = 1; + } else { + // Wait input done + for (auto *in : inputs_) { + auto &p = static_cast(in)->place_; + in->generated_op_->Wait(dev_ctxes_[p]); + } + + auto &var_name = static_cast(this->inputs_[0])->name_; + int dtype = -1; + size_t numel = 0; + + std::vector> all_reduce_calls; + + for (size_t i = 0; i < local_scopes_.size(); ++i) { + auto &p = places_[i]; + auto *s = local_scopes_[i]; + int dev_id = boost::get(p).device; + + auto &lod_tensor = s->FindVar(var_name)->Get(); + void *buffer = const_cast(lod_tensor.data()); + + if (dtype == -1) { + dtype = platform::ToNCCLDataType(lod_tensor.type()); + } + + if (numel == 0) { + numel = static_cast(lod_tensor.numel()); + } + + auto &nccl_ctx = nccl_ctxs_.at(dev_id); + auto stream = nccl_ctx.stream(); + auto comm = nccl_ctx.comm_; + all_reduce_calls.emplace_back([=] { + PADDLE_ENFORCE(platform::dynload::ncclAllReduce( + buffer, buffer, numel, static_cast(dtype), ncclSum, + comm, stream)); + }); + } + + platform::NCCLGroupGuard guard; + for (auto &call : all_reduce_calls) { + call(); + } + } +} + +std::string NCCLAllReduceOpHandle::Name() const { return "NCCL AllReduce"; } +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/nccl_all_reduce_op_handle.h b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..045070bb6a97e90600cd24d9f43cd2a10a4bc1f5 --- /dev/null +++ b/paddle/fluid/framework/details/nccl_all_reduce_op_handle.h @@ -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. + +#pragma once + +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/platform/nccl_helper.h" + +namespace paddle { +namespace framework { +namespace details { + +struct NCCLAllReduceOpHandle : public OpHandleBase { + const std::vector &local_scopes_; + const std::vector &places_; + const platform::NCCLContextMap &nccl_ctxs_; + + NCCLAllReduceOpHandle(const std::vector &local_scopes, + const std::vector &places, + const platform::NCCLContextMap &ctxs); + + std::string Name() const override; + + protected: + void RunImpl() override; +}; + +} // 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 new file mode 100644 index 0000000000000000000000000000000000000000..e4194a7442f677ec8970dbc387bb01ebbbf579f1 --- /dev/null +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -0,0 +1,102 @@ +// 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_handle_base.h" + +namespace paddle { +namespace framework { +namespace details { +std::string OpHandleBase::DebugString() const { + std::stringstream ss; + ss << "("; + for (auto *var : inputs_) { + ss << var->DebugString() << ", "; + } + ss << ") --> ("; + for (auto *var : outputs_) { + ss << var->DebugString() << ", "; + } + ss << ")\n"; + return ss.str(); +} + +OpHandleBase::~OpHandleBase() { +#ifdef PADDLE_WITH_CUDA + for (auto &ev : events_) { + PADDLE_ENFORCE(cudaEventDestroy(ev.second)); + } +#endif +} + +void OpHandleBase::Run(bool use_event) { +#ifdef PADDLE_WITH_CUDA + if (events_.empty() && use_event) { + for (auto &p : dev_ctxes_) { + int dev_id = boost::get(p.first).device; + PADDLE_ENFORCE(cudaSetDevice(dev_id)); + PADDLE_ENFORCE( + cudaEventCreateWithFlags(&events_[dev_id], cudaEventDisableTiming)); + } + } +#else + PADDLE_ENFORCE(!use_event); +#endif + + RunImpl(); + +#ifdef PADDLE_WITH_CUDA + if (use_event) { + for (auto &p : dev_ctxes_) { + int dev_id = boost::get(p.first).device; + auto stream = + static_cast(p.second)->stream(); + PADDLE_ENFORCE(cudaEventRecord(events_.at(dev_id), stream)); + } + } +#endif +} + +void OpHandleBase::Wait(platform::DeviceContext *waited_dev) { +#ifdef PADDLE_WITH_CUDA + if (platform::is_cpu_place(waited_dev->GetPlace()) || events_.empty()) { + for (auto &dev_ctx : dev_ctxes_) { + dev_ctx.second->Wait(); + } + } else { + auto stream = + static_cast(waited_dev)->stream(); + for (auto &ev : events_) { + PADDLE_ENFORCE(cudaStreamWaitEvent(stream, ev.second, 0)); + } + } +#else + for (auto &dev_ctx : dev_ctxes_) { + dev_ctx.second->Wait(); + } +#endif +} + +void OpHandleBase::AddInput(VarHandleBase *in) { + this->inputs_.emplace_back(in); + in->pending_ops_.insert(this); +} + +void OpHandleBase::AddOutput(VarHandleBase *out) { + outputs_.emplace_back(out); + out->generated_op_ = this; +} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h new file mode 100644 index 0000000000000000000000000000000000000000..71672fd24c65ee654fb9f703ea5808c31ee8fbb0 --- /dev/null +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -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. + +#pragma once + +#include "paddle/fluid/framework/details/var_handle.h" +#include "paddle/fluid/platform/device_context.h" +#include "paddle/fluid/platform/macros.h" + +namespace paddle { +namespace framework { +namespace details { + +class OpHandleBase { + private: + DISABLE_COPY_AND_ASSIGN(OpHandleBase); + + public: + std::vector inputs_; + std::vector outputs_; + std::unordered_map + dev_ctxes_; + +#ifdef PADDLE_WITH_CUDA + std::unordered_map events_; +#endif + + OpHandleBase() {} + + std::string DebugString() const; + + virtual std::string Name() const = 0; + + virtual ~OpHandleBase(); + + void Run(bool use_event); + + virtual void Wait(platform::DeviceContext *waited_dev); + + void AddInput(VarHandleBase *in); + + void AddOutput(VarHandleBase *out); + + protected: + virtual void RunImpl() = 0; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc b/paddle/fluid/framework/details/scale_loss_grad_op_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..0a6f6129b812ca84db7573957b1ee0a32c1ef5c4 --- /dev/null +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.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 "paddle/fluid/framework/details/scale_loss_grad_op_handle.h" + +namespace paddle { +namespace framework { +namespace details { +ScaleLossGradOpHandle::ScaleLossGradOpHandle(size_t num_dev, Scope *scope, + platform::Place place, + platform::DeviceContext *dev_ctx) + : coeff_(static_cast(1.0 / num_dev)), scope_(scope), place_(place) { + dev_ctxes_[place_] = dev_ctx; +} + +ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {} + +void ScaleLossGradOpHandle::RunImpl() { + std::string var_name = static_cast(this->outputs_[0])->name_; + + float *tmp = + scope_->FindVar(var_name)->GetMutable()->mutable_data( + make_ddim({1}), place_); + + if (platform::is_cpu_place(place_)) { + *tmp = coeff_; + } else { +#ifdef PADDLE_WITH_CUDA + auto stream = + static_cast(this->dev_ctxes_[place_]) + ->stream(); + memory::Copy(boost::get(place_), tmp, + platform::CPUPlace(), &coeff_, sizeof(float), stream); +#endif + } +} + +std::string ScaleLossGradOpHandle::Name() const { return "Scale LossGrad"; } +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/scale_loss_grad_op_handle.h b/paddle/fluid/framework/details/scale_loss_grad_op_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..ab7353a4fc56bebfe04696efd838dc4559218058 --- /dev/null +++ b/paddle/fluid/framework/details/scale_loss_grad_op_handle.h @@ -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. + +#pragma once + +#include "paddle/fluid/framework/details/op_handle_base.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" + +namespace paddle { +namespace framework { +namespace details { + +struct ScaleLossGradOpHandle : public OpHandleBase { + float coeff_; + Scope *scope_; + platform::Place place_; + + ScaleLossGradOpHandle(size_t num_dev, Scope *scope, platform::Place place, + platform::DeviceContext *context); + + ~ScaleLossGradOpHandle() final; + + std::string Name() const override; + + protected: + void RunImpl() override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph.cc b/paddle/fluid/framework/details/ssa_graph.cc new file mode 100644 index 0000000000000000000000000000000000000000..1b8c889449059c563ea39f86250075ac2537cdbe --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph.cc @@ -0,0 +1,15 @@ +// 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/ssa_graph.h" diff --git a/paddle/fluid/framework/details/ssa_graph.h b/paddle/fluid/framework/details/ssa_graph.h new file mode 100644 index 0000000000000000000000000000000000000000..ac3e2d86993aee31b79f4481c4d5a47cd9cdf5b4 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph.h @@ -0,0 +1,35 @@ +// 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/details/op_handle_base.h" +#include "paddle/fluid/framework/details/var_handle.h" + +namespace paddle { +namespace framework { +namespace details { + +struct SSAGraph { + std::vector>> vars_; + // aux variables to represent dependency. Useful to resolve data hazard. + std::unordered_set> dep_vars_; + std::vector> ops_; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_builder.cc b/paddle/fluid/framework/details/ssa_graph_builder.cc new file mode 100644 index 0000000000000000000000000000000000000000..361ba6d39721eed406a30fea325b3b4508ec45d0 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_builder.cc @@ -0,0 +1,141 @@ +// 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/ssa_graph_builder.h" + +namespace paddle { +namespace framework { +namespace details { +void SSAGraphBuilder::PolishGraphToSupportDataHazards(SSAGraph *graph) { + for (auto &var_map : graph->vars_) { + for (auto &name_pair : var_map) { + if (name_pair.second.size() <= 1) { + continue; + } + auto it_new = name_pair.second.rbegin(); + auto it_old = name_pair.second.rbegin(); + ++it_old; + for (; it_old != name_pair.second.rend(); it_new = it_old, ++it_old) { + auto *write_op = it_new->second.generated_op_; + auto &read_ops = it_old->second.pending_ops_; + + for (auto *read_op : read_ops) { + // Manually add a dependency var from read_op to write_op; + if (read_op == write_op) { + // Read Write is the same op. + continue; + } + + auto *dep_var = new DummyVarHandle(); + read_op->AddOutput(dep_var); + write_op->AddInput(dep_var); + graph->dep_vars_.emplace(dep_var); + } + } + } + } +} + +VarHandle *SSAGraphBuilder::CreateOrGetLatestVarHandle( + SSAGraph *graph, const std::string &each_var_name, + const platform::Place &place, size_t place_offset) { + auto &var_holders = graph->vars_[place_offset]; + auto &var_holder = var_holders[each_var_name]; + VarHandle *var = nullptr; + if (var_holder.empty()) { + auto &init_var = var_holder[0]; + init_var.place_ = place; + init_var.name_ = each_var_name; + init_var.generated_op_ = nullptr; + init_var.version_ = 0; + var = &init_var; + } else { + var = &var_holder.rbegin()->second; + } + return var; +} + +void SSAGraphBuilder::CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle, + const std::string &each_var_name, + const platform::Place &place, + size_t place_offset) { + auto &vars = graph->vars_[place_offset][each_var_name]; + size_t version = vars.size(); + auto &var = vars[version]; + var.version_ = version; + var.name_ = each_var_name; + var.place_ = place; + op_handle->AddOutput(&var); +} + +template +void IterAllVar(const SSAGraph &graph, Callback callback) { + for (auto &each : graph.vars_) { + for (auto &pair1 : each) { + for (auto &pair2 : pair1.second) { + callback(pair2.second); + } + } + } + + for (auto &var : graph.dep_vars_) { + callback(*var); + } +} + +void SSAGraphBuilder::PrintGraphviz(const SSAGraph &graph, std::ostream &sout) { + size_t var_id = 0; + std::unordered_map vars; + + sout << "digraph G {\n"; + + IterAllVar(graph, [&](const VarHandleBase &var) { + auto *var_ptr = &var; + auto *var_handle_ptr = dynamic_cast(var_ptr); + auto *dummy_ptr = dynamic_cast(var_ptr); + + size_t cur_var_id = var_id++; + vars[var_ptr] = cur_var_id; + + if (var_handle_ptr) { + sout << "var_" << cur_var_id << " [label=\"" << var_handle_ptr->name_ + << "\\n" + << var_handle_ptr->place_ << "\\n" + << var_handle_ptr->version_ << "\"]" << std::endl; + } else if (dummy_ptr) { + sout << "var_" << cur_var_id << " [label=\"dummy\"]" << std::endl; + } + }); + + size_t op_id = 0; + for (auto &op : graph.ops_) { + std::string op_name = "op_" + std::to_string(op_id++); + sout << op_name << " [label=\"" << op->Name() << "\", shape=rect]" + << std::endl; + for (auto in : op->inputs_) { + std::string var_name = "var_" + std::to_string(vars[in]); + sout << var_name << " -> " << op_name << std::endl; + } + + for (auto out : op->outputs_) { + std::string var_name = "var_" + std::to_string(vars[out]); + sout << op_name << " -> " << var_name << std::endl; + } + } + + sout << "}\n"; +} +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_builder.h b/paddle/fluid/framework/details/ssa_graph_builder.h new file mode 100644 index 0000000000000000000000000000000000000000..bf20e7164a100718c1dcfe3ef971cfff60bbbaa2 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_builder.h @@ -0,0 +1,59 @@ +// 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/details/ssa_graph.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/platform/place.h" + +#include +#include + +namespace paddle { +namespace framework { +namespace details { + +class SSAGraphBuilder { + public: + SSAGraphBuilder() {} + virtual ~SSAGraphBuilder() {} + virtual std::unique_ptr Build(const ProgramDesc &program) const = 0; + + DISABLE_COPY_AND_ASSIGN(SSAGraphBuilder); + + protected: + /** + * We only handle write after read(WAR), since it should not have a write + * after write in program. If there are write after write operators, we need + * prune them. + * + * https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR) + */ + static void PolishGraphToSupportDataHazards(SSAGraph *graph); + + static VarHandle *CreateOrGetLatestVarHandle(SSAGraph *graph, + const std::string &each_var_name, + const platform::Place &place, + size_t place_offset); + + static void CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle, + const std::string &each_var_name, + const platform::Place &place, size_t place_offset); + + static void PrintGraphviz(const SSAGraph &graph, std::ostream &sout); +}; +} // 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 new file mode 100644 index 0000000000000000000000000000000000000000..8da6ca889b89999e0f6f974503cea476c9de97f3 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_executor.cc @@ -0,0 +1,28 @@ +// 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/ssa_graph_executor.h" + +namespace paddle { +namespace framework { +namespace details { + +SSAGraphExecutor::SSAGraphExecutor(std::unique_ptr &&graph) + : graph_(std::move(graph)) {} + +SSAGraphExecutor::~SSAGraphExecutor() {} + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/ssa_graph_executor.h b/paddle/fluid/framework/details/ssa_graph_executor.h new file mode 100644 index 0000000000000000000000000000000000000000..3b818b1a45b56351e34f9e52ec22b6d02a0c1591 --- /dev/null +++ b/paddle/fluid/framework/details/ssa_graph_executor.h @@ -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. + +#pragma once + +#include +#include "paddle/fluid/framework/details/ssa_graph.h" +#include "paddle/fluid/framework/feed_fetch_type.h" + +namespace paddle { +namespace framework { +namespace details { + +class SSAGraphExecutor { + DISABLE_COPY_AND_ASSIGN(SSAGraphExecutor); + + public: + // Steal graph inside + explicit SSAGraphExecutor(std::unique_ptr &&graph); + + virtual ~SSAGraphExecutor(); + + virtual FeedFetchList Run(const std::vector &fetch_tensors) = 0; + + protected: + std::unique_ptr graph_; +}; +} // 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 new file mode 100644 index 0000000000000000000000000000000000000000..3f8655147b688239509dea98925df310a46cbef8 --- /dev/null +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc @@ -0,0 +1,205 @@ +// 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/threaded_ssa_graph_executor.h" + +#include "paddle/fluid/framework/details/fetch_op_handle.h" + +namespace paddle { +namespace framework { +namespace details { +ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor( + size_t num_threads, bool use_event, + const std::vector &local_scopes, + const std::vector &places, + std::unique_ptr &&graph) + : SSAGraphExecutor(std::move(graph)), + pool_(num_threads >= 2 ? new ::ThreadPool(num_threads) : nullptr), + local_scopes_(local_scopes), + places_(places), + fetch_ctxs_(places), + use_event_(use_event) {} + +FeedFetchList ThreadedSSAGraphExecutor::Run( + const std::vector &fetch_tensors) { + std::unordered_map pending_ops; + std::unordered_set pending_vars; + + BlockingQueue ready_vars; + + std::unordered_set ready_ops; + + auto InsertPendingVar = [&pending_vars, &ready_vars](VarHandleBase &var) { + pending_vars.insert(&var); + if (var.generated_op_ == nullptr) { + ready_vars.Push(&var); + } + }; + + auto InsertPendingOp = [&pending_ops](OpHandleBase &op_instance) { + pending_ops.insert({&op_instance, op_instance.inputs_.size()}); + }; + + // Transform SSAGraph to pending_ops & pending_vars + for (auto &var_map : graph_->vars_) { + for (auto &name_pair : var_map) { + for (auto &version_pair : name_pair.second) { + InsertPendingVar(version_pair.second); + } + } + } + for (auto &var : graph_->dep_vars_) { + InsertPendingVar(*var); + } + + for (auto &op : graph_->ops_) { + if (op->inputs_.empty()) { // Special case, Op has no input. + ready_ops.insert(op.get()); + } else { + InsertPendingOp(*op); + } + } + + // Step 2. Insert FetchOps + std::vector> fetch_ops; + std::vector dummy_vars; + FeedFetchList fetch_data(fetch_tensors.size()); + + std::unordered_map> fetched_vars; + + for (auto &fetch_var_name : fetch_tensors) { + for (auto &var_map : graph_->vars_) { + auto it = var_map.find(fetch_var_name); + if (it != var_map.end()) { + fetched_vars[fetch_var_name].push_back(&it->second.rbegin()->second); + } + } + } + + for (size_t i = 0; i < fetch_tensors.size(); ++i) { + auto &var_name = fetch_tensors[i]; + auto &vars = fetched_vars.at(var_name); + auto *op = new FetchOpHandle(&fetch_data, i, &local_scopes_); + fetch_ops.emplace_back(op); + + // FIXME: Use new device context + for (auto &p : places_) { + op->dev_ctxes_[p] = fetch_ctxs_.Get(p); + } + + for (auto *var : vars) { + op->AddInput(var); + } + InsertPendingOp(*op); + } + + auto run_all_ready_ops = [&] { + for (auto *op : ready_ops) { + RunOp(ready_vars, op); + } + ready_ops.clear(); + }; + + // Create local scopes. + for (auto &scope : local_scopes_) { + auto &local_scope = scope->NewScope(); + *scope->Var("@TMP_SCOPE@")->GetMutable() = &local_scope; + } + + // Step 3. Execution + while (!pending_vars.empty()) { + // 1. Run All Ready ops + run_all_ready_ops(); + + // 2. Find ready variable + bool timeout; + auto cur_ready_vars = ready_vars.PopAll(1000, &timeout); + + if (timeout) { + if (exception_) { + throw * exception_; + } else { + continue; + } + } + // 3. Remove the dependency of ready_var. + // Find the ready_ops after the ready_var. + for (auto ready_var : cur_ready_vars) { + pending_vars.erase(ready_var); + for (auto *op : ready_var->pending_ops_) { + auto &deps = pending_ops[op]; + --deps; + if (deps == 0) { + ready_ops.insert(op); + } + } + } + // Keep loop until all vars are ready. + } + + ++computation_count_; + + auto sync_computation = [&] { + computation_count_ = 0; + // Wait All computational streams + for (auto p : this->places_) { + platform::DeviceContextPool::Instance().Get(p)->Wait(); + } + for (auto &scope : local_scopes_) { + scope->DropKids(); + } + }; + + // Wait FetchOps. + if (!fetch_ops.empty()) { + fetch_ops.clear(); + sync_computation(); + } + + if (computation_count_ == max_async_computation) { + sync_computation(); + } + + // NOTE: the temp scope can be dropped lazily if needed. + // Drop tmp scopes; + for (auto &scope : local_scopes_) { + auto &kid = *scope->Var("@TMP_SCOPE@")->GetMutable(); + kid = nullptr; + } + + return fetch_data; +} + +void ThreadedSSAGraphExecutor::RunOp( + BlockingQueue &ready_var_q, details::OpHandleBase *op) { + auto op_run = [&ready_var_q, op, this] { + try { + VLOG(10) << op->Name() << " : " << op->DebugString(); + op->Run(use_event_); + ready_var_q.Extend(op->outputs_); + } catch (platform::EnforceNotMet ex) { + exception_.reset(new platform::EnforceNotMet(ex)); + } catch (...) { + LOG(FATAL) << "Unknown exception catched"; + } + }; + if (pool_) { + pool_->enqueue(op_run); + } else { + op_run(); + } +} +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h new file mode 100644 index 0000000000000000000000000000000000000000..2ea57ac8f96bc9c2b5c98bcd25d9ce921c3683cd --- /dev/null +++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.h @@ -0,0 +1,99 @@ +// 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 "ThreadPool.h" // ThreadPool in thrird party +#include "paddle/fluid/framework/details/ssa_graph_executor.h" + +namespace paddle { +namespace framework { +class Scope; + +namespace details { + +template +class BlockingQueue { + public: + void Push(const T &item) { + { + std::lock_guard g(mutex_); + q_.emplace_back(item); + } + cv_.notify_one(); + } + + template + void Extend(const U &items) { + { + std::lock_guard g(mutex_); + for (auto &item : items) { + q_.emplace_back(item); + } + } + cv_.notify_all(); + } + + std::deque PopAll(size_t ms, bool *timeout) { + auto time = + std::chrono::system_clock::now() + std::chrono::milliseconds(ms); + std::unique_lock lock(mutex_); + *timeout = !cv_.wait_until(lock, time, [this] { return !q_.empty(); }); + std::deque ret; + if (!*timeout) { + std::swap(ret, q_); + } + return ret; + } + + private: + std::mutex mutex_; + std::condition_variable cv_; + std::deque q_; +}; + +class ThreadedSSAGraphExecutor : public SSAGraphExecutor { + public: + ThreadedSSAGraphExecutor(size_t num_threads, bool use_event, + const std::vector &local_scopes, + const std::vector &places, + std::unique_ptr &&graph); + + // Run a SSAGraph by a thread pool + // Use topological sort algorithm + FeedFetchList Run(const std::vector &fetch_tensors) override; + + ~ThreadedSSAGraphExecutor() {} + + private: + void RunOp(BlockingQueue &ready_var_q, + details::OpHandleBase *op); + + private: + std::unique_ptr<::ThreadPool> pool_; + std::vector local_scopes_; + std::vector places_; + platform::DeviceContextPool fetch_ctxs_; + const bool use_event_; + std::unique_ptr exception_; + + size_t computation_count_{0}; + size_t max_async_computation{100}; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/var_handle.cc b/paddle/fluid/framework/details/var_handle.cc new file mode 100644 index 0000000000000000000000000000000000000000..6f00abd9473a84a77ed1a39015e2ae079e00be79 --- /dev/null +++ b/paddle/fluid/framework/details/var_handle.cc @@ -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. + +#include "paddle/fluid/framework/details/var_handle.h" + +namespace paddle { +namespace framework { +namespace details { + +VarHandleBase::~VarHandleBase() {} + +std::string VarHandle::DebugString() const { + std::stringstream ss; + ss << name_ << ":" << place_; + return ss.str(); +} + +std::string DummyVarHandle::DebugString() const { return "dummy"; } +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/var_handle.h b/paddle/fluid/framework/details/var_handle.h new file mode 100644 index 0000000000000000000000000000000000000000..569dda17c6e91d5658c4f8b9ba0b8c8fbd966832 --- /dev/null +++ b/paddle/fluid/framework/details/var_handle.h @@ -0,0 +1,64 @@ +// 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/platform/place.h" + +namespace paddle { +namespace framework { +namespace details { +class OpHandleBase; + +// VarHandleBase is the var node in the dependency graph. +// A variable can only be generated by a single operator. i.e. +// This is a single assignment graph. +struct VarHandleBase { + virtual ~VarHandleBase(); + virtual std::string DebugString() const = 0; + + // The operator who generate this variable. nullptr if the variable + // is a root node. + OpHandleBase *generated_op_; + + // Operators which depend on this variable ready. + std::unordered_set pending_ops_; +}; + +// VarHandle is actually a single version of Runtime Variable. +// Variable in Runtime mapped to many VarHandles in Graph. +// Each assignment will generate a new var handle with newer version. +// +// NOTE: runtime variables have place. +struct VarHandle : public VarHandleBase { + std::string DebugString() const override; + + // version field currently is not used, however, just store the version to + // debug easily. + size_t version_; + std::string name_; + platform::Place place_; +}; + +// Dummy Variable. It is used to represent dependencies between operators +struct DummyVarHandle : public VarHandleBase { + std::string DebugString() const override; +}; + +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 5de2b78f3ae7b50e804fefbb9a70a1a24e49adb6..c394122f3018abd84264f0de8f6ea4df3720fcaa 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -46,7 +46,7 @@ ExecutorPrepareContext::~ExecutorPrepareContext() { Executor::Executor(const platform::Place& place) : place_(place) {} -static void CreateTensor(Variable* var, proto::VarType::Type var_type) { +void InitializeVariable(Variable* var, proto::VarType::Type var_type) { if (var_type == proto::VarType::LOD_TENSOR) { var->GetMutable(); } else if (var_type == proto::VarType::SELECTED_ROWS) { @@ -110,12 +110,12 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, if (var->Persistable()) { auto* ptr = const_cast(ancestor_scope)->Var(var->Name()); - CreateTensor(ptr, var->GetType()); + InitializeVariable(ptr, var->GetType()); VLOG(3) << "Create Variable " << var->Name() << " global, which pointer is " << ptr; } else { auto* ptr = scope->Var(var->Name()); - CreateTensor(ptr, var->GetType()); + InitializeVariable(ptr, var->GetType()); VLOG(3) << "Create Variable " << var->Name() << " locally, which pointer is " << ptr; } @@ -123,7 +123,7 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, } else { for (auto& var : global_block.AllVars()) { auto* ptr = scope->Var(var->Name()); - CreateTensor(ptr, var->GetType()); + InitializeVariable(ptr, var->GetType()); VLOG(3) << "Create variable " << var->Name() << ", which pointer is " << ptr; } @@ -319,12 +319,13 @@ std::unique_ptr Executor::Prepare( void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope, bool create_local_scope, bool create_vars) { Scope* local_scope = scope; - if (create_vars && create_local_scope) { - local_scope = &scope->NewScope(); + if (create_vars) { + if (create_local_scope) { + local_scope = &scope->NewScope(); + } + CreateVariables(ctx->prog_, local_scope, ctx->block_id_); } - CreateVariables(ctx->prog_, local_scope, ctx->block_id_); - for (auto& op : ctx->ops_) { VLOG(3) << place_ << " " << op->DebugStringEx(local_scope); op->Run(*local_scope, place_); diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 75ea9a885cc3ff8d5bcc30d85f5656adec12db92..0276f4a7a624201f792c12eb56fad76eee8fc8e4 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -22,6 +22,7 @@ limitations under the License. */ namespace paddle { namespace framework { +extern void InitializeVariable(Variable* var, proto::VarType::Type var_type); struct ExecutorPrepareContext { ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id); diff --git a/paddle/fluid/framework/mixed_vector.h b/paddle/fluid/framework/mixed_vector.h index 6a6fa538718837a958b7d82c37f583f62f4bf96e..d99a15547b77a0e0d71b14bd1c798cd1485720b0 100644 --- a/paddle/fluid/framework/mixed_vector.h +++ b/paddle/fluid/framework/mixed_vector.h @@ -176,7 +176,7 @@ class Vector { // resize the vector void resize(size_t size) { - if (size + 1 < capacity()) { + if (size + 1 <= capacity()) { size_ = size; } else { MutableCPU(); diff --git a/paddle/fluid/framework/mixed_vector_test.cu b/paddle/fluid/framework/mixed_vector_test.cu index 4bf78499f2fda2d2631e05ddcbbd0bc49498af1a..d57f82510833d6a0cea7009cf1f0b49543812f8d 100644 --- a/paddle/fluid/framework/mixed_vector_test.cu +++ b/paddle/fluid/framework/mixed_vector_test.cu @@ -104,3 +104,11 @@ TEST(mixed_vector, ForEach) { for (auto& v : tmp) { } } + +TEST(mixed_vector, Reserve) { + paddle::framework::Vector vec; + vec.reserve(1); + vec.push_back(0); + vec.push_back(0); + vec.push_back(0); +} diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index b39a1164dbd9877d9f45cc6415d74f930921a42f..f6a43804ef2fd73c4a2c2c3b3dfbb90bff1c451b 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -517,6 +517,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope, // do data transform Scope& new_scope = scope.NewScope(); + std::vector inplace_vars; for (auto& var_name_item : this->Inputs()) { for (auto& var_name : var_name_item.second) { auto* var = scope.FindVar(var_name); @@ -529,10 +530,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope, auto out_var_names = OutputVars(true); if (std::find(out_var_names.begin(), out_var_names.end(), var_name) != out_var_names.end()) { - PADDLE_THROW( - "var %s is both input and output, " - "does not support transform", - var_name); + inplace_vars.push_back(var_name); } VLOG(3) << "Transform Variable " << var_name << " from " << kernel_type_for_var << " to " << expected_kernel_key; @@ -551,6 +549,13 @@ void OperatorWithKernel::RunImpl(const Scope& scope, kernel_iter->second->Compute( ExecutionContext(*this, new_scope, *new_dev_ctx)); + 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(new_scope.FindVar(var_name)); + original_tensor->ShareDataWith(*transformed_tensor); + } + /*For profiling/benchmark only*/ if (FLAGS_benchmark) { new_dev_ctx->Wait(); diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc new file mode 100644 index 0000000000000000000000000000000000000000..577eea92d217f9feb54085b4d9f071494c3c5165 --- /dev/null +++ b/paddle/fluid/framework/parallel_executor.cc @@ -0,0 +1,160 @@ +/* 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/parallel_executor.h" + +#include +#include + +#ifdef PADDLE_WITH_CUDA +#include "paddle/fluid/platform/nccl_helper.h" +#endif + +#include "paddle/fluid/framework/details/multi_devices_graph_builder.h" +#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h" + +namespace paddle { +namespace framework { + +class ParallelExecutorPrivate { + public: + explicit ParallelExecutorPrivate(const std::vector &places) + : places_(places) {} + + std::vector places_; + std::vector local_scopes_; + Scope *global_scope_; + std::unique_ptr executor_; + +#ifdef PADDLE_WITH_CUDA + std::unique_ptr nccl_ctxs_; +#endif +}; + +ParallelExecutor::ParallelExecutor( + size_t num_threads, bool use_event, + const std::vector &places, + const std::unordered_set ¶ms, + const ProgramDesc &startup_program, const ProgramDesc &main_program, + const std::string &loss_var_name, Scope *scope) + : member_(new ParallelExecutorPrivate(places)) { + member_->global_scope_ = scope; + + // Step 1. RunStartupProgram and Bcast the params to devs. + Executor exe(places[0]); + exe.Run(startup_program, scope, 0); + // Create local scopes + for (size_t i = 0; i < member_->places_.size(); ++i) { + member_->local_scopes_.push_back(&scope->NewScope()); + } + +// Bcast Parameters to all GPUs +#ifdef PADDLE_WITH_CUDA + member_->nccl_ctxs_.reset(new platform::NCCLContextMap(member_->places_)); +#endif + if (platform::is_gpu_place(places[0]) && + member_->local_scopes_.size() != 1) { // Is CUDA + BCastParamsToGPUs(startup_program); + } +// Startup Program has been run. All local scopes has correct parameters. + +// Step 2. Convert main_program to SSA form and dependency graph. Also, insert +// ncclOp +#ifdef PADDLE_WITH_CUDA + details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name, + params, member_->local_scopes_, + member_->nccl_ctxs_.get()); +#else + details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name, + params, member_->local_scopes_); +#endif + auto graph = builder.Build(main_program); + + member_->executor_.reset(new details::ThreadedSSAGraphExecutor( + num_threads, use_event, member_->local_scopes_, places, + std::move(graph))); + + // Step 3. Create vars in each scope; + for (auto *scope : member_->local_scopes_) { + for (auto *var : main_program.Block(0).AllVars()) { + if (scope->FindVar(var->Name()) != nullptr) { + continue; + } + + InitializeVariable(scope->Var(var->Name()), var->GetType()); + } + } +} + +void ParallelExecutor::BCastParamsToGPUs( + const ProgramDesc &startup_program) const { +#ifdef PADDLE_WITH_CUDA + auto *main_scope = member_->local_scopes_[0]; + + for (auto *var_desc : startup_program.Block(0).AllVars()) { + size_t idx = var_desc->Name().find("@GRAD"); + if (idx != std::string::npos) continue; + if (var_desc->GetType() == proto::VarType::LOD_TENSOR) { + auto &main_tensor = + main_scope->FindVar(var_desc->Name())->Get(); + + auto &dims = main_tensor.dims(); + + if (paddle::platform::is_gpu_place(main_tensor.place())) { + size_t numel = main_tensor.numel(); + ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type()); + platform::NCCLGroupGuard guard; + for (size_t i = 0; i < member_->places_.size(); ++i) { + auto place = member_->places_[i]; + void *buffer; + if (i == 0) { + buffer = const_cast(main_tensor.data()); + } else { + auto local_scope = member_->local_scopes_[i]; + auto *t = + local_scope->Var(var_desc->Name())->GetMutable(); + t->Resize(dims); + buffer = t->mutable_data(place, main_tensor.type()); + } + auto &nccl_ctx = member_->nccl_ctxs_->at(place); + platform::dynload::ncclBcast(buffer, numel, data_type, 0, + nccl_ctx.comm_, nccl_ctx.stream()); + } + } else { + platform::CPUPlace cpu; + for (size_t i = 1; i < member_->places_.size(); ++i) { + auto local_scope = member_->local_scopes_[i]; + auto *t = local_scope->Var(var_desc->Name())->GetMutable(); + t->Resize(dims); + t->mutable_data(cpu, main_tensor.type()); + paddle::framework::TensorCopy(main_tensor, cpu, t); + } + } + } + member_->nccl_ctxs_->WaitAll(); + } +#else + PADDLE_THROW("Not compiled with CUDA"); +#endif +} + +void ParallelExecutor::Run(const std::vector &fetch_tensors, + const std::string &fetched_var_name) { + auto fetch_data = member_->executor_->Run(fetch_tensors); + *member_->global_scope_->Var(fetched_var_name)->GetMutable() = + fetch_data; +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/parallel_executor.h b/paddle/fluid/framework/parallel_executor.h new file mode 100644 index 0000000000000000000000000000000000000000..503efa2e447b0ac70f6302aa0a89cc55e5afcb81 --- /dev/null +++ b/paddle/fluid/framework/parallel_executor.h @@ -0,0 +1,52 @@ +/* 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/framework/executor.h" +#include "paddle/fluid/framework/op_info.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/platform/device_context.h" + +namespace paddle { +namespace framework { + +class ParallelExecutorPrivate; + +class ParallelExecutor { + DISABLE_COPY_AND_ASSIGN(ParallelExecutor); + + public: + explicit ParallelExecutor(size_t num_threads, bool use_event, + const std::vector& places, + const std::unordered_set& params, + const ProgramDesc& startup_program, + const ProgramDesc& main_program, + const std::string& loss_var_name, Scope* scope); + + void Run(const std::vector& fetch_tensors, + const std::string& fetched_var_name = "fetched_var"); + + private: + ParallelExecutorPrivate* member_; + + void BCastParamsToGPUs(const ProgramDesc& startup_program) const; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/reader.cc b/paddle/fluid/framework/reader.cc index fa00c08e0d5791ee1187aed38b4d140564b7c97d..56bf00e5f91700f0cffa917aad8608caaab0a7fe 100644 --- a/paddle/fluid/framework/reader.cc +++ b/paddle/fluid/framework/reader.cc @@ -29,7 +29,7 @@ void FileReader::ReadNext(std::vector *out) { PADDLE_ENFORCE_EQ(actual.size(), expect.size()); for (int j = 0; j < actual.size(); ++j) { - PADDLE_ENFORCE(actual[i] == expect[i] || expect[i] == -1); + // PADDLE_ENFORCE(actual[i] == expect[i] || expect[i] == -1); } } } diff --git a/paddle/fluid/framework/tensor.h b/paddle/fluid/framework/tensor.h index 6f878541e6de1deec1829145b1b325ecd176a034..f7a6b5ba84ca1762bd903790aa3c0346b22ed035 100644 --- a/paddle/fluid/framework/tensor.h +++ b/paddle/fluid/framework/tensor.h @@ -45,10 +45,11 @@ class Tensor { friend struct EigenVector; public: - Tensor() : offset_(0) {} + Tensor() : offset_(0), is_pinned_(false) {} /*! Constructor with place should only be used in pybind. */ - explicit Tensor(const platform::Place& place) : offset_(0) { + explicit Tensor(const platform::Place& place) + : offset_(0), is_pinned_(false) { holder_->set_place(place); } @@ -69,11 +70,12 @@ class Tensor { * @note If not exist, then allocation. */ template - inline T* mutable_data(platform::Place place); + inline T* mutable_data(platform::Place place, bool is_pinned = false); - inline void* mutable_data(platform::Place place, std::type_index type); + inline void* mutable_data(platform::Place place, std::type_index type, + bool is_pinned = false); - inline void* mutable_data(platform::Place place); + inline void* mutable_data(platform::Place place, bool is_pinned = false); /** * @brief Return a pointer to mutable memory block. @@ -84,7 +86,8 @@ class Tensor { * @note If not exist, then allocation. */ template - inline T* mutable_data(DDim dims, platform::Place place); + inline T* mutable_data(DDim dims, platform::Place place, + bool is_pinned = false); /*! Return the dimensions of the memory block. */ inline const DDim& dims() const; @@ -92,6 +95,9 @@ class Tensor { /*! Return the numel of the memory block. */ inline int64_t numel() const; + /*! Return the numel of the memory block. */ + inline bool isPinned() const; + /*! Resize the dimensions of the memory block. */ inline Tensor& Resize(const DDim& dims); @@ -146,12 +152,14 @@ class Tensor { template struct PlaceholderImpl : public Placeholder { - PlaceholderImpl(Place place, size_t size, std::type_index type) - : ptr_(static_cast(memory::Alloc(place, size)), - memory::PODDeleter(place)), + PlaceholderImpl(Place place, size_t size, std::type_index type, + bool is_pinned = false) + : ptr_(static_cast(memory::Alloc(place, size, is_pinned)), + memory::PODDeleter(place, is_pinned)), place_(place), size_(size), - type_(type) { + type_(type), + is_pinned_(is_pinned) { PADDLE_ENFORCE_NOT_NULL(ptr_, "Insufficient %s memory to allocation.", (is_cpu_place(place_) ? "CPU" : "GPU")); } @@ -174,6 +182,9 @@ class Tensor { /* the current type of memory */ std::type_index type_; + + /*! use pinned memory or not. */ + bool is_pinned_; }; /*! holds the memory block if allocated. */ @@ -208,6 +219,7 @@ class Tensor { * PlaceHolder::ptr_ and where the tensor data really begins. */ size_t offset_; + bool is_pinned_; }; inline void Tensor::switch_place(platform::Place new_place) { diff --git a/paddle/fluid/framework/tensor_impl.h b/paddle/fluid/framework/tensor_impl.h index 638bd0db9d7025199c31a9327b96062512aa5adb..113814971e115fa88bd0ded34017fa26a9dd5803 100644 --- a/paddle/fluid/framework/tensor_impl.h +++ b/paddle/fluid/framework/tensor_impl.h @@ -101,52 +101,55 @@ inline T* Tensor::data() { } template -inline T* Tensor::mutable_data(DDim dims, platform::Place place) { +inline T* Tensor::mutable_data(DDim dims, platform::Place place, + bool is_pinned) { static_assert(std::is_pod::value, "T must be POD"); Resize(dims); - return mutable_data(place); + return mutable_data(place, is_pinned); } template -inline T* Tensor::mutable_data(platform::Place place) { +inline T* Tensor::mutable_data(platform::Place place, bool is_pinned) { static_assert(std::is_pod::value, "T must be POD"); - return reinterpret_cast(mutable_data(place, typeid(T))); + return reinterpret_cast(mutable_data(place, typeid(T), is_pinned)); } -inline void* Tensor::mutable_data(platform::Place place, std::type_index type) { +inline void* Tensor::mutable_data(platform::Place place, std::type_index type, + bool is_pinned) { if (holder_ != nullptr) { holder_->set_type(type); } - PADDLE_ENFORCE_GT( - numel(), 0, - "When calling this method, the Tensor's numel must be larger than zero. " - "Please check Tensor::Resize has been called first."); + PADDLE_ENFORCE_GE(numel(), 0, + "When calling this method, the Tensor's numel must be " + "equal or larger than zero. " + "Please check Tensor::Resize has been called first."); int64_t size = numel() * SizeOfType(type); /* some versions of boost::variant don't have operator!= */ if (holder_ == nullptr || !(holder_->place() == place) || holder_->size() < size + offset_) { if (platform::is_cpu_place(place)) { holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); + boost::get(place), size, type, is_pinned)); } else if (platform::is_gpu_place(place)) { #ifndef PADDLE_WITH_CUDA PADDLE_THROW("'CUDAPlace' is not supported in CPU only device."); } #else holder_.reset(new PlaceholderImpl( - boost::get(place), size, type)); + boost::get(place), size, type, is_pinned)); } #endif offset_ = 0; + is_pinned_ = is_pinned; } return reinterpret_cast(reinterpret_cast(holder_->ptr()) + offset_); } -inline void* Tensor::mutable_data(platform::Place place) { +inline void* Tensor::mutable_data(platform::Place place, bool is_pinned) { PADDLE_ENFORCE(this->holder_ != nullptr, "Cannot invoke mutable data if current hold nothing"); - return mutable_data(place, holder_->type()); + return mutable_data(place, holder_->type(), is_pinned); } inline Tensor& Tensor::ShareDataWith(const Tensor& src) { @@ -188,6 +191,8 @@ inline const DDim& Tensor::dims() const { return dims_; } inline int64_t Tensor::numel() const { return product(dims_); } +inline bool Tensor::isPinned() const { return is_pinned_; } + inline Tensor ReshapeToMatrix(const Tensor& src, int num_col_dims) { Tensor res; res.ShareDataWith(src); diff --git a/paddle/fluid/framework/threadpool.h b/paddle/fluid/framework/threadpool.h index df51fb24a588c84788d7d0b671f932ff4c40f9c2..f9dce7105e32ff0ba03d03f8faaac3a4ed1a3595 100644 --- a/paddle/fluid/framework/threadpool.h +++ b/paddle/fluid/framework/threadpool.h @@ -32,6 +32,8 @@ namespace framework { // number of threads. class ThreadPool { public: + explicit ThreadPool(int num_threads); + using Task = std::packaged_task()>; // Returns the singleton of ThreadPool. @@ -103,8 +105,6 @@ class ThreadPool { DISABLE_COPY_AND_ASSIGN(ThreadPool); - explicit ThreadPool(int num_threads); - // 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. diff --git a/paddle/fluid/memory/detail/system_allocator.cc b/paddle/fluid/memory/detail/system_allocator.cc index 8ac8978120ad5930cd80272189ac0a83a77b2617..22f6f506748735d1a0fe75375aeea22bd92b8b7e 100644 --- a/paddle/fluid/memory/detail/system_allocator.cc +++ b/paddle/fluid/memory/detail/system_allocator.cc @@ -79,7 +79,18 @@ void* GPUAllocator::Alloc(size_t& index, size_t size) { // if size is 0. We just make sure it does. if (size <= 0) return nullptr; void* p; + int prev_id; + cudaGetDevice(&prev_id); + if (prev_id != gpu_id_) { + cudaSetDevice(gpu_id_); + } + cudaError_t result = cudaMalloc(&p, size); + + if (prev_id != gpu_id_) { + cudaSetDevice(prev_id); + } + if (result == cudaSuccess) { index = 0; gpu_alloc_size_ += size; @@ -119,6 +130,50 @@ void GPUAllocator::Free(void* p, size_t size, size_t index) { bool GPUAllocator::UseGpu() const { return true; } +// PINNED memory allows direct DMA transfers by the GPU to and from system +// memory. It’s locked to a physical address. +void* CUDAPinnedAllocator::Alloc(size_t& index, size_t size) { + if (size <= 0) return nullptr; + void* p; + // NOTE: here, we use GpuMaxAllocSize() as the maximum memory size + // of host pinned allocation. Allocates too much would reduce + // the amount of memory available to the underlying system for paging. + + size_t usable = paddle::platform::GpuMaxAllocSize() - fallback_alloc_size_; + + if (size > usable) return nullptr; + + // PINNED memory is visible to all CUDA contexts. + cudaError_t result = cudaMallocHost(&p, size); + if (result == cudaSuccess) { + index = 1; + fallback_alloc_size_ += size; + return p; + } + + return nullptr; +} + +void CUDAPinnedAllocator::Free(void* p, size_t size, size_t index) { + cudaError_t err; + PADDLE_ASSERT(index == 1); + + PADDLE_ASSERT(fallback_alloc_size_ >= size); + fallback_alloc_size_ -= size; + err = cudaFreeHost(p); + + // Purposefully allow cudaErrorCudartUnloading, because + // that is returned if you ever call cudaFreeHost after the + // driver has already shutdown. This happens only if the + // process is terminating, in which case we don't care if + // cudaFreeHost succeeds. + if (err != cudaErrorCudartUnloading) { + PADDLE_ENFORCE(err, "cudaFreeHost failed in GPUPinnedAllocator::Free."); + } +} + +bool CUDAPinnedAllocator::UseGpu() const { return true; } + #endif } // namespace detail diff --git a/paddle/fluid/memory/detail/system_allocator.h b/paddle/fluid/memory/detail/system_allocator.h index e93c2c1e3231f7f42794dd78121072dbdb6abc41..e8479e73f433f1d741b2933da4843c0ba80276d5 100644 --- a/paddle/fluid/memory/detail/system_allocator.h +++ b/paddle/fluid/memory/detail/system_allocator.h @@ -43,6 +43,8 @@ class CPUAllocator : public SystemAllocator { #ifdef PADDLE_WITH_CUDA class GPUAllocator : public SystemAllocator { public: + explicit GPUAllocator(int gpu_id) : gpu_id_(gpu_id) {} + virtual void* Alloc(size_t& index, size_t size); virtual void Free(void* p, size_t size, size_t index); virtual bool UseGpu() const; @@ -50,6 +52,19 @@ class GPUAllocator : public SystemAllocator { private: size_t gpu_alloc_size_ = 0; size_t fallback_alloc_size_ = 0; + int gpu_id_; +}; + +class CUDAPinnedAllocator : public SystemAllocator { + public: + virtual void* Alloc(size_t& index, size_t size); + virtual void Free(void* p, size_t size, size_t index); + virtual bool UseGpu() const; + + private: + size_t gpu_alloc_size_ = + 0; // TODO(zcd): how to define the upper limit of CUDAPinnedMemory? + size_t fallback_alloc_size_ = 0; }; #endif diff --git a/paddle/fluid/memory/detail/system_allocator_test.cc b/paddle/fluid/memory/detail/system_allocator_test.cc index d5df9e6897e9e788f14d2625e424c13949eeaa26..3e1926f632c57b7906e4a76f43ff7a753d71d97f 100644 --- a/paddle/fluid/memory/detail/system_allocator_test.cc +++ b/paddle/fluid/memory/detail/system_allocator_test.cc @@ -58,7 +58,7 @@ TEST(CPUAllocator, LockMem) { #ifdef PADDLE_WITH_CUDA TEST(GPUAllocator, Alloc) { - paddle::memory::detail::GPUAllocator a; + paddle::memory::detail::GPUAllocator a(0); TestAllocator(a, 2048); TestAllocator(a, 0); } diff --git a/paddle/fluid/memory/memory.cc b/paddle/fluid/memory/memory.cc index d07f89439a1ec37682f79799d5569cad2ab75818..56593653a622bce323306d86156d140c46f58d18 100644 --- a/paddle/fluid/memory/memory.cc +++ b/paddle/fluid/memory/memory.cc @@ -38,7 +38,8 @@ BuddyAllocator* GetCPUBuddyAllocator() { } template <> -void* Alloc(platform::CPUPlace place, size_t size) { +void* Alloc(platform::CPUPlace place, size_t size, + bool is_pinned) { VLOG(10) << "Allocate " << size << " bytes on " << platform::Place(place); void* p = GetCPUBuddyAllocator()->Alloc(size); VLOG(10) << " pointer=" << p; @@ -46,7 +47,8 @@ void* Alloc(platform::CPUPlace place, size_t size) { } template <> -void Free(platform::CPUPlace place, void* p) { +void Free(platform::CPUPlace place, void* p, + bool is_pinned) { VLOG(10) << "Free pointer=" << p << " on " << platform::Place(place); GetCPUBuddyAllocator()->Free(p); } @@ -69,7 +71,31 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { } platform::SetDeviceId(gpu_id); if (!as[gpu_id]) { - as[gpu_id] = new BuddyAllocator(new detail::GPUAllocator, + as[gpu_id] = new BuddyAllocator(new detail::GPUAllocator(gpu_id), + 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"; + } + return as[gpu_id]; +} + +BuddyAllocator* GetCUDAPinnedBuddyAllocator(int gpu_id) { + static BuddyAllocator** as = NULL; + if (as == NULL) { + int gpu_num = platform::GetCUDADeviceCount(); + as = new BuddyAllocator*[gpu_num]; + for (int gpu = 0; gpu < gpu_num; gpu++) { + as[gpu] = nullptr; + } + } + platform::SetDeviceId(gpu_id); + if (!as[gpu_id]) { + as[gpu_id] = new BuddyAllocator(new detail::CUDAPinnedAllocator, platform::GpuMinChunkSize(), platform::GpuMaxChunkSize()); VLOG(10) << "\n\nNOTE: each GPU device use " @@ -88,9 +114,17 @@ size_t Used(platform::CUDAPlace place) { } template <> -void* Alloc(platform::CUDAPlace place, size_t size) { - auto* buddy_allocator = GetGPUBuddyAllocator(place.device); - auto* ptr = buddy_allocator->Alloc(size); +void* Alloc(platform::CUDAPlace place, size_t size, + bool is_pinned) { + void* ptr; + if (is_pinned) { + auto* buddy_allocator = GetCUDAPinnedBuddyAllocator(place.device); + ptr = buddy_allocator->Alloc(size); + } else { + auto* buddy_allocator = GetGPUBuddyAllocator(place.device); + ptr = buddy_allocator->Alloc(size); + } + if (ptr == nullptr) { int cur_dev = platform::GetCurrentDeviceId(); platform::SetDeviceId(place.device); @@ -108,8 +142,13 @@ void* Alloc(platform::CUDAPlace place, size_t size) { } template <> -void Free(platform::CUDAPlace place, void* p) { - GetGPUBuddyAllocator(place.device)->Free(p); +void Free(platform::CUDAPlace place, void* p, + bool is_pinned) { + if (is_pinned) { + GetCUDAPinnedBuddyAllocator(place.device)->Free(p); + } else { + GetGPUBuddyAllocator(place.device)->Free(p); + } } #endif diff --git a/paddle/fluid/memory/memory.h b/paddle/fluid/memory/memory.h index 7c5db815d6543f026ab99f7cf895a87db4e5a3d8..062bfc880e78dc5d90c567ffe5c4e521704c9ca6 100644 --- a/paddle/fluid/memory/memory.h +++ b/paddle/fluid/memory/memory.h @@ -33,7 +33,7 @@ namespace memory { * address is valid or not. */ template -void* Alloc(Place place, size_t size); +void* Alloc(Place place, size_t size, bool is_pinned = false); /** * \brief Free memory block in one place. @@ -43,7 +43,7 @@ void* Alloc(Place place, size_t size); * */ template -void Free(Place place, void* ptr); +void Free(Place place, void* ptr, bool is_pinned = false); /** * \brief Total size of used memory in one place. @@ -74,11 +74,13 @@ class PODDeleter { static_assert(std::is_pod::value, "T must be POD"); public: - explicit PODDeleter(Place place) : place_(place) {} - void operator()(T* ptr) { Free(place_, static_cast(ptr)); } + explicit PODDeleter(Place place, bool is_pinned = false) + : place_(place), is_pinned_(is_pinned) {} + void operator()(T* ptr) { Free(place_, static_cast(ptr), is_pinned_); } private: Place place_; + bool is_pinned_; }; /** diff --git a/paddle/fluid/memory/memory_test.cc b/paddle/fluid/memory/memory_test.cc index ae98d0d52542c49620a5d598b1089c168d39ede4..eb27a52b254c1cda065197746eb179bbd1d7f2f1 100644 --- a/paddle/fluid/memory/memory_test.cc +++ b/paddle/fluid/memory/memory_test.cc @@ -59,7 +59,7 @@ TEST(BuddyAllocator, CPUMultAlloc) { EXPECT_EQ(total_size, 0UL); for (auto size : - {128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) { + {0, 128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) { ps[paddle::memory::Alloc(cpu, size)] = size; // Buddy Allocator doesn't manage too large memory chunk @@ -117,7 +117,7 @@ TEST(BuddyAllocator, GPUMultAlloc) { EXPECT_EQ(total_size, 0UL); for (auto size : - {128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) { + {0, 128, 256, 1024, 4096, 16384, 65536, 262144, 1048576, 4194304}) { ps[paddle::memory::Alloc(gpu, size)] = size; // Buddy Allocator doesn't manage too large memory chunk diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index d30124d4a3b89b802a4abaae07a33b76526f163d..9ed79453b962b8702a88cea888a860cd5d8d64d1 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -12,6 +12,8 @@ function(op_library TARGET) set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE) set(cc_srcs) set(cu_srcs) + set(hip_cu_srcs) + set(miopen_hip_cc_srcs) set(cu_cc_srcs) set(cudnn_cu_cc_srcs) set(CUDNN_FILE) @@ -36,10 +38,19 @@ 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}.hip.cu) + list(APPEND hip_cu_srcs ${TARGET}.hip.cu) + endif() string(REPLACE "_op" "_cudnn_op" CUDNN_FILE "${TARGET}") if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${CUDNN_FILE}.cu.cc) list(APPEND cudnn_cu_cc_srcs ${CUDNN_FILE}.cu.cc) endif() + if(WITH_AMD_GPU) + string(REPLACE "_op" "_miopen_op" MIOPEN_FILE "${TARGET}") + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MIOPEN_FILE}.hip.cc) + list(APPEND miopen_hip_cc_srcs ${MIOPEN_FILE}.hip.cc) + endif() + endif() if(WITH_MKLDNN) string(REPLACE "_op" "_mkldnn_op" MKLDNN_FILE "${TARGET}") if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${MKLDNN_FILE}.cc) @@ -48,10 +59,14 @@ function(op_library TARGET) endif() else() foreach(src ${op_library_SRCS}) - if (${src} MATCHES ".*\\.cu$") + if (${src} MATCHES ".*\\.hip.cu$") + list(APPEND hip_cu_srcs ${src}) + elseif (${src} MATCHES ".*\\.cu$") list(APPEND cu_srcs ${src}) elseif(${src} MATCHES ".*_cudnn_op.cu.cc$") list(APPEND cudnn_cu_cc_srcs ${src}) + elseif(WITH_AMD_GPU AND ${src} MATCHES ".*_miopen_op.hip.cc$") + list(APPEND miopen_hip_cc_srcs ${src}) elseif(WITH_MKLDNN AND ${src} MATCHES ".*_mkldnn_op.cc$") list(APPEND mkldnn_cc_srcs ${src}) elseif(${src} MATCHES ".*\\.cu.cc$") @@ -76,6 +91,9 @@ function(op_library TARGET) if (WITH_GPU) nv_library(${TARGET} SRCS ${cc_srcs} ${cu_cc_srcs} ${cudnn_cu_cc_srcs} ${mkldnn_cc_srcs} ${cu_srcs} DEPS ${op_library_DEPS} ${op_common_deps}) + elseif (WITH_AMD_GPU) + hip_library(${TARGET} SRCS ${cc_srcs} ${hip_cu_srcs} ${miopen_hip_cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS} + ${op_common_deps}) else() cc_library(${TARGET} SRCS ${cc_srcs} ${mkldnn_cc_srcs} DEPS ${op_library_DEPS} ${op_common_deps}) @@ -88,7 +106,7 @@ function(op_library TARGET) endif() endforeach() - # The registration of USE_OP, please refer to paddle/framework/op_registry.h. + # The registration of USE_OP, please refer to paddle/fluid/framework/op_registry.h. # Note that it's enough to just adding one operator to pybind in a *_op.cc file. # And for detail pybind information, please see generated paddle/pybind/pybind.h. file(READ ${TARGET}.cc TARGET_CONTENT) @@ -114,7 +132,10 @@ function(op_library TARGET) list(LENGTH cu_srcs cu_srcs_len) list(LENGTH cu_cc_srcs cu_cc_srcs_len) list(LENGTH mkldnn_cc_srcs mkldnn_cc_srcs_len) - if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0) + list(LENGTH hip_cu_srcs hip_cu_srcs_len) + list(LENGTH miopen_hip_cc_srcs miopen_hip_cc_srcs_len) + if (${pybind_flag} EQUAL 0 AND ${mkldnn_cc_srcs_len} EQUAL 0 AND ${cu_srcs_len} EQUAL 0 AND ${cu_cc_srcs_len} EQUAL 0 AND + ${hip_cu_srcs_len} EQUAL 0 AND ${miopen_hip_cc_srcs_len} EQUAL 0) file(APPEND ${pybind_file} "USE_CPU_ONLY_OP(${TARGET});\n") set(pybind_flag 1) endif() @@ -125,9 +146,19 @@ function(op_library TARGET) file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, CUDNN);\n") endif() + # pybind USE_OP_DEVICE_KERNEL for MIOPEN + if (WITH_AMD_GPU AND ${miopen_hip_cc_srcs_len} GREATER 0) + file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MIOPEN);\n") + endif() + # pybind USE_OP_DEVICE_KERNEL for MKLDNN if (WITH_MKLDNN AND ${mkldnn_cc_srcs_len} GREATER 0) + # Append first implemented MKLDNN activation operator + if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op") + file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n") + else() file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n") + endif() endif() # pybind USE_OP @@ -152,13 +183,19 @@ if(WITH_DISTRIBUTE) set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") op_library(send_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(send_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + op_library(prefetch_op DEPS ${DISTRIBUTE_DEPS}) + set_source_files_properties(prefetch_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) op_library(recv_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(recv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) op_library(listen_and_serv_op DEPS ${DISTRIBUTE_DEPS}) set_source_files_properties(listen_and_serv_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS send_op listen_and_serv_op sum_op executor) + op_library(send_vars_op DEPS ${DISTRIBUTE_DEPS}) + set_source_files_properties(send_vars_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + op_library(send_barrier_op DEPS ${DISTRIBUTE_DEPS}) + set_source_files_properties(send_barrier_op.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) else() - set(DEPS_OPS ${DEPS_OPS} send_op recv_op listen_and_serv_op) + set(DEPS_OPS ${DEPS_OPS} send_op prefetch_op recv_op listen_and_serv_op send_vars_op send_barrier_op) endif() op_library(cond_op DEPS framework_proto tensor net_op) @@ -229,3 +266,4 @@ cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memor 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) nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) +nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor) diff --git a/paddle/fluid/operators/activation_mkldnn_op.cc b/paddle/fluid/operators/activation_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..6ff363d766db7dd97e1bc193ef7b4a095a7b7c24 --- /dev/null +++ b/paddle/fluid/operators/activation_mkldnn_op.cc @@ -0,0 +1,193 @@ +/* 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 "mkldnn.hpp" +#include "mkldnn_activation_op.h" +#include "paddle/fluid/operators/activation_op.h" + +namespace paddle { +namespace operators { + +using paddle::framework::Tensor; +using paddle::platform::MKLDNNDeviceContext; + +namespace { +template +void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm, + const T alpha = 0, const T beta = 0) { + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "It must use CPUPlace."); + + auto &dev_ctx = ctx.template device_context(); + const auto &mkldnn_engine = dev_ctx.GetEngine(); + + // get buffers + const auto *src = ctx.template Input("X"); + const auto *src_data = src->template data(); + + auto *dst = ctx.template Output("Out"); + const T *dst_data = dst->template mutable_data(ctx.GetPlace()); + + // get memory dim + PADDLE_ENFORCE(src->dims().size() == 4, + "Input dim must be with 4, i.e. NCHW"); + std::vector src_tz = framework::vectorize2int(src->dims()); + + // create memory description + // TODO(kbinias-intel): support more formats + auto data_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + + // create memory primitives + auto src_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)src_data); + auto dst_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)dst_data); + + auto forward_desc = mkldnn::eltwise_forward::desc( + mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta); + + // save prim desc into global device context to be referred in backward path + const std::string key = ctx.op().Output("Out"); + const std::string key_eltwise_pd = key + "@eltwise_pd"; + auto forward_pd = std::make_shared( + forward_desc, mkldnn_engine); + dev_ctx.SetBlob(key_eltwise_pd, forward_pd); + + auto eltwise = mkldnn::eltwise_forward(*forward_pd, src_memory, dst_memory); + + // push primitive to stream and wait until it's executed + std::vector pipeline = {eltwise}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); +} + +template +void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm, + const T alpha = 0, const T beta = 0) { + auto &dev_ctx = ctx.template device_context(); + const auto &mkldnn_engine = dev_ctx.GetEngine(); + + // get buffers + const auto *x = ctx.template Input("X"); + const auto *src = x->template data(); + + auto *dout = ctx.template Input(framework::GradVarName("Out")); + const auto *diff_dst = dout->template data(); + + auto *dx = + ctx.template Output(framework::GradVarName("X")); + const T *diff_src = dx->template mutable_data(ctx.GetPlace()); + + // get memory dim + std::vector src_tz = framework::vectorize2int(x->dims()); + + // create memory description + auto data_md = platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32, + mkldnn::memory::format::nchw); + + // create memory primitives + auto src_memory = mkldnn::memory({data_md, mkldnn_engine}, (void *)src); + auto diff_src_memory = + mkldnn::memory({data_md, mkldnn_engine}, (void *)diff_src); + auto diff_dst_memory = + mkldnn::memory({data_md, mkldnn_engine}, (void *)diff_dst); + + auto backward_desc = + mkldnn::eltwise_backward::desc(algorithm, data_md, data_md, alpha, beta); + + // retrieve eltwise primitive desc from device context + const std::string key = ctx.op().Input("Out"); + const std::string key_eltwise_pd = key + "@eltwise_pd"; + const std::shared_ptr forward_pd = dev_ctx.GetBlob(key_eltwise_pd); + PADDLE_ENFORCE(forward_pd != nullptr, + "Fail to find eltwise_pd in device context"); + auto *p_forward_pd = + static_cast(forward_pd.get()); + + auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc( + backward_desc, mkldnn_engine, *p_forward_pd); + + auto eltwise_bwd = mkldnn::eltwise_backward(eltwise_bwd_prim_desc, src_memory, + diff_dst_memory, diff_src_memory); + + // push primitive to stream and wait until it's executed + std::vector pipeline = {eltwise_bwd}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); +} +} // anonymous namespace + +template +struct MKLDNNActivationFunc : public BaseActivationFunctor { + template + void operator()(const ExecContext &ctx) const { + eltwise_forward(ctx, algorithm); + } +}; + +template +struct MKLDNNActivationGradFunc : public BaseActivationFunctor { + template + void operator()(const ExecContext &ctx) const { + eltwise_grad(ctx, algorithm); + } +}; + +template +using ReluMkldnnFunctor = + MKLDNNActivationFunc; + +template +using TanhMkldnnFunctor = + MKLDNNActivationFunc; + +template +using SqrtMkldnnFunctor = + MKLDNNActivationFunc; + +template +using AbsMkldnnFunctor = + MKLDNNActivationFunc; + +template +using ReluMkldnnGradFunctor = + MKLDNNActivationGradFunc; + +template +using TanhMkldnnGradFunctor = + MKLDNNActivationGradFunc; + +template +using SqrtMkldnnGradFunctor = + MKLDNNActivationGradFunc; + +template +using AbsMkldnnGradFunctor = + MKLDNNActivationGradFunc; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +#define REGISTER_ACTIVATION_MKLDNN_KERNEL(act_type, functor, grad_functor) \ + REGISTER_OP_KERNEL(act_type, MKLDNN, ::paddle::platform::CPUPlace, \ + ops::MKLDNNActivationKernel>); \ + REGISTER_OP_KERNEL( \ + act_type##_grad, MKLDNN, ::paddle::platform::CPUPlace, \ + ops::MKLDNNActivationGradKernel>); + +#define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro) \ + __macro(relu, ReluMkldnnFunctor, ReluMkldnnGradFunctor); \ + __macro(tanh, TanhMkldnnFunctor, TanhMkldnnGradFunctor); \ + __macro(sqrt, SqrtMkldnnFunctor, SqrtMkldnnGradFunctor); \ + __macro(abs, AbsMkldnnFunctor, AbsMkldnnGradFunctor); + +FOR_EACH_MKLDNN_KERNEL_FUNCTOR(REGISTER_ACTIVATION_MKLDNN_KERNEL); diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index ec637658c03ad94624ee9a4f5def6a84387d293e..a6d9ce0f041b859ecf6b3de902a9d1f132a4c76e 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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. @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/activation_op.h" +#include "paddle/fluid/operators/mkldnn_activation_op.h" namespace paddle { namespace operators { @@ -87,6 +88,9 @@ class ReluOpMaker : public framework::OpProtoAndCheckerMaker { : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Relu operator"); AddOutput("Out", "Output of Relu operator"); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Relu Activation Operator. @@ -140,6 +144,9 @@ class TanhOpMaker : public framework::OpProtoAndCheckerMaker { : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Tanh operator"); AddOutput("Out", "Output of Tanh operator"); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Tanh Activation Operator. @@ -193,6 +200,9 @@ class SqrtOpMaker : public framework::OpProtoAndCheckerMaker { : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Sqrt operator"); AddOutput("Out", "Output of Sqrt operator"); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Sqrt Activation Operator. @@ -208,6 +218,9 @@ class AbsOpMaker : public framework::OpProtoAndCheckerMaker { : framework::OpProtoAndCheckerMaker(proto, op_checker) { AddInput("X", "Input of Abs operator"); AddOutput("Out", "Output of Abs operator"); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); AddComment(R"DOC( Abs Activation Operator. @@ -247,6 +260,36 @@ $out = floor(x)$ } }; +class CosOpMaker : public framework::OpProtoAndCheckerMaker { + public: + CosOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Cosine operator"); + AddOutput("Out", "Output of Cosine operator"); + AddComment(R"DOC( +Cosine Activation Operator. + +$out = cos(x)$ + +)DOC"); + } +}; + +class SinOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SinOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : framework::OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "Input of Sine operator"); + AddOutput("Out", "Output of Sine operator"); + AddComment(R"DOC( +Sine Activation Operator. + +$out = sin(x)$ + +)DOC"); + } +}; + class RoundOpMaker : public framework::OpProtoAndCheckerMaker { public: RoundOpMaker(OpProto *proto, OpAttrChecker *op_checker) @@ -524,11 +567,11 @@ REGISTER_OP(logsigmoid, ops::ActivationOp, ops::LogSigmoidOpMaker, REGISTER_OP(exp, ops::ActivationOp, ops::ExpOpMaker, exp_grad, ops::ActivationOpGrad); -REGISTER_OP(relu, ops::ActivationOp, ops::ReluOpMaker, relu_grad, - ops::ActivationOpGrad); +REGISTER_OP(relu, ops::ActivationWithMKLDNNOp, ops::ReluOpMaker, relu_grad, + ops::ActivationWithMKLDNNOpGrad); -REGISTER_OP(tanh, ops::ActivationOp, ops::TanhOpMaker, tanh_grad, - ops::ActivationOpGrad); +REGISTER_OP(tanh, ops::ActivationWithMKLDNNOp, ops::TanhOpMaker, tanh_grad, + ops::ActivationWithMKLDNNOpGrad); REGISTER_OP(tanh_shrink, ops::ActivationOp, ops::TanhShrinkOpMaker, tanh_shrink_grad, ops::ActivationOpGrad); @@ -536,11 +579,11 @@ REGISTER_OP(tanh_shrink, ops::ActivationOp, ops::TanhShrinkOpMaker, REGISTER_OP(softshrink, ops::ActivationOp, ops::SoftShrinkOpMaker, softshrink_grad, ops::ActivationOpGrad); -REGISTER_OP(sqrt, ops::ActivationOp, ops::SqrtOpMaker, sqrt_grad, - ops::ActivationOpGrad); +REGISTER_OP(sqrt, ops::ActivationWithMKLDNNOp, ops::SqrtOpMaker, sqrt_grad, + ops::ActivationWithMKLDNNOpGrad); -REGISTER_OP(abs, ops::ActivationOp, ops::AbsOpMaker, abs_grad, - ops::ActivationOpGrad); +REGISTER_OP(abs, ops::ActivationWithMKLDNNOp, ops::AbsOpMaker, abs_grad, + ops::ActivationWithMKLDNNOpGrad); REGISTER_OP(ceil, ops::ActivationOp, ops::CeilOpMaker, ceil_grad, ops::ActivationOpGrad); @@ -548,6 +591,12 @@ REGISTER_OP(ceil, ops::ActivationOp, ops::CeilOpMaker, ceil_grad, REGISTER_OP(floor, ops::ActivationOp, ops::FloorOpMaker, floor_grad, ops::ActivationOpGrad); +REGISTER_OP(cos, ops::ActivationOp, ops::CosOpMaker, cos_grad, + ops::ActivationOpGrad); + +REGISTER_OP(sin, ops::ActivationOp, ops::SinOpMaker, sin_grad, + ops::ActivationOpGrad); + REGISTER_OP(round, ops::ActivationOp, ops::RoundOpMaker, round_grad, ops::ActivationOpGrad); diff --git a/paddle/fluid/operators/activation_op.h b/paddle/fluid/operators/activation_op.h index b95e793586219b7c413d0c7adb835081874d9363..7fbe4efc045b6539b498389af94769e5bdb1f82e 100644 --- a/paddle/fluid/operators/activation_op.h +++ b/paddle/fluid/operators/activation_op.h @@ -1,4 +1,4 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. +/* 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. @@ -17,6 +17,10 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + namespace paddle { namespace operators { @@ -327,6 +331,54 @@ struct FloorFunctor : public BaseActivationFunctor { } }; +template +struct Sine { + HOSTDEVICE T operator()(const T& val) const { return sin(val); } +}; + +template +struct Cosine { + HOSTDEVICE T operator()(const T& val) const { return cos(val); } +}; + +// cosine'(x) = -sin(x) +template +struct CosGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = -dout * x.unaryExpr(Sine()); + } +}; + +// cosine(x) = cos(x) +template +struct CosFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.unaryExpr(Cosine()); + } +}; + +// sine'(x) = cos(x) +template +struct SinGradFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out, dOut dout, dX dx) const { + dx.device(d) = dout * x.unaryExpr(Cosine()); + } +}; + +// sine(x) = sin(x) +template +struct SinFunctor : public BaseActivationFunctor { + template + void operator()(Device d, X x, Out out) const { + out.device(d) = x.unaryExpr(Sine()); + } +}; + // round(x) = [x] template struct RoundFunctor : public BaseActivationFunctor { @@ -778,6 +830,8 @@ struct SwishGradFunctor : public BaseActivationFunctor { __macro(abs, AbsFunctor, AbsGradFunctor); \ __macro(ceil, CeilFunctor, ZeroGradFunctor); \ __macro(floor, FloorFunctor, ZeroGradFunctor); \ + __macro(cos, CosFunctor, CosGradFunctor); \ + __macro(sin, SinFunctor, SinGradFunctor); \ __macro(round, RoundFunctor, ZeroGradFunctor); \ __macro(reciprocal, ReciprocalFunctor, ReciprocalGradFunctor); \ __macro(log, LogFunctor, LogGradFunctor); \ diff --git a/paddle/fluid/operators/batch_norm_op.cc b/paddle/fluid/operators/batch_norm_op.cc index 5d27f5b60c7115a32aeeca5ec2a6654471c310c7..36049ee6a4a0d2a251b6d10cf1ff05a9d9845089 100644 --- a/paddle/fluid/operators/batch_norm_op.cc +++ b/paddle/fluid/operators/batch_norm_op.cc @@ -457,12 +457,39 @@ class BatchNormGradKernel } }; +class BatchNormGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + protected: + std::unique_ptr Apply() const override { + auto *op = new framework::OpDesc(); + op->SetType("batch_norm_grad"); + op->SetInput("X", Input("X")); + op->SetInput(framework::GradVarName("Y"), OutputGrad("Y")); + + op->SetInput("Scale", Input("Scale")); + op->SetInput("SavedMean", Output("SavedMean")); + op->SetInput("SavedVariance", Output("SavedVariance")); + + op->SetAttrMap(Attrs()); + + op->SetOutput(framework::GradVarName("X"), InputGrad("X")); + op->SetOutput(framework::GradVarName("Scale"), InputGrad("Scale")); + op->SetOutput(framework::GradVarName("Bias"), InputGrad("Bias")); + + return std::unique_ptr(op); + } +}; + } // namespace operators } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker, - batch_norm_grad, ops::BatchNormGradOp); +REGISTER_OPERATOR(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker, + ops::BatchNormGradMaker); +REGISTER_OPERATOR(batch_norm_grad, ops::BatchNormGradOp); + REGISTER_OP_CPU_KERNEL( batch_norm, ops::BatchNormKernel); diff --git a/paddle/fluid/operators/box_coder_op.cc b/paddle/fluid/operators/box_coder_op.cc index eccdd408a17a07a541480705242b137f8207c139..ec416f725e75fae57484751ee8a066c0b9da8a70 100644 --- a/paddle/fluid/operators/box_coder_op.cc +++ b/paddle/fluid/operators/box_coder_op.cc @@ -126,6 +126,7 @@ width and height. } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker); +REGISTER_OPERATOR(box_coder, ops::BoxCoderOp, ops::BoxCoderOpMaker, + paddle::framework::EmptyGradOpMaker); REGISTER_OP_CPU_KERNEL(box_coder, ops::BoxCoderKernel, ops::BoxCoderKernel); diff --git a/paddle/fluid/operators/channel_send_op.cc b/paddle/fluid/operators/channel_send_op.cc index 47cf7d7efc9996e8a8db11b79c0310f77c2435a4..66d33617ede5bef8a95de14f5b447c0910fe3eb4 100644 --- a/paddle/fluid/operators/channel_send_op.cc +++ b/paddle/fluid/operators/channel_send_op.cc @@ -23,21 +23,10 @@ limitations under the License. */ static constexpr char Channel[] = "Channel"; static constexpr char X[] = "X"; -static constexpr char Status[] = "Status"; -static constexpr char copy[] = "copy"; namespace paddle { namespace operators { -void SetSendStatus(const platform::Place &dev_place, - framework::Variable &status_var, bool status) { - auto cpu = platform::CPUPlace(); - auto status_tensor = - status_var.GetMutable()->mutable_data({1}, - cpu); - status_tensor[0] = status; -} - class ChannelSendOp : public framework::OperatorBase { public: ChannelSendOp(const std::string &type, @@ -51,9 +40,6 @@ class ChannelSendOp : public framework::OperatorBase { "Input(Channel) of ChannelSendOp should not be null."); PADDLE_ENFORCE(ctx->HasInput(X), "Input(X) of ChannelSendOp should not be null."); - PADDLE_ENFORCE(ctx->HasOutput(Status), - "Output(Status) of ChannelSendOp should not be null."); - ctx->SetOutputDim("Status", {1}); } private: @@ -65,10 +51,7 @@ class ChannelSendOp : public framework::OperatorBase { auto input_var = scope.FindVar(Input(X)); // Send the input data through the channel. - bool ok = concurrency::ChannelSend(ch, input_var); - - // Set the status output of the `ChannelSend` call. - SetSendStatus(dev_place, *scope.FindVar(Output(Status)), ok); + concurrency::ChannelSend(ch, input_var); } }; @@ -82,12 +65,6 @@ class ChannelSendOpMaker : public framework::OpProtoAndCheckerMaker { .AsDuplicable(); AddInput(X, "(Variable) The value which gets sent by the channel.") .AsDuplicable(); - AddOutput(Status, - "(Tensor) An LoD Tensor that returns a boolean status of the" - "result of the send operation.") - .AsDuplicable(); - AddAttr(copy, "(bool, default false) Should copy before send") - .SetDefault(false); AddComment(R"DOC( )DOC"); } diff --git a/paddle/fluid/operators/compare_op.cc b/paddle/fluid/operators/compare_op.cc index 86f7046058c7001fcaa588727b1cdc0f3f20c35f..9a139ab27ec53395a8d1ab1347dbce93ea68fd8e 100644 --- a/paddle/fluid/operators/compare_op.cc +++ b/paddle/fluid/operators/compare_op.cc @@ -29,6 +29,11 @@ class CompareOpProtoMaker : public framework::OpProtoAndCheckerMaker { AddInput("Y", string::Sprintf( "(LoDTensor) the right hand operand of %s operator", comment.type)); + AddAttr("force_cpu", + "(bool, default false) Force fill output variable to cpu " + "memory. Otherwise, fill output variable to the running " + "device") + .SetDefault(false); AddOutput("Out", string::Sprintf( "(LoDTensor) n-dim bool tensor. Each element is %s", comment.equation)); @@ -75,7 +80,9 @@ class CompareOp : public framework::OperatorWithKernel { const framework::ExecutionContext &ctx) const override { framework::OpKernelType kt = OperatorWithKernel::GetExpectedKernelType(ctx); // CompareOp kernel's device type is decided by input tensor place - kt.place_ = ctx.Input("X")->place(); + bool force_cpu = ctx.Attr("force_cpu"); + kt.place_ = force_cpu ? platform::CPUPlace() + : ctx.Input("X")->place(); return kt; } }; diff --git a/paddle/fluid/operators/concurrency/channel_util.cc b/paddle/fluid/operators/concurrency/channel_util.cc index a483af7affd824da7d18676d934dc959167ef71f..246c99489c45efec16babb1d3980606318236605 100644 --- a/paddle/fluid/operators/concurrency/channel_util.cc +++ b/paddle/fluid/operators/concurrency/channel_util.cc @@ -17,20 +17,20 @@ limitations under the License. */ namespace poc = paddle::operators::concurrency; -bool poc::ChannelSend(framework::ChannelHolder *ch, framework::Variable *var) { +void poc::ChannelSend(framework::ChannelHolder *ch, framework::Variable *var) { auto type = framework::ToVarType(var->Type()); if (type == framework::proto::VarType_Type_LOD_TENSOR) - return ch->Send(var->GetMutable()); + ch->Send(var->GetMutable()); else if (type == framework::proto::VarType_Type_LOD_RANK_TABLE) - return ch->Send(var->GetMutable()); + ch->Send(var->GetMutable()); else if (type == framework::proto::VarType_Type_LOD_TENSOR_ARRAY) - return ch->Send(var->GetMutable()); + ch->Send(var->GetMutable()); else if (type == framework::proto::VarType_Type_SELECTED_ROWS) - return ch->Send(var->GetMutable()); + ch->Send(var->GetMutable()); else if (type == framework::proto::VarType_Type_READER) - return ch->Send(var->GetMutable()); + ch->Send(var->GetMutable()); else if (type == framework::proto::VarType_Type_CHANNEL) - return ch->Send(var->GetMutable()); + ch->Send(var->GetMutable()); else PADDLE_THROW("ChannelSend:Unsupported type"); } diff --git a/paddle/fluid/operators/concurrency/channel_util.h b/paddle/fluid/operators/concurrency/channel_util.h index c3674bd9815df451751707bfa84d18dbb5fa0f6b..cd18ca78c6fdecdc6c72748611ccdd9c2690ef46 100644 --- a/paddle/fluid/operators/concurrency/channel_util.h +++ b/paddle/fluid/operators/concurrency/channel_util.h @@ -21,7 +21,7 @@ namespace paddle { namespace operators { namespace concurrency { -bool ChannelSend(framework::ChannelHolder *ch, framework::Variable *var); +void ChannelSend(framework::ChannelHolder *ch, framework::Variable *var); bool ChannelReceive(framework::ChannelHolder *ch, framework::Variable *var); void ChannelAddToSendQ(framework::ChannelHolder *ch, const void *referrer, diff --git a/paddle/fluid/operators/conditional_block_op.cc b/paddle/fluid/operators/conditional_block_op.cc index 337b34e8f0bf4cb89753235205be9eb058dd01ab..bff2c34ec893d0e6212426b108dd98b0d0d0fb48 100644 --- a/paddle/fluid/operators/conditional_block_op.cc +++ b/paddle/fluid/operators/conditional_block_op.cc @@ -54,7 +54,18 @@ class ConditionalOp : public framework::OperatorBase { "numel should be 1, actual numel is %d", ips[0]->numel()); } - return ips[0]->data()[0]; + bool res = false; + if (platform::is_gpu_place(ips[0]->place())) { +#ifdef PADDLE_WITH_CUDA + framework::LoDTensor cpu_tensor; + framework::TensorCopy(*ips[0], platform::CPUPlace(), &cpu_tensor); + platform::DeviceContextPool::Instance().Get(ips[0]->place())->Wait(); + res = cpu_tensor.data()[0]; +#endif + } else { + res = ips[0]->data()[0]; + } + return res; } }; diff --git a/paddle/fluid/operators/detail/CMakeLists.txt b/paddle/fluid/operators/detail/CMakeLists.txt index 94395ccfbcbd74ee40552a5c70dc8b8063a5f851..3adeeda90645ca983d9d9229b4cc1c4c90302206 100644 --- a/paddle/fluid/operators/detail/CMakeLists.txt +++ b/paddle/fluid/operators/detail/CMakeLists.txt @@ -1,6 +1,9 @@ if(WITH_DISTRIBUTE) - grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc grpc_server.cc PROTO send_recv.proto DEPS lod_tensor selected_rows) + grpc_library(sendrecvop_grpc SRCS bytebuffer_stream.cc sendrecvop_utils.cc grpc_client.cc + grpc_server.cc variable_response.cc PROTO send_recv.proto DEPS lod_tensor selected_rows) set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor") - set_source_files_properties(test_serde.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) - cc_test(serde_test SRCS test_serde.cc DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc) + set_source_files_properties(serde_test.cc grpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) + cc_test(serde_test SRCS serde_test.cc variable_response.cc DEPS grpc++_unsecure grpc_unsecure gpr + cares zlib protobuf sendrecvop_grpc) + cc_test(grpc_server_test SRCS grpc_server_test.cc DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf) endif() diff --git a/paddle/fluid/operators/detail/bytebuffer_stream.h b/paddle/fluid/operators/detail/bytebuffer_stream.h index 099deb12d0e436427c147ab9b1eb553b712e14fb..1791a48aab1b66147f645c90757b35ef5f6e001b 100644 --- a/paddle/fluid/operators/detail/bytebuffer_stream.h +++ b/paddle/fluid/operators/detail/bytebuffer_stream.h @@ -23,9 +23,107 @@ limitations under the License. */ #include "google/protobuf/io/coded_stream.h" #include "google/protobuf/io/zero_copy_stream.h" +namespace grpc { +// A ZeroCopyInputStream that reads from grpc_byte_buffer +class GrpcBufferReader final + : public ::google::protobuf::io::ZeroCopyInputStream { + typedef void (CoreCodegenInterface::*OldReaderInitAPI)( + grpc_byte_buffer_reader* reader, grpc_byte_buffer* buffer); + typedef int (CoreCodegenInterface::*NewReaderInitAPI)( + grpc_byte_buffer_reader* reader, grpc_byte_buffer* buffer); + void ReaderInit(OldReaderInitAPI ptr, grpc_byte_buffer_reader* reader, + grpc_byte_buffer* buffer) { + (g_core_codegen_interface->*ptr)(reader, buffer); + } + void ReaderInit(NewReaderInitAPI ptr, grpc_byte_buffer_reader* reader, + grpc_byte_buffer* buffer) { + int result = (g_core_codegen_interface->*ptr)(reader, buffer); + (void)result; + } + + public: + explicit GrpcBufferReader(grpc_byte_buffer* buffer) + : byte_count_(0), backup_count_(0) { + ReaderInit(&CoreCodegenInterface::grpc_byte_buffer_reader_init, &reader_, + buffer); + } + ~GrpcBufferReader() override { + g_core_codegen_interface->grpc_byte_buffer_reader_destroy(&reader_); + } + + bool Next(const void** data, int* size) override { + if (backup_count_ > 0) { + *data = GRPC_SLICE_START_PTR(slice_) + GRPC_SLICE_LENGTH(slice_) - + backup_count_; + GPR_CODEGEN_ASSERT(backup_count_ <= INT_MAX); + *size = (int)backup_count_; + backup_count_ = 0; + return true; + } + if (!g_core_codegen_interface->grpc_byte_buffer_reader_next(&reader_, + &slice_)) { + return false; + } + g_core_codegen_interface->grpc_slice_unref(slice_); + *data = GRPC_SLICE_START_PTR(slice_); + // On win x64, int is only 32bit + GPR_CODEGEN_ASSERT(GRPC_SLICE_LENGTH(slice_) <= INT_MAX); + byte_count_ += * size = (int)GRPC_SLICE_LENGTH(slice_); + return true; + } + + void BackUp(int count) override { backup_count_ = count; } + + bool Skip(int count) override { + const void* data; + int size; + while (Next(&data, &size)) { + if (size >= count) { + BackUp(size - count); + return true; + } + // size < count; + count -= size; + } + // error or we have too large count; + return false; + } + + ::google::protobuf::int64 ByteCount() const override { + return byte_count_ - backup_count_; + } + + private: + int64_t byte_count_; + int64_t backup_count_; + grpc_byte_buffer_reader reader_; + grpc_slice slice_; +}; + +}; // namespace grpc + namespace paddle { namespace operators { namespace detail { +// Source provides a way for a particular RPC implementation to provide +// received data to ParseFrom. +class Source { + public: + virtual ~Source() {} + + // Return the stream that contains the data to be parsed. + // Note that this method might be invoked more than once if + // ParseFrom needs to fall back to a more expensive parsing method. + // Every call must return a stream pointing at the beginning of + // the serialized RecvTensorResponse. + // + // Note that a subsequent call to contents() invalidates previous + // results of contents(). + // + // Ownership of the returned stream is retained by the Source and + // should not be deleted by the caller. + virtual ::google::protobuf::io::ZeroCopyInputStream* contents() = 0; +}; // A ZeroCopyInputStream that reads from a grpc::ByteBuffer. class GrpcByteBufferSource @@ -46,6 +144,43 @@ class GrpcByteBufferSource ::google::protobuf::int64 byte_count_; }; +class GrpcByteBufferSourceWrapper : public Source { + public: + explicit GrpcByteBufferSourceWrapper(GrpcByteBufferSource* source) + : source_(source) {} + ::google::protobuf::io::ZeroCopyInputStream* contents() override { + return source_; + } + + private: + GrpcByteBufferSource* source_; +}; + +class GrpcByteSource : public Source { + public: + explicit GrpcByteSource(grpc_byte_buffer* buffer) : buffer_(buffer) {} + ~GrpcByteSource() override { DeleteStream(); } + + typedef ::grpc::GrpcBufferReader Reader; + + ::google::protobuf::io::ZeroCopyInputStream* contents() override { + DeleteStream(); + stream_ = new (&space_) Reader(buffer_); + return stream_; + } + + private: + void DeleteStream() { + if (stream_) { + stream_->~Reader(); + } + } + + grpc_byte_buffer* buffer_; // Not owned + Reader* stream_ = nullptr; // Points into space_ if non-nullptr + char space_[sizeof(Reader)]; +}; + } // namespace detail } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/detail/grpc_client.cc b/paddle/fluid/operators/detail/grpc_client.cc index ddeeebec58e02f1686fd2e3d3e5ac1a4c4fd3c59..d79ba6d291950e1f089eb11713bd1c3e4d154b27 100644 --- a/paddle/fluid/operators/detail/grpc_client.cc +++ b/paddle/fluid/operators/detail/grpc_client.cc @@ -12,8 +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 "grpc_client.h" +#include "paddle/fluid/operators/detail/grpc_client.h" + +#include + #include "paddle/fluid/framework/threadpool.h" + namespace paddle { namespace operators { namespace detail { @@ -31,8 +35,9 @@ bool RPCClient::AsyncSendVariable(const std::string& ep, framework::Async([var_name_val, p_ctx, ep_val, p_scope, time_out, ch, this] { auto* var = p_scope->FindVar(var_name_val); - sendrecv::VariableMessage req; - SerializeToMessage(var_name_val, var, *p_ctx, &req); + + ::grpc::ByteBuffer req; + SerializeToByteBuffer(var_name_val, var, *p_ctx, &req); // varhandle VarHandle var_h; @@ -46,8 +51,10 @@ bool RPCClient::AsyncSendVariable(const std::string& ep, s->Prepare(var_h, time_out); s->response_call_back_ = NULL; - auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); - rpc->Finish(&s->reply_, &s->status_, (void*)s); + auto call = s->stub_g_.PrepareUnaryCall( + s->context_.get(), "/sendrecv.SendRecvService/SendVariable", req, &cq_); + call->StartCall(); + call->Finish(&s->reply_, &s->status_, static_cast(s)); }); req_count_++; @@ -56,9 +63,18 @@ bool RPCClient::AsyncSendVariable(const std::string& ep, } void ProcGetResponse(const VarHandle& var_h, - const sendrecv::VariableMessage& ret_msg) { - auto* outvar = var_h.scope->FindVar(var_h.name); - DeserializeFromMessage(ret_msg, *var_h.ctx, outvar); + // const sendrecv::VariableMessage& ret_msg) { + const ::grpc::ByteBuffer& ret_msg) { + framework::Variable* outvar = NULL; + DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, outvar); +} + +template +void RequestToByteBuffer(const T& proto, ::grpc::ByteBuffer* result) { + ::grpc::Slice slice(proto.ByteSizeLong()); + proto.SerializeWithCachedSizesToArray(const_cast(slice.begin())); + ::grpc::ByteBuffer tmp(&slice, 1); + result->Swap(&tmp); } bool RPCClient::AsyncGetVariable(const std::string& ep, @@ -73,10 +89,13 @@ bool RPCClient::AsyncGetVariable(const std::string& ep, const auto ch = GetChannel(ep_val); framework::Async([var_name_val, ep_val, p_scope, p_ctx, time_out, ch, this] { + // prepare input sendrecv::VariableMessage req; req.set_varname(var_name_val); + ::grpc::ByteBuffer buf; + RequestToByteBuffer(req, &buf); - // varhandle + // var handle VarHandle var_h; var_h.ep = ep_val; var_h.scope = p_scope; @@ -88,8 +107,10 @@ bool RPCClient::AsyncGetVariable(const std::string& ep, s->Prepare(var_h, time_out); s->response_call_back_ = ProcGetResponse; - auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); - rpc->Finish(&s->reply_, &s->status_, (void*)s); + auto call = s->stub_g_.PrepareUnaryCall( + s->context_.get(), "/sendrecv.SendRecvService/GetVariable", buf, &cq_); + call->StartCall(); + call->Finish(&s->reply_, &s->status_, static_cast(s)); }); req_count_++; @@ -97,6 +118,49 @@ bool RPCClient::AsyncGetVariable(const std::string& ep, return true; } +bool RPCClient::AsyncPrefetchVariable(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out) { + const platform::DeviceContext* p_ctx = &ctx; + const std::string ep_val = ep; + const std::string in_var_name_val = in_var_name; + const std::string out_var_name_val = out_var_name; + const framework::Scope* p_scope = &scope; + const auto ch = GetChannel(ep_val); + + framework::Async([in_var_name_val, out_var_name_val, ep_val, p_scope, p_ctx, + time_out, ch, this] { + auto* var = p_scope->FindVar(in_var_name_val); + + ::grpc::ByteBuffer req; + SerializeToByteBuffer(in_var_name_val, var, *p_ctx, &req); + + // var handle + VarHandle var_h; + var_h.ep = ep_val; + var_h.scope = p_scope; + var_h.name = out_var_name_val; + var_h.ctx = p_ctx; + + // stub context + GetProcessor* s = new GetProcessor(ch); + s->Prepare(var_h, time_out); + s->response_call_back_ = ProcGetResponse; + + 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)); + }); + + req_count_++; + return true; +} + void RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { const auto ch = GetChannel(ep); @@ -106,7 +170,7 @@ void RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) { sendrecv::VariableMessage req; req.set_varname(BATCH_BARRIER_MESSAGE); auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_); - rpc->Finish(&s->reply_, &s->status_, (void*)s); + rpc->Finish(&s->reply_, &s->status_, static_cast(s)); req_count_++; } @@ -118,7 +182,7 @@ void RPCClient::AsyncSendFetchBarrier(const std::string& ep, int64_t time_out) { sendrecv::VariableMessage req; req.set_varname(FETCH_BARRIER_MESSAGE); auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_); - rpc->Finish(&s->reply_, &s->status_, (void*)s); + rpc->Finish(&s->reply_, &s->status_, static_cast(s)); req_count_++; } @@ -184,7 +248,6 @@ std::shared_ptr RPCClient::GetChannel(const std::string& ep) { } grpc::ChannelArguments args; - args.SetInt("grpc.testing.fixed_reconnect_backoff_ms", 5000); args.SetCompressionAlgorithm(GRPC_COMPRESS_NONE); args.SetMaxSendMessageSize(std::numeric_limits::max()); args.SetMaxReceiveMessageSize(std::numeric_limits::max()); diff --git a/paddle/fluid/operators/detail/grpc_client.h b/paddle/fluid/operators/detail/grpc_client.h index f520367dd981288416631fdad15241fb5d811d07..fe237e54ef61fb5b6e9bfa46fbe6b3df3dd40265 100644 --- a/paddle/fluid/operators/detail/grpc_client.h +++ b/paddle/fluid/operators/detail/grpc_client.h @@ -25,6 +25,11 @@ limitations under the License. */ #include #include +#include +#include +#include +#include + #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" @@ -49,15 +54,11 @@ struct VarHandle { } }; -void ProcGetResponse(const VarHandle& var_h, - const sendrecv::VariableMessage& msg); +void ProcGetResponse(const VarHandle& var_h, const grpc::ByteBuffer& msg); class BaseProcessor { public: - explicit BaseProcessor(std::shared_ptr ch) { - stub_ = sendrecv::SendRecvService::NewStub(ch); - context_ = NULL; - } + explicit BaseProcessor(std::shared_ptr ch) { context_ = NULL; } virtual ~BaseProcessor() {} @@ -82,19 +83,18 @@ class BaseProcessor { virtual void Process() = 0; - std::unique_ptr stub_; std::unique_ptr context_; grpc::Status status_; VarHandle var_h_; }; -typedef std::function +typedef std::function RequestSendCallBack; class SendProcessor : public BaseProcessor { public: explicit SendProcessor(std::shared_ptr ch) - : BaseProcessor(ch) {} + : BaseProcessor(ch), stub_g_(ch) {} virtual ~SendProcessor() {} @@ -104,17 +104,18 @@ class SendProcessor : public BaseProcessor { } } - sendrecv::VoidMessage reply_; + ::grpc::GenericStub stub_g_; + ::grpc::ByteBuffer reply_; RequestSendCallBack response_call_back_ = NULL; }; -typedef std::function +typedef std::function RequestGetCallBack; class GetProcessor : public BaseProcessor { public: explicit GetProcessor(std::shared_ptr ch) - : BaseProcessor(ch) {} + : BaseProcessor(ch), stub_g_(ch) {} virtual ~GetProcessor() {} @@ -124,30 +125,37 @@ class GetProcessor : public BaseProcessor { } } - sendrecv::VariableMessage reply_; + ::grpc::ByteBuffer reply_; + ::grpc::GenericStub stub_g_; RequestGetCallBack response_call_back_ = ProcGetResponse; }; class BatchBarrierProcessor : public BaseProcessor { public: explicit BatchBarrierProcessor(std::shared_ptr ch) - : BaseProcessor(ch) {} + : BaseProcessor(ch) { + stub_ = sendrecv::SendRecvService::NewStub(ch); + } virtual ~BatchBarrierProcessor() {} virtual void Process() {} sendrecv::VoidMessage reply_; + std::unique_ptr stub_; }; class FetchBarrierProcessor : public BaseProcessor { public: explicit FetchBarrierProcessor(std::shared_ptr ch) - : BaseProcessor(ch) {} + : BaseProcessor(ch) { + stub_ = sendrecv::SendRecvService::NewStub(ch); + } virtual ~FetchBarrierProcessor() {} virtual void Process() {} sendrecv::VariableMessage reply_; + std::unique_ptr stub_; }; class RPCClient { @@ -164,6 +172,13 @@ class RPCClient { const std::string& var_name, int64_t time_out = 600 * 1000); + bool AsyncPrefetchVariable(const std::string& ep, + const platform::DeviceContext& ctx, + const framework::Scope& scope, + const std::string& in_var_name, + const std::string& out_var_name, + int64_t time_out = 600 * 1000); + void AsyncSendBatchBarrier(const std::string& ep, int64_t time_out = 600 * 1000); diff --git a/paddle/fluid/operators/detail/grpc_server.cc b/paddle/fluid/operators/detail/grpc_server.cc index 8fff430cc4890925e4edba2fadb8eb7fc647d181..7c978b28b6873d05afb435de4caf7f4ce5d33193 100644 --- a/paddle/fluid/operators/detail/grpc_server.cc +++ b/paddle/fluid/operators/detail/grpc_server.cc @@ -14,7 +14,10 @@ limitations under the License. */ #include "paddle/fluid/operators/detail/grpc_server.h" -using grpc::ServerAsyncResponseWriter; +#include +#include + +using ::grpc::ServerAsyncResponseWriter; namespace paddle { namespace operators { @@ -26,9 +29,10 @@ enum CallStatus { PROCESS = 0, FINISH }; // https://stackoverflow.com/questions/41732884/grpc-multiple-services-in-cpp-async-server class RequestBase { public: - explicit RequestBase(sendrecv::SendRecvService::AsyncService* service, - grpc::ServerCompletionQueue* cq) - : service_(service), cq_(cq), status_(PROCESS) { + explicit RequestBase(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + const platform::DeviceContext* dev_ctx) + : service_(service), cq_(cq), status_(PROCESS), dev_ctx_(dev_ctx) { PADDLE_ENFORCE(cq_); } virtual ~RequestBase() {} @@ -42,55 +46,58 @@ class RequestBase { } protected: - grpc::ServerContext ctx_; - sendrecv::SendRecvService::AsyncService* service_; - grpc::ServerCompletionQueue* cq_; + ::grpc::ServerContext ctx_; + GrpcService::AsyncService* service_; + ::grpc::ServerCompletionQueue* cq_; CallStatus status_; + const platform::DeviceContext* dev_ctx_; }; -typedef std::pair MessageWithName; - class RequestSend final : public RequestBase { public: - explicit RequestSend(sendrecv::SendRecvService::AsyncService* service, - grpc::ServerCompletionQueue* cq, - SimpleBlockQueue* queue) - : RequestBase(service, cq), queue_(queue), responder_(&ctx_) { - service_->RequestSendVariable(&ctx_, &request_, &responder_, cq_, cq_, - this); + explicit RequestSend(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + framework::Scope* scope, ReceivedQueue* queue, + const platform::DeviceContext* dev_ctx) + : RequestBase(service, cq, dev_ctx), queue_(queue), responder_(&ctx_) { + request_.reset(new VariableResponse(scope, dev_ctx_)); + int method_id = static_cast(detail::GrpcMethod::kSendVariable); + service_->RequestAsyncUnary(method_id, &ctx_, request_.get(), &responder_, + cq_, cq_, this); } virtual ~RequestSend() {} - virtual std::string GetReqName() { return request_.varname(); } + virtual std::string GetReqName() { return request_->Varname(); } virtual void Process() { - MessageWithName msg_with_name = - std::make_pair(request_.varname(), std::move(request_)); - queue_->Push(std::move(msg_with_name)); - responder_.Finish(reply_, grpc::Status::OK, this); + queue_->Push(std::make_pair(request_->Varname(), request_)); + + sendrecv::VoidMessage reply; + responder_.Finish(reply, ::grpc::Status::OK, this); status_ = FINISH; } protected: - sendrecv::VariableMessage request_; - sendrecv::VoidMessage reply_; - SimpleBlockQueue* queue_; + std::shared_ptr request_; + ReceivedQueue* queue_; ServerAsyncResponseWriter responder_; }; class RequestGet final : public RequestBase { public: - explicit RequestGet(sendrecv::SendRecvService::AsyncService* service, - grpc::ServerCompletionQueue* cq, framework::Scope* scope, + explicit RequestGet(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + framework::Scope* scope, const platform::DeviceContext* dev_ctx, SimpleBlockQueue* queue) - : RequestBase(service, cq), + : RequestBase(service, cq, dev_ctx), responder_(&ctx_), scope_(scope), - dev_ctx_(dev_ctx), queue_(queue) { - service_->RequestGetVariable(&ctx_, &request_, &responder_, cq_, cq_, this); + int method_id = static_cast(detail::GrpcMethod::kGetVariable); + service_->RequestAsyncUnary(method_id, &ctx_, &request_, &responder_, cq_, + cq_, this); } virtual ~RequestGet() {} @@ -101,27 +108,72 @@ class RequestGet final : public RequestBase { // proc request. std::string var_name = request_.varname(); auto* var = scope_->FindVar(var_name); + + ::grpc::ByteBuffer reply; if (var_name != FETCH_BARRIER_MESSAGE) { - SerializeToMessage(var_name, var, *dev_ctx_, &reply_); + SerializeToByteBuffer(var_name, var, *dev_ctx_, &reply); } - // TODO(gongwb): check var's info. - responder_.Finish(reply_, grpc::Status::OK, this); + + responder_.Finish(reply, ::grpc::Status::OK, this); status_ = FINISH; - MessageWithName msg_with_name = - // request name reply - std::make_pair(var_name, std::move(reply_)); - queue_->Push(msg_with_name); + + if (var_name == FETCH_BARRIER_MESSAGE) { + sendrecv::VariableMessage msg; + MessageWithName msg_with_name = std::make_pair(var_name, msg); + queue_->Push(msg_with_name); + } } protected: sendrecv::VariableMessage request_; - sendrecv::VariableMessage reply_; - ServerAsyncResponseWriter responder_; + ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; framework::Scope* scope_; - const platform::DeviceContext* dev_ctx_; SimpleBlockQueue* queue_; }; +class RequestPrefetch final : public RequestBase { + public: + explicit RequestPrefetch(GrpcService::AsyncService* service, + ::grpc::ServerCompletionQueue* cq, + framework::Scope* scope, + const platform::DeviceContext* dev_ctx, + framework::Executor* executor, + framework::ProgramDesc* program, int blkid) + : RequestBase(service, cq, dev_ctx), + responder_(&ctx_), + scope_(scope), + executor_(executor), + program_(program), + blkid_(blkid) { + int method_id = static_cast(detail::GrpcMethod::kPrefetchVariable); + service_->RequestAsyncUnary(method_id, &ctx_, &request_, &responder_, cq_, + cq_, this); + } + + virtual ~RequestPrefetch() {} + + virtual std::string GetReqName() { return request_.varname(); } + + virtual void Process() { + // prefetch process... + ::grpc::ByteBuffer reply; + // TODO(Yancey1989): execute the Block which containers prefetch ops + + VLOG(3) << "RequestPrefetch Process in"; + + responder_.Finish(reply, ::grpc::Status::OK, this); + status_ = FINISH; + } + + protected: + sendrecv::VariableMessage request_; + ServerAsyncResponseWriter<::grpc::ByteBuffer> responder_; + framework::Scope* scope_; + framework::Executor* executor_; + framework::ProgramDesc* program_; + int blkid_; +}; + void AsyncGRPCServer::WaitClientGet(int count) { int fetch_barriers = 0; while (fetch_barriers < count) { @@ -133,14 +185,15 @@ void AsyncGRPCServer::WaitClientGet(int count) { } void AsyncGRPCServer::RunSyncUpdate() { - grpc::ServerBuilder builder; - builder.AddListeningPort(address_, grpc::InsecureServerCredentials()); + ::grpc::ServerBuilder builder; + builder.AddListeningPort(address_, ::grpc::InsecureServerCredentials()); builder.SetMaxSendMessageSize(std::numeric_limits::max()); builder.SetMaxReceiveMessageSize(std::numeric_limits::max()); builder.RegisterService(&service_); cq_send_ = builder.AddCompletionQueue(); cq_get_ = builder.AddCompletionQueue(); + cq_prefetch_ = builder.AddCompletionQueue(); server_ = builder.BuildAndStart(); LOG(INFO) << "Server listening on " << address_ << std::endl; @@ -149,6 +202,8 @@ void AsyncGRPCServer::RunSyncUpdate() { std::bind(&AsyncGRPCServer::TryToRegisterNewSendOne, this); std::function get_register = std::bind(&AsyncGRPCServer::TryToRegisterNewGetOne, this); + std::function prefetch_register = + std::bind(&AsyncGRPCServer::TryToRegisterNewPrefetchOne, this); t_send_.reset( new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, @@ -157,39 +212,45 @@ void AsyncGRPCServer::RunSyncUpdate() { t_get_.reset( new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, cq_get_.get(), "cq_get", get_register))); - + t_prefetch_.reset(new std::thread( + std::bind(&AsyncGRPCServer::HandleRequest, this, cq_prefetch_.get(), + "cq_prefetch", prefetch_register))); // wait server server_->Wait(); t_send_->join(); t_get_->join(); + t_prefetch_->join(); } void AsyncGRPCServer::ShutdownQueue() { std::unique_lock lock(cq_mutex_); cq_send_->Shutdown(); cq_get_->Shutdown(); - is_shut_down_ = true; + cq_prefetch_->Shutdown(); } // This URL explains why shutdown is complicate: void AsyncGRPCServer::ShutDown() { - server_->Shutdown(); + is_shut_down_ = true; ShutdownQueue(); + server_->Shutdown(); } void AsyncGRPCServer::TryToRegisterNewSendOne() { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { + VLOG(3) << "shutdown, do not TryToRegisterNewSendOne"; return; } - RequestSend* send = - new RequestSend(&service_, cq_send_.get(), &var_recv_queue_); + RequestSend* send = new RequestSend(&service_, cq_send_.get(), scope_, + &var_recv_queue_, dev_ctx_); VLOG(4) << "Create RequestSend status:" << send->Status(); } void AsyncGRPCServer::TryToRegisterNewGetOne() { std::unique_lock lock(cq_mutex_); if (is_shut_down_) { + VLOG(3) << "shutdown, do not TryToRegisterNewGetOne"; return; } RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_, dev_ctx_, @@ -197,33 +258,49 @@ void AsyncGRPCServer::TryToRegisterNewGetOne() { VLOG(4) << "Create RequestGet status:" << get->Status(); } +void AsyncGRPCServer::TryToRegisterNewPrefetchOne() { + std::unique_lock lock(cq_mutex_); + if (is_shut_down_) { + VLOG(3) << "shutdown, do not TryToRegisterNewPrefetchOne"; + return; + } + RequestPrefetch* prefetch = + new RequestPrefetch(&service_, cq_prefetch_.get(), scope_, dev_ctx_, + executor_, program_, prefetch_blk_id_); + + VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status(); +} + // FIXME(typhoonzero): change cq_name to enum. -void AsyncGRPCServer::HandleRequest(grpc::ServerCompletionQueue* cq, +void AsyncGRPCServer::HandleRequest(::grpc::ServerCompletionQueue* cq, std::string cq_name, std::function TryToRegisterNewOne) { TryToRegisterNewOne(); void* tag = NULL; bool ok = false; + while (true) { + VLOG(3) << "HandleRequest for " << cq_name << " while in"; if (!cq->Next(&tag, &ok)) { - LOG(INFO) << cq_name << " get CompletionQueue shutdown!"; + LOG(INFO) << cq_name << " CompletionQueue shutdown!"; break; } + VLOG(3) << "HandleRequest for " << cq_name << " while after Next"; PADDLE_ENFORCE(tag); // FIXME(typhoonzero): de-couple the barriers with recv_op - if (cq_name == "cq_get") WaitCond(1); - if (cq_name == "cq_send") WaitCond(0); + if (!is_shut_down_ && cq_name == "cq_get") WaitCond(1); + if (!is_shut_down_ && cq_name == "cq_send") WaitCond(0); - RequestBase* base = (RequestBase*)tag; + RequestBase* base = reinterpret_cast(tag); // reference: // https://github.com/tensorflow/tensorflow/issues/5596 // https://groups.google.com/forum/#!topic/grpc-io/xftlRy-IQwM // https://groups.google.com/forum/#!topic/grpc-io/ywATt88Ef_I if (!ok) { - LOG(WARNING) << cq_name << " recv no regular event:argument name" - << base->GetReqName(); + LOG(WARNING) << cq_name << " recv no regular event:argument name[" + << base->GetReqName() << "]"; TryToRegisterNewOne(); delete base; continue; diff --git a/paddle/fluid/operators/detail/grpc_server.h b/paddle/fluid/operators/detail/grpc_server.h index b6666bcf96e484b0b17b935c0efb2930f19b19f2..b0596d3cd1e108f28e8f1485d6b5c989c55be7e9 100644 --- a/paddle/fluid/operators/detail/grpc_server.h +++ b/paddle/fluid/operators/detail/grpc_server.h @@ -14,28 +14,34 @@ limitations under the License. */ #pragma once +#include +#include +#include + +#include "paddle/fluid/framework/executor.h" #include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" #include "paddle/fluid/framework/var_type.h" -#include "paddle/fluid/operators/detail/simple_block_queue.h" - +#include "paddle/fluid/operators/detail/grpc_service.h" #include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" #include "paddle/fluid/operators/detail/send_recv.pb.h" - -#include -#include -#include #include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/operators/detail/simple_block_queue.h" namespace paddle { namespace operators { namespace detail { +typedef std::pair> + ReceivedMessage; +typedef SimpleBlockQueue ReceivedQueue; + typedef std::pair MessageWithName; class RequestBase; -class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { +class AsyncGRPCServer final { public: explicit AsyncGRPCServer(const std::string &address) : address_(address) {} @@ -50,34 +56,46 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { void SetDevCtx(const platform::DeviceContext *dev_ctx) { dev_ctx_ = dev_ctx; } - const MessageWithName Get() { return this->var_recv_queue_.Pop(); } + void SetProgram(framework::ProgramDesc *program) { program_ = program; } + + void SetPrefetchBlkdId(int blkid) { prefetch_blk_id_ = blkid; } + + void SetExecutor(framework::Executor *executor) { executor_ = executor; } - void Push(const MessageWithName &msg) { this->var_recv_queue_.Push(msg); } + const ReceivedMessage Get() { return this->var_recv_queue_.Pop(); } + + void Push(const std::string &msg_name) { + this->var_recv_queue_.Push(std::make_pair(msg_name, nullptr)); + } void ShutDown(); protected: - void HandleRequest(grpc::ServerCompletionQueue *cq, std::string cq_name, + void HandleRequest(::grpc::ServerCompletionQueue *cq, std::string cq_name, std::function TryToRegisterNewOne); void TryToRegisterNewSendOne(); void TryToRegisterNewGetOne(); + void TryToRegisterNewPrefetchOne(); void ShutdownQueue(); private: std::mutex cq_mutex_; volatile bool is_shut_down_ = false; - std::unique_ptr cq_send_; - std::unique_ptr cq_get_; + std::unique_ptr<::grpc::ServerCompletionQueue> cq_send_; + std::unique_ptr<::grpc::ServerCompletionQueue> cq_get_; + std::unique_ptr<::grpc::ServerCompletionQueue> cq_prefetch_; - sendrecv::SendRecvService::AsyncService service_; - std::unique_ptr server_; + GrpcService::AsyncService service_; + std::unique_ptr<::grpc::Server> server_; std::string address_; framework::Scope *scope_; const platform::DeviceContext *dev_ctx_; + // received variable from RPC, operators fetch variable from this queue. - SimpleBlockQueue var_recv_queue_; SimpleBlockQueue var_get_queue_; + // client send variable to this queue. + ReceivedQueue var_recv_queue_; // condition of the sub program std::mutex barrier_mutex_; @@ -86,6 +104,11 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service { std::unique_ptr t_send_; std::unique_ptr t_get_; + std::unique_ptr t_prefetch_; + + int prefetch_blk_id_; + framework::ProgramDesc *program_; + framework::Executor *executor_; }; }; // namespace detail diff --git a/paddle/fluid/operators/detail/grpc_server_test.cc b/paddle/fluid/operators/detail/grpc_server_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..1ad62863a1a98c28cb08f47dfa8a5bfae463ba91 --- /dev/null +++ b/paddle/fluid/operators/detail/grpc_server_test.cc @@ -0,0 +1,62 @@ +/* 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 +#include +#include + +#include "gtest/gtest.h" +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/detail/grpc_server.h" + +namespace framework = paddle::framework; +namespace platform = paddle::platform; +namespace detail = paddle::operators::detail; + +std::unique_ptr rpc_service_; + +void StartServer(const std::string& endpoint) { + rpc_service_.reset(new detail::AsyncGRPCServer(endpoint)); + rpc_service_->RunSyncUpdate(); +} + +TEST(PREFETCH, CPU) { + // start up a server instance backend + // TODO(Yancey1989): Need to start a server with optimize blocks and + // prefetch blocks. + std::thread server_thread(StartServer, "127.0.0.1:8889"); + framework::Scope scope; + platform::CPUPlace place; + platform::CPUDeviceContext ctx(place); + // create var on local scope + std::string in_var_name("in"); + std::string out_var_name("out"); + auto* in_var = scope.Var(in_var_name); + auto* in_tensor = in_var->GetMutable(); + in_tensor->Resize({10, 10}); + VLOG(3) << "before mutable_data"; + in_tensor->mutable_data(place); + + scope.Var(out_var_name); + + VLOG(3) << "before fetch"; + detail::RPCClient client; + client.AsyncPrefetchVariable("127.0.0.1:8889", ctx, scope, in_var_name, + out_var_name); + client.Wait(); + + rpc_service_->ShutDown(); + server_thread.join(); + rpc_service_.reset(nullptr); +} diff --git a/paddle/fluid/operators/detail/grpc_service.h b/paddle/fluid/operators/detail/grpc_service.h new file mode 100644 index 0000000000000000000000000000000000000000..e6dab2f5a3a4280f3979417c3ca2d884a0b8ff2f --- /dev/null +++ b/paddle/fluid/operators/detail/grpc_service.h @@ -0,0 +1,121 @@ +// 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 +#include +#include +#include +#include +#include "paddle/fluid/operators/detail/variable_response.h" + +// NOTE: This method was originally created by tensorflow +// (https://github.com/tensorflow/tensorflow/) we borrow this +// method and did some modifications so that we can parse gRPC +// requests without too much copying of the tensor data. + +namespace grpc { +class CompletionQueue; +class Channel; +class RpcService; +class ServerCompletionQueue; +class ServerContext; + +// Support parsing/unparsing of tensorflow::VariableResponse. +// Wire-format is identical to RecvVariableResponse. +template <> +class SerializationTraits { + public: + static Status Serialize( + const paddle::operators::detail::VariableResponse& msg, + grpc_byte_buffer** bp, bool* own_buffer) { + PADDLE_ENFORCE(false, "SerializationTraits::Serialize not implemented!"); + return Status(); + } + static Status Deserialize(grpc_byte_buffer* buffer, + paddle::operators::detail::VariableResponse* msg, + int max_message_size = INT_MAX) { + if (buffer == nullptr) { + return Status(StatusCode::INTERNAL, "No payload"); + } + + Status result = g_core_codegen_interface->ok(); + if (result.ok()) { + paddle::operators::detail::GrpcByteSource source(buffer); + int ret = msg->Parse(&source); + if (ret != 0) { + result = Status(StatusCode::INTERNAL, "VariableResponse parse error"); + } + } + g_core_codegen_interface->grpc_byte_buffer_destroy(buffer); + return result; + } +}; +} // namespace grpc + +namespace paddle { +namespace operators { +namespace detail { + +enum class GrpcMethod { + kSendVariable, + kGetVariable, + kPrefetchVariable, +}; + +static const int kGrpcNumMethods = + static_cast(GrpcMethod::kPrefetchVariable) + 1; + +inline const char* GrpcMethodName(GrpcMethod id) { + switch (id) { + case GrpcMethod::kSendVariable: + return "/sendrecv.SendRecvService/SendVariable"; + case GrpcMethod::kGetVariable: + return "/sendrecv.SendRecvService/GetVariable"; + case GrpcMethod::kPrefetchVariable: + return "/sendrecv.SendRecvService/PrefetchVariable"; + } + + // Shouldn't be reached. + PADDLE_ENFORCE(false, "Invalid id: not found valid method name"); + return nullptr; +} + +class GrpcService final { + public: + class AsyncService : public ::grpc::Service { + public: + AsyncService() { + for (int i = 0; i < kGrpcNumMethods; ++i) { + AddMethod(new ::grpc::internal::RpcServiceMethod( + GrpcMethodName(static_cast(i)), + ::grpc::internal::RpcMethod::NORMAL_RPC, nullptr)); + ::grpc::Service::MarkMethodAsync(i); + } + } + virtual ~AsyncService() {} + + // Make RequestAsyncUnary public for grpc_call.h + using ::grpc::Service::RequestAsyncUnary; + }; +}; + +} // namespace detail +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/detail/send_recv.proto b/paddle/fluid/operators/detail/send_recv.proto index b0215d4a80c9440f09c35434903fd6166b03e8b0..fc12e82a7e6bd10262092d1ca367980df64e91c2 100644 --- a/paddle/fluid/operators/detail/send_recv.proto +++ b/paddle/fluid/operators/detail/send_recv.proto @@ -21,6 +21,8 @@ service SendRecvService { rpc SendVariable(VariableMessage) returns (VoidMessage) {} // Argument VariableMessage for GetVariable should only contain varname. rpc GetVariable(VariableMessage) returns (VariableMessage) {} + // Prefetch variable by Ids + rpc PrefetchVariable(VariableMessage) returns (VariableMessage) {} } // VariableMessage is serialized paddle variable message. @@ -32,6 +34,9 @@ enum VarType { SELECTED_ROWS = 1; } +// NOTICE(gongwb):don't modify this proto if you are not +// not familar with how we serialize in sendrecvop_utils.h +// and deserilize it in variable_response.h. message VariableMessage { enum Type { // Pod Types @@ -45,7 +50,6 @@ message VariableMessage { } message LodData { repeated int64 lod_data = 1; } - string varname = 1; // TODO(Yancey1989): reference framework::proto::VarDesc::VarType VarType type = 2; @@ -57,10 +61,12 @@ message VariableMessage { // lod details: int64 lod_level = 5; repeated LodData lod = 6; + // selected_rows height, aka. original dim0 + int64 slr_height = 7; // tensor data - bytes serialized = 7; + bytes serialized = 8; // selected_rows data - bytes rows = 8; + bytes rows = 9; } message VoidMessage {} diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.cc b/paddle/fluid/operators/detail/sendrecvop_utils.cc index 39117eeeb611b025c426938c60ddf82c6af232ca..7e3f015dabdb3fd6190d1ca2f422aa526e8889cd 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.cc +++ b/paddle/fluid/operators/detail/sendrecvop_utils.cc @@ -13,61 +13,19 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include +#include #include "google/protobuf/io/coded_stream.h" #include "google/protobuf/io/zero_copy_stream.h" #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/operators/detail/bytebuffer_stream.h" #include "paddle/fluid/operators/detail/proto_encoder_helper.h" +#include "paddle/fluid/operators/detail/variable_response.h" namespace paddle { namespace operators { namespace detail { -void SerializeToMessage(const std::string& name, const framework::Variable* var, - const platform::DeviceContext& ctx, - sendrecv::VariableMessage* msg) { - msg->set_varname(name); - std::ostringstream oss; - switch (framework::ToVarType(var->Type())) { - case framework::proto::VarType_Type_LOD_TENSOR: - msg->set_type(sendrecv::VarType::LOD_TENSOR); - framework::SerializeToStream(oss, var->Get(), ctx); - break; - case framework::proto::VarType_Type_SELECTED_ROWS: - msg->set_type(sendrecv::VarType::SELECTED_ROWS); - framework::SerializeToStream(oss, var->Get(), - ctx); - break; - default: { - PADDLE_THROW("Serialize does not support type: %s", - typeid(var->Type()).name()); - break; - } - } - msg->set_serialized(oss.str()); -} - -void DeserializeFromMessage(const sendrecv::VariableMessage& msg, - const platform::DeviceContext& ctx, - framework::Variable* var) { - std::istringstream iss(msg.serialized()); - switch (msg.type()) { - case sendrecv::VarType::LOD_TENSOR: - DeserializeFromStream(iss, var->GetMutable(), ctx); - break; - case sendrecv::VarType::SELECTED_ROWS: { - DeserializeFromStream(iss, var->GetMutable(), - ctx); - break; - } - default: { - PADDLE_THROW("Deserialize does not support type: %s", - typeid(var->Type()).name()); - break; - } - } -} - void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, ::grpc::ByteBuffer* msg) { @@ -123,6 +81,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, static_cast(ctx); auto copy_size = tensor.memory_size(); payload = memory::Alloc(cpu, copy_size); + memory::Copy(cpu, payload, boost::get(tensor.place()), reinterpret_cast(tensor.data()), @@ -132,6 +91,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, platform::CPUPlace cpu; memory::Free(cpu, backing); }; + #endif } else { payload = tensor.data(); @@ -148,6 +108,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, e.WriteUint64(VarMsg::kDimsFieldNumber, dim); } e.WriteUint64(VarMsg::kLodLevelFieldNumber, 0); + e.WriteUint64(VarMsg::kSlrHeightFieldNumber, slr->height()); auto* tensor = slr->mutable_value(); if (platform::is_gpu_place(ctx.GetPlace())) { #ifdef PADDLE_WITH_CUDA @@ -194,7 +155,7 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, ProtoEncodeHelper e2((char*)buf, 128); // NOTE: rows is of type int64_t size_t rows_memory_size = - slr->rows().capacity() * framework::SizeOfType(typeid(int64_t)); + slr->rows().size() * framework::SizeOfType(typeid(int64_t)); e2.WriteVarlengthBeginning(VarMsg::kRowsFieldNumber, rows_memory_size); slices[2] = ::grpc::Slice(e2.size()); memcpy(const_cast(slices[2].begin()), e2.data(), e2.size()); @@ -219,80 +180,11 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, - framework::Variable* var) { - sendrecv::VariableMessage meta; - GrpcByteBufferSource source; - source.Init(msg); - ::google::protobuf::io::CodedInputStream input(&source); - // do zerocopy parsing - PADDLE_ENFORCE(meta.ParseFromCodedStream(&input)); - PADDLE_ENFORCE(input.ConsumedEntireMessage()); - // dims is needed by both tensor and selectedrows - std::vector vecdims; - for (auto& d : meta.dims()) { - vecdims.push_back(d); - } - framework::DDim dims = framework::make_ddim(vecdims); - - if (meta.type() == sendrecv::LOD_TENSOR) { - auto* tensor = var->GetMutable(); - tensor->Resize(dims); - void* tensor_data = tensor->mutable_data( - ctx.GetPlace(), - paddle::operators::detail::ToTypeIndex(meta.data_type())); - framework::LoD lod; - for (int i = 0; i < meta.lod_level(); ++i) { - framework::Vector v; - for (int j = 0; j < meta.lod(i).lod_data_size(); ++j) { - v.push_back(meta.lod(i).lod_data(j)); - } - lod.push_back(v); - } - tensor->set_lod(lod); - // How to avoid copying and use the message buffer directly? - // Maybe need to find a way to release all memory except tensor content. - if (platform::is_gpu_place(ctx.GetPlace())) { -#ifdef PADDLE_WITH_CUDA - platform::CPUPlace cpu; - auto& gpu_dev_ctx = static_cast(ctx); - memory::Copy(boost::get(tensor->place()), - tensor_data, cpu, - reinterpret_cast(meta.serialized().data()), - meta.serialized().size(), gpu_dev_ctx.stream()); - ctx.Wait(); -#endif - } else { - memcpy(tensor_data, - reinterpret_cast(meta.serialized().data()), - meta.serialized().size()); - } - } else if (meta.type() == sendrecv::SELECTED_ROWS) { - auto* slr = var->GetMutable(); - auto* tensor = slr->mutable_value(); - int64_t* rows_data = slr->mutable_rows()->data(); - tensor->Resize(dims); - void* tensor_data = tensor->mutable_data( - ctx.GetPlace(), - paddle::operators::detail::ToTypeIndex(meta.data_type())); - if (platform::is_gpu_place(ctx.GetPlace())) { -#ifdef PADDLE_WITH_CUDA - platform::CPUPlace cpu; - auto& gpu_dev_ctx = static_cast(ctx); - memory::Copy(boost::get(tensor->place()), - tensor_data, cpu, - reinterpret_cast(meta.serialized().data()), - meta.serialized().size(), gpu_dev_ctx.stream()); - ctx.Wait(); -#endif - } else { - memcpy(tensor_data, - reinterpret_cast(meta.serialized().data()), - meta.serialized().size()); - } - // copy rows CPU data, GPU data will be copied lazly - memcpy(rows_data, reinterpret_cast(meta.rows().data()), - meta.rows().size()); - } + const framework::Scope* scope, + framework::Variable*& var) { + operators::detail::VariableResponse resp(scope, &ctx); + PADDLE_ENFORCE(resp.Parse(msg) == 0, "parse bytebuffer to tensor error!"); + var = resp.GetVar(); } } // namespace detail diff --git a/paddle/fluid/operators/detail/sendrecvop_utils.h b/paddle/fluid/operators/detail/sendrecvop_utils.h index 4fa6aefd3e0b1bd45ac52b1eff3b29126d79f03a..b3b2b8469c8f19313038f2551ab04708a05656d5 100644 --- a/paddle/fluid/operators/detail/sendrecvop_utils.h +++ b/paddle/fluid/operators/detail/sendrecvop_utils.h @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once +#include #include #include #include @@ -21,6 +22,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/scope.h" #include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/framework/var_type.h" #include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" @@ -34,15 +36,13 @@ namespace detail { #define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV" #define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV" -typedef void (*DestroyCallback)(void*); - -void SerializeToMessage(const std::string& name, const framework::Variable* var, - const platform::DeviceContext& ctx, - sendrecv::VariableMessage* msg); +static int64_t GetTimestamp() { + struct timeval tp; + gettimeofday(&tp, NULL); + return tp.tv_sec * 1000 + tp.tv_usec / 1000; +} -void DeserializeFromMessage(const sendrecv::VariableMessage& msg, - const platform::DeviceContext& ctx, - framework::Variable* var); +typedef void (*DestroyCallback)(void*); void SerializeToByteBuffer(const std::string& name, framework::Variable* var, const platform::DeviceContext& ctx, @@ -50,7 +50,8 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var, void DeserializeFromByteBuffer(const ::grpc::ByteBuffer& msg, const platform::DeviceContext& ctx, - framework::Variable* var); + const framework::Scope* scope, + framework::Variable*& var); inline std::type_index ToTypeIndex(sendrecv::VariableMessage::Type type) { switch (type) { diff --git a/paddle/fluid/operators/detail/test_serde.cc b/paddle/fluid/operators/detail/serde_test.cc similarity index 73% rename from paddle/fluid/operators/detail/test_serde.cc rename to paddle/fluid/operators/detail/serde_test.cc index 2f06e5a686b996858d21930a1afa2861efca4a9b..ea1670e56f3c2fedc2617db1425472e52c6519f5 100644 --- a/paddle/fluid/operators/detail/test_serde.cc +++ b/paddle/fluid/operators/detail/serde_test.cc @@ -16,11 +16,13 @@ limitations under the License. */ #include #include +#include #include "gtest/gtest.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/framework/variable.h" #include "paddle/fluid/operators/detail/sendrecvop_utils.h" +#include "paddle/fluid/operators/detail/variable_response.h" #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/platform/place.h" #include "paddle/fluid/string/printf.h" @@ -31,19 +33,21 @@ namespace operators = paddle::operators; namespace math = paddle::operators::math; namespace memory = paddle::memory; -void RunSerdeTestTensor(platform::Place place) { - // serialize var to ByteBuffer - framework::Variable var; - auto* tensor = var.GetMutable(); - tensor->Resize(framework::make_ddim({4, 8, 4, 2})); - framework::LoD lod; - lod.push_back(framework::Vector({1, 3, 8})); - tensor->set_lod(lod); - int tensor_numel = 4 * 8 * 4 * 2; +void RunSerdeTestSelectedRows(platform::Place place) { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); + + // serialize var to ByteBuffer + framework::Variable var; + auto* slr = var.GetMutable(); + slr->set_height(1000); + auto* tensor = slr->mutable_value(); + auto* rows = slr->mutable_rows(); + tensor->Resize(framework::make_ddim({564, 128})); tensor->mutable_data(place); - math::set_constant(ctx, tensor, 31.9); + int tensor_numel = 564 * 128; + math::set_constant(ctx, tensor, 32.7); + for (int i = 0; i < 564; ++i) rows->push_back(i); ::grpc::ByteBuffer msg; operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg); @@ -56,62 +60,72 @@ void RunSerdeTestTensor(platform::Place place) { for (const auto& s : slices) { tmp.append(reinterpret_cast(s.begin()), s.size()); } + sendrecv::VariableMessage varmsg; EXPECT_TRUE(varmsg.ParseFromString(tmp)); + + // deserialize bytebuffer EXPECT_EQ(varmsg.varname(), "myvar"); - EXPECT_EQ(varmsg.type(), 0); - EXPECT_EQ(varmsg.dims()[0], 4); - EXPECT_EQ(varmsg.dims()[1], 8); - EXPECT_EQ(varmsg.dims()[2], 4); - EXPECT_EQ(varmsg.dims()[3], 2); - EXPECT_EQ(varmsg.lod_level(), 1); - EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); - EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); - EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); + EXPECT_EQ(varmsg.type(), 1); const float* tensor_data = reinterpret_cast(varmsg.serialized().data()); + const int64_t* rows_data = + reinterpret_cast(varmsg.rows().data()); for (int i = 0; i < tensor_numel; ++i) { - EXPECT_FLOAT_EQ(tensor_data[i], 31.9); + EXPECT_FLOAT_EQ(tensor_data[i], 32.7); + } + for (int i = 0; i < 564; ++i) { + EXPECT_EQ(rows_data[i], i); } // deserialize zero-copy - framework::Variable var2; - operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2); - auto tensor2 = var2.Get(); + // framework::Variable var2; + // operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2); + framework::Scope scope; + scope.Var("myvar"); + operators::detail::VariableResponse resp(&scope, &ctx); + EXPECT_EQ(resp.Parse(msg), 0); + + framework::Variable* var2 = resp.GetVar(); + + auto* slr2 = var2->GetMutable(); + auto* tensor2 = slr2->mutable_value(); + auto* rows2 = slr2->mutable_rows(); float* tensor_data2 = nullptr; framework::Tensor tmp_tensor; if (platform::is_gpu_place(ctx.GetPlace())) { platform::CPUPlace cpu; - framework::TensorCopy(tensor2, cpu, &tmp_tensor); + framework::TensorCopy(*tensor2, cpu, &tmp_tensor); tensor_data2 = tmp_tensor.data(); } else { - tensor_data2 = const_cast(tensor2.data()); + tensor_data2 = const_cast(tensor2->data()); } + const int64_t* rows_data2 = rows2->data(); - EXPECT_EQ(varmsg.lod_level(), 1); - EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); - EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); - EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); - for (int i = 0; i < tensor_numel; ++i) EXPECT_FLOAT_EQ(tensor_data2[i], 31.9); + for (int i = 0; i < tensor_numel; ++i) { + EXPECT_FLOAT_EQ(tensor_data2[i], 32.7); + } + for (int i = 0; i < rows2->size(); ++i) { + EXPECT_EQ(rows_data2[i], i); + } + EXPECT_EQ(slr2->height(), 1000); } -void RunSerdeTestSelectedRows(platform::Place place) { - platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); - auto& ctx = *pool.Get(place); - +void RunTestLodTensor(platform::Place place, int from_type = 0) { // serialize var to ByteBuffer framework::Variable var; - auto* slr = var.GetMutable(); - auto* tensor = slr->mutable_value(); - auto* rows = slr->mutable_rows(); - tensor->Resize(framework::make_ddim({2, 10})); + auto* tensor = var.GetMutable(); + tensor->Resize(framework::make_ddim({4, 8, 4, 2})); + framework::LoD lod; + lod.push_back(framework::Vector({1, 3, 8})); + tensor->set_lod(lod); + int tensor_numel = 4 * 8 * 4 * 2; + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); tensor->mutable_data(place); - int tensor_numel = 2 * 10; - math::set_constant(ctx, tensor, 32.7); - rows->push_back(3); - rows->push_back(10); + math::set_constant(ctx, tensor, 31.9); ::grpc::ByteBuffer msg; operators::detail::SerializeToByteBuffer("myvar", &var, ctx, &msg); @@ -126,61 +140,82 @@ void RunSerdeTestSelectedRows(platform::Place place) { } sendrecv::VariableMessage varmsg; EXPECT_TRUE(varmsg.ParseFromString(tmp)); - EXPECT_EQ(varmsg.varname(), "myvar"); - EXPECT_EQ(varmsg.type(), 1); + EXPECT_EQ(varmsg.type(), 0); + EXPECT_EQ(varmsg.dims()[0], 4); + EXPECT_EQ(varmsg.dims()[1], 8); + EXPECT_EQ(varmsg.dims()[2], 4); + EXPECT_EQ(varmsg.dims()[3], 2); + EXPECT_EQ(varmsg.lod_level(), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); + EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); const float* tensor_data = reinterpret_cast(varmsg.serialized().data()); - const int64_t* rows_data = - reinterpret_cast(varmsg.rows().data()); for (int i = 0; i < tensor_numel; ++i) { - EXPECT_FLOAT_EQ(tensor_data[i], 32.7); + EXPECT_FLOAT_EQ(tensor_data[i], 31.9); } - EXPECT_EQ(rows_data[0], 3); - EXPECT_EQ(rows_data[1], 10); + + // message binary + std::string str; + varmsg.SerializeToString(&str); + + // message bytebuffer + ::grpc::Slice slices_2[1]; + int num_slices = 1; + slices_2[0] = ::grpc::Slice(str.length()); + memcpy(const_cast(slices_2[0].begin()), str.c_str(), str.length()); + ::grpc::ByteBuffer bytebuffer2(&slices_2[0], num_slices); + // deserialize zero-copy - framework::Variable var2; - operators::detail::DeserializeFromByteBuffer(msg, ctx, &var2); + framework::Scope scope; + scope.Var("myvar"); + operators::detail::VariableResponse resp(&scope, &ctx); + if (from_type == 0) { + EXPECT_EQ(resp.Parse(msg), 0); + } else { + EXPECT_EQ(resp.Parse(bytebuffer2), 0); + } - auto* slr2 = var2.GetMutable(); - auto* tensor2 = slr2->mutable_value(); - auto* rows2 = slr2->mutable_rows(); + framework::Variable* var2 = resp.GetVar(); + + auto tensor2 = var2->Get(); float* tensor_data2 = nullptr; framework::Tensor tmp_tensor; if (platform::is_gpu_place(ctx.GetPlace())) { platform::CPUPlace cpu; - framework::TensorCopy(*tensor2, cpu, &tmp_tensor); + framework::TensorCopy(tensor2, cpu, &tmp_tensor); tensor_data2 = tmp_tensor.data(); } else { - tensor_data2 = const_cast(tensor2->data()); + tensor_data2 = const_cast(tensor2.data()); } - const int64_t* rows_data2 = rows2->data(); - for (int i = 0; i < tensor_numel; ++i) { - EXPECT_FLOAT_EQ(tensor_data2[i], 32.7); - } - EXPECT_EQ(rows_data2[0], 3); - EXPECT_EQ(rows_data2[1], 10); + EXPECT_EQ(varmsg.lod_level(), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(0), 1); + EXPECT_EQ(varmsg.lod(0).lod_data(1), 3); + EXPECT_EQ(varmsg.lod(0).lod_data(2), 8); + for (int i = 0; i < tensor_numel; ++i) EXPECT_FLOAT_EQ(tensor_data2[i], 31.9); } -TEST(SelectedRows, CPU) { +TEST(LodTensor, Run) { platform::CPUPlace place; - RunSerdeTestSelectedRows(place); + RunTestLodTensor(place); + RunTestLodTensor(place, 1); +#ifdef PADDLE_WITH_CUDA + platform::CUDAPlace gpu(0); + RunTestLodTensor(gpu); + RunTestLodTensor(gpu, 1); +#endif } -TEST(SelectedRows, GPU) { - platform::CUDAPlace place; +TEST(SelectedRows, Run) { + platform::CPUPlace place; RunSerdeTestSelectedRows(place); -} -TEST(Tensor, CPU) { - platform::CPUPlace place; - RunSerdeTestTensor(place); +#ifdef PADDLE_WITH_CUDA + platform::CUDAPlace gpu; + RunSerdeTestSelectedRows(gpu); +#endif } - -TEST(Tensor, GPU) { - platform::CUDAPlace place; - RunSerdeTestTensor(place); -} \ No newline at end of file diff --git a/paddle/fluid/operators/detail/variable_response.cc b/paddle/fluid/operators/detail/variable_response.cc new file mode 100644 index 0000000000000000000000000000000000000000..f59c9b50bb36c12c9abc0a52e0d11c6a73217047 --- /dev/null +++ b/paddle/fluid/operators/detail/variable_response.cc @@ -0,0 +1,426 @@ +// 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/detail/variable_response.h" +#include +#include "paddle/fluid/operators/detail/send_recv.pb.h" +#include "paddle/fluid/operators/detail/sendrecvop_utils.h" + +namespace paddle { +namespace operators { +namespace detail { + +enum WireType { + WIRETYPE_VARINT = 0, + WIRETYPE_LENGTH_DELIMITED = 2, +}; + +inline int GetTagFieldNumber(uint32_t tag) { return tag >> 3; } + +inline WireType GetTagWireType(uint32_t tag) { + return static_cast(tag & 0x7); +} + +bool ReadVarintSizeAsInt(::google::protobuf::io::CodedInputStream* input, + int* result) { + uint64_t v; + if (input->ReadVarint64(&v) && v <= static_cast(INT_MAX)) { + *result = static_cast(v); + return true; + } else { + return false; + } +} + +bool ReadRaw(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& dev_ctx, platform::Place place, + void* dest, int size) { + const void* data = NULL; + int size_to_write = 0; + int length = size; + int total_written = 0; + + if (platform::is_gpu_place(place)) { +#ifdef PADDLE_WITH_CUDA + auto& gpu_dev_ctx = + static_cast(dev_ctx); + platform::CPUPlace cpu; + + char* p = reinterpret_cast(dest); + while (total_written < length) { + if (!input->GetDirectBufferPointer(&data, &size_to_write)) { + return false; + } + // NOTE: if raw buffer is large and have two neighbor fields of raw + // buffers GetDirectBufferPointer can get all of them, use length to + // truncate it. + if (total_written + size_to_write > length) { + size_to_write = length - total_written; + } + memory::Copy(boost::get(place), + reinterpret_cast(p), cpu, data, size_to_write, + gpu_dev_ctx.stream()); + p += size_to_write; + total_written += size_to_write; + + input->Skip(size_to_write); + } + gpu_dev_ctx.Wait(); +#else + PADDLE_THROW("Unexpected branch"); +#endif + return true; + } + + char* p = reinterpret_cast(dest); + while (total_written < length) { + if (!input->GetDirectBufferPointer(&data, &size_to_write)) { + return false; + } + // NOTE: if raw buffer is large and have two neighbor fields of raw buffers + // GetDirectBufferPointer can get all of them, use length to truncate it. + if (total_written + size_to_write > length) { + size_to_write = length - total_written; + } + // TODO(gongwb): can we avoid copy? + platform::CPUPlace cpu; + memory::Copy(cpu, reinterpret_cast(p), cpu, data, size_to_write); + + p += size_to_write; + total_written += size_to_write; + + input->Skip(size_to_write); + } + + return true; +} + +bool VariableResponse::CopyLodTensorData( + ::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, framework::DDim& dims, int length) { + auto var = scope_->FindVar(meta_.varname()); + auto* tensor = var->GetMutable(); + tensor->Resize(dims); + + framework::LoD lod; + for (int i = 0; i < meta_.lod_level(); ++i) { + framework::Vector v; + for (int j = 0; j < meta_.lod(i).lod_data_size(); ++j) { + v.push_back(meta_.lod(i).lod_data(j)); + } + lod.push_back(v); + } + tensor->set_lod(lod); + + void* tensor_data = + tensor->mutable_data(ctx.GetPlace(), ToTypeIndex(meta_.data_type())); + + if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) { + return false; + } + + return true; +} + +inline framework::DDim GetDims( + const ::google::protobuf::RepeatedField<::google::protobuf::int64>& dims) { + std::vector vecdims; + for (auto& d : dims) { + vecdims.push_back(d); + } + return framework::make_ddim(vecdims); +} + +bool VariableResponse::CopySelectRowsTensorData( + ::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, framework::DDim& dims, int length) { + auto var = scope_->FindVar(meta_.varname()); + auto* slr = var->GetMutable(); + slr->set_height(meta_.slr_height()); + auto* tensor = slr->mutable_value(); + tensor->Resize(dims); + PADDLE_ENFORCE_EQ( + tensor->numel(), + length / framework::SizeOfType( + paddle::operators::detail::ToTypeIndex(meta_.data_type()))); + void* tensor_data = tensor->mutable_data( + ctx.GetPlace(), + paddle::operators::detail::ToTypeIndex(meta_.data_type())); + + if (!ReadRaw(input, ctx, tensor->place(), tensor_data, length)) { + return false; + } + + return true; +} + +bool VariableResponse::CopySelectRowsData( + ::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, int length) { + auto var = scope_->FindVar(meta_.varname()); + auto* slr = var->GetMutable(); + slr->mutable_rows()->resize(length / + framework::SizeOfType(typeid(int64_t))); // int64 + int64_t* rows_data = slr->mutable_rows()->data(); + + // copy rows CPU data, GPU data will be copied lazily. + platform::CPUPlace cpu; + if (!ReadRaw(input, ctx, cpu, rows_data, length)) { + return false; + } + + return true; +} + +bool ParseLodData(::google::protobuf::io::CodedInputStream* input, + std::vector* lod) { + while (true) { + auto p = input->ReadTagWithCutoff(127); + int tag = GetTagFieldNumber(p.first); + WireType wt = GetTagWireType(p.first); + + if (!p.second) { + return (tag == 0); + } + + switch (tag) { + case sendrecv::VariableMessage_LodData::kLodDataFieldNumber: { + uint64_t v; + if (wt == WIRETYPE_VARINT) { + if (!input->ReadVarint64(&v)) { + return false; + } + lod->push_back(v); + break; + } + + if (wt == WIRETYPE_LENGTH_DELIMITED) { + int length = 0; + if (!input->ReadVarintSizeAsInt(&length)) { + return tag; + } + + for (int i = 0; i < length; i++) { + uint64_t v; + if (!input->ReadVarint64(&v)) { + return false; + } + lod->push_back(v); + } + break; + } + + return false; + } + default: { return false; } + } + } + + return true; +} + +int VariableResponse::Parse(const ::grpc::ByteBuffer& byte_buffer) { + GrpcByteBufferSource source; + source.Init(byte_buffer); + GrpcByteBufferSourceWrapper r(&source); + + return Parse(&r); +} + +int VariableResponse::Parse(Source* source) { + ::google::protobuf::io::ZeroCopyInputStream* input_stream = + source->contents(); + ::google::protobuf::io::CodedInputStream input(input_stream); + input.SetTotalBytesLimit(INT_MAX, INT_MAX); + + while (true) { + auto p = input.ReadTagWithCutoff(127); + int tag = GetTagFieldNumber(p.first); + WireType wt = GetTagWireType(p.first); + if (!p.second) { + if (tag != 0) { + return -1; + } + return 0; + } + + switch (tag) { + case sendrecv::VariableMessage::kVarnameFieldNumber: { + uint32_t length; + if ((wt != WIRETYPE_LENGTH_DELIMITED) || !input.ReadVarint32(&length)) { + return tag; + } + + std::string temp; + if (!input.ReadString(&temp, length)) { + return tag; + } + + meta_.set_varname(temp); + break; + } + case sendrecv::VariableMessage::kTypeFieldNumber: { + uint64_t v; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + return tag; + } + + meta_.set_type(static_cast<::sendrecv::VarType>(v)); + break; + } + case sendrecv::VariableMessage::kDataTypeFieldNumber: { + uint64_t v = 0; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + return tag; + } + + meta_.set_data_type(static_cast<::sendrecv::VariableMessage_Type>(v)); + break; + } + case sendrecv::VariableMessage::kDimsFieldNumber: { + // not packed + if (wt == WIRETYPE_VARINT) { + uint64_t v; + if (!input.ReadVarint64(&v)) { + return tag; + } + meta_.add_dims(v); + break; + } + + // packed + if (wt == WIRETYPE_LENGTH_DELIMITED) { + int length = 0; + if (!input.ReadVarintSizeAsInt(&length)) { + return tag; + } + for (int i = 0; i < length; i++) { + uint64_t v; + if (!input.ReadVarint64(&v)) { + return tag; + } + meta_.add_dims(v); + } + break; + } + + return tag; + } + case sendrecv::VariableMessage::kLodLevelFieldNumber: { + uint64_t v = 0; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + return tag; + } + meta_.set_lod_level(static_cast(v)); + break; + } + case sendrecv::VariableMessage::kLodFieldNumber: { + int length = 0; + if (wt != WIRETYPE_LENGTH_DELIMITED || + !ReadVarintSizeAsInt(&input, &length)) { + return tag; + } + + std::pair<::google::protobuf::io::CodedInputStream::Limit, int> p = + input.IncrementRecursionDepthAndPushLimit(length); + + std::vector lod_data; + if (p.second < 0 || !ParseLodData(&input, &lod_data)) { + return tag; + } + + if (!input.DecrementRecursionDepthAndPopLimit(p.first)) { + return false; + } + + if (lod_data.size() == 0) { + break; + } + + auto lod = meta_.add_lod(); + for (uint32_t i = 0; i < lod_data.size(); i++) { + lod->add_lod_data(lod_data[i]); + } + break; + } + case sendrecv::VariableMessage::kSlrHeightFieldNumber: { + uint64_t v = 0; + if ((wt != WIRETYPE_VARINT) || !input.ReadVarint64(&v)) { + return tag; + } + meta_.set_slr_height(static_cast(v)); + break; + } + case sendrecv::VariableMessage::kSerializedFieldNumber: { + PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || + meta_.type() == sendrecv::LOD_TENSOR) && + meta_.varname() != "", + "meta info should be got first!"); + + int length = 0; + if (wt != WIRETYPE_LENGTH_DELIMITED || + !ReadVarintSizeAsInt(&input, &length)) { + return tag; + } + + framework::DDim dims = GetDims(meta_.dims()); + if (meta_.type() == sendrecv::LOD_TENSOR) { + PADDLE_ENFORCE(meta_.lod_size() >= 0, + "lod info should be got first!"); + if (!CopyLodTensorData(&input, *dev_ctx_, dims, length)) { + return tag; + } + break; + } + + if (meta_.type() == sendrecv::SELECTED_ROWS) { + if (!CopySelectRowsTensorData(&input, *dev_ctx_, dims, length)) { + return tag; + } + break; + } + + return tag; + } + case sendrecv::VariableMessage::kRowsFieldNumber: { + PADDLE_ENFORCE((meta_.type() == sendrecv::SELECTED_ROWS || + meta_.type() == sendrecv::LOD_TENSOR) && + meta_.varname() != "", + "meta info should be got first!"); + + int length = 0; + if (wt != WIRETYPE_LENGTH_DELIMITED || + !ReadVarintSizeAsInt(&input, &length)) { + return tag; + } + + if (!CopySelectRowsData(&input, *dev_ctx_, length)) { + return tag; + } + break; + } + + default: { + // Unknown tag, return unknown error. + return -1; + } + } + } + + return 0; +} + +}; // namespace detail +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/detail/variable_response.h b/paddle/fluid/operators/detail/variable_response.h new file mode 100644 index 0000000000000000000000000000000000000000..e121ed7bce966d7dea94f71087f2187dcaa17cec --- /dev/null +++ b/paddle/fluid/operators/detail/variable_response.h @@ -0,0 +1,81 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/scope.h" +#include "paddle/fluid/framework/selected_rows.h" +#include "paddle/fluid/framework/var_type.h" + +#include "paddle/fluid/operators/detail/send_recv.grpc.pb.h" +#include "paddle/fluid/operators/detail/send_recv.pb.h" + +#include "google/protobuf/io/coded_stream.h" +#include "google/protobuf/io/zero_copy_stream.h" +#include "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/detail/bytebuffer_stream.h" + +namespace paddle { +namespace operators { +namespace detail { + +class VariableResponse { + public: + VariableResponse(const framework::Scope* scope, + const platform::DeviceContext* dev_ctx) + : scope_(scope), dev_ctx_(dev_ctx) {} + + virtual ~VariableResponse() {} + + // return: + // 0:ok. + // -1: unkown error. + // other: number of error field. + int Parse(Source* source); + + // return: + // 0:ok. + // -1: unkown error. + // other: number of error field. + int Parse(const ::grpc::ByteBuffer& byte_buffer); + + inline std::string Varname() { return meta_.varname(); } + + // should call parse first. + framework::Variable* GetVar() { return scope_->FindVar(meta_.varname()); } + + private: + bool CopySelectRowsTensorData(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, + framework::DDim& dims, int length); + + bool CopySelectRowsData(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, int length); + + bool CopyLodTensorData(::google::protobuf::io::CodedInputStream* input, + const platform::DeviceContext& ctx, + framework::DDim& dims, int length); + + private: + const framework::Scope* scope_; + const platform::DeviceContext* dev_ctx_; + // only Skeleton + sendrecv::VariableMessage meta_; +}; + +}; // namespace detail +}; // namespace operators +}; // namespace paddle diff --git a/paddle/fluid/operators/detection_map_op.cc b/paddle/fluid/operators/detection_map_op.cc index 73c84c2fe0155d21d7059938330e44fa3668c6df..93ef15b9332168a9c62abfd4d0827207173ece45 100644 --- a/paddle/fluid/operators/detection_map_op.cc +++ b/paddle/fluid/operators/detection_map_op.cc @@ -188,8 +188,8 @@ The general steps are as follows. First, calculate the true positive and } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(detection_map, ops::DetectionMAPOp, - ops::DetectionMAPOpMaker); +REGISTER_OPERATOR(detection_map, ops::DetectionMAPOp, ops::DetectionMAPOpMaker, + paddle::framework::EmptyGradOpMaker); REGISTER_OP_CPU_KERNEL( detection_map, ops::DetectionMAPOpKernel, ops::DetectionMAPOpKernel); diff --git a/paddle/fluid/operators/dropout_op.cu b/paddle/fluid/operators/dropout_op.cu index f6c85a2a537b37feb20e6d62729dc5075af2a5d9..184c095e487a302ebc4d251dd6f332333c415c6d 100644 --- a/paddle/fluid/operators/dropout_op.cu +++ b/paddle/fluid/operators/dropout_op.cu @@ -33,6 +33,7 @@ __global__ void RandomGenerator(const size_t n, const int seed, int idx = blockDim.x * blockIdx.x + threadIdx.x; for (; idx < n; idx += blockDim.x * gridDim.x) { + rng.discard(idx); if (dist(rng) < dropout_prob) { mask_data[idx] = static_cast(0); } else { @@ -54,9 +55,6 @@ class GPUDropoutKernel : public framework::OpKernel { y->mutable_data(context.GetPlace()); float dropout_prob = context.Attr("dropout_prob"); - auto X = EigenMatrix::Reshape(*x, 1); - auto Y = EigenMatrix::Reshape(*y, 1); - auto& place = *context.template device_context().eigen_device(); if (!context.Attr("is_test")) { auto* mask = context.Output("Mask"); @@ -75,6 +73,8 @@ class GPUDropoutKernel : public framework::OpKernel { T><<>>( size, seed, dropout_prob, x_data, mask_data, y_data); } else { + auto X = EigenMatrix::Reshape(*x, 1); + auto Y = EigenMatrix::Reshape(*y, 1); Y.device(place) = X * static_cast(1.0f - dropout_prob); } } diff --git a/paddle/fluid/operators/dropout_op.h b/paddle/fluid/operators/dropout_op.h index b5ee86ae2d11dfc835e1a3a6826ce016baf38a29..0628b4b826d2730a8e3fb4842e4ae550b8c00569 100644 --- a/paddle/fluid/operators/dropout_op.h +++ b/paddle/fluid/operators/dropout_op.h @@ -11,9 +11,10 @@ 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" diff --git a/paddle/fluid/operators/dropout_op_test.cc b/paddle/fluid/operators/dropout_op_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..424d273c34b7e8d70c88b591c4fe45db61465f38 --- /dev/null +++ b/paddle/fluid/operators/dropout_op_test.cc @@ -0,0 +1,104 @@ +/* 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 + +#include +#include // NOLINT +#include + +#include "gtest/gtest.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/operator.h" +#include "paddle/fluid/framework/program_desc.h" +#include "paddle/fluid/operators/dropout_op.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/string/printf.h" + +namespace f = paddle::framework; +namespace p = paddle::platform; +namespace m = paddle::operators::math; + +USE_OP(dropout); + +void Compare(f::Scope* scope, const p::DeviceContext& ctx) { + // init + auto var = scope->Var("X"); + auto tensor = var->GetMutable(); + tensor->Resize({10, 10}); + + std::vector init; + for (int64_t i = 0; i < 10 * 10; ++i) { + init.push_back(1.0); + } + + TensorFromVector(init, ctx, tensor); + + auto place = ctx.GetPlace(); + auto out_var = scope->Var("Out"); + auto out_tensor = out_var->GetMutable(); + out_tensor->Resize({10, 10}); + out_tensor->mutable_data(place); // allocate + + auto mask_var = scope->Var("Mask"); + auto mask_tensor = mask_var->GetMutable(); + mask_tensor->Resize({10, 10}); + mask_tensor->mutable_data(place); // allocate + + // run + f::AttributeMap attrs; + float dropout_prob = 0.5; + attrs.insert({"fix_seed", 1}); + attrs.insert({"seed", 3}); + attrs.insert({"dropout_prob", dropout_prob}); + auto dropout_op = f::OpRegistry::CreateOp( + "dropout", {{"X", {"X"}}}, {{"Out", {"Out"}}, {"Mask", {"Mask"}}}, attrs); + + dropout_op->Run(*scope, place); + + std::vector out_vec; + TensorToVector(*out_tensor, ctx, &out_vec); + + std::vector std_out = { + 0, 0, 1, 1, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 1, 1, 0, 1, + 1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0, 1, 1, 0, + 1, 0, 1, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, + 1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 1, 0, + 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 0, 0, 1, 1}; + + EXPECT_EQ(out_vec.size(), std_out.size()); + for (uint32_t i = 0; i < out_vec.size(); i++) { + EXPECT_EQ(out_vec[i], std_out[i]); + } +} + +// TODO(wyi): Due to +// https://github.com/PaddlePaddle/Paddle/issues/9507, I temporarily +// disable this test to remove the prevention of the merge of +// unrelated PRs. +/* +TEST(Dropout, CPUDense) { + f::Scope scope; + p::CPUPlace place; + p::CPUDeviceContext ctx(place); + Compare(scope, ctx); +} + +TEST(Dropout, GPUDense) { + f::Scope scope; + p::CUDAPlace place; + p::CUDADeviceContext ctx(place); + Compare(scope, ctx); +} +*/ diff --git a/paddle/fluid/operators/increment_op.cc b/paddle/fluid/operators/increment_op.cc index 6b5c3db13c0929ae0dd2fb2c981867df0a36c1ce..ec2e641679fedec776d48716f13445f44375ce3d 100644 --- a/paddle/fluid/operators/increment_op.cc +++ b/paddle/fluid/operators/increment_op.cc @@ -1,71 +1,46 @@ -/* 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" +// 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/increment_op.h" namespace paddle { namespace operators { -class IncrementInferShape : public framework::InferShapeBase { +class IncrementOp : public framework::OperatorWithKernel { public: - void operator()(framework::InferShapeContext *ctx) const override { + IncrementOp(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 IncrementOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of IncrementOp should not be null."); PADDLE_ENFORCE_EQ(1, framework::product(ctx->GetInputDim("X"))); ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", "Out"); } -}; - -struct IncrementFunctor { - IncrementFunctor(const framework::LoDTensor &x, framework::LoDTensor *out, - float value) - : x_(x), out_(out), value_(value) {} - - template - void operator()() const { - *out_->data() = *x_.data() + static_cast(value_); - } - - const framework::LoDTensor &x_; - framework::LoDTensor *out_; - float value_; -}; - -class IncrementOp : public framework::OperatorBase { - public: - IncrementOp(const std::string &type, const framework::VariableNameMap &inputs, - const framework::VariableNameMap &outputs, - const framework::AttributeMap &attrs) - : OperatorBase(type, inputs, outputs, attrs) {} - - private: - void RunImpl(const framework::Scope &scope, - const platform::Place &place) const override { - auto &x = scope.FindVar(Input("X"))->Get(); - auto &out = - *scope.FindVar(Output("Out"))->GetMutable(); - PADDLE_ENFORCE(platform::is_cpu_place(x.place())); - out.Resize(x.dims()); - out.mutable_data(x.place(), x.type()); - float value = Attr("step"); - VLOG(10) << Output("Out") << " increase " << Input("X") << " with " - << value; - framework::VisitDataType(framework::ToDataType(out.type()), - IncrementFunctor(x, &out, value)); + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + framework::OpKernelType kt = OperatorWithKernel::GetExpectedKernelType(ctx); + // IncrementOp kernel's device type is decided by input tensor place + kt.place_ = ctx.Input("X")->place(); + return kt; } }; @@ -108,5 +83,10 @@ class IncrementGradOpMaker : public framework::SingleGradOpDescMaker { } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(increment, ops::IncrementOp, ops::IncrementInferShape, - ops::IncrementOpMaker, ops::IncrementGradOpMaker); +REGISTER_OPERATOR(increment, ops::IncrementOp, ops::IncrementOpMaker, + ops::IncrementGradOpMaker); +REGISTER_OP_CPU_KERNEL( + increment, ops::IncrementKernel, + ops::IncrementKernel, + ops::IncrementKernel, + ops::IncrementKernel) diff --git a/paddle/fluid/operators/increment_op.cu b/paddle/fluid/operators/increment_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..7fb6425fe994751c4d7a025bb62e43a84c8d95c2 --- /dev/null +++ b/paddle/fluid/operators/increment_op.cu @@ -0,0 +1,22 @@ +// 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/increment_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL( + increment, ops::IncrementKernel, + ops::IncrementKernel, + ops::IncrementKernel, + ops::IncrementKernel) diff --git a/paddle/fluid/operators/increment_op.h b/paddle/fluid/operators/increment_op.h new file mode 100644 index 0000000000000000000000000000000000000000..d0e8c66255ef68b975701fb6b3c145be2590e271 --- /dev/null +++ b/paddle/fluid/operators/increment_op.h @@ -0,0 +1,39 @@ +// 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" + +namespace paddle { +namespace operators { + +template +class IncrementKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + auto* x_tensor = context.Input("X"); + auto* out_tensor = context.Output("Out"); + float step = context.Attr("step"); + + out_tensor->mutable_data(context.GetPlace()); + auto& dev = + *context.template device_context().eigen_device(); + framework::EigenScalar::From(*out_tensor).device(dev) = + framework::EigenScalar::From(*x_tensor) + static_cast(step); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/iou_similarity_op.cc b/paddle/fluid/operators/iou_similarity_op.cc index ffbd7c7814c3fdec9fef0580ccd1ea3661ac0012..4b78ec510d1fb73592ee8af9a641622f4d713f8d 100755 --- a/paddle/fluid/operators/iou_similarity_op.cc +++ b/paddle/fluid/operators/iou_similarity_op.cc @@ -87,8 +87,9 @@ $$ } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(iou_similarity, ops::IOUSimilarityOp, - ops::IOUSimilarityOpMaker); +REGISTER_OPERATOR(iou_similarity, ops::IOUSimilarityOp, + ops::IOUSimilarityOpMaker, + paddle::framework::EmptyGradOpMaker); REGISTER_OP_CPU_KERNEL( iou_similarity, diff --git a/paddle/fluid/operators/layer_norm_op.h b/paddle/fluid/operators/layer_norm_op.h index 605b5c258ca57b1a63c9b741a1a30dcb9fca2248..7b84ba0a7daf10e9e636f62eea6bd759ebec9541 100644 --- a/paddle/fluid/operators/layer_norm_op.h +++ b/paddle/fluid/operators/layer_norm_op.h @@ -22,6 +22,103 @@ limitations under the License. */ namespace paddle { namespace operators { +// Wrap RowwiseMean and ColwiseMean. +// Reuse the cpu codes and replace the gpu codes with cublas_gemv, which is +// significantly faster. Unlike the RowwiseMean and ColwiseMean, the +// implementation only considers 2D. +template +struct RowwiseMean2D { + RowwiseMean2D(int left, int right, const platform::DeviceContext& dev_ctx); + + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor* vec); +}; + +#ifdef PADDLE_WITH_CUDA +template +class RowwiseMean2D { + public: + RowwiseMean2D(int left, int right, const platform::DeviceContext& dev_ctx) + : left_(left), right_(right) { + framework::DDim ones_dim({right_}); + divisor_.mutable_data(ones_dim, dev_ctx.GetPlace()); + math::set_constant(dev_ctx, &divisor_, 1.0 / right); + } + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor& input, framework::Tensor* out) { + math::gemv( + context, false, left_, right_, 1., input.data(), divisor_.data(), + 0., out->data()); + } + + private: + int left_; + int right_; + framework::Tensor divisor_; +}; +#endif + +template +class RowwiseMean2D { + public: + RowwiseMean2D(int left, int right, const platform::DeviceContext& dev_ctx) {} + + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, framework::Tensor* out) { + row_mean_(context, input, out); + } + + private: + math::RowwiseMean row_mean_; +}; + +template +struct ColwiseSum2D { + ColwiseSum2D(int left, int right, const platform::DeviceContext& dev_ctx); + + void operator()(const platform::DeviceContext& context, + const framework::Tensor& input, framework::Tensor* vec); +}; + +#ifdef PADDLE_WITH_CUDA +template +class ColwiseSum2D { + public: + ColwiseSum2D(int left, int right, const platform::DeviceContext& dev_ctx) + : left_(left), right_(right) { + framework::DDim ones_dim({left_}); + divisor_.mutable_data(ones_dim, dev_ctx.GetPlace()); + math::set_constant(dev_ctx, &divisor_, 1.0); + } + + void operator()(const platform::CUDADeviceContext& context, + const framework::Tensor& input, framework::Tensor* out) { + math::gemv( + context, true, left_, right_, 1., input.data(), divisor_.data(), + 0., out->data()); + } + + private: + int left_; + int right_; + framework::Tensor divisor_; +}; +#endif + +template +class ColwiseSum2D { + public: + ColwiseSum2D(int left, int right, const platform::DeviceContext& dev_ctx) {} + + void operator()(const platform::CPUDeviceContext& context, + const framework::Tensor& input, framework::Tensor* out) { + col_wise_(context, input, out); + } + + private: + math::ColwiseSum col_wise_; +}; + template struct SubAndSquareFunctor { inline HOSTDEVICE T operator()(T a, T b) const { return (a - b) * (a - b); } @@ -67,15 +164,15 @@ using DataLayout = framework::DataLayout; template class LayerNormKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext &ctx) const override { + void Compute(const framework::ExecutionContext& ctx) const override { const float epsilon = ctx.Attr("epsilon"); - auto *scale = ctx.Input("Scale"); - auto *bias = ctx.Input("Bias"); + auto* scale = ctx.Input("Scale"); + auto* bias = ctx.Input("Bias"); auto x = *ctx.Input("X"); - auto *y = ctx.Output("Y"); - auto *mean = ctx.Output("Mean"); - auto *var = ctx.Output("Variance"); + auto* y = ctx.Output("Y"); + auto* mean = ctx.Output("Mean"); + auto* var = ctx.Output("Variance"); const auto begin_norm_axis = ctx.Attr("begin_norm_axis"); const auto x_dims = x.dims(); @@ -94,8 +191,8 @@ class LayerNormKernel : public framework::OpKernel { out.ShareDataWith(*y); out.Resize(matrix_shape); - auto &dev_ctx = ctx.template device_context(); - math::RowwiseMean row_mean; + auto& dev_ctx = ctx.template device_context(); + RowwiseMean2D row_mean(left, right, ctx.device_context()); // get mean row_mean(dev_ctx, x, mean); @@ -126,31 +223,32 @@ class LayerNormKernel : public framework::OpKernel { template class LayerNormGradKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext &ctx) const override { + void Compute(const framework::ExecutionContext& ctx) const override { const float epsilon = ctx.Attr("epsilon"); auto x = *ctx.Input("X"); - auto *y = ctx.Input("Y"); - auto *mean = ctx.Input("Mean"); - auto *var = ctx.Input("Variance"); - auto *scale = ctx.Input("Scale"); - auto *bias = ctx.Input("Bias"); + auto* y = ctx.Input("Y"); + auto* mean = ctx.Input("Mean"); + auto* var = ctx.Input("Variance"); + auto* scale = ctx.Input("Scale"); + auto* bias = ctx.Input("Bias"); auto d_y = *ctx.Input(framework::GradVarName("Y")); const auto begin_norm_axis = ctx.Attr("begin_norm_axis"); // init output - auto *d_x = ctx.Output(framework::GradVarName("X")); - auto *d_scale = ctx.Output(framework::GradVarName("Scale")); - auto *d_bias = ctx.Output(framework::GradVarName("Bias")); + auto* d_x = ctx.Output(framework::GradVarName("X")); + auto* d_scale = ctx.Output(framework::GradVarName("Scale")); + auto* d_bias = ctx.Output(framework::GradVarName("Bias")); - const auto &x_dims = x.dims(); + const auto& x_dims = x.dims(); auto matrix_dim = framework::flatten_to_2d(x_dims, begin_norm_axis); int left = static_cast(matrix_dim[0]); int right = static_cast(matrix_dim[1]); framework::DDim matrix_shape({left, right}); d_y.Resize(matrix_shape); - auto &dev_ctx = ctx.template device_context(); - math::ColwiseSum colwise_sum; + auto& dev_ctx = ctx.template device_context(); + ColwiseSum2D colwise_sum(left, right, + ctx.device_context()); Tensor temp; Tensor temp_norm; @@ -190,7 +288,8 @@ class LayerNormGradKernel : public framework::OpKernel { Tensor temp_vec; temp_vec.mutable_data(vec_shape, ctx.GetPlace()); - math::RowwiseMean row_mean; + RowwiseMean2D row_mean(left, right, + ctx.device_context()); if (d_scale) { // dy_dx diff --git a/paddle/fluid/operators/listen_and_serv_op.cc b/paddle/fluid/operators/listen_and_serv_op.cc index a594de67e05acd28ffedc5407beecfaea1281444..b19add24e2bd325896a96be53d3d9762abfe217c 100644 --- a/paddle/fluid/operators/listen_and_serv_op.cc +++ b/paddle/fluid/operators/listen_and_serv_op.cc @@ -13,22 +13,13 @@ See the License for the specific language governing permissions and limitations under the License. */ #include -#include #include -#include - -#include #include "paddle/fluid/framework/executor.h" -#include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/framework/proto_desc.h" #include "paddle/fluid/framework/threadpool.h" #include "paddle/fluid/operators/detail/grpc_server.h" -#include "paddle/fluid/operators/detail/sendrecvop_utils.h" -#include "paddle/fluid/operators/detail/simple_block_queue.h" -#include "paddle/fluid/string/printf.h" namespace paddle { namespace operators { @@ -54,6 +45,24 @@ static void CreateTensorFromMessageType(framework::Variable *var, } } +static void ParallelExecuteBlocks(const std::vector ¶llel_blkids, + framework::Executor *executor, + framework::ProgramDesc *program, + framework::Scope *scope) { + std::vector> fs; + for (size_t idx : parallel_blkids) { + fs.push_back(framework::Async([&executor, &program, &scope, idx]() { + int run_block = idx; // thread local + try { + executor->Run(*program, scope, run_block, false, false); + } catch (std::exception &e) { + LOG(ERROR) << "run sub program error " << e.what(); + } + })); + } + for (size_t i = 0; i < fs.size(); ++i) fs[i].wait(); +} + class ListenAndServOp : public framework::OperatorBase { public: ListenAndServOp(const std::string &type, @@ -69,10 +78,7 @@ class ListenAndServOp : public framework::OperatorBase { } void Stop() override { - detail::MessageWithName term_msg; - term_msg.first = LISTEN_TERMINATE_MESSAGE; - rpc_service_->Push(term_msg); - rpc_service_->ShutDown(); + rpc_service_->Push(LISTEN_TERMINATE_MESSAGE); server_thread_->join(); } @@ -96,6 +102,11 @@ class ListenAndServOp : public framework::OperatorBase { framework::Executor executor(dev_place); + // TODO(qiao) set proper fields for table lookup and update + rpc_service_->SetExecutor(&executor); + rpc_service_->SetPrefetchBlkdId(0); + rpc_service_->SetProgram(program); + // TODO(typhoonzero): change this to a while_op for every cluster-batch. bool exit_flag = false; // Record received sparse variables, so that @@ -108,7 +119,7 @@ class ListenAndServOp : public framework::OperatorBase { size_t recv_var_cnt = 0; int batch_barrier = 0; while (batch_barrier != fan_in) { - const detail::MessageWithName &v = rpc_service_->Get(); + const detail::ReceivedMessage v = rpc_service_->Get(); auto recv_var_name = v.first; if (recv_var_name == LISTEN_TERMINATE_MESSAGE) { LOG(INFO) << "received terminate message and exit"; @@ -121,12 +132,11 @@ class ListenAndServOp : public framework::OperatorBase { } else { VLOG(3) << "received grad: " << recv_var_name; recv_var_cnt++; - auto *var = recv_scope.FindVar(recv_var_name); + auto var = v.second->GetVar(); if (var == nullptr) { LOG(ERROR) << "Can not find server side var: " << recv_var_name; PADDLE_THROW("Can not find server side var"); } - detail::DeserializeFromMessage(v.second, dev_ctx, var); if (var->IsType()) { sparse_vars.push_back(var); } @@ -138,34 +148,29 @@ class ListenAndServOp : public framework::OperatorBase { break; } - // put optimize blocks in the thread pool to start run, the last block - // should be global ops. // NOTE: if is_gpu_place, CUDA kernels are laugched by multiple threads // and this will still work. - std::vector> fs; - // block0 contains only listen_and_serv op, start run from block1. - for (int blkid = 1; blkid < num_blocks - 1; ++blkid) { - fs.push_back(framework::Async([&executor, &program, &recv_scope, - blkid]() { - int run_block = blkid; // thread local - try { - executor.Run(*program, &recv_scope, run_block, - false /*create_local_scope*/, false /*create_vars*/); - } catch (std::exception &e) { - LOG(ERROR) << "run sub program error " << e.what(); - } - })); - } - for (int i = 0; i < num_blocks - 2; ++i) fs[i].wait(); - // Run global block at final step, or block1 if there are only 2 blocks - if (num_blocks >= 2) { - try { - executor.Run(*program, &recv_scope, num_blocks - 1, - false /*create_local_scope*/, false /*create_vars*/); - } catch (std::exception &e) { - LOG(ERROR) << "run sub program error " << e.what(); + + // The optimize blocks which have the same parent ID would run parallel + // TODO(Yancey1989): need to use ParallelExecutor for future + size_t last_parent_blkid = program->Block(1).Parent(); + std::vector parallel_blkids; + parallel_blkids.push_back(1); + double ts = detail::GetTimestamp(); + for (size_t blkid = 2; blkid < num_blocks; ++blkid) { + if (program->Block(blkid).Parent() != last_parent_blkid) { + for (size_t idx : parallel_blkids) VLOG(3) << idx; + ParallelExecuteBlocks(parallel_blkids, &executor, program, + &recv_scope); + parallel_blkids.clear(); + last_parent_blkid = program->Block(blkid).Parent(); } + parallel_blkids.push_back(blkid); } + ParallelExecuteBlocks(parallel_blkids, &executor, program, &recv_scope); + + VLOG(3) << "run all blocks spent " << detail::GetTimestamp() - ts + << "(ms)"; // Reset the received sparse variables, the sum operator would not // sum the input sparse variables which rows is empty at the next diff --git a/paddle/fluid/operators/lrn_mkldnn_op.cc b/paddle/fluid/operators/lrn_mkldnn_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..0a18882e8199c2a375a230a693b8b01d12aabfa0 --- /dev/null +++ b/paddle/fluid/operators/lrn_mkldnn_op.cc @@ -0,0 +1,212 @@ +/* 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 "paddle/fluid/framework/tensor.h" +#include "paddle/fluid/operators/lrn_op.h" +#include "paddle/fluid/platform/mkldnn_helper.h" + +namespace paddle { +namespace operators { + +using paddle::framework::Tensor; +using paddle::platform::MKLDNNDeviceContext; + +namespace { +template +std::shared_ptr insert_to_context(const std::string& key, + const MKLDNNDeviceContext& dev_ctx, + Args&&... args) { + auto p = std::static_pointer_cast(dev_ctx.GetBlob(key)); + + if (!p) { + p = std::make_shared(args...); + dev_ctx.SetBlob(key, std::static_pointer_cast(p)); + } + + return p; +} + +template +void run_primitive(Args&&... args) { + auto forward_op = mkldnn::lrn_forward{args...}; + + std::vector pipeline = {forward_op}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); +} +} // namespace + +template +class LRNMKLDNNOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(std::is_same::value, + "MKLDNN LRN must use float data."); + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "MKLDNN LRN must use CPUPlace."); + + auto& dev_ctx = ctx.template device_context(); + const auto& mkldnn_engine = dev_ctx.GetEngine(); + + auto x = ctx.Input("X"); + auto out = ctx.Output("Out"); + auto mid = ctx.Output("MidOut"); + + auto input_data = x->data(); + auto output_data = out->mutable_data(ctx.GetPlace()); + mid->mutable_data(ctx.GetPlace()); + + const int n = ctx.Attr("n"); + const float alpha = ctx.Attr("alpha"); + const float beta = ctx.Attr("beta"); + const float k = ctx.Attr("k"); + const bool is_test = ctx.Attr("is_test"); + + auto e_mid = framework::EigenTensor::From(*mid); + e_mid = e_mid.constant(k); + + auto dims = paddle::framework::vectorize2int(x->dims()); + + auto src_md = paddle::platform::MKLDNNMemDesc( + dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + + auto dst_md = paddle::platform::MKLDNNMemDesc( + dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + + auto forward_desc = mkldnn::lrn_forward::desc{mkldnn::prop_kind::forward, + mkldnn::lrn_across_channels, + src_md, + n, + alpha, + beta, + k}; + + auto src_memory_pd = mkldnn::memory::primitive_desc{src_md, mkldnn_engine}; + auto dst_memory = mkldnn::memory{{dst_md, mkldnn_engine}, + static_cast(output_data)}; + + if (!is_test) { + const std::string key = ctx.op().Output("Out"); + const std::string key_src_memory = key + "@lrn_src_memory"; + const std::string key_pd = key + "@lrn_pd"; + const std::string key_workspace_memory = key + "@lrn_workspace_memory"; + + auto forward_pd = insert_to_context( + key_pd, dev_ctx, forward_desc, mkldnn_engine); + + auto src_memory = insert_to_context( + key_src_memory, dev_ctx, src_memory_pd); + + src_memory->set_data_handle( + static_cast(const_cast(input_data))); + + auto workspace_memory = insert_to_context( + key_workspace_memory, dev_ctx, + forward_pd->workspace_primitive_desc()); + + run_primitive(*forward_pd, *src_memory, *workspace_memory, dst_memory); + } else { + auto forward_pd = + mkldnn::lrn_forward::primitive_desc{forward_desc, mkldnn_engine}; + auto src_memory = mkldnn::memory{ + src_memory_pd, static_cast(const_cast(input_data))}; + auto workspace_memory = + mkldnn::memory{forward_pd.workspace_primitive_desc()}; + + run_primitive(forward_pd, src_memory, workspace_memory, dst_memory); + } + } +}; + +template +class LRNMKLDNNGradOpKernel : public paddle::framework::OpKernel { + public: + void Compute(const paddle::framework::ExecutionContext& ctx) const override { + PADDLE_ENFORCE(std::is_same::value, + "MKLDNN LRN must use float data."); + PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()), + "MKLDNN LRN must use CPUPlace."); + PADDLE_ENFORCE( + !ctx.Attr("is_test"), + "is_test attribute should be set to False in training phase."); + + auto x = ctx.Input("X"); + + auto out_grad = ctx.Input(framework::GradVarName("Out")); + auto x_grad = ctx.Output(framework::GradVarName("X")); + + const std::string key = ctx.op().Input("Out"); + const std::string key_src_memory = key + "@lrn_src_memory"; + const std::string key_pd = key + "@lrn_pd"; + const std::string key_workspace_memory = key + "@lrn_workspace_memory"; + + const int n = ctx.Attr("n"); + const float alpha = ctx.Attr("alpha"); + const float beta = ctx.Attr("beta"); + const float k = ctx.Attr("k"); + + auto& dev_ctx = ctx.template device_context(); + const auto& mkldnn_engine = dev_ctx.GetEngine(); + + auto x_grad_data = x_grad->mutable_data(ctx.GetPlace()); + auto out_grad_data = out_grad->data(); + + auto dims = paddle::framework::vectorize2int(x->dims()); + + auto src_md = paddle::platform::MKLDNNMemDesc( + dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + + auto diff_src_md = paddle::platform::MKLDNNMemDesc( + dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + + auto diff_dst_md = paddle::platform::MKLDNNMemDesc( + dims, mkldnn::memory::data_type::f32, mkldnn::memory::format::nchw); + + auto diff_dst_memory = + mkldnn::memory{{diff_dst_md, mkldnn_engine}, + static_cast(const_cast(out_grad_data))}; + + auto diff_src_memory = mkldnn::memory{{diff_src_md, mkldnn_engine}, + static_cast(x_grad_data)}; + + auto backward_desc = mkldnn::lrn_backward::desc{ + mkldnn::lrn_across_channels, src_md, diff_src_md, n, alpha, beta, k}; + + auto forward_pd = dev_ctx.GetBlob(key_pd); + + auto backward_pd = mkldnn::lrn_backward::primitive_desc{ + backward_desc, mkldnn_engine, + *static_cast(forward_pd.get())}; + + std::shared_ptr workspace_memory = + dev_ctx.GetBlob(key_workspace_memory); + + auto src_memory = dev_ctx.GetBlob(key_src_memory); + auto backward_op = mkldnn::lrn_backward{ + backward_pd, *static_cast(src_memory.get()), + diff_dst_memory, *static_cast(workspace_memory.get()), + diff_src_memory}; + + std::vector pipeline = {backward_op}; + mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait(); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OP_KERNEL(lrn, MKLDNN, paddle::platform::CPUPlace, + ops::LRNMKLDNNOpKernel); +REGISTER_OP_KERNEL(lrn_grad, MKLDNN, paddle::platform::CPUPlace, + ops::LRNMKLDNNGradOpKernel); diff --git a/paddle/fluid/operators/lrn_op.cc b/paddle/fluid/operators/lrn_op.cc index 692e85dcffa583abcb22a1629953badc67489efa..cb1568398125bbb57da974096da527200c1e0975 100644 --- a/paddle/fluid/operators/lrn_op.cc +++ b/paddle/fluid/operators/lrn_op.cc @@ -13,6 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/lrn_op.h" +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif namespace paddle { namespace operators { @@ -116,6 +119,26 @@ struct LRNGradFunctor { template struct LRNGradFunctor; template struct LRNGradFunctor; +namespace { +framework::OpKernelType GetExpectedLRNKernel( + const framework::ExecutionContext& ctx) { + framework::LibraryType library_{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_MKLDNN + if (library_ == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library_ = framework::LibraryType::kMKLDNN; + } +#endif + + std::string data_format = ctx.Attr("data_format"); + // TODO(pzelazko-intel): enable MKLDNN layout when it's ready + framework::DataLayout layout_ = framework::StringToDataLayout(data_format); + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), ctx.GetPlace(), + layout_, library_); +} +} // namespace + class LRNOp : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; @@ -132,8 +155,13 @@ class LRNOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_EQ(x_dim.size(), 4, "Input(X)'rank of LRNOp should be 4."); ctx->SetOutputDim("Out", x_dim); - ctx->SetOutputDim("MidOut", x_dim); ctx->ShareLoD("X", /*->*/ "Out"); + ctx->SetOutputDim("MidOut", x_dim); + } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return GetExpectedLRNKernel(ctx); } }; @@ -176,6 +204,20 @@ class LRNOpMaker : public framework::OpProtoAndCheckerMaker { "beta is the power number.") .SetDefault(0.75) .GreaterThan(0.0); + AddAttr("use_mkldnn", + "(bool, default false) Only used in mkldnn kernel") + .SetDefault(false); + AddAttr( + "data_format", + "(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"); + AddAttr("is_test", + "Turns on memory optimization that optimizes away " + "unnecessary memory allocations. Used by MKLDNN.") + .SetDefault(false); AddComment(R"DOC( Local Response Normalization Operator. @@ -223,8 +265,12 @@ class LRNOpGrad : public framework::OperatorWithKernel { auto x_dims = ctx->GetInputDim("X"); ctx->SetOutputDim(framework::GradVarName("X"), x_dims); } -}; + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return GetExpectedLRNKernel(ctx); + } +}; } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/lrn_op.h b/paddle/fluid/operators/lrn_op.h index 95796f7eecd2bcd61aab7944f557ca568b03e027..0fd3175e8579df9e61368cc151a94fa45e433884 100644 --- a/paddle/fluid/operators/lrn_op.h +++ b/paddle/fluid/operators/lrn_op.h @@ -121,6 +121,10 @@ class LRNGradKernel : public framework::OpKernel { T alpha = ctx.Attr("alpha"); T beta = ctx.Attr("beta"); + PADDLE_ENFORCE( + !ctx.Attr("is_test"), + "is_test attribute should be set to False in training phase."); + LRNGradFunctor f; f(ctx, x, out, mid, x_g, out_g, N, C, H, W, n, alpha, beta); } diff --git a/paddle/fluid/operators/math/CMakeLists.txt b/paddle/fluid/operators/math/CMakeLists.txt index 547d081006f1c28ba73cb02d38e36bb612cea494..ee0e91132bce52998e9c45b37335618e4354e1cd 100644 --- a/paddle/fluid/operators/math/CMakeLists.txt +++ b/paddle/fluid/operators/math/CMakeLists.txt @@ -6,6 +6,7 @@ function(math_library TARGET) # But it handle split GPU/CPU code and link some common library. set(cc_srcs) set(cu_srcs) + set(hip_srcs) set(math_common_deps device_context framework_proto) set(multiValueArgs DEPS) cmake_parse_arguments(math_library "${options}" "${oneValueArgs}" @@ -17,10 +18,15 @@ function(math_library TARGET) if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.cu) list(APPEND cu_srcs ${TARGET}.cu) endif() + if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${TARGET}.hip.cu) + list(APPEND hip_srcs ${TARGET}.hip.cu) + endif() list(LENGTH cc_srcs cc_srcs_len) if (WITH_GPU) nv_library(${TARGET} SRCS ${cc_srcs} ${cu_srcs} DEPS ${math_library_DEPS} ${math_common_deps}) + elseif (WITH_AMD_GPU) + hip_library(${TARGET} SRCS ${cc_srcs} ${hip_srcs} DEPS ${math_library_DEPS} ${math_common_deps}) elseif(${cc_srcs_len} GREATER 0) cc_library(${TARGET} SRCS ${cc_srcs} DEPS ${math_library_DEPS} ${math_common_deps}) endif() diff --git a/paddle/fluid/operators/math/concat.cc b/paddle/fluid/operators/math/concat.cc index b672c79afd97e36894af647fd4bc6edfb885ff13..bfce56f9fdcafa0800c9742b9fae41fd6a572b40 100644 --- a/paddle/fluid/operators/math/concat.cc +++ b/paddle/fluid/operators/math/concat.cc @@ -20,7 +20,7 @@ namespace math { /* * All tensors' dimension should be the same and the values of - * each dimension are the same, except the axis dimension. + * each dimension must be the same, except the axis dimension. */ template class ConcatFunctor { @@ -63,7 +63,7 @@ class ConcatFunctor { /* * All tensors' dimension should be the same and the values of - * each dimension are the same, except the axis dimension. + * each dimension must be the same, except the axis dimension. */ template class ConcatGradFunctor { diff --git a/paddle/fluid/operators/math/concat.cu b/paddle/fluid/operators/math/concat.cu index 60b266f08fb2d4217c5933902d69de96fc2abe22..c0786757b34195d47c3b1cadc938f0e9fcfd6038 100644 --- a/paddle/fluid/operators/math/concat.cu +++ b/paddle/fluid/operators/math/concat.cu @@ -66,68 +66,66 @@ __global__ void KernelConcat(T** inputs, const int* input_cols, int col_size, } template -__global__ void KernelConcat(T** inputs, const int input_col, - const int output_rows, const int output_cols, - T* output) { +__global__ void KernelConcat(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; - double inv_input_col = 1.0 / input_col; - for (; tid_x < output_cols; tid_x += blockDim.x * gridDim.x) { - int split = tid_x * inv_input_col; - int in_offset = tid_x - split * input_col; - T* input_ptr = inputs[split]; + for (; tid_x < out_cols; tid_x += blockDim.x * gridDim.x) { + int split = tid_x * 1.0 / fixed_in_col; + int in_offset = tid_x - split * fixed_in_col; + T* input_ptr = inputs_data[split]; int tid_y = blockIdx.y * blockDim.y + threadIdx.y; - for (; tid_y < output_rows; tid_y += blockDim.y * gridDim.y) { - output[tid_y * output_cols + tid_x] = - input_ptr[tid_y * input_col + in_offset]; + for (; tid_y < out_rows; tid_y += blockDim.y * gridDim.y) { + output_data[tid_y * out_cols + tid_x] = + input_ptr[tid_y * fixed_in_col + in_offset]; } } } template -__global__ void KernelConcatGrad(const T* input, const int input_row, - const int input_col, const int* output_cols, - int col_size, T** outputs) { +__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) { int tid_x = blockIdx.x * blockDim.x + threadIdx.x; - int segment = upper_bound(output_cols, col_size, tid_x) - 1; - int curr_offset = output_cols[segment]; + int segment = upper_bound(out_cols, out_cols_size, tid_x) - 1; + int curr_offset = out_cols[segment]; int curr_segment = segment; - for (; tid_x < input_col; tid_x += blockDim.x * gridDim.x) { + for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) { T curr_col_offset; - while ((curr_col_offset = output_cols[curr_segment + 1]) <= tid_x) { + while ((curr_col_offset = out_cols[curr_segment + 1]) <= tid_x) { curr_offset = curr_col_offset; ++curr_segment; } int local_col = tid_x - curr_offset; int segment_width = curr_col_offset - curr_offset; - T* output_ptr = outputs[curr_segment]; + T* output_ptr = outputs_data[curr_segment]; int tid_y = blockIdx.y * blockDim.y + threadIdx.y; - for (; tid_y < input_row; tid_y += blockDim.y * gridDim.y) + for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) output_ptr[tid_y * segment_width + local_col] = - input[tid_y * input_col + tid_x]; + input_data[tid_y * in_col + tid_x]; } } template -__global__ void KernelConcatGrad(const T* input, const int input_row, - const int input_col, const int output_cols, - T** outputs) { +__global__ void KernelConcatGrad(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; - double inv_input_col = 1.0 / input_col; - for (; tid_x < input_col; tid_x += blockDim.x * gridDim.x) { - int split = tid_x * inv_input_col; - int in_offset = tid_x - split * input_col; - T* output_ptr = outputs[split]; + for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) { + int split = tid_x / fixed_out_col; + int in_offset = tid_x - split * fixed_out_col; + T* output_ptr = outputs_data[split]; int tid_y = blockIdx.y * blockDim.y + threadIdx.y; - for (; tid_y < input_row; tid_y += blockDim.y * gridDim.y) - output_ptr[tid_y * output_cols + in_offset] = - input[tid_y * input_col + tid_x]; + for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y) + output_ptr[tid_y * fixed_out_col + in_offset] = + input_data[tid_y * in_col + tid_x]; } } /* * All tensors' dimension should be the same and the values of - * each dimension are the same, except the axis dimension. + * each dimension must be the same, except the axis dimension. */ template class ConcatFunctor { @@ -136,41 +134,40 @@ class ConcatFunctor { const std::vector& input, const int axis, framework::Tensor* output) { // TODO(zcd): Add input data validity checking - int num = input.size(); - int rows = 1; + int in_num = input.size(); + int in_row = 1; auto dim_0 = input[0].dims(); for (int i = 0; i < axis; ++i) { - rows *= dim_0[i]; + in_row *= dim_0[i]; } - int cols = input[0].numel() / rows; - int out_rows = rows, out_cols = 0; + int in_col = input[0].numel() / in_row; + int out_row = in_row, out_col = 0; - framework::Vector inputs_data(num * sizeof(T*) / 2); - framework::Vector inputs_cols(num + 1); - inputs_cols[0] = 0; + framework::Vector inputs_data(in_num * sizeof(T*) / 2); + framework::Vector inputs_col(in_num + 1); T** inputs_ptr = reinterpret_cast(inputs_data.data()); + inputs_col[0] = 0; bool sameShape = true; - for (int i = 0; i < num; ++i) { - int t_cols = input[i].numel() / rows; + for (int i = 0; i < in_num; ++i) { + int t_cols = input[i].numel() / in_row; if (sameShape) { - if (t_cols != cols) sameShape = false; + if (t_cols != in_col) sameShape = false; } - out_cols += t_cols; - inputs_cols[i + 1] = out_cols; + out_col += t_cols; + inputs_col[i + 1] = out_col; inputs_ptr[i] = const_cast(input[i].data()); } - T** ins_gpu = + T** dev_ins_data = reinterpret_cast(inputs_data.CUDAMutableData(context.GetPlace())); - const int* ins_col_gpu = inputs_cols.CUDAData(context.GetPlace()); // computation // set the thread block and grid according to CurrentDeviceId const int kThreadsPerBlock = 1024; int block_cols = kThreadsPerBlock; - if (out_cols < kThreadsPerBlock) { // block_cols is aligned by 32. - block_cols = ((out_cols + 31) >> 5) << 5; + if (out_col < kThreadsPerBlock) { // block_cols is aligned by 32. + block_cols = ((out_col + 31) >> 5) << 5; } int block_rows = kThreadsPerBlock / block_cols; dim3 block_size = dim3(block_cols, block_rows, 1); @@ -179,25 +176,26 @@ class ConcatFunctor { int max_blocks = std::max(max_threads / kThreadsPerBlock, 1); int grid_cols = - std::min((out_cols + block_cols - 1) / block_cols, max_blocks); + std::min((out_col + block_cols - 1) / block_cols, max_blocks); int grid_rows = - std::min(max_blocks / grid_cols, std::max(out_rows / block_rows, 1)); + std::min(max_blocks / grid_cols, std::max(out_row / block_rows, 1)); dim3 grid_size = dim3(grid_cols, grid_rows, 1); if (sameShape) { KernelConcat<<>>( - ins_gpu, cols, out_rows, out_cols, output->data()); + dev_ins_data, in_col, out_row, out_col, output->data()); } else { + const int* dev_ins_col_data = inputs_col.CUDAData(context.GetPlace()); KernelConcat<<>>( - ins_gpu, ins_col_gpu, static_cast(inputs_cols.size()), out_rows, - out_cols, output->data()); + dev_ins_data, dev_ins_col_data, static_cast(inputs_col.size()), + out_row, out_col, output->data()); } } }; /* * All tensors' dimension should be the same and the values of - * each dimension are the same, except the axis dimension. + * each dimension must be the same, except the axis dimension. */ template class ConcatGradFunctor { @@ -206,41 +204,40 @@ class ConcatGradFunctor { const framework::Tensor& input, const int axis, std::vector& outputs) { // TODO(zcd): Add input data validity checking - int num = outputs.size(); - int input_row = 1; + int o_num = outputs.size(); + int out_row = 1; auto dim_0 = outputs[0].dims(); for (int i = 0; i < axis; ++i) { - input_row *= dim_0[i]; + out_row *= dim_0[i]; } - int output_col_0 = outputs[0].numel() / input_row; - int input_col = 0; + int out_col = outputs[0].numel() / out_row; + int in_col = 0, in_row = out_row; bool sameShape = true; - framework::Vector outputs_data(num * sizeof(T*) / 2); - framework::Vector outputs_cols(num + 1); - outputs_cols[0] = 0; + framework::Vector outputs_data(o_num * sizeof(T*) / 2); + framework::Vector outputs_cols(o_num + 1); T** outputs_ptr = reinterpret_cast(outputs_data.data()); - for (int i = 0; i < num; ++i) { - int t_col = outputs[i].numel() / input_row; + outputs_cols[0] = 0; + for (int i = 0; i < o_num; ++i) { + int t_col = outputs[i].numel() / out_row; if (sameShape) { - if (t_col != output_col_0) sameShape = false; + if (t_col != out_col) sameShape = false; } - input_col += t_col; - outputs_cols[i + 1] = input_col; + in_col += t_col; + outputs_cols[i + 1] = in_col; outputs_ptr[i] = outputs[i].data(); } - T** outs_gpu = + T** dev_out_gpu_data = reinterpret_cast(outputs_data.CUDAMutableData(context.GetPlace())); - const int* outs_col_gpu = outputs_cols.CUDAData(context.GetPlace()); // computation const int kThreadsPerBlock = 1024; int block_cols = kThreadsPerBlock; - if (input_col < kThreadsPerBlock) { // block_cols is aligned by 32. - block_cols = ((input_col + 31) >> 5) << 5; + if (in_col < kThreadsPerBlock) { // block_cols is aligned by 32. + block_cols = ((in_col + 31) >> 5) << 5; } int block_rows = kThreadsPerBlock / block_cols; dim3 block_size = dim3(block_cols, block_rows, 1); @@ -249,18 +246,19 @@ class ConcatGradFunctor { int max_blocks = std::max(max_threads / kThreadsPerBlock, 1); int grid_cols = - std::min((input_col + block_cols - 1) / block_cols, max_blocks); + std::min((in_col + block_cols - 1) / block_cols, max_blocks); int grid_rows = - std::min(max_blocks / grid_cols, std::max(input_row / block_rows, 1)); + std::min(max_blocks / grid_cols, std::max(out_row / block_rows, 1)); dim3 grid_size = dim3(grid_cols, grid_rows, 1); if (sameShape) { KernelConcatGrad<<>>( - input.data(), input_row, input_col, output_col_0, outs_gpu); + input.data(), in_row, in_col, out_col, dev_out_gpu_data); } else { + const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace()); KernelConcatGrad<<>>( - input.data(), input_row, input_col, outs_col_gpu, - static_cast(outputs_cols.size()), outs_gpu); + input.data(), in_row, in_col, dev_outs_col_data, + static_cast(outputs_cols.size()), dev_out_gpu_data); } } }; diff --git a/paddle/fluid/operators/math/concat.h b/paddle/fluid/operators/math/concat.h index 22147d79e4b1eeee76f7445dd963bf5062049a34..c0e983e4aa7abbdd87649f5a3147d2a464993bce 100644 --- a/paddle/fluid/operators/math/concat.h +++ b/paddle/fluid/operators/math/concat.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/data_type.h" #include "paddle/fluid/framework/tensor.h" namespace paddle { diff --git a/paddle/fluid/operators/math/concat.hip.cu b/paddle/fluid/operators/math/concat.hip.cu new file mode 100644 index 0000000000000000000000000000000000000000..eacef0438883891671fec6e4001f862f619723cb --- /dev/null +++ b/paddle/fluid/operators/math/concat.hip.cu @@ -0,0 +1,15 @@ +/* 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 diff --git a/paddle/fluid/operators/math/math_function.h b/paddle/fluid/operators/math/math_function.h index 47e2386d0578265330088eeac6c57fe2518f951a..cdbc7bfb37e83c6c2b696ba010277c9eec49f2a8 100644 --- a/paddle/fluid/operators/math/math_function.h +++ b/paddle/fluid/operators/math/math_function.h @@ -19,13 +19,6 @@ limitations under the License. */ #include #endif -#ifdef PADDLE_USE_ATLAS -extern "C" { -#include -#include -} -#endif - #ifdef PADDLE_USE_OPENBLAS #include #include diff --git a/paddle/fluid/operators/math/sequence_pooling.cc b/paddle/fluid/operators/math/sequence_pooling.cc index f7a6f2bdf4e3b7896df39acfa51fa20577b20f3b..5ae42ab973c81d3794fbbbe088e37ab02168c8dc 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cc +++ b/paddle/fluid/operators/math/sequence_pooling.cc @@ -19,8 +19,17 @@ namespace paddle { namespace operators { namespace math { +using Tensor = framework::Tensor; +using LoDTensor = framework::LoDTensor; +template +using EigenVector = framework::EigenVector; +template +using EigenMatrix = framework::EigenMatrix; + template -class MaxSeqPoolFunctor { +class MaxSeqPoolFunctor { public: void operator()(const platform::CPUDeviceContext& context, const framework::LoDTensor& input, framework::Tensor* output, @@ -60,7 +69,7 @@ class MaxSeqPoolFunctor { }; template -class MaxSeqPoolGradFunctor { +class MaxSeqPoolGradFunctor { public: void operator()(const platform::CPUDeviceContext& context, const framework::Tensor& out_grad, @@ -93,10 +102,101 @@ class MaxSeqPoolGradFunctor { } }; -template class MaxSeqPoolFunctor; -template class MaxSeqPoolFunctor; -template class MaxSeqPoolGradFunctor; -template class MaxSeqPoolGradFunctor; +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* index = nullptr) { + if (pooltype == "MAX") { + math::MaxSeqPoolFunctor max_pool; + max_pool(context, input, output, index); + return; + } + auto lod = input.lod()[0]; + auto& place = *context.eigen_device(); + 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])); + Tensor out_t = output->Slice(i, i + 1); + int64_t h = static_cast(lod[i + 1] - lod[i]); + int64_t w = input.numel() / input.dims()[0]; + auto in_e = EigenMatrix::From(in_t, framework::make_ddim({h, w})); + auto out_e = EigenVector::Flatten(out_t); + 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}})); + } else if (pooltype == "SQRT") { + out_e.device(place) = in_e.sum(Eigen::array({{0}})) / + std::sqrt(static_cast(h)); + } else if (pooltype == "LAST") { + out_e.device(place) = in_e.chip(h - 1, 0); + } else if (pooltype == "FIRST") { + out_e.device(place) = in_e.chip(0, 0); + } else { + PADDLE_THROW("unsupported pooling pooltype"); + } + } + } +}; + +template +class SequencePoolGradFunctor { + public: + void operator()(const platform::CPUDeviceContext& context, + const std::string pooltype, const framework::Tensor& out_grad, + framework::LoDTensor* in_grad, + /* max pool has index */ + const framework::Tensor* index = nullptr) { + if (pooltype == "MAX") { + math::MaxSeqPoolGradFunctor max_pool_grad; + max_pool_grad(context, out_grad, *index, in_grad); + return; + } + + if (pooltype == "LAST" || pooltype == "FIRST") { + // set X@Grad be zero at first when pooltype is LAST/FIRST + math::SetConstant functor; + functor(context, in_grad, 0); + } + auto lod = in_grad->lod()[0]; + auto& place = *context.eigen_device(); + for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { + auto in_g_t = in_grad->Slice(static_cast(lod[i]), + static_cast(lod[i + 1])); + auto out_g_t = out_grad.Slice(i, i + 1); + int64_t h = static_cast(lod[i + 1] - lod[i]); + int64_t w = in_grad->numel() / in_grad->dims()[0]; + auto in_g_e = EigenMatrix::From(in_g_t, {h, w}); + auto out_g_e = EigenMatrix::From(out_g_t, {1, w}); + auto out_g_e_v = EigenVector::Flatten(out_g_t); + Eigen::DSizes bcast(h, 1); + + 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); + } else if (pooltype == "LAST") { + in_g_e.chip(h - 1, 0).device(place) = out_g_e_v; + } else if (pooltype == "FIRST") { + in_g_e.chip(0, 0).device(place) = out_g_e_v; + } else { + PADDLE_THROW("unsupported pooling pooltype"); + } + } + } +}; + +template class SequencePoolFunctor; +template class SequencePoolFunctor; +template class SequencePoolGradFunctor; +template class SequencePoolGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/fluid/operators/math/sequence_pooling.cu b/paddle/fluid/operators/math/sequence_pooling.cu index d61407c020142f046f41f71a56702fd6106df628..1935364da37e9a9881651455d2da4ecef1b1e266 100644 --- a/paddle/fluid/operators/math/sequence_pooling.cu +++ b/paddle/fluid/operators/math/sequence_pooling.cu @@ -14,6 +14,7 @@ limitations under the License. */ #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/sequence_pooling.h" +#include "paddle/fluid/platform/cuda_helper.h" namespace paddle { namespace operators { @@ -22,113 +23,331 @@ namespace math { #define FLT_MAX __FLT_MAX__ template -__global__ void KeMaxSequencePool(const T* input, const size_t* starts, - T* output, int* index, int64_t num_seq, - int64_t dim) { - int dim_idx = threadIdx.x; - int seq_id = blockIdx.x; - if (seq_id >= num_seq) return; - size_t start = starts[seq_id]; - size_t end = starts[seq_id + 1]; - - for (int64_t i = dim_idx; i < dim; i += blockDim.x) { - T max_val = static_cast(-FLT_MAX); - int max_id = -1; - for (size_t step_id = start; step_id < end; step_id++) { - if (max_val < input[step_id * dim + i]) { - max_val = input[step_id * dim + i]; - max_id = step_id; +struct MaxPoolFunctor { + HOSTDEVICE void operator()(const T* input, const size_t start, + const size_t end, const size_t item_dim, T* output, + int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + T max_val = static_cast(-FLT_MAX); + int max_index = -1; + for (int i = start; i < end; ++i) { + if (max_val < input[item_dim * i + tid]) { + max_val = input[item_dim * i + tid]; + max_index = i; + } } + output[tid] = max_val; + index[tid] = max_index; } - output[seq_id * dim + i] = max_val; - index[seq_id * dim + i] = max_id; } -} +}; template -class MaxSeqPoolFunctor { - public: - void operator()(const platform::CUDADeviceContext& context, - const framework::LoDTensor& input, framework::Tensor* output, - framework::Tensor* index) { - auto in_dims = input.dims(); - auto out_dims = output->dims(); - auto idx_dims = index->dims(); - PADDLE_ENFORCE_GT(in_dims.size(), static_cast(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]); +struct AvgPoolFunctor { + HOSTDEVICE void operator()(const T* input, const size_t start, + const size_t end, const size_t item_dim, T* output, + int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + T val = static_cast(0); + for (int i = start; i < end; ++i) { + val += input[item_dim * i + tid]; + } + // end, start is lod, so end - start != 0 + output[tid] = val / static_cast(end - start); } - PADDLE_ENFORCE_EQ(idx_dims, out_dims); + } +}; - auto starts = input.lod()[0]; - const T* in_data = input.data(); - T* out_data = output->data(); - int* max_index = index->data(); +template +struct SumPoolFunctor { + HOSTDEVICE void operator()(const T* input, const size_t start, + const size_t end, const size_t item_dim, T* output, + int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + T val = static_cast(0); + for (int i = start; i < end; ++i) { + val += input[item_dim * i + tid]; + } + output[tid] = val; + } + } +}; - int64_t num_seq = out_dims[0]; - int64_t dim = output->numel() / num_seq; +template +struct SqrtPoolFunctor { + HOSTDEVICE void operator()(const T* input, const size_t start, + const size_t end, const size_t item_dim, T* output, + int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + T val = static_cast(0); + for (int i = start; i < end; ++i) { + val += input[item_dim * i + tid]; + } + // end, start is lod, so end - start != 0 + output[tid] = val / sqrt(end - start); + } + } +}; - dim3 threads(256, 1); - dim3 grid(num_seq, 1); - auto stream = context.stream(); - KeMaxSequencePool<<>>( - in_data, starts.CUDAData(context.GetPlace()), out_data, max_index, - num_seq, dim); +template +struct LastPoolFunctor { + HOSTDEVICE void operator()(const T* input, const size_t start, + const size_t end, const size_t item_dim, T* output, + int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + output[tid] = input[item_dim * (end - 1) + tid]; + } } }; template -__global__ void KeMaxSequencePoolGrad(const T* out_grad, const int* max_index, - T* in_grad, int64_t num_seq, - int64_t dim) { - int idx = threadIdx.x + blockIdx.x * blockDim.x; - int col_idx = idx % dim; - if (idx < num_seq * dim) { - int step_id = max_index[idx]; - in_grad[step_id * dim + col_idx] = out_grad[idx]; +struct FirstPoolFunctor { + HOSTDEVICE void operator()(const T* input, const size_t start, + const size_t end, const size_t item_dim, T* output, + int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + output[tid] = input[item_dim * start + tid]; + } } +}; + +template +__global__ void sequence_pool_kernel(Range_OP op, const T* input, + const size_t* lod, const size_t lod_size, + const size_t item_dim, T* output, + int* index) { + int bid = blockIdx.x; + if (bid >= lod_size - 1) return; + size_t start = lod[bid]; + size_t end = lod[bid + 1]; + int* index_offset = nullptr; + if (index != nullptr) { + index_offset = &index[bid * item_dim]; + } + op(input, start, end, item_dim, &output[bid * item_dim], index_offset); } template -class MaxSeqPoolGradFunctor { +class SequencePoolFunctor { public: void operator()(const platform::CUDADeviceContext& context, - const framework::Tensor& out_grad, - const framework::Tensor& index, - framework::LoDTensor* in_grad) { - auto og_dims = out_grad.dims(); - auto idx_dims = index.dims(); - auto ig_dims = in_grad->dims(); - PADDLE_ENFORCE_GT(og_dims.size(), static_cast(1)); - PADDLE_ENFORCE_GT(ig_dims.size(), static_cast(1)); - for (int64_t i = 1; i < og_dims.size(); ++i) { - PADDLE_ENFORCE_EQ(og_dims[i], ig_dims[i]); + const std::string pooltype, const framework::LoDTensor& input, + framework::Tensor* output, + framework::Tensor* index = nullptr) { + auto lod = input.lod()[0]; + const size_t item_dim = output->numel() / output->dims()[0]; + dim3 threads(1024, 1); + dim3 grid(lod.size(), 1); + if (pooltype == "MAX") { + sequence_pool_kernel< + T, MaxPoolFunctor><<>>( + MaxPoolFunctor(), input.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + output->mutable_data(context.GetPlace()), index->data()); + } else if (pooltype == "AVERAGE") { + sequence_pool_kernel< + T, AvgPoolFunctor><<>>( + AvgPoolFunctor(), input.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + output->mutable_data(context.GetPlace()), nullptr); + } else if (pooltype == "SUM") { + sequence_pool_kernel< + T, SumPoolFunctor><<>>( + SumPoolFunctor(), input.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + output->mutable_data(context.GetPlace()), nullptr); + } else if (pooltype == "SQRT") { + sequence_pool_kernel< + T, SqrtPoolFunctor><<>>( + SqrtPoolFunctor(), input.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + output->mutable_data(context.GetPlace()), nullptr); + } else if (pooltype == "LAST") { + sequence_pool_kernel< + T, LastPoolFunctor><<>>( + LastPoolFunctor(), input.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + output->mutable_data(context.GetPlace()), nullptr); + } else if (pooltype == "FIRST") { + sequence_pool_kernel< + T, FirstPoolFunctor><<>>( + FirstPoolFunctor(), input.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + output->mutable_data(context.GetPlace()), nullptr); + } else { + PADDLE_THROW("unsupported pooling pooltype"); } - PADDLE_ENFORCE_EQ(idx_dims, og_dims); + } +}; - const T* og_data = out_grad.data(); - const int* max_index = index.data(); - T* ig_data = in_grad->data(); +template +struct MaxPoolGradFunctor { + HOSTDEVICE void operator()(const T* out_grad, const size_t start, + const size_t end, const size_t item_dim, + T* in_grad, const int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + for (int i = start; i < end; ++i) { + if (i == index[tid]) { + in_grad[item_dim * i + tid] = out_grad[tid]; + } else { + in_grad[item_dim * i + tid] = static_cast(0); + } + } + } + } +}; - SetConstant set_zero; - set_zero(context, in_grad, static_cast(0.0)); - int64_t num_seq = og_dims[0]; - int64_t dim = out_grad.numel() / num_seq; +template +struct AvgPoolGradFunctor { + HOSTDEVICE void operator()(const T* out_grad, const size_t start, + const size_t end, const size_t item_dim, + T* in_grad, const int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + for (int i = start; i < end; ++i) { + in_grad[item_dim * i + tid] = out_grad[tid] / (end - start); + } + } + } +}; - unsigned int blocks = (num_seq * dim + 128 - 1) / 128; - dim3 threads(128, 1); - dim3 grid(blocks, 1); - auto stream = context.stream(); - KeMaxSequencePoolGrad<<>>( - og_data, max_index, ig_data, num_seq, dim); +template +struct SumPoolGradFunctor { + HOSTDEVICE void operator()(const T* out_grad, const size_t start, + const size_t end, const size_t item_dim, + T* in_grad, const int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + for (int i = start; i < end; ++i) { + in_grad[item_dim * i + tid] = out_grad[tid]; + } + } + } +}; + +template +struct SqrtPoolGradFunctor { + HOSTDEVICE void operator()(const T* out_grad, const size_t start, + const size_t end, const size_t item_dim, + T* in_grad, const int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + for (int i = start; i < end; ++i) { + in_grad[item_dim * i + tid] = + out_grad[tid] / (sqrt(static_cast(end - start))); + } + } + } +}; + +template +struct LastPoolGradFunctor { + HOSTDEVICE void operator()(const T* out_grad, const size_t start, + const size_t end, const size_t item_dim, + T* in_grad, const int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + for (int i = start; i < end; ++i) { + if (i == end - 1) { + in_grad[item_dim * i + tid] = out_grad[tid]; + } else { + in_grad[item_dim * i + tid] = static_cast(0); + } + } + } + } +}; + +template +struct FirstPoolGradFunctor { + HOSTDEVICE void operator()(const T* out_grad, const size_t start, + const size_t end, const size_t item_dim, + T* in_grad, const int* index) { + for (int tid = threadIdx.x; tid < item_dim; tid += blockDim.x) { + for (int i = start; i < end; ++i) { + if (i == start) { + in_grad[item_dim * i + tid] = out_grad[tid]; + } else { + in_grad[item_dim * i + tid] = static_cast(0); + } + } + } + } +}; + +template +__global__ void sequence_pool_grad_kernel(Range_OP op, const T* out_grad, + const size_t* lod, + const size_t lod_size, + const size_t item_dim, T* in_grad, + const int* index) { + int bid = blockIdx.x; + if (bid >= lod_size - 1) return; + size_t start = lod[bid]; + size_t end = lod[bid + 1]; + const int* index_offset = nullptr; + if (index != nullptr) { + index_offset = &index[bid * item_dim]; + } + op(&out_grad[bid * item_dim], start, end, item_dim, in_grad, index_offset); +} + +template +class SequencePoolGradFunctor { + public: + void operator()(const platform::CUDADeviceContext& context, + const std::string pooltype, const framework::Tensor& out_grad, + framework::LoDTensor* in_grad, + /* max pool has index */ + const framework::Tensor* index = nullptr) { + auto lod = in_grad->lod()[0]; + const size_t item_dim = in_grad->numel() / in_grad->dims()[0]; + dim3 threads(1024, 1); + dim3 grid(lod.size(), 1); + if (pooltype == "MAX") { + sequence_pool_grad_kernel< + T, MaxPoolGradFunctor><<>>( + MaxPoolGradFunctor(), out_grad.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + in_grad->mutable_data(context.GetPlace()), index->data()); + } else if (pooltype == "AVERAGE") { + sequence_pool_grad_kernel< + T, AvgPoolGradFunctor><<>>( + AvgPoolGradFunctor(), out_grad.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + in_grad->mutable_data(context.GetPlace()), nullptr); + } else if (pooltype == "SUM") { + sequence_pool_grad_kernel< + T, SumPoolGradFunctor><<>>( + SumPoolGradFunctor(), out_grad.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + in_grad->mutable_data(context.GetPlace()), nullptr); + } else if (pooltype == "SQRT") { + sequence_pool_grad_kernel< + T, SqrtPoolGradFunctor><<>>( + SqrtPoolGradFunctor(), out_grad.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + in_grad->mutable_data(context.GetPlace()), nullptr); + } else if (pooltype == "LAST") { + sequence_pool_grad_kernel< + T, LastPoolGradFunctor><<>>( + LastPoolGradFunctor(), out_grad.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + in_grad->mutable_data(context.GetPlace()), nullptr); + } else if (pooltype == "FIRST") { + sequence_pool_grad_kernel< + T, FirstPoolGradFunctor><<>>( + FirstPoolGradFunctor(), out_grad.data(), + lod.CUDAData(context.GetPlace()), lod.size(), item_dim, + in_grad->mutable_data(context.GetPlace()), nullptr); + + } else { + PADDLE_THROW("unsupported pooling pooltype"); + } } }; -template class MaxSeqPoolFunctor; -template class MaxSeqPoolFunctor; -template class MaxSeqPoolGradFunctor; -template class MaxSeqPoolGradFunctor; +// sequence pooling +template class SequencePoolFunctor; +template class SequencePoolFunctor; +template class SequencePoolGradFunctor; +template class SequencePoolGradFunctor; } // namespace math } // namespace operators diff --git a/paddle/fluid/operators/math/sequence_pooling.h b/paddle/fluid/operators/math/sequence_pooling.h index ecb76884f670df1aee64ed65c3bb0cf09c5beaff..38e780222955644c14e5bbbf16dee720c7758f5c 100644 --- a/paddle/fluid/operators/math/sequence_pooling.h +++ b/paddle/fluid/operators/math/sequence_pooling.h @@ -21,23 +21,23 @@ namespace paddle { namespace operators { namespace math { -#define FLT_MAX __FLT_MAX__ - template -class MaxSeqPoolFunctor { +class SequencePoolFunctor { public: - void operator()(const DeviceContext& context, + /* max pool has index output */ + void operator()(const DeviceContext& context, const std::string pooltype, const framework::LoDTensor& input, framework::Tensor* output, - framework::Tensor* index); + framework::Tensor* index = nullptr); }; -template -class MaxSeqPoolGradFunctor { +template +class SequencePoolGradFunctor { public: - void operator()(const DeviceContext& context, + void operator()(const DeviceContext& context, const std::string pooltype, const framework::Tensor& out_grad, - const framework::Tensor& index, - framework::LoDTensor* in_grad); + framework::LoDTensor* in_grad, + /* max pool has index */ + const framework::Tensor* index = nullptr); }; } // namespace math diff --git a/paddle/fluid/operators/mine_hard_examples_op.cc b/paddle/fluid/operators/mine_hard_examples_op.cc index 0e81d60878dce747b047abbe4641b71462373b2b..277901cff493445e1e85e92e22ea0ada0e1cba43 100644 --- a/paddle/fluid/operators/mine_hard_examples_op.cc +++ b/paddle/fluid/operators/mine_hard_examples_op.cc @@ -324,8 +324,9 @@ MatchIndices elements with value -1. } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(mine_hard_examples, ops::MineHardExamplesOp, - ops::MineHardExamplesOpMaker); +REGISTER_OPERATOR(mine_hard_examples, ops::MineHardExamplesOp, + ops::MineHardExamplesOpMaker, + paddle::framework::EmptyGradOpMaker); REGISTER_OP_CPU_KERNEL( mine_hard_examples, diff --git a/paddle/fluid/operators/mkldnn_activation_op.h b/paddle/fluid/operators/mkldnn_activation_op.h new file mode 100644 index 0000000000000000000000000000000000000000..083d03ebe610521c5a4beb7b977a8179700bcf40 --- /dev/null +++ b/paddle/fluid/operators/mkldnn_activation_op.h @@ -0,0 +1,111 @@ +/* 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" + +#ifdef PADDLE_WITH_MKLDNN +#include "paddle/fluid/platform/mkldnn_helper.h" +#endif + +namespace paddle { +namespace operators { + +template +class MKLDNNActivationKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + PADDLE_ENFORCE(context.Input("X") != nullptr, + "Cannot get input tensor X, variable name = %s", + context.op().Input("X")); + PADDLE_ENFORCE(context.Output("Out") != nullptr, + "Cannot find output tensor Out, variable name = %s", + context.op().Output("Out")); + Functor functor; + + auto attrs = functor.GetAttrs(); + for (auto& attr : attrs) { + *attr.second = context.Attr(attr.first); + } + functor(context); + } +}; + +template +class MKLDNNActivationGradKernel + : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + Functor functor; + + auto attrs = functor.GetAttrs(); + for (auto& attr : attrs) { + *attr.second = context.Attr(attr.first); + } + functor(context); + } +}; + +namespace { +framework::OpKernelType GetKernelType( + const framework::ExecutionContext& ctx, + const framework::OperatorWithKernel& oper) { + framework::LibraryType library{framework::LibraryType::kPlain}; +#ifdef PADDLE_WITH_MKLDNN + if (library == framework::LibraryType::kPlain && + platform::CanMKLDNNBeUsed(ctx)) { + library = framework::LibraryType::kMKLDNN; + } +#endif + framework::DataLayout layout = framework::DataLayout::kAnyLayout; + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.GetPlace(), layout, library); +} +} // anonymous namespace + +class ActivationWithMKLDNNOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareLoD("X", /*->*/ "Out"); + } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return GetKernelType(ctx, *this); + } +}; + +class ActivationWithMKLDNNOpGrad : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); + } + + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override { + return GetKernelType(ctx, *this); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/parallel_do_op.cc b/paddle/fluid/operators/parallel_do_op.cc index 4001b9a130348b4e3ea99f3017eae6d85e41fc6e..b28c16b13fce30c6e9be9953009b53e722cf4885 100644 --- a/paddle/fluid/operators/parallel_do_op.cc +++ b/paddle/fluid/operators/parallel_do_op.cc @@ -144,7 +144,12 @@ class ParallelDoOp : public framework::OperatorBase { PADDLE_ENFORCE(scope.FindVar(param)->IsType(), "Only support parameter type as LoDTensor"); auto &src = scope.FindVar(param)->Get(); - for (size_t i = 0; i < sub_scopes.size(); ++i) { + + auto *sub_scope0 = sub_scopes[0]; + auto *dst0 = sub_scope0->Var(param)->GetMutable(); + dst0->ShareDataWith(src); + + for (size_t i = 1; i < sub_scopes.size(); ++i) { auto &place = places[i]; auto *sub_scope = sub_scopes[i]; auto *dst = sub_scope->Var(param)->GetMutable(); diff --git a/paddle/fluid/operators/prefetch_op.cc b/paddle/fluid/operators/prefetch_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..09ab7da663b5ef5f099b9f65b0df661ceea0d9e2 --- /dev/null +++ b/paddle/fluid/operators/prefetch_op.cc @@ -0,0 +1,115 @@ +/* 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/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/send_recv_util.h" + +namespace paddle { +namespace operators { + +class PrefetchOp : public framework::OperatorBase { + public: + PrefetchOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void RunImpl(const framework::Scope& scope, + const platform::Place& place) const override { + auto ins = Inputs("X"); + auto outs = Outputs("Out"); + + std::vector epmap = Attr>("epmap"); + + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + + auto client_var_name = Output("RPCClient"); + PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name), + "Can not find variable '%s' in the scope.", + client_var_name); + auto* client_var = scope.FindVar(client_var_name); + detail::RPCClient* rpc_client = client_var->GetMutable(); + + 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"; + rpc_client->AsyncPrefetchVariable(epmap[i], ctx, scope, ins[i], + outs[i]); + } else { + VLOG(3) << "don't send no-initialied variable: " << ins[i]; + } + } + PADDLE_ENFORCE(rpc_client->Wait()); + } +}; + +class PrefetchOpMaker : public framework::OpProtoAndCheckerMaker { + public: + PrefetchOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "(LoDTensor) Input Id variables to be sent").AsDuplicable(); + AddOutput("RPCClient", + "(RPCClient) The RPC client object which will be" + "initialized at most once."); + AddOutput("Out", + "(SelectedRows) result " + "to be fetched from parameter server") + .AsDuplicable(); + AddAttr>( + "epmap", + "(string vector, default 127.0.0.1:6164)" + "Server endpoints in the order of input variables for mapping") + .SetDefault({"127.0.0.1:6164"}); + AddComment(R"DOC( +Prefetch operator + +This operator will send Ids variables to listen_and_serve op at +the parameter server and fetch result back. +)DOC"); + } +}; + +class PrefetchOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto out_var_name = op_desc.Output("RPCClient").front(); + auto& out_var = block->FindRecursiveOrCreateVar(out_var_name); + auto var_type = framework::proto::VarType::RAW; + out_var.SetType(var_type); + } +}; + +class PrefetchOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override {} +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(prefetch, ops::PrefetchOp, + paddle::framework::EmptyGradOpMaker, ops::PrefetchOpMaker, + ops::PrefetchOpVarTypeInference, + ops::PrefetchOpShapeInference); diff --git a/paddle/fluid/operators/prior_box_op.cc b/paddle/fluid/operators/prior_box_op.cc index 7ba55437cb20f802cc12ceea7777d7d78bba62a6..c22a55bce263423d5c17fffdb06b7ece02ae26da 100644 --- a/paddle/fluid/operators/prior_box_op.cc +++ b/paddle/fluid/operators/prior_box_op.cc @@ -168,7 +168,9 @@ https://arxiv.org/abs/1512.02325. } // namespace paddle namespace ops = paddle::operators; -REGISTER_OP_WITHOUT_GRADIENT(prior_box, ops::PriorBoxOp, ops::PriorBoxOpMaker); +REGISTER_OPERATOR(prior_box, ops::PriorBoxOp, ops::PriorBoxOpMaker, + paddle::framework::EmptyGradOpMaker); + REGISTER_OP_CPU_KERNEL( prior_box, ops::PriorBoxOpKernel, ops::PriorBoxOpKernel); diff --git a/paddle/fluid/operators/read_op.cc b/paddle/fluid/operators/read_op.cc index 2a5605e0d378a184ae132e657b2872279784855d..2925b8a85da1b0d19672124e49c8fd22c8b4e6bf 100644 --- a/paddle/fluid/operators/read_op.cc +++ b/paddle/fluid/operators/read_op.cc @@ -14,6 +14,7 @@ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/reader.h" +#include "paddle/fluid/operators/detail/safe_ref.h" namespace paddle { namespace operators { @@ -59,7 +60,9 @@ class ReadOp : public framework::OperatorBase { void RunImpl(const framework::Scope& scope, const platform::Place& dev_place) const override { framework::ReaderHolder* reader = - scope.FindVar(Input("Reader"))->GetMutable(); + detail::Ref(scope.FindVar(Input("Reader")), + "Cannot find reader variable %s", Input("Reader")) + .GetMutable(); std::vector out_arg_names = Outputs("Out"); std::vector ins; reader->ReadNext(&ins); diff --git a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc index 76cdb794ccdb4a015ae8630940a5c26845e7a7b3..f9a8058f2a32b6736d6513b017b761a31ddc2e37 100644 --- a/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc +++ b/paddle/fluid/operators/reader/create_double_buffer_reader_op.cc @@ -20,12 +20,29 @@ namespace paddle { namespace operators { namespace reader { -static constexpr size_t kDoubleBufferSize = 2; +// 'Double buffer' means we shall maintain two batches of input data at the same +// time. So the kCacheSize shoul be at least 2. +static constexpr size_t kCacheSize = 2; +// There will be two bacthes out of the channel during training: +// 1. the one waiting to be sent to the channel +// 2. the one just be received from the channel, which is also being used by +// subsequent operators. +// So the channel size should be kChacheSize - 2 +static constexpr size_t kChannelSize = 0; // kCacheSize - 2 class DoubleBufferReader : public framework::DecoratedReader { public: struct Item { Item() : ctx_(nullptr) {} + Item(Item&& b) { + payloads_ = std::move(b.payloads_); + ctx_ = std::move(b.ctx_); + } + Item& operator=(Item&& b) { + payloads_ = std::move(b.payloads_); + ctx_ = std::move(b.ctx_); + return *this; + } std::vector payloads_; platform::DeviceContext* ctx_; @@ -34,42 +51,44 @@ class DoubleBufferReader : public framework::DecoratedReader { explicit DoubleBufferReader( ReaderBase* reader, platform::Place target_place = platform::CPUPlace()) : DecoratedReader(reader), place_(target_place) { - for (size_t i = 0; i < kDoubleBufferSize; ++i) { - if (platform::is_gpu_place(place_)) { #ifdef PADDLE_WITH_CUDA + for (size_t i = 0; i < kCacheSize; ++i) { + if (platform::is_gpu_place(place_)) { ctxs_.emplace_back(new platform::CUDADeviceContext( boost::get(place_))); -#endif } } - - start_thread(); - } - - void start_thread() { - buffer_ = framework::MakeChannel(kDoubleBufferSize); - prefetcher_ = std::thread([this] { PrefetchThreadFunc(); }); +#endif + StartPrefetcher(); } + bool HasNext() const override; void ReadNext(std::vector* out) override; void ReInit() override; - ~DoubleBufferReader() { - buffer_->Close(); - prefetcher_.join(); - delete buffer_; + ~DoubleBufferReader() { EndPrefetcher(); } + + private: + void StartPrefetcher() { + channel_ = framework::MakeChannel(kChannelSize); + prefetcher_ = std::thread([this] { PrefetchThreadFunc(); }); } - bool HasNext() const override; + void EndPrefetcher() { + channel_->Close(); + if (prefetcher_.joinable()) { + prefetcher_.join(); + } + delete channel_; + channel_ = nullptr; + } - private: void PrefetchThreadFunc(); std::thread prefetcher_; - framework::Channel* buffer_; + framework::Channel* channel_; platform::Place place_; std::vector> ctxs_; - mutable Item local_buffer_; }; class CreateDoubleBufferReaderOp : public framework::OperatorBase { @@ -123,68 +142,70 @@ class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase { } }; +bool DoubleBufferReader::HasNext() const { + while (!channel_->IsClosed() && !channel_->CanReceive()) { + } + return channel_->CanReceive(); +} + void DoubleBufferReader::ReadNext(std::vector* out) { if (!HasNext()) { PADDLE_THROW("There is no next data!"); } - if (local_buffer_.payloads_.empty()) { - buffer_->Receive(&local_buffer_); - } - *out = local_buffer_.payloads_; - local_buffer_.payloads_.clear(); - if (local_buffer_.ctx_) { - local_buffer_.ctx_->Wait(); + Item batch; + channel_->Receive(&batch); + *out = batch.payloads_; + if (batch.ctx_) { + batch.ctx_->Wait(); } } void DoubleBufferReader::ReInit() { reader_->ReInit(); - buffer_->Close(); - prefetcher_.join(); - delete buffer_; - start_thread(); + EndPrefetcher(); + StartPrefetcher(); } void DoubleBufferReader::PrefetchThreadFunc() { VLOG(5) << "A new prefetch thread starts."; - size_t gpu_ctx_offset = 0; + std::vector> cpu_tensor_cache(kCacheSize); + std::vector> gpu_tensor_cache(kCacheSize); + size_t cached_tensor_id = 0; + while (reader_->HasNext()) { Item batch; - reader_->ReadNext(&batch.payloads_); + auto& cpu_batch = cpu_tensor_cache[cached_tensor_id]; + reader_->ReadNext(&cpu_batch); if (platform::is_gpu_place(place_)) { - std::vector gpu_batch; - auto& gpu_ctx = this->ctxs_[gpu_ctx_offset++]; - gpu_ctx_offset %= this->ctxs_.size(); - gpu_batch.resize(batch.payloads_.size()); - for (size_t i = 0; i < batch.payloads_.size(); ++i) { - framework::TensorCopy(batch.payloads_[i], place_, *gpu_ctx, - &gpu_batch[i]); - gpu_batch[i].set_lod(batch.payloads_[i].lod()); + auto& gpu_batch = gpu_tensor_cache[cached_tensor_id]; + auto* gpu_ctx = ctxs_[cached_tensor_id].get(); + gpu_batch.resize(cpu_batch.size()); + for (size_t i = 0; i < cpu_batch.size(); ++i) { + framework::TensorCopy(cpu_batch[i], place_, *gpu_ctx, &gpu_batch[i]); + gpu_batch[i].set_lod(cpu_batch[i].lod()); } - batch.ctx_ = gpu_ctx.get(); - std::swap(gpu_batch, batch.payloads_); + batch.payloads_ = gpu_batch; + batch.ctx_ = gpu_ctx; + } else { + // CPUPlace + batch.payloads_ = cpu_batch; } + ++cached_tensor_id; + cached_tensor_id %= kCacheSize; - if (!buffer_->Send(&batch)) { + try { + channel_->Send(&batch); + } catch (paddle::platform::EnforceNotMet e) { VLOG(5) << "WARNING: The double buffer channel has been closed. The " "prefetch thread will terminate."; break; } } - buffer_->Close(); + channel_->Close(); VLOG(5) << "Prefetch thread terminates."; } -bool DoubleBufferReader::HasNext() const { - if (local_buffer_.payloads_.empty()) { - bool ok = buffer_->Receive(&local_buffer_); - return ok; - } else { - return true; - } -} - } // namespace reader } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc index 4d4e9fb909eafea5328491a4097276577f28a5ba..47d9989bc8748840ec2d39587fde24355d90b6b4 100644 --- a/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc +++ b/paddle/fluid/operators/reader/create_multi_pass_reader_op.cc @@ -81,10 +81,10 @@ class CreateMultiPassReaderOpMaker : public DecoratedReaderMakerBase { This operator creates a multi-pass reader. A multi-pass reader is used to yield data for several pass training continuously. - It takes the the number of pass to run as one of its attributes + It takes the number of passes to run as one of its attributes ('pass_num'), and maintains a pass counter to record how many - passes it has completed. When the underlying reader reach the EOF, - the multi-pass reader checks whether it has completed training + passes it has completed. When the underlying reader reaches the + EOF, the multi-pass reader checks whether it has completed training of the given number of pass. If not, the underlying reader will be re-initialized and starts a new pass automatically. )DOC"); diff --git a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc index c4aa29c7206dbd3fe6a99b2a6c5ac6f083621944..adaa0b9e5f1ffcfbf3e9cd8fd060153575f270a6 100644 --- a/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc +++ b/paddle/fluid/operators/reader/create_recordio_file_reader_op.cc @@ -12,12 +12,15 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include +#include #include "paddle/fluid/operators/reader/reader_op_registry.h" #include "paddle/fluid/recordio/scanner.h" namespace paddle { namespace operators { namespace reader { +template class RecordIOFileReader : public framework::FileReader { public: explicit RecordIOFileReader(const std::string& filename, @@ -25,7 +28,12 @@ class RecordIOFileReader : public framework::FileReader { : FileReader(dims), scanner_(filename), dev_ctx_(*platform::DeviceContextPool::Instance().Get( - platform::CPUPlace())) {} + platform::CPUPlace())) { + if (ThreadSafe) { + mutex_.reset(new std::mutex()); + } + LOG(INFO) << "Creating file reader" << filename; + } bool HasNext() const override { return scanner_.HasNext(); } @@ -33,10 +41,16 @@ class RecordIOFileReader : public framework::FileReader { protected: void ReadNextImpl(std::vector* out) override { - *out = framework::ReadFromRecordIO(scanner_, dev_ctx_); + if (ThreadSafe) { + std::lock_guard guard(*mutex_); + *out = framework::ReadFromRecordIO(scanner_, dev_ctx_); + } else { + *out = framework::ReadFromRecordIO(scanner_, dev_ctx_); + } } private: + std::unique_ptr mutex_; recordio::Scanner scanner_; const platform::DeviceContext& dev_ctx_; }; @@ -59,8 +73,9 @@ class CreateRecordIOReaderOp : public framework::OperatorBase { auto* out = scope.FindVar(Output("Out")) ->template GetMutable(); - out->Reset( - new RecordIOFileReader(filename, RestoreShapes(shape_concat, ranks))); + + out->Reset(new RecordIOFileReader( + filename, RestoreShapes(shape_concat, ranks))); } }; @@ -87,4 +102,4 @@ REGISTER_FILE_READER_OPERATOR(create_recordio_file_reader, reader::CreateRecordIOReaderOp, reader::CreateRecordIOReaderOpMaker); -REGISTER_FILE_READER(recordio, reader::RecordIOFileReader); +REGISTER_FILE_READER(recordio, reader::RecordIOFileReader); diff --git a/paddle/fluid/operators/reader/open_files_op.cc b/paddle/fluid/operators/reader/open_files_op.cc index 414c76fea0bb916dfeafe38c0448a7a800889e03..b6ac7b21d56f7760b3f4814581c90b0ff2cc4a6a 100644 --- a/paddle/fluid/operators/reader/open_files_op.cc +++ b/paddle/fluid/operators/reader/open_files_op.cc @@ -146,14 +146,19 @@ void MultipleReader::PrefetchThreadFunc(std::string file_name, while (reader->HasNext()) { std::vector ins; reader->ReadNext(&ins); - if (!buffer_->Send(&ins)) { + try { + buffer_->Send(&ins); + } catch (paddle::platform::EnforceNotMet e) { VLOG(5) << "WARNING: The buffer channel has been closed. The prefetch " "thread of file '" << file_name << "' will terminate."; break; } } - if (!available_thread_idx_->Send(&thread_idx)) { + + try { + available_thread_idx_->Send(&thread_idx); + } catch (paddle::platform::EnforceNotMet e) { VLOG(5) << "WARNING: The available_thread_idx_ channel has been closed. " "Fail to send thread_idx."; } diff --git a/paddle/fluid/operators/reshape_op.cc b/paddle/fluid/operators/reshape_op.cc index 832509641cc3d5178ff090e05437484d395bfe51..b87b8e6b26cdeb017e700870998a53c1b295988c 100644 --- a/paddle/fluid/operators/reshape_op.cc +++ b/paddle/fluid/operators/reshape_op.cc @@ -17,90 +17,66 @@ limitations under the License. */ namespace paddle { namespace operators { -class ReshapeOp : public framework::OperatorWithKernel { - public: - ReshapeOp(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 { - // input check - PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of ReshapeOp should not be null."); - PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of ReshapeOp should not be null."); - - auto shape = ctx->Attrs().Get>("shape"); - PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty."); - auto x_dims = ctx->GetInputDim("X"); - - std::vector neg_dims_idx; - // set some dimension to -1 if it is unknown - const int unknown_size = -1; - for (size_t i = 0; i < shape.size(); ++i) { - PADDLE_ENFORCE(shape[i] > 0 || shape[i] == unknown_size, - "Each dimension of Attr(shape) must be positive or %d.", - unknown_size); - if (shape[i] == unknown_size) { - neg_dims_idx.push_back(i); - PADDLE_ENFORCE(neg_dims_idx.size() <= 1, - "Only one dimension of Attr(shape) can be unknown."); - } - } - - int64_t capacity = - std::accumulate(shape.begin(), shape.end(), 1, std::multiplies()); - int64_t in_size = framework::product(x_dims); - if (neg_dims_idx.size() == 1) { - // dim infer - shape[neg_dims_idx[0]] = in_size / (-capacity); - // recalculate capacity - capacity = shape[neg_dims_idx[0]] * (-capacity); - } - // capacity check - PADDLE_ENFORCE(capacity == in_size, - "The size of Input(X) mismatches with Attr(shape)."); - // resize output - std::vector shape_int64(shape.size(), 0); - std::transform(shape.begin(), shape.end(), shape_int64.begin(), - [](int a) { return static_cast(a); }); - auto out_dims = framework::make_ddim(shape_int64); - ctx->SetOutputDim("Out", out_dims); - if (shape[0] == x_dims[0]) { - // Only pass LoD when the first dimension is equal between - // output and input. - ctx->ShareLoD("X", /*->*/ "Out"); - } - } -}; - class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker { public: ReshapeOpMaker(OpProto *proto, OpAttrChecker *op_checker) : OpProtoAndCheckerMaker(proto, op_checker) { - AddInput("X", "The input tensor of reshape operator."); - AddOutput("Out", "The output tensor of reshape operator."); - AddAttr>("shape", - "(vector) " - "Target shape of reshape operator."); + AddInput("X", "(Tensor). The input tensor of reshape operator."); + AddInput("Shape", + "(Tensor, optional). If provided, reshape according to " + "this given shape. That is to say it has a higher priority than " + "the shape attribute, while the shape attribute still should be " + "set correctly to gurantee shape inference in compile time.") + .AsDispensable(); + AddOutput("Out", "(Tensor). The output tensor of reshape operator."); + AddAttr>( + "shape", "(std::vector) Target shape of reshape operator."); AddAttr("inplace", - "Change the source tensor's shape without copy memory.") - .SetDefault(true); + "(default: false) Change the source tensor's shape without " + "memory copy. When Attr(inplace) is set true, the output " + "tensor shares memory with Input(X), otherwise, a new output " + "tensor is created, and its data are copied from Input(x).") + .SetDefault(false); AddComment(R"DOC( Reshape Operator. -Reshape Input(X) into the shape specified by Attr(shape). +Reshape Input(X) into the shape specified by Attr(shape) or Input(Shape). The +data in Input(X) are unchanged. + +Examples: -An example: -Given a 2-D tensor X with 2 rows and 2 columns : [[1, 2], [3, 4]] +1. Given a 3-D tensor Input(X) with a shape [2, 4, 6], and the target shape +specified by Attr(shape) is [6, 8], the reshape operator will transform Input(X) +into a 2-D tensor with shape [6, 8] and leaving Input(X)'s data unchanged. -and target shape = [1, 4], the reshape operator will transform -the tensor X into a 2-D tensor: [[1, 2, 3, 4]] +2. Given a 3-D tensor Input(X) with a shape [2, 4, 6], and the target shape +specified by Attr(shape) is [2, 3, -1, 2], the reshape operator will transform +Input(X) into a 4-D tensor with shape [2, 3, 4, 2] and leaving Input(X)'s data +unchanged. In this case, one and only dimension of Attr(shape) can be set to -1, +the value of this dimension is inferred from the total element number of +Input(X) and remaining dimensions. + +3. Given a 3-D tensor Input(X) with a shape [2, 4, 6], and the target shape +specified by Attr(shape) is [-1, 0, 3, 2], the reshape operator will transform +Input(X) into a 4-D tensor with shape [2, 4, 3, 2] and leaving Input(X)'s data +unchanged. In this case, besides -1, 0 means the actual dimension value is going +to be copied from the corresponding dimension of Input(X). + +Note: + +1. One and only one dimension in Attr(shape) can be set -1. In this case, +the actual dimension value will be infered from the total element number of +Input(X) and remaining dimensions. + +2. More than one dimensions in Attr(shape) can be set to 0, which means the real +dimension value will be copied from Input(X) at runtime. Note that the index of +0 can not exceed Rank(X). For example, Input(X) is a 3-D tensor with shape +[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 +compile-time. -One dimension in the target shape can be set -1, representing that its -size is unknown. In this case, the real dimension will be infered from -the original shape of Input(X) and other dimensions in the target shape. )DOC"); } }; @@ -119,6 +95,14 @@ class ReshapeGradOp : public framework::OperatorWithKernel { "Input(Out@GRAD) shouldn't be null."); ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } }; } // namespace operators diff --git a/paddle/fluid/operators/reshape_op.h b/paddle/fluid/operators/reshape_op.h index eacb0a0cf21a60ffbdef5787434859ac549388bc..871b4d38d56f10f3c0c178caa566508ab75f316c 100644 --- a/paddle/fluid/operators/reshape_op.h +++ b/paddle/fluid/operators/reshape_op.h @@ -20,17 +20,129 @@ limitations under the License. */ namespace paddle { namespace operators { +class ReshapeOp : public framework::OperatorWithKernel { + public: + ReshapeOp(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 ReshapeOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of ReshapeOp should not be null."); + + const std::vector &shape = ctx->Attrs().Get>("shape"); + PADDLE_ENFORCE(!shape.empty(), + "The shape information must be set by Attr(shape)."); + + if (ctx->HasInput("Shape") && ctx->IsRuntime()) { + // If true, set the shape of Output(Out) according to Input(Shape) in + // ReshapeKernel with ExecutionContext. Also check LoD in ReshapeKernel. + ctx->ShareLoD("X", /*->*/ "Out"); + return; + } + + auto x_dims = ctx->GetInputDim("X"); + auto out_dims = ValidateShape(shape, x_dims); + 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"); + } + } + + static framework::DDim ValidateShape(const std::vector shape, + const framework::DDim &in_dims) { + const int64_t in_size = framework::product(in_dims); + // only one dimension canbe set to -1, whose size will be automatically + // infered. + const int64_t unk_dim_val = -1; + const int64_t copy_dim_val = 0; + + std::vector output_shape(shape.size(), 0); + int64_t capacity = 1; + int unk_dim_idx = -1; + for (size_t i = 0; i < shape.size(); ++i) { + if (shape[i] == unk_dim_val) { + PADDLE_ENFORCE( + unk_dim_idx == -1, + "Only one input dimension of Attr(shape) can be unknown."); + unk_dim_idx = i; + } else if (shape[i] == copy_dim_val) { + PADDLE_ENFORCE( + static_cast(i) < in_dims.size(), + "The index of dimension to copy from input shape must be less " + "than the size of input shape."); + } else { + PADDLE_ENFORCE( + shape[i] > 0, + "Each input dimension of Attr(shape) must not be negtive except " + "one unknown dimension."); + } + + capacity *= (shape[i] ? shape[i] : in_dims[i]); + output_shape[i] = + (shape[i] ? static_cast(shape[i]) : in_dims[i]); + } + + if (unk_dim_idx != -1) { + output_shape[unk_dim_idx] = -in_size / capacity; + PADDLE_ENFORCE_EQ(output_shape[unk_dim_idx] * capacity, -in_size, + "Invalid shape is given."); + } else { + PADDLE_ENFORCE_EQ(capacity, in_size, "Invalid shape is given."); + } + return framework::make_ddim(output_shape); + } + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext &ctx) const override { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); + } +}; + template class ReshapeKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const { - auto* out = ctx.Output("Out"); - auto* in = ctx.Input("X"); + void Compute(const framework::ExecutionContext &ctx) const { + auto *out = ctx.Output("Out"); + auto *in = ctx.Input("X"); + auto *shape_tensor = ctx.Input("Shape"); + + framework::DDim out_dims = out->dims(); + if (shape_tensor) { + auto *shape_data = shape_tensor->data(); + if (platform::is_gpu_place(ctx.GetPlace())) { + framework::Tensor cpu_shape_tensor; + TensorCopy(*shape_tensor, platform::CPUPlace(), ctx.device_context(), + &cpu_shape_tensor); + shape_data = cpu_shape_tensor.data(); + } + auto shape = + std::vector(shape_data, shape_data + shape_tensor->numel()); + out_dims = ReshapeOp::ValidateShape(shape, in->dims()); + } + if (!in->lod().empty()) { + PADDLE_ENFORCE_EQ( + out_dims[0], in->dims()[0], + "Reshape operator cannot reshape an input sequence batch " + "into an output sequence batch that has a different " + "number of time steps. Please consider using " + "sequence_reshape op."); + } + bool inplace = ctx.Attr("inplace"); - auto out_dims = out->dims(); + out->Resize(out_dims); if (!inplace) { out->mutable_data(ctx.GetPlace()); framework::TensorCopy(*in, ctx.GetPlace(), ctx.device_context(), out); + // TensorCopy will resize to in_dims. out->Resize(out_dims); } else { out->ShareDataWith(*in); @@ -42,9 +154,10 @@ class ReshapeKernel : public framework::OpKernel { template class ReshapeGradKernel : public framework::OpKernel { public: - void Compute(const framework::ExecutionContext& ctx) const { - auto* d_out = ctx.Input(framework::GradVarName("Out")); - auto* d_x = ctx.Output(framework::GradVarName("X")); + void Compute(const framework::ExecutionContext &ctx) const { + auto *d_out = ctx.Input(framework::GradVarName("Out")); + auto *d_x = ctx.Output(framework::GradVarName("X")); + d_x->mutable_data(ctx.GetPlace()); bool inplace = ctx.Attr("inplace"); diff --git a/paddle/fluid/operators/select_op.cc b/paddle/fluid/operators/select_op.cc index 8344a239df7b3fcbe91f91a17a3c5958013b55a6..c0bf0ff927481bc4da9cd6c4bb9b0c4a6841c891 100644 --- a/paddle/fluid/operators/select_op.cc +++ b/paddle/fluid/operators/select_op.cc @@ -27,6 +27,7 @@ namespace operators { static constexpr char kX[] = "X"; static constexpr char kCaseToExecute[] = "case_to_execute"; +static constexpr char kOutputs[] = "Out"; static constexpr char kCases[] = "cases"; static constexpr char kCasesBlock[] = "sub_block"; @@ -388,6 +389,10 @@ class SelectOpMaker : public framework::OpProtoAndCheckerMaker { "(Int) The variable the sets the index of the case to execute, " "after evaluating the channels being sent to and received from") .AsDuplicable(); + AddOutput(kOutputs, + "A set of variables, which will be assigned with values " + "generated by the operators inside the cases of Select Op.") + .AsDuplicable(); AddAttr>(kCases, "(String vector) Serialized list of" "all cases in the select op. Each" diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..8d02a6f29177536562e38372eb0336424aa0a47c --- /dev/null +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -0,0 +1,103 @@ +/* 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 + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/framework.pb.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" + +#include +#include "paddle/fluid/operators/detail/grpc_client.h" + +namespace paddle { +namespace operators { + +class SendBarrierOp : public framework::OperatorBase { + public: + SendBarrierOp(const std::string& type, + const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void RunImpl(const framework::Scope& scope, + const platform::Place& place) const override { + std::vector eps = Attr>("endpoints"); + + auto client_var_name = Output("RPCClient"); + PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name), + "Can not find variable '%s' in the scope.", + client_var_name); + auto* client_var = scope.FindVar(client_var_name); + detail::RPCClient* rpc_client = client_var->GetMutable(); + + // need to wait before sending send_barrier message + PADDLE_ENFORCE(rpc_client->Wait()); + + for (auto& ep : eps) { + VLOG(3) << "send barrier, ep: " << ep; + rpc_client->AsyncSendBatchBarrier(ep); + } + PADDLE_ENFORCE(rpc_client->Wait()); + } +}; + +class SendBarrierOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SendBarrierOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddOutput("RPCClient", + "(RPCClient) The RPC client object which is" + "initialized at most once."); + AddComment(R"DOC( +SendBarrier operator + +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>("endpoints", + "(string vector, default 127.0.0.1:6164)" + "Server endpoints to send variables to.") + .SetDefault({"127.0.0.1:6164"}); + } +}; + +class SendBarrierOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto out_var_name = op_desc.Output("RPCClient").front(); + auto& out_var = block->FindRecursiveOrCreateVar(out_var_name); + auto var_type = framework::proto::VarType::RAW; + out_var.SetType(var_type); + } +}; + +class SendBarrierOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override {} +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(send_barrier, ops::SendBarrierOp, + paddle::framework::EmptyGradOpMaker, ops::SendBarrierOpMaker, + ops::SendBarrierOpVarTypeInference, + ops::SendBarrierOpShapeInference); diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index 443f40e803ea31c3961ed77842bd0775e0f74f35..d47f66de2161dce7ed162db4c2e23859e19596cb 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -12,34 +12,19 @@ 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/data_type.h" #include "paddle/fluid/framework/framework.pb.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" - -#include #include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/send_recv_util.h" +#include "paddle/fluid/platform/profiler.h" namespace paddle { namespace operators { -static bool NeedSend(const framework::Scope& scope, - const std::string& varname) { - auto* var = scope.FindVar(varname); - PADDLE_ENFORCE_NOT_NULL(var, "Can not find variable '%s' in the send side.", - varname); - if (var->IsType()) { - return var->Get().IsInitialized(); - } else if (var->IsType()) { - return var->Get().rows().size() > 0UL; - } else { - PADDLE_THROW( - "Variable type in send side should be in " - "[LodTensor, SelectedRows]"); - } - return false; -} class SendOp : public framework::OperatorBase { public: @@ -59,6 +44,9 @@ class SendOp : public framework::OperatorBase { platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); auto& ctx = *pool.Get(place); + // For profiling + platform::RecordEvent record_event(Type(), &ctx); + auto client_var_name = Output("RPCClient"); PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name), "Can not find variable '%s' in the scope.", @@ -84,13 +72,13 @@ class SendOp : public framework::OperatorBase { if (outs.size() > 0) { for (size_t i = 0; i < outs.size(); i++) { - VLOG(3) << "getting " << outs[i] << " from " << epmap[i]; + VLOG(2) << "getting " << outs[i] << " from " << epmap[i]; rpc_client->AsyncGetVariable(epmap[i], ctx, scope, outs[i]); } PADDLE_ENFORCE(rpc_client->Wait()); // tell pservers that current trainer have called fetch for (auto& ep : endpoints) { - VLOG(3) << "send fetch barrier, ep: " << ep; + VLOG(2) << "send fetch barrier, ep: " << ep; rpc_client->AsyncSendFetchBarrier(ep); } PADDLE_ENFORCE(rpc_client->Wait()); diff --git a/paddle/fluid/operators/send_recv_op_test.cc b/paddle/fluid/operators/send_recv_op_test.cc index e9fb845b475ff5776bf948ab120a44c16ed87aa0..04392b3e05fa2d8b602946ba03672bf2491dcfbc 100644 --- a/paddle/fluid/operators/send_recv_op_test.cc +++ b/paddle/fluid/operators/send_recv_op_test.cc @@ -122,7 +122,8 @@ void StartServerNet(bool is_sparse) { // sub program run in listen_and_serv_op, for simple test we use sum f::ProgramDesc program; - f::BlockDesc *optimize_block = program.MutableBlock(0); + const auto &root_block = program.Block(0); + auto *optimize_block = program.AppendBlock(root_block); // X for server side tensors, RX for received tensers, must be of same shape. AddOp("sum", {{"X", {"x0", "x1"}}}, {{"Out", {"Out"}}}, {}, optimize_block); diff --git a/paddle/fluid/operators/send_recv_util.h b/paddle/fluid/operators/send_recv_util.h new file mode 100644 index 0000000000000000000000000000000000000000..196f56f6340a75b599b8dd15957dfe6835f9bf59 --- /dev/null +++ b/paddle/fluid/operators/send_recv_util.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. */ + +namespace paddle { +namespace operators { + +inline bool NeedSend(const framework::Scope& scope, + const std::string& varname) { + auto* var = scope.FindVar(varname); + PADDLE_ENFORCE_NOT_NULL(var, "Can not find variable '%s' in the send side.", + varname); + if (var->IsType()) { + return var->Get().IsInitialized(); + } else if (var->IsType()) { + return var->Get().rows().size() > 0UL; + } else { + PADDLE_THROW( + "Variable type in send side should be in " + "[LodTensor, SelectedRows]"); + } + return false; +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/send_vars_op.cc b/paddle/fluid/operators/send_vars_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..2cbd9e2394800dc3b9c5be1163d16bbec435c533 --- /dev/null +++ b/paddle/fluid/operators/send_vars_op.cc @@ -0,0 +1,117 @@ +/* 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 +#include + +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/detail/grpc_client.h" +#include "paddle/fluid/operators/send_recv_util.h" + +namespace paddle { +namespace operators { + +class SendVarsOp : public framework::OperatorBase { + public: + SendVarsOp(const std::string& type, const framework::VariableNameMap& inputs, + const framework::VariableNameMap& outputs, + const framework::AttributeMap& attrs) + : OperatorBase(type, inputs, outputs, attrs) {} + + void RunImpl(const framework::Scope& scope, + const platform::Place& place) const override { + auto ins = Inputs("X"); + + std::vector epmap = Attr>("epmap"); + int sync_send = Attr("sync_sent"); + + platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance(); + auto& ctx = *pool.Get(place); + + auto client_var_name = Output("RPCClient"); + PADDLE_ENFORCE_NOT_NULL(scope.FindVar(client_var_name), + "Can not find variable '%s' in the scope.", + client_var_name); + auto* client_var = scope.FindVar(client_var_name); + detail::RPCClient* rpc_client = client_var->GetMutable(); + + for (size_t i = 0; i < ins.size(); i++) { + if (NeedSend(scope, ins[i])) { + VLOG(3) << "sending " << ins[i] << " to " << epmap[i]; + // TODO(Yancey1989): we need to use an IO threadpool which has + // a larger number of threads than the computing threadpool. + rpc_client->AsyncSendVariable(epmap[i], ctx, scope, ins[i]); + } else { + VLOG(3) << "don't send no-initialied variable: " << ins[i]; + } + } + if (sync_send) { + rpc_client->Wait(); + } + } +}; + +class SendVarsOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SendVarsOpMaker(OpProto* proto, OpAttrChecker* op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("X", "(Tensor, SelectedRows) Input variables to be sent") + .AsDuplicable(); + AddOutput("RPCClient", + "(RPCClient) The RPC client object which will be" + "initialized at most once."); + AddComment(R"DOC( +Send operator + +This operator will send variables to listen_and_serve op at the parameter server. +)DOC"); + AddAttr("sync_send", + "(int, default 0)" + "sync send or async send.") + .SetDefault(0); + AddAttr>("epmap", + "(string vector, default 127.0.0.1:6164)" + "Server endpoints in the order of input " + "variables for mapping") + .SetDefault({"127.0.0.1:6164"}); + } +}; + +class SendVarsOpVarTypeInference : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto out_var_name = op_desc.Output("RPCClient").front(); + auto& out_var = block->FindRecursiveOrCreateVar(out_var_name); + auto var_type = framework::proto::VarType::RAW; + out_var.SetType(var_type); + } +}; + +class SendVarsOpShapeInference : public framework::InferShapeBase { + public: + void operator()(framework::InferShapeContext* ctx) const override {} +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; + +REGISTER_OPERATOR(send_vars, ops::SendVarsOp, + paddle::framework::EmptyGradOpMaker, ops::SendVarsOpMaker, + ops::SendVarsOpVarTypeInference, + ops::SendVarsOpShapeInference); diff --git a/paddle/fluid/operators/sequence_pool_op.h b/paddle/fluid/operators/sequence_pool_op.h index 8706ff14aa20714e77d5625fc1f6287ee9b4a8a6..c58d677c92b7a20eb54dc5f9a447566e91bdc3d4 100644 --- a/paddle/fluid/operators/sequence_pool_op.h +++ b/paddle/fluid/operators/sequence_pool_op.h @@ -23,12 +23,6 @@ namespace operators { using Tensor = framework::Tensor; using LoDTensor = framework::LoDTensor; -template -using EigenVector = framework::EigenVector; -template -using EigenMatrix = framework::EigenMatrix; template class SequencePoolKernel : public framework::OpKernel { @@ -37,11 +31,13 @@ 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(); - int64_t w = in->numel() / dims[0]; - // InferShape by lod PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now."); PADDLE_ENFORCE_GE( @@ -50,45 +46,14 @@ class SequencePoolKernel : public framework::OpKernel { "The first dimension of Input(X) must be large than batch size."); dims[0] = lod[0].size() - 1; out->Resize({dims}); - - auto lod_level_0 = lod[0]; - out->mutable_data(context.GetPlace()); - auto& dev_ctx = context.template device_context(); if (pooltype == "MAX") { - math::MaxSeqPoolFunctor max_pool; - auto* index = context.Output("MaxIndex"); index->Resize({dims}); index->mutable_data(context.GetPlace()); - max_pool(dev_ctx, *in, out, index); - return; - } - - auto& place = - *context.template device_context().eigen_device(); - for (int i = 0; i < static_cast(lod_level_0.size()) - 1; ++i) { - Tensor in_t = in->Slice(static_cast(lod_level_0[i]), - static_cast(lod_level_0[i + 1])); - Tensor out_t = out->Slice(i, i + 1); - int64_t h = static_cast(lod_level_0[i + 1] - lod_level_0[i]); - auto in_e = EigenMatrix::From(in_t, framework::make_ddim({h, w})); - auto out_e = EigenVector::Flatten(out_t); - - 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}})); - } else if (pooltype == "SQRT") { - out_e.device(place) = in_e.sum(Eigen::array({{0}})) / - std::sqrt(static_cast(h)); - } else if (pooltype == "LAST") { - out_e.device(place) = in_e.chip(h - 1, 0); - } else if (pooltype == "FIRST") { - out_e.device(place) = in_e.chip(0, 0); - } else { - PADDLE_THROW("unsupported pooling pooltype"); - } } + math::SequencePoolFunctor pool; + pool(context.template device_context(), pooltype, *in, out, + index); } }; @@ -96,58 +61,17 @@ template class SequencePoolGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* in = context.Input("X"); auto* out_g = context.Input(framework::GradVarName("Out")); auto* in_g = context.Output(framework::GradVarName("X")); std::string pooltype = context.Attr("pooltype"); - - auto dims = in->dims(); - auto lod = in->lod()[0]; - int64_t w = in->numel() / dims[0]; - - in_g->mutable_data(context.GetPlace()); - auto& dev_ctx = context.template device_context(); - + const Tensor* index = nullptr; if (pooltype == "MAX") { - math::MaxSeqPoolGradFunctor max_pool_grad; - auto* index = context.Input("MaxIndex"); - max_pool_grad(dev_ctx, *out_g, *index, in_g); - return; - } - - if (pooltype == "LAST" || pooltype == "FIRST") { - // set X@Grad be zero at first when pooltype is LAST/FIRST - math::SetConstant functor; - functor(dev_ctx, in_g, 0); - } - auto& place = - *context.template device_context().eigen_device(); - - for (int i = 0; i < static_cast(lod.size()) - 1; ++i) { - auto in_g_t = - in_g->Slice(static_cast(lod[i]), static_cast(lod[i + 1])); - auto out_g_t = out_g->Slice(i, i + 1); - int64_t h = static_cast(lod[i + 1] - lod[i]); - auto in_g_e = EigenMatrix::From(in_g_t, {h, w}); - auto out_g_e = EigenMatrix::From(out_g_t, {1, w}); - auto out_g_e_v = EigenVector::Flatten(out_g_t); - Eigen::DSizes bcast(h, 1); - - 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); - } else if (pooltype == "LAST") { - in_g_e.chip(h - 1, 0).device(place) = out_g_e_v; - } else if (pooltype == "FIRST") { - in_g_e.chip(0, 0).device(place) = out_g_e_v; - } else { - PADDLE_THROW("unsupported pooling pooltype"); - } + index = context.Input("MaxIndex"); } + in_g->mutable_data(context.GetPlace()); + math::SequencePoolGradFunctor pool; + pool(context.template device_context(), pooltype, *out_g, + in_g, index); } }; diff --git a/paddle/fluid/operators/split_ids_op.cc b/paddle/fluid/operators/split_ids_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a54f8a2878c8606e6b487552324d1e7dfa94b9b8 --- /dev/null +++ b/paddle/fluid/operators/split_ids_op.cc @@ -0,0 +1,76 @@ +/* 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/split_ids_op.h" + +namespace paddle { +namespace operators { + +class SplitIdsOpMaker : public framework::OpProtoAndCheckerMaker { + public: + SplitIdsOpMaker(OpProto *proto, OpAttrChecker *op_checker) + : OpProtoAndCheckerMaker(proto, op_checker) { + AddInput("Ids", "(LoDTensor) the input ids with shape{batch_num, 1}"); + AddOutput("Out", "(LoDTensor) 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] + + Out(3 output): + out0 = [3, 6] + out1 = [1, 4] + out2 = [2, 5] +)DOC"); + } +}; + +class SplitIdsOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext *ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("Ids"), "SplitIdsOp must has input Ids."); + PADDLE_ENFORCE(ctx->HasOutputs("Out"), "SplitIdsOp must has output Out."); + + auto ids_var_type = ctx->GetInputsVarType("Ids").front(); + PADDLE_ENFORCE_EQ(ids_var_type, framework::proto::VarType::LOD_TENSOR); + + auto ids_dims = ctx->GetInputDim("Ids"); + PADDLE_ENFORCE_EQ(ids_dims.size(), 2); + PADDLE_ENFORCE_EQ(ids_dims[1], 1); + } +}; + +class SplitIdsOpInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc &op_desc, + framework::BlockDesc *block) const override { + for (auto &out_var : op_desc.Output("Out")) { + block->Var(out_var)->SetType(framework::proto::VarType::LOD_TENSOR); + } + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(split_ids, ops::SplitIdsOp, ops::SplitIdsOpMaker, + ops::SplitIdsOpInferVarType); +REGISTER_OP_CPU_KERNEL( + split_ids, ops::SplitIdsOpKernel); diff --git a/paddle/fluid/operators/split_ids_op.h b/paddle/fluid/operators/split_ids_op.h new file mode 100644 index 0000000000000000000000000000000000000000..3e750ed2d171876ce2d3c232f5d34234217b3c3e --- /dev/null +++ b/paddle/fluid/operators/split_ids_op.h @@ -0,0 +1,65 @@ +/* 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/operators/math/selected_rows_functor.h" + +namespace paddle { +namespace operators { + +template +class SplitIdsOpKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto place = ctx.GetPlace(); + if (!platform::is_cpu_place(place)) { + PADDLE_THROW("SplitIds do not support GPU kernel"); + } + + const auto* ids_t = ctx.Input("Ids"); + auto& ids_dims = ids_t->dims(); + auto outs = ctx.MultiOutput("Out"); + + const T* ids = ids_t->data(); + + const size_t shard_num = outs.size(); + + std::vector> out_ids; + out_ids.resize(outs.size()); + + // split id by their shard_num. + for (size_t i = 0; i < ids_dims[0]; ++i) { + T id = ids[i]; + size_t shard_id = static_cast(id) % shard_num; + out_ids[shard_id].push_back(id); + } + + // create tensor for each shard and send to parameter server + for (size_t i = 0; i < out_ids.size(); ++i) { + auto* shard_t = outs[i]; + std::vector ids = out_ids[i]; + auto* shard_data = shard_t->mutable_data( + framework::make_ddim({static_cast(ids.size()), 1}), place); + for (size_t i = 0; i < ids.size(); ++i) { + shard_data[i] = ids[i]; + } + } + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/target_assign_op.cc b/paddle/fluid/operators/target_assign_op.cc index a894b12fa35a121eff0b8f9d2d0eecc5ae5185f3..33ff967e5e8f5afbaa62ba39ce596687ae0a71cd 100644 --- a/paddle/fluid/operators/target_assign_op.cc +++ b/paddle/fluid/operators/target_assign_op.cc @@ -153,8 +153,8 @@ template struct NegTargetAssignFunctor, diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt index 7eec6ab657723c6390dfa14a78d6c49a76f2a279..686c0889140f0050b37192542ca98e2f3e5f23df 100644 --- a/paddle/fluid/platform/CMakeLists.txt +++ b/paddle/fluid/platform/CMakeLists.txt @@ -49,7 +49,7 @@ nv_test(device_context_test SRCS device_context_test.cu DEPS device_context gpu_ nv_test(cudnn_helper_test SRCS cudnn_helper_test.cc DEPS dynload_cuda) nv_test(transform_test SRCS transform_test.cu DEPS paddle_memory place device_context) -cc_library(device_tracer SRCS device_tracer.cc DEPS profiler_proto ${GPU_CTX_DEPS}) +cc_library(device_tracer SRCS device_tracer.cc DEPS boost profiler_proto ${GPU_CTX_DEPS}) cc_library(profiler SRCS profiler.cc DEPS device_context device_tracer) cc_test(profiler_test SRCS profiler_test.cc DEPS profiler) diff --git a/paddle/fluid/platform/nccl_helper.h b/paddle/fluid/platform/nccl_helper.h new file mode 100644 index 0000000000000000000000000000000000000000..29990043206509e4192bfff84832f09ef127d9dd --- /dev/null +++ b/paddle/fluid/platform/nccl_helper.h @@ -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. + +#pragma once + +#include +#include +#include "paddle/fluid/platform/dynload/nccl.h" +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace platform { + +inline ncclDataType_t ToNCCLDataType(std::type_index type) { + if (type == typeid(float)) { // NOLINT + return ncclFloat; + } else if (type == typeid(double)) { // NOLINT + return ncclDouble; + } else if (type == typeid(int)) { // NOLINT + return ncclInt; + } else { + PADDLE_THROW("Not supported"); + } +} + +class NCCLGroupGuard { + public: + inline NCCLGroupGuard() { + mutex().lock(); + PADDLE_ENFORCE(dynload::ncclGroupStart()); + } + + inline ~NCCLGroupGuard() { + PADDLE_ENFORCE(dynload::ncclGroupEnd()); + mutex().unlock(); + } + + private: + static std::mutex &mutex() { + static std::mutex mtx; + return mtx; + } +}; + +struct NCCLContext { + std::unique_ptr ctx_; + ncclComm_t comm_; + + explicit NCCLContext(int dev_id) + : ctx_(new CUDADeviceContext(CUDAPlace(dev_id))) {} + + cudaStream_t stream() const { return ctx_->stream(); } + + int device_id() const { + return boost::get(ctx_->GetPlace()).device; + } + + static void InitNCCLContext(std::unordered_map &contexts, + const std::vector &places) { + std::vector comms; + std::vector devs; + comms.resize(contexts.size()); + devs.reserve(contexts.size()); + + for (auto &p : places) { + devs.push_back(boost::get(p).device); + } + + PADDLE_ENFORCE(platform::dynload::ncclCommInitAll( + &comms[0], static_cast(contexts.size()), &devs[0])); + + int i = 0; + for (auto &dev_id : devs) { + contexts.at(dev_id).comm_ = comms[i++]; + } + } +}; + +struct NCCLContextMap { + std::unordered_map contexts_; + std::vector order_; + + NCCLContextMap(const std::vector &places) { + order_.reserve(places.size()); + for (auto &p : places) { + int dev_id = boost::get(p).device; + order_.emplace_back(dev_id); + contexts_.emplace(dev_id, NCCLContext(dev_id)); + } + PADDLE_ENFORCE_EQ( + order_.size(), contexts_.size(), + "NCCL Context Map does not support contain two or more same device"); + + std::vector comms; + comms.resize(order_.size()); + + PADDLE_ENFORCE(platform::dynload::ncclCommInitAll( + &comms[0], static_cast(order_.size()), &order_[0])); + + int i = 0; + for (auto &dev_id : order_) { + contexts_.at(dev_id).comm_ = comms[i++]; + } + } + + CUDADeviceContext *DevCtx(int dev_id) const { return at(dev_id).ctx_.get(); } + + CUDADeviceContext *DevCtx(platform::Place p) const { + return DevCtx(boost::get(p).device); + } + + const NCCLContext &at(platform::Place p) const { + return this->at(boost::get(p).device); + } + + const NCCLContext &at(int dev_id) const { return contexts_.at(dev_id); } + + void WaitAll() { + for (auto &p : contexts_) { + p.second.ctx_->Wait(); + } + } +}; + +} // namespace platform +} // namespace paddle diff --git a/paddle/fluid/platform/profiler_test.cc b/paddle/fluid/platform/profiler_test.cc index fc77e0f3213da776e0b05ad5b5da9081665cdf6e..45cc271bb888fc3a07ecc5daea6b549cb88b6d21 100644 --- a/paddle/fluid/platform/profiler_test.cc +++ b/paddle/fluid/platform/profiler_test.cc @@ -13,6 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/platform/profiler.h" +#ifdef PADDLE_WITH_CUDA +#include "cuda_runtime.h" +#endif #include "gtest/gtest.h" TEST(Event, CpuElapsedTime) { @@ -157,3 +160,13 @@ TEST(RecordEvent, RecordEvent) { // Will remove parsing-related code from test later DisableProfiler(EventSortingKey::kTotal, "/tmp/profiler"); } + +#ifdef PADDLE_WITH_CUDA +TEST(TMP, stream_wait) { + cudaStream_t stream; + cudaStreamCreate(&stream); + cudaStreamSynchronize(stream); + cudaStreamSynchronize(stream); + cudaStreamSynchronize(stream); +} +#endif diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt index 8942b5c9430ffa4e499b0ad1d2b5acf6d18ec0ab..ada69ea4a425f70dc085ad9046bb6b930136803d 100644 --- a/paddle/fluid/pybind/CMakeLists.txt +++ b/paddle/fluid/pybind/CMakeLists.txt @@ -1,9 +1,18 @@ if(WITH_PYTHON) - cc_library(paddle_pybind SHARED - SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc - DEPS pybind python backward proto_desc paddle_memory executor prune init profiler feed_fetch_method - ${GLOB_OP_LIB}) - if(NOT APPLE AND NOT ANDROID) - target_link_libraries(paddle_pybind rt) - endif(NOT APPLE AND NOT ANDROID) + if(WITH_AMD_GPU) + hip_library(paddle_pybind SHARED + SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc + DEPS pybind python backward proto_desc paddle_memory executor prune init profiler feed_fetch_method + parallel_executor + ${GLOB_OP_LIB}) + else() + cc_library(paddle_pybind SHARED + SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc + DEPS pybind python backward proto_desc paddle_memory executor prune init profiler feed_fetch_method + parallel_executor + ${GLOB_OP_LIB}) + if(NOT APPLE AND NOT ANDROID) + target_link_libraries(paddle_pybind rt) + endif(NOT APPLE AND NOT ANDROID) + endif(WITH_AMD_GPU) endif(WITH_PYTHON) diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 6c05442466f5f3d8e04a8f0a2206443b1007a107..e1b1bbec97985aa839c62a0a82b81b020faf0008 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -25,6 +25,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_rank_table.h" #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/lod_tensor_array.h" +#include "paddle/fluid/framework/parallel_executor.h" #include "paddle/fluid/framework/prune.h" #include "paddle/fluid/framework/reader.h" #include "paddle/fluid/framework/selected_rows.h" @@ -496,6 +497,20 @@ All parameter, weight, gradient are variables in Paddle. m.def("disable_profiler", platform::DisableProfiler); m.def("reset_profiler", platform::ResetProfiler); + py::class_(m, "ParallelExecutor") + .def("__init__", + [](ParallelExecutor &self, size_t num_threads, bool use_event, + const std::vector &places, + const std::unordered_set ¶ms, + const ProgramDesc &startup_program, + const ProgramDesc &main_program, const std::string &loss_var_name, + Scope *scope) { + new (&self) ParallelExecutor(num_threads, use_event, places, + params, startup_program, main_program, + loss_var_name, scope); + }) + .def("run", &ParallelExecutor::Run); + BindRecordIOWriter(m); return m.ptr(); } diff --git a/paddle/math/MathFunctions.cpp b/paddle/math/MathFunctions.cpp index b2ff4bc3232a8e5d5d7b49bf49c62fe756d303f4..de404cad89fba8021b8645a40e25c1f5b7e86596 100644 --- a/paddle/math/MathFunctions.cpp +++ b/paddle/math/MathFunctions.cpp @@ -59,17 +59,10 @@ void* lapack_dso_handle = nullptr; } __name; // struct DynLoad__##__name #endif -#ifdef PADDLE_USE_ATLAS - #define PADDLE_SGETRF clapack_sgetrf - #define PADDLE_DGETRF clapack_dgetrf - #define PADDLE_SGETRI clapack_sgetri - #define PADDLE_DGETRI clapack_dgetri -#else - #define PADDLE_SGETRF LAPACKE_sgetrf - #define PADDLE_DGETRF LAPACKE_dgetrf - #define PADDLE_SGETRI LAPACKE_sgetri - #define PADDLE_DGETRI LAPACKE_dgetri -#endif +#define PADDLE_SGETRF LAPACKE_sgetrf +#define PADDLE_DGETRF LAPACKE_dgetrf +#define PADDLE_SGETRI LAPACKE_sgetri +#define PADDLE_DGETRI LAPACKE_dgetri #define LAPACK_ROUTINE_EACH(__macro) \ __macro(PADDLE_SGETRF) \ diff --git a/paddle/math/MathFunctions.h b/paddle/math/MathFunctions.h index f4cf6bd6c2c06f95cda098af389b37b7ff2983eb..f3d8b1a39e849d5f5a9e79cf33252b60170ced81 100644 --- a/paddle/math/MathFunctions.h +++ b/paddle/math/MathFunctions.h @@ -21,7 +21,7 @@ limitations under the License. */ #include #endif -#if defined(PADDLE_USE_ATLAS) || defined(PADDLE_USE_VECLIB) +#if defined(PADDLE_USE_VECLIB) extern "C" { #include #include diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh old mode 100644 new mode 100755 index 6be2bd8fad9e33cf4e1dcafdd6b8f39111bdbe88..f916295cd7bc762e2052553b321344845f504648 --- a/paddle/scripts/docker/build.sh +++ b/paddle/scripts/docker/build.sh @@ -35,8 +35,9 @@ function cmake_gen() { -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release} ${PYTHON_FLAGS} -DWITH_DSO=ON - -DWITH_DOC=OFF + -DWITH_DOC=${WITH_DOC:-OFF} -DWITH_GPU=${WITH_GPU:-OFF} + -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF} -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF} -DWITH_MKL=${WITH_MKL:-ON} -DWITH_AVX=${WITH_AVX:-OFF} @@ -50,7 +51,9 @@ function cmake_gen() { -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} -DWITH_TESTING=${WITH_TESTING:-ON} -DWITH_FAST_BUNDLE_TEST=ON + -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON + -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} ======================================== EOF # Disable UNITTEST_USE_VIRTUALENV in docker because @@ -60,8 +63,9 @@ EOF -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE:-Release} \ ${PYTHON_FLAGS} \ -DWITH_DSO=ON \ - -DWITH_DOC=OFF \ + -DWITH_DOC=${WITH_DOC:-OFF} \ -DWITH_GPU=${WITH_GPU:-OFF} \ + -DWITH_AMD_GPU=${WITH_AMD_GPU:-OFF} \ -DWITH_DISTRIBUTE=${WITH_DISTRIBUTE:-OFF} \ -DWITH_MKL=${WITH_MKL:-ON} \ -DWITH_AVX=${WITH_AVX:-OFF} \ @@ -74,6 +78,8 @@ EOF -DWITH_STYLE_CHECK=${WITH_STYLE_CHECK:-ON} \ -DWITH_TESTING=${WITH_TESTING:-ON} \ -DWITH_FAST_BUNDLE_TEST=ON \ + -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \ + -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON } @@ -119,9 +125,8 @@ EOF -DWITH_AVX=${WITH_AVX:-ON} \ -DWITH_SWIG_PY=ON \ -DWITH_STYLE_CHECK=OFF - make -j `nproc` gen_proto_py framework_py_proto - make -j `nproc` copy_paddle_pybind - make -j `nproc` paddle_docs paddle_docs_cn paddle_api_docs + + make -j `nproc` paddle_docs paddle_apis popd fi @@ -231,7 +236,7 @@ gen_capi_package gen_fluid_inference_lib if [[ ${WITH_C_API:-OFF} == "ON" ]]; then - printf "PaddlePaddle C-API libraries was generated on build/paddle.tgz\n" + printf "PaddlePaddle C-API libraries was generated on build/paddle.tgz\n" else printf "If you need to install PaddlePaddle in develop docker image," printf "please make install or pip install build/python/dist/*.whl.\n" diff --git a/paddle/scripts/submit_local.sh.in b/paddle/scripts/submit_local.sh.in index 80fa0c72af65cbdc21ba955389318a233e02657c..1283de9d957a46b848c7bb6caf9c5f49398468e2 100755 --- a/paddle/scripts/submit_local.sh.in +++ b/paddle/scripts/submit_local.sh.in @@ -153,9 +153,15 @@ if [ $? -ne 0 ]; then exit 1 fi -INSTALLED_VERSION=`pip freeze 2>/dev/null | grep '^paddle' | sed 's/.*==//g'` +if [ "@WITH_GPU@" == "ON" ]; then + PADDLE_NAME="paddlepaddle-gpu" +else + PADDLE_NAME="paddlepaddle" +fi + +INSTALLED_VERSION=`pip freeze 2>/dev/null | grep "^${PADDLE_NAME}==" | sed 's/.*==//g'` -if [ -z ${INSTALLED_VERSION} ]; then +if [ -z "${INSTALLED_VERSION}" ]; then INSTALLED_VERSION="0.0.0" # not installed fi cat <'] @@ -145,7 +145,7 @@ def test(word_idx, n, data_type=DataType.NGRAM): def fetch(): - paddle.v2.dataset.common.download(URL, "imikolov", MD5) + paddle.dataset.common.download(URL, "imikolov", MD5) def convert(path): @@ -154,8 +154,7 @@ def convert(path): """ N = 5 word_dict = build_dict() - paddle.v2.dataset.common.convert(path, - train(word_dict, N), 1000, - "imikolov_train") - paddle.v2.dataset.common.convert(path, - test(word_dict, N), 1000, "imikolov_test") + paddle.dataset.common.convert(path, + train(word_dict, N), 1000, "imikolov_train") + paddle.dataset.common.convert(path, + test(word_dict, N), 1000, "imikolov_test") diff --git a/python/paddle/v2/dataset/mnist.py b/python/paddle/dataset/mnist.py similarity index 76% rename from python/paddle/v2/dataset/mnist.py rename to python/paddle/dataset/mnist.py index 9f675bed895223e054cd3bb6e504fe1607f19858..6a1b8b5fac223c0d134cae69a61a0c2c00bc1feb 100644 --- a/python/paddle/v2/dataset/mnist.py +++ b/python/paddle/dataset/mnist.py @@ -17,7 +17,7 @@ MNIST dataset. This module will download dataset from http://yann.lecun.com/exdb/mnist/ and parse training set and test set into paddle reader creators. """ -import paddle.v2.dataset.common +import paddle.dataset.common import subprocess import numpy import platform @@ -85,10 +85,10 @@ def train(): :rtype: callable """ return reader_creator( - paddle.v2.dataset.common.download(TRAIN_IMAGE_URL, 'mnist', - TRAIN_IMAGE_MD5), - paddle.v2.dataset.common.download(TRAIN_LABEL_URL, 'mnist', - TRAIN_LABEL_MD5), 100) + paddle.dataset.common.download(TRAIN_IMAGE_URL, 'mnist', + TRAIN_IMAGE_MD5), + paddle.dataset.common.download(TRAIN_LABEL_URL, 'mnist', + TRAIN_LABEL_MD5), 100) def test(): @@ -102,22 +102,21 @@ def test(): :rtype: callable """ return reader_creator( - paddle.v2.dataset.common.download(TEST_IMAGE_URL, 'mnist', - TEST_IMAGE_MD5), - paddle.v2.dataset.common.download(TEST_LABEL_URL, 'mnist', - TEST_LABEL_MD5), 100) + paddle.dataset.common.download(TEST_IMAGE_URL, 'mnist', TEST_IMAGE_MD5), + paddle.dataset.common.download(TEST_LABEL_URL, 'mnist', TEST_LABEL_MD5), + 100) def fetch(): - paddle.v2.dataset.common.download(TRAIN_IMAGE_URL, 'mnist', TRAIN_IMAGE_MD5) - paddle.v2.dataset.common.download(TRAIN_LABEL_URL, 'mnist', TRAIN_LABEL_MD5) - paddle.v2.dataset.common.download(TEST_IMAGE_URL, 'mnist', TEST_IMAGE_MD5) - paddle.v2.dataset.common.download(TEST_LABEL_URL, 'mnist', TRAIN_LABEL_MD5) + paddle.dataset.common.download(TRAIN_IMAGE_URL, 'mnist', TRAIN_IMAGE_MD5) + paddle.dataset.common.download(TRAIN_LABEL_URL, 'mnist', TRAIN_LABEL_MD5) + paddle.dataset.common.download(TEST_IMAGE_URL, 'mnist', TEST_IMAGE_MD5) + paddle.dataset.common.download(TEST_LABEL_URL, 'mnist', TRAIN_LABEL_MD5) def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train(), 1000, "minist_train") - paddle.v2.dataset.common.convert(path, test(), 1000, "minist_test") + paddle.dataset.common.convert(path, train(), 1000, "minist_train") + paddle.dataset.common.convert(path, test(), 1000, "minist_test") diff --git a/python/paddle/v2/dataset/movielens.py b/python/paddle/dataset/movielens.py similarity index 95% rename from python/paddle/v2/dataset/movielens.py rename to python/paddle/dataset/movielens.py index 5b61a9420af1bb81e1d826f8a7b69f34c306d382..ab11716202a8298c182e23b661eb1d2ac74bf4da 100644 --- a/python/paddle/v2/dataset/movielens.py +++ b/python/paddle/dataset/movielens.py @@ -23,7 +23,7 @@ set and test set into paddle reader creators. """ import zipfile -import paddle.v2.dataset.common +import paddle.dataset.common import re import random import functools @@ -100,7 +100,7 @@ USER_INFO = None def __initialize_meta_info__(): - fn = paddle.v2.dataset.common.download(URL, "movielens", MD5) + fn = paddle.dataset.common.download(URL, "movielens", MD5) global MOVIE_INFO if MOVIE_INFO is None: pattern = re.compile(r'^(.*)\((\d+)\)$') @@ -247,15 +247,15 @@ def unittest(): def fetch(): - paddle.v2.dataset.common.download(URL, "movielens", MD5) + paddle.dataset.common.download(URL, "movielens", MD5) def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train(), 1000, "movielens_train") - paddle.v2.dataset.common.convert(path, test(), 1000, "movielens_test") + paddle.dataset.common.convert(path, train(), 1000, "movielens_train") + paddle.dataset.common.convert(path, test(), 1000, "movielens_test") if __name__ == '__main__': diff --git a/python/paddle/v2/dataset/mq2007.py b/python/paddle/dataset/mq2007.py similarity index 100% rename from python/paddle/v2/dataset/mq2007.py rename to python/paddle/dataset/mq2007.py diff --git a/python/paddle/v2/dataset/sentiment.py b/python/paddle/dataset/sentiment.py similarity index 87% rename from python/paddle/v2/dataset/sentiment.py rename to python/paddle/dataset/sentiment.py index b0b9757c1a75d215cf8945b5cedbb1239fd43af7..f5461164fe6b816356e42fc7b7dcf388eccfadfb 100644 --- a/python/paddle/v2/dataset/sentiment.py +++ b/python/paddle/dataset/sentiment.py @@ -26,7 +26,7 @@ from itertools import chain import nltk from nltk.corpus import movie_reviews -import paddle.v2.dataset.common +import paddle.dataset.common __all__ = ['train', 'test', 'get_word_dict', 'convert'] NUM_TRAINING_INSTANCES = 1600 @@ -39,13 +39,13 @@ def download_data_if_not_yet(): """ try: # make sure that nltk can find the data - if paddle.v2.dataset.common.DATA_HOME not in nltk.data.path: - nltk.data.path.append(paddle.v2.dataset.common.DATA_HOME) + if paddle.dataset.common.DATA_HOME not in nltk.data.path: + nltk.data.path.append(paddle.dataset.common.DATA_HOME) movie_reviews.categories() except LookupError: print "Downloading movie_reviews data set, please wait....." nltk.download( - 'movie_reviews', download_dir=paddle.v2.dataset.common.DATA_HOME) + 'movie_reviews', download_dir=paddle.dataset.common.DATA_HOME) print "Download data set success....." print "Path is " + nltk.data.find('corpora/movie_reviews').path @@ -129,13 +129,12 @@ def test(): def fetch(): - nltk.download( - 'movie_reviews', download_dir=paddle.v2.dataset.common.DATA_HOME) + nltk.download('movie_reviews', download_dir=paddle.dataset.common.DATA_HOME) def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train, 1000, "sentiment_train") - paddle.v2.dataset.common.convert(path, test, 1000, "sentiment_test") + paddle.dataset.common.convert(path, train, 1000, "sentiment_train") + paddle.dataset.common.convert(path, test, 1000, "sentiment_test") diff --git a/python/paddle/dataset/tests/CMakeLists.txt b/python/paddle/dataset/tests/CMakeLists.txt new file mode 100644 index 0000000000000000000000000000000000000000..485c38a13b573664d8033c237272a10ebb7c9701 --- /dev/null +++ b/python/paddle/dataset/tests/CMakeLists.txt @@ -0,0 +1 @@ +py_test(test_image SRCS test_image.py) diff --git a/python/paddle/v2/tests/cat.jpg b/python/paddle/dataset/tests/cat.jpg similarity index 100% rename from python/paddle/v2/tests/cat.jpg rename to python/paddle/dataset/tests/cat.jpg diff --git a/python/paddle/v2/dataset/tests/cifar_test.py b/python/paddle/dataset/tests/cifar_test.py similarity index 88% rename from python/paddle/v2/dataset/tests/cifar_test.py rename to python/paddle/dataset/tests/cifar_test.py index e0e18229da7818be5752ee592e094a00da286ad9..839125b09dd5c6432e3572374a7345a77a43f7cf 100644 --- a/python/paddle/v2/dataset/tests/cifar_test.py +++ b/python/paddle/dataset/tests/cifar_test.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.cifar +import paddle.dataset.cifar import unittest @@ -29,25 +29,25 @@ class TestCIFAR(unittest.TestCase): def test_test10(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.cifar.test10()) + paddle.dataset.cifar.test10()) self.assertEqual(instances, 10000) self.assertEqual(max_label_value, 9) def test_train10(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.cifar.train10()) + paddle.dataset.cifar.train10()) self.assertEqual(instances, 50000) self.assertEqual(max_label_value, 9) def test_test100(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.cifar.test100()) + paddle.dataset.cifar.test100()) self.assertEqual(instances, 10000) self.assertEqual(max_label_value, 99) def test_train100(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.cifar.train100()) + paddle.dataset.cifar.train100()) self.assertEqual(instances, 50000) self.assertEqual(max_label_value, 99) diff --git a/python/paddle/v2/dataset/tests/common_test.py b/python/paddle/dataset/tests/common_test.py similarity index 81% rename from python/paddle/v2/dataset/tests/common_test.py rename to python/paddle/dataset/tests/common_test.py index cfa194eba38ea70311c4deeac2635dc0a0103576..e7cc02aa83061599ffefa18de6cb02ac0fc9e9b7 100644 --- a/python/paddle/v2/dataset/tests/common_test.py +++ b/python/paddle/dataset/tests/common_test.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.common +import paddle.dataset.common import unittest import tempfile import glob @@ -24,14 +24,14 @@ class TestCommon(unittest.TestCase): with open(temp_path, 'w') as f: f.write("Hello\n") self.assertEqual('09f7e02f1290be211da707a266f153b3', - paddle.v2.dataset.common.md5file(temp_path)) + paddle.dataset.common.md5file(temp_path)) def test_download(self): yi_avatar = 'https://avatars0.githubusercontent.com/u/1548775?v=3&s=460' self.assertEqual( - paddle.v2.dataset.common.DATA_HOME + '/test/1548775?v=3&s=460', - paddle.v2.dataset.common.download( - yi_avatar, 'test', 'f75287202d6622414c706c36c16f8e0d')) + paddle.dataset.common.DATA_HOME + '/test/1548775?v=3&s=460', + paddle.dataset.common.download(yi_avatar, 'test', + 'f75287202d6622414c706c36c16f8e0d')) def test_split(self): def test_reader(): @@ -42,7 +42,7 @@ class TestCommon(unittest.TestCase): return reader _, temp_path = tempfile.mkstemp() - paddle.v2.dataset.common.split( + paddle.dataset.common.split( test_reader(), 4, suffix=temp_path + '/test-%05d.pickle') files = glob.glob(temp_path + '/test-%05d.pickle') self.assertEqual(len(files), 3) @@ -52,7 +52,7 @@ class TestCommon(unittest.TestCase): for x in xrange(5): with open(temp_path + '/%05d.test' % x) as f: f.write('%d\n' % x) - reader = paddle.v2.dataset.common.cluster_files_reader( + reader = paddle.dataset.common.cluster_files_reader( temp_path + '/*.test', 5, 0) for idx, e in enumerate(reader()): self.assertEqual(e, str("0")) @@ -69,9 +69,9 @@ class TestCommon(unittest.TestCase): return reader path = tempfile.mkdtemp() - paddle.v2.dataset.common.convert(path, - test_reader(), num_shards, - 'random_images') + paddle.dataset.common.convert(path, + test_reader(), num_shards, + 'random_images') files = glob.glob(path + '/random_images-*') self.assertEqual(len(files), num_shards) diff --git a/python/paddle/v2/dataset/tests/flowers_test.py b/python/paddle/dataset/tests/flowers_test.py similarity index 89% rename from python/paddle/v2/dataset/tests/flowers_test.py rename to python/paddle/dataset/tests/flowers_test.py index a8ae9a07acc22eb9d3c0cc5ebb07f8f11ed21233..06260fd796ce0271b7cec2f42a8a5a255a02dc24 100644 --- a/python/paddle/v2/dataset/tests/flowers_test.py +++ b/python/paddle/dataset/tests/flowers_test.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.flowers +import paddle.dataset.flowers import unittest @@ -30,19 +30,19 @@ class TestFlowers(unittest.TestCase): def test_train(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.flowers.train()) + paddle.dataset.flowers.train()) self.assertEqual(instances, 6149) self.assertEqual(max_label_value, 102) def test_test(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.flowers.test()) + paddle.dataset.flowers.test()) self.assertEqual(instances, 1020) self.assertEqual(max_label_value, 102) def test_valid(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.flowers.valid()) + paddle.dataset.flowers.valid()) self.assertEqual(instances, 1020) self.assertEqual(max_label_value, 102) diff --git a/python/paddle/v2/dataset/tests/imdb_test.py b/python/paddle/dataset/tests/imdb_test.py similarity index 77% rename from python/paddle/v2/dataset/tests/imdb_test.py rename to python/paddle/dataset/tests/imdb_test.py index c4d82f26895d77d05c6e936bd636b1239e1a0cd8..539da049449cd273db0a9e260851ed40e1be0f04 100644 --- a/python/paddle/v2/dataset/tests/imdb_test.py +++ b/python/paddle/dataset/tests/imdb_test.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.imdb +import paddle.dataset.imdb import unittest import re @@ -30,15 +30,13 @@ class TestIMDB(unittest.TestCase): def test_build_dict(self): if self.word_idx == None: - self.word_idx = paddle.v2.dataset.imdb.build_dict(TRAIN_PATTERN, - 150) + self.word_idx = paddle.dataset.imdb.build_dict(TRAIN_PATTERN, 150) self.assertEqual(len(self.word_idx), 7036) def check_dataset(self, dataset, expected_size): if self.word_idx == None: - self.word_idx = paddle.v2.dataset.imdb.build_dict(TRAIN_PATTERN, - 150) + self.word_idx = paddle.dataset.imdb.build_dict(TRAIN_PATTERN, 150) sum = 0 for l in dataset(self.word_idx): @@ -47,10 +45,10 @@ class TestIMDB(unittest.TestCase): self.assertEqual(sum, expected_size) def test_train(self): - self.check_dataset(paddle.v2.dataset.imdb.train, 25000) + self.check_dataset(paddle.dataset.imdb.train, 25000) def test_test(self): - self.check_dataset(paddle.v2.dataset.imdb.test, 25000) + self.check_dataset(paddle.dataset.imdb.test, 25000) if __name__ == '__main__': diff --git a/python/paddle/v2/dataset/tests/imikolov_test.py b/python/paddle/dataset/tests/imikolov_test.py similarity index 79% rename from python/paddle/v2/dataset/tests/imikolov_test.py rename to python/paddle/dataset/tests/imikolov_test.py index 714a75d6f1ff31697eec2d893d350a726d6390fe..233fd9fc8cea4cd0b5cd052580030fc8c993693c 100644 --- a/python/paddle/v2/dataset/tests/imikolov_test.py +++ b/python/paddle/dataset/tests/imikolov_test.py @@ -12,10 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.imikolov +import paddle.dataset.imikolov import unittest -WORD_DICT = paddle.v2.dataset.imikolov.build_dict() +WORD_DICT = paddle.dataset.imikolov.build_dict() class TestMikolov(unittest.TestCase): @@ -25,7 +25,7 @@ class TestMikolov(unittest.TestCase): def test_train(self): n = 5 - self.check_reader(paddle.v2.dataset.imikolov.train(WORD_DICT, n), n) + self.check_reader(paddle.dataset.imikolov.train(WORD_DICT, n), n) first_line = 'aer banknote berlitz calloway centrust cluett fromstein '\ 'gitano guterman hydro-quebec ipo kia memotec mlx nahb punts '\ @@ -34,16 +34,16 @@ class TestMikolov(unittest.TestCase): WORD_DICT.get(ch, WORD_DICT['']) for ch in first_line.split(' ') ] - for l in paddle.v2.dataset.imikolov.train( + for l in paddle.dataset.imikolov.train( WORD_DICT, n=-1, - data_type=paddle.v2.dataset.imikolov.DataType.SEQ)(): + data_type=paddle.dataset.imikolov.DataType.SEQ)(): read_line = l[0][1:] break self.assertEqual(first_line, read_line) def test_test(self): n = 5 - self.check_reader(paddle.v2.dataset.imikolov.test(WORD_DICT, n), n) + self.check_reader(paddle.dataset.imikolov.test(WORD_DICT, n), n) first_line = 'consumers may want to move their telephones a little '\ 'closer to the tv set' @@ -51,9 +51,9 @@ class TestMikolov(unittest.TestCase): WORD_DICT.get(ch, WORD_DICT['']) for ch in first_line.split(' ') ] - for l in paddle.v2.dataset.imikolov.test( + for l in paddle.dataset.imikolov.test( WORD_DICT, n=-1, - data_type=paddle.v2.dataset.imikolov.DataType.SEQ)(): + data_type=paddle.dataset.imikolov.DataType.SEQ)(): read_line = l[0][1:] break self.assertEqual(first_line, read_line) diff --git a/python/paddle/v2/dataset/tests/mnist_test.py b/python/paddle/dataset/tests/mnist_test.py similarity index 91% rename from python/paddle/v2/dataset/tests/mnist_test.py rename to python/paddle/dataset/tests/mnist_test.py index 1d344cac3e7483a351033570fbec75a4d19f4a55..8ada19d3f2ee13e194d08e19a4b86b558c69a0a7 100644 --- a/python/paddle/v2/dataset/tests/mnist_test.py +++ b/python/paddle/dataset/tests/mnist_test.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.mnist +import paddle.dataset.mnist import unittest @@ -29,13 +29,13 @@ class TestMNIST(unittest.TestCase): def test_train(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.mnist.train()) + paddle.dataset.mnist.train()) self.assertEqual(instances, 60000) self.assertEqual(max_label_value, 9) def test_test(self): instances, max_label_value = self.check_reader( - paddle.v2.dataset.mnist.test()) + paddle.dataset.mnist.test()) self.assertEqual(instances, 10000) self.assertEqual(max_label_value, 9) diff --git a/python/paddle/v2/dataset/tests/mq2007_test.py b/python/paddle/dataset/tests/mq2007_test.py similarity index 85% rename from python/paddle/v2/dataset/tests/mq2007_test.py rename to python/paddle/dataset/tests/mq2007_test.py index 59847b6c18eadb12123cae824e8bce1051a69d4c..fba388724a8e84591df7150b41f8ea39a850fc31 100644 --- a/python/paddle/v2/dataset/tests/mq2007_test.py +++ b/python/paddle/dataset/tests/mq2007_test.py @@ -12,19 +12,19 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.mq2007 +import paddle.dataset.mq2007 import unittest class TestMQ2007(unittest.TestCase): def test_pairwise(self): - for label, query_left, query_right in paddle.v2.dataset.mq2007.test( + for label, query_left, query_right in paddle.dataset.mq2007.test( format="pairwise"): self.assertEqual(query_left.shape(), (46, )) self.assertEqual(query_right.shape(), (46, )) def test_listwise(self): - for label_array, query_array in paddle.v2.dataset.mq2007.test( + for label_array, query_array in paddle.dataset.mq2007.test( format="listwise"): self.assertEqual(len(label_array), len(query_array)) diff --git a/python/paddle/v2/tests/test_image.py b/python/paddle/dataset/tests/test_image.py similarity index 97% rename from python/paddle/v2/tests/test_image.py rename to python/paddle/dataset/tests/test_image.py index c78bbdc40a25878b21ba7e678afedf9d8f0a87cf..8bd56607ae1998935a3b3aaa0e3279515c2a540c 100644 --- a/python/paddle/v2/tests/test_image.py +++ b/python/paddle/dataset/tests/test_image.py @@ -15,7 +15,7 @@ import unittest import numpy as np -import paddle.v2.image as image +import paddle.dataset.image as image class Image(unittest.TestCase): diff --git a/python/paddle/v2/dataset/tests/test_sentiment.py b/python/paddle/dataset/tests/test_sentiment.py similarity index 97% rename from python/paddle/v2/dataset/tests/test_sentiment.py rename to python/paddle/dataset/tests/test_sentiment.py index 407405290734609059c1767600748d530e8a13a6..543f4b7378b583ea3857bf785cf330c43e535c2a 100644 --- a/python/paddle/v2/dataset/tests/test_sentiment.py +++ b/python/paddle/dataset/tests/test_sentiment.py @@ -17,7 +17,7 @@ import unittest import nltk -import paddle.v2.dataset.sentiment as st +import paddle.dataset.sentiment as st from nltk.corpus import movie_reviews diff --git a/python/paddle/v2/dataset/tests/voc2012_test.py b/python/paddle/dataset/tests/voc2012_test.py similarity index 82% rename from python/paddle/v2/dataset/tests/voc2012_test.py rename to python/paddle/dataset/tests/voc2012_test.py index 31e72ebf5eac0508d12783f9ceaa6eef0fa6d353..0d285461a8ae8a9cc69fbec0dcf5efc106b594f0 100644 --- a/python/paddle/v2/dataset/tests/voc2012_test.py +++ b/python/paddle/dataset/tests/voc2012_test.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.voc2012 +import paddle.dataset.voc2012 import unittest @@ -26,15 +26,15 @@ class TestVOC(unittest.TestCase): return sum def test_train(self): - count = self.check_reader(paddle.v2.dataset.voc_seg.train()) + count = self.check_reader(paddle.dataset.voc_seg.train()) self.assertEqual(count, 2913) def test_test(self): - count = self.check_reader(paddle.v2.dataset.voc_seg.test()) + count = self.check_reader(paddle.dataset.voc_seg.test()) self.assertEqual(count, 1464) def test_val(self): - count = self.check_reader(paddle.v2.dataset.voc_seg.val()) + count = self.check_reader(paddle.dataset.voc_seg.val()) self.assertEqual(count, 1449) diff --git a/python/paddle/v2/dataset/tests/wmt16_test.py b/python/paddle/dataset/tests/wmt16_test.py similarity index 89% rename from python/paddle/v2/dataset/tests/wmt16_test.py rename to python/paddle/dataset/tests/wmt16_test.py index cef6c3216e7de8d9785a063976e63f88d90b24df..8b949d8bf5212d51016a33da322095bde2038200 100644 --- a/python/paddle/v2/dataset/tests/wmt16_test.py +++ b/python/paddle/dataset/tests/wmt16_test.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2.dataset.wmt16 +import paddle.dataset.wmt16 import unittest @@ -34,28 +34,28 @@ class TestWMT16(unittest.TestCase): def test_train(self): for idx, sample in enumerate( - paddle.v2.dataset.wmt16.train( + paddle.dataset.wmt16.train( src_dict_size=100000, trg_dict_size=100000)()): if idx >= 10: break self.checkout_one_sample(sample) def test_test(self): for idx, sample in enumerate( - paddle.v2.dataset.wmt16.test( + paddle.dataset.wmt16.test( src_dict_size=1000, trg_dict_size=1000)()): if idx >= 10: break self.checkout_one_sample(sample) def test_val(self): for idx, sample in enumerate( - paddle.v2.dataset.wmt16.validation( + paddle.dataset.wmt16.validation( src_dict_size=1000, trg_dict_size=1000)()): if idx >= 10: break self.checkout_one_sample(sample) def test_get_dict(self): dict_size = 1000 - word_dict = paddle.v2.dataset.wmt16.get_dict("en", dict_size, True) + word_dict = paddle.dataset.wmt16.get_dict("en", dict_size, True) self.assertEqual(len(word_dict), dict_size) self.assertEqual(word_dict[0], "") self.assertEqual(word_dict[1], "") diff --git a/python/paddle/v2/dataset/uci_housing.py b/python/paddle/dataset/uci_housing.py similarity index 82% rename from python/paddle/v2/dataset/uci_housing.py rename to python/paddle/dataset/uci_housing.py index f10bf7e42a1ead09b3eba0d61e55701215e4360f..6a56e9d5563c76ab6f524ccea9191693dc227010 100644 --- a/python/paddle/v2/dataset/uci_housing.py +++ b/python/paddle/dataset/uci_housing.py @@ -21,8 +21,7 @@ parse training set and test set into paddle reader creators. import numpy as np import os -import paddle.v2.dataset.common -from paddle.v2.parameters import Parameters +import paddle.dataset.common __all__ = ['train', 'test'] @@ -85,7 +84,7 @@ def train(): :rtype: callable """ global UCI_TRAIN_DATA - load_data(paddle.v2.dataset.common.download(URL, 'uci_housing', MD5)) + load_data(paddle.dataset.common.download(URL, 'uci_housing', MD5)) def reader(): for d in UCI_TRAIN_DATA: @@ -105,7 +104,7 @@ def test(): :rtype: callable """ global UCI_TEST_DATA - load_data(paddle.v2.dataset.common.download(URL, 'uci_housing', MD5)) + load_data(paddle.dataset.common.download(URL, 'uci_housing', MD5)) def reader(): for d in UCI_TEST_DATA: @@ -114,21 +113,13 @@ def test(): return reader -def model(): - tar_file = paddle.v2.dataset.common.download(URL_MODEL, 'fit_a_line.tar', - MD5_MODEL) - with open(tar_file, 'r') as f: - parameters = Parameters.from_tar(f) - return parameters - - def fetch(): - paddle.v2.dataset.common.download(URL, 'uci_housing', MD5) + paddle.dataset.common.download(URL, 'uci_housing', MD5) def convert(path): """ Converts dataset to recordio format """ - paddle.v2.dataset.common.convert(path, train(), 1000, "uci_housing_train") - paddle.v2.dataset.common.convert(path, test(), 1000, "uci_houseing_test") + paddle.dataset.common.convert(path, train(), 1000, "uci_housing_train") + paddle.dataset.common.convert(path, test(), 1000, "uci_houseing_test") diff --git a/python/paddle/v2/dataset/voc2012.py b/python/paddle/dataset/voc2012.py similarity index 97% rename from python/paddle/v2/dataset/voc2012.py rename to python/paddle/dataset/voc2012.py index 617e212d67fbe37f9d9663e9c83c62045411fa77..9c945574dbcc15f5cee370206ed7e70ba8ab5014 100644 --- a/python/paddle/v2/dataset/voc2012.py +++ b/python/paddle/dataset/voc2012.py @@ -22,8 +22,8 @@ with segmentation has been increased from 7,062 to 9,993. import tarfile import io import numpy as np -from paddle.v2.dataset.common import download -from paddle.v2.image import * +from paddle.dataset.common import download +from paddle.dataset.image import * from PIL import Image __all__ = ['train', 'test', 'val'] diff --git a/python/paddle/v2/dataset/wmt14.py b/python/paddle/dataset/wmt14.py similarity index 84% rename from python/paddle/v2/dataset/wmt14.py rename to python/paddle/dataset/wmt14.py index 5104e29051e4480f3a7eb18421f1b519841b009b..f0908c737874fa7335cca5b5f0cba83190c9f90f 100644 --- a/python/paddle/v2/dataset/wmt14.py +++ b/python/paddle/dataset/wmt14.py @@ -22,8 +22,7 @@ parse training set and test set into paddle reader creators. import tarfile import gzip -import paddle.v2.dataset.common -from paddle.v2.parameters import Parameters +import paddle.dataset.common __all__ = [ 'train', @@ -123,7 +122,7 @@ def train(dict_size): :rtype: callable """ return reader_creator( - paddle.v2.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN), + paddle.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN), 'train/train', dict_size) @@ -139,27 +138,20 @@ def test(dict_size): :rtype: callable """ return reader_creator( - paddle.v2.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN), + paddle.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN), 'test/test', dict_size) def gen(dict_size): return reader_creator( - paddle.v2.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN), + paddle.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN), 'gen/gen', dict_size) -def model(): - tar_file = paddle.v2.dataset.common.download(URL_MODEL, 'wmt14', MD5_MODEL) - with gzip.open(tar_file, 'r') as f: - parameters = Parameters.from_tar(f) - return parameters - - def get_dict(dict_size, reverse=True): # if reverse = False, return dict = {'a':'001', 'b':'002', ...} # else reverse = true, return dict = {'001':'a', '002':'b', ...} - tar_file = paddle.v2.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN) + tar_file = paddle.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN) src_dict, trg_dict = __read_to_dict(tar_file, dict_size) if reverse: src_dict = {v: k for k, v in src_dict.items()} @@ -168,8 +160,8 @@ def get_dict(dict_size, reverse=True): def fetch(): - paddle.v2.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN) - paddle.v2.dataset.common.download(URL_MODEL, 'wmt14', MD5_MODEL) + paddle.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN) + paddle.dataset.common.download(URL_MODEL, 'wmt14', MD5_MODEL) def convert(path): @@ -177,6 +169,5 @@ def convert(path): Converts dataset to recordio format """ dict_size = 30000 - paddle.v2.dataset.common.convert(path, - train(dict_size), 1000, "wmt14_train") - paddle.v2.dataset.common.convert(path, test(dict_size), 1000, "wmt14_test") + paddle.dataset.common.convert(path, train(dict_size), 1000, "wmt14_train") + paddle.dataset.common.convert(path, test(dict_size), 1000, "wmt14_test") diff --git a/python/paddle/v2/dataset/wmt16.py b/python/paddle/dataset/wmt16.py similarity index 94% rename from python/paddle/v2/dataset/wmt16.py rename to python/paddle/dataset/wmt16.py index c8818f715beadd9499ae588f2c19a57fbf26f372..ad23338a96df6856c7e15cb5e3bb713021a55bf0 100644 --- a/python/paddle/v2/dataset/wmt16.py +++ b/python/paddle/dataset/wmt16.py @@ -33,7 +33,7 @@ import tarfile import gzip from collections import defaultdict -import paddle.v2.dataset.common +import paddle.dataset.common __all__ = [ "train", @@ -76,7 +76,7 @@ def __build_dict(tar_file, dict_size, save_path, lang): def __load_dict(tar_file, dict_size, lang, reverse=False): - dict_path = os.path.join(paddle.v2.dataset.common.DATA_HOME, + dict_path = os.path.join(paddle.dataset.common.DATA_HOME, "wmt16/%s_%d.dict" % (lang, dict_size)) if not os.path.exists(dict_path) or ( len(open(dict_path, "r").readlines()) != dict_size): @@ -178,8 +178,8 @@ def train(src_dict_size, trg_dict_size, src_lang="en"): src_lang) return reader_creator( - tar_file=paddle.v2.dataset.common.download(DATA_URL, "wmt16", DATA_MD5, - "wmt16.tar.gz"), + tar_file=paddle.dataset.common.download(DATA_URL, "wmt16", DATA_MD5, + "wmt16.tar.gz"), file_name="wmt16/train", src_dict_size=src_dict_size, trg_dict_size=trg_dict_size, @@ -227,8 +227,8 @@ def test(src_dict_size, trg_dict_size, src_lang="en"): src_lang) return reader_creator( - tar_file=paddle.v2.dataset.common.download(DATA_URL, "wmt16", DATA_MD5, - "wmt16.tar.gz"), + tar_file=paddle.dataset.common.download(DATA_URL, "wmt16", DATA_MD5, + "wmt16.tar.gz"), file_name="wmt16/test", src_dict_size=src_dict_size, trg_dict_size=trg_dict_size, @@ -274,8 +274,8 @@ def validation(src_dict_size, trg_dict_size, src_lang="en"): src_lang) return reader_creator( - tar_file=paddle.v2.dataset.common.download(DATA_URL, "wmt16", DATA_MD5, - "wmt16.tar.gz"), + tar_file=paddle.dataset.common.download(DATA_URL, "wmt16", DATA_MD5, + "wmt16.tar.gz"), file_name="wmt16/val", src_dict_size=src_dict_size, trg_dict_size=trg_dict_size, @@ -303,12 +303,12 @@ def get_dict(lang, dict_size, reverse=False): if lang == "en": dict_size = min(dict_size, TOTAL_EN_WORDS) else: dict_size = min(dict_size, TOTAL_DE_WORDS) - dict_path = os.path.join(paddle.v2.dataset.common.DATA_HOME, + dict_path = os.path.join(paddle.dataset.common.DATA_HOME, "wmt16/%s_%d.dict" % (lang, dict_size)) assert os.path.exists(dict_path), "Word dictionary does not exist. " "Please invoke paddle.dataset.wmt16.train/test/validation first " "to build the dictionary." - tar_file = os.path.join(paddle.v2.dataset.common.DATA_HOME, "wmt16.tar.gz") + tar_file = os.path.join(paddle.dataset.common.DATA_HOME, "wmt16.tar.gz") return __load_dict(tar_file, dict_size, lang, reverse) @@ -323,7 +323,7 @@ def convert(path, src_dict_size, trg_dict_size, src_lang): """Converts dataset to recordio format. """ - paddle.v2.dataset.common.convert( + paddle.dataset.common.convert( path, train( src_dict_size=src_dict_size, @@ -331,7 +331,7 @@ def convert(path, src_dict_size, trg_dict_size, src_lang): src_lang=src_lang), 1000, "wmt16_train") - paddle.v2.dataset.common.convert( + paddle.dataset.common.convert( path, test( src_dict_size=src_dict_size, @@ -339,7 +339,7 @@ def convert(path, src_dict_size, trg_dict_size, src_lang): src_lang=src_lang), 1000, "wmt16_test") - paddle.v2.dataset.common.convert( + paddle.dataset.common.convert( path, validation( src_dict_size=src_dict_size, diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index fcea28220485039c9daf3c5fa2688c31f9f34c42..5ea4d977f4d8d9eb56b1fefa16f429df6e2a15bb 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -41,6 +41,7 @@ from memory_optimization_transpiler import memory_optimize, release_memory import profiler import unique_name import recordio_writer +from parallel_executor import ParallelExecutor Tensor = LoDTensor @@ -68,6 +69,7 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + [ 'profiler', 'unique_name', 'recordio_writer', + 'ParallelExecutor', ] diff --git a/python/paddle/fluid/concurrency.py b/python/paddle/fluid/concurrency.py index 3e4292d23550b853ea73de787a1c053e1f2c80fd..470dd0df524936a773f6e740c8079f0efa8ef7b4 100644 --- a/python/paddle/fluid/concurrency.py +++ b/python/paddle/fluid/concurrency.py @@ -12,7 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. -from layers.control_flow import BlockGuard, Select +from layers.control_flow import BlockGuard, equal +from .framework import Operator from layer_helper import LayerHelper, unique_name from layers import fill_constant import core @@ -75,6 +76,206 @@ class Go(BlockGuard): attrs={'sub_block': go_block}) +class SelectCase(object): + DEFAULT = 0 + SEND = 1 + RECEIVE = 2 + + def __init__(self, + select, + case_idx, + case_to_execute, + channel_action_fn=None, + channel=None, + value=None, + is_copy=False): + self.select = select + self.helper = LayerHelper('conditional_block') + self.main_program = self.helper.main_program + self.is_scalar_condition = True + + self.case_to_execute = case_to_execute + self.idx = case_idx + + # Since we aren't going to use the `channel_send` or `channel_recv` + # functions directly, we just need to capture the name. + self.action = (self.SEND + if channel_action_fn.__name__ == ('channel_send') else + self.RECEIVE) if channel_action_fn else self.DEFAULT + + X = value + if self.action == self.SEND and is_copy: + # We create of copy of the data we want to send + copied_X = self.select.parent_block.create_var( + name=unique_name.generate(value.name + '_copy'), + type=value.type, + dtype=value.dtype, + shape=value.shape, + lod_level=value.lod_level, + capacity=value.capacity + if hasattr(value, 'capacity') else None, ) + + self.select.parent_block.append_op( + type="assign", inputs={"X": value}, outputs={"Out": copied_X}) + X = copied_X + + self.value = X + self.channel = channel + + def __enter__(self): + self.block = self.main_program.create_block() + + def construct_op(self): + main_program = self.helper.main_program + cases_block = main_program.current_block() + + inner_outputs = set() + input_set = set() + params = set() + + for op in self.block.ops: + # Iterate over all operators, get all the inputs + # and add as input to the SelectCase operator. + for iname in op.input_names: + for in_var_name in op.input(iname): + if in_var_name not in inner_outputs: + input_set.add(in_var_name) + + for oname in op.output_names: + for out_var_name in op.output(oname): + inner_outputs.add(out_var_name) + + param_list = [ + cases_block.var(each_name) for each_name in params + if each_name not in input_set + ] + + # Iterate over all operators, get all the outputs + # add to the output list of SelectCase operator only if + # they exist in the parent block. + out_vars = [] + for inner_out_name in inner_outputs: + if inner_out_name in cases_block.vars: + out_vars.append(cases_block.var(inner_out_name)) + + # First, create an op that will determine whether or not this is the + # conditional variable to execute. + should_execute_block = equal( + fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=self.idx), + self.case_to_execute) + + step_scope = cases_block.create_var( + type=core.VarDesc.VarType.STEP_SCOPES) + + cases_block.append_op( + type='conditional_block', + inputs={'X': [should_execute_block], + 'Params': param_list}, + outputs={'Out': out_vars, + 'Scope': [step_scope]}, + attrs={ + 'sub_block': self.block, + 'is_scalar_condition': self.is_scalar_condition + }) + + return '%s,%s,%s,%s' % (self.idx, self.action, self.channel.name + if self.channel else '', self.value.name + if self.value else '') + + def __exit__(self, exc_type, exc_val, exc_tb): + self.main_program.rollback() + if exc_type is not None: + return False # re-raise exception + return True + + +class Select(BlockGuard): + def __init__(self, name=None): + self.helper = LayerHelper('select', name=name) + self.parent_block = self.helper.main_program.current_block() + self.cases = [] + + super(Select, self).__init__(self.helper.main_program) + self.case_to_execute = fill_constant( + shape=[1], dtype=core.VarDesc.VarType.INT32, value=-1) + + def __enter__(self): + super(Select, self).__enter__() + return self + + def case(self, channel_action_fn, channel, value, is_copy=False): + """Create a new block for this condition. + """ + select_case = SelectCase(self, + len(self.cases), self.case_to_execute, + channel_action_fn, channel, value, is_copy) + + self.cases.append(select_case) + + return select_case + + def default(self): + """Create a default case block for this condition. + """ + default_case = SelectCase(self, len(self.cases), self.case_to_execute) + + self.cases.append(default_case) + + return default_case + + def __exit__(self, exc_type, exc_val, exc_tb): + if exc_type is not None: + return False + + # Create a select op and another block to wrap its + # case blocks. + select_block = self.helper.main_program.current_block() + parent_block = self.helper.main_program.block(select_block.parent_idx) + + # Construct each case op, inside the newly created select block. + serialized_cases = [] + for case in self.cases: + serialized_cases.append(case.construct_op()) + + intermediate = set() + params = set() + + for case_block in select_block.ops: + if case_block.attrs and 'sub_block' in case_block.attrs: + for each_op in case_block.attrs['sub_block'].ops: + assert isinstance(each_op, Operator) + for iname in each_op.input_names: + for in_var_name in each_op.input(iname): + if in_var_name not in intermediate: + params.add(in_var_name) + + for oname in each_op.output_names: + for out_var_name in each_op.output(oname): + intermediate.add(out_var_name) + + out_list = [ + parent_block.var(var_name) for var_name in parent_block.vars + if var_name in intermediate + ] + + X = [select_block.var_recursive(x_name) for x_name in params] + + # Needs to be used by `equal` inside the cases block. + X.append(self.case_to_execute) + + # Construct the select op. + parent_block.append_op( + type='select', + inputs={'X': X, + 'case_to_execute': self.case_to_execute}, + attrs={'sub_block': select_block, + 'cases': serialized_cases}, + outputs={'Out': out_list}) + + return super(Select, self).__exit__(exc_type, exc_val, exc_tb) + + def make_channel(dtype, capacity=0): """ Helps implementation of a concurrent program by creating a "channel" of @@ -159,35 +360,26 @@ def channel_send(channel, value, is_copy=False): main_program = helper.main_program channel_send_block = main_program.current_block() - status = helper.create_variable( - name=unique_name.generate('status'), - type=core.VarDesc.VarType.LOD_TENSOR, - dtype=core.VarDesc.VarType.BOOL) - X = value - if is_copy is True: + if is_copy: copied_X = helper.create_variable( name=unique_name.generate(value.name + '_copy'), type=value.type, dtype=value.dtype, shape=value.shape, lod_level=value.lod_level, - capacity=value.capacity) + capacity=value.capacity if hasattr(value, 'capacity') else None) assign_op = channel_send_block.append_op( - type="assign_op", inputs={"X": value}, outputs={"Out": copied_X}) + type="assign", inputs={"X": value}, outputs={"Out": copied_X}) X = copied_X - channel_send_op = channel_send_block.append_op( - type="channel_send", - inputs={ + channel_send_block.append_op( + type="channel_send", inputs={ "Channel": channel, "X": X, - }, - outputs={"Status": status}) - - return status + }) def channel_recv(channel, return_value): diff --git a/python/paddle/fluid/debuger.py b/python/paddle/fluid/debuger.py index 97fa182c4007cc730c06e9f95259a2509e01ecdf..7b4afa9bf65e1369329cd4648c1f5c4bd8fa8357 100644 --- a/python/paddle/fluid/debuger.py +++ b/python/paddle/fluid/debuger.py @@ -16,7 +16,6 @@ import sys import re from graphviz import GraphPreviewGenerator import proto.framework_pb2 as framework_pb2 -import paddle.fluid.core as core _vartype2str_ = [ "UNK", @@ -126,7 +125,6 @@ def pprint_block_codes(block_desc, show_backward=False): def is_var_backward(var_desc): return "@GRAD" in var_desc.name - #print(type(block_desc)) if type(block_desc) is not framework_pb2.BlockDesc: block_desc = framework_pb2.BlockDesc.FromString( block_desc.serialize_to_string()) diff --git a/python/paddle/fluid/distribute_transpiler.py b/python/paddle/fluid/distribute_transpiler.py index ad655ee96cee0744e7bedb17163faf7d8d1d8877..9311fc9904eb730aa56e94a4e45a1479a67df641 100644 --- a/python/paddle/fluid/distribute_transpiler.py +++ b/python/paddle/fluid/distribute_transpiler.py @@ -20,6 +20,7 @@ from layer_helper import LayerHelper from distributed_spliter import * import math from . import core +import debuger class VarBlock: @@ -275,20 +276,26 @@ class DistributeTranspiler: suff_idx = v.name.find(".trainer_") if suff_idx >= 0: orig_var_name = v.name[:suff_idx] - pserver_program.global_block().create_var( + else: + orig_var_name = v.name + single_trainer_var = pserver_program.global_block().create_var( name=orig_var_name, persistable=True, type=v.type, dtype=v.dtype, shape=v.shape) - for trainer_id in xrange(self.trainers): - var = pserver_program.global_block().create_var( - name="%s.trainer_%d" % (orig_var_name, trainer_id), - persistable=False, - type=v.type, - dtype=v.dtype, - shape=v.shape) - recv_inputs.append(var) + if self.trainers > 1: + for trainer_id in xrange(self.trainers): + var = pserver_program.global_block().create_var( + name="%s.trainer_%d" % (orig_var_name, trainer_id), + persistable=False, + type=v.type, + dtype=v.dtype, + shape=v.shape) + recv_inputs.append(var) + else: + recv_inputs.append(single_trainer_var) + # step3 optimize_block = pserver_program.create_block(0) # step 4 @@ -336,15 +343,24 @@ class DistributeTranspiler: else: self._append_pserver_non_opt_ops(block, op) + append_block = optimize_block + # append lr decay ops to the child block if exits + lr_ops = self._get_lr_ops() + if len(lr_ops) > 0: + for _, op in enumerate(lr_ops): + self._append_pserver_non_opt_ops(append_block, op) + + append_block = pserver_program.create_block(append_block.idx) + # append op to the current block - per_opt_block = optimize_block + per_opt_block = append_block for _, opt_op in enumerate(opt_op_on_pserver): for _, op in enumerate(self.optimize_ops): # optimizer is connected to itself if ufind.is_connected(op, opt_op) and \ op not in global_ops: __append_optimize_op__(op, per_opt_block) - per_opt_block = pserver_program.create_block(0) + per_opt_block = pserver_program.create_block(append_block.idx) # append global ops for glb_op in global_ops: @@ -500,8 +516,11 @@ class DistributeTranspiler: def _append_split_op(self, program, gradblocks): # Split variables that need to be split and append respective ops + add_suffix = False + if self.trainers > 1: + add_suffix = True var_mapping = self._create_vars_from_blocklist( - program, gradblocks, add_trainer_suffix=True) + program, gradblocks, add_trainer_suffix=add_suffix) for varname, splited_vars in var_mapping.iteritems(): # variable that don't need to split have empty splited_vars if len(splited_vars) <= 1: @@ -563,6 +582,8 @@ class DistributeTranspiler: orig_var_name = "" if suff_idx >= 0: orig_var_name = varname[:suff_idx] + else: + orig_var_name = varname return orig_var_name def _append_pserver_ops(self, optimize_block, opt_op, endpoint, @@ -577,7 +598,8 @@ class DistributeTranspiler: grad_block = None for g in self.param_grad_ep_mapping[endpoint]["grads"]: if same_or_split_var( - self._orig_varname(g.name), opt_op.input(key)[0]): + self._orig_varname(g.name), + self._orig_varname(opt_op.input(key)[0])): grad_block = g break if not grad_block: @@ -748,7 +770,7 @@ class DistributeTranspiler: param_names = [ p.name for p in self.param_grad_ep_mapping[endpoint]["params"] ] - if op.input("Param") in param_names: + if op.input("Param")[0] in param_names: return True else: for n in param_names: @@ -781,3 +803,33 @@ class DistributeTranspiler: else: iomap[key] = vars return iomap + + def _get_lr_ops(self): + lr_ops = [] + # find learning rate variables by optimize op + lr_vars = set() + for op in self.optimize_ops: + if self._is_opt_op(op): + lr_vars.add(op.input("LearningRate")[0]) + + find_ops = [] + # find ops which output is lr var + block = self.program.global_block() + for op in block.ops: + if set(op.output_arg_names) & lr_vars: + find_ops.append(op) + # make a union find struct by the ops in default_main_program + ufind = UnionFind(block.ops) + for op1 in block.ops: + for op2 in block.ops: + # NOTE: we need to skip all optimize ops, since it is connected + # with forward/backward ops and lr ops, we only need the lr ops. + if op1 != op2 and self._is_op_connected(op1, op2) and \ + not self._is_opt_op(op1) and not self._is_opt_op(op2): + ufind.union(op1, op2) + # find all ops which is related with lr var + for op1 in block.ops: + for op2 in find_ops: + if ufind.is_connected(op1, op2): + lr_ops.append(op1) + return lr_ops diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index 2612fb1ae41986ae0d5c6e942cc3accebcb00e19..54d0a12bcdbb1b6c13e584dd1a3a5d73cddd4af7 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -48,8 +48,7 @@ def as_numpy(tensor): assert isinstance(tensor, core.LoDTensor) lod = tensor.lod() if len(lod) > 0: - raise RuntimeError( - "Some of your featched tensors hold LoD information. \ + raise RuntimeError("Some of your fetched tensors hold LoD information. \ They can not be completely cast to Python ndarray. \ Please set the parameter 'return_numpy' as 'False' to \ return LoDTensor itself directly.") @@ -180,60 +179,24 @@ def get_program_cache_key(feed, fetch_list): class Executor(object): - def __init__(self, places): - if not isinstance(places, list) and not isinstance(places, tuple): - places = [places] - - act_places = [] - for each in places: - p = core.Place() - p.set_place(each) - act_places.append(p) - - # TODO(dzhwinter) : only use the first place - self.executor = core.Executor(act_places[0]) - self.places = places + def __init__(self, place): + self.place = place + p = core.Place() + p.set_place(place) + self.executor = core.Executor(p) self.program_caches = dict() - def aslodtensor(self, data): - def accumulate(data): - if not isinstance(data, list): - return 1 - return sum([accumulate(sub) for sub in data]) - - def parselod(data): - seq_lens = [accumulate(seq) for seq in data] - cur_len = 0 - lod = [cur_len] - for l in seq_lens: - cur_len += l - lod.append(cur_len) - return lod - - assert len(self.places) != 0 - if not isinstance(data, list): - # pure tensor case - tensor = core.LoDTensor() - tensor.set(data, self.places[0]) - return tensor - else: - raise RuntimeError("Current implementation lacks unittests") - # lodtensor case - lod = [] - if not isinstance(data[0], list): - lod.append(parselod(data)) - flattened_data = np.concatenate(data, axis=0).astype("int64") - else: - while isinstance(data[0], list): - lod.append(parselod(seq)) - flattened_data = [item for seq in data for item in seq] - data = flattened_data - flattened_data = np.concatenate(data, axis=0).astype("int64") - flattened_data = flattened_data.reshape([len(flattened_data), 1]) - tensor = core.LoDTensor() - tensor.set(flattened_data, self.places[0]) - tensor.set_lod(lod) - return tensor + def as_lodtensor(self, data): + if isinstance(data, list): + raise RuntimeError("Some of your feed data hold LoD information. \ + They can not be completely cast from a list of Python \ + ndarray to LoDTensor. Please convert data to LoDTensor \ + directly before feeding the data.\ + ") + # single tensor case + tensor = core.LoDTensor() + tensor.set(data, self.place) + return tensor def _get_program_cache(self, program_cache_key): return self.program_caches.get(program_cache_key, None) @@ -293,7 +256,7 @@ class Executor(object): feed_target_name = op.desc.output('Out')[0] cur_feed = feed[feed_target_name] if not isinstance(cur_feed, core.LoDTensor): - cur_feed = self.aslodtensor(cur_feed) + cur_feed = self.as_lodtensor(cur_feed) idx = op.desc.attr('col') core.set_feed_variable(scope, cur_feed, feed_var_name, idx) else: diff --git a/python/paddle/fluid/layer_helper.py b/python/paddle/fluid/layer_helper.py index 58b668227168c5c5e080f3928035ad98303bbae9..d771837fc545167f7c32fcf914dd1c3c3ae64fb3 100644 --- a/python/paddle/fluid/layer_helper.py +++ b/python/paddle/fluid/layer_helper.py @@ -403,6 +403,8 @@ class LayerHelper(object): if 'use_mkldnn' in self.kwargs: act['use_mkldnn'] = self.kwargs.get('use_mkldnn') act_type = act.pop('type') + if 'use_mkldnn' in self.kwargs: + act['use_mkldnn'] = self.kwargs.get('use_mkldnn') 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 02cd0a05a11d8d1d52d42c2b62799f1093d0abc2..b9a53eda9144e9e56cf9bc626db40cf4225bd87f 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -16,8 +16,9 @@ import contextlib from layer_function_generator import autodoc from tensor import assign, fill_constant from .. import core -from ..framework import Program, Variable, Operator, Block +from ..framework import Program, Variable, Operator from ..layer_helper import LayerHelper, unique_name +from ..initializer import force_init_on_cpu from ops import logical_and, logical_not, logical_or __all__ = [ @@ -29,7 +30,6 @@ __all__ = [ 'WhileGuard', 'While', 'Switch', - 'Select', 'lod_rank_table', 'max_sequence_len', 'topk', @@ -950,7 +950,7 @@ def create_array(dtype): dtype=dtype) -def less_than(x, y, cond=None, **ignored): +def less_than(x, y, force_cpu=True, cond=None, **ignored): """ **Less than** @@ -959,6 +959,7 @@ def less_than(x, y, cond=None, **ignored): Args: x(Variable): First operand of *less_than* y(Variable): Second operand of *less_than* + force_cpu(Bool|True): The output data will be on CPU if set true. cond(Variable|None): Optional output variable to store the result of *less_than* Returns: @@ -975,8 +976,11 @@ def less_than(x, y, cond=None, **ignored): cond.stop_gradient = True helper.append_op( - type='less_than', inputs={'X': [x], - 'Y': [y]}, outputs={'Out': [cond]}) + type='less_than', + inputs={'X': [x], + 'Y': [y]}, + outputs={'Out': [cond]}, + attrs={'force_cpu': force_cpu or force_init_on_cpu()}) return cond @@ -1212,186 +1216,6 @@ class Switch(object): return True -class SelectCase(object): - DEFAULT = 0 - SEND = 1 - RECEIVE = 2 - - def __init__(self, - case_idx, - case_to_execute, - channel_action_fn=None, - channel=None, - value=None): - self.helper = LayerHelper('conditional_block') - self.main_program = self.helper.main_program - self.is_scalar_condition = True - - self.case_to_execute = case_to_execute - self.idx = case_idx - - # Since we aren't going to use the `channel_send` or `channel_recv` - # functions directly, we just need to capture the name. - self.action = (self.SEND - if channel_action_fn.__name__ == ('channel_send') else - self.RECEIVE) if channel_action_fn else (self.DEFAULT) - self.value = value - self.channel = channel - - def __enter__(self): - self.block = self.main_program.create_block() - - def construct_op(self): - main_program = self.helper.main_program - cases_block = main_program.current_block() - - inner_outputs = set() - input_set = set() - params = set() - - for op in self.block.ops: - # Iterate over all operators, get all the inputs - # and add as input to the SelectCase operator. - for iname in op.input_names: - for in_var_name in op.input(iname): - if in_var_name not in inner_outputs: - input_set.add(in_var_name) - - for oname in op.output_names: - for out_var_name in op.output(oname): - inner_outputs.add(out_var_name) - - param_list = [ - cases_block.var(each_name) for each_name in params - if each_name not in input_set - ] - - # Iterate over all operators, get all the outputs - # add to the output list of SelectCase operator only if - # they exist in the parent block. - out_vars = [] - for inner_out_name in inner_outputs: - if inner_out_name in cases_block.vars: - out_vars.append(cases_block.var(inner_out_name)) - - # First, create an op that will determine whether or not this is the - # conditional variable to execute. - should_execute_block = equal( - fill_constant( - shape=[1], dtype=core.VarDesc.VarType.INT32, value=self.idx), - self.case_to_execute) - - step_scope = cases_block.create_var( - type=core.VarDesc.VarType.STEP_SCOPES) - - cases_block.append_op( - type='conditional_block', - inputs={'X': [should_execute_block], - 'Params': param_list}, - outputs={'Out': out_vars, - 'Scope': [step_scope]}, - attrs={ - 'sub_block': self.block, - 'is_scalar_condition': self.is_scalar_condition - }) - - return '%s,%s,%s,%s' % (self.idx, self.action, self.channel.name - if self.channel else '', self.value.name - if self.value else '') - - def __exit__(self, exc_type, exc_val, exc_tb): - self.main_program.rollback() - if exc_type is not None: - return False # re-raise exception - return True - - -class Select(BlockGuard): - def __init__(self, name=None): - self.helper = LayerHelper('select', name=name) - self.cases = [] - - super(Select, self).__init__(self.helper.main_program) - self.case_to_execute = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.INT32, value=-1) - - def __enter__(self): - super(Select, self).__enter__() - return self - - def case(self, channel_action_fn, channel, value): - """Create a new block for this condition. - """ - select_case = SelectCase( - len(self.cases), self.case_to_execute, channel_action_fn, channel, - value) - - self.cases.append(select_case) - - return select_case - - def default(self): - """Create a default case block for this condition. - """ - default_case = SelectCase(len(self.cases), self.case_to_execute) - - self.cases.append(default_case) - - return default_case - - def __exit__(self, exc_type, exc_val, exc_tb): - if exc_type is not None: - return False - - # Create a select op and another block to wrap its - # case blocks. - select_block = self.helper.main_program.current_block() - parent_block = self.helper.main_program.block(select_block.parent_idx) - - # Construct each case op, inside the newly created select block. - serialized_cases = [] - for case in self.cases: - serialized_cases.append(case.construct_op()) - - intermediate = set() - params = set() - - for case_block in select_block.ops: - if case_block.attrs and 'sub_block' in case_block.attrs: - for each_op in case_block.attrs['sub_block'].ops: - assert isinstance(each_op, Operator) - for iname in each_op.input_names: - for in_var_name in each_op.input(iname): - if in_var_name not in intermediate: - params.add(in_var_name) - - for oname in each_op.output_names: - for out_var_name in each_op.output(oname): - intermediate.add(out_var_name) - - # TODO(varunarora): Figure out if defining output is needed. - out_list = [ - parent_block.var(var_name) for var_name in parent_block.vars - if var_name in intermediate - ] - - X = [select_block.var_recursive(x_name) for x_name in params] - - # Needs to be used by `equal` inside the cases block. - X.append(self.case_to_execute) - - # Construct the select op. - parent_block.append_op( - type='select', - inputs={'X': X, - 'case_to_execute': self.case_to_execute}, - attrs={'sub_block': select_block, - 'cases': serialized_cases}, - outputs={}) - - return super(Select, self).__exit__(exc_type, exc_val, exc_tb) - - class IfElseBlockGuard(object): def __init__(self, is_true, ifelse): if not isinstance(ifelse, IfElse): @@ -1577,7 +1401,8 @@ class DynamicRNN(object): type='less_than', inputs={'X': self.step_idx, 'Y': self.max_seq_len}, - outputs={'Out': self.cond}) + outputs={'Out': self.cond}, + attrs={'force_cpu': True}) input_array = parent_block.create_var( name=unique_name.generate('dynamic_rnn_input_array'), @@ -1626,7 +1451,11 @@ class DynamicRNN(object): for new_mem, mem_array in self.mem_link: array_write(x=new_mem, i=self.step_idx, array=mem_array) - less_than(x=self.step_idx, y=self.max_seq_len, cond=self.cond) + less_than( + x=self.step_idx, + y=self.max_seq_len, + force_cpu=True, + cond=self.cond) self.status = DynamicRNN.AFTER_RNN for each_array in self.output_array: diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index a889ab6bdc6ac9494ef992a97292b7a2536c41c4..a5938fe494265778ef7032c56a8d6d35acd729c5 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -19,7 +19,6 @@ from layer_function_generator import generate_layer_fn from layer_function_generator import autodoc from ..layer_helper import LayerHelper import tensor -import ops import nn import math @@ -58,7 +57,7 @@ def detection_output(loc, This operation is to get the detection results by performing following two steps: - + 1. Decode input bounding box predictions according to the prior boxes. 2. Get the final detection results by applying multi-class non maximum suppression (NMS). @@ -129,13 +128,12 @@ def detection_output(loc, prior_box_var=prior_box_var, target_box=loc, code_type='decode_center_size') - old_shape = scores.shape - scores = ops.reshape(x=scores, shape=(-1, old_shape[-1])) + scores = nn.reshape(x=scores, shape=(-1, old_shape[-1])) scores = nn.softmax(input=scores) - scores = ops.reshape(x=scores, shape=old_shape) + scores = nn.reshape(x=scores, shape=old_shape) scores = nn.transpose(scores, perm=[0, 2, 1]) - + scores.stop_gradient = True nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype) helper.append_op( type="multiclass_nms", @@ -150,6 +148,7 @@ def detection_output(loc, 'score_threshold': score_threshold, 'nms_eta': 1.0 }) + nmsed_outs.stop_gradient = True return nmsed_outs @@ -463,7 +462,7 @@ def ssd_loss(location, num, num_prior, num_class = confidence.shape def __reshape_to_2d(var): - return ops.reshape(x=var, shape=[-1, var.shape[-1]]) + return nn.reshape(x=var, shape=[-1, var.shape[-1]]) # 1. Find matched boundding box by prior box. # 1.1 Compute IOU similarity between ground-truth boxes and prior boxes. @@ -474,7 +473,8 @@ def ssd_loss(location, # 2. Compute confidence for mining hard examples # 2.1. Get the target label based on matched indices - gt_label = ops.reshape(x=gt_label, shape=gt_label.shape + (1, )) + gt_label = nn.reshape(x=gt_label, shape=gt_label.shape + (1, )) + gt_label.stop_gradient = True target_label, _ = target_assign( gt_label, matched_indices, mismatch_value=background_label) # 2.2. Compute confidence loss. @@ -482,10 +482,12 @@ def ssd_loss(location, confidence = __reshape_to_2d(confidence) target_label = tensor.cast(x=target_label, dtype='int64') target_label = __reshape_to_2d(target_label) + target_label.stop_gradient = True conf_loss = nn.softmax_with_cross_entropy(confidence, target_label) # 3. Mining hard examples - conf_loss = ops.reshape(x=conf_loss, shape=(num, num_prior)) + conf_loss = nn.reshape(x=conf_loss, shape=(num, num_prior)) + conf_loss.stop_gradient = True neg_indices = helper.create_tmp_variable(dtype='int32') dtype = matched_indices.dtype updated_matched_indices = helper.create_tmp_variable(dtype=dtype) @@ -553,7 +555,7 @@ def ssd_loss(location, # 5.3 Compute overall weighted loss. loss = conf_loss_weight * conf_loss + loc_loss_weight * loc_loss # reshape to [N, Np], N is the batch size and Np is the prior box number. - loss = ops.reshape(x=loss, shape=[-1, num_prior]) + loss = nn.reshape(x=loss, shape=[-1, num_prior]) loss = nn.reduce_sum(loss, dim=1, keep_dim=True) if normalize: normalizer = nn.reduce_sum(target_loc_weight) @@ -695,6 +697,8 @@ def multi_box_head(inputs, outputs={"Boxes": box, "Variances": var}, attrs=attrs, ) + box.stop_gradient = True + var.stop_gradient = True return box, var def _reshape_with_axis_(input, axis=1): @@ -704,7 +708,7 @@ def multi_box_head(inputs, new_shape = [ -1, reduce(lambda x, y: x * y, input.shape[axis:len(input.shape)]) ] - out = ops.reshape(x=input, shape=new_shape) + out = nn.reshape(x=input, shape=new_shape) return out def _is_list_or_tuple_(data): @@ -798,7 +802,7 @@ def multi_box_head(inputs, mbox_loc.shape[0], mbox_loc.shape[1] * mbox_loc.shape[2] * mbox_loc.shape[3] / 4, 4 ] - mbox_loc_flatten = ops.reshape(mbox_loc, shape=new_shape) + mbox_loc_flatten = nn.reshape(mbox_loc, shape=new_shape) mbox_locs.append(mbox_loc_flatten) # get conf @@ -814,7 +818,7 @@ def multi_box_head(inputs, conf_loc.shape[0], conf_loc.shape[1] * conf_loc.shape[2] * conf_loc.shape[3] / num_classes, num_classes ] - conf_loc_flatten = ops.reshape(conf_loc, shape=new_shape) + conf_loc_flatten = nn.reshape(conf_loc, shape=new_shape) mbox_confs.append(conf_loc_flatten) if len(box_results) == 1: @@ -834,4 +838,6 @@ def multi_box_head(inputs, mbox_locs_concat = tensor.concat(mbox_locs, axis=1) mbox_confs_concat = tensor.concat(mbox_confs, axis=1) + box.stop_gradient = True + var.stop_gradient = True return mbox_locs_concat, mbox_confs_concat, box, var diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index bc5e291ad811315ddc9d101853d69c7f5ab5082d..bd7e9c30fed2c38a206bf17a646d8a4433af4099 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -113,9 +113,9 @@ class ListenAndServ(object): which can receive variables from clients and run a block. """ - def __init__(self, endpoint, fan_in=1, optimizer_mode=True): + def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True): self.helper = LayerHelper("listen_and_serv") - self.inputs = [] + self.inputs = inputs self.outputs = [] self.endpoint = endpoint self.fan_in = fan_in @@ -160,18 +160,13 @@ class ListenAndServ(object): current_block = main_program.current_block() parent_block = self.parent_block() - params, grads = self.get_params_and_grads() - param_names = [p.name for p in params] - grad_names = [g.name for g in grads] parent_block.append_op( type='listen_and_serv', - inputs={}, + inputs={"X": self.inputs}, outputs={}, attrs={ 'endpoint': self.endpoint, 'Fanin': self.fan_in, - 'ParamList': param_names, - 'GradList': grad_names, 'OptimizeBlock': current_block }) @@ -196,10 +191,14 @@ def Send(endpoints, send_vars, get_vars): endpoints = list(set(epmap)) helper = LayerHelper("Send", **locals()) + rpc_client_var = default_main_program().global_block().create_var( + name="RPC_CLIENT_VAR", persistable=True, type=core.VarDesc.VarType.RAW) + helper.append_op( type="send", inputs={"X": send_vars}, - outputs={"Out": get_vars}, + outputs={"Out": get_vars, + "RPCClient": rpc_client_var}, attrs={"endpoints": endpoints, "epmap": epmap}) diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index 679de6ce2aa67abe1322702fcb371eded0130698..e59ee251202d4985f0aa9ce46408fc7f0d197e1d 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -73,7 +73,9 @@ __all__ = [ 'smooth_l1', 'one_hot', 'autoincreased_step_counter', + 'reshape', 'lod_reset', + 'lrn', ] @@ -1482,6 +1484,7 @@ def batch_norm(input, param_attr=None, bias_attr=None, data_layout='NCHW', + in_place=False, name=None, moving_mean_name=None, moving_variance_name=None): @@ -1537,7 +1540,7 @@ def batch_norm(input, saved_mean = helper.create_tmp_variable(dtype=dtype, stop_gradient=True) saved_variance = helper.create_tmp_variable(dtype=dtype, stop_gradient=True) - batch_norm_out = helper.create_tmp_variable(dtype) + batch_norm_out = input if in_place else helper.create_tmp_variable(dtype) helper.append_op( type="batch_norm", @@ -3263,6 +3266,8 @@ def one_hot(input, depth): The one-hot tensor or LodTensor, same as input. Examples: + .. code-block:: python + X is a LoDTensor: X.lod = [[0, 1, 4]] X.shape = [4, 1] @@ -3317,6 +3322,101 @@ 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): + """ + Gives a new shape to the input Tensor without changing its data. + + The target shape can be given by :attr:`shape` or :attr:`actual_shape`. + :attr:`shape` is a list of integer while :attr:`actual_shape` is a tensor + variable. :attr:`actual_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 compile-time. + + Some tricks exist when specifying the target shape. + + 1. -1 means the value of this dimension is inferred from the total element + number of x and remaining dimensions. Thus one and only one dimension can + be set -1. + + 2. 0 means the actual dimension value is going to be copied from the + corresponding dimension of x. The indice of 0s in shape can not exceed + Rank(X). + + Here are some examples to explain it. + + 1. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape + is [6, 8], the reshape operator will transform x into a 2-D tensor with + shape [6, 8] and leaving x's data unchanged. + + 2. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape + specified is [2, 3, -1, 2], the reshape operator will transform x into a + 4-D tensor with shape [2, 3, 4, 2] and leaving x's data unchanged. In this + case, one dimension of the target shape is set to -1, the value of this + dimension is inferred from the total element number of x and remaining + dimensions. + + 3. Given a 3-D tensor x with a shape [2, 4, 6], and the target shape + is [-1, 0, 3, 2], the reshape operator will transform x into a 4-D tensor + with shape [2, 4, 3, 2] and leaving x's data unchanged. In this case, + besides -1, 0 means the actual dimension value is going to be copied from + the corresponding dimension of x. + + Args: + input(variable): The input tensor. + shape(list): The new shape. At most one dimension of the new shape can + be -1. + actual_shape(variable): An optional input. If provided, reshape + according to this given shape rather than + :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, a new output tensor is created + whose data is copied from input x, otherwise the output + shares data with input without copying. + + Returns(variable): The output tensor. + + Examples: + .. code-block:: python + 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) + """ + + if not (isinstance(shape, list) or isinstance(shape, tuple)): + raise ValueError("Input shape must be a python lsit or tuple.") + + # Validate the shape + unk_dim_idx = -1 + for dim_idx, dim_size in enumerate(shape): + if dim_size == -1: + assert unk_dim_idx == -1, ( + "Only one dimension in shape can be unknown.") + unk_dim_idx = dim_idx + elif dim_size == 0: + assert dim_idx < len(x.shape), ( + "The indice of 0s in shape can not exceed Rank(X).") + else: + assert dim_size > 0, ( + "Each dimension size given in shape must not be negtive " + "except one unknown dimension.") + + helper = LayerHelper("reshape", **locals()) + reshaped = helper.create_tmp_variable(dtype=x.dtype) + helper.append_op( + type="reshape", + inputs={"X": x, + "Shape": actual_shape} + if isinstance(actual_shape, Variable) else {"X": x}, + attrs={"shape": shape, + "inplace": inplace}, + outputs={"Out": reshaped}) + + return helper.append_activation(reshaped) + + def lod_reset(x, y=None, target_lod=None): """ LoD Reset Operator. Set LoD of **x** to a new one specified by **y** or @@ -3410,3 +3510,73 @@ def lod_reset(x, y=None, target_lod=None): raise ValueError("y and target_lod should not be both None.") return out + + +def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None): + """ + Local Response Normalization Layer. This layer performs a type of + "lateral inhibition" by normalizing over local input regions. + + The formula is as follows: + + .. math:: + + Output(i, x, y) = Input(i, x, y) / \left( + k + \alpha \sum\limits^{\min(C, c + n/2)}_{j = \max(0, c - n/2)} + (Input(j, x, y))^2 \right)^{\beta} + + In the above equation: + + * :math:`n`: The number of channels to sum over. + * :math:`k`: The offset (avoid being divided by 0). + * :math:`alpha`: The scaling parameter. + * :math:`beta`: The exponent parameter. + + Refer to `ImageNet Classification with Deep Convolutional Neural Networks + `_ + + Args: + input (Variable): The input tensor of this layer, and the dimension of input tensor must be 4. + n (int, default 5): The number of channels to sum over. + k (float, default 1.0): An offset (usually positive to avoid dividing by 0). + alpha (float, default 1e-4): The scaling parameter. + beta (float, default 0.75): The exponent. + name (str, default None): A name for this operation. + + Raises: + ValueError: If rank of the input tensor is not 4. + + Returns: + A tensor variable storing the transformation result. + + Examples: + .. code-block:: python + + data = fluid.layers.data(name="data", shape=[3, 112, 112], dtype="float32") + lrn = fluid.layers.lrn(input=data) + """ + helper = LayerHelper('lrn', **locals()) + dtype = helper.input_dtype() + input_shape = input.shape + dims = len(input_shape) + + if dims != 4: + raise ValueError( + "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) + helper.append_op( + type="lrn", + inputs={"X": input}, + outputs={ + "Out": lrn_out, + "MidOut": mid_out, + }, + attrs={"n": n, + "k": k, + "alpha": alpha, + "beta": beta}) + + return lrn_out diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index f5c6b47d243dcf4ba985cfb41fc23b44d3ed809f..a9fe25744cc0b385479c9366af1b731ec221dd5a 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -25,6 +25,8 @@ __activations__ = [ 'abs', 'ceil', 'floor', + 'cos', + 'sin', 'round', 'reciprocal', 'log', @@ -47,7 +49,6 @@ __activations__ = [ __all__ = [ 'mean', 'mul', - 'reshape', 'scale', 'sigmoid_cross_entropy_with_logits', 'elementwise_add', diff --git a/python/paddle/fluid/nets.py b/python/paddle/fluid/nets.py index 3b2e1a3073251a6d6460450dc957e1b5c7a873c5..bbedf6fde0872fd32d81c103bf5fe61449b7f57b 100644 --- a/python/paddle/fluid/nets.py +++ b/python/paddle/fluid/nets.py @@ -98,7 +98,7 @@ def img_conv_group(input, use_mkldnn=use_mkldnn) if conv_with_batchnorm[i]: - tmp = layers.batch_norm(input=tmp, act=conv_act) + tmp = layers.batch_norm(input=tmp, act=conv_act, in_place=True) drop_rate = conv_batchnorm_drop_rate[i] if abs(drop_rate) > 1e-5: tmp = layers.dropout(x=tmp, dropout_prob=drop_rate) diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py new file mode 100644 index 0000000000000000000000000000000000000000..5e0588fa73241a8752e1b3195a123820165f070d --- /dev/null +++ b/python/paddle/fluid/parallel_executor.py @@ -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. + +import core +import multiprocessing +import framework +import executor + +__all__ = ['ParallelExecutor'] + + +class ParallelExecutor(object): + def __init__(self, loss_name, use_cuda, num_threads=None): + places = [] + if use_cuda: + for i in xrange(core.get_cuda_device_count()): + p = core.Place() + p.set_place(core.CUDAPlace(i)) + places.append(p) + else: + for i in xrange(multiprocessing.cpu_count()): + p = core.Place() + p.set_place(core.CPUPlace()) + places.append(p) + + if num_threads is None: + num_threads = min(len(places) * 2, multiprocessing.cpu_count()) + + startup = framework.default_startup_program() + main = framework.default_main_program() + scope = executor.global_scope() + + self.executor = core.ParallelExecutor( + num_threads, + True if use_cuda else False, # use_event + places, + set([ + p.name for p in main.global_block().iter_parameters() + if not p.stop_gradient + ]), + startup.desc, + main.desc, + loss_name, + scope) + self.scope = scope + + def run(self, fetch_list): + fetch_var_name = '@FETCHED_VAR_NAME@' + self.executor.run(fetch_list, fetch_var_name) + arr = self.scope.find_var(fetch_var_name).get_lod_tensor_array() + return [arr[i] for i in range(len(arr))] diff --git a/python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py b/python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py index 983f8f4dbeac83566839de25ec9765eb248be768..ce640dece8a5067bd10f410a2bb58874b7cc0908 100644 --- a/python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py +++ b/python/paddle/fluid/tests/book/notest_rnn_encoder_decoer.py @@ -13,7 +13,7 @@ # limitations under the License. import numpy as np -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import paddle.fluid.core as core import paddle.fluid.framework as framework diff --git a/python/paddle/fluid/tests/book/test_fit_a_line.py b/python/paddle/fluid/tests/book/test_fit_a_line.py index 93ef66851b0efd65361122853dadeefe11992ed5..6dfc2997ae0328a41fe22d13dfa8fc51d4d021a6 100644 --- a/python/paddle/fluid/tests/book/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/test_fit_a_line.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import contextlib import numpy diff --git a/python/paddle/fluid/tests/book/test_image_classification.py b/python/paddle/fluid/tests/book/test_image_classification.py index b01c1875d64d7fc14e0141672f7e8eab2b6a0394..e8bb082be196b6342b1719235f1264bbe3d776ac 100644 --- a/python/paddle/fluid/tests/book/test_image_classification.py +++ b/python/paddle/fluid/tests/book/test_image_classification.py @@ -14,7 +14,7 @@ from __future__ import print_function -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import contextlib import math 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 f488527e0bc69059bc44422aa28188441f3d5b54..c0a6df831acbfe2654a5941cf95c91343992ef13 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -15,8 +15,8 @@ import math import numpy as np -import paddle.v2 as paddle -import paddle.v2.dataset.conll05 as conll05 +import paddle +import paddle.dataset.conll05 as conll05 import paddle.fluid as fluid from paddle.fluid.initializer import init_on_cpu import contextlib diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index 3a1a0859ecfd4ac5337e2112f8b22e32d8474f22..830d78df8b9e56b45f7e928562ef4b89e88f696d 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -14,7 +14,7 @@ import contextlib import numpy as np -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import paddle.fluid.framework as framework import paddle.fluid.layers as pd diff --git a/python/paddle/fluid/tests/book/test_recognize_digits.py b/python/paddle/fluid/tests/book/test_recognize_digits.py index e85b97a7f430b6d752baa179f27a7d15bc4d9a81..e4997b4069f60ff4382b4254bc026ae8ae29b345 100644 --- a/python/paddle/fluid/tests/book/test_recognize_digits.py +++ b/python/paddle/fluid/tests/book/test_recognize_digits.py @@ -14,7 +14,7 @@ from __future__ import print_function import argparse import paddle.fluid as fluid -import paddle.v2 as paddle +import paddle import sys import numpy import unittest diff --git a/python/paddle/fluid/tests/book/test_recommender_system.py b/python/paddle/fluid/tests/book/test_recommender_system.py index 2ce66d32c993672793b0db213267d1f80b5c49dd..2172c275b8082689a6ff5f2c3c27a2ff4e92275a 100644 --- a/python/paddle/fluid/tests/book/test_recommender_system.py +++ b/python/paddle/fluid/tests/book/test_recommender_system.py @@ -16,7 +16,7 @@ import math import sys import os import numpy as np -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import paddle.fluid.framework as framework import paddle.fluid.layers as layers diff --git a/python/paddle/fluid/tests/book/test_understand_sentiment.py b/python/paddle/fluid/tests/book/test_understand_sentiment.py index d2f3f7404697feb0768f873070b97aeb3ba0cd64..dedd153778d7ad9caeb5fa7090a980bc7f177dea 100644 --- a/python/paddle/fluid/tests/book/test_understand_sentiment.py +++ b/python/paddle/fluid/tests/book/test_understand_sentiment.py @@ -15,7 +15,7 @@ from __future__ import print_function import unittest import paddle.fluid as fluid -import paddle.v2 as paddle +import paddle import contextlib import math import numpy as np diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py index 26b97c3e254f54b83515436660e44d4908c98fbe..8929779de9448d036e1528b64330b37463ab3988 100644 --- a/python/paddle/fluid/tests/book/test_word2vec.py +++ b/python/paddle/fluid/tests/book/test_word2vec.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import unittest import os diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py index ad79e96b958b36a06c8a3cc990dbe3608e32c9ac..8818cf96fa8f08036f9e23aae786f67b5614b2b9 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py @@ -13,7 +13,7 @@ # limitations under the License. import numpy as np -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import math import sys diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py index 204669d7e6176e9e8250e8aebc2d10441fa24b67..dfebb9a06ea4f290f128c486dcaccaeccdcef8c4 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py @@ -16,7 +16,7 @@ from __future__ import print_function import sys -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import math import sys diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py index a24834a6f0b19d1265f6c8d7089d31583af82d1f..a1ca6d981fafb401985d03e9f2d63d1cb41b21b5 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py @@ -13,7 +13,7 @@ # limitations under the License. import numpy as np -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import paddle.fluid.core as core import paddle.fluid.framework as framework diff --git a/python/paddle/fluid/tests/demo/fc_gan.py b/python/paddle/fluid/tests/demo/fc_gan.py index 7452ea2a34aa0c75d8e0990639b29705033af98b..8ea1b2b15cc0c0eb5bca67a9c5a6ac6c6774e7e2 100644 --- a/python/paddle/fluid/tests/demo/fc_gan.py +++ b/python/paddle/fluid/tests/demo/fc_gan.py @@ -19,7 +19,7 @@ import os import matplotlib import numpy -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid matplotlib.use('Agg') diff --git a/python/paddle/fluid/tests/test_concurrency.py b/python/paddle/fluid/tests/test_concurrency.py index 924895a9afac610059bac5f617c49712441339cc..e8f6cfb4a907b2c01e9662e7e9bf2cb0fbd6cb1b 100644 --- a/python/paddle/fluid/tests/test_concurrency.py +++ b/python/paddle/fluid/tests/test_concurrency.py @@ -173,16 +173,10 @@ class TestRoutineOp(unittest.TestCase): with while_op.block(): result2 = fill_constant( shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) - x_to_send_tmp = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.INT32, value=0) - - # TODO(abhinav): Need to perform copy when doing a channel send. - # Once this is complete, we can remove these lines - assign(input=x, output=x_to_send_tmp) with fluid.Select() as select: - with select.case(fluid.channel_send, channel, - x_to_send_tmp): + with select.case( + fluid.channel_send, channel, x, is_copy=True): assign(input=x, output=x_tmp) assign(input=y, output=x) assign(elementwise_add(x=x_tmp, y=y), output=y) @@ -230,21 +224,12 @@ class TestRoutineOp(unittest.TestCase): core.VarDesc.VarType.LOD_TENSOR, core.VarDesc.VarType.FP64) - pong_result = self._create_tensor('pong_return_value', - core.VarDesc.VarType.LOD_TENSOR, - core.VarDesc.VarType.FP64) - def ping(ch, message): - message_to_send_tmp = fill_constant( - shape=[1], dtype=core.VarDesc.VarType.FP64, value=0) - - assign(input=message, output=message_to_send_tmp) - fluid.channel_send(ch, message_to_send_tmp) + fluid.channel_send(ch, message, is_copy=True) def pong(ch1, ch2): fluid.channel_recv(ch1, ping_result) - assign(input=ping_result, output=pong_result) - fluid.channel_send(ch2, pong_result) + fluid.channel_send(ch2, ping_result, is_copy=True) pings = fluid.make_channel( dtype=core.VarDesc.VarType.LOD_TENSOR, capacity=1) diff --git a/python/paddle/fluid/tests/test_cpp_reader.py b/python/paddle/fluid/tests/test_cpp_reader.py index 4b0d039b7e05a55980946a8949e32802e9e57c20..e54c73b2956dd99ee57804318130c261e133d21a 100644 --- a/python/paddle/fluid/tests/test_cpp_reader.py +++ b/python/paddle/fluid/tests/test_cpp_reader.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid import numpy as np import sys diff --git a/python/paddle/fluid/tests/test_error_clip.py b/python/paddle/fluid/tests/test_error_clip.py index b2fd5ae29c724da52df0a5d3cb56d2ec9e5530f3..89f4c64975802dc1827ec17ed3626b91e36d6971 100644 --- a/python/paddle/fluid/tests/test_error_clip.py +++ b/python/paddle/fluid/tests/test_error_clip.py @@ -14,7 +14,7 @@ from __future__ import print_function import numpy as np -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid BATCH_SIZE = 128 diff --git a/python/paddle/fluid/tests/test_gradient_clip.py b/python/paddle/fluid/tests/test_gradient_clip.py index 68b682f68b1fd147b821cfdb1e0866cf8aa04bff..d530601f13be6810a8a99b13c92faf584df568f9 100644 --- a/python/paddle/fluid/tests/test_gradient_clip.py +++ b/python/paddle/fluid/tests/test_gradient_clip.py @@ -13,7 +13,7 @@ # limitations under the License. import numpy as np -import paddle.v2 as paddle +import paddle import paddle.fluid as fluid BATCH_SIZE = 128 diff --git a/python/paddle/fluid/tests/test_mnist_if_else_op.py b/python/paddle/fluid/tests/test_mnist_if_else_op.py index 94395f6cfb4648967558ed265e798e3505c20fc1..d34f52db5ffc889f17513d034ad2c99f696b0cdf 100644 --- a/python/paddle/fluid/tests/test_mnist_if_else_op.py +++ b/python/paddle/fluid/tests/test_mnist_if_else_op.py @@ -12,12 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +import paddle import paddle.fluid.layers as layers from paddle.fluid.framework import Program, program_guard, default_main_program, default_startup_program from paddle.fluid.executor import Executor from paddle.fluid.optimizer import MomentumOptimizer import paddle.fluid.core as core -import paddle.v2 as paddle import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/.gitignore b/python/paddle/fluid/tests/unittests/.gitignore index ad02bdecf436bba925e2e3b7efb20c878df70dfd..3538a9c2009bb133609153427981fb66974377fa 100644 --- a/python/paddle/fluid/tests/unittests/.gitignore +++ b/python/paddle/fluid/tests/unittests/.gitignore @@ -2,3 +2,5 @@ mnist.recordio mnist_0.recordio mnist_1.recordio mnist_2.recordio +flowers.recordio +wmt16.recordio diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index 8393f7827b1c7d361ebea72f2cfc6033268772f0..299ab8e51f017e1980a8b40e3830fc42b1ff7ccc 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -334,7 +334,7 @@ class OpTest(unittest.TestCase): np.allclose( actual_t, expect_t, atol=atol), "Output (" + out_name + ") has diff at " + str(place) + - str(actual_t) + str(expect_t)) + str(actual_t) + "\n" + str(expect_t)) if isinstance(expect, tuple): self.assertListEqual(actual.lod(), expect[1], "Output (" + out_name + @@ -568,6 +568,6 @@ class OpTest(unittest.TestCase): fetch_list = [g for p, g in param_grad_list] executor = Executor(place) - return map( - np.array, - executor.run(prog, feed_dict, fetch_list, return_numpy=False)) + return map(np.array, + executor.run(prog, feed_dict, fetch_list, + return_numpy=False)) diff --git a/python/paddle/fluid/tests/unittests/test_activation_op.py b/python/paddle/fluid/tests/unittests/test_activation_op.py index 1e3decfbaf0691e912b96b415b68353e626cf51e..fb162f8b7315936824ad40aca0c99e4dd09f9734 100644 --- a/python/paddle/fluid/tests/unittests/test_activation_op.py +++ b/python/paddle/fluid/tests/unittests/test_activation_op.py @@ -196,6 +196,34 @@ class TestFloor(OpTest): self.check_grad(['X'], 'Out', max_relative_error=0.007) +class TestCos(OpTest): + def setUp(self): + self.op_type = "cos" + x = np.random.uniform(-1, 1, [4, 4]).astype("float32") + self.inputs = {'X': x} + self.outputs = {'Out': np.cos(self.inputs['X'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out', max_relative_error=0.007) + + +class TestSin(OpTest): + def setUp(self): + self.op_type = "sin" + x = np.random.uniform(-1, 1, [4, 4]).astype("float32") + self.inputs = {'X': x} + self.outputs = {'Out': np.sin(self.inputs['X'])} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(['X'], 'Out', max_relative_error=0.007) + + class TestRound(OpTest): def setUp(self): self.op_type = "round" @@ -506,5 +534,54 @@ class TestSwish(OpTest): self.check_grad(['X'], 'Out', max_relative_error=0.008) +#--------------------test MKLDNN-------------------- +class TestMKLDNNRelu(TestRelu): + def setUp(self): + super(TestMKLDNNRelu, self).setUp() + + x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32") + # The same reason with TestAbs + x[np.abs(x) < 0.005] = 0.02 + out = np.maximum(x, 0) + + self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)} + self.outputs = {'Out': out} + self.attrs = {"use_mkldnn": True} + + +class TestMKLDNNTanh(TestTanh): + def setUp(self): + super(TestMKLDNNTanh, self).setUp() + + self.inputs = { + 'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32") + } + self.outputs = {'Out': np.tanh(self.inputs['X'])} + self.attrs = {"use_mkldnn": True} + + +class TestMKLDNNSqrt(TestSqrt): + def setUp(self): + super(TestMKLDNNSqrt, self).setUp() + + self.inputs = { + 'X': np.random.uniform(0.1, 1, [2, 4, 3, 5]).astype("float32") + } + self.outputs = {'Out': np.sqrt(self.inputs['X'])} + self.attrs = {"use_mkldnn": True} + + +class TestMKLDNNAbs(TestAbs): + def setUp(self): + super(TestMKLDNNAbs, self).setUp() + + x = np.random.uniform(-1, 1, [2, 4, 3, 5]).astype("float32") + # The same reason with TestAbs + x[np.abs(x) < 0.005] = 0.02 + self.inputs = {'X': x} + self.outputs = {'Out': np.abs(self.inputs['X'])} + self.attrs = {"use_mkldnn": True} + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_concat_op.py b/python/paddle/fluid/tests/unittests/test_concat_op.py index 558f3a4dcbb8fe39c427d8b100f4488440e8b8cb..1e00d67d5480bfa77a60e1aed52cafac6e8242ca 100644 --- a/python/paddle/fluid/tests/unittests/test_concat_op.py +++ b/python/paddle/fluid/tests/unittests/test_concat_op.py @@ -20,19 +20,35 @@ from op_test import OpTest class TestConcatOp(OpTest): def setUp(self): self.op_type = "concat" - x0 = np.random.random((2, 1, 4, 5)).astype('float32') - x1 = np.random.random((2, 2, 4, 5)).astype('float32') - x2 = np.random.random((2, 3, 4, 5)).astype('float32') - axis = 1 - self.inputs = {'X': [('x0', x0), ('x1', x1), ('x2', x2)]} - self.attrs = {'axis': axis} - self.outputs = {'Out': np.concatenate((x0, x1, x2), axis=axis)} + self.init_test_data() + self.inputs = {'X': [('x0', self.x0), ('x1', self.x1), ('x2', self.x2)]} + self.attrs = {'axis': self.axis} + self.outputs = { + 'Out': np.concatenate( + (self.x0, self.x1, self.x2), axis=self.axis) + } def test_check_output(self): self.check_output() def test_check_grad(self): self.check_grad(['x0'], 'Out') + self.check_grad(['x1'], 'Out') + self.check_grad(['x2'], 'Out') + + def init_test_data(self): + self.x0 = np.random.random((2, 1, 4, 5)).astype('float32') + self.x1 = np.random.random((2, 2, 4, 5)).astype('float32') + self.x2 = np.random.random((2, 3, 4, 5)).astype('float32') + self.axis = 1 + + +class TestConcatOp2(OpTest): + def init_test_data(self): + self.x0 = np.random.random((2, 3, 4, 5)).astype('float32') + self.x1 = np.random.random((2, 3, 4, 5)).astype('float32') + self.x2 = np.random.random((2, 3, 4, 5)).astype('float32') + self.axis = 1 if __name__ == '__main__': diff --git a/python/paddle/fluid/tests/unittests/test_dyn_rnn.py b/python/paddle/fluid/tests/unittests/test_dyn_rnn.py index df7ab0d29bdfc9410cd7dd4a8f2a7cd440ef4aba..0faed94deb4808783027d776e0f4c61da0db457a 100644 --- a/python/paddle/fluid/tests/unittests/test_dyn_rnn.py +++ b/python/paddle/fluid/tests/unittests/test_dyn_rnn.py @@ -13,7 +13,7 @@ # limitations under the License. import paddle.fluid as fluid -import paddle.v2 as paddle +import paddle import unittest import numpy diff --git a/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py b/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py index b03a70f1b9e61162d37541ffeba8510fc11c605a..d3f63ee2c414a71309be8f0af6d3e5912078ecdb 100644 --- a/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py +++ b/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py @@ -13,7 +13,7 @@ # limitations under the License. import unittest -import paddle.v2 as paddle +import paddle import paddle.fluid.core as core import paddle.fluid as fluid from paddle.fluid.backward import append_backward diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index b5fd59cf3a1bea50b799c3ace8f3b9cea088b9d5..2179826d81f715d6d280aea28a76f919330dd644 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -231,6 +231,13 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(layers.softmax(hid)) print(str(program)) + def test_lrn(self): + program = Program() + with program_guard(program): + data = layers.data(name='data', shape=[6, 2, 2], dtype='float32') + self.assertIsNotNone(layers.lrn(data)) + print(str(program)) + def test_get_places(self): program = Program() with program_guard(program): diff --git a/python/paddle/fluid/tests/unittests/test_lrn_op.py b/python/paddle/fluid/tests/unittests/test_lrn_op.py index eaff45cbb2a58798e9d55149510bec72eea370cd..8fa480b9bce84d2936f23cce9e41e8e54014b074 100644 --- a/python/paddle/fluid/tests/unittests/test_lrn_op.py +++ b/python/paddle/fluid/tests/unittests/test_lrn_op.py @@ -87,5 +87,34 @@ class TestLRNOp(OpTest): self.check_grad(['X'], 'Out', max_relative_error=0.01) +class TestLRNMKLDNNOp(TestLRNOp): + def get_attrs(self): + attrs = TestLRNOp.get_attrs(self) + attrs['use_mkldnn'] = True + return attrs + + def test_check_output(self): + self.check_output(atol=0.002) + + +class TestLRNMKLDNNOpWithIsTest(TestLRNMKLDNNOp): + def get_attrs(self): + attrs = TestLRNMKLDNNOp.get_attrs(self) + attrs['is_test'] = True + return attrs + + def test_check_grad_normal(self): + def check_raise_is_test(): + try: + self.check_grad(['X'], 'Out', max_relative_error=0.01) + except Exception as e: + t = \ + "is_test attribute should be set to False in training phase." + if t in str(e): + raise AttributeError + + self.assertRaises(AttributeError, check_raise_is_test) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py b/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py old mode 100755 new mode 100644 diff --git a/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py b/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py index 8add353303e3626bbce68199a100306d4858766a..0b7a29075939a548320185947b5afa7261029d49 100644 --- a/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py +++ b/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py @@ -15,8 +15,8 @@ import unittest import paddle.fluid as fluid -import paddle.v2 as paddle -import paddle.v2.dataset.mnist as mnist +import paddle +import paddle.dataset.mnist as mnist class TestMultipleReader(unittest.TestCase): diff --git a/python/paddle/fluid/tests/unittests/test_multiple_reader.py b/python/paddle/fluid/tests/unittests/test_multiple_reader.py index 69f8acf81efaba8fc0f3df4cfe3a42dc4e477df2..a60a5d6c4af2b6b3652d0fe2089018b9403eee25 100644 --- a/python/paddle/fluid/tests/unittests/test_multiple_reader.py +++ b/python/paddle/fluid/tests/unittests/test_multiple_reader.py @@ -15,8 +15,8 @@ import unittest import paddle.fluid as fluid -import paddle.v2 as paddle -import paddle.v2.dataset.mnist as mnist +import paddle +import paddle.dataset.mnist as mnist from shutil import copyfile diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor.py b/python/paddle/fluid/tests/unittests/test_parallel_executor.py new file mode 100644 index 0000000000000000000000000000000000000000..95d0f9da47e97e94ff97eb3647ac5244d5409ca3 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor.py @@ -0,0 +1,429 @@ +# 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 numpy +import unittest + +import paddle.fluid as fluid +import paddle +import paddle.dataset.mnist as mnist +import paddle.dataset.wmt16 as wmt16 + + +def simple_fc_net(): + reader = fluid.layers.open_recordio_file( + filename='./mnist.recordio', + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64']) + img, label = fluid.layers.read_file(reader) + hidden = img + for _ in xrange(4): + hidden = fluid.layers.fc( + hidden, + size=200, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + prediction = fluid.layers.fc(hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +def fc_with_batchnorm(): + reader = fluid.layers.open_recordio_file( + filename='./mnist.recordio', + shapes=[[-1, 784], [-1, 1]], + lod_levels=[0, 0], + dtypes=['float32', 'int64']) + img, label = fluid.layers.read_file(reader) + hidden = img + for _ in xrange(1): + hidden = fluid.layers.fc( + hidden, + size=200, + act='tanh', + bias_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=1.0))) + + hidden = fluid.layers.batch_norm(input=hidden) + + prediction = fluid.layers.fc(hidden, size=10, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +def squeeze_excitation(input, num_channels, reduction_ratio): + # pool = fluid.layers.pool2d( + # input=input, pool_size=0, pool_type='avg', global_pooling=True) + conv = input + shape = conv.shape + reshape = fluid.layers.reshape( + x=conv, shape=[-1, shape[1], shape[2] * shape[3]]) + pool = fluid.layers.reduce_mean(input=reshape, dim=2) + + squeeze = fluid.layers.fc(input=pool, + size=num_channels / reduction_ratio, + act='relu') + excitation = fluid.layers.fc(input=squeeze, + size=num_channels, + act='sigmoid') + scale = fluid.layers.elementwise_mul(x=input, y=excitation, axis=0) + return scale + + +def conv_bn_layer(input, num_filters, filter_size, stride=1, groups=1, + act=None): + conv = fluid.layers.conv2d( + input=input, + num_filters=num_filters, + filter_size=filter_size, + stride=stride, + padding=(filter_size - 1) / 2, + groups=groups, + act=None, + bias_attr=False) + return fluid.layers.batch_norm(input=conv, act=act, momentum=0.1) + + +def shortcut(input, ch_out, stride): + ch_in = input.shape[1] + if ch_in != ch_out: + if stride == 1: + filter_size = 1 + else: + filter_size = 3 + return conv_bn_layer(input, ch_out, filter_size, stride) + else: + return input + + +def bottleneck_block(input, num_filters, stride, cardinality, reduction_ratio): + # The number of first 1x1 convolutional channels for each bottleneck build block + # was halved to reduce the compution cost. + conv0 = conv_bn_layer( + input=input, num_filters=num_filters, filter_size=1, act='relu') + conv1 = conv_bn_layer( + input=conv0, + num_filters=num_filters * 2, + filter_size=3, + stride=stride, + groups=cardinality, + act='relu') + conv2 = conv_bn_layer( + input=conv1, num_filters=num_filters * 2, filter_size=1, act=None) + scale = squeeze_excitation( + input=conv2, + num_channels=num_filters * 2, + reduction_ratio=reduction_ratio) + + short = shortcut(input, num_filters * 2, stride) + + return fluid.layers.elementwise_add(x=short, y=scale, act='relu') + + +def SE_ResNeXt152(batch_size=4): + img = fluid.layers.fill_constant( + shape=[batch_size, 3, 224, 224], dtype='float32', value=0.0) + label = fluid.layers.fill_constant( + shape=[batch_size, 1], dtype='int64', value=0.0) + + conv = conv_bn_layer( + input=img, num_filters=64, filter_size=3, stride=2, act='relu') + conv = conv_bn_layer( + input=conv, num_filters=64, filter_size=3, stride=1, act='relu') + conv = conv_bn_layer( + input=conv, num_filters=128, filter_size=3, stride=1, act='relu') + conv = fluid.layers.pool2d( + input=conv, pool_size=3, pool_stride=2, pool_padding=1, pool_type='max') + + cardinality = 64 + reduction_ratio = 16 + depth = [3, 8, 36, 3] + num_filters = [128, 256, 512, 1024] + + for block in range(len(depth)): + for i in range(depth[block]): + conv = bottleneck_block( + input=conv, + num_filters=num_filters[block], + stride=2 if i == 0 and block != 0 else 1, + cardinality=cardinality, + reduction_ratio=reduction_ratio) + + shape = conv.shape + reshape = fluid.layers.reshape( + x=conv, shape=[-1, shape[1], shape[2] * shape[3]]) + pool = fluid.layers.reduce_mean(input=reshape, dim=2) + dropout = fluid.layers.dropout(x=pool, dropout_prob=0.2) + # Classifier layer: + prediction = fluid.layers.fc(input=dropout, size=1000, act='softmax') + loss = fluid.layers.cross_entropy(input=prediction, label=label) + loss = fluid.layers.mean(loss) + return loss + + +import time + + +class TestParallelExecutorBase(unittest.TestCase): + def check_network_convergence(self, + method, + memory_opt=True, + iter=10, + batch_size=None): + main = fluid.Program() + startup = fluid.Program() + with fluid.program_guard(main, startup): + loss = method() + adam = fluid.optimizer.Adam() + adam.minimize(loss) + if memory_opt: + fluid.memory_optimize(main) + + exe = fluid.ParallelExecutor(loss_name=loss.name, use_cuda=True) + if batch_size is not None: + batch_size *= fluid.core.get_cuda_device_count() + begin = time.time() + first_loss, = exe.run([loss.name]) + first_loss = numpy.array(first_loss) + + for i in xrange(iter): + exe.run([]) + + last_loss, = exe.run([loss.name]) + end = time.time() + + if batch_size is not None: + print "%.4f Instance per second" % ( + (batch_size * iter + 2) / (end - begin)) + + last_loss = numpy.array(last_loss) + + print first_loss, last_loss + # self.assertGreater(first_loss[0], last_loss[0]) + + +class TestMNIST(TestParallelExecutorBase): + @classmethod + def setUpClass(cls): + # Convert mnist to recordio file + with fluid.program_guard(fluid.Program(), fluid.Program()): + reader = paddle.batch(mnist.train(), batch_size=32) + 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', reader, feeder) + + def test_simple_fc(self): + self.check_network_convergence(simple_fc_net) + + def test_batchnorm_fc(self): + self.check_network_convergence(fc_with_batchnorm) + + +class TestResnet(TestParallelExecutorBase): + # @classmethod + # def setUpClass(cls): + # # import os + # # if os.path.exists('./flowers.recordio'): + # # return + # with fluid.program_guard(fluid.Program(), fluid.Program()): + # reader = paddle.batch(flowers.train(), batch_size=4) + # feeder = fluid.DataFeeder( + # feed_list=[ + # fluid.layers.data( + # name='image', shape=[3, 224, 224]), + # fluid.layers.data( + # name='label', shape=[1], dtype='int64'), + # ], + # place=fluid.CPUPlace()) + # fluid.recordio_writer.convert_reader_to_recordio_file( + # "./flowers.recordio", reader, feeder, compressor=fluid.core.RecordIOWriter.Compressor.NoCompress) + + def test_resnet(self): + import functools + batch_size = 4 + self.check_network_convergence( + functools.partial( + SE_ResNeXt152, batch_size=batch_size), + iter=20, + batch_size=batch_size) + + +class ModelHyperParams(object): + # Dictionary size for source and target language. This model directly uses + # paddle.dataset.wmt16 in which , and token has + # alreay been added, but the token is not added. Transformer requires + # sequences in a mini-batch are padded to have the same length. A token is + # added into the original dictionary in paddle.dateset.wmt16. + + # size of source word dictionary. + src_vocab_size = 10000 + # index for token in source language. + src_pad_idx = src_vocab_size + + # size of target word dictionay + trg_vocab_size = 10000 + # index for token in target language. + trg_pad_idx = trg_vocab_size + + # position value corresponding to the token. + pos_pad_idx = 0 + + # max length of sequences. It should plus 1 to include position + # padding token for position encoding. + max_length = 50 + + # the dimension for word embeddings, which is also the last dimension of + # the input and output of multi-head attention, position-wise feed-forward + # networks, encoder and decoder. + + d_model = 512 + # size of the hidden layer in position-wise feed-forward networks. + d_inner_hid = 1024 + # the dimension that keys are projected to for dot-product attention. + d_key = 64 + # the dimension that values are projected to for dot-product attention. + d_value = 64 + # number of head used in multi-head attention. + n_head = 8 + # number of sub-layers to be stacked in the encoder and decoder. + n_layer = 6 + # dropout rate used by all dropout layers. + dropout = 0.1 + + +import numpy as np + + +def prepare_batch_input(insts, src_pad_idx, trg_pad_idx, n_head): + """ + Pad the instances to the max sequence length in batch, and generate the + corresponding position data and attention bias. Then, convert the numpy + data to tensors and return a dict mapping names to tensors. + """ + + def __pad_batch_data(insts, + pad_idx, + is_target=False, + return_pos=True, + return_attn_bias=True, + return_max_len=True): + """ + Pad the instances to the max sequence length in batch, and generate the + corresponding position data and attention bias. + """ + return_list = [] + max_len = max(len(inst) for inst in insts) + inst_data = np.array( + [inst + [pad_idx] * (max_len - len(inst)) for inst in insts]) + return_list += [inst_data.astype("int64").reshape([-1, 1])] + if return_pos: + inst_pos = np.array([[ + pos_i + 1 if w_i != pad_idx else 0 + for pos_i, w_i in enumerate(inst) + ] for inst in inst_data]) + + return_list += [inst_pos.astype("int64").reshape([-1, 1])] + if return_attn_bias: + if is_target: + # This is used to avoid attention on paddings and subsequent + # words. + slf_attn_bias_data = np.ones((inst_data.shape[0], max_len, + max_len)) + slf_attn_bias_data = np.triu(slf_attn_bias_data, 1).reshape( + [-1, 1, max_len, max_len]) + slf_attn_bias_data = np.tile(slf_attn_bias_data, + [1, n_head, 1, 1]) * [-1e9] + else: + # This is used to avoid attention on paddings. + slf_attn_bias_data = np.array([[0] * len(inst) + [-1e9] * + (max_len - len(inst)) + for inst in insts]) + slf_attn_bias_data = np.tile( + slf_attn_bias_data.reshape([-1, 1, 1, max_len]), + [1, n_head, max_len, 1]) + return_list += [slf_attn_bias_data.astype("float32")] + if return_max_len: + return_list += [max_len] + return return_list if len(return_list) > 1 else return_list[0] + + def data_to_tensor(data_list, name_list, input_dict, place): + assert len(data_list) == len(name_list) + for i in range(len(name_list)): + tensor = fluid.LoDTensor() + tensor.set(data_list[i], place) + input_dict[name_list[i]] = tensor + + src_word, src_pos, src_slf_attn_bias, src_max_len = __pad_batch_data( + [inst[0] for inst in insts], src_pad_idx, is_target=False) + trg_word, trg_pos, trg_slf_attn_bias, trg_max_len = __pad_batch_data( + [inst[1] for inst in insts], trg_pad_idx, is_target=True) + trg_src_attn_bias = np.tile(src_slf_attn_bias[:, :, ::src_max_len, :], + [1, 1, trg_max_len, 1]).astype("float32") + lbl_word = __pad_batch_data([inst[2] for inst in insts], trg_pad_idx, False, + False, False, False) + lbl_weight = (lbl_word != trg_pad_idx).astype("float32").reshape([-1, 1]) + + return [ + src_word, src_pos, trg_word, trg_pos, src_slf_attn_bias, + trg_slf_attn_bias, trg_src_attn_bias, lbl_word, lbl_weight + ] + + +import transformer_model + + +def transformer(): + return transformer_model.transformer( + ModelHyperParams.src_vocab_size + 1, + ModelHyperParams.trg_vocab_size + 1, ModelHyperParams.max_length + 1, + ModelHyperParams.n_layer, ModelHyperParams.n_head, + ModelHyperParams.d_key, ModelHyperParams.d_value, + ModelHyperParams.d_model, ModelHyperParams.d_inner_hid, + ModelHyperParams.dropout, ModelHyperParams.src_pad_idx, + ModelHyperParams.trg_pad_idx, ModelHyperParams.pos_pad_idx) + + +class TestTransformer(TestParallelExecutorBase): + @classmethod + def setUpClass(cls): + reader = paddle.batch( + wmt16.train(ModelHyperParams.src_vocab_size, + ModelHyperParams.trg_vocab_size), + batch_size=transformer_model.batch_size) + + with fluid.recordio_writer.create_recordio_writer( + "./wmt16.recordio") as writer: + for batch in reader(): + for tensor in prepare_batch_input( + batch, ModelHyperParams.src_pad_idx, + ModelHyperParams.trg_pad_idx, ModelHyperParams.n_head): + t = fluid.LoDTensor() + t.set(tensor, fluid.CPUPlace()) + writer.append_tensor(t) + writer.complete_append_tensor() + + @unittest.skip("transformer is buggy in multi gpu") + def test_main(self): + self.check_network_convergence(transformer) diff --git a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py index 309ea2b9b7ede442da3ac897ce8d1a4b9aa68233..da85786d0c085a4e97d9ac272feed251296ad52d 100644 --- a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py +++ b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py @@ -186,6 +186,34 @@ class TestBlockDesc(unittest.TestCase): all_ops.append(block.op(idx)) self.assertEqual(all_ops, [op0, op1, op2]) + def test_remove_op(self): + prog = core.ProgramDesc() + self.assertIsNotNone(prog) + block = prog.block(0) + self.assertIsNotNone(block) + op1 = block.append_op() + op2 = block.append_op() + var1 = block.var("var1") + var2 = block.var("var2") + var3 = block.var("var3") + var4 = block.var("var4") + var5 = block.var("var5") + op1.set_input("X", ["var1", "var2"]) + op1.set_output("Y", ["var3", "var4"]) + op2.set_input("X", ["var1"]) + op2.set_output("Y", ["var4", "var5"]) + + # remove op1, its input var2 and output var3 will be removed at the same time, + # but its input var1 and output var4 will not be removed since they are used for op2. + block.remove_op(0, 1) + + all_ops = [] + for idx in xrange(0, block.op_size()): + all_ops.append(block.op(idx)) + self.assertEqual(all_ops, [op2]) + all_vars = block.all_vars() + self.assertEqual(set(all_vars), {var1, var4, var5}) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_recordio_reader.py b/python/paddle/fluid/tests/unittests/test_recordio_reader.py index 24a0074d9b9621d902d12eb8cb29d9b65be22ed3..640264d82f0dc7fa71bf882d5549e30b87b8d7c5 100644 --- a/python/paddle/fluid/tests/unittests/test_recordio_reader.py +++ b/python/paddle/fluid/tests/unittests/test_recordio_reader.py @@ -15,8 +15,8 @@ import unittest import paddle.fluid as fluid -import paddle.v2 as paddle -import paddle.v2.dataset.mnist as mnist +import paddle +import paddle.dataset.mnist as mnist class TestRecordIO(unittest.TestCase): diff --git a/python/paddle/fluid/tests/unittests/test_recv_op.py b/python/paddle/fluid/tests/unittests/test_recv_op.py index 985d892c568472614c5f3e6691f54807ddccc4bd..854238c6279528d8f3adf173140a47e233134f43 100644 --- a/python/paddle/fluid/tests/unittests/test_recv_op.py +++ b/python/paddle/fluid/tests/unittests/test_recv_op.py @@ -38,14 +38,15 @@ class TestRecvOp(unittest.TestCase): def init_serv(self, place): main = fluid.Program() with fluid.program_guard(main): - x = layers.data( - shape=[32, 32], - dtype='float32', - name="X", - append_batch_size=False) - fluid.initializer.Constant(value=1.0)(x, main.global_block()) - serv = layers.ListenAndServ("127.0.0.1:6174", optimizer_mode=False) + serv = layers.ListenAndServ( + "127.0.0.1:6174", ["X"], optimizer_mode=False) with serv.do(): + x = layers.data( + shape=[32, 32], + dtype='float32', + name="X", + append_batch_size=False) + fluid.initializer.Constant(value=1.0)(x, main.global_block()) o = layers.scale(x=x, scale=10.0) main.global_block().create_var( name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape) diff --git a/python/paddle/fluid/tests/unittests/test_reshape_op.py b/python/paddle/fluid/tests/unittests/test_reshape_op.py index 11f35c74d41146118525a5efa6c211d528e255fe..f51b5a7e9907294a5b91c920a363830d8b9a7137 100644 --- a/python/paddle/fluid/tests/unittests/test_reshape_op.py +++ b/python/paddle/fluid/tests/unittests/test_reshape_op.py @@ -14,15 +14,19 @@ import unittest import numpy as np + from op_test import OpTest class TestReshapeOp(OpTest): def setUp(self): + ori_shape = (2, 25) + new_shape = (5, 10) + self.op_type = "reshape" - self.inputs = {'X': np.random.random((10, 20)).astype("float32")} - self.attrs = {'shape': [10 * 20]} - self.outputs = {'Out': self.inputs['X'].reshape(self.attrs['shape'])} + self.inputs = {"X": np.random.random(ori_shape).astype("float32")} + self.attrs = {"shape": new_shape, "inplace": False} + self.outputs = {"Out": self.inputs["X"].reshape(new_shape)} def test_check_output(self): self.check_output() @@ -31,12 +35,33 @@ class TestReshapeOp(OpTest): self.check_grad(["X"], "Out") -class TestReshapeOpDimInfer(OpTest): +class TestReshapeOpDimInfer1(OpTest): def setUp(self): + ori_shape = (5, 10) + new_shape = (5, -1, 5) + self.op_type = "reshape" - self.inputs = {'X': np.random.random((10, 20)).astype("float32")} - self.attrs = {'shape': [4, -1, 5]} - self.outputs = {'Out': self.inputs['X'].reshape(self.attrs['shape'])} + self.inputs = {"X": np.random.random(ori_shape).astype("float32")} + self.attrs = {"shape": new_shape, "inplace": False} + self.outputs = {"Out": self.inputs["X"].reshape(self.attrs["shape"])} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + +class TestReshapeOpDimInfer2(OpTest): + def setUp(self): + ori_shape = (2, 2, 6) + new_shape = (2, 0, 3, -1) + infered_shape = (2, 2, 3, -1) + + self.op_type = "reshape" + self.inputs = {"X": np.random.random(ori_shape).astype("float32")} + self.attrs = {"shape": new_shape, "inplace": False} + self.outputs = {"Out": self.inputs["X"].reshape(infered_shape)} def test_check_output(self): self.check_output() @@ -47,10 +72,30 @@ class TestReshapeOpDimInfer(OpTest): class TestReshapeOpInplace(OpTest): def setUp(self): + ori_shape = (2, 25) + new_shape = (5, 10) + + self.op_type = "reshape" + self.inputs = {"X": np.random.random(ori_shape).astype("float32")} + self.attrs = {"shape": new_shape} + self.outputs = {"Out": self.inputs["X"].reshape(new_shape)} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + +class TestReshapeOpDimInferInplace1(OpTest): + def setUp(self): + ori_shape = (5, 10) + new_shape = (5, -1, 5) + self.op_type = "reshape" - self.inputs = {'X': np.random.random((10, 20)).astype("float32")} - self.attrs = {'shape': [10 * 20], 'inplace': True} - self.outputs = {'Out': self.inputs['X'].reshape(self.attrs['shape'])} + self.inputs = {"X": np.random.random(ori_shape).astype("float32")} + self.attrs = {"shape": new_shape} + self.outputs = {"Out": self.inputs["X"].reshape(new_shape)} def test_check_output(self): self.check_output() @@ -59,12 +104,38 @@ class TestReshapeOpInplace(OpTest): self.check_grad(["X"], "Out") -class TestReshapeOpDimInferInplace(OpTest): +class TestReshapeOpDimInferInplace2(OpTest): def setUp(self): + ori_shape = (2, 2, 6) + new_shape = (2, 0, 3, -1) + infered_shape = (2, 2, 3, -1) + + self.op_type = "reshape" + self.inputs = {"X": np.random.random(ori_shape).astype("float32")} + self.attrs = {"shape": new_shape} + self.outputs = {"Out": self.inputs["X"].reshape(infered_shape)} + + def test_check_output(self): + self.check_output() + + def test_check_grad(self): + self.check_grad(["X"], "Out") + + +class TestReshapeOpWithInputShape(OpTest): + def setUp(self): + ori_shape = (6, 5) + new_shape = (0, -1, 5) + actual_shape = (2, 3, 5) + self.op_type = "reshape" - self.inputs = {'X': np.random.random((10, 20)).astype("float32")} - self.attrs = {'shape': [4, -1, 5], 'inplace': True} - self.outputs = {'Out': self.inputs['X'].reshape(self.attrs['shape'])} + self.inputs = { + "X": np.random.random(ori_shape).astype("float32"), + "Shape": np.array( + actual_shape, dtype="int32") + } + self.attrs = {"shape": new_shape} + self.outputs = {"Out": self.inputs["X"].reshape(actual_shape)} def test_check_output(self): self.check_output() @@ -73,5 +144,5 @@ class TestReshapeOpDimInferInplace(OpTest): self.check_grad(["X"], "Out") -if __name__ == '__main__': +if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_seq_pool.py b/python/paddle/fluid/tests/unittests/test_seq_pool.py index 04884757216bc29a96eb97a6db403c3925472294..2e48ef0e880839f6d5b4e515a174f427a35e7e6f 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_pool.py +++ b/python/paddle/fluid/tests/unittests/test_seq_pool.py @@ -49,6 +49,61 @@ class TestSeqAvgPool(OpTest): self.check_grad(["X"], "Out") +class TestSeqSumPool(TestSeqAvgPool): + def compute(self, x, lod, out): + self.attrs = {'pooltype': "SUM"} + for i in range(4): + sub_x = x[lod[0][i]:lod[0][i + 1], :] + out[i] = sub_x.sum(axis=0) + + +class TestSeqMaxPool(TestSeqAvgPool): + def set_data(self): + self.op_type = 'sequence_pool' + x = np.random.uniform(0.1, 1, [13, 23]).astype('float32') + lod = [[0, 4, 5, 8, 13]] + for i in range(4): + l = lod[0][i + 1] - lod[0][i] + x[lod[0][i] + np.random.randint(l), :] += 2.0 + + self.inputs = {'X': (x, lod)} + + out = np.zeros((4, 23)).astype('float32') + self.outputs = {'Out': out} + return x, lod, out + + def compute(self, x, lod, out): + self.attrs = {'pooltype': "MAX"} + for i in range(4): + sub_x = x[lod[0][i]:lod[0][i + 1], :] + out[i] = np.amax(sub_x, axis=0) + + +class TestSeqSqrtPool(TestSeqAvgPool): + def compute(self, x, lod, out): + self.attrs = {'pooltype': "SQRT"} + for i in range(4): + sub_x = x[lod[0][i]:lod[0][i + 1], :] + len = lod[0][i + 1] - lod[0][i] + out[i] = sub_x.sum(axis=0) / np.sqrt(len) + + +class TestSeqLastPool(TestSeqAvgPool): + def compute(self, x, lod, out): + self.attrs = {'pooltype': "LAST"} + for i in range(4): + sub_x = x[lod[0][i]:lod[0][i + 1], :] + out[i] = sub_x[-1, :] + + +class TestSeqFirstPool(TestSeqAvgPool): + def compute(self, x, lod, out): + self.attrs = {'pooltype': "FIRST"} + for i in range(4): + sub_x = x[lod[0][i]:lod[0][i + 1], :] + out[i] = sub_x[0, :] + + class TestSeqAvgPool2D(TestSeqAvgPool): def set_data(self): self.op_type = 'sequence_pool' @@ -68,14 +123,6 @@ class TestSeqAvgPool2D(TestSeqAvgPool): out[i] = np.reshape(sub_x.mean(axis=0), (3, 17)) -class TestSeqSumPool(TestSeqAvgPool): - def compute(self, x, lod, out): - self.attrs = {'pooltype': "SUM"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] - out[i] = sub_x.sum(axis=0) - - class TestSeqSumPool2D(TestSeqAvgPool2D): def compute(self, x, lod, out): self.attrs = {'pooltype': "SUM"} @@ -84,15 +131,6 @@ class TestSeqSumPool2D(TestSeqAvgPool2D): out[i] = np.reshape(sub_x.sum(axis=0), (3, 17)) -class TestSeqSqrtPool(TestSeqAvgPool): - def compute(self, x, lod, out): - self.attrs = {'pooltype': "SQRT"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] - len = lod[0][i + 1] - lod[0][i] - out[i] = sub_x.sum(axis=0) / np.sqrt(len) - - class TestSeqSqrtPool2D(TestSeqAvgPool2D): def compute(self, x, lod, out): self.attrs = {'pooltype': "SQRT"} @@ -108,28 +146,6 @@ class TestSeqSqrtPool2D(TestSeqAvgPool2D): self.check_grad(["X"], "Out", max_relative_error=0.06) -class TestSeqMaxPool(TestSeqAvgPool): - def set_data(self): - self.op_type = 'sequence_pool' - x = np.random.uniform(0.1, 1, [13, 23]).astype('float32') - lod = [[0, 4, 5, 8, 13]] - for i in range(4): - l = lod[0][i + 1] - lod[0][i] - x[lod[0][i] + np.random.randint(l), :] += 2.0 - - self.inputs = {'X': (x, lod)} - - out = np.zeros((4, 23)).astype('float32') - self.outputs = {'Out': out} - return x, lod, out - - def compute(self, x, lod, out): - self.attrs = {'pooltype': "MAX"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] - out[i] = np.amax(sub_x, axis=0) - - class TestSeqMaxPool2D(TestSeqAvgPool2D): def set_data(self): self.op_type = 'sequence_pool' @@ -151,14 +167,6 @@ class TestSeqMaxPool2D(TestSeqAvgPool2D): out[i] = np.reshape(np.amax(sub_x, axis=0), (3, 11)) -class TestSeqLastPool(TestSeqAvgPool): - def compute(self, x, lod, out): - self.attrs = {'pooltype': "LAST"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] - out[i] = sub_x[-1, :] - - class TestSeqLastPool2D(TestSeqAvgPool2D): def compute(self, x, lod, out): self.attrs = {'pooltype': "LAST"} @@ -167,14 +175,6 @@ class TestSeqLastPool2D(TestSeqAvgPool2D): out[i] = np.reshape(sub_x[-1, :], (3, 17)) -class TestSeqFirstPool(TestSeqAvgPool): - def compute(self, x, lod, out): - self.attrs = {'pooltype': "FIRST"} - for i in range(4): - sub_x = x[lod[0][i]:lod[0][i + 1], :] - out[i] = sub_x[0, :] - - class TestSeqFirstPool2D(TestSeqAvgPool2D): def compute(self, x, lod, out): self.attrs = {'pooltype': "FIRST"} diff --git a/python/paddle/fluid/tests/unittests/test_split_ids_op.py b/python/paddle/fluid/tests/unittests/test_split_ids_op.py new file mode 100644 index 0000000000000000000000000000000000000000..e9f0a06a56b42952800411d548bb3fc1732e031e --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_split_ids_op.py @@ -0,0 +1,35 @@ +# 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 TestSplitIdsOp(OpTest): + def setUp(self): + self.op_type = "split_ids" + ids = np.array([[0], [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} + self.outputs = {'Out': [('out0', out0), ('out1', out1), ('out2', out2)]} + + def test_check_output(self): + self.check_output() + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_target_assign_op.py b/python/paddle/fluid/tests/unittests/test_target_assign_op.py old mode 100755 new mode 100644 diff --git a/python/paddle/fluid/tests/unittests/test_tensor.py b/python/paddle/fluid/tests/unittests/test_tensor.py index a369783245ae2e35a9743ef1f4321ac919e58283..379081c3287ce81dbf2bd7307cb5eac2620b13db 100644 --- a/python/paddle/fluid/tests/unittests/test_tensor.py +++ b/python/paddle/fluid/tests/unittests/test_tensor.py @@ -126,7 +126,6 @@ class TestTensor(unittest.TestCase): def test_lod_tensor_gpu_init(self): if not core.is_compiled_with_cuda(): return - scope = core.Scope() place = core.CUDAPlace(0) lod_py = [[0, 2, 5], [0, 2, 4, 5]] lod_tensor = core.LoDTensor() @@ -144,6 +143,25 @@ class TestTensor(unittest.TestCase): self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1]) self.assertListEqual(lod_py, lod_tensor.lod()) + def test_empty_tensor(self): + place = core.CPUPlace() + scope = core.Scope() + var = scope.var("test_tensor") + + tensor = var.get_tensor() + + tensor.set_dims([0, 1]) + tensor.alloc_float(place) + + tensor_array = numpy.array(tensor) + self.assertEqual((0, 1), tensor_array.shape) + + if core.is_compiled_with_cuda(): + gpu_place = core.CUDAPlace(0) + tensor.alloc_float(gpu_place) + tensor_array = numpy.array(tensor) + self.assertEqual((0, 1), tensor_array.shape) + if __name__ == '__main__': unittest.main() diff --git a/python/paddle/fluid/tests/unittests/transformer_model.py b/python/paddle/fluid/tests/unittests/transformer_model.py new file mode 100644 index 0000000000000000000000000000000000000000..c62792face3c353db1f2e3c77eaf4bd32fbded69 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/transformer_model.py @@ -0,0 +1,487 @@ +# 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 functools import partial +import numpy as np + +import paddle.fluid as fluid +import paddle.fluid.layers as layers + +pos_enc_param_names = ( + "src_pos_enc_table", + "trg_pos_enc_table", ) + +batch_size = 64 + + +def position_encoding_init(n_position, d_pos_vec): + """ + Generate the initial values for the sinusoid position encoding table. + """ + position_enc = np.array([[ + pos / np.power(10000, 2 * (j // 2) / d_pos_vec) + for j in range(d_pos_vec) + ] if pos != 0 else np.zeros(d_pos_vec) for pos in range(n_position)]) + position_enc[1:, 0::2] = np.sin(position_enc[1:, 0::2]) # dim 2i + position_enc[1:, 1::2] = np.cos(position_enc[1:, 1::2]) # dim 2i+1 + return position_enc.astype("float32") + + +def multi_head_attention(queries, + keys, + values, + attn_bias, + d_key, + d_value, + d_model, + n_head=1, + dropout_rate=0.): + """ + Multi-Head Attention. Note that attn_bias is added to the logit before + computing softmax activiation to mask certain selected positions so that + they will not considered in attention weights. + """ + if not (len(queries.shape) == len(keys.shape) == len(values.shape) == 3): + raise ValueError( + "Inputs: quries, keys and values should all be 3-D tensors.") + + def __compute_qkv(queries, keys, values, n_head, d_key, d_value): + """ + Add linear projection to queries, keys, and values. + """ + q = layers.fc(input=queries, + size=d_key * n_head, + param_attr=fluid.initializer.Xavier( + uniform=False, + fan_in=d_model * d_key, + fan_out=n_head * d_key), + bias_attr=False, + num_flatten_dims=2) + k = layers.fc(input=keys, + size=d_key * n_head, + param_attr=fluid.initializer.Xavier( + uniform=False, + fan_in=d_model * d_key, + fan_out=n_head * d_key), + bias_attr=False, + num_flatten_dims=2) + v = layers.fc(input=values, + size=d_value * n_head, + param_attr=fluid.initializer.Xavier( + uniform=False, + fan_in=d_model * d_value, + fan_out=n_head * d_value), + bias_attr=False, + num_flatten_dims=2) + return q, k, v + + def __split_heads(x, n_head): + """ + Reshape the last dimension of inpunt tensor x so that it becomes two + dimensions and then transpose. Specifically, input a tensor with shape + [bs, max_sequence_length, n_head * hidden_dim] then output a tensor + with shape [bs, n_head, max_sequence_length, hidden_dim]. + """ + if n_head == 1: + return x + + hidden_size = x.shape[-1] + # FIXME(guosheng): Decouple the program desc with batch_size. + reshaped = layers.reshape( + x=x, shape=[batch_size, -1, n_head, hidden_size // n_head]) + + # permuate the dimensions into: + # [batch_size, n_head, max_sequence_len, hidden_size_per_head] + return layers.transpose(x=reshaped, perm=[0, 2, 1, 3]) + + def __combine_heads(x): + """ + Transpose and then reshape the last two dimensions of inpunt tensor x + so that it becomes one dimension, which is reverse to __split_heads. + """ + if len(x.shape) == 3: return x + if len(x.shape) != 4: + raise ValueError("Input(x) should be a 4-D Tensor.") + + trans_x = layers.transpose(x, perm=[0, 2, 1, 3]) + # FIXME(guosheng): Decouple the program desc with batch_size. + return layers.reshape( + x=trans_x, + shape=map(int, + [batch_size, -1, trans_x.shape[2] * trans_x.shape[3]])) + + def scaled_dot_product_attention(q, k, v, attn_bias, d_model, dropout_rate): + """ + Scaled Dot-Product Attention + """ + + # FIXME(guosheng): Optimize the shape in reshape_op or softmax_op. + + # The current implementation of softmax_op only supports 2D tensor, + # consequently it cannot be directly used here. + # If to use the reshape_op, Besides, the shape of product inferred in + # compile-time is not the actual shape in run-time. It cann't be used + # to set the attribute of reshape_op. + # So, here define the softmax for temporary solution. + + def __softmax(x, eps=1e-9): + exp_out = layers.exp(x=x) + sum_out = layers.reduce_sum(exp_out, dim=-1, keep_dim=False) + return layers.elementwise_div(x=exp_out, y=sum_out, axis=0) + + scaled_q = layers.scale(x=q, scale=d_model**-0.5) + product = layers.matmul(x=scaled_q, y=k, transpose_y=True) + weights = __softmax(layers.elementwise_add(x=product, y=attn_bias)) + if dropout_rate: + weights = layers.dropout( + weights, dropout_prob=dropout_rate, is_test=False) + out = layers.matmul(weights, v) + return out + + q, k, v = __compute_qkv(queries, keys, values, n_head, d_key, d_value) + + q = __split_heads(q, n_head) + k = __split_heads(k, n_head) + v = __split_heads(v, n_head) + + ctx_multiheads = scaled_dot_product_attention(q, k, v, attn_bias, d_model, + dropout_rate) + + out = __combine_heads(ctx_multiheads) + + # Project back to the model size. + proj_out = layers.fc(input=out, + size=d_model, + param_attr=fluid.initializer.Xavier(uniform=False), + bias_attr=False, + num_flatten_dims=2) + return proj_out + + +def positionwise_feed_forward(x, d_inner_hid, d_hid): + """ + Position-wise Feed-Forward Networks. + This module consists of two linear transformations with a ReLU activation + in between, which is applied to each position separately and identically. + """ + hidden = layers.fc(input=x, + size=d_inner_hid, + num_flatten_dims=2, + param_attr=fluid.initializer.Uniform( + low=-(d_hid**-0.5), high=(d_hid**-0.5)), + act="relu") + out = layers.fc(input=hidden, + size=d_hid, + num_flatten_dims=2, + param_attr=fluid.initializer.Uniform( + low=-(d_inner_hid**-0.5), high=(d_inner_hid**-0.5))) + return out + + +def pre_post_process_layer(prev_out, out, process_cmd, dropout=0.): + """ + Add residual connection, layer normalization and droput to the out tensor + optionally according to the value of process_cmd. + + This will be used before or after multi-head attention and position-wise + feed-forward networks. + """ + for cmd in process_cmd: + if cmd == "a": # add residual connection + out = out + prev_out if prev_out else out + elif cmd == "n": # add layer normalization + out = layers.layer_norm( + out, + begin_norm_axis=len(out.shape) - 1, + param_attr=fluid.initializer.Constant(1.), + bias_attr=fluid.initializer.Constant(0.)) + elif cmd == "d": # add dropout + if dropout: + out = layers.dropout(out, dropout_prob=dropout, is_test=False) + return out + + +pre_process_layer = partial(pre_post_process_layer, None) +post_process_layer = pre_post_process_layer + + +def prepare_encoder(src_word, + src_pos, + src_vocab_size, + src_emb_dim, + src_pad_idx, + src_max_len, + dropout=0., + pos_pad_idx=0, + pos_enc_param_name=None): + """Add word embeddings and position encodings. + The output tensor has a shape of: + [batch_size, max_src_length_in_batch, d_model]. + + This module is used at the bottom of the encoder stacks. + """ + src_word_emb = layers.embedding( + src_word, + size=[src_vocab_size, src_emb_dim], + padding_idx=src_pad_idx, + param_attr=fluid.initializer.Normal(0., 1.)) + src_pos_enc = layers.embedding( + src_pos, + size=[src_max_len, src_emb_dim], + padding_idx=pos_pad_idx, + param_attr=fluid.ParamAttr( + name=pos_enc_param_name, trainable=False)) + enc_input = src_word_emb + src_pos_enc + + # FIXME(guosheng): Decouple the program desc with batch_size. + enc_input = layers.reshape(x=enc_input, shape=[batch_size, -1, src_emb_dim]) + return layers.dropout( + enc_input, dropout_prob=dropout, + is_test=False) if dropout else enc_input + + +prepare_encoder = partial( + prepare_encoder, pos_enc_param_name=pos_enc_param_names[0]) +prepare_decoder = partial( + prepare_encoder, pos_enc_param_name=pos_enc_param_names[1]) + + +def encoder_layer(enc_input, + attn_bias, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate=0.): + """The encoder layers that can be stacked to form a deep encoder. + + This module consits of a multi-head (self) attention followed by + position-wise feed-forward networks and both the two components companied + with the post_process_layer to add residual connection, layer normalization + and droput. + """ + attn_output = multi_head_attention(enc_input, enc_input, enc_input, + attn_bias, d_key, d_value, d_model, + n_head, dropout_rate) + attn_output = post_process_layer(enc_input, attn_output, "dan", + dropout_rate) + ffd_output = positionwise_feed_forward(attn_output, d_inner_hid, d_model) + return post_process_layer(attn_output, ffd_output, "dan", dropout_rate) + + +def encoder(enc_input, + attn_bias, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate=0.): + """ + The encoder is composed of a stack of identical layers returned by calling + encoder_layer. + """ + for i in range(n_layer): + enc_output = encoder_layer(enc_input, attn_bias, n_head, d_key, d_value, + d_model, d_inner_hid, dropout_rate) + enc_input = enc_output + return enc_output + + +def decoder_layer(dec_input, + enc_output, + slf_attn_bias, + dec_enc_attn_bias, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate=0.): + """ The layer to be stacked in decoder part. + + The structure of this module is similar to that in the encoder part except + a multi-head attention is added to implement encoder-decoder attention. + """ + slf_attn_output = multi_head_attention( + dec_input, + dec_input, + dec_input, + slf_attn_bias, + d_key, + d_value, + d_model, + n_head, + dropout_rate, ) + slf_attn_output = post_process_layer( + dec_input, + slf_attn_output, + "dan", # residual connection + dropout + layer normalization + dropout_rate, ) + enc_attn_output = multi_head_attention( + slf_attn_output, + enc_output, + enc_output, + dec_enc_attn_bias, + d_key, + d_value, + d_model, + n_head, + dropout_rate, ) + enc_attn_output = post_process_layer( + slf_attn_output, + enc_attn_output, + "dan", # residual connection + dropout + layer normalization + dropout_rate, ) + ffd_output = positionwise_feed_forward( + enc_attn_output, + d_inner_hid, + d_model, ) + dec_output = post_process_layer( + enc_attn_output, + ffd_output, + "dan", # residual connection + dropout + layer normalization + dropout_rate, ) + return dec_output + + +def decoder(dec_input, + enc_output, + dec_slf_attn_bias, + dec_enc_attn_bias, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate=0.): + """ + The decoder is composed of a stack of identical decoder_layer layers. + """ + for i in range(n_layer): + dec_output = decoder_layer( + dec_input, + enc_output, + dec_slf_attn_bias, + dec_enc_attn_bias, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, ) + dec_input = dec_output + return dec_output + + +def transformer( + src_vocab_size, + trg_vocab_size, + max_length, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, + src_pad_idx, + trg_pad_idx, + pos_pad_idx, ): + file_obj = fluid.layers.open_recordio_file( + filename='./wmt16.recordio', + shapes=[ + [batch_size * max_length, 1], + [batch_size * max_length, 1], + [batch_size * max_length, 1], + [batch_size * max_length, 1], + [batch_size, n_head, max_length, max_length], + [batch_size, n_head, max_length, max_length], + [batch_size, n_head, max_length, max_length], + [batch_size * max_length, 1], + [batch_size * max_length, 1], + ], + dtypes=[ + 'int64', + 'int64', + 'int64', + 'int64', + 'float32', + 'float32', + 'float32', + 'int64', + 'float32', + ], + lod_levels=[0] * 9) + + src_word, src_pos, trg_word, trg_pos, src_slf_attn_bias, trg_slf_attn_bias, trg_src_attn_bias, gold, weights = fluid.layers.read_file( + file_obj) + + enc_input = prepare_encoder( + src_word, + src_pos, + src_vocab_size, + d_model, + src_pad_idx, + max_length, + dropout_rate, ) + enc_output = encoder( + enc_input, + src_slf_attn_bias, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, ) + + dec_input = prepare_decoder( + trg_word, + trg_pos, + trg_vocab_size, + d_model, + trg_pad_idx, + max_length, + dropout_rate, ) + dec_output = decoder( + dec_input, + enc_output, + trg_slf_attn_bias, + trg_src_attn_bias, + n_layer, + n_head, + d_key, + d_value, + d_model, + d_inner_hid, + dropout_rate, ) + + # TODO(guosheng): Share the weight matrix between the embedding layers and + # the pre-softmax linear transformation. + predict = layers.reshape( + x=layers.fc(input=dec_output, + size=trg_vocab_size, + param_attr=fluid.initializer.Xavier(uniform=False), + bias_attr=False, + num_flatten_dims=2), + shape=[-1, trg_vocab_size], + act="softmax") + + cost = layers.cross_entropy(input=predict, label=gold) + weighted_cost = cost * weights + return layers.reduce_sum(weighted_cost) diff --git a/python/paddle/v2/reader/__init__.py b/python/paddle/reader/__init__.py similarity index 100% rename from python/paddle/v2/reader/__init__.py rename to python/paddle/reader/__init__.py diff --git a/python/paddle/v2/reader/creator.py b/python/paddle/reader/creator.py similarity index 62% rename from python/paddle/v2/reader/creator.py rename to python/paddle/reader/creator.py index 421f6c933d7032e4103f504fc509e2d5c89149b2..4c905d959fad4e8c1a8826ce8dc60c5fa834514d 100644 --- a/python/paddle/v2/reader/creator.py +++ b/python/paddle/reader/creator.py @@ -16,7 +16,7 @@ Creator package contains some simple reader creator, which could be used in user program. """ -__all__ = ['np_array', 'text_file', "cloud_reader"] +__all__ = ['np_array', 'text_file', 'recordio'] def np_array(x): @@ -66,7 +66,7 @@ def recordio(paths, buf_size=100): """ import recordio as rec - import paddle.v2.reader.decorator as dec + import paddle.reader.decorator as dec import cPickle as pickle def reader(): @@ -83,48 +83,3 @@ def recordio(paths, buf_size=100): f.close() return dec.buffered(reader, buf_size) - - -pass_num = 0 - - -def cloud_reader(paths, etcd_endpoints, timeout_sec=5, buf_size=64): - """ - Create a data reader that yield a record one by one from - the paths: - :paths: path of recordio files, can be a string or a string list. - :etcd_endpoints: the endpoints for etcd cluster - :returns: data reader of recordio files. - - .. code-block:: python - from paddle.v2.reader.creator import cloud_reader - etcd_endpoints = "http://127.0.0.1:2379" - trainer.train.( - reader=cloud_reader(["/work/dataset/uci_housing/uci_housing*"], etcd_endpoints), - ) - """ - import os - import cPickle as pickle - import paddle.v2.master as master - c = master.client(etcd_endpoints, timeout_sec, buf_size) - - if isinstance(paths, basestring): - path = [paths] - else: - path = paths - c.set_dataset(path) - - def reader(): - global pass_num - c.paddle_start_get_records(pass_num) - pass_num += 1 - - while True: - r, e = c.next_record() - if not r: - if e != -2: - print "get record error: ", e - break - yield pickle.loads(r) - - return reader diff --git a/python/paddle/v2/reader/decorator.py b/python/paddle/reader/decorator.py similarity index 100% rename from python/paddle/v2/reader/decorator.py rename to python/paddle/reader/decorator.py diff --git a/python/paddle/v2/reader/tests/CMakeLists.txt b/python/paddle/reader/tests/CMakeLists.txt similarity index 100% rename from python/paddle/v2/reader/tests/CMakeLists.txt rename to python/paddle/reader/tests/CMakeLists.txt diff --git a/python/paddle/v2/reader/tests/__init__.py b/python/paddle/reader/tests/__init__.py similarity index 100% rename from python/paddle/v2/reader/tests/__init__.py rename to python/paddle/reader/tests/__init__.py diff --git a/python/paddle/v2/reader/tests/creator_test.py b/python/paddle/reader/tests/creator_test.py similarity index 92% rename from python/paddle/v2/reader/tests/creator_test.py rename to python/paddle/reader/tests/creator_test.py index 7fe374e663607607cd0839eb6ca9c70c4d15eef8..c4238c12a74759d52eb09f31ce1126cc93dd3489 100644 --- a/python/paddle/v2/reader/tests/creator_test.py +++ b/python/paddle/reader/tests/creator_test.py @@ -28,14 +28,14 @@ import os import unittest import numpy as np -import paddle.v2.reader.creator +import paddle.reader.creator class TestNumpyArray(unittest.TestCase): def test_numpy_array(self): l = [[1, 2, 3], [4, 5, 6]] x = np.array(l, np.int32) - reader = paddle.v2.reader.creator.np_array(x) + reader = paddle.reader.creator.np_array(x) for idx, e in enumerate(reader()): self.assertItemsEqual(e, l[idx]) @@ -43,14 +43,14 @@ class TestNumpyArray(unittest.TestCase): class TestTextFile(unittest.TestCase): def test_text_file(self): path = os.path.join(os.path.dirname(__file__), "test_data_creator.txt") - reader = paddle.v2.reader.creator.text_file(path) + reader = paddle.reader.creator.text_file(path) for idx, e in enumerate(reader()): self.assertEqual(e, str(idx * 2) + " " + str(idx * 2 + 1)) class TestRecordIO(unittest.TestCase): def do_test(self, path): - reader = paddle.v2.reader.creator.recordio(path) + reader = paddle.reader.creator.recordio(path) idx = 0 for e in reader(): if idx == 0: diff --git a/python/paddle/v2/reader/tests/decorator_test.py b/python/paddle/reader/tests/decorator_test.py similarity index 81% rename from python/paddle/v2/reader/tests/decorator_test.py rename to python/paddle/reader/tests/decorator_test.py index 6b680e39f3fb299a14e7d8162470996d1d16b83d..bee24d3b6579db5e99ec66931df201fdf9e1af07 100644 --- a/python/paddle/v2/reader/tests/decorator_test.py +++ b/python/paddle/reader/tests/decorator_test.py @@ -15,7 +15,7 @@ import time import unittest -import paddle.v2.reader +import paddle.reader def reader_creator_10(dur): @@ -39,7 +39,7 @@ class TestMap(unittest.TestCase): yield "h" yield "i" - r = paddle.v2.reader.map_readers(tokenize, read) + r = paddle.reader.map_readers(tokenize, read) for i, e in enumerate(r()): self.assertEqual(e, i) @@ -47,7 +47,7 @@ class TestMap(unittest.TestCase): class TestBuffered(unittest.TestCase): def test_read(self): for size in range(20): - b = paddle.v2.reader.buffered(reader_creator_10(0), size) + b = paddle.reader.buffered(reader_creator_10(0), size) c = 0 for i in b(): self.assertEqual(i, c) @@ -56,7 +56,7 @@ class TestBuffered(unittest.TestCase): def test_buffering(self): # read have 30ms delay. - b = paddle.v2.reader.buffered(reader_creator_10(0.03), 10) + b = paddle.reader.buffered(reader_creator_10(0.03), 10) last_time = time.time() for idx, i in enumerate(b()): elapsed_time = time.time() - last_time @@ -70,17 +70,17 @@ class TestBuffered(unittest.TestCase): class TestCompose(unittest.TestCase): def test_compse(self): - reader = paddle.v2.reader.compose( + reader = paddle.reader.compose( reader_creator_10(0), reader_creator_10(0)) for idx, e in enumerate(reader()): self.assertEqual(e, (idx, idx)) def test_compose_not_aligned(self): total = 0 - reader = paddle.v2.reader.compose( - paddle.v2.reader.chain(reader_creator_10(0), reader_creator_10(0)), + reader = paddle.reader.compose( + paddle.reader.chain(reader_creator_10(0), reader_creator_10(0)), reader_creator_10(0)) - with self.assertRaises(paddle.v2.reader.ComposeNotAligned): + with self.assertRaises(paddle.reader.ComposeNotAligned): for e in reader(): total += 1 # expecting 10, not 20 @@ -88,8 +88,8 @@ class TestCompose(unittest.TestCase): def test_compose_not_aligned_no_check(self): total = 0 - reader = paddle.v2.reader.compose( - paddle.v2.reader.chain(reader_creator_10(0), reader_creator_10(0)), + reader = paddle.reader.compose( + paddle.reader.chain(reader_creator_10(0), reader_creator_10(0)), reader_creator_10(0), check_alignment=False) for e in reader(): @@ -100,7 +100,7 @@ class TestCompose(unittest.TestCase): class TestChain(unittest.TestCase): def test_chain(self): - c = paddle.v2.reader.chain(reader_creator_10(0), reader_creator_10(0)) + c = paddle.reader.chain(reader_creator_10(0), reader_creator_10(0)) idx = 0 for e in c(): self.assertEqual(e, idx % 10) @@ -113,7 +113,7 @@ class TestShuffle(unittest.TestCase): case = [(0, True), (1, True), (10, False), (100, False)] a = reader_creator_10(0) for size, checkEq in case: - s = paddle.v2.reader.shuffle(a, size) + s = paddle.reader.shuffle(a, size) total = 0 for idx, e in enumerate(s()): if checkEq: @@ -133,9 +133,9 @@ class TestXmap(unittest.TestCase): for order in orders: for tNum in thread_nums: for size in buffered_size: - reader = paddle.v2.reader.xmap_readers(mapper, - reader_creator_10(0), - tNum, size, order) + reader = paddle.reader.xmap_readers(mapper, + reader_creator_10(0), + tNum, size, order) for n in xrange(3): result = [] for i in reader(): @@ -150,7 +150,7 @@ class TestPipeReader(unittest.TestCase): def test_pipe_reader(self): def example_reader(myfiles): for f in myfiles: - pr = paddle.v2.reader.PipeReader("cat %s" % f, bufsize=128) + pr = paddle.reader.PipeReader("cat %s" % f, bufsize=128) for l in pr.get_line(): yield l diff --git a/python/paddle/v2/reader/tests/test_data_creator.txt b/python/paddle/reader/tests/test_data_creator.txt similarity index 100% rename from python/paddle/v2/reader/tests/test_data_creator.txt rename to python/paddle/reader/tests/test_data_creator.txt diff --git a/python/paddle/v2/reader/tests/test_reader_recordio.dat b/python/paddle/reader/tests/test_reader_recordio.dat similarity index 100% rename from python/paddle/v2/reader/tests/test_reader_recordio.dat rename to python/paddle/reader/tests/test_reader_recordio.dat diff --git a/python/paddle/v2/reader/tests/test_recordio_creator.dat b/python/paddle/reader/tests/test_recordio_creator.dat similarity index 100% rename from python/paddle/v2/reader/tests/test_recordio_creator.dat rename to python/paddle/reader/tests/test_recordio_creator.dat diff --git a/python/paddle/trainer_config_helpers/activations.py b/python/paddle/trainer_config_helpers/activations.py index 00efc01c0592107314f5b23c951706d039d49a88..3683968262266a2d654d2480b828173bc761152b 100644 --- a/python/paddle/trainer_config_helpers/activations.py +++ b/python/paddle/trainer_config_helpers/activations.py @@ -77,7 +77,7 @@ class SoftmaxActivation(BaseActivation): .. math:: - P(y=j|x) = \\frac{e^{x_j}} {\\sum^K_{k=1} e^{x_j} } + P(y=j|x) = \\frac{e^{x_j}} {\\sum^K_{k=1} e^{x_k} } """ def __init__(self): diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index eac2cb316835fda0a52ac9895eaa80914d0f1e5b..3684d1e8f73a21d9c6f2a5985f8b40ed6984057b 100644 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -2747,17 +2747,17 @@ def img_pool_layer(input, .. math:: - w & = 1 + \\frac{ceil(input\_width + 2 * padding - pool\_size)}{stride} + w & = 1 + ceil(\\frac{input\_width + 2 * padding - pool\_size}{stride}) - h & = 1 + \\frac{ceil(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y} + h & = 1 + ceil(\\frac{input\_height + 2 * padding\_y - pool\_size\_y}{stride\_y}) - ceil_mode=False: .. math:: - w & = 1 + \\frac{floor(input\_width + 2 * padding - pool\_size)}{stride} + w & = 1 + floor(\\frac{input\_width + 2 * padding - pool\_size}{stride}) - h & = 1 + \\frac{floor(input\_height + 2 * padding\_y - pool\_size\_y)}{stride\_y} + h & = 1 + floor(\\frac{input\_height + 2 * padding\_y - pool\_size\_y}{stride\_y}) The example usage is: diff --git a/python/paddle/v2/__init__.py b/python/paddle/v2/__init__.py index df710c33d0c0ca16d358dac1eb42327e9cd4c7ae..02b0d077eefa431baed05c421a367ebe3581626c 100644 --- a/python/paddle/v2/__init__.py +++ b/python/paddle/v2/__init__.py @@ -22,17 +22,13 @@ import data_type import topology import networks import evaluator -from . import dataset -from . import reader from . import plot import attr import op import pooling import inference import networks -import minibatch import plot -import image import paddle.trainer.config_parser as cp __all__ = [ @@ -48,14 +44,11 @@ __all__ = [ 'data_type', 'attr', 'pooling', - 'dataset', - 'reader', 'topology', 'networks', 'infer', 'plot', 'evaluator', - 'image', 'master', ] @@ -153,4 +146,3 @@ def init(**kwargs): infer = inference.infer -batch = minibatch.batch diff --git a/python/paddle/v2/inference.py b/python/paddle/v2/inference.py index 52f5b947fdec55eea45b9d34eddd576c981fa97c..14b64742fd09bf6c197c5d1aa2354271293df239 100644 --- a/python/paddle/v2/inference.py +++ b/python/paddle/v2/inference.py @@ -15,7 +15,7 @@ import numpy import collections import topology -import minibatch +import paddle import cPickle __all__ = ['infer', 'Inference'] @@ -80,7 +80,7 @@ class Inference(object): for each_sample in input: yield each_sample - reader = minibatch.batch(__reader_impl__, batch_size=batch_size) + reader = paddle.batch(__reader_impl__, batch_size=batch_size) self.__gradient_machine__.start() for data_batch in reader(): diff --git a/python/paddle/v2/layer.py b/python/paddle/v2/layer.py index 6a2bb8d337b7667aa2b1e3ef0815bb80f6e38d6a..a188a03eb3698c972de92c9807f1bdb71a249330 100644 --- a/python/paddle/v2/layer.py +++ b/python/paddle/v2/layer.py @@ -20,7 +20,7 @@ The primary usage shows below. .. code-block:: python - import paddle.v2 as paddle + import paddle img = paddle.layer.data(name='img', type=paddle.data_type.dense_vector(784)) hidden = paddle.layer.fc(input=img, size=200) diff --git a/python/paddle/v2/tests/CMakeLists.txt b/python/paddle/v2/tests/CMakeLists.txt index b4333ed530ce464095ec38d72706949cc464fbe4..46e4feb8e1ce1d12f214f5c49b1b589a46110603 100644 --- a/python/paddle/v2/tests/CMakeLists.txt +++ b/python/paddle/v2/tests/CMakeLists.txt @@ -1,5 +1,4 @@ py_test(test_op SRCS test_op.py) -py_test(test_image SRCS test_image.py) py_test(test_layer SRCS test_layer.py) py_test(test_topology SRCS test_topology.py) py_test(test_rnn_layer SRCS test_rnn_layer.py) diff --git a/python/paddle/v2/tests/test_paramconf_order.py b/python/paddle/v2/tests/test_paramconf_order.py index 264442be182ea69c95b39b3bdb4c389d52eff66e..8320217da2795da756cf12a80f39279182789eef 100644 --- a/python/paddle/v2/tests/test_paramconf_order.py +++ b/python/paddle/v2/tests/test_paramconf_order.py @@ -27,6 +27,7 @@ # limitations under the License. import unittest import math +import paddle.dataset as dataset import paddle.v2 as paddle @@ -40,7 +41,7 @@ def wordemb(inlayer): def train(): - word_dict = paddle.dataset.imikolov.build_dict() + word_dict = dataset.imikolov.build_dict() dict_size = len(word_dict) # Every layer takes integer value of range [0, dict_size) firstword = paddle.layer.data( diff --git a/python/setup.py.in b/python/setup.py.in index 4cb5409524457b7bc5a99c88a0dbbfc8834923fa..d73a3a6a1c41b87efb9600ac59983bd16547ec6a 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -63,18 +63,18 @@ write_version_py(filename='@PADDLE_SOURCE_DIR@/python/paddle/version.py') packages=['paddle', 'paddle.utils', + 'paddle.dataset', + 'paddle.reader', 'paddle.fluid', 'paddle.fluid.proto', 'paddle.fluid.proto.profiler', 'paddle.fluid.layers'] -if '${WITH_FLUID}'== 'OFF': +if '${WITH_FLUID_ONLY}'== 'OFF': packages+=['paddle.proto', 'paddle.trainer', 'paddle.trainer_config_helpers', 'paddle.v2', - 'paddle.v2.dataset', - 'paddle.v2.reader', 'paddle.v2.master', 'paddle.v2.plot', 'py_paddle'] @@ -87,7 +87,7 @@ if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']: # the prefix is sys.prefix which should always be usr paddle_bins = '' -if '${WITH_FLUID}'== 'OFF': +if '${WITH_FLUID_ONLY}'== 'OFF': paddle_bin_dir = 'opt/paddle/bin' paddle_bins = ['${PADDLE_BINARY_DIR}/paddle/trainer/paddle_trainer', '${PADDLE_BINARY_DIR}/paddle/trainer/paddle_merge_model', @@ -95,7 +95,7 @@ if '${WITH_FLUID}'== 'OFF': '${PADDLE_BINARY_DIR}/paddle/scripts/paddle'] package_data={'paddle.fluid': ['core.so']} -if '${WITH_FLUID}'== 'OFF': +if '${WITH_FLUID_ONLY}'== 'OFF': package_data['paddle.v2.master']=['libpaddle_master.so'] package_data['py_paddle']=['*.py','_swig_paddle.so'] @@ -106,7 +106,7 @@ package_dir={ 'paddle.fluid.proto.profiler': '${PADDLE_BINARY_DIR}/paddle/fluid/platform', 'paddle.fluid.proto': '${PADDLE_BINARY_DIR}/paddle/fluid/framework', } -if '${WITH_FLUID}'== 'OFF': +if '${WITH_FLUID_ONLY}'== 'OFF': package_dir['py_paddle']='${PADDLE_SOURCE_DIR}/paddle/py_paddle' diff --git a/tools/codestyle/cpplint_pre_commit.hook b/tools/codestyle/cpplint_pre_commit.hook new file mode 100755 index 0000000000000000000000000000000000000000..94d1e23ce716f7f1d723bad5f1f4c60030f19eb7 --- /dev/null +++ b/tools/codestyle/cpplint_pre_commit.hook @@ -0,0 +1,12 @@ +#!/bin/bash + +TOTAL_ERRORS=0 + +# The trick to remove deleted files: https://stackoverflow.com/a/2413151 +for file in $(git diff --cached --name-status | awk '$1 != "D" {print $2}'); do + cpplint $file; + TOTAL_ERRORS=$(expr $TOTAL_ERRORS + $?); +done + +exit $TOTAL_ERRORS +