Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
1 年多 前同步成功

通知 696
Star 11112
Fork 2696
P
PaddleDetection
提交 65d355a7 编写于 作者: M minqiyang
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add_huber_regression_loss_op 

test=develop
上级 c550e0ce e2130502
展开全部 隐藏空白更改
内联 并排
Showing with 15979 addition and 2813 deletion
CMakeLists.txt
浏览文件 @ 65d355a7
......@@ -54,7 +54,7 @@ option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON)
option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF)
option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF)
option(WITH_TIMER "Compile PaddlePaddle with stats timer" OFF)
option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler" OFF)
option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools" OFF)
option(WITH_DOC "Compile PaddlePaddle with documentation" OFF)
option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF)
option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF)
......@@ -254,6 +254,12 @@ elseif()
set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in MKL only now." FORCE)
endif()
if (WITH_PROFILER)
find_package(Gperftools REQUIRED)
include_directories(${GPERFTOOLS_INCLUDE_DIR})
add_definitions(-DWITH_GPERFTOOLS)
endif()
include(generic) # simplify cmake module
include(package) # set paddle packages
include(ccache) # set ccache for compilation
......
README.md
浏览文件 @ 65d355a7
......@@ -2,8 +2,8 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1)
### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### Install Latest Stable Release:
```
# Linux CPU
......@@ -27,9 +27,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.1.0.post87
pip install paddlepaddle-gpu==1.2.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.1.0.post85
pip install paddlepaddle-gpu==1.2.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
......@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==1.1.0.post85
## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website.
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation.
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation.
- [Deep Learning 101](https://github.com/PaddlePaddle/book)
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html)
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
You can run distributed training jobs on MPI clusters.
- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html)
- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
Our new API enables much shorter programs.
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html)
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
We appreciate your contributions!
......
cmake/FindGperftools.cmake 0 → 100644
浏览文件 @ 65d355a7
# Tries to find Gperftools.
#
# Usage of this module as follows:
#
# find_package(Gperftools)
#
# Variables used by this module, they can change the default behaviour and need
# to be set before calling find_package:
#
# Gperftools_ROOT_DIR Set this variable to the root installation of
# Gperftools if the module has problems finding
# the proper installation path.
#
# Variables defined by this module:
#
# GPERFTOOLS_FOUND System has Gperftools libs/headers
# GPERFTOOLS_LIBRARIES The Gperftools libraries (tcmalloc & profiler)
# GPERFTOOLS_INCLUDE_DIR The location of Gperftools headers
find_library(GPERFTOOLS_TCMALLOC
NAMES tcmalloc
HINTS ${Gperftools_ROOT_DIR}/lib)
find_library(GPERFTOOLS_PROFILER
NAMES profiler
HINTS ${Gperftools_ROOT_DIR}/lib)
find_library(GPERFTOOLS_TCMALLOC_AND_PROFILER
NAMES tcmalloc_and_profiler
HINTS ${Gperftools_ROOT_DIR}/lib)
find_path(GPERFTOOLS_INCLUDE_DIR
NAMES gperftools/heap-profiler.h
HINTS ${Gperftools_ROOT_DIR}/include)
set(GPERFTOOLS_LIBRARIES ${GPERFTOOLS_TCMALLOC_AND_PROFILER})
include(FindPackageHandleStandardArgs)
find_package_handle_standard_args(
Gperftools
DEFAULT_MSG
GPERFTOOLS_LIBRARIES
GPERFTOOLS_INCLUDE_DIR)
mark_as_advanced(
Gperftools_ROOT_DIR
GPERFTOOLS_TCMALLOC
GPERFTOOLS_PROFILER
GPERFTOOLS_TCMALLOC_AND_PROFILER
GPERFTOOLS_LIBRARIES
GPERFTOOLS_INCLUDE_DIR)
# create IMPORTED targets
if (Gperftools_FOUND AND NOT TARGET gperftools::tcmalloc)
add_library(gperftools::tcmalloc UNKNOWN IMPORTED)
set_target_properties(gperftools::tcmalloc PROPERTIES
IMPORTED_LOCATION ${GPERFTOOLS_TCMALLOC}
INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}")
add_library(gperftools::profiler UNKNOWN IMPORTED)
set_target_properties(gperftools::profiler PROPERTIES
IMPORTED_LOCATION ${GPERFTOOLS_PROFILER}
INTERFACE_INCLUDE_DIRECTORIES "${GPERFTOOLS_INCLUDE_DIR}")
endif()
cmake/configure.cmake
浏览文件 @ 65d355a7
......@@ -86,6 +86,7 @@ endif(NOT WITH_GOLANG)
if(WITH_GPU)
add_definitions(-DPADDLE_WITH_CUDA)
add_definitions(-DEIGEN_USE_GPU)
FIND_PACKAGE(CUDA REQUIRED)
......
cmake/external/gzstream.cmake
浏览文件 @ 65d355a7
......@@ -27,13 +27,14 @@ SET(GZSTREAM_INCLUDE_DIR "${GZSTREAM_INSTALL_DIR}/include/" CACHE PATH "gzstream
ExternalProject_Add(
extern_gzstream
DEPENDS zlib
GIT_REPOSITORY "https://github.com/jacquesqiao/gzstream.git"
GIT_TAG ""
PREFIX ${GZSTREAM_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
BUILD_COMMAND make -j8
BUILD_COMMAND make EXTERN_CPPFLAGS="-I${THIRD_PARTY_PATH}/install/zlib/include" EXTERM_LDFLAGS="-L${THIRD_PARTY_PATH}/install/zlib/lib" -j8
INSTALL_COMMAND mkdir -p ${GZSTREAM_INSTALL_DIR}/lib/ && mkdir -p ${GZSTREAM_INSTALL_DIR}/include/
&& cp ${GZSTREAM_SOURCES_DIR}/src/extern_gzstream/libgzstream.a ${GZSTREAM_INSTALL_DIR}/lib
&& cp -r ${GZSTREAM_SOURCES_DIR}/src/extern_gzstream/gzstream.h ${GZSTREAM_INSTALL_DIR}/include
......
cmake/external/ngraph.cmake
浏览文件 @ 65d355a7
......@@ -32,6 +32,8 @@ IF(NOT ${WITH_NGRAPH})
return()
ENDIF()
INCLUDE(GNUInstallDirs)
INCLUDE(ExternalProject)
SET(NGRAPH_PROJECT "extern_ngraph")
......@@ -40,10 +42,14 @@ SET(NGRAPH_GIT_TAG "f9fd9d4cc318dc59dd4b68448e7fbb5f67a28bd0")
SET(NGRAPH_SOURCES_DIR ${THIRD_PARTY_PATH}/ngraph)
SET(NGRAPH_INSTALL_DIR ${THIRD_PARTY_PATH}/install/ngraph)
SET(NGRAPH_INC_DIR ${NGRAPH_INSTALL_DIR}/include)
SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
SET(NGRAPH_SHARED_LIB_NAME libngraph.so.${NGRAPH_VERSION})
SET(NGRAPH_CPU_LIB_NAME libcpu_backend.so)
SET(NGRAPH_TBB_LIB_NAME libtbb.so.2)
SET(NGRAPH_GIT_REPO "https://github.com/NervanaSystems/ngraph.git")
SET(NGRAPH_SHARED_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME})
SET(NGRAPH_CPU_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME})
SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
ExternalProject_Add(
${NGRAPH_PROJECT}
......@@ -63,18 +69,6 @@ ExternalProject_Add(
CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/lib
)
if(UNIX AND NOT APPLE)
include(GNUInstallDirs)
SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR})
else()
SET(NGRAPH_LIB_DIR ${NGRAPH_INSTALL_DIR}/lib)
endif()
MESSAGE(STATUS "nGraph lib will be installed at: ${NGRAPH_LIB_DIR}")
SET(NGRAPH_SHARED_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_SHARED_LIB_NAME})
SET(NGRAPH_CPU_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_CPU_LIB_NAME})
SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
# Workaround for nGraph expecting mklml to be in mkldnn install directory.
ExternalProject_Add_Step(
${NGRAPH_PROJECT}
......
cmake/generic.cmake
浏览文件 @ 65d355a7
......@@ -110,6 +110,14 @@ function(find_fluid_modules TARGET_NAME)
endif()
endfunction(find_fluid_modules)
function(common_link TARGET_NAME)
if (WITH_PROFILER)
target_link_libraries(${TARGET_NAME} gperftools::profiler)
endif()
endfunction()
# find all third_party modules is used for paddle static library
# for reduce the dependency when building the inference libs.
set_property(GLOBAL PROPERTY FLUID_THIRD_PARTY)
......@@ -274,6 +282,7 @@ function(cc_library TARGET_NAME)
endif()
target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
common_link(${TARGET_NAME})
endif()
# cpplint code style
......@@ -340,6 +349,7 @@ function(cc_binary TARGET_NAME)
if(cc_binary_DEPS)
target_link_libraries(${TARGET_NAME} ${cc_binary_DEPS})
add_dependencies(${TARGET_NAME} ${cc_binary_DEPS})
common_link(${TARGET_NAME})
endif()
endfunction(cc_binary)
......@@ -362,6 +372,7 @@ function(cc_test TARGET_NAME)
target_link_libraries(${TARGET_NAME} ${win32_deps})
endif(WIN32)
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
common_link(${TARGET_NAME})
add_test(NAME ${TARGET_NAME}
COMMAND ${TARGET_NAME} ${cc_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
......@@ -420,6 +431,7 @@ function(nv_binary TARGET_NAME)
if(nv_binary_DEPS)
target_link_libraries(${TARGET_NAME} ${nv_binary_DEPS})
add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
common_link(${TARGET_NAME})
endif()
endif()
endfunction(nv_binary)
......@@ -433,6 +445,7 @@ function(nv_test TARGET_NAME)
cuda_add_executable(${TARGET_NAME} ${nv_test_SRCS})
target_link_libraries(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
common_link(${TARGET_NAME})
add_test(${TARGET_NAME} ${TARGET_NAME})
if (nv_test_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
......@@ -499,6 +512,7 @@ function(hip_binary TARGET_NAME)
if(hip_binary_DEPS)
target_link_libraries(${TARGET_NAME} ${hip_binary_DEPS})
add_dependencies(${TARGET_NAME} ${hip_binary_DEPS})
common_link(${TARGET_NAME})
endif()
endif()
endfunction(hip_binary)
......@@ -518,6 +532,7 @@ function(hip_test TARGET_NAME)
set_target_properties(${TARGET_NAME} PROPERTIES LINKER_LANGUAGE HIP)
target_link_libraries(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
add_dependencies(${TARGET_NAME} ${hip_test_DEPS} paddle_gtest_main memory gtest gflags)
common_link(${TARGET_NAME})
add_test(${TARGET_NAME} ${TARGET_NAME})
endif()
endfunction(hip_test)
......@@ -560,6 +575,7 @@ function(go_library TARGET_NAME)
endif()
if(go_library_DEPS)
add_dependencies(${TARGET_NAME} ${go_library_DEPS})
common_link(${TARGET_NAME})
endif(go_library_DEPS)
# The "source file" of the library is `${dummyfile}` which never
......
cmake/inference_lib.cmake
浏览文件 @ 65d355a7
......@@ -129,6 +129,15 @@ if (WITH_MKLDNN)
)
endif ()
if (WITH_NGRAPH)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/ngraph")
copy(ngraph_lib
SRCS ${NGRAPH_INC_DIR} ${NGRAPH_LIB_DIR}
DSTS ${dst_dir} ${dst_dir}
DEPS ngraph
)
endif ()
if (NOT WIN32)
if (NOT MOBILE_INFERENCE AND NOT RPI)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/snappy")
......
cmake/operators.cmake
浏览文件 @ 65d355a7
......@@ -166,6 +166,8 @@ function(op_library TARGET)
# Append first implemented MKLDNN activation operator
if (${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(relu, MKLDNN);\n")
elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op")
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n")
else()
file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(${TARGET}, MKLDNN);\n")
endif()
......
paddle/fluid/API.spec
浏览文件 @ 65d355a7
......@@ -32,6 +32,13 @@ paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.c
paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy) -> None
paddle.fluid.create_lod_tensor ArgSpec(args=['data', 'recursive_seq_lens', 'place'], varargs=None, keywords=None, defaults=None)
paddle.fluid.create_random_int_lodtensor ArgSpec(args=['recursive_seq_lens', 'base_shape', 'place', 'low', 'high'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DataFeedDesc.__init__ ArgSpec(args=['self', 'proto_file'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DataFeedDesc.desc ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DataFeedDesc.set_batch_size ArgSpec(args=['self', 'batch_size'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DataFeedDesc.set_dense_slots ArgSpec(args=['self', 'dense_slots_name'], varargs=None, keywords=None, defaults=None)
paddle.fluid.DataFeedDesc.set_use_slots ArgSpec(args=['self', 'use_slots_name'], varargs=None, keywords=None, defaults=None)
paddle.fluid.AsyncExecutor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.AsyncExecutor.run ArgSpec(args=['self', 'program', 'data_feed', 'filelist', 'thread_num', 'fetch', 'debug'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.io.save_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.io.save_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.io.save_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None))
......@@ -59,6 +66,7 @@ paddle.fluid.layers.linear_chain_crf ArgSpec(args=['input', 'label', 'param_attr
paddle.fluid.layers.crf_decoding ArgSpec(args=['input', 'param_attr', 'label'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.cos_sim ArgSpec(args=['X', 'Y'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.cross_entropy ArgSpec(args=['input', 'label', 'soft_label', 'ignore_index'], varargs=None, keywords=None, defaults=(False, -100))
paddle.fluid.layers.bpr_loss ArgSpec(args=['input', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.square_error_cost ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', 'num_chunk_types', 'excluded_chunk_types'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))
......@@ -69,7 +77,7 @@ paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name']
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.pool3d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name', 'exclusive'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None, True))
paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False))
paddle.fluid.layers.batch_norm ArgSpec(args=['input', 'act', 'is_test', 'momentum', 'epsilon', 'param_attr', 'bias_attr', 'data_layout', 'in_place', 'name', 'moving_mean_name', 'moving_variance_name', 'do_model_average_for_mean_and_var', 'fuse_with_relu', 'use_global_stats'], varargs=None, keywords=None, defaults=(None, False, 0.9, 1e-05, None, None, 'NCHW', False, None, None, None, False, False, False))
paddle.fluid.layers.beam_search_decode ArgSpec(args=['ids', 'scores', 'beam_size', 'end_id', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
......@@ -175,7 +183,7 @@ paddle.fluid.layers.clip ArgSpec(args=['x', 'min', 'max', 'name'], varargs=None,
paddle.fluid.layers.clip_by_norm ArgSpec(args=['x', 'max_norm', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None))
paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'ignore_index', 'name'], varargs=None, keywords=None, defaults=(-100, None))
paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
......@@ -187,6 +195,10 @@ paddle.fluid.layers.grid_sampler ArgSpec(args=['x', 'grid', 'name'], varargs=Non
paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.bilinear_tensor_product ArgSpec(args=['x', 'y', 'size', 'act', 'name', 'param_attr', 'bias_attr'], varargs=None, keywords=None, defaults=(None, None, None, None))
paddle.fluid.layers.merge_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.get_tensor_from_selected_rows ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.lstm ArgSpec(args=['input', 'init_h', 'init_c', 'max_len', 'hidden_size', 'num_layers', 'dropout_prob', 'is_bidirec', 'is_test', 'name', 'default_initializer', 'seed'], varargs=None, keywords=None, defaults=(0.0, False, False, None, None, -1))
paddle.fluid.layers.psroi_pool ArgSpec(args=['input', 'rois', 'output_channels', 'spatial_scale', 'pooled_height', 'pooled_width', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
......@@ -291,6 +303,7 @@ paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'i
paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None))
paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.yolov3_loss ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'class_num', 'ignore_thresh', 'loss_weight_xy', 'loss_weight_wh', 'loss_weight_conf_target', 'loss_weight_conf_notarget', 'loss_weight_class', 'name'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None))
paddle.fluid.layers.accuracy ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None))
paddle.fluid.layers.auc ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1))
paddle.fluid.layers.exponential_decay ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,))
......@@ -411,3 +424,17 @@ paddle.fluid.Scope.drop_kids drop_kids(self: paddle.fluid.core.Scope) -> None
paddle.fluid.Scope.find_var find_var(self: paddle.fluid.core.Scope, arg0: unicode) -> paddle.fluid.core.Variable
paddle.fluid.Scope.new_scope new_scope(self: paddle.fluid.core.Scope) -> paddle.fluid.core.Scope
paddle.fluid.Scope.var var(self: paddle.fluid.core.Scope, arg0: unicode) -> paddle.fluid.core.Variable
paddle.reader.map_readers ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None)
paddle.reader.buffered ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None)
paddle.reader.compose ArgSpec(args=[], varargs='readers', keywords='kwargs', defaults=None)
paddle.reader.chain ArgSpec(args=[], varargs='readers', keywords=None, defaults=None)
paddle.reader.shuffle ArgSpec(args=['reader', 'buf_size'], varargs=None, keywords=None, defaults=None)
paddle.reader.firstn ArgSpec(args=['reader', 'n'], varargs=None, keywords=None, defaults=None)
paddle.reader.xmap_readers ArgSpec(args=['mapper', 'reader', 'process_num', 'buffer_size', 'order'], varargs=None, keywords=None, defaults=(False,))
paddle.reader.PipeReader.__init__ ArgSpec(args=['self', 'command', 'bufsize', 'file_type'], varargs=None, keywords=None, defaults=(8192, 'plain'))
paddle.reader.PipeReader.get_line ArgSpec(args=['self', 'cut_lines', 'line_break'], varargs=None, keywords=None, defaults=(True, '\n'))
paddle.reader.multiprocess_reader ArgSpec(args=['readers', 'use_pipe', 'queue_size'], varargs=None, keywords=None, defaults=(True, 1000))
paddle.reader.Fake.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.reader.creator.np_array ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
paddle.reader.creator.text_file ArgSpec(args=['path'], varargs=None, keywords=None, defaults=None)
paddle.reader.creator.recordio ArgSpec(args=['paths', 'buf_size'], varargs=None, keywords=None, defaults=(100,))
paddle/fluid/CMakeLists.txt
浏览文件 @ 65d355a7
add_subdirectory(memory)
add_subdirectory(platform)
add_subdirectory(framework)
add_subdirectory(imperative)
add_subdirectory(operators)
add_subdirectory(string)
add_subdirectory(recordio)
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ 65d355a7
......@@ -34,6 +34,7 @@ add_subdirectory(ir)
add_subdirectory(details)
# ddim lib
proto_library(framework_proto SRCS framework.proto)
proto_library(async_executor_proto SRCS data_feed.proto)
cc_library(ddim SRCS ddim.cc DEPS eigen3 boost)
cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
......@@ -71,6 +72,8 @@ cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory)
cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
cc_test(reader_test SRCS reader_test.cc DEPS reader)
......@@ -117,8 +120,9 @@ cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context)
cc_library(op_kernel_type SRCS op_kernel_type.cc DEPS device_context place)
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler transfer_scope_cache)
shape_inference data_transform lod_tensor profiler transfer_scope_cache op_kernel_type)
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)
......@@ -126,16 +130,19 @@ cc_library(version SRCS version.cc)
cc_test(version_test SRCS version_test.cc DEPS version)
cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS shape_inference op_info operator glog version)
cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto)
if(NOT WIN32)
cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler)
endif(NOT WIN32)
if(WITH_NGRAPH)
if(NOT WIN32)
cc_library(ngraph_bridge SRCS ngraph_bridge.cc DEPS operator framework_proto ngraph)
cc_library(ngraph_operator SRCS ngraph_operator.cc DEPS ngraph_bridge operator op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler ngraph)
endif(NOT WIN32)
endif(WITH_NGRAPH)
cc_library(op_registry SRCS op_registry.cc DEPS op_proto_maker op_info operator glog proto_desc)
nv_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
py_proto_compile(framework_py_proto SRCS framework.proto)
py_proto_compile(framework_py_proto SRCS framework.proto data_feed.proto)
# Generate an empty __init__.py to make framework_py_proto as a valid python module.
add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
add_dependencies(framework_py_proto framework_py_proto_init)
......@@ -157,27 +164,37 @@ endif(NOT WIN32)
cc_library(lod_rank_table SRCS lod_rank_table.cc DEPS lod_tensor)
cc_library(feed_fetch_method SRCS feed_fetch_method.cc DEPS lod_tensor scope glog)
cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor)
cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
if(WITH_DISTRIBUTE)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method sendrecvop_grpc cares grpc++_unsecure grpc_unsecure gpr graph_to_program_pass variable_helper)
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
else()
if(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph_operator)
else(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass)
endif(NOT WIN32)
if(WITH_NGRAPH)
if(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass ngraph ngraph_operator variable_helper)
else(NOT WIN32)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif(NOT WIN32)
else(WITH_NGRAPH)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif(WITH_NGRAPH)
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif()
target_link_libraries(executor garbage_collector)
cc_library(parallel_executor SRCS parallel_executor.cc DEPS
threaded_ssa_graph_executor scope_buffered_ssa_graph_executor
graph build_strategy
fast_threaded_ssa_graph_executor)
fast_threaded_ssa_graph_executor variable_helper)
cc_library(async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc executor_thread_worker.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass async_executor_proto variable_helper)
cc_test(data_feed_test SRCS data_feed_test.cc DEPS async_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
......@@ -185,7 +202,7 @@ cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
cc_library(selected_rows SRCS selected_rows.cc DEPS tensor)
cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto op_kernel_type)
cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
cc_test(tuple_test SRCS tuple_test.cc )
......
paddle/fluid/framework/async_executor.cc 0 → 100644
浏览文件 @ 65d355a7
/* 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/async_executor.h"
#include "google/protobuf/io/zero_copy_stream_impl.h"
#include "google/protobuf/message.h"
#include "google/protobuf/text_format.h"
#include "gflags/gflags.h"
#include "paddle/fluid/framework/data_feed_factory.h"
#include "paddle/fluid/framework/executor_thread_worker.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/pybind/pybind.h"
namespace paddle {
namespace framework {
AsyncExecutor::AsyncExecutor(Scope* scope, const platform::Place& place)
: root_scope_(scope), place_(place) {}
void AsyncExecutor::CreateThreads(
ExecutorThreadWorker* worker, const ProgramDesc& main_program,
const std::shared_ptr<DataFeed>& reader,
const std::vector<std::string>& fetch_var_names, Scope* root_scope,
const int thread_index, const bool debug) {
worker->SetThreadId(thread_index);
worker->SetDebug(debug);
worker->SetRootScope(root_scope);
worker->CreateThreadResource(main_program, place_);
worker->SetDataFeed(reader);
worker->SetFetchVarNames(fetch_var_names);
worker->BindingDataFeedMemory();
}
void PrepareReaders(std::vector<std::shared_ptr<DataFeed>>& readers, // NOLINT
const int thread_num, const DataFeedDesc& data_feed_desc,
const std::vector<std::string>& filelist) {
readers.resize(thread_num);
for (size_t i = 0; i < readers.size(); ++i) {
readers[i] = DataFeedFactory::CreateDataFeed(data_feed_desc.name());
readers[i]->Init(data_feed_desc); // set batch_size and queue_size here
}
readers[0]->SetFileList(filelist);
}
void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
const std::string& data_feed_desc_str,
const std::vector<std::string>& filelist,
const int thread_num,
const std::vector<std::string>& fetch_var_names,
const bool debug) {
std::vector<std::thread> threads;
auto& block = main_program.Block(0);
for (auto var_name : fetch_var_names) {
auto var_desc = block.FindVar(var_name);
auto shapes = var_desc->GetShape();
PADDLE_ENFORCE(shapes[shapes.size() - 1] == 1,
"var %s: Fetched var has wrong shape, "
"only variables with the last dimension size 1 supported",
var_name);
}
DataFeedDesc data_feed_desc;
google::protobuf::TextFormat::ParseFromString(data_feed_desc_str,
&data_feed_desc);
int actual_thread_num = thread_num;
int file_cnt = filelist.size();
PADDLE_ENFORCE(file_cnt > 0, "File list cannot be empty");
if (actual_thread_num > file_cnt) {
VLOG(1) << "Thread num = " << thread_num << ", file num = " << file_cnt
<< ". Changing thread_num = " << file_cnt;
actual_thread_num = file_cnt;
}
/*
readerDesc: protobuf description for reader initlization
argument: class_name, batch_size, use_slot, queue_size, buffer_size,
padding_index
reader:
1) each thread has a reader, reader will read input data and
put it into input queue
2) each reader has a Next() iterface, that can fetch an instance
from the input queue
*/
// todo: should be factory method for creating datafeed
std::vector<std::shared_ptr<DataFeed>> readers;
PrepareReaders(readers, actual_thread_num, data_feed_desc, filelist);
std::vector<std::shared_ptr<ExecutorThreadWorker>> workers;
workers.resize(actual_thread_num);
for (auto& worker : workers) {
worker.reset(new ExecutorThreadWorker);
}
// prepare thread resource here
for (int thidx = 0; thidx < actual_thread_num; ++thidx) {
CreateThreads(workers[thidx].get(), main_program, readers[thidx],
fetch_var_names, root_scope_, thidx, debug);
}
// start executing ops in multiple threads
for (int thidx = 0; thidx < actual_thread_num; ++thidx) {
threads.push_back(
std::thread(&ExecutorThreadWorker::TrainFiles, workers[thidx].get()));
}
for (auto& th : threads) {
th.join();
}
root_scope_->DropKids();
return;
}
} // einit_modelnd namespace framework
} // end namespace paddle
paddle/fluid/framework/async_executor.h 0 → 100644
浏览文件 @ 65d355a7
/* 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 <map>
#include <memory>
#include <mutex> // NOLINT
#include <set>
#include <string>
#include <thread> // NOLINT
#include <typeinfo>
#include <vector>
#include "paddle/fluid/framework/data_feed.pb.h"
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/executor_thread_worker.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
namespace paddle {
namespace framework {
class AsyncExecutor {
public:
AsyncExecutor(Scope* scope, const platform::Place& place);
virtual ~AsyncExecutor() {}
void RunFromFile(const ProgramDesc& main_program,
const std::string& data_feed_desc_str,
const std::vector<std::string>& filelist,
const int thread_num,
const std::vector<std::string>& fetch_names,
const bool debug = false);
private:
void CreateThreads(ExecutorThreadWorker* worker,
const ProgramDesc& main_program,
const std::shared_ptr<DataFeed>& reader,
const std::vector<std::string>& fetch_var_names,
Scope* root_scope, const int thread_index,
const bool debug);
public:
Scope* root_scope_;
platform::Place place_;
};
} // namespace framework
} // namespace paddle
paddle/fluid/framework/data_feed.cc 0 → 100644
浏览文件 @ 65d355a7
/* 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 "google/protobuf/io/zero_copy_stream_impl.h"
#include "google/protobuf/message.h"
#include "google/protobuf/text_format.h"
#include "gflags/gflags.h"
#include "paddle/fluid/framework/data_feed.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
namespace paddle {
namespace framework {
std::vector<std::string> DataFeed::filelist_;
size_t DataFeed::file_idx_;
std::mutex DataFeed::mutex_for_pick_file_;
bool DataFeed::finish_set_filelist_;
void DataFeed::AddFeedVar(Variable* var, const std::string& name) {
CheckInit();
for (size_t i = 0; i < use_slots_.size(); ++i) {
if (name == use_slots_[i]) {
feed_vec_[i] = var->GetMutable<LoDTensor>();
}
}
}
bool DataFeed::SetFileList(const std::vector<std::string>& files) {
std::unique_lock<std::mutex> lock(mutex_for_pick_file_);
CheckInit();
if (finish_set_filelist_) {
VLOG(3) << "info: you have set the filelist.";
return false;
}
PADDLE_ENFORCE(files.size(), "You have set an empty filelist.");
filelist_.assign(files.begin(), files.end());
file_idx_ = 0;
finish_set_filelist_ = true;
return true;
}
void DataFeed::SetBatchSize(int batch_size) {
PADDLE_ENFORCE(batch_size > 0, "Illegal batch size: %d.", batch_size);
default_batch_size_ = batch_size;
}
bool DataFeed::PickOneFile(std::string* filename) {
std::unique_lock<std::mutex> lock(mutex_for_pick_file_);
if (file_idx_ == filelist_.size()) {
return false;
}
*filename = filelist_[file_idx_++];
return true;
}
void DataFeed::CheckInit() {
PADDLE_ENFORCE(finish_init_, "Initialization did not succeed.");
}
void DataFeed::CheckSetFileList() {
PADDLE_ENFORCE(finish_set_filelist_, "Set filelist did not succeed.");
}
void DataFeed::CheckStart() {
PADDLE_ENFORCE(finish_start_, "Datafeed has not started running yet.");
}
template <typename T>
void PrivateQueueDataFeed<T>::SetQueueSize(int queue_size) {
PADDLE_ENFORCE(queue_size > 0, "Illegal queue size: %d.", queue_size);
queue_size_ = queue_size;
queue_ = std::unique_ptr<paddle::operators::reader::BlockingQueue<T>>(
new paddle::operators::reader::BlockingQueue<T>(queue_size_));
}
template <typename T>
bool PrivateQueueDataFeed<T>::Start() {
CheckSetFileList();
read_thread_ = std::thread(&PrivateQueueDataFeed::ReadThread, this);
read_thread_.detach();
finish_start_ = true;
return true;
}
template <typename T>
void PrivateQueueDataFeed<T>::ReadThread() {
std::string filename;
while (PickOneFile(&filename)) {
file_.open(filename.c_str()); // is_text_feed
PADDLE_ENFORCE(file_.good(), "Open file<%s> fail.", filename.c_str());
T instance;
while (ParseOneInstance(&instance)) {
queue_->Send(instance);
}
file_.close();
}
queue_->Close();
}
template <typename T>
int PrivateQueueDataFeed<T>::Next() {
CheckStart();
int index = 0;
T instance;
T ins_vec;
while (index < default_batch_size_) {
if (!queue_->Receive(&instance)) {
break;
}
AddInstanceToInsVec(&ins_vec, instance, index++);
}
batch_size_ = index;
if (batch_size_ != 0) {
PutToFeedVec(ins_vec);
}
return batch_size_;
}
#ifdef _WIN32
template class PrivateQueueDataFeed<std::vector<MultiSlotType>>;
#endif
void MultiSlotDataFeed::Init(
const paddle::framework::DataFeedDesc& data_feed_desc) {
finish_init_ = false;
finish_set_filelist_ = false;
finish_start_ = false;
PADDLE_ENFORCE(data_feed_desc.has_multi_slot_desc(),
"Multi_slot_desc has not been set.");
paddle::framework::MultiSlotDesc multi_slot_desc =
data_feed_desc.multi_slot_desc();
SetBatchSize(data_feed_desc.batch_size());
SetQueueSize(data_feed_desc.batch_size());
size_t all_slot_num = multi_slot_desc.slots_size();
all_slots_.resize(all_slot_num);
all_slots_type_.resize(all_slot_num);
use_slots_index_.resize(all_slot_num);
use_slots_.clear();
use_slots_is_dense_.clear();
for (size_t i = 0; i < all_slot_num; ++i) {
const auto& slot = multi_slot_desc.slots(i);
all_slots_[i] = slot.name();
all_slots_type_[i] = slot.type();
use_slots_index_[i] = slot.is_used() ? use_slots_.size() : -1;
if (slot.is_used()) {
use_slots_.push_back(all_slots_[i]);
use_slots_is_dense_.push_back(slot.is_dense());
}
}
feed_vec_.resize(use_slots_.size());
finish_init_ = true;
}
bool MultiSlotDataFeed::CheckFile(const char* filename) {
CheckInit(); // get info of slots
std::ifstream fin(filename);
if (!fin.good()) {
VLOG(1) << "error: open file<" << filename << "> fail";
return false;
}
std::string line;
int instance_cout = 0;
std::string all_slots_alias = "";
for (const auto& alias : all_slots_) {
all_slots_alias += alias + " ";
}
std::string use_slots_alias = "";
for (const auto& alias : use_slots_) {
use_slots_alias += alias + " ";
}
VLOG(3) << "total slots num: " << all_slots_.size();
VLOG(3) << "total slots alias: " << all_slots_alias;
VLOG(3) << "used slots num: " << use_slots_.size();
VLOG(3) << "used slots alias: " << use_slots_alias;
while (getline(fin, line)) {
++instance_cout;
const char* str = line.c_str();
char* endptr = const_cast<char*>(str);
int len = line.length();
for (size_t i = 0; i < all_slots_.size(); ++i) {
int num = strtol(endptr, &endptr, 10);
if (num < 0) {
VLOG(0) << "error: the number of ids is a negative number: " << num;
VLOG(0) << "please check line<" << instance_cout << "> in file<"
<< filename << ">";
return false;
} else if (num == 0) {
VLOG(0)
<< "error: the number of ids can not be zero, you need "
"padding it in data generator; or if there is something wrong"
" with the data, please check if the data contains unresolvable "
"characters.";
VLOG(0) << "please check line<" << instance_cout << "> in file<"
<< filename << ">";
return false;
} else if (errno == ERANGE || num > INT_MAX) {
VLOG(0) << "error: the number of ids greater than INT_MAX";
VLOG(0) << "please check line<" << instance_cout << "> in file<"
<< filename << ">";
return false;
}
if (all_slots_type_[i] == "float") {
for (int i = 0; i < num; ++i) {
strtof(endptr, &endptr);
if (errno == ERANGE) {
VLOG(0) << "error: the value is out of the range of "
"representable values for float";
VLOG(0) << "please check line<" << instance_cout << "> in file<"
<< filename << ">";
return false;
}
if (i + 1 != num && endptr - str == len) {
VLOG(0) << "error: there is a wrong with the number of ids.";
VLOG(0) << "please check line<" << instance_cout << "> in file<"
<< filename << ">";
return false;
}
}
} else if (all_slots_type_[i] == "uint64") {
for (int i = 0; i < num; ++i) {
strtoull(endptr, &endptr, 10);
if (errno == ERANGE) {
VLOG(0) << "error: the value is out of the range of "
"representable values for uint64_t";
VLOG(0) << "please check line<" << instance_cout << "> in file<"
<< filename << ">";
return false;
}
if (i + 1 != num && endptr - str == len) {
VLOG(0) << "error: there is a wrong with the number of ids.";
VLOG(0) << "please check line<" << instance_cout << "> in file<"
<< filename << ">";
return false;
}
}
} else {
VLOG(0) << "error: this type<" << all_slots_type_[i]
<< "> is not supported";
return false;
}
}
// It may be added '\t' character to the end of the output of reduce
// task when processes data by Hadoop(when the output of the reduce
// task of Hadoop has only one field, it will add a '\t' at the end
// of the line by default, and you can use this option to avoid it:
// `-D mapred.textoutputformat.ignoreseparator=true`), which does
// not affect the correctness of the data. Therefore, it should be
// judged that the data is not normal when the end of each line of
// data contains characters which are not spaces.
while (endptr - str != len) {
if (!isspace(*(endptr++))) {
VLOG(0)
<< "error: there is some extra characters at the end of the line.";
VLOG(0) << "please check line<" << instance_cout << "> in file<"
<< filename << ">";
return false;
}
}
}
VLOG(3) << "instances cout: " << instance_cout;
VLOG(3) << "The file format is correct";
return true;
}
bool MultiSlotDataFeed::ParseOneInstance(std::vector<MultiSlotType>* instance) {
std::string line;
if (getline(file_, line)) {
int use_slots_num = use_slots_.size();
instance->resize(use_slots_num);
// parse line
const char* str = line.c_str();
char* endptr = const_cast<char*>(str);
int pos = 0;
for (size_t i = 0; i < use_slots_index_.size(); ++i) {
int idx = use_slots_index_[i];
int num = strtol(&str[pos], &endptr, 10);
PADDLE_ENFORCE(
num,
"The number of ids can not be zero, you need padding "
"it in data generator; or if there is something wrong with "
"the data, please check if the data contains unresolvable "
"characters.\nplease check this error line: %s",
str);
if (idx != -1) {
(*instance)[idx].Init(all_slots_type_[i]);
if ((*instance)[idx].GetType()[0] == 'f') { // float
for (int j = 0; j < num; ++j) {
float feasign = strtof(endptr, &endptr);
(*instance)[idx].AddValue(feasign);
}
} else if ((*instance)[idx].GetType()[0] == 'u') { // uint64
for (int j = 0; j < num; ++j) {
uint64_t feasign = (uint64_t)strtoull(endptr, &endptr, 10);
(*instance)[idx].AddValue(feasign);
}
}
pos = endptr - str;
} else {
for (int j = 0; j <= num; ++j) {
pos = line.find_first_of(' ', pos + 1);
}
}
}
} else {
return false;
}
return true;
}
void MultiSlotDataFeed::AddInstanceToInsVec(
std::vector<MultiSlotType>* ins_vec,
const std::vector<MultiSlotType>& instance, int index) {
if (index == 0) {
ins_vec->resize(instance.size());
for (size_t i = 0; i < instance.size(); ++i) {
(*ins_vec)[i].Init(instance[i].GetType());
(*ins_vec)[i].InitOffset();
}
}
for (size_t i = 0; i < instance.size(); ++i) {
(*ins_vec)[i].AddIns(instance[i]);
}
}
void MultiSlotDataFeed::PutToFeedVec(
const std::vector<MultiSlotType>& ins_vec) {
for (size_t i = 0; i < use_slots_.size(); ++i) {
const auto& type = ins_vec[i].GetType();
const auto& offset = ins_vec[i].GetOffset();
int total_instance = static_cast<int>(offset.back());
if (type[0] == 'f') { // float
const auto& feasign = ins_vec[i].GetFloatData();
float* tensor_ptr = feed_vec_[i]->mutable_data<float>(
{total_instance, 1}, platform::CPUPlace());
memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
} else if (type[0] == 'u') { // uint64
// no uint64_t type in paddlepaddle
const auto& feasign = ins_vec[i].GetUint64Data();
int64_t* tensor_ptr = feed_vec_[i]->mutable_data<int64_t>(
{total_instance, 1}, platform::CPUPlace());
memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
}
LoD data_lod{offset};
feed_vec_[i]->set_lod(data_lod);
if (use_slots_is_dense_[i]) {
int dim = total_instance / batch_size_;
feed_vec_[i]->Resize({batch_size_, dim});
}
}
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/data_feed.h 0 → 100644
浏览文件 @ 65d355a7
/* 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 <fstream>
#include <memory>
#include <mutex> // NOLINT
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "paddle/fluid/framework/data_feed.pb.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/operators/reader/blocking_queue.h"
namespace paddle {
namespace framework {
// DataFeed is the base virtual class for all ohther DataFeeds.
// It is used to read files and parse the data for subsequent trainer.
// Example:
// DataFeed* reader =
// paddle::framework::DataFeedFactory::CreateDataFeed(data_feed_name);
// reader->Init(data_feed_desc); // data_feed_desc is a protobuf object
// reader->SetFileList(filelist);
// const std::vector<std::string> & use_slot_alias =
// reader->GetUseSlotAlias();
// for (auto name: use_slot_alias){ // for binding memory
// reader->AddFeedVar(scope->Var(name), name);
// }
// reader->Start();
// while (reader->Next()) {
// // trainer do something
// }
class DataFeed {
public:
DataFeed() {}
virtual ~DataFeed() {}
virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc) = 0;
virtual bool CheckFile(const char* filename) {
PADDLE_THROW("This function(CheckFile) is not implemented.");
}
// Set filelist for DataFeed.
// Pay attention that it must init all readers before call this function.
// Otherwise, Init() function will init finish_set_filelist_ flag.
virtual bool SetFileList(const std::vector<std::string>& files);
virtual bool Start() = 0;
// The trainer calls the Next() function, and the DataFeed will load a new
// batch to the feed_vec. The return value of this function is the batch
// size of the current batch.
virtual int Next() = 0;
// Get all slots' alias which defined in protofile
virtual const std::vector<std::string>& GetAllSlotAlias() {
return all_slots_;
}
// Get used slots' alias which defined in protofile
virtual const std::vector<std::string>& GetUseSlotAlias() {
return use_slots_;
}
// This function is used for binding feed_vec memory
virtual void AddFeedVar(Variable* var, const std::string& name);
protected:
// The following three functions are used to check if it is executed in this
// order:
// Init() -> SetFileList() -> Start() -> Next()
virtual void CheckInit();
virtual void CheckSetFileList();
virtual void CheckStart();
virtual void SetBatchSize(
int batch); // batch size will be set in Init() function
// This function is used to pick one file from the global filelist(thread
// safe).
virtual bool PickOneFile(std::string* filename);
static std::vector<std::string> filelist_;
static size_t file_idx_;
static std::mutex mutex_for_pick_file_;
// the alias of used slots, and its order is determined by
// data_feed_desc(proto object)
std::vector<std::string> use_slots_;
std::vector<bool> use_slots_is_dense_;
// the alias of all slots, and its order is determined by data_feed_desc(proto
// object)
std::vector<std::string> all_slots_;
std::vector<std::string> all_slots_type_;
std::vector<int>
use_slots_index_; // -1: not used; >=0: the index of use_slots_
// The data read by DataFeed will be stored here
std::vector<LoDTensor*> feed_vec_;
// the batch size defined by user
int default_batch_size_;
// current batch size
int batch_size_;
bool finish_init_;
static bool finish_set_filelist_;
bool finish_start_;
};
// PrivateQueueDataFeed is the base virtual class for ohther DataFeeds.
// It use a read-thread to read file and parse data to a private-queue
// (thread level), and get data from this queue when trainer call Next().
template <typename T>
class PrivateQueueDataFeed : public DataFeed {
public:
PrivateQueueDataFeed() {}
virtual ~PrivateQueueDataFeed() {}
virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc) = 0;
virtual bool Start();
virtual int Next();
protected:
// The thread implementation function for reading file and parse.
virtual void ReadThread();
// This function is used to set private-queue size, and the most
// efficient when the queue size is close to the batch size.
virtual void SetQueueSize(int queue_size);
// The reading and parsing method called in the ReadThread.
virtual bool ParseOneInstance(T* instance) = 0;
// This function is used to put instance to vec_ins
virtual void AddInstanceToInsVec(T* vec_ins, const T& instance,
int index) = 0;
// This function is used to put ins_vec to feed_vec
virtual void PutToFeedVec(const T& ins_vec) = 0;
// The thread for read files
std::thread read_thread_;
// using ifstream one line and one line parse is faster
// than using fread one buffer and one buffer parse.
// for a 601M real data:
// ifstream one line and one line parse: 6034 ms
// fread one buffer and one buffer parse: 7097 ms
std::ifstream file_;
size_t queue_size_;
// The queue for store parsed data
std::unique_ptr<paddle::operators::reader::BlockingQueue<T>> queue_;
};
// This class define the data type of instance(ins_vec) in MultiSlotDataFeed
class MultiSlotType {
public:
MultiSlotType() {}
~MultiSlotType() {}
void Init(const std::string& type) {
CheckType(type);
if (type_[0] == 'f') {
float_feasign_.clear();
} else if (type_[0] == 'u') {
uint64_feasign_.clear();
}
type_ = type;
}
void InitOffset() {
offset_.resize(1);
// LoDTensor' lod is counted from 0, the size of lod
// is one size larger than the size of data.
offset_[0] = 0;
}
const std::vector<size_t>& GetOffset() const { return offset_; }
void AddValue(const float v) {
CheckFloat();
float_feasign_.push_back(v);
}
void AddValue(const uint64_t v) {
CheckUint64();
uint64_feasign_.push_back(v);
}
void AddIns(const MultiSlotType& ins) {
if (ins.GetType()[0] == 'f') { // float
CheckFloat();
auto& vec = ins.GetFloatData();
offset_.push_back(offset_.back() + vec.size());
float_feasign_.insert(float_feasign_.end(), vec.begin(), vec.end());
} else if (ins.GetType()[0] == 'u') { // uint64
CheckUint64();
auto& vec = ins.GetUint64Data();
offset_.push_back(offset_.back() + vec.size());
uint64_feasign_.insert(uint64_feasign_.end(), vec.begin(), vec.end());
}
}
const std::vector<float>& GetFloatData() const { return float_feasign_; }
const std::vector<uint64_t>& GetUint64Data() const { return uint64_feasign_; }
const std::string& GetType() const { return type_; }
private:
void CheckType(const std::string& type) const {
PADDLE_ENFORCE((type == "uint64") || (type == "float"),
"There is no this type<%s>.", type);
}
void CheckFloat() const {
PADDLE_ENFORCE(type_[0] == 'f', "Add %s value to float slot.", type_);
}
void CheckUint64() const {
PADDLE_ENFORCE(type_[0] == 'u', "Add %s value to uint64 slot.", type_);
}
std::vector<float> float_feasign_;
std::vector<uint64_t> uint64_feasign_;
std::string type_;
std::vector<size_t> offset_;
};
// This DataFeed is used to feed multi-slot type data.
// The format of multi-slot type data:
// [n feasign_0 feasign_1 ... feasign_n]*
class MultiSlotDataFeed
: public PrivateQueueDataFeed<std::vector<MultiSlotType>> {
public:
MultiSlotDataFeed() {}
virtual ~MultiSlotDataFeed() {}
virtual void Init(const paddle::framework::DataFeedDesc& data_feed_desc);
virtual bool CheckFile(const char* filename);
protected:
virtual void AddInstanceToInsVec(std::vector<MultiSlotType>* vec_ins,
const std::vector<MultiSlotType>& instance,
int index);
virtual bool ParseOneInstance(std::vector<MultiSlotType>* instance);
virtual void PutToFeedVec(const std::vector<MultiSlotType>& ins_vec);
};
} // namespace framework
} // namespace paddle
paddle/fluid/framework/data_feed.proto 0 → 100644
浏览文件 @ 65d355a7
/* 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. */
syntax = "proto2";
package paddle.framework;
message Slot {
required string name = 1;
required string type = 2;
optional bool is_dense = 3 [ default = false ];
optional bool is_used = 4 [ default = false ];
}
message MultiSlotDesc { repeated Slot slots = 1; }
message DataFeedDesc {
optional string name = 1;
optional int32 batch_size = 2 [ default = 32 ];
optional MultiSlotDesc multi_slot_desc = 3;
}
paddle/fluid/framework/data_feed_factory.cc 0 → 100644
浏览文件 @ 65d355a7
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/data_feed_factory.h"
#include <memory>
#include <string>
#include <unordered_map>
#include "paddle/fluid/framework/data_feed.h"
namespace paddle {
namespace framework {
typedef std::shared_ptr<DataFeed> (*Createdata_feedFunction)();
typedef std::unordered_map<std::string, Createdata_feedFunction> data_feedMap;
data_feedMap g_data_feed_map;
#define REGISTER_DATAFEED_CLASS(data_feed_class) \
namespace { \
std::shared_ptr<DataFeed> Creator_##data_feed_class() { \
return std::shared_ptr<DataFeed>(new data_feed_class); \
} \
class __Registerer_##data_feed_class { \
public: \
__Registerer_##data_feed_class() { \
g_data_feed_map[#data_feed_class] = &Creator_##data_feed_class; \
} \
}; \
__Registerer_##data_feed_class g_registerer_##data_feed_class; \
} // namespace
std::string DataFeedFactory::DataFeedTypeList() {
std::string data_feed_types;
for (auto iter = g_data_feed_map.begin(); iter != g_data_feed_map.end();
++iter) {
if (iter != g_data_feed_map.begin()) {
data_feed_types += ", ";
}
data_feed_types += iter->first;
}
return data_feed_types;
}
std::shared_ptr<DataFeed> DataFeedFactory::CreateDataFeed(
std::string data_feed_class) {
if (g_data_feed_map.count(data_feed_class) < 1) {
exit(-1);
}
return g_data_feed_map[data_feed_class]();
}
REGISTER_DATAFEED_CLASS(MultiSlotDataFeed);
} // namespace framework
} // namespace paddle
paddle/fluid/framework/data_feed_factory.h 0 → 100644
浏览文件 @ 65d355a7
/* 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 <memory>
#include <string>
#include "paddle/fluid/framework/data_feed.h"
namespace paddle {
namespace framework {
class DataFeedFactory {
public:
static std::string DataFeedTypeList();
static std::shared_ptr<DataFeed> CreateDataFeed(std::string data_feed_class);
};
} // namespace framework
} // namespace paddle
paddle/fluid/framework/data_feed_test.cc 0 → 100644
浏览文件 @ 65d355a7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/data_feed.h"
#include <fcntl.h>
#include <chrono> // NOLINT
#include <fstream>
#include <iostream>
#include <map>
#include <mutex> // NOLINT
#include <set>
#include <thread> // NOLINT
#include <utility>
#include <vector>
#include "google/protobuf/io/zero_copy_stream_impl.h"
#include "google/protobuf/text_format.h"
#include "gtest/gtest.h"
#include "paddle/fluid/framework/data_feed_factory.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
paddle::framework::DataFeedDesc load_datafeed_param_from_file(
const char* filename) {
paddle::framework::DataFeedDesc data_feed_desc;
int file_descriptor = open(filename, O_RDONLY);
PADDLE_ENFORCE(file_descriptor != -1, "Can not open %s.", filename);
google::protobuf::io::FileInputStream fileInput(file_descriptor);
google::protobuf::TextFormat::Parse(&fileInput, &data_feed_desc);
close(file_descriptor);
return data_feed_desc;
}
const std::vector<std::string> load_filelist_from_file(const char* filename) {
std::vector<std::string> filelist;
std::ifstream fin(filename);
PADDLE_ENFORCE(fin.good(), "Can not open %s.", filename);
std::string line;
while (getline(fin, line)) {
filelist.push_back(line);
}
fin.close();
return filelist;
}
void GenerateFileForTest(const char* protofile, const char* filelist) {
std::ofstream w_protofile(protofile);
w_protofile << "name: \"MultiSlotDataFeed\"\n"
"batch_size: 2\n"
"multi_slot_desc {\n"
" slots {\n"
" name: \"uint64_sparse_slot\"\n"
" type: \"uint64\"\n"
" is_dense: false\n"
" is_used: true\n"
" }\n"
" slots {\n"
" name: \"float_sparse_slot\"\n"
" type: \"float\"\n"
" is_dense: false\n"
" is_used: true\n"
" }\n"
" slots {\n"
" name: \"uint64_dense_slot\"\n"
" type: \"uint64\"\n"
" is_dense: true\n"
" is_used: true\n"
" }\n"
" slots {\n"
" name: \"float_dense_slot\"\n"
" type: \"float\"\n"
" is_dense: true\n"
" is_used: true\n"
" }\n"
" slots {\n"
" name: \"not_used_slot\"\n"
" type: \"uint64\"\n"
" is_dense: false\n"
" is_used: false\n"
" }\n"
"}";
w_protofile.close();
std::ofstream w_filelist(filelist);
int total_file = 4;
for (int i = 0; i < total_file; ++i) {
std::string filename = "TestMultiSlotDataFeed.data." + std::to_string(i);
w_filelist << filename;
if (i + 1 != total_file) {
w_filelist << std::endl;
}
std::ofstream w_datafile(filename.c_str());
w_datafile << "3 3978 620 82 1 1926.08 1 1926 1 6.02 1 1996\n"
"2 1300 2983353 1 985.211 1 8 1 0.618 1 12\n"
"1 19260827 2 3.14 2.718 1 27 1 2.236 1 28\n";
w_datafile.close();
}
w_filelist.close();
}
class MultiTypeSet {
public:
MultiTypeSet() {
uint64_set_.clear();
float_set_.clear();
}
~MultiTypeSet() {}
void AddValue(uint64_t v) { uint64_set_.insert(v); }
void AddValue(float v) { float_set_.insert(v); }
const std::set<uint64_t>& GetUint64Set() const { return uint64_set_; }
const std::set<float>& GetFloatSet() const { return float_set_; }
private:
std::set<uint64_t> uint64_set_;
std::set<float> float_set_;
};
void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
const paddle::framework::DataFeedDesc& data_feed_desc,
const std::vector<std::string>& filelist,
const int thread_num) {
int used_slot_num = 0;
for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) {
if (data_feed_desc.multi_slot_desc().slots(i).is_used()) {
++used_slot_num;
}
}
reader_elem_set->resize(used_slot_num);
std::vector<std::thread> threads;
std::vector<std::shared_ptr<paddle::framework::DataFeed>> readers;
readers.resize(thread_num);
for (int i = 0; i < thread_num; ++i) {
readers[i] = paddle::framework::DataFeedFactory::CreateDataFeed(
data_feed_desc.name());
readers[i]->Init(data_feed_desc);
}
readers[0]->SetFileList(filelist);
std::mutex mu;
for (int idx = 0; idx < thread_num; ++idx) {
threads.emplace_back(std::thread([&, idx] {
std::unique_ptr<paddle::framework::Scope> scope(
new paddle::framework::Scope());
const auto& multi_slot_desc = data_feed_desc.multi_slot_desc();
std::map<std::string, const paddle::framework::LoDTensor*>
lodtensor_targets;
for (int i = 0; i < multi_slot_desc.slots_size(); ++i) {
const auto& slot = multi_slot_desc.slots(i);
if (slot.is_used()) {
const auto& name = slot.name();
readers[idx]->AddFeedVar(scope->Var(name), name);
lodtensor_targets[name] =
&scope->FindVar(name)->Get<paddle::framework::LoDTensor>();
}
}
readers[idx]->Start();
while (readers[idx]->Next()) {
int index = 0;
for (int k = 0; k < multi_slot_desc.slots_size(); ++k) {
const auto& slot = multi_slot_desc.slots(k);
if (!slot.is_used()) {
continue;
}
const paddle::framework::LoDTensor* tens =
lodtensor_targets[slot.name()];
if (slot.is_dense()) { // dense branch
if (slot.type() == "uint64") {
const int64_t* data = tens->data<int64_t>();
int batch_size = tens->dims()[0];
int dim = tens->dims()[1];
for (int i = 0; i < batch_size; ++i) {
for (int j = 0; j < dim; ++j) {
std::lock_guard<std::mutex> lock(mu);
(*reader_elem_set)[index].AddValue(
(uint64_t)data[i * dim + j]);
}
}
} else if (slot.type() == "float") {
const float* data = tens->data<float>();
int batch_size = tens->dims()[0];
int dim = tens->dims()[1];
for (int i = 0; i < batch_size; ++i) {
for (int j = 0; j < dim; ++j) {
std::lock_guard<std::mutex> lock(mu);
(*reader_elem_set)[index].AddValue(data[i * dim + j]);
}
}
} else {
PADDLE_THROW("Error type in proto file.");
}
} else { // sparse branch
if (slot.type() == "uint64") {
const int64_t* data = tens->data<int64_t>();
for (size_t i = 0; i < tens->NumElements(); ++i) {
std::pair<size_t, size_t> element = tens->lod_element(0, i);
for (size_t j = element.first; j < element.second; ++j) {
std::lock_guard<std::mutex> lock(mu);
(*reader_elem_set)[index].AddValue((uint64_t)data[j]);
}
}
} else if (slot.type() == "float") {
const float* data = tens->data<float>();
for (size_t i = 0; i < tens->NumElements(); ++i) {
std::pair<size_t, size_t> element = tens->lod_element(0, i);
for (size_t j = element.first; j < element.second; ++j) {
std::lock_guard<std::mutex> lock(mu);
(*reader_elem_set)[index].AddValue(data[j]);
}
}
} else {
PADDLE_THROW("Error type in proto file.");
}
} // end sparse branch
++index;
} // end slots loop
} // end while Next()
})); // end anonymous function
}
for (auto& th : threads) {
th.join();
}
}
void CheckIsUnorderedSame(const std::vector<MultiTypeSet>& s1,
const std::vector<MultiTypeSet>& s2) {
EXPECT_EQ(s1.size(), s2.size());
for (size_t i = 0; i < s1.size(); ++i) {
// check for uint64
const std::set<uint64_t>& uint64_s1 = s1[i].GetUint64Set();
const std::set<uint64_t>& uint64_s2 = s2[i].GetUint64Set();
EXPECT_EQ(uint64_s1.size(), uint64_s2.size());
auto uint64_it1 = uint64_s1.begin();
auto uint64_it2 = uint64_s2.begin();
while (uint64_it1 != uint64_s1.end()) {
EXPECT_EQ(*uint64_it1, *uint64_it2);
++uint64_it1;
++uint64_it2;
}
// check for float
const std::set<float>& float_s1 = s1[i].GetFloatSet();
const std::set<float>& float_s2 = s2[i].GetFloatSet();
EXPECT_EQ(float_s1.size(), float_s2.size());
auto float_it1 = float_s1.begin();
auto float_it2 = float_s2.begin();
while (float_it1 != float_s1.end()) {
EXPECT_EQ(*float_it1, *float_it2);
++float_it1;
++float_it2;
}
}
}
void GetElemSetFromFile(std::vector<MultiTypeSet>* file_elem_set,
const paddle::framework::DataFeedDesc& data_feed_desc,
const std::vector<std::string>& filelist) {
int used_slot_num = 0;
for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) {
if (data_feed_desc.multi_slot_desc().slots(i).is_used()) {
++used_slot_num;
}
}
file_elem_set->resize(used_slot_num);
for (const auto& file : filelist) {
std::ifstream fin(file.c_str());
PADDLE_ENFORCE(fin.good(), "Can not open %s.", file.c_str());
while (1) {
bool end_flag = false;
int index = 0;
for (auto i = 0; i < data_feed_desc.multi_slot_desc().slots_size(); ++i) {
int num;
if (fin >> num) {
auto slot = data_feed_desc.multi_slot_desc().slots(i);
auto type = slot.type();
if (type == "uint64") {
while (num--) {
uint64_t feasign;
fin >> feasign;
if (slot.is_used()) {
(*file_elem_set)[index].AddValue(feasign);
}
}
} else if (type == "float") {
while (num--) {
float feasign;
fin >> feasign;
if (slot.is_used()) {
(*file_elem_set)[index].AddValue(feasign);
}
}
} else {
PADDLE_THROW("Error type in proto file.");
}
if (slot.is_used()) {
++index;
}
} else {
end_flag = true;
break;
}
}
if (end_flag) {
break;
}
}
fin.close();
}
}
TEST(DataFeed, MultiSlotUnitTest) {
const char* protofile = "data_feed_desc.prototxt";
const char* filelist_name = "filelist.txt";
GenerateFileForTest(protofile, filelist_name);
const std::vector<std::string> filelist =
load_filelist_from_file(filelist_name);
paddle::framework::DataFeedDesc data_feed_desc =
load_datafeed_param_from_file(protofile);
std::vector<MultiTypeSet> reader_elem_set;
std::vector<MultiTypeSet> file_elem_set;
GetElemSetFromReader(&reader_elem_set, data_feed_desc, filelist, 4);
GetElemSetFromFile(&file_elem_set, data_feed_desc, filelist);
CheckIsUnorderedSame(reader_elem_set, file_elem_set);
}
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @ 65d355a7
......@@ -151,19 +151,22 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
auto out_format =
platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
void* in_data = GetDataFromTensor(in, in_type);
// output tensor has the same dims as input. Reorder don't change dims
out->Resize(in.dims());
auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
auto in_memory = memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
auto out_memory =
memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
if (in_format != out_format) {
void* in_data = GetDataFromTensor(in, in_type);
auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
platform::Reorder(in_memory, out_memory);
auto in_memory =
memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
auto out_memory =
memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
platform::Reorder(in_memory, out_memory);
} else {
out->ShareDataWith(in);
}
out->set_layout(out_layout);
// reset format since the out tensor will be feed to non-MKLDNN OPkernel
out->set_format(memory::format::format_undef);
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ 65d355a7
......@@ -15,14 +15,26 @@ cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_ro
if(WITH_GPU)
nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim dynload_cuda)
if(WITH_DISTRIBUTE)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim dynload_cuda selected_rows_functor sendrecvop_grpc)
else()
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim dynload_cuda selected_rows_functor)
endif()
nv_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor dynload_cuda)
nv_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
else()
cc_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
variable_visitor)
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope ddim)
if(WITH_DISTRIBUTE)
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim selected_rows_functor sendrecvop_grpc)
else()
cc_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim selected_rows_functor)
endif()
cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
endif()
......@@ -33,10 +45,10 @@ cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base s
cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
if (WITH_GPU)
cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle scale_loss_grad_op_handle rpc_op_handle
all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
endif()
cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle)
cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)
cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass)
cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle graph graph_helper pass op_graph_view reference_count_pass_helper)
cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)
cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
......@@ -44,10 +56,7 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass)
if (WITH_GPU)
list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif()
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
......
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ 65d355a7
......@@ -48,7 +48,14 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
void AllReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
// FIXME(typhoonzero): If scope0(global scope) have NCCL_ID_VAR,
// this is a distributed or inter-process call, find a better way.
#ifdef PADDLE_WITH_CUDA
if (NoDummyInputSize() == 1 &&
local_scopes_[0]->FindLocalVar(NCCL_ID_VARNAME) == nullptr) {
#else
if (NoDummyInputSize() == 1) {
#endif
return; // No need to all reduce when GPU count = 1;
} else {
// Wait input done
......
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ 65d355a7
......@@ -58,10 +58,23 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
}
}
CollectiveContext *context = CollectiveContext::GetInstance();
context->endpoints_ = strategy_.trainers_endpoints_;
context->trainer_id_ = strategy_.trainer_id_;
PADDLE_ENFORCE(strategy_.trainer_id_ >= 0, "trainer_id_ >= 0");
if (strategy_.trainer_id_ > 0) {
PADDLE_ENFORCE((unsigned)(strategy_.trainer_id_) <
strategy_.trainers_endpoints_.size(),
"trainer_id_ < endpoints_ size");
}
VLOG(1) << "CollectiveContext:" << context->String();
// Convert graph to run on multi-devices.
auto multi_devices_pass = AppendPass("multi_devices_pass");
multi_devices_pass->SetNotOwned<const BuildStrategy>("strategy",
&strategy_);
multi_devices_pass->Set<int>("num_trainers",
new int(strategy_.num_trainers_));
// Add a graph print pass to record a graph with device info.
if (!strategy_.debug_graphviz_path_.empty()) {
......@@ -133,16 +146,16 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
#endif
} else if (pass->Type() == "sequential_execution_pass") {
VLOG(1) << "set enable_sequential_execution:"
<< enable_sequential_execution_;
LOG(INFO) << "set enable_sequential_execution:"
<< enable_sequential_execution_;
pass->Erase(kAllOpDescs);
pass->Set<const std::vector<OpDesc *>>(
kAllOpDescs,
new std::vector<OpDesc *>(main_program.Block(0).AllOps()));
} else if (pass->Type() == "all_reduce_deps_pass") {
VLOG(1) << "SeqOnlyAllReduceOps:" << SeqOnlyAllReduceOps(*this)
<< ", num_trainers:" << num_trainers_;
LOG(INFO) << "SeqOnlyAllReduceOps:" << SeqOnlyAllReduceOps(*this)
<< ", num_trainers:" << num_trainers_;
pass->Erase(kAllOpDescs);
pass->Set<const std::vector<OpDesc *>>(
......
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ 65d355a7
......@@ -74,6 +74,8 @@ struct BuildStrategy {
bool fuse_broadcast_op_{false};
int num_trainers_{1};
int trainer_id_{0};
std::vector<std::string> trainers_endpoints_;
bool remove_unnecessary_lock_{false};
// NOTE:
......
paddle/fluid/framework/details/computation_op_handle.cc
浏览文件 @ 65d355a7
......@@ -20,11 +20,13 @@ namespace paddle {
namespace framework {
namespace details {
ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
platform::Place place)
platform::Place place,
size_t scope_idx)
: OpHandleBase(node),
op_(framework::OpRegistry::CreateOp(*node->Op())),
scope_(scope),
place_(place) {}
place_(place),
scope_idx_(scope_idx) {}
void ComputationOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
......
paddle/fluid/framework/details/computation_op_handle.h
浏览文件 @ 65d355a7
......@@ -28,7 +28,8 @@ namespace framework {
namespace details {
struct ComputationOpHandle : public OpHandleBase {
public:
ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place);
ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,
size_t scope_idx);
std::string Name() const override;
......@@ -38,6 +39,8 @@ struct ComputationOpHandle : public OpHandleBase {
void SetLockAndRecordEventFree(bool b) { is_lock_and_record_event_free_ = b; }
size_t GetScopeIdx() const { return scope_idx_; }
protected:
void RunImpl() override;
......@@ -47,6 +50,7 @@ struct ComputationOpHandle : public OpHandleBase {
std::unique_ptr<OperatorBase> op_;
Scope *scope_;
platform::Place place_;
size_t scope_idx_;
bool is_lock_and_record_event_free_{false};
};
} // namespace details
......
paddle/fluid/framework/details/eager_deletion_op_handle.cc 0 → 100644
浏览文件 @ 65d355a7
// 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/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
namespace paddle {
namespace framework {
namespace details {
EagerDeletionOpHandle::EagerDeletionOpHandle(
ir::Node *node, const Scope *scope, const platform::Place &place,
const std::unordered_set<std::string> &var_names, GarbageCollector *gc,
AtomicReferenceCountMap *ref_cnts)
: OpHandleBase(node),
scope_(scope),
var_names_(var_names),
gc_(gc),
ref_cnts_(ref_cnts) {
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(place)) {
dev_ctx_ = reinterpret_cast<platform::CUDADeviceContext *>(
platform::DeviceContextPool::Instance().Get(place));
if (dynamic_cast<StreamGarbageCollector *>(gc_)) {
platform::CUDADeviceGuard guard(
boost::get<platform::CUDAPlace>(place).device);
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
PADDLE_ENFORCE_NOT_NULL(event_);
}
}
#endif
}
EagerDeletionOpHandle::~EagerDeletionOpHandle() {
#ifdef PADDLE_WITH_CUDA
if (event_) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
platform::CUDADeviceGuard guard(gpu_place.device);
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
#endif
}
std::string EagerDeletionOpHandle::Name() const { return "eager_deletion"; }
void EagerDeletionOpHandle::RunImpl() {
auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
std::deque<std::shared_ptr<memory::Allocation>> garbages;
for (auto &name : var_names_) {
auto it = ref_cnts_->find(name);
// Var not found, not reference count has not decreased to 0
if (it == ref_cnts_->end() || it->second.fetch_sub(1) != 1) {
continue;
}
auto *var = exec_scope->FindVar(name);
if (var == nullptr) {
continue;
}
VLOG(2) << "Erase variable " << name;
if (var->IsType<LoDTensor>()) {
garbages.emplace_back(var->GetMutable<LoDTensor>()->MoveMemoryHolder());
} else if (var->IsType<SelectedRows>()) {
garbages.emplace_back(
var->GetMutable<SelectedRows>()->mutable_value()->MoveMemoryHolder());
} else if (var->IsType<LoDTensorArray>()) {
auto *tensor_arr = var->GetMutable<LoDTensorArray>();
for (auto &t : *tensor_arr) {
garbages.emplace_back(t.MoveMemoryHolder());
}
} else {
PADDLE_THROW("Type %s of %s is not supported eager deletion",
var->Type().name(), name);
}
}
if (!garbages.empty()) {
ClearGarbages(&garbages);
}
}
void EagerDeletionOpHandle::ClearGarbages(
std::deque<std::shared_ptr<memory::Allocation>> *garbages) {
#ifdef PADDLE_WITH_CUDA
if (event_) {
auto compute_stream = dev_ctx_->stream();
auto callback_stream =
reinterpret_cast<StreamGarbageCollector *>(gc_)->stream();
auto callback_func = [=]() {
PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream));
PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0));
};
gc_->Add(std::move(*garbages), callback_func);
} else {
#endif
gc_->Add(std::move(*garbages));
#ifdef PADDLE_WITH_CUDA
}
#endif
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/eager_deletion_op_handle.h 0 → 100644
浏览文件 @ 65d355a7
// 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 <deque>
#include <string>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
namespace paddle {
namespace framework {
class Scope;
namespace details {
class EagerDeletionOpHandle : public OpHandleBase {
public:
EagerDeletionOpHandle(ir::Node *node, const Scope *scope,
const platform::Place &place,
const std::unordered_set<std::string> &var_names,
GarbageCollector *gc,
AtomicReferenceCountMap *ref_cnts);
~EagerDeletionOpHandle();
std::string Name() const override;
protected:
void RunImpl() override;
private:
void ClearGarbages(std::deque<std::shared_ptr<memory::Allocation>> *garbages);
const Scope *scope_;
std::unordered_set<std::string> var_names_;
GarbageCollector *gc_; // not own
AtomicReferenceCountMap *ref_cnts_; // not own
#ifdef PADDLE_WITH_CUDA
platform::CUDADeviceContext *dev_ctx_{nullptr};
cudaEvent_t event_{nullptr};
#endif
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/eager_deletion_pass.cc 0 → 100644
浏览文件 @ 65d355a7
// 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 <queue>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_pass.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts =
Get<std::vector<AtomicReferenceCountMap>>(kRuntimeReferenceCount);
PADDLE_ENFORCE(ref_cnts.empty(),
"kRuntimeReferenceCount should be initialized here!");
const auto &vars = graph->Get<GraphVars>(kGraphVars);
ref_cnts.resize(vars.size());
const auto &last_live_ops =
Get<std::vector<LastLiveOpsOfVars>>(kLastLiveOpsOfVars);
const auto &gcs = Get<GarbageCollectorMap>(kGarbageCollector);
const auto &places = Get<std::vector<platform::Place>>(kAllPlaces);
// a reverse map of last_live_ops
// i.e., last op --> variable names which can be deleted.
std::unordered_map<ComputationOpHandle *, std::unordered_set<std::string>>
op_vars_map;
for (auto &var_ops_map : last_live_ops) {
for (auto &var_ops_pair : var_ops_map) {
const std::string &var_name = var_ops_pair.first;
for (auto *op : var_ops_pair.second) {
op_vars_map[op].insert(var_name);
}
}
}
for (auto &pair : op_vars_map) {
auto *op = pair.first;
auto &var_names = pair.second;
auto *eager_deletion_node =
graph->CreateEmptyNode("eager_deletion", ir::Node::Type::kOperation);
auto *eager_deletion_op = new EagerDeletionOpHandle(
eager_deletion_node, op->GetScope(), op->GetPlace(), var_names,
gcs.at(places[op->GetScopeIdx()]).get(),
&(ref_cnts[op->GetScopeIdx()]));
auto it = std::find_if(
op->Outputs().begin(), op->Outputs().end(), [](VarHandleBase *var) {
return dynamic_cast<DummyVarHandle *>(var) != nullptr;
});
if (it != op->Outputs().end()) {
eager_deletion_op->AddInput(*it);
} else {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
op->AddOutput(dep_var);
eager_deletion_op->AddInput(dep_var);
}
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
eager_deletion_op->AddOutput(dummy_leaf);
}
VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(eager_deletion_pass,
paddle::framework::details::EagerDeletionPass)
.RequirePassAttr(paddle::framework::details::kRuntimeReferenceCount)
.RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars)
.RequirePassAttr(paddle::framework::details::kAllPlaces)
.RequirePassAttr(paddle::framework::details::kGarbageCollector);
paddle/fluid/framework/details/eager_deletion_pass.h 0 → 100644
浏览文件 @ 65d355a7
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class EagerDeletionPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ 65d355a7
......@@ -133,6 +133,7 @@ static const char kPlaces[] = "places";
static const char kParams[] = "params";
static const char kLocalScopes[] = "local_scopes";
static const char kStrategy[] = "strategy";
static const char kNumTrainers[] = "num_trainers";
void MultiDevSSAGraphBuilder::Init() const {
all_vars_.clear();
......@@ -299,6 +300,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
auto nodes = graph->ReleaseNodes();
ir::Graph &result = *graph;
int num_trainers = Get<int>(kNumTrainers);
for (auto &node : nodes) {
if (node->IsVar() && node->Var()) {
all_vars_.emplace(node->Name(), node->Var());
......@@ -383,7 +386,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
CreateComputationalOps(&result, node, places_.size());
}
if (!is_forwarding && places_.size() > 1) {
if (!is_forwarding && (places_.size() > 1 || num_trainers > 1)) {
// Currently, we assume that once gradient is generated, it can be
// broadcast, and each gradient is only broadcast once.
if (static_cast<bool>(boost::get<int>(node->Op()->GetAttr(
......@@ -562,7 +565,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
int dev_id) const {
result->Get<GraphOps>(kGraphOps).emplace_back(
new ComputationOpHandle(result->CreateOpNode(node->Op()),
local_scopes_[dev_id], places_[dev_id]));
local_scopes_[dev_id], places_[dev_id], dev_id));
CreateOpHandleIOs(result, node, dev_id);
}
......@@ -685,8 +688,8 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result,
for (size_t scope_idx = 0; scope_idx < num_places; ++scope_idx) {
auto p = places_[scope_idx];
auto s = local_scopes_[scope_idx];
result->Get<GraphOps>(kGraphOps).emplace_back(
new ComputationOpHandle(result->CreateOpNode(node->Op()), s, p));
result->Get<GraphOps>(kGraphOps).emplace_back(new ComputationOpHandle(
result->CreateOpNode(node->Op()), s, p, scope_idx));
CreateOpHandleIOs(result, node, scope_idx);
}
}
......@@ -862,7 +865,7 @@ int MultiDevSSAGraphBuilder::CreateRPCOp(
if (node->Op()->Type() == "fetch_barrier") {
outvar_dev_id =
GetVarDeviceID(*result, output->Name(), *sharded_var_device);
PADDLE_ENFORCE_NE(outvar_dev_id, -1);
PADDLE_ENFORCE_NE(outvar_dev_id, -1, "output name %s", output->Name());
}
p = places_[outvar_dev_id];
ir::Node *new_node = nullptr;
......@@ -895,4 +898,5 @@ REGISTER_PASS(multi_devices_pass,
.RequirePassAttr(paddle::framework::details::kPlaces)
.RequirePassAttr(paddle::framework::details::kParams)
.RequirePassAttr(paddle::framework::details::kLocalScopes)
.RequirePassAttr(paddle::framework::details::kStrategy);
.RequirePassAttr(paddle::framework::details::kStrategy)
.RequirePassAttr(paddle::framework::details::kNumTrainers);
paddle/fluid/framework/details/op_graph_view.cc
浏览文件 @ 65d355a7
......@@ -23,6 +23,8 @@ namespace details {
OpGraphView::OpGraphView(const std::vector<OpHandleBase *> &ops) { Build(ops); }
void OpGraphView::Build(const std::vector<OpHandleBase *> &ops) {
preceding_ops_.clear();
pending_ops_.clear();
for (auto &op : ops) {
preceding_ops_[op];
pending_ops_[op];
......@@ -40,6 +42,7 @@ void OpGraphView::Build(const std::vector<OpHandleBase *> &ops) {
std::unordered_set<OpHandleBase *> OpGraphView::AllOps() const {
std::unordered_set<OpHandleBase *> ret;
ret.reserve(preceding_ops_.size());
for (auto &pair : preceding_ops_) {
ret.insert(pair.first);
}
......
paddle/fluid/framework/details/op_graph_view.h
浏览文件 @ 65d355a7
......@@ -14,7 +14,7 @@
#pragma once
#include <memory>
#include <queue>
#include <unordered_map>
#include <unordered_set>
#include <vector>
......@@ -34,6 +34,11 @@ class OpGraphView {
bool HasOp(OpHandleBase *op) const;
// Use a visitor to visit all pending ops of op
// Stop when callback returns false
template <typename Callback>
bool VisitAllPendingOps(OpHandleBase *op, Callback &&callback) const;
private:
void Build(const std::vector<OpHandleBase *> &ops);
void EnforceHasOp(OpHandleBase *op) const;
......@@ -44,6 +49,28 @@ class OpGraphView {
pending_ops_;
};
template <typename Callback>
bool OpGraphView::VisitAllPendingOps(OpHandleBase *op,
Callback &&callback) const {
EnforceHasOp(op);
std::unordered_set<OpHandleBase *> visited;
std::queue<OpHandleBase *> q;
q.push(op);
do {
op = q.front();
q.pop();
for (auto &pending_op : pending_ops_.at(op)) {
if (visited.count(pending_op) == 0) {
visited.insert(pending_op);
if (!callback(pending_op)) {
return false;
}
}
}
} while (!q.empty());
return true;
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/reduce_and_gather.h
浏览文件 @ 65d355a7
......@@ -53,7 +53,7 @@ struct ReduceLoDTensor {
}
};
inline void GatherSelectedRows(
inline void GatherLocalSelectedRows(
const std::vector<const SelectedRows *> &src_selecte_rows_,
const std::vector<platform::Place> &in_places,
const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes,
......
paddle/fluid/framework/details/reduce_op_handle.cc
浏览文件 @ 65d355a7
......@@ -16,6 +16,12 @@
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/reduce_and_gather.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
#include "paddle/fluid/operators/distributed/collective_client.h"
#include "paddle/fluid/operators/distributed/collective_server.h"
#include "paddle/fluid/operators/distributed/request_handler.h"
#endif
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/platform/profiler.h"
DEFINE_bool(
......@@ -26,6 +32,112 @@ namespace paddle {
namespace framework {
namespace details {
std::once_flag CollectiveContext::init_flag_;
std::unique_ptr<CollectiveContext> CollectiveContext::context_;
static inline std::string GetRemoteVarName(const std::string &var_name,
int trainer_id) {
return string::Sprintf("%s_merged_tmp@trainer_%d", var_name, trainer_id);
}
void ReduceOpHandle::Wait(
const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes) {
// TODO(gongwb): use event wait?
for (auto &dev_ctx : dev_ctxes) {
dev_ctx.second->Wait();
}
}
#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
template <typename DevCtx, typename DataType>
void ReduceOpHandle::GatherSelectedRows(
const std::vector<const SelectedRows *> &src_selected_rows,
const std::vector<platform::Place> &in_places,
const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes,
VarHandle *out_var_handle, const platform::Place &out_place,
SelectedRows *dst_selected_rows) {
const CollectiveContext &collective_context =
*CollectiveContext::GetInstance();
// 1. gather local selected rows, merge them
std::string gathered_var_name = out_var_handle->name_ + "_gathered_tmp";
auto scope = local_scopes_.at(out_var_handle->scope_idx_);
auto gathered_var_mid = scope->Var(gathered_var_name);
auto gathered_select_rows =
gathered_var_mid->GetMutable<framework::SelectedRows>();
GatherLocalSelectedRows(src_selected_rows, in_places, dev_ctxes, out_place,
gathered_select_rows);
// FIXME(gongwb): remove this Wait.
Wait(dev_ctxes);
// merge them
auto merged_dev_ctx = dynamic_cast<DevCtx *>(dev_ctxes.at(out_place));
std::string merged_var_name =
GetRemoteVarName(out_var_handle->name_, collective_context.trainer_id_);
auto merged_select_rows =
scope->Var(merged_var_name)->GetMutable<SelectedRows>();
operators::math::scatter::MergeAdd<DevCtx, DataType> merge_func;
merge_func(*merged_dev_ctx, *gathered_select_rows, merged_select_rows);
// 2. start collective server if it doesn't exist
operators::distributed::CollectiveServer *server =
operators::distributed::CollectiveServer::GetInstance(
collective_context.endpoints_[collective_context.trainer_id_],
collective_context.endpoints_.size() - 1);
auto rpc_server = server->GetRPCServer();
rpc_server->RegisterVar(merged_var_name,
operators::distributed::kRequestGetMonomerVariable,
scope, merged_dev_ctx);
// 3. gather them from all remote nodes.
std::vector<const SelectedRows *> remote;
operators::distributed::CollectiveClient *client =
operators::distributed::CollectiveClient::GetInstance();
std::vector<operators::distributed::RemoteVar> vars;
for (unsigned int i = 0; i < collective_context.endpoints_.size(); i++) {
if (i == (unsigned)collective_context.trainer_id_) continue;
operators::distributed::RemoteVar var;
var.trainer_id_ = i;
var.var_name_ = GetRemoteVarName(out_var_handle->name_, i);
var.ep_ = collective_context.endpoints_[i];
vars.push_back(var);
VLOG(4) << "gather from:" << var.String();
}
// erase gathered vars
merged_dev_ctx->Wait();
scope->EraseVars(std::vector<std::string>{gathered_var_name});
PADDLE_ENFORCE(client->Gather(vars, &remote, *merged_dev_ctx, scope));
PADDLE_ENFORCE(remote.size() == vars.size());
// 4. merged local selected rows.
std::vector<const SelectedRows *> all;
all.resize(collective_context.endpoints_.size());
for (auto v : vars) {
all[v.trainer_id_] =
scope->FindVar(v.var_name_)->GetMutable<SelectedRows>();
}
all[collective_context.trainer_id_] = merged_select_rows;
merge_func(*merged_dev_ctx, all, dst_selected_rows);
rpc_server->WaitVarBarrier(merged_var_name);
rpc_server->ClearVar(merged_var_name);
// 5. clear mid vars
std::vector<std::string> tmp_vars{merged_var_name};
for (auto r : vars) {
tmp_vars.push_back(r.var_name_);
}
scope->EraseVars(tmp_vars);
}
#endif
void ReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
......@@ -90,8 +202,36 @@ void ReduceOpHandle::RunImpl() {
this->RunAndRecordEvent([&] {
std::vector<const SelectedRows *> in_selected_rows =
GetInputValues<SelectedRows>(in_var_handles, var_scopes);
GatherSelectedRows(in_selected_rows, in_places, dev_ctxes_, t_out_p,
out_var->GetMutable<framework::SelectedRows>());
const CollectiveContext &collective_context =
*CollectiveContext::GetInstance();
VLOG(10) << "GatherSelectedRows CollectiveContext:"
<< collective_context.String();
// TODO(gongwb): add cpu support
if (collective_context.endpoints_.size() <= 1 ||
is_cpu_place(in_places[0]) || is_cpu_place(t_out_p)) {
GatherLocalSelectedRows(in_selected_rows, in_places, dev_ctxes_,
t_out_p,
out_var->GetMutable<framework::SelectedRows>());
return;
}
#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
if (framework::IsType<const float>(in_selected_rows[0]->value().type())) {
GatherSelectedRows<platform::CUDADeviceContext, float>(
in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
out_var->GetMutable<framework::SelectedRows>());
} else if (framework::IsType<const double>(
in_selected_rows[0]->value().type())) {
GatherSelectedRows<platform::CUDADeviceContext, double>(
in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
out_var->GetMutable<framework::SelectedRows>());
} else {
PADDLE_ENFORCE(false,
"only support double or float when gahter SelectedRows");
}
#endif
});
} else {
std::vector<const LoDTensor *> lod_tensors =
......
paddle/fluid/framework/details/reduce_op_handle.h
浏览文件 @ 65d355a7
......@@ -30,6 +30,32 @@
namespace paddle {
namespace framework {
namespace details {
struct CollectiveContext {
std::vector<std::string> endpoints_;
int trainer_id_{0};
std::string String() const {
std::stringstream ss;
ss << "endpoints_:";
for (auto e : endpoints_) {
ss << e << ",";
}
ss << "trainer_id_:" << trainer_id_;
return ss.str();
}
static CollectiveContext *GetInstance() {
std::call_once(init_flag_,
[&]() { context_.reset(new CollectiveContext()); });
return context_.get();
}
private:
static std::once_flag init_flag_;
static std::unique_ptr<CollectiveContext> context_;
};
struct ReduceOpHandle : public OpHandleBase {
std::vector<Scope *> local_scopes_;
......@@ -64,6 +90,19 @@ struct ReduceOpHandle : public OpHandleBase {
protected:
void RunImpl() override;
#if defined PADDLE_WITH_CUDA && defined PADDLE_WITH_DISTRIBUTE
template <typename DevCtx, typename DataType>
void GatherSelectedRows(
const std::vector<const SelectedRows *> &src_selecte_rows_,
const std::vector<platform::Place> &in_places,
const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes,
VarHandle *out_var_handle, const platform::Place &out_place,
SelectedRows *dst_selecte_rows);
#endif
void Wait(
const std::map<platform::Place, platform::DeviceContext *> &dev_ctxes);
template <typename T>
std::vector<const T *> GetInputValues(
const std::vector<VarHandle *> &in_var_handles,
......
paddle/fluid/framework/details/reference_count_op_handle.h 已删除 100644 → 0
浏览文件 @ c550e0ce
// 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 <atomic>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/garbage_collector.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/framework/tensor.h"
namespace paddle {
namespace framework {
namespace details {
using ReferenceCountMap = std::unordered_map<std::string, int>;
using AtomicReferenceCountMap =
std::unordered_map<std::string, std::atomic<int>>;
using DeviceReferenceCountMap =
std::unordered_map<int, std::unique_ptr<ReferenceCountMap>>;
using AtomicDeviceReferenceCountMap =
std::unordered_map<int, std::unique_ptr<AtomicReferenceCountMap>>;
using DeviceGarbageCollectorMap =
std::unordered_map<int,
std::unique_ptr<GarbageCollector<framework::Tensor>>>;
class ReferenceCountOpHandle : public OpHandleBase {
public:
ReferenceCountOpHandle(ir::Node *node, const Scope *scope,
const platform::CUDAPlace &place,
const std::vector<std::string> &var_names,
GarbageCollector<Tensor> *gc,
AtomicReferenceCountMap *ref_cnts)
: OpHandleBase(node), scope_(scope), gc_(gc), ref_cnts_(ref_cnts) {
dev_ctx_ = static_cast<platform::CUDADeviceContext *>(
platform::DeviceContextPool::Instance().Get(place));
if (IsStreamGarabageCollector()) {
platform::SetDeviceId(place.device);
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
}
for (auto &name : var_names) AddVar(name);
}
~ReferenceCountOpHandle() {
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
platform::SetDeviceId(gpu_place.device);
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
}
std::string Name() const override { return "reference_count"; }
void AddVar(const std::string &name) {
auto it = var_names_.find(name);
if (it != var_names_.end())
++(it->second);
else
var_names_[name] = 1;
}
protected:
void RunImpl() override {
auto *exec_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
std::vector<Tensor *> tensors;
for (auto &pair : var_names_) {
auto &name = pair.first;
auto it = ref_cnts_->find(name);
if (it == ref_cnts_->end()) continue;
auto *var = exec_scope->FindVar(name);
if (var == nullptr) continue;
if (var->IsType<LoDTensor>()) {
if (it->second.fetch_sub(pair.second) <= pair.second) {
tensors.emplace_back(var->GetMutable<LoDTensor>());
}
} else if (var->IsType<SelectedRows>()) {
if (it->second.fetch_sub(pair.second) <= pair.second) {
tensors.emplace_back(
var->GetMutable<SelectedRows>()->mutable_value());
}
}
}
if (!tensors.empty()) {
ClearTensors(tensors);
}
}
private:
void ClearTensors(const std::vector<Tensor *> &tensors) {
auto *gc = dynamic_cast<StreamGarbageCollector<Tensor> *>(gc_);
if (gc != nullptr) {
auto compute_stream = dev_ctx_->stream();
auto callback_stream = gc->stream();
auto callback_func = [=]() {
PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream));
PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0));
};
gc_->Add(tensors, callback_func);
} else {
gc_->Add(tensors);
}
}
bool IsStreamGarabageCollector() const {
return dynamic_cast<const StreamGarbageCollector<Tensor> *>(gc_) != nullptr;
}
const Scope *scope_;
platform::CUDADeviceContext *dev_ctx_;
std::unordered_map<std::string, int> var_names_;
GarbageCollector<Tensor> *gc_; // not own
AtomicReferenceCountMap *ref_cnts_; // not own
cudaEvent_t event_;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/reference_count_pass.cc
浏览文件 @ 65d355a7
......@@ -14,187 +14,240 @@
#include <queue>
#include <string>
#include <type_traits>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/op_graph_view.h"
#include "paddle/fluid/framework/details/reference_count_pass.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
static ComputationOpHandle *FindNextComputationOpHandle(VarHandle *var_in) {
std::queue<VarHandleBase *> queue;
queue.push(var_in);
do {
auto *var = queue.front();
queue.pop();
for (auto *op : var->PendingOps()) {
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op != nullptr && compute_op->GetPlace() == var_in->place_) {
return compute_op;
// A functor to shrink/remove operators who depend on other operators in a set
class ShrinkDepsOpFunctor {
private:
enum RelationShip { kSame = 0, kNoDeps = 1, kBefore = 2, kAfter = 3 };
public:
explicit ShrinkDepsOpFunctor(const std::vector<OpHandleBase *> &all_ops)
: graph_(all_ops) {}
template <typename OpSet>
OpSet operator()(const OpSet &op_set) const {
using KeyType = typename OpSet::key_type;
static_assert(
std::is_base_of<OpHandleBase,
typename std::remove_pointer<KeyType>::type>::value,
"Key type of OpSet must be OpHandleBase, or derived of OpHandleBase");
if (op_set.size() <= 1) return op_set;
std::vector<OpHandleBase *> ops(op_set.begin(), op_set.end());
OpSet ret;
auto rels = GetRelations(ops);
auto not_before = [](RelationShip r) { return r != kBefore; };
for (size_t i = 0; i < rels.size(); ++i) {
if (std::all_of(rels[i].begin(), rels[i].end(), not_before)) {
ret.emplace(static_cast<KeyType>(ops[i]));
}
for (auto *out_var : op->Outputs()) {
queue.push(out_var);
}
return ret;
}
private:
std::vector<std::vector<RelationShip>> GetRelations(
const std::vector<OpHandleBase *> &ops) const {
std::unordered_map<OpHandleBase *, size_t> op_to_idx;
for (size_t i = 0; i < ops.size(); ++i) {
PADDLE_ENFORCE(graph_.HasOp(ops[i]), "Op does not exist in graph");
op_to_idx[ops[i]] = i;
}
PADDLE_ENFORCE(op_to_idx.size() == ops.size(), "Duplicate ops");
std::vector<std::vector<RelationShip>> ret(ops.size());
for (auto &e : ret) {
e.assign(ops.size(), kSame);
}
size_t found_num = ops.size();
size_t total_num = ops.size() * ops.size();
auto visitor = [&](OpHandleBase *op, size_t i) {
auto it = op_to_idx.find(op);
if (it != op_to_idx.end()) {
size_t j = it->second;
if (i != j && ret[i][j] == kSame) {
ret[i][j] = kBefore;
ret[j][i] = kAfter;
found_num += 2;
if (found_num == total_num) {
return false;
}
}
}
return true;
};
for (size_t i = 0; i < ops.size(); ++i) {
auto sub_visitor = [&, i](OpHandleBase *op) { return visitor(op, i); };
if (!graph_.VisitAllPendingOps(ops[i], sub_visitor)) {
break;
}
}
for (size_t i = 0; i < ops.size(); ++i) {
for (size_t j = i + 1; j < ops.size(); ++j) {
if (ret[i][j] != kSame) continue;
ret[i][j] = kNoDeps;
ret[j][i] = kNoDeps;
}
}
return ret;
}
const OpGraphView graph_;
};
/**
* Find the nearest downstream computation op handle. If the op is a
* computation op, just return itself.
*/
static ComputationOpHandle *FindNextComputationOpHandleOrReturnItself(
OpHandleBase *op, size_t scope_idx) {
std::queue<OpHandleBase *> q;
std::unordered_set<OpHandleBase *> visited;
q.push(op);
do {
auto *op = q.front();
q.pop();
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op != nullptr && compute_op->GetScopeIdx() == scope_idx) {
return compute_op;
}
for (auto *out_var : op->Outputs()) {
for (auto *pending_op : out_var->PendingOps()) {
if (visited.count(pending_op)) continue;
visited.insert(pending_op);
}
}
} while (!queue.empty());
} while (!q.empty());
return nullptr;
}
static void AddDependencyBetween(OpHandleBase *in, OpHandleBase *out,
ir::Graph *graph) {
auto it = std::find_if(
in->Outputs().begin(), in->Outputs().end(), [](VarHandleBase *var) {
return dynamic_cast<DummyVarHandle *>(var) != nullptr;
});
if (it != in->Outputs().end()) {
out->AddInput(*it);
} else {
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
in->AddOutput(dep_var);
out->AddInput(dep_var);
static std::unordered_set<ComputationOpHandle *>
ExtractComputationOpFromLastLivedVar(VarHandle *var, size_t scope_idx,
const ShrinkDepsOpFunctor &shrink_func,
bool *ok) {
// stage one. Get last op for variable.
std::unordered_set<OpHandleBase *> candidates;
{
if (var->PendingOps().empty() && var->GeneratedOp()) {
// No operator depends on this variable. So the last operator is the op
// who generates this variable.
candidates.emplace(var->GeneratedOp());
} else {
candidates = var->PendingOps();
}
// No pending ops or generated op is nullptr
if (candidates.empty()) {
*ok = false;
return {};
}
}
// stage two. Try to cast them to computation op.
// return (*ok=false) when failed.
//
// The reason why we cannot make any types of op handle to be the last lived
// op is:
// some op handle may operate on many DeviceContext, however, our garbage
// collector can only wait one DeviceContext for now. So currently, we wait
// the nearest compute op.
std::unordered_set<ComputationOpHandle *> computation_op;
{
for (auto *op : candidates) {
auto *compute_op =
FindNextComputationOpHandleOrReturnItself(op, scope_idx);
if (compute_op == nullptr) {
*ok = false;
return {};
}
computation_op.emplace(compute_op);
}
}
// stage three. Try to shrink computation op if they depend on each other.
// Get the smallest set of the most ops.
*ok = true;
return shrink_func(computation_op);
}
static VarDesc *TryGetLatestVarDesc(const std::vector<VarHandle *> &vars) {
VarDesc *var_desc = nullptr;
std::find_if(vars.rbegin(), vars.rend(), [&](VarHandle *var_handle) -> bool {
var_desc = var_handle->Node()->Var();
return var_desc != nullptr;
});
return var_desc;
}
std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts = Get<DeviceReferenceCountMap>(kGlobalReferenceCount);
auto &cur_ref_cnts = Get<AtomicDeviceReferenceCountMap>(kCurReferenceCount);
auto &gcs = Get<DeviceGarbageCollectorMap>(kGarbageCollector);
// It is not easy to find the right reference counts of varaibles in graph
// Step 1: Find all variables in computation ops
// Step 2: Find all variables in non-computation ops which refers to variables
// in computation ops
std::unordered_set<std::string> names;
std::unordered_map<OpHandleBase *, ReferenceCountOpHandle *>
compute_ref_cnt_map;
auto get_ref_cnts_from_compute_op = [&](
OpHandleBase *op, const std::vector<VarHandleBase *> &vars) {
std::vector<std::string> var_names_in_op;
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
if (compute_op == nullptr ||
!platform::is_gpu_place(compute_op->GetPlace()))
return var_names_in_op;
auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
for (VarHandleBase *var_handle_base : vars) {
auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
if (!platform::is_gpu_place(var_handle->place_) ||
boost::get<platform::CUDAPlace>(var_handle->place_) != place)
continue;
auto &ref_cnts = Get<std::vector<ReferenceCountMap>>(kGlobalReferenceCount);
auto &last_live_ops_of_vars =
Get<std::vector<LastLiveOpsOfVars>>(kLastLiveOpsOfVars);
PADDLE_ENFORCE(last_live_ops_of_vars.empty() && ref_cnts.empty(),
"Last Live Ops and Reference Counts of vars should be "
"initialized at here.");
VarDesc *var_desc = var_handle->Node()->Var();
auto var_name = var_handle->Node()->Name();
const auto &vars = graph->Get<GraphVars>(kGraphVars);
// This is weird but there is really some variables without var_desc
// in computation_op
if (var_desc == nullptr) {
var_desc = compute_op->Node()->Op()->Block()->FindVar(var_name);
if (var_desc == nullptr) continue;
last_live_ops_of_vars.resize(vars.size());
ref_cnts.resize(vars.size());
ShrinkDepsOpFunctor shrink_func(
ir::FilterByNodeWrapper<OpHandleBase>(*graph));
for (size_t i = 0; i < vars.size(); ++i) {
for (auto &name_var_pair : vars[i]) {
// Whether this variable can be reused or deleted? If not, we do not
// compute reference counts and dependencies.
VarDesc *var_desc = TryGetLatestVarDesc(name_var_pair.second);
if (var_desc == nullptr || var_desc->Persistable()) {
continue;
}
if (var_desc->Persistable()) continue;
auto var_type = var_desc->Proto()->type().type();
if (var_type != proto::VarType::LOD_TENSOR &&
var_type != proto::VarType::SELECTED_ROWS) {
var_type != proto::VarType::SELECTED_ROWS &&
var_type != proto::VarType::LOD_TENSOR_ARRAY) {
// Var type cannot be deleted
continue;
}
// compute op only runs in one device
if (ref_cnts[place.device]->count(var_name))
++(*ref_cnts[place.device])[var_name];
else
(*ref_cnts[place.device])[var_name] = 1;
bool ok;
auto result = ExtractComputationOpFromLastLivedVar(
name_var_pair.second.back(), i, shrink_func, &ok);
names.insert(var_name);
var_names_in_op.push_back(var_name);
}
return var_names_in_op;
};
auto update_ref_cnts_from_non_compute_op = [&](
OpHandleBase *op, const std::vector<VarHandleBase *> &vars) {
if (dynamic_cast<ComputationOpHandle *>(op) != nullptr) return;
for (VarHandleBase *var_handle_base : vars) {
auto *var_handle = dynamic_cast<VarHandle *>(var_handle_base);
if (var_handle == nullptr || !var_handle->Node()->IsVar()) continue;
auto var_name = var_handle->Node()->Name();
auto var_place = var_handle->place_;
if (!platform::is_gpu_place(var_place)) continue;
auto place = boost::get<platform::CUDAPlace>(var_place);
if (names.count(var_name) == 0) continue;
if (ref_cnts.count(place.device) &&
ref_cnts[place.device]->count(var_name)) {
++(*ref_cnts[place.device])[var_name];
auto *next_compute_op = FindNextComputationOpHandle(var_handle);
if (next_compute_op != nullptr) {
if (compute_ref_cnt_map.count(next_compute_op)) {
compute_ref_cnt_map[next_compute_op]->AddVar(var_name);
VLOG(5) << "Add reference count of " << var_name << " to Operator "
<< next_compute_op->Name();
} else {
// Create new reference_count_op_handle
ir::Node *ref_cnt_node = graph->CreateEmptyNode(
"reference_count", ir::Node::Type::kOperation);
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, next_compute_op->GetScope(), place, {var_name},
gcs[place.device].get(), cur_ref_cnts[place.device].get());
AddDependencyBetween(next_compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[next_compute_op] = ref_cnt_handle;
}
}
if (ok) {
auto &var_name = name_var_pair.first;
PADDLE_ENFORCE(!result.empty(), "Last living ops of %s cannot be empty",
var_name);
ref_cnts[i].emplace(var_name, result.size());
last_live_ops_of_vars[i].emplace(var_name, std::move(result));
}
}
};
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto &op : all_ops) {
auto in_var_names = get_ref_cnts_from_compute_op(op, op->Inputs());
auto out_var_names = get_ref_cnts_from_compute_op(op, op->Outputs());
if (in_var_names.empty() && out_var_names.empty()) continue;
in_var_names.insert(in_var_names.end(), out_var_names.begin(),
out_var_names.end());
auto *compute_op = dynamic_cast<ComputationOpHandle *>(op);
auto place = boost::get<platform::CUDAPlace>(compute_op->GetPlace());
ir::Node *ref_cnt_node =
graph->CreateEmptyNode("reference_count", ir::Node::Type::kOperation);
auto *ref_cnt_handle = new ReferenceCountOpHandle(
ref_cnt_node, compute_op->GetScope(), place, in_var_names,
gcs[place.device].get(), cur_ref_cnts[place.device].get());
AddDependencyBetween(compute_op, ref_cnt_handle, graph.get());
compute_ref_cnt_map[compute_op] = ref_cnt_handle;
}
for (auto &op : all_ops) {
update_ref_cnts_from_non_compute_op(op, op->Inputs());
update_ref_cnts_from_non_compute_op(op, op->Outputs());
}
std::vector<OpHandleBase *> new_all_ops;
new_all_ops.reserve(compute_ref_cnt_map.size() + all_ops.size());
for (auto &op : all_ops) {
new_all_ops.emplace_back(std::move(op));
auto it = compute_ref_cnt_map.find(new_all_ops.back());
if (it != compute_ref_cnt_map.end()) {
// Add LeafNode to ReferenceCountOpHandle
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
it->second->AddOutput(dummy_leaf);
new_all_ops.emplace_back(std::move(it->second));
}
}
all_ops.swap(new_all_ops);
return graph;
}
......@@ -205,5 +258,4 @@ std::unique_ptr<ir::Graph> ReferenceCountPass::ApplyImpl(
REGISTER_PASS(reference_count_pass,
paddle::framework::details::ReferenceCountPass)
.RequirePassAttr(paddle::framework::details::kGlobalReferenceCount)
.RequirePassAttr(paddle::framework::details::kCurReferenceCount)
.RequirePassAttr(paddle::framework::details::kGarbageCollector);
.RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars);
paddle/fluid/framework/details/reference_count_pass.h
浏览文件 @ 65d355a7
......@@ -14,7 +14,6 @@
#pragma once
#include "paddle/fluid/framework/details/reference_count_op_handle.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
......@@ -22,10 +21,6 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kGlobalReferenceCount[] = "reference_count";
constexpr char kCurReferenceCount[] = "current_reference_count";
constexpr char kGarbageCollector[] = "garbage_collector";
class ReferenceCountPass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
......
paddle/fluid/framework/details/reference_count_pass_helper.cc 0 → 100644
浏览文件 @ 65d355a7
// 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/reference_count_pass_helper.h"
namespace paddle {
namespace framework {
namespace details {} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/reference_count_pass_helper.h 0 → 100644
浏览文件 @ 65d355a7
// 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 <atomic>
#include <map>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/garbage_collector.h"
namespace paddle {
namespace framework {
namespace details {
class ComputationOpHandle;
using ReferenceCountMap = std::unordered_map<std::string, size_t>;
using AtomicReferenceCountMap =
std::unordered_map<std::string, std::atomic<size_t>>;
using GarbageCollectorMap =
std::map<platform::Place, std::unique_ptr<GarbageCollector>>;
const char kGlobalReferenceCount[] = "global_reference_count";
const char kRuntimeReferenceCount[] = "runtime_reference_count";
const char kGarbageCollector[] = "garbage_collector";
const char kAllPlaces[] = "all_places";
using LastLiveOpsOfVars =
std::unordered_map<std::string, std::unordered_set<ComputationOpHandle*>>;
const char kLastLiveOpsOfVars[] = "last_live_ops_of_var";
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
浏览文件 @ 65d355a7
......@@ -16,11 +16,8 @@
#include <stdexcept>
#include <string>
#include <vector>
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/platform/profiler.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/framework/details/reference_count_op_handle.h"
#endif
namespace paddle {
namespace framework {
......@@ -69,27 +66,12 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr);
drop_scope_counter_ += 1;
#ifdef PADDLE_WITH_CUDA
const std::string gc_name = "garbage_collector";
DeviceGarbageCollectorMap *gc =
Graph().Has(gc_name) ? &(Graph().Get<DeviceGarbageCollectorMap>(gc_name))
: nullptr;
#endif
if (!fetch_tensors.empty() ||
drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {
drop_scope_counter_ = 0;
// Wait All computational streams
for (auto p : places_) {
platform::DeviceContextPool::Instance().Get(p)->Wait();
#ifdef PADDLE_WITH_CUDA
if (gc != nullptr && platform::is_gpu_place(p)) {
auto gpu_place = boost::get<platform::CUDAPlace>(p);
auto &gc_at_place = gc->at(gpu_place.device);
gc_at_place->Wait();
gc_at_place->Reset();
}
#endif
}
for (auto &scope : local_scopes_) {
auto &local_scope =
......
paddle/fluid/framework/executor.cc
浏览文件 @ 65d355a7
......@@ -13,18 +13,23 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/executor.h"
#include <deque>
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/ngraph_operator.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/transfer_scope_cache.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/operators/detail/macros.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/platform/profiler.h"
#ifdef PADDLE_WITH_NGRAPH
#include "paddle/fluid/framework/ngraph_operator.h"
#endif
DECLARE_bool(benchmark);
DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run");
DEFINE_bool(use_ngraph, false, "Use NGRAPH to run");
......@@ -37,11 +42,43 @@ namespace {
int kProgramId = -1;
} // namespace
static std::unordered_map<std::string, size_t> GetNonPersistableReferenceCounts(
const BlockDesc& block, const std::vector<std::string>& skip_var_list) {
std::unordered_map<std::string, size_t> ref_cnts;
std::unordered_set<std::string> skip_vars(skip_var_list.begin(),
skip_var_list.end());
auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
for (auto& name_pair : name_map) {
for (auto& name : name_pair.second) {
if (skip_vars.count(name)) continue;
auto* var_desc = block.FindVar(name);
if (var_desc == nullptr || var_desc->Persistable()) continue;
auto type = var_desc->Proto()->type().type();
if (type != proto::VarType::LOD_TENSOR &&
type != proto::VarType::SELECTED_ROWS &&
type != proto::VarType::LOD_TENSOR_ARRAY) {
continue;
}
++ref_cnts[name];
}
}
};
for (auto op_desc : block.AllOps()) {
update_ref_cnts(op_desc, op_desc->Inputs());
update_ref_cnts(op_desc, op_desc->Outputs());
}
return ref_cnts;
}
ExecutorPrepareContext::ExecutorPrepareContext(
const framework::ProgramDesc& prog, size_t block_id)
const framework::ProgramDesc& prog, size_t block_id,
const std::vector<std::string>& skip_ref_cnt_vars)
: prog_(prog), block_id_(block_id) {
if (GetEagerDeletionThreshold() >= 0) {
ref_cnts_ = GetNonPersistableReferenceCount<int>(prog_, block_id_);
global_ref_cnts_ = GetNonPersistableReferenceCounts(prog.Block(block_id),
skip_ref_cnt_vars);
}
}
......@@ -49,28 +86,40 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
VLOG(5) << "destroy ExecutorPrepareContext";
}
template <typename RefCntMap>
static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
GarbageCollector<Tensor>* gc,
RefCntMap* ref_cnts) {
std::unordered_set<Tensor*> erase_tensors;
static void DeleteUnusedTensors(
const Scope& scope, const OperatorBase* op, GarbageCollector* gc,
std::unordered_map<std::string, size_t>* ref_cnts) {
std::deque<std::shared_ptr<memory::Allocation>> garbages;
auto handler = [&](const VariableNameMap& name_map) {
for (auto& name_pair : name_map) {
for (auto& name : name_pair.second) {
auto it = ref_cnts->find(name);
if (it == ref_cnts->end()) continue;
if ((it->second)-- == 1) {
auto* var = scope.FindVar(name);
if (var != nullptr) {
VLOG(10) << "Erase tensor \'" << name << "\'";
if (var->IsType<LoDTensor>()) {
erase_tensors.insert(var->GetMutable<LoDTensor>());
} else if (var->IsType<SelectedRows>()) {
erase_tensors.insert(
var->GetMutable<SelectedRows>()->mutable_value());
}
if (--(it->second) != 0) {
continue;
}
auto* var = scope.FindVar(name);
if (var != nullptr) {
continue;
}
VLOG(2) << "Erase variable " << name;
if (var->IsType<LoDTensor>()) {
garbages.emplace_back(
var->GetMutable<LoDTensor>()->MoveMemoryHolder());
} else if (var->IsType<SelectedRows>()) {
garbages.emplace_back(var->GetMutable<SelectedRows>()
->mutable_value()
->MoveMemoryHolder());
} else if (var->IsType<LoDTensorArray>()) {
auto* lod_tensor_arr = var->GetMutable<LoDTensorArray>();
for (auto& t : *lod_tensor_arr) {
garbages.emplace_back(t.MoveMemoryHolder());
}
} else {
PADDLE_THROW("Type %s of %s is not supported eager deletion",
var->Type().name(), name);
}
}
}
......@@ -79,19 +128,19 @@ static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
handler(op->Inputs());
handler(op->Outputs());
if (!erase_tensors.empty()) {
gc->Add(erase_tensors);
if (!garbages.empty()) {
gc->Add(std::move(garbages));
}
}
static void EnableFusedOp(ExecutorPrepareContext* ctx) {
#ifdef PADDLE_WITH_NGRAPH
VLOG(3) << "use_ngraph=True";
auto intervals = FusedOperator::FusedOpIntervals(&ctx->ops_);
auto intervals = NgraphOperator::NgraphOpIntervals(&ctx->ops_);
for (auto& interval : intervals) {
auto* fused_op = new FusedOperator(ctx->prog_, ctx->block_id_,
interval.at(0), interval.at(1));
*interval[0] = std::unique_ptr<OperatorBase>(fused_op);
auto* ng_op = new NgraphOperator(ctx->prog_, ctx->block_id_, interval.at(0),
interval.at(1));
*interval[0] = std::unique_ptr<OperatorBase>(ng_op);
}
for (auto it = intervals.rbegin(); it != intervals.rend(); ++it) {
ctx->ops_.erase(it->at(0) + 1, it->at(1));
......@@ -114,36 +163,6 @@ void Executor::Close() {
#endif
}
void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
if (var_type == proto::VarType::LOD_TENSOR) {
var->GetMutable<LoDTensor>();
} else if (var_type == proto::VarType::SELECTED_ROWS) {
var->GetMutable<SelectedRows>();
} else if (var_type == proto::VarType::FEED_MINIBATCH) {
var->GetMutable<FeedFetchList>();
} else if (var_type == proto::VarType::FETCH_LIST) {
var->GetMutable<FeedFetchList>();
} else if (var_type == proto::VarType::STEP_SCOPES) {
var->GetMutable<std::vector<framework::Scope*>>();
} else if (var_type == proto::VarType::LOD_RANK_TABLE) {
var->GetMutable<LoDRankTable>();
} else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
var->GetMutable<LoDTensorArray>();
} else if (var_type == proto::VarType::PLACE_LIST) {
var->GetMutable<platform::PlaceList>();
} else if (var_type == proto::VarType::READER) {
var->GetMutable<ReaderHolder>();
} else if (var_type == proto::VarType::RAW) {
// GetMutable will be called in operator
} else {
PADDLE_THROW(
"Variable type %d is not in "
"[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
"LOD_RANK_TABLE, PLACE_LIST, READER, RAW]",
var_type);
}
}
void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope,
int block_id) {
auto& global_block = pdesc.Block(block_id);
......@@ -351,9 +370,10 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
}
std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
const ProgramDesc& program, int block_id) {
const ProgramDesc& program, int block_id,
const std::vector<std::string>& skip_ref_cnt_vars) {
std::unique_ptr<ExecutorPrepareContext> ctx(
new ExecutorPrepareContext(program, block_id));
new ExecutorPrepareContext(program, block_id, skip_ref_cnt_vars));
PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
auto& block = program.Block(block_id);
for (auto& op_desc : block.AllOps()) {
......@@ -364,16 +384,28 @@ std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
}
std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
const ProgramDesc& program, const std::vector<int>& block_ids) {
const ProgramDesc& program, const std::vector<int>& block_ids,
const std::vector<std::vector<std::string>>& skip_ref_cnt_vars) {
PADDLE_ENFORCE(
skip_ref_cnt_vars.empty() || skip_ref_cnt_vars.size() == block_ids.size(),
"skip_ref_cnt_vars should be either empty or equals to block number %d",
block_ids.size());
std::vector<std::shared_ptr<ExecutorPrepareContext>> result;
size_t idx = 0;
for (auto& bid : block_ids) {
auto* ctx = new ExecutorPrepareContext(program, bid);
ExecutorPrepareContext* ctx;
if (skip_ref_cnt_vars.empty()) {
ctx = new ExecutorPrepareContext(program, bid);
} else {
ctx = new ExecutorPrepareContext(program, bid, skip_ref_cnt_vars[idx]);
}
PADDLE_ENFORCE_LT(static_cast<size_t>(bid), program.Size());
auto& block = program.Block(bid);
for (auto& op_desc : block.AllOps()) {
ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
}
result.push_back(std::shared_ptr<ExecutorPrepareContext>(ctx));
++idx;
}
return result;
}
......@@ -391,22 +423,23 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
}
int64_t max_memory_size = GetEagerDeletionThreshold();
std::unique_ptr<GarbageCollector<Tensor>> gc;
// WhileOp would set keep_kids to true,
// because WhileGradOp needs the scopes created in WhileOp.
// Perhaps, we should not perform eager deletion in WhileOp
// The scopes and variables created by WhileOp would be deleted
// in WhileGradOp.
std::unique_ptr<GarbageCollector> gc;
// skip while_op and while_grad_op temporarily
if (max_memory_size >= 0 && !keep_kids) {
ctx->ResetReferenceCount();
#ifdef PADDLE_WITH_CUDA
if (platform::is_gpu_place(place_)) {
gc.reset(new DefaultStreamGarbageCollector<Tensor>(
boost::get<platform::CUDAPlace>(place_), max_memory_size));
} else {
if (IsFastEagerDeletionModeEnabled()) {
gc.reset(new UnsafeFastGPUGarbageCollector(
boost::get<platform::CUDAPlace>(place_), max_memory_size));
} else {
gc.reset(new DefaultStreamGarbageCollector(
boost::get<platform::CUDAPlace>(place_), max_memory_size));
}
} else if (platform::is_cpu_place(place_)) {
#endif
gc.reset(new CPUGarbageCollector<Tensor>(
boost::get<platform::CPUPlace>(place_), max_memory_size));
gc.reset(new CPUGarbageCollector(boost::get<platform::CPUPlace>(place_),
max_memory_size));
#ifdef PADDLE_WITH_CUDA
}
#endif
......@@ -415,17 +448,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
for (auto& op : ctx->ops_) {
op->Run(*local_scope, place_);
if (gc != nullptr) {
if (gc) {
DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
&(ctx->cur_ref_cnts_));
&(ctx->runtime_ref_cnts_));
}
}
if (gc != nullptr) {
gc->Wait();
} else {
platform::DeviceContextPool::Instance().Get(place_)->Wait();
}
platform::DeviceContextPool::Instance().Get(place_)->Wait();
if (local_scope != scope) {
scope->DeleteScope(local_scope);
......
paddle/fluid/framework/executor.h
浏览文件 @ 65d355a7
......@@ -26,54 +26,22 @@ limitations under the License. */
namespace paddle {
namespace framework {
extern void InitializeVariable(Variable* var, proto::VarType::Type var_type);
template <typename T>
std::unordered_map<std::string, T> GetNonPersistableReferenceCount(
const ProgramDesc& prog, size_t block_id) {
auto& block = prog.Block(block_id);
std::unordered_map<std::string, T> ref_cnts;
auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
for (auto& name_pair : name_map) {
for (auto& name : name_pair.second) {
auto* var_desc = block.FindVar(name);
if (var_desc == nullptr || var_desc->Persistable()) continue;
auto type = var_desc->Proto()->type().type();
if (type != proto::VarType::LOD_TENSOR &&
type != proto::VarType::SELECTED_ROWS) {
continue;
}
auto it = ref_cnts.find(name);
if (it != ref_cnts.end()) {
++it->second;
} else {
ref_cnts[name] = 1;
}
}
}
};
for (auto op_desc : block.AllOps()) {
update_ref_cnts(op_desc, op_desc->Inputs());
update_ref_cnts(op_desc, op_desc->Outputs());
}
return ref_cnts;
}
struct ExecutorPrepareContext {
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id);
ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>());
~ExecutorPrepareContext();
void ResetReferenceCount() { cur_ref_cnts_ = ref_cnts_; }
void ResetReferenceCount() { runtime_ref_cnts_ = global_ref_cnts_; }
const framework::ProgramDesc& prog_;
size_t block_id_;
std::vector<std::unique_ptr<OperatorBase>> ops_;
std::unordered_map<std::string, int> ref_cnts_;
std::unordered_map<std::string, int> cur_ref_cnts_;
std::unordered_map<std::string, size_t> global_ref_cnts_;
std::unordered_map<std::string, size_t> runtime_ref_cnts_;
};
class Executor {
......@@ -109,10 +77,14 @@ class Executor {
const std::string& fetch_holder_name = "fetch");
static std::unique_ptr<ExecutorPrepareContext> Prepare(
const ProgramDesc& program, int block_id);
const ProgramDesc& program, int block_id,
const std::vector<std::string>& skip_ref_cnt_vars =
std::vector<std::string>());
static std::vector<std::shared_ptr<ExecutorPrepareContext>> Prepare(
const ProgramDesc& program, const std::vector<int>& block_ids);
const ProgramDesc& program, const std::vector<int>& block_ids,
const std::vector<std::vector<std::string>>& skip_ref_cnt_vars =
std::vector<std::vector<std::string>>());
void CreateVariables(const ProgramDesc& pdesc, Scope* scope, int block_id);
......
paddle/fluid/framework/executor_thread_worker.cc 0 → 100644
浏览文件 @ 65d355a7
/* 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/executor_thread_worker.h"
#include "google/protobuf/io/zero_copy_stream_impl.h"
#include "google/protobuf/message.h"
#include "google/protobuf/text_format.h"
#include "gflags/gflags.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/cpu_helper.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/pybind/pybind.h"
namespace paddle {
namespace framework {
void ExecutorThreadWorker::CreateThreadOperators(const ProgramDesc& program) {
auto& block = program.Block(0);
op_names_.clear();
for (auto& op_desc : block.AllOps()) {
std::unique_ptr<OperatorBase> local_op = OpRegistry::CreateOp(*op_desc);
op_names_.push_back(op_desc->Type());
OperatorBase* local_op_ptr = local_op.release();
ops_.push_back(local_op_ptr);
continue;
}
}
void ExecutorThreadWorker::CreateThreadResource(
const framework::ProgramDesc& program,
const paddle::platform::Place& place) {
CreateThreadScope(program);
CreateThreadOperators(program);
SetMainProgram(program);
SetPlace(place);
}
void ExecutorThreadWorker::CreateThreadScope(const ProgramDesc& program) {
auto& block = program.Block(0);
PADDLE_ENFORCE_NOT_NULL(
root_scope_, "root_scope should be set before creating thread scope");
thread_scope_ = &root_scope_->NewScope();
for (auto& var : block.AllVars()) {
if (var->Persistable()) {
auto* ptr = root_scope_->Var(var->Name());
InitializeVariable(ptr, var->GetType());
} else {
auto* ptr = thread_scope_->Var(var->Name());
InitializeVariable(ptr, var->GetType());
}
}
}
void ExecutorThreadWorker::SetDataFeed(
const std::shared_ptr<DataFeed>& datafeed) {
thread_reader_ = datafeed;
}
void ExecutorThreadWorker::BindingDataFeedMemory() {
const std::vector<std::string>& input_feed =
thread_reader_->GetUseSlotAlias();
for (auto name : input_feed) {
thread_reader_->AddFeedVar(thread_scope_->Var(name), name);
}
}
void ExecutorThreadWorker::SetFetchVarNames(
const std::vector<std::string>& fetch_var_names) {
fetch_var_names_.clear();
fetch_var_names_.insert(fetch_var_names_.end(), fetch_var_names.begin(),
fetch_var_names.end());
}
void ExecutorThreadWorker::SetDevice() {
#if defined _WIN32 || defined __APPLE__
return;
#else
static unsigned concurrency_cap = std::thread::hardware_concurrency();
int thread_id = this->thread_id_;
if (static_cast<unsigned>(thread_id) < concurrency_cap) {
unsigned proc = thread_id;
cpu_set_t mask;
CPU_ZERO(&mask);
CPU_SET(proc, &mask);
if (-1 == sched_setaffinity(0, sizeof(mask), &mask)) {
VLOG(1) << "WARNING: Failed to set thread affinity for thread "
<< thread_id;
} else {
CPU_ZERO(&mask);
if ((0 != sched_getaffinity(0, sizeof(mask), &mask)) ||
(CPU_ISSET(proc, &mask) == 0)) {
VLOG(3) << "WARNING: Failed to set thread affinity for thread "
<< thread_id;
}
}
} else {
VLOG(1) << "WARNING: Failed to set thread affinity for thread "
<< thread_id;
}
#endif
}
template <typename T>
void print_lod_tensor(std::string var_name, const LoDTensor& lod_tensor) {
auto inspect = lod_tensor.data<T>();
auto element_num = lod_tensor.numel();
std::ostringstream sstream;
sstream << var_name << " (element num " << element_num << "): [";
sstream << inspect[0];
for (int j = 1; j < element_num; ++j) {
sstream << " " << inspect[j];
}
sstream << "]";
std::cout << sstream.str() << std::endl;
}
void print_fetch_var(Scope* scope, std::string var_name) {
const LoDTensor& tensor = scope->FindVar(var_name)->Get<LoDTensor>();
if (std::type_index(tensor.type()) ==
std::type_index(typeid(platform::float16))) {
print_lod_tensor<platform::float16>(var_name, tensor);
} else if (std::type_index(tensor.type()) == std::type_index(typeid(float))) {
print_lod_tensor<float>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(double))) {
print_lod_tensor<double>(var_name, tensor);
} else if (std::type_index(tensor.type()) == std::type_index(typeid(int))) {
print_lod_tensor<int>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(int64_t))) {
print_lod_tensor<int64_t>(var_name, tensor);
} else if (std::type_index(tensor.type()) == std::type_index(typeid(bool))) {
print_lod_tensor<bool>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(uint8_t))) {
print_lod_tensor<uint8_t>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(int16_t))) {
print_lod_tensor<int16_t>(var_name, tensor);
} else if (std::type_index(tensor.type()) ==
std::type_index(typeid(int8_t))) {
print_lod_tensor<int8_t>(var_name, tensor);
} else {
VLOG(1) << "print_fetch_var: unrecognized data type:"
<< tensor.type().name();
}
return;
}
void ExecutorThreadWorker::TrainFiles() {
platform::SetNumThreads(1);
// todo: configurable
SetDevice();
int fetch_var_num = fetch_var_names_.size();
fetch_values_.clear();
fetch_values_.resize(fetch_var_num);
thread_reader_->Start();
int cur_batch;
int batch_cnt = 0;
while ((cur_batch = thread_reader_->Next()) > 0) {
// executor run here
for (auto& op : ops_) {
op->Run(*thread_scope_, place_);
}
++batch_cnt;
thread_scope_->DropKids();
if (debug_ == false || thread_id_ != 0) {
continue;
}
for (int i = 0; i < fetch_var_num; ++i) {
print_fetch_var(thread_scope_, fetch_var_names_[i]);
} // end for (int i = 0...)
} // end while ()
}
void ExecutorThreadWorker::SetThreadId(int tid) { thread_id_ = tid; }
void ExecutorThreadWorker::SetPlace(const platform::Place& place) {
place_ = place;
}
void ExecutorThreadWorker::SetMainProgram(
const ProgramDesc& main_program_desc) {
main_program_.reset(new ProgramDesc(main_program_desc));
}
void ExecutorThreadWorker::SetRootScope(Scope* g_scope) {
root_scope_ = g_scope;
}
} // einit_modelnd namespace framework
} // end namespace paddle
paddle/fluid/framework/executor_thread_worker.h 0 → 100644
浏览文件 @ 65d355a7
/* 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 <map>
#include <memory>
#include <mutex> // NOLINT
#include <set>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "paddle/fluid/framework/data_feed.h"
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
namespace paddle {
namespace framework {
void CreateTensor(Variable* var, proto::VarType::Type var_type);
class ExecutorThreadWorker {
public:
ExecutorThreadWorker()
: thread_id_(-1), root_scope_(NULL), thread_scope_(NULL), debug_(false) {}
~ExecutorThreadWorker() {}
void CreateThreadResource(const framework::ProgramDesc& program,
const paddle::platform::Place& place);
void SetThreadId(int tid);
void SetDebug(const bool debug) { debug_ = debug; }
void SetRootScope(Scope* g_scope);
// set cpu device in this function
// cpu binding is used by default
void SetDevice();
// since we read data into memory that can not be accessed by program
// we need to bind memory of data with corresponding variables in program
// this function should be called after data feed is set
void BindingDataFeedMemory();
// set data feed declared in executor
void SetDataFeed(const std::shared_ptr<DataFeed>& datafeed);
// A multi-thread training function
void TrainFiles();
// set fetch variable names from python interface assigned by users
void SetFetchVarNames(const std::vector<std::string>& fetch_var_names);
private:
void CreateThreadScope(const framework::ProgramDesc& program);
void CreateThreadOperators(const framework::ProgramDesc& program);
void SetMainProgram(const ProgramDesc& main_program_desc);
void SetPlace(const paddle::platform::Place& place);
protected:
// thread index
std::shared_ptr<DataFeed> thread_reader_; // shared queue, thread buffer
int thread_id_;
// operator name
std::vector<std::string> op_names_;
// thread level, local operators for forward and backward
std::vector<OperatorBase*> ops_;
// main program for training
std::unique_ptr<framework::ProgramDesc> main_program_;
// execution place
platform::Place place_;
// root scope for model parameters
Scope* root_scope_;
// a thread scope, father scope is global score which is shared
Scope* thread_scope_;
private:
std::vector<std::string> fetch_var_names_;
std::vector<std::vector<float>> fetch_values_;
bool debug_;
};
} // namespace framework
} // namespace paddle
paddle/fluid/framework/feed_fetch_method.cc
浏览文件 @ 65d355a7
......@@ -16,7 +16,9 @@ limitations under the License. */
#include <string>
#include <vector>
#include "glog/logging.h"
#include "paddle/fluid/framework/var_type.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/platform/place.h"
namespace paddle {
namespace framework {
......@@ -53,5 +55,12 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
return tensor;
}
LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name) {
Variable* var = scope.FindVar(var_name);
PADDLE_ENFORCE(var, "%s no in scope", var_name);
PADDLE_ENFORCE(var->IsType<LoDTensor>(), "Only support lod tensor now.");
return *var->GetMutable<LoDTensor>();
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/feed_fetch_method.h
浏览文件 @ 65d355a7
......@@ -27,5 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
size_t index);
LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name);
} // namespace framework
} // namespace paddle
paddle/fluid/framework/garbage_collector.cc 0 → 100644
浏览文件 @ 65d355a7
// 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 <algorithm>
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
#include "paddle/fluid/framework/garbage_collector.h"
namespace paddle {
namespace framework {
GarbageCollector::GarbageCollector(const platform::Place &place,
size_t max_memory_size)
: max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
garbages_.reset(new GarbageQueue());
dev_ctx_ = platform::DeviceContextPool::Instance().Get(place);
}
CPUGarbageCollector::CPUGarbageCollector(const platform::CPUPlace &place,
size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void CPUGarbageCollector::ClearCallback(const std::function<void()> &callback) {
callback();
}
#ifdef PADDLE_WITH_CUDA
UnsafeFastGPUGarbageCollector::UnsafeFastGPUGarbageCollector(
const platform::CUDAPlace &place, size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void UnsafeFastGPUGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
callback();
}
DefaultStreamGarbageCollector::DefaultStreamGarbageCollector(
const platform::CUDAPlace &place, size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {}
void DefaultStreamGarbageCollector::Wait() const {
static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
->WaitStreamCallback();
}
void DefaultStreamGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
->AddStreamCallback(callback);
}
StreamGarbageCollector::StreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size)
: GarbageCollector(place, max_memory_size) {
platform::CUDADeviceGuard guard(place.device);
PADDLE_ENFORCE(cudaStreamCreate(&stream_));
callback_manager_.reset(new platform::StreamCallbackManager(stream_));
}
StreamGarbageCollector::~StreamGarbageCollector() {
auto place = boost::get<platform::CUDAPlace>(this->dev_ctx_->GetPlace());
platform::CUDADeviceGuard guard(place.device);
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
PADDLE_ENFORCE(cudaStreamDestroy(stream_));
}
cudaStream_t StreamGarbageCollector::stream() const { return stream_; }
void StreamGarbageCollector::Wait() const { callback_manager_->Wait(); }
void StreamGarbageCollector::ClearCallback(
const std::function<void()> &callback) {
callback_manager_->AddCallback(callback);
}
#endif
} // namespace framework
} // namespace paddle
paddle/fluid/framework/garbage_collector.h
浏览文件 @ 65d355a7
......@@ -14,7 +14,6 @@
#pragma once
#include <algorithm>
#include <deque>
#include <functional>
#include <memory>
......@@ -24,134 +23,74 @@
namespace paddle {
namespace framework {
// T should have memory_size() and clear() method
template <typename T>
class GarbageCollector {
public:
GarbageCollector(const platform::Place &place, size_t max_memory_size)
: max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
garbages_.reset(new std::deque<T *>());
dev_ctx_ = platform::DeviceContextPool::Instance().Get(place);
}
using GarbageQueue = std::deque<std::shared_ptr<memory::Allocation>>;
virtual ~GarbageCollector() {}
GarbageCollector(const platform::Place &place, size_t max_memory_size);
void Reset() {
std::lock_guard<std::mutex> guard(mutex_);
garbages_.reset(new std::deque<T *>());
cur_memory_size_ = 0;
}
virtual ~GarbageCollector() = default;
virtual void Wait() const {}
template <typename Container>
void Add(const Container &objs) {
Add(objs, []() {});
}
void Add(Container &&objs);
template <typename Container, typename Callback>
void Add(const Container &objs, Callback &&callback) {
std::shared_ptr<std::deque<T *>> clear_deque;
{
std::lock_guard<std::mutex> guard(mutex_);
for (auto *obj : objs) {
garbages_->push_back(obj);
cur_memory_size_ += obj->memory_size();
}
if (cur_memory_size_ >= max_memory_size_) {
cur_memory_size_ = 0;
clear_deque = garbages_;
garbages_.reset(new std::deque<T *>());
}
}
if (clear_deque != nullptr) {
callback();
ClearCallback([=]() {
for (auto *obj : *clear_deque) obj->clear();
});
}
}
virtual void Wait() const {}
void Add(Container &&objs, Callback &&callback);
protected:
virtual void ClearCallback(const std::function<void()> &callback) = 0;
platform::DeviceContext *dev_ctx_;
std::shared_ptr<std::deque<T *>> garbages_;
std::unique_ptr<GarbageQueue> garbages_;
mutable std::mutex mutex_;
const size_t max_memory_size_;
size_t cur_memory_size_ = 0;
size_t cur_memory_size_{0};
};
template <typename T>
class CPUGarbageCollector : public GarbageCollector<T> {
class CPUGarbageCollector : public GarbageCollector {
public:
CPUGarbageCollector(const platform::CPUPlace &place, size_t max_memory_size)
: GarbageCollector<T>(place, max_memory_size) {}
CPUGarbageCollector(const platform::CPUPlace &place, size_t max_memory_size);
protected:
void ClearCallback(const std::function<void()> &callback) override {
callback();
}
void ClearCallback(const std::function<void()> &callback) override;
};
#ifdef PADDLE_WITH_CUDA
template <typename T>
class DefaultStreamGarbageCollector : public GarbageCollector<T> {
class UnsafeFastGPUGarbageCollector : public GarbageCollector {
public:
DefaultStreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size)
: GarbageCollector<T>(place, max_memory_size) {}
UnsafeFastGPUGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size);
cudaStream_t stream() const {
return static_cast<const platform::CUDADeviceContext *>(this->dev_ctx_)
->stream();
}
protected:
void ClearCallback(const std::function<void()> &callback) override;
};
void Wait() const override {
this->dev_ctx_->Wait();
static_cast<const platform::CUDADeviceContext *>(this->dev_ctx_)
->WaitStreamCallback();
}
class DefaultStreamGarbageCollector : public GarbageCollector {
public:
DefaultStreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size);
void Wait() const override;
protected:
void ClearCallback(const std::function<void()> &callback) override {
static_cast<platform::CUDADeviceContext *>(this->dev_ctx_)
->AddStreamCallback(callback);
}
void ClearCallback(const std::function<void()> &callback) override;
};
template <typename T>
class StreamGarbageCollector : public GarbageCollector<T> {
class StreamGarbageCollector : public GarbageCollector {
public:
StreamGarbageCollector(const platform::CUDAPlace &place,
size_t max_memory_size)
: GarbageCollector<T>(place, max_memory_size) {
PADDLE_ENFORCE(cudaSetDevice(place.device));
PADDLE_ENFORCE(cudaStreamCreate(&stream_));
callback_manager_.reset(new platform::StreamCallbackManager(stream_));
}
size_t max_memory_size);
~StreamGarbageCollector() {
auto place = boost::get<platform::CUDAPlace>(this->dev_ctx_->GetPlace());
PADDLE_ENFORCE(cudaSetDevice(place.device));
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
PADDLE_ENFORCE(cudaStreamDestroy(stream_));
}
~StreamGarbageCollector();
void Wait() const override {
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
std::lock_guard<std::mutex> guard(this->mutex_);
callback_manager_->Wait();
}
void Wait() const override;
cudaStream_t stream() const { return stream_; }
cudaStream_t stream() const;
protected:
void ClearCallback(const std::function<void()> &callback) override {
std::lock_guard<std::mutex> guard(this->mutex_);
callback_manager_->AddCallback(callback);
}
void ClearCallback(const std::function<void()> &callback) override;
private:
cudaStream_t stream_;
......@@ -159,5 +98,33 @@ class StreamGarbageCollector : public GarbageCollector<T> {
};
#endif
template <typename Container>
void GarbageCollector::Add(Container &&objs) {
Add(std::forward<Container>(objs), []() {});
}
template <typename Container, typename Callback>
void GarbageCollector::Add(Container &&objs, Callback &&callback) {
GarbageQueue *garbage_queue = nullptr;
{
std::lock_guard<std::mutex> guard(mutex_);
for (auto &obj : objs) {
if (!obj) continue;
cur_memory_size_ += obj->size();
garbages_->push_back(std::move(obj));
}
if (cur_memory_size_ >= max_memory_size_) {
cur_memory_size_ = 0;
garbage_queue = garbages_.release();
garbages_.reset(new GarbageQueue());
}
}
if (garbage_queue) {
callback();
ClearCallback([garbage_queue]() { delete garbage_queue; });
}
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.cc
浏览文件 @ 65d355a7
......@@ -46,14 +46,16 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
auto* scope = param_scope();
PADDLE_ENFORCE(scope);
std::string type = is_conv3d() ? "conv3d" : "conv2d";
GraphPatternDetector gpd;
auto* conv_input =
gpd.mutable_pattern()
->NewNode(patterns::PDNodeName(name_scope_, "conv_input"))
->AsInput()
->assert_is_op_input("conv2d", "Input");
->assert_is_op_input(type, "Input");
patterns::ConvBias conv_bias_pattern(gpd.mutable_pattern(), name_scope_);
conv_bias_pattern(conv_input);
conv_bias_pattern(conv_input, is_conv3d());
int found_conv_bias_count = 0;
auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph,
Graph* g) {
......@@ -109,7 +111,7 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
desc.SetInput("Filter", std::vector<std::string>({conv_weight->Name()}));
desc.SetInput("Bias", std::vector<std::string>({eltwise_bias->Name()}));
desc.SetOutput("Output", std::vector<std::string>({eltwise_out->Name()}));
desc.SetType("conv2d");
desc.SetType(type);
for (auto& attr : conv->Op()->GetAttrMap()) {
desc.SetAttr(attr.first, attr.second);
......@@ -135,3 +137,5 @@ std::unique_ptr<ir::Graph> ConvBiasFusePass::ApplyImpl(
} // namespace paddle
REGISTER_PASS(conv_bias_mkldnn_fuse_pass,
paddle::framework::ir::ConvBiasFusePass);
REGISTER_PASS(conv3d_bias_mkldnn_fuse_pass,
paddle::framework::ir::Conv3DBiasFusePass);
paddle/fluid/framework/ir/conv_bias_mkldnn_fuse_pass.h
浏览文件 @ 65d355a7
......@@ -26,11 +26,19 @@ namespace ir {
class ConvBiasFusePass : public FusePassBase {
public:
virtual ~ConvBiasFusePass() {}
virtual bool is_conv3d() const { return false; }
protected:
std::unique_ptr<ir::Graph> ApplyImpl(std::unique_ptr<ir::Graph> graph) const;
const std::string name_scope_{"conv_bias_mkldnn_fuse"};
};
/*
* Fuse the Conv3D and Elementwise_add to a Conv3DBiasOp.
*/
class Conv3DBiasFusePass : public ConvBiasFusePass {
public:
bool is_conv3d() const override { return true; }
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/graph.cc
浏览文件 @ 65d355a7
......@@ -38,9 +38,8 @@ void CheckProgram(const ProgramDesc &program) {
switch (role_id) {
case _INT(OpRole::kForward):
if (visit.find(_INT(OpRole::kBackward)) != visit.end()) {
LOG(ERROR)
<< "Cannot add backward operator before forward operator %s."
<< op->Type();
LOG(ERROR) << "Cannot add backward operator before forward operator "
<< op->Type();
}
break;
case _INT(OpRole::kBackward):
......
paddle/fluid/framework/ir/graph.h
浏览文件 @ 65d355a7
......@@ -73,14 +73,21 @@ class Graph {
}
bool Has(const std::string &attr_name) const {
return attrs_.find(attr_name) != attrs_.end();
return attrs_.count(attr_name) > 0;
}
template <typename AttrType>
AttrType &Get(const std::string &attr_name) const {
PADDLE_ENFORCE(Has(attr_name), "%s attr not registered for graph.",
attr_name);
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
try {
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
} catch (boost::bad_any_cast &) {
PADDLE_THROW(
"Invalid attribute type of %s error, expected: %s, actual: %s",
attr_name, typeid(AttrType *).name(),
attrs_.at(attr_name).type().name());
}
}
template <typename AttrType>
......@@ -177,14 +184,13 @@ class Graph {
return nullptr;
}
const ProgramDesc &program() const { return program_; }
std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
const ProgramDesc &program);
void ResolveHazard(
const std::map<std::string, std::vector<ir::Node *>> &var_nodes);
private:
std::map<std::string, std::vector<ir::Node *>> InitFromProgram(
const ProgramDesc &program);
// This method takes ownership of `node`.
ir::Node *AddNode(ir::Node *node) {
PADDLE_ENFORCE(node_set_.find(node) == node_set_.end());
......
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ 65d355a7
......@@ -1030,10 +1030,11 @@ PDNode *patterns::ElewiseAddActInplaceGrad::operator()(
}
PDNode *patterns::ConvBias::operator()(
paddle::framework::ir::PDNode *conv_input) {
paddle::framework::ir::PDNode *conv_input, bool is_conv3d) {
std::string type = is_conv3d ? "conv3d" : "conv2d";
// Create Operators
conv_input->assert_is_op_input("conv2d", "Input");
auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op("conv2d");
conv_input->assert_is_op_input(type, "Input");
auto *conv_op = pattern->NewNode(conv_repr())->assert_is_op(type);
auto *eltiwse_op =
pattern->NewNode(eltwise_repr())->assert_is_op("elementwise_add");
// Create variables
......@@ -1041,11 +1042,11 @@ PDNode *patterns::ConvBias::operator()(
auto *conv_weight_var = pattern->NewNode(conv_weight_repr())
->AsInput()
->assert_is_persistable_var()
->assert_is_op_input("conv2d", "Filter");
->assert_is_op_input(type, "Filter");
// intermediate variable, will be removed in the IR after fuse.
auto *conv_out_var = pattern->NewNode(conv_out_repr())
->AsIntermediate()
->assert_is_only_output_of_op("conv2d")
->assert_is_only_output_of_op(type)
->assert_is_op_input("elementwise_add");
// Bias stored in elementwise_add
auto *eltwise_bias_var = pattern->NewNode(eltwise_bias_repr())
......
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @ 65d355a7
......@@ -623,7 +623,7 @@ struct ElewiseAddActInplaceGrad : public PatternBase {
struct ConvBias : public PatternBase {
ConvBias(PDPattern* pattern, const std::string& name_scope)
: PatternBase(pattern, name_scope, "conv_bias") {}
PDNode* operator()(PDNode* conv_input);
PDNode* operator()(PDNode* conv_input, bool is_conv3d = false);
// declare operator node's name
PATTERN_DECL_NODE(conv);
PATTERN_DECL_NODE(eltwise);
......
paddle/fluid/framework/ir/is_test_pass.cc
浏览文件 @ 65d355a7
......@@ -38,7 +38,7 @@ std::unique_ptr<ir::Graph> IsTestPass::ApplyImpl(
for (const Node* n : graph->Nodes()) {
if (n->IsOp()) {
auto* op = n->Op();
if (op->HasAttr("is_test")) {
if (op->HasAttr("is_test") || op->HasProtoAttr("is_test")) {
op->SetAttr("is_test", true);
} else if (std::find(begin(op_list), end(op_list), op->Type()) !=
end(op_list)) {
......
paddle/fluid/framework/ir/mkldnn_placement_pass.cc
浏览文件 @ 65d355a7
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/ir/mkldnn_placement_pass.h"
#include <string>
namespace paddle {
namespace framework {
......@@ -21,9 +22,19 @@ namespace ir {
std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
VLOG(3) << "Aplies MKL-DNN placement strategy.";
const auto& op_types_list =
Get<std::unordered_set<std::string>>("mkldnn_enabled_op_types");
for (const Node* n : graph->Nodes()) {
if (n->IsOp() && n->Op()->HasAttr("use_mkldnn")) {
n->Op()->SetAttr("use_mkldnn", true);
if (n->IsOp()) {
auto* op = n->Op();
if (op->HasAttr("use_mkldnn") || op->HasProtoAttr("use_mkldnn")) {
if (op_types_list.empty()) {
op->SetAttr("use_mkldnn", true);
} else if (std::find(op_types_list.begin(), op_types_list.end(),
n->Name()) != op_types_list.end()) {
op->SetAttr("use_mkldnn", true);
}
}
}
}
return graph;
......@@ -33,5 +44,5 @@ std::unique_ptr<ir::Graph> MKLDNNPlacementPass::ApplyImpl(
} // namespace framework
} // namespace paddle
REGISTER_PASS(mkldnn_placement_pass,
paddle::framework::ir::MKLDNNPlacementPass);
REGISTER_PASS(mkldnn_placement_pass, paddle::framework::ir::MKLDNNPlacementPass)
.RequirePassAttr("mkldnn_enabled_op_types");
paddle/fluid/framework/ir/node.cc
浏览文件 @ 65d355a7
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/ir/node.h"
#include "paddle/fluid/framework/op_info.h"
namespace paddle {
namespace framework {
......@@ -24,10 +25,11 @@ constexpr char Node::kControlDepVarName[];
const char Node::kControlDepVarName[] = "__control_var";
#endif
std::unique_ptr<Node> CreateNodeForTest(const std::string& name,
std::unique_ptr<Node> CreateNodeForTest(const std::string &name,
Node::Type type) {
return std::unique_ptr<Node>(new Node(name, type));
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/pass.h
浏览文件 @ 65d355a7
......@@ -51,11 +51,18 @@ class Pass {
AttrType &Get(const std::string &attr_name) const {
PADDLE_ENFORCE(attrs_.find(attr_name) != attrs_.end(),
"%s attr not registered for pass.", attr_name);
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
try {
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
} catch (boost::bad_any_cast &) {
PADDLE_THROW(
"Invalid attribute type of %s error, expected: %s, actual: %s",
attr_name, typeid(AttrType *).name(),
attrs_.at(attr_name).type().name());
}
}
bool Has(const std::string &attr_name) const {
return attrs_.find(attr_name) != attrs_.end();
return attrs_.count(attr_name) > 0;
}
void Erase(const std::string &attr_name) {
......
paddle/fluid/framework/naive_executor.cc
浏览文件 @ 65d355a7
......@@ -21,42 +21,11 @@
#include "paddle/fluid/framework/naive_executor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/framework/variable_helper.h"
#include "paddle/fluid/string/pretty_log.h"
namespace paddle {
namespace framework {
// These code can be shared with Executor.
static void InitializeVariable(Variable *var, proto::VarType::Type var_type) {
if (var_type == proto::VarType::LOD_TENSOR) {
var->GetMutable<LoDTensor>();
} else if (var_type == proto::VarType::SELECTED_ROWS) {
var->GetMutable<SelectedRows>();
} else if (var_type == proto::VarType::FEED_MINIBATCH) {
var->GetMutable<FeedFetchList>();
} else if (var_type == proto::VarType::FETCH_LIST) {
var->GetMutable<FeedFetchList>();
} else if (var_type == proto::VarType::STEP_SCOPES) {
var->GetMutable<std::vector<framework::Scope *>>();
} else if (var_type == proto::VarType::LOD_RANK_TABLE) {
var->GetMutable<LoDRankTable>();
} else if (var_type == proto::VarType::LOD_TENSOR_ARRAY) {
var->GetMutable<LoDTensorArray>();
} else if (var_type == proto::VarType::PLACE_LIST) {
var->GetMutable<platform::PlaceList>();
} else if (var_type == proto::VarType::READER) {
var->GetMutable<ReaderHolder>();
} else if (var_type == proto::VarType::RAW) {
// GetMutable will be called in operator
} else {
PADDLE_THROW(
"Variable type %d is not in "
"[LOD_TENSOR, SELECTED_ROWS, FEED_MINIBATCH, FETCH_LIST, "
"LOD_RANK_TABLE, PLACE_LIST, READER, CHANNEL, RAW]",
var_type);
}
}
void NaiveExecutor::Prepare(Scope *scope, const ProgramDesc &program_desc,
int block_id, bool with_feed_fetch_ops) {
if (!scope) {
......
paddle/fluid/framework/ngraph_bridge.cc
浏览文件 @ 65d355a7
......@@ -12,28 +12,109 @@ 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. */
#ifdef PADDLE_WITH_NGRAPH
#include <algorithm>
#include <functional>
#include <vector>
#include "paddle/fluid/framework/ngraph_bridge.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/enforce.h"
#include "ngraph/ngraph.hpp"
namespace paddle {
namespace framework {
static std::shared_ptr<ngraph::Node> GetNode(
const std::shared_ptr<OperatorBase>& op, const std::string name,
const VariableNameMap& var_map,
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
ngb_node_map) {
auto& var_names = var_map.at(name);
PADDLE_ENFORCE_EQ(var_names.size(), 1,
"op %s name %s expects one associated var", op->Type(),
name);
if (ngb_node_map->find(var_names[0]) != ngb_node_map->end()) {
return (*ngb_node_map)[var_names[0]];
} else {
return nullptr;
}
}
static std::shared_ptr<ngraph::Node> GetInputNode(
const std::shared_ptr<OperatorBase>& op, const std::string name,
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
ngb_node_map) {
return GetNode(op, name, op->Inputs(), ngb_node_map);
}
static std::shared_ptr<ngraph::Node> GetOutputNode(
const std::shared_ptr<OperatorBase>& op, const std::string name,
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
ngb_node_map) {
return GetNode(op, name, op->Outputs(), ngb_node_map);
}
static void SetOutputNode(
const std::shared_ptr<OperatorBase>& op, const std::string name,
std::shared_ptr<ngraph::Node> node,
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
ngb_node_map) {
auto& var_names = op->Outputs().at(name);
if (var_names.size() == 1) {
(*ngb_node_map)[var_names[0]] = node;
} else if (var_names.size() == 0) {
(*ngb_node_map)[""] = node;
} else {
PADDLE_THROW("name %s has more than 1 var_names.", name);
}
}
static bool HasOutput(const std::shared_ptr<OperatorBase>& op,
const std::string name) {
auto& outputs = op->Outputs();
if (outputs.find(name) == outputs.end()) return false;
return outputs.at(name).size() > 0;
}
template <typename T>
static void BuildBinaryNode(
const std::shared_ptr<OperatorBase>& op,
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
ngb_node_map) {
auto x = GetInputNode(op, "X", ngb_node_map);
auto y = GetInputNode(op, "Y", ngb_node_map);
auto out = std::make_shared<T>(x, y);
SetOutputNode(op, "Out", out, ngb_node_map);
}
template <typename T>
static void BuildUnaryNode(
const std::shared_ptr<OperatorBase>& op,
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
ngb_node_map) {
auto input = GetInputNode(op, "X", ngb_node_map);
auto out = std::make_shared<T>(input);
SetOutputNode(op, "Out", out, ngb_node_map);
}
std::map<std::string,
std::function<void(const std::shared_ptr<OperatorBase>&,
std::shared_ptr<std::unordered_map<
std::string, std::shared_ptr<ngraph::Node>>>)>>
NgraphBridge::NG_NODE_MAP = {};
NgraphBridge::NG_NODE_MAP = {{"relu", BuildUnaryNode<ngraph::op::Relu>},
{"tanh", BuildUnaryNode<ngraph::op::Tanh>}};
void NgraphBridge::build_graph(const std::shared_ptr<OperatorBase>& op) {
void NgraphBridge::BuildNgNode(const std::shared_ptr<OperatorBase>& op) {
auto& op_type = op->Type();
NG_NODE_MAP[op_type](op, ngb_node_map);
NG_NODE_MAP[op_type](op, ngb_node_map_);
}
} // namespace framework
} // namespace paddle
#endif
paddle/fluid/framework/ngraph_bridge.h
浏览文件 @ 65d355a7
......@@ -14,22 +14,18 @@ limitations under the License. */
#pragma once
#ifdef PADDLE_WITH_NGRAPH
#include <algorithm>
#include <map>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/enforce.h"
#include "ngraph/ngraph.hpp"
#include "ngraph/node.hpp"
namespace paddle {
namespace framework {
class OperatorBase;
class NgraphBridge {
public:
static std::map<
......@@ -43,16 +39,15 @@ class NgraphBridge {
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
var_node_map)
: ngb_node_map(var_node_map) {}
: ngb_node_map_(var_node_map) {}
void build_graph(const std::shared_ptr<OperatorBase>& op);
void BuildNgNode(const std::shared_ptr<OperatorBase>& op);
private:
std::shared_ptr<
std::unordered_map<std::string, std::shared_ptr<ngraph::Node>>>
ngb_node_map;
ngb_node_map_;
};
} // namespace framework
} // namespace paddle
#endif
paddle/fluid/framework/ngraph_operator.h
浏览文件 @ 65d355a7
......@@ -14,39 +14,32 @@ limitations under the License. */
#pragma once
#ifdef PADDLE_WITH_NGRAPH
#include <algorithm>
#include <atomic>
#include <string>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/attribute.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/ngraph_bridge.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_kernel_type.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/variant.h"
#include "ngraph/ngraph.hpp"
#include "ngraph/type/element_type.hpp"
namespace paddle {
namespace framework {
class FusedOperator : public OperatorBase {
class NgraphOperator : public OperatorBase {
public:
static std::vector<
std::vector<std::vector<std::unique_ptr<OperatorBase>>::iterator>>
FusedOpIntervals(
NgraphOpIntervals(
std::vector<std::unique_ptr<paddle::framework::OperatorBase>>* ops);
explicit FusedOperator(
explicit NgraphOperator(
const ProgramDesc& prog, size_t block_id,
std::vector<std::unique_ptr<OperatorBase>>::iterator start,
std::vector<std::unique_ptr<OperatorBase>>::iterator end,
......@@ -69,4 +62,3 @@ class FusedOperator : public OperatorBase {
};
} // namespace framework
} // namespace paddle
#endif
paddle/fluid/framework/op_desc.cc
浏览文件 @ 65d355a7
......@@ -237,6 +237,23 @@ void OpDesc::SetOutput(const std::string &param_name,
this->outputs_[param_name] = args;
}
bool OpDesc::HasProtoAttr(const std::string &name) const {
auto &op_info = OpInfoMap::Instance();
if (op_info.Has(desc_.type())) {
auto op_info_ptr = op_info.Get(desc_.type());
if (op_info_ptr.HasOpProtoAndChecker()) {
const proto::OpProto &proto = op_info_ptr.Proto();
for (int i = 0; i != proto.attrs_size(); ++i) {
const proto::OpProto::Attr &attr = proto.attrs(i);
if (attr.name() == name) {
return true;
}
}
}
}
return false;
}
proto::AttrType OpDesc::GetAttrType(const std::string &name) const {
auto it = attrs_.find(name);
PADDLE_ENFORCE(it != attrs_.end(), "Attribute %s is not found", name);
......
paddle/fluid/framework/op_desc.h
浏览文件 @ 65d355a7
......@@ -65,6 +65,8 @@ class OpDesc {
return attrs_.find(name) != attrs_.end();
}
bool HasProtoAttr(const std::string &name) const;
proto::AttrType GetAttrType(const std::string &name) const;
std::vector<std::string> AttrNames() const;
......
paddle/fluid/framework/op_kernel_type.cc 0 → 100644
浏览文件 @ 65d355a7
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/op_kernel_type.h"
namespace paddle {
namespace framework {
size_t OpKernelType::Hash::operator()(const OpKernelType& key) const {
int cur_loc = 0;
int place = key.place_.which();
cur_loc += OpKernelType::kPlaceBits;
int data_type = static_cast<int>(key.data_type_) << cur_loc;
cur_loc += OpKernelType::kPrimaryDTypeBits;
int data_layout = static_cast<int>(key.data_layout_) << cur_loc;
cur_loc += OpKernelType::kLayoutBits;
int library_type = static_cast<int>(key.library_type_) << cur_loc;
cur_loc += OpKernelType::kLibBits;
int customized_value = key.customized_type_value_;
PADDLE_ENFORCE(customized_value < (1 << OpKernelType::kCustomizeBits));
customized_value = customized_value << cur_loc;
cur_loc += OpKernelType::kCustomizeBits;
PADDLE_ENFORCE(cur_loc < 64);
std::hash<int> hasher;
return hasher(place + data_type + data_layout + library_type +
customized_value);
}
bool OpKernelType::operator==(const OpKernelType& o) const {
return platform::places_are_same_class(place_, o.place_) &&
data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
library_type_ == o.library_type_ &&
customized_type_value_ == o.customized_type_value_;
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/op_kernel_type.h
浏览文件 @ 65d355a7
......@@ -24,54 +24,55 @@ limitations under the License. */
namespace paddle {
namespace framework {
struct OpKernelType {
struct Hash {
size_t operator()(const OpKernelType& key) const {
int place = key.place_.which();
int data_type = static_cast<int>(key.data_type_) << LEFT_SHIFT;
int data_layout = static_cast<int>(key.data_layout_) << (LEFT_SHIFT * 2);
int library_type = static_cast<int>(key.library_type_)
<< (LEFT_SHIFT * 3);
std::hash<int> hasher;
return hasher(place + data_type + data_layout + library_type);
}
};
class OpKernelType {
public:
constexpr static int kDefaultCustomizedTypeValue = 0;
// place, data_type, library_type kinds less than 2^8
constexpr static int LEFT_SHIFT = 8;
proto::VarType::Type data_type_;
DataLayout data_layout_;
platform::Place place_;
LibraryType library_type_;
// In total should be smaller than 64.
constexpr static int kPlaceBits = 4;
constexpr static int kPrimaryDTypeBits = 8;
constexpr static int kLayoutBits = 4;
constexpr static int kLibBits = 4;
constexpr static int kCustomizeBits = 4;
OpKernelType(proto::VarType::Type data_type, platform::Place place,
DataLayout data_layout = DataLayout::kAnyLayout,
LibraryType library_type = LibraryType::kPlain)
LibraryType library_type = LibraryType::kPlain,
int customized_type_value = kDefaultCustomizedTypeValue)
: data_type_(data_type),
data_layout_(data_layout),
place_(place),
library_type_(library_type) {}
library_type_(library_type),
customized_type_value_(customized_type_value) {}
OpKernelType(proto::VarType::Type data_type,
const platform::DeviceContext& dev_ctx,
DataLayout data_layout = DataLayout::kAnyLayout,
LibraryType library_type = LibraryType::kPlain)
LibraryType library_type = LibraryType::kPlain,
int customized_type_value = kDefaultCustomizedTypeValue)
: data_type_(data_type),
data_layout_(data_layout),
place_(dev_ctx.GetPlace()),
library_type_(library_type) {}
library_type_(library_type),
customized_type_value_(customized_type_value) {}
virtual ~OpKernelType() {}
struct Hash {
size_t operator()(const OpKernelType& key) const;
};
size_t hash_key() const { return Hash()(*this); }
bool operator==(const OpKernelType& o) const {
return platform::places_are_same_class(place_, o.place_) &&
data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
library_type_ == o.library_type_;
}
bool operator==(const OpKernelType& o) const;
bool operator!=(const OpKernelType& o) const { return !(*this == o); }
proto::VarType::Type data_type_;
DataLayout data_layout_;
platform::Place place_;
LibraryType library_type_;
int customized_type_value_;
};
inline std::ostream& operator<<(std::ostream& os,
......
paddle/fluid/framework/operator.cc
浏览文件 @ 65d355a7
......@@ -695,6 +695,12 @@ static void CheckTensorNANOrInf(const std::string& name,
"Tensor %s contains NAN", name);
}
void OperatorWithKernel::RuntimeInferShape(const Scope& scope,
const platform::Place& place) const {
RuntimeInferShapeContext infer_shape_ctx(*this, scope);
this->InferShape(&infer_shape_ctx);
}
void OperatorWithKernel::RunImpl(const Scope& scope,
const platform::Place& place) const {
RuntimeInferShapeContext infer_shape_ctx(*this, scope);
......@@ -873,6 +879,8 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
t = &(var->Get<SelectedRows>().value());
}
if (t != nullptr) {
PADDLE_ENFORCE(t->IsInitialized(), "Input %s is not initialized: %s",
ipt_name, DebugString());
int tmp = static_cast<int>(ToDataType(t->type()));
PADDLE_ENFORCE(
tmp == data_type || data_type == -1,
......
paddle/fluid/framework/operator.h
浏览文件 @ 65d355a7
......@@ -128,6 +128,8 @@ class OperatorBase {
virtual std::vector<std::string> OutputVars(bool has_intermediate) const;
void SetIsCalledByExecutor(bool x) { run_by_executor_ = x; }
virtual void RuntimeInferShape(const Scope& scope,
const platform::Place& place) const {}
protected:
std::string type_;
......@@ -348,6 +350,9 @@ class OperatorWithKernel : public OperatorBase {
OpInfoMap::Instance().Get(Type()).infer_shape_(ctx);
}
void RuntimeInferShape(const Scope& scope,
const platform::Place& place) const override;
protected:
virtual OpKernelType GetExpectedKernelType(const ExecutionContext& ctx) const;
virtual OpKernelType GetKernelTypeForVar(
......
paddle/fluid/framework/parallel_executor.h
浏览文件 @ 65d355a7
......@@ -14,7 +14,6 @@ limitations under the License. */
#pragma once
#include <atomic>
#include <string>
#include <unordered_map>
#include <unordered_set>
......@@ -29,10 +28,6 @@ limitations under the License. */
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/device_context.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/framework/details/reference_count_pass.h"
#endif
namespace paddle {
namespace framework {
......@@ -75,24 +70,7 @@ class ParallelExecutor {
private:
void BCastParamsToDevices(const std::unordered_set<std::string> &vars) const;
std::unique_ptr<ParallelExecutorPrivate> member_;
#ifdef PADDLE_WITH_CUDA
// ref_cnts_ is only initialized when ParallelExecutor constructs, and then
// keeps unchanged
// Before each iteration, cur_ref_cnts_ is reset to ref_cnts_
details::DeviceReferenceCountMap ref_cnts_;
details::AtomicDeviceReferenceCountMap cur_ref_cnts_;
details::DeviceGarbageCollectorMap gcs_;
void ResetReferenceCount() {
for (auto &pair1 : ref_cnts_) {
for (auto &pair2 : *(pair1.second)) {
(*(cur_ref_cnts_[pair1.first]))[pair2.first] = pair2.second;
}
}
}
#endif
ParallelExecutorPrivate *member_;
};
} // namespace framework
......
paddle/fluid/framework/scope.cc
浏览文件 @ 65d355a7
......@@ -38,6 +38,10 @@ DEFINE_double(
"Memory size threshold (GB) when the garbage collector clear tensors."
"Disabled when this value is less than 0");
DEFINE_bool(fast_eager_deletion_mode, false,
"Fast eager deletion mode. If enabled, memory would release "
"immediately without waiting GPU kernel ends.");
// When in inference scenario, the scopes will not be written by two threads in
// a mean time, but a scope may be read by multiple threads concurrently, and
// the mutex will cause serious performance issue.
......@@ -58,6 +62,8 @@ int64_t GetEagerDeletionThreshold() {
(static_cast<int64_t>(1) << 30));
}
bool IsFastEagerDeletionModeEnabled() { return FLAGS_fast_eager_deletion_mode; }
Scope::~Scope() { DropKids(); }
Scope& Scope::NewScope() const {
......
paddle/fluid/framework/scope.h
浏览文件 @ 65d355a7
......@@ -27,6 +27,7 @@ namespace paddle {
namespace framework {
int64_t GetEagerDeletionThreshold();
bool IsFastEagerDeletionModeEnabled();
class Scope;
......
paddle/fluid/framework/selected_rows.h
浏览文件 @ 65d355a7
......@@ -32,8 +32,7 @@ namespace framework {
class SelectedRows {
/*
* @brief We can use the SelectedRows structure to reproduce a sparse table.
* A sparse table is a key-value structure that the key is an `int64_t`
* number,
* A sparse table is a key-value structure that the key is an `int64_t`,
* and the value is a Tensor which the first dimension is 0.
* You can use the following interface to operate the sparse table, and you
* can find
......
paddle/fluid/framework/tensor.h
浏览文件 @ 65d355a7
......@@ -158,6 +158,10 @@ class Tensor {
const std::shared_ptr<memory::Allocation>& Holder() const { return holder_; }
size_t offset() const { return offset_; }
std::shared_ptr<memory::Allocation> MoveMemoryHolder() {
return std::move(holder_);
}
private:
/*! holds the memory block if allocated. */
std::shared_ptr<memory::Allocation> holder_;
......
paddle/fluid/framework/variable_helper.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/framework/variable_helper.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/imperative/CMakeLists.txt 0 → 100644
浏览文件 @ 65d355a7
cc_library(layer SRCS layer.cc DEPS proto_desc operator)
cc_library(tracer SRCS tracer.cc DEPS proto_desc)
cc_library(engine SRCS engine.cc)
paddle/fluid/imperative/engine.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/imperative/engine.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/imperative/layer.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/imperative/layer.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/imperative/tracer.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/imperative/tracer.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/inference/analysis/analysis_pass.h
浏览文件 @ 65d355a7
......@@ -46,8 +46,6 @@ class AnalysisPass {
protected:
// User should implement these.
virtual void RunImpl(Argument* argument) = 0;
Argument* argument_{nullptr};
};
} // namespace analysis
......
paddle/fluid/inference/io.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/inference/io.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/inference/utils/visualizer.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/inference/utils/visualizer.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/bpr_loss_op.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/bpr_loss_op.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/concat_mkldnn_op.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/concat_op.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/conv_op.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/conv_transpose_mkldnn_op.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/crop_op.cu
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/cudnn_lstm_op.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/cudnn_lstm_op.cu.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/distributed/collective_client.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/distributed/collective_client.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/distributed/collective_server.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/distributed/collective_server.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/distributed/parameter_prefetch.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/expand_op.cu
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/math/blas.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/math/prelu.cu 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/math/prelu.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/mean_op.cu
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/merge_selected_rows_op.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/merge_selected_rows_op.cu.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/merge_selected_rows_op.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/pad2d_op.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/pad2d_op.cu
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/pad_op.cu
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/prelu_op.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/prelu_op.cu 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.cu 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/psroi_pool_op.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/softmax_op.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/sum_op.cu
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/yolov3_loss_op.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/operators/yolov3_loss_op.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/assert.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/cpu_info.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/cpu_info.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/enforce.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/float16.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/gpu_info.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/gpu_info.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/init.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/mkldnn_reuse.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/nccl_helper.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/platform/stream_callback_manager.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/pybind/async_executor_py.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/pybind/async_executor_py.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/pybind/imperative.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/pybind/imperative.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/string/CMakeLists.txt
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/string/split.h 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/fluid/string/split_test.cc 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/dataset/image.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/dataset/wmt16.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/async_executor.py 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/average.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/clip.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/data_feed_desc.py 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/data_feeder.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/imperative/__init__.py 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/imperative/base.py 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/imperative/layers.py 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/initializer.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/metrics.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/param_attr.py
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/tests/demo/async_executor.py 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/tests/unittests/test_bpr_loss_op.py 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
python/paddle/fluid/tests/unittests/test_imperative.py 0 → 100644
浏览文件 @ 65d355a7
此差异已折叠。
python/setup.py.in
浏览文件 @ 65d355a7
此差异已折叠。
tools/print_signatures.py
浏览文件 @ 65d355a7
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或