m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #16 from PaddlePaddle/develop

Browse Source 
merge to local
revert-16807-engine2-interface
lujun 7 years ago committed by GitHub
parent 8ba177593b ed61d67c73
commit cc29bec6e6
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
382 changed files with 13382 additions and 2265 deletions
Show Stats Download Patch File Download Diff File

8
CMakeLists.txt
Unescape View File

@ -64,6 +64,7 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(WITH_PSLIB "Compile with pslib support" OFF) option(WITH_PSLIB "Compile with pslib support" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF) option(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF) option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
# TODO(Superjomn) Remove WITH_ANAKIN option if not needed latter.
option(WITH_ANAKIN "Compile with Anakin library" OFF) option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF) option(ANAKIN_BUILD_FAT_BIN "Build anakin cuda fat-bin lib for all device plantform, ignored when WITH_ANAKIN=OFF" OFF)
option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON) option(ANAKIN_BUILD_CROSS_PLANTFORM "Build anakin lib for any nvidia device plantform. ignored when WITH_ANAKIN=OFF" ON)
@ -190,7 +191,14 @@ include(configure) # add paddle env configuration
if(WITH_GPU) if(WITH_GPU)
include(cuda) include(cuda)
include(tensorrt) include(tensorrt)
include(anakin_subgraph)
endif() endif()
if(WITH_GPU AND NOT WIN32)
message(STATUS "add dgc lib.")
include(external/dgc)
endif()
if(WITH_MKL OR WITH_MKLML) if(WITH_MKL OR WITH_MKLML)
include(external/anakin) include(external/anakin)
elseif() elseif()

8
CONTRIBUTING.md
Unescape View File

@ -156,7 +156,7 @@ python \
This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the verbose range of 0 to 3, so you will see above example VLOG message, which is in level 3. This suggests that we output overall messages in lower verbose levels, so they display with higher probability. When coding C++, please follow the verbose level convention as follows: This will enable VLOG messages generated by `buddy_allocator.{h,cc}` and in the verbose range of 0 to 3, so you will see above example VLOG message, which is in level 3. This suggests that we output overall messages in lower verbose levels, so they display with higher probability. When coding C++, please follow the verbose level convention as follows:
- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework) - verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/framework)
- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators) - verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators)
- verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/platform) - verbose level 5: [memory](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/memory), [platform](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/platform)
- verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/math) - verbose level 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/fluid/operators/math/)

32
cmake/anakin_subgraph.cmake
Unescape View File

@ -0,0 +1,32 @@
if(NOT WITH_GPU)
return()
endif()
set(ANAKIN_ROOT "/usr" CACHE PATH "ANAKIN ROOT")
find_path(ANAKIN_INCLUDE_DIR anakin_config.h
PATHS ${ANAKIN_ROOT} ${ANAKIN_ROOT}/include
$ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/include
NO_DEFAULT_PATH
)
find_library(ANAKIN_LIBRARY NAMES libanakin_saber_common.so libanakin.so
PATHS ${ANAKIN_ROOT}
$ENV{ANAKIN_ROOT} $ENV{ANAKIN_ROOT}/lib
NO_DEFAULT_PATH
DOC "Path to ANAKIN library.")
if(ANAKIN_INCLUDE_DIR AND ANAKIN_LIBRARY)
if(WITH_DSO)
set(ANAKIN_FOUND ON)
endif(WITH_DSO)
else()
set(ANAKIN_FOUND OFF)
endif()
if(ANAKIN_FOUND)
message(STATUS "Current ANAKIN header is ${ANAKIN_INCLUDE_DIR}/anakin_config.h. ")
include_directories(${ANAKIN_ROOT}/include)
include_directories(${ANAKIN_ROOT}/include/saber)
link_directories(${ANAKIN_ROOT})
add_definitions(-DPADDLE_WITH_ANAKIN)
endif()

42
cmake/external/dgc.cmake vendored
Unescape View File

@ -0,0 +1,42 @@
# 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(ExternalProject)
SET(DGC_SOURCES_DIR "${THIRD_PARTY_PATH}/dgc")
SET(DGC_INSTALL_DIR "${THIRD_PARTY_PATH}/install/dgc")
SET(DGC_INCLUDE_DIR "${DGC_INSTALL_DIR}/include" CACHE PATH "dgc include directory." FORCE)
SET(DGC_LIBRARIES "${DGC_INSTALL_DIR}/lib/libdgc.a" CACHE FILEPATH "dgc library." FORCE)
INCLUDE_DIRECTORIES(${DGC_INCLUDE_DIR})
ExternalProject_Add(
extern_dgc
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/PaddlePaddle/Fleet"
GIT_TAG "2d04dc3800cdd0601f1b65d547dabcc60b0cf9dc"
SOURCE_DIR "${DGC_SOURCES_DIR}"
CONFIGURE_COMMAND ""
BUILD_COMMAND cd collective && make -j
INSTALL_COMMAND mkdir -p ${DGC_INSTALL_DIR}/lib/ ${DGC_INCLUDE_DIR}/dgc
&& cp ${DGC_SOURCES_DIR}/collective/build/lib/libdgc.a ${DGC_LIBRARIES}
&& cp ${DGC_SOURCES_DIR}/collective/build/include/dgc.h ${DGC_INCLUDE_DIR}/dgc/
BUILD_IN_SOURCE 1
)
ADD_LIBRARY(dgc SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET dgc PROPERTY IMPORTED_LOCATION ${DGC_LIBRARIES})
ADD_DEPENDENCIES(dgc extern_dgc)
LIST(APPEND external_project_dependencies dgc)

33
cmake/external/ngraph.cmake vendored
Unescape View File

@ -57,20 +57,25 @@ SET(NGRAPH_TBB_LIB ${NGRAPH_LIB_DIR}/${NGRAPH_TBB_LIB_NAME})
ExternalProject_Add( ExternalProject_Add(
${NGRAPH_PROJECT} ${NGRAPH_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS} ${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_PROJECT} ${MKLML_PROJECT} DEPENDS ${MKLDNN_PROJECT} ${MKLML_PROJECT}
GIT_REPOSITORY ${NGRAPH_GIT_REPO} GIT_REPOSITORY ${NGRAPH_GIT_REPO}
GIT_TAG ${NGRAPH_GIT_TAG} GIT_TAG ${NGRAPH_GIT_TAG}
PREFIX ${NGRAPH_SOURCES_DIR} PREFIX ${NGRAPH_SOURCES_DIR}
UPDATE_COMMAND "" UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR} CMAKE_GENERATOR ${CMAKE_GENERATOR}
CMAKE_ARGS -DNGRAPH_UNIT_TEST_ENABLE=FALSE CMAKE_GENERATOR_PLATFORM ${CMAKE_GENERATOR_PLATFORM}
CMAKE_ARGS -DNGRAPH_TOOLS_ENABLE=FALSE CMAKE_GENERATOR_TOOLSET ${CMAKE_GENERATOR_TOOLSET}
CMAKE_ARGS -DNGRAPH_INTERPRETER_ENABLE=FALSE CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
CMAKE_ARGS -DNGRAPH_DEX_ONLY=TRUE CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE} CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${NGRAPH_INSTALL_DIR}
CMAKE_ARGS -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR} CMAKE_ARGS -DNGRAPH_UNIT_TEST_ENABLE=FALSE
CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR} CMAKE_ARGS -DNGRAPH_TOOLS_ENABLE=FALSE
CMAKE_ARGS -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib CMAKE_ARGS -DNGRAPH_INTERPRETER_ENABLE=FALSE
CMAKE_ARGS -DNGRAPH_DEX_ONLY=TRUE
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
CMAKE_ARGS -DMKLDNN_INCLUDE_DIR=${MKLDNN_INC_DIR}
CMAKE_ARGS -DMKLDNN_LIB_DIR=${MKLDNN_INSTALL_DIR}/${CMAKE_INSTALL_LIBDIR}
CMAKE_ARGS -DMKLML_LIB_DIR=${MKLML_INSTALL_DIR}/lib
) )
add_dependencies(ngraph ${NGRAPH_PROJECT}) add_dependencies(ngraph ${NGRAPH_PROJECT})

9
cmake/inference_lib.cmake
Unescape View File

@ -131,6 +131,15 @@ elseif (NOT CBLAS_FOUND OR WIN32)
) )
endif () endif ()
if (WITH_GPU AND NOT WIN32)
set(dgc_dir "${FLUID_INSTALL_DIR}/third_party/install/dgc")
copy(dgc_lib
SRCS ${DGC_INSTALL_DIR}/lib ${DGC_INSTALL_DIR}/include
DSTS ${dgc_dir} ${dgc_dir}
DEPS dgc)
endif()
if (WITH_MKLDNN) if (WITH_MKLDNN)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn") set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
copy(mkldnn_lib copy(mkldnn_lib

2
cmake/operators.cmake
Unescape View File

@ -110,7 +110,7 @@ function(op_library TARGET)
# Define operators that don't need pybind here. # Define operators that don't need pybind here.
foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op"
"tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op" "tensor_array_read_write_op" "tensorrt_engine_op" "conv_fusion_op"
"fusion_transpose_flatten_concat_op" "fusion_conv_inception_op" "sync_batch_norm_op") "fusion_transpose_flatten_concat_op" "fusion_conv_inception_op" "sync_batch_norm_op" "dgc_op")
if ("${TARGET}" STREQUAL "${manual_pybind_op}") if ("${TARGET}" STREQUAL "${manual_pybind_op}")
set(pybind_flag 1) set(pybind_flag 1)
endif() endif()

1
cmake/tensorrt.cmake
Unescape View File

@ -33,5 +33,6 @@ if(TENSORRT_FOUND)
message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. " message(STATUS "Current TensorRT header is ${TENSORRT_INCLUDE_DIR}/NvInfer.h. "
"Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ") "Current TensorRT version is v${TENSORRT_MAJOR_VERSION}. ")
include_directories(${TENSORRT_INCLUDE_DIR}) include_directories(${TENSORRT_INCLUDE_DIR})
link_directories(${TENSORRT_LIBRARY})
add_definitions(-DPADDLE_WITH_TENSORRT) add_definitions(-DPADDLE_WITH_TENSORRT)
endif() endif()

8
paddle/fluid/API.spec
Unescape View File

@ -211,7 +211,7 @@ paddle.fluid.layers.mean (ArgSpec(args=['x', 'name'], varargs=None, keywords=Non
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)), ('document', 'ccd37fa6b53f074adbfb732d738c4c2d')) 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)), ('document', 'ccd37fa6b53f074adbfb732d738c4c2d'))
paddle.fluid.layers.sigmoid_cross_entropy_with_logits (ArgSpec(args=['x', 'label', 'ignore_index', 'name', 'normalize'], varargs=None, keywords=None, defaults=(-100, None, False)), ('document', '180c284317ea45ef89a460d8d79c0b72')) paddle.fluid.layers.sigmoid_cross_entropy_with_logits (ArgSpec(args=['x', 'label', 'ignore_index', 'name', 'normalize'], varargs=None, keywords=None, defaults=(-100, None, False)), ('document', '180c284317ea45ef89a460d8d79c0b72'))
paddle.fluid.layers.maxout (ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '891870d069a6aea746d34cc53b61690c')) paddle.fluid.layers.maxout (ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '891870d069a6aea746d34cc53b61690c'))
paddle.fluid.layers.space_to_depth (ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '5f207ae10589ebe38a63575ef6ff8e1e')) paddle.fluid.layers.space_to_depth (ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'a9221eaef53884a00654e028551b78e2'))
paddle.fluid.layers.affine_grid (ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '51def402b8910e163cbace9d0c0526ed')) paddle.fluid.layers.affine_grid (ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '51def402b8910e163cbace9d0c0526ed'))
paddle.fluid.layers.sequence_reverse (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '77a6d80aa5551ca70324fc975c44507f')) paddle.fluid.layers.sequence_reverse (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '77a6d80aa5551ca70324fc975c44507f'))
paddle.fluid.layers.affine_channel (ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name', 'act'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None, None)), ('document', 'ab84fdc6dc60f3ad9aa397e6007e3bf9')) paddle.fluid.layers.affine_channel (ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name', 'act'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None, None)), ('document', 'ab84fdc6dc60f3ad9aa397e6007e3bf9'))
@ -483,6 +483,11 @@ paddle.fluid.optimizer.LarsMomentumOptimizer.apply_gradients (ArgSpec(args=['sel
paddle.fluid.optimizer.LarsMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f')) paddle.fluid.optimizer.LarsMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.LarsMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754')) paddle.fluid.optimizer.LarsMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.LarsMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea')) paddle.fluid.optimizer.LarsMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.optimizer.DGCMomentumOptimizer.__init__ (ArgSpec(args=['self', 'learning_rate', 'momentum', 'rampup_begin_step', 'rampup_step', 'sparsity', 'use_nesterov', 'local_grad_clip_norm', 'num_trainers', 'regularization', 'name'], varargs=None, keywords=None, defaults=(1, [0.999], False, None, None, None, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.DGCMomentumOptimizer.apply_gradients (ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None), ('document', 'bfe7305918552aaecfdaa22411dbe871'))
paddle.fluid.optimizer.DGCMomentumOptimizer.backward (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None)), ('document', 'ba3a113d0229ff7bc9d39bda0a6d947f'))
paddle.fluid.optimizer.DGCMomentumOptimizer.get_opti_var_name_list (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.optimizer.DGCMomentumOptimizer.minimize (ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '35fd5d3330c97903528c7e0dacc7f6ea'))
paddle.fluid.backward.append_backward (ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '1a79bd7d10ae54ca763ec81bca36ba24')) paddle.fluid.backward.append_backward (ArgSpec(args=['loss', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '1a79bd7d10ae54ca763ec81bca36ba24'))
paddle.fluid.regularizer.L1DecayRegularizer.__init__ (ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754')) paddle.fluid.regularizer.L1DecayRegularizer.__init__ (ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.regularizer.L2DecayRegularizer.__init__ (ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754')) paddle.fluid.regularizer.L2DecayRegularizer.__init__ (ArgSpec(args=['self', 'regularization_coeff'], varargs=None, keywords=None, defaults=(0.0,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@ -520,6 +525,7 @@ paddle.fluid.unique_name.guard (ArgSpec(args=['new_generator'], varargs=None, ke
paddle.fluid.recordio_writer.convert_reader_to_recordio_file (ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', '65c7523e86f0c50bb729b01667f36310')) paddle.fluid.recordio_writer.convert_reader_to_recordio_file (ArgSpec(args=['filename', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', '65c7523e86f0c50bb729b01667f36310'))
paddle.fluid.recordio_writer.convert_reader_to_recordio_files (ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', 'bc643f0f5f1b9db57ff0d8a57d379bd7')) paddle.fluid.recordio_writer.convert_reader_to_recordio_files (ArgSpec(args=['filename', 'batch_per_file', 'reader_creator', 'feeder', 'compressor', 'max_num_records', 'feed_order'], varargs=None, keywords=None, defaults=(Compressor.Snappy, 1000, None)), ('document', 'bc643f0f5f1b9db57ff0d8a57d379bd7'))
paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope paddle.fluid.Scope Scope() -> paddle.fluid.core._Scope
paddle.fluid.install_check.run_check (ArgSpec(args=[], varargs=None, keywords=None, defaults=None), ('document', '66b7c84a17ed32fec2df9628367be2b9'))
paddle.reader.cache (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', '1676886070eb607cb608f7ba47be0d3c')) paddle.reader.cache (ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None), ('document', '1676886070eb607cb608f7ba47be0d3c'))
paddle.reader.map_readers (ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None), ('document', '77cbadb09df588e21e5cc0819b69c87d')) paddle.reader.map_readers (ArgSpec(args=['func'], varargs='readers', keywords=None, defaults=None), ('document', '77cbadb09df588e21e5cc0819b69c87d'))
paddle.reader.buffered (ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None), ('document', '0d6186f109feceb99f60ec50a0a624cb')) paddle.reader.buffered (ArgSpec(args=['reader', 'size'], varargs=None, keywords=None, defaults=None), ('document', '0d6186f109feceb99f60ec50a0a624cb'))

7
paddle/fluid/framework/CMakeLists.txt
Unescape View File

@ -63,7 +63,7 @@ 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) 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) 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(garbage_collector SRCS garbage_collector.cc DEPS device_context memory gflags glog)
cc_library(reader SRCS reader.cc DEPS lod_tensor ddim) cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
cc_test(reader_test SRCS reader_test.cc DEPS reader) cc_test(reader_test SRCS reader_test.cc DEPS reader)
@ -164,6 +164,8 @@ else()
set(NGRAPH_EXE_DEPS) set(NGRAPH_EXE_DEPS)
endif() endif()
cc_library(executor_gc_helper SRCS executor_gc_helper.cc DEPS scope proto_desc operator garbage_collector)
if(WITH_DISTRIBUTE) if(WITH_DISTRIBUTE)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog
lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS}) lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS})
@ -174,7 +176,7 @@ else()
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op) cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif() endif()
target_link_libraries(executor garbage_collector while_op_helper) target_link_libraries(executor while_op_helper executor_gc_helper)
cc_library(parallel_executor SRCS parallel_executor.cc DEPS cc_library(parallel_executor SRCS parallel_executor.cc DEPS
threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor
@ -194,6 +196,7 @@ cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_con
cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry cc_test(var_type_inference_test SRCS var_type_inference_test.cc DEPS op_registry
proto_desc) proto_desc)
cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS op_registry proto_desc op_info memory_optimize_helper) cc_test(inplace_op_inference_test SRCS inplace_op_inference_test.cc DEPS op_registry proto_desc op_info memory_optimize_helper)
cc_library(selected_rows SRCS selected_rows.cc DEPS tensor) 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(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)

5
paddle/fluid/framework/details/CMakeLists.txt
Unescape View File

@ -5,6 +5,7 @@ cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_h
cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory) cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry) cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry) cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(fetch_barrier_op_handle SRCS fetch_barrier_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper) cc_library(multi_devices_helper SRCS multi_devices_helper.cc DEPS graph graph_helper)
cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper) cc_library(multi_devices_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper)
@ -22,7 +23,7 @@ endif()
if(WITH_GPU) if(WITH_GPU)
nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory 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) dynload_cuda variable_visitor dgc)
nv_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory nv_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor) dynload_cuda variable_visitor)
if(WITH_DISTRIBUTE) if(WITH_DISTRIBUTE)
@ -72,7 +73,7 @@ cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS grap
cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass) cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle 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 fused_broadcast_op_handle) scale_loss_grad_op_handle rpc_op_handle fetch_barrier_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle) cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle)

38
paddle/fluid/framework/details/all_reduce_deps_pass.cc
Unescape View File

@ -13,6 +13,7 @@
// limitations under the License. // limitations under the License.
#include <algorithm> #include <algorithm>
#include <memory>
#include <string> #include <string>
#include <unordered_map> #include <unordered_map>
#include <unordered_set> #include <unordered_set>
@ -41,8 +42,7 @@ VarHandle* GetValidInput(const OpHandleBase* a) {
return nullptr; return nullptr;
} }
std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl( void AllReduceDepsPass::ApplyImpl(ir::Graph* graph) const {
std::unique_ptr<ir::Graph> graph) const {
auto graph_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph); auto graph_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
// get vars order // get vars order
@ -52,13 +52,28 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
// Note that must assert topology sort is stable // Note that must assert topology sort is stable
auto& ops = graph->Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs); auto& ops = graph->Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
for (auto* op_desc : ops) { for (auto* op_desc : ops) {
auto outputs = op_desc->Outputs(); try {
for (auto& o_it : outputs) { bool is_bk_op =
for (auto& v : o_it.second) { // values static_cast<bool>(boost::get<int>(op_desc->GetAttr(
vars[v] = order; OpProtoAndCheckerMaker::OpRoleAttrName())) &
static_cast<int>(OpRole::kBackward));
if (!is_bk_op) continue;
auto backward_vars =
boost::get<std::vector<std::string>>(op_desc->GetNullableAttr(
OpProtoAndCheckerMaker::OpRoleVarAttrName()));
PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
auto outputs = op_desc->Outputs();
for (auto& o_it : outputs) {
for (auto& v : o_it.second) { // values
vars[v] = order;
VLOG(1) << "in all_reduce_deps_pass:" << v;
}
} }
order++;
} catch (boost::bad_get e) {
} }
order++;
} }
std::vector<OpHandleBase*> dist_ops; std::vector<OpHandleBase*> dist_ops;
@ -70,7 +85,8 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
} }
} }
VLOG(10) << "dist_ops size:" << dist_ops.size() << std::endl; VLOG(10) << "dist_ops size:" << dist_ops.size()
<< ", outputs size:" << vars.size() << ", ops size:" << ops.size();
std::sort(dist_ops.begin(), dist_ops.end(), [&](OpHandleBase* op1, std::sort(dist_ops.begin(), dist_ops.end(), [&](OpHandleBase* op1,
OpHandleBase* op2) { OpHandleBase* op2) {
@ -83,6 +99,10 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
auto l_it = vars.find(i0->name()); auto l_it = vars.find(i0->name());
auto r_it = vars.find(i1->name()); auto r_it = vars.find(i1->name());
PADDLE_ENFORCE(l_it != vars.end() && r_it != vars.end(),
"can't find var's name %s and %s in opdesc", i0->name(),
i1->name());
if (l_it->second < r_it->second) return true; if (l_it->second < r_it->second) return true;
if (l_it->second == r_it->second) { if (l_it->second == r_it->second) {
@ -110,8 +130,6 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
VLOG(10) << "pre_op:" << pre_op->DebugString() VLOG(10) << "pre_op:" << pre_op->DebugString()
<< ", op:" << op->DebugString(); << ", op:" << op->DebugString();
} }
return graph;
} }
} // namespace details } // namespace details

3
paddle/fluid/framework/details/all_reduce_deps_pass.h
Unescape View File

@ -24,8 +24,7 @@ namespace details {
// TODO(gongwb): overlap allreduce with backward computation. // TODO(gongwb): overlap allreduce with backward computation.
class AllReduceDepsPass : public ir::Pass { class AllReduceDepsPass : public ir::Pass {
protected: protected:
std::unique_ptr<ir::Graph> ApplyImpl( void ApplyImpl(ir::Graph* graph) const override;
std::unique_ptr<ir::Graph> graph) const override;
}; };
} // namespace details } // namespace details

200
paddle/fluid/framework/details/all_reduce_op_handle.cc
Unescape View File

@ -16,6 +16,13 @@
#include "paddle/fluid/framework/details/container_cast.h" #include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/reduce_and_gather.h" #include "paddle/fluid/framework/details/reduce_and_gather.h"
#include "paddle/fluid/framework/details/variable_visitor.h" #include "paddle/fluid/framework/details/variable_visitor.h"
#include "paddle/fluid/framework/operator.h"
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "dgc/dgc.h"
#endif
#include "paddle/fluid/platform/gpu_info.h"
#include "paddle/fluid/platform/profiler.h" #include "paddle/fluid/platform/profiler.h"
// asynchronous nccl allreduce or synchronous issue: // asynchronous nccl allreduce or synchronous issue:
@ -33,11 +40,14 @@ namespace details {
AllReduceOpHandle::AllReduceOpHandle(ir::Node *node, AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
const std::vector<Scope *> &local_scopes, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places, const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs) const platform::NCCLContextMap *ctxs,
bool is_encoded, int nranks)
: OpHandleBase(node), : OpHandleBase(node),
local_scopes_(local_scopes), local_scopes_(local_scopes),
places_(places), places_(places),
nccl_ctxs_(ctxs) { nccl_ctxs_(ctxs),
is_encoded_(is_encoded),
nranks_(nranks) {
if (nccl_ctxs_) { if (nccl_ctxs_) {
for (auto &p : places_) { for (auto &p : places_) {
this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p)); this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
@ -51,7 +61,185 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
: OpHandleBase(node), local_scopes_(local_scopes), places_(places) {} : OpHandleBase(node), local_scopes_(local_scopes), places_(places) {}
#endif #endif
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
void AllReduceOpHandle::RunImplEncoded() {
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(
in_var_handles.size(), places_.size(),
"The NoDummyInputSize should be equal to the number of places.");
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
"The NoDummyInputSize and NoDummyOutputSize should be equal.");
std::vector<const LoDTensor *> ins;
std::vector<LoDTensor *> outs;
int k = -1;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &local_scope =
local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto original_name =
paddle::framework::GradOriginalVarName(in_var_handles[i]->name());
auto encode_var_name = original_name + g_dgc_encoded;
auto *in_var = local_scope->FindVar(encode_var_name);
PADDLE_ENFORCE_NOT_NULL(in_var);
auto &in = in_var->Get<LoDTensor>();
ins.emplace_back(&in);
auto *out = local_scope->FindVar(out_var_handles[i]->name())
->GetMutable<LoDTensor>();
outs.emplace_back(out);
if (k < 0) {
k = GetKValue(in_var_handles[i]->name());
}
}
PADDLE_ENFORCE(platform::is_gpu_place(ins[0]->place()));
PADDLE_ENFORCE(platform::is_gpu_place(outs[0]->place()));
PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
int dtype = -1;
size_t in_numel = 0;
size_t out_numel = 0;
PADDLE_ENFORCE(nranks_ > 1);
std::vector<std::function<void()>> all_reduce_calls;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &place = places_[i];
auto &in = *ins[i];
void *in_tensor_buf = const_cast<void *>(in.data<void>());
auto &out = *outs[i];
float *out_tensor_buf = out.data<float>();
dtype = (dtype == -1) ? platform::ToNCCLDataType(in.type()) : dtype;
in_numel = (in_numel == 0) ? static_cast<size_t>(in.numel()) : in_numel;
PADDLE_ENFORCE(in_numel % 2 == 0);
PADDLE_ENFORCE(in_numel / 2 == static_cast<size_t>(k));
out_numel = (out_numel == 0) ? static_cast<size_t>(out.numel()) : out_numel;
int dev_id = boost::get<platform::CUDAPlace>(place).device;
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
auto comm = nccl_ctx.comm_;
auto &allocator =
platform::DeviceTemporaryAllocator::Instance().Get(place, stream);
int encode_size = 2 * k * sizeof(int);
// dgc use ncclAllGather to get all the encoded data
// so the buffer need nranks.
int buf_size = nranks_ * encode_size;
auto tmp_ious_data = allocator.Allocate(buf_size);
void *gather_buff = reinterpret_cast<void *>(tmp_ious_data->ptr());
VLOG(10) << "in_numel:" << in_numel << ", out_numel:" << out_numel
<< ", nranks:" << nranks_ << ", gather_buf size:" << buf_size
<< ", k:" << k << ", place:" << place << ", dtype:" << dtype;
all_reduce_calls.emplace_back([=] {
PADDLE_ENFORCE(paddle::communication::dgc::sparseAllGReduce(
in_tensor_buf, gather_buff, k, out_tensor_buf, out_numel, comm,
stream));
});
}
this->RunAndRecordEvent([&] {
if (all_reduce_calls.size() == 1UL) {
// Do not use NCCLGroup when manage NCCL by per thread per device
all_reduce_calls[0]();
} else {
platform::NCCLGroupGuard guard;
for (auto &call : all_reduce_calls) {
call();
}
}
});
if (FLAGS_sync_nccl_allreduce) {
for (auto &p : places_) {
int dev_id = boost::get<platform::CUDAPlace>(p).device;
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
cudaError_t e_sync = cudaStreamSynchronize(stream);
if (e_sync != 0) {
LOG(FATAL) << "cudaStreamSynchronize " << cudaGetErrorString(e_sync);
}
cudaError_t e_get = cudaGetLastError();
if (e_get != 0) {
LOG(FATAL) << "cudaGetLastError " << cudaGetErrorString(e_get)
<< " errno:" << e_get;
}
}
}
}
int AllReduceOpHandle::GetKValue(const std::string &grad_name) {
auto original_name = paddle::framework::GradOriginalVarName(grad_name);
auto var_name = original_name + g_dgc_k;
PADDLE_ENFORCE(local_scopes_.size() > 0);
auto *scope = local_scopes_[0];
auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto var = local_scope->FindVar(var_name);
PADDLE_ENFORCE_NOT_NULL(var);
auto tensor = var->Get<LoDTensor>().data<float>();
return *tensor;
}
#endif
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
bool AllReduceOpHandle::IsEncoded() {
if (!is_encoded_) {
return false;
}
auto counter_name = g_dgc_counter_name;
auto step_name = g_dgc_rampup_begin_step;
PADDLE_ENFORCE(local_scopes_.size() > 0);
auto *scope = local_scopes_[0];
auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto count_var = local_scope->FindVar(counter_name);
auto step_var = local_scope->FindVar(step_name);
if (count_var == nullptr || step_var == nullptr) {
PADDLE_THROW("not find count_var:%s or step_var:%s", counter_name,
step_var);
}
float count = *count_var->Get<LoDTensor>().data<float>();
float step = *step_var->Get<LoDTensor>().data<float>();
if (static_cast<int>(count) < static_cast<int>(step)) {
VLOG(10) << "in all_reduce currentstep:" << count
<< " < rampup_begin_step:" << step
<< " so not use sparse all reduce";
return false;
}
return true;
}
#else
bool AllReduceOpHandle::IsEncoded() { return false; }
#endif
void AllReduceOpHandle::RunImpl() { void AllReduceOpHandle::RunImpl() {
if (!IsEncoded()) {
RunImplNormal();
return;
}
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
RunImplEncoded();
#else
PADDLE_THROW("Not compiled with CUDA");
#endif
}
void AllReduceOpHandle::RunImplNormal() {
platform::RecordEvent record_event(Name()); platform::RecordEvent record_event(Name());
WaitInputVarGenerated(); WaitInputVarGenerated();
@ -72,6 +260,8 @@ void AllReduceOpHandle::RunImpl() {
auto &lod_tensor = auto &lod_tensor =
local_scope.FindVar(in_var_handles[i]->name())->Get<LoDTensor>(); local_scope.FindVar(in_var_handles[i]->name())->Get<LoDTensor>();
lod_tensors.emplace_back(&lod_tensor); lod_tensors.emplace_back(&lod_tensor);
VLOG(10) << "place:" << i << ", input_name:" << in_var_handles[i]->name()
<< ", out_name:" << out_var_handles[i]->name();
PADDLE_ENFORCE_EQ(in_var_handles[i]->name(), out_var_handles[i]->name(), PADDLE_ENFORCE_EQ(in_var_handles[i]->name(), out_var_handles[i]->name(),
"The name of input and output should be equal."); "The name of input and output should be equal.");
} }
@ -99,13 +289,17 @@ void AllReduceOpHandle::RunImpl() {
auto &nccl_ctx = nccl_ctxs_->at(dev_id); auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream(); auto stream = nccl_ctx.stream();
auto comm = nccl_ctx.comm_; auto comm = nccl_ctx.comm_;
VLOG(10) << "before all reduce buffer:" << buffer << ", numel:" << numel
<< ", dev_id:" << dev_id << ", dtype:" << dtype
<< ", place:" << p;
all_reduce_calls.emplace_back([=] { all_reduce_calls.emplace_back([=] {
PADDLE_ENFORCE(platform::dynload::ncclAllReduce( PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
buffer, buffer, numel, static_cast<ncclDataType_t>(dtype), ncclSum, buffer, buffer, numel, static_cast<ncclDataType_t>(dtype), ncclSum,
comm, stream)); comm, stream));
}); });
} }
this->RunAndRecordEvent([&] { this->RunAndRecordEvent([&] {
if (all_reduce_calls.size() == 1UL) { if (all_reduce_calls.size() == 1UL) {
// Do not use NCCLGroup when manage NCCL by per thread per device // Do not use NCCLGroup when manage NCCL by per thread per device

16
paddle/fluid/framework/details/all_reduce_op_handle.h
Unescape View File

@ -28,11 +28,19 @@ namespace paddle {
namespace framework { namespace framework {
namespace details { namespace details {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
constexpr char g_dgc_counter_name[] = "__g_dgc_counter__";
constexpr char g_dgc_rampup_begin_step[] = "__g_rampup_begin_step__";
constexpr char g_dgc_encoded[] = "__dgc_encoded__";
constexpr char g_dgc_k[] = "__dgc_k__";
#endif
struct AllReduceOpHandle : public OpHandleBase { struct AllReduceOpHandle : public OpHandleBase {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes, AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places, const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs); const platform::NCCLContextMap *ctxs,
bool is_encoded = false, int nranks = -1);
#else #else
AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes, AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places); const std::vector<platform::Place> &places);
@ -50,8 +58,14 @@ struct AllReduceOpHandle : public OpHandleBase {
std::vector<Scope *> local_scopes_; std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_; std::vector<platform::Place> places_;
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
void RunImplEncoded();
const platform::NCCLContextMap *nccl_ctxs_; const platform::NCCLContextMap *nccl_ctxs_;
bool is_encoded_{false};
int nranks_{-1};
int GetKValue(const std::string &grad_name);
#endif #endif
void RunImplNormal();
bool IsEncoded();
}; };
} // namespace details } // namespace details

7
paddle/fluid/framework/details/alloc_continuous_space_for_grad_pass.cc
Unescape View File

@ -46,8 +46,7 @@ static framework::proto::VarType::Type kDefaultDtype =
class AllocContinuousSpaceForGradPass : public ir::Pass { class AllocContinuousSpaceForGradPass : public ir::Pass {
protected: protected:
std::unique_ptr<ir::Graph> ApplyImpl( void ApplyImpl(ir::Graph *graph) const override {
std::unique_ptr<ir::Graph> graph) const override {
ir::Graph &result = *graph; ir::Graph &result = *graph;
auto &places = Get<const std::vector<platform::Place>>(kPlaces); auto &places = Get<const std::vector<platform::Place>>(kPlaces);
@ -65,7 +64,7 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
if (params_grads.size() == 0) { if (params_grads.size() == 0) {
VLOG(10) << "Doesn't find gradients"; VLOG(10) << "Doesn't find gradients";
return std::move(graph); return;
} }
std::unordered_map<std::string, ir::Node *> vars; std::unordered_map<std::string, ir::Node *> vars;
@ -124,8 +123,6 @@ class AllocContinuousSpaceForGradPass : public ir::Pass {
InitFusedVarsAndAllocSpaceForVars(places, local_scopes, vars, InitFusedVarsAndAllocSpaceForVars(places, local_scopes, vars,
fused_var_name, params_grads); fused_var_name, params_grads);
return std::move(graph);
} }
template <typename AttrType> template <typename AttrType>

17
paddle/fluid/framework/details/build_strategy.cc
Unescape View File

@ -204,15 +204,16 @@ bool BuildStrategy::IsMultiDevPass(const std::string &pass_name) const {
return framework::details::MultiDevSSAGraphBuilder().count(pass_name) > 0; return framework::details::MultiDevSSAGraphBuilder().count(pass_name) > 0;
} }
std::unique_ptr<ir::Graph> BuildStrategy::Apply( ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
std::unique_ptr<ir::Graph> graph, const std::vector<platform::Place> &places,
const std::vector<platform::Place> &places, const std::string &loss_var_name,
const std::string &loss_var_name, const std::vector<Scope *> &local_scopes, const std::vector<Scope *> &local_scopes,
const size_t &nranks, const size_t &nranks,
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const { const bool use_cuda,
platform::NCCLContextMap *nccl_ctxs) const {
#else #else
const bool use_cuda) const { const bool use_cuda) const {
#endif #endif
// Create a default one if not finalized by user. // Create a default one if not finalized by user.
CreatePassesFromStrategy(false); CreatePassesFromStrategy(false);
@ -265,7 +266,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
} }
} }
VLOG(3) << "Start Apply Pass " << pass->Type(); VLOG(3) << "Start Apply Pass " << pass->Type();
graph = pass->Apply(std::move(graph)); graph = pass->Apply(graph);
VLOG(3) << "Finish Apply Pass " << pass->Type(); VLOG(3) << "Finish Apply Pass " << pass->Type();
} }
return graph; return graph;

15
paddle/fluid/framework/details/build_strategy.h
Unescape View File

@ -120,16 +120,15 @@ struct BuildStrategy {
// Apply the passes built by the pass_builder_. The passes will be // Apply the passes built by the pass_builder_. The passes will be
// applied to the Program and output an ir::Graph. // applied to the Program and output an ir::Graph.
std::unique_ptr<ir::Graph> Apply(std::unique_ptr<ir::Graph> graph, ir::Graph *Apply(ir::Graph *graph, const std::vector<platform::Place> &places,
const std::vector<platform::Place> &places, const std::string &loss_var_name,
const std::string &loss_var_name, const std::vector<Scope *> &local_scopes,
const std::vector<Scope *> &local_scopes, const size_t &nranks,
const size_t &nranks,
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32) #if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const bool use_cuda, const bool use_cuda,
platform::NCCLContextMap *nccl_ctxs) const; platform::NCCLContextMap *nccl_ctxs) const;
#else #else
const bool use_cuda) const; const bool use_cuda) const;
#endif #endif
// If set true, ParallelExecutor would build the main_program into multiple // If set true, ParallelExecutor would build the main_program into multiple

23
paddle/fluid/framework/details/eager_deletion_pass.cc
Unescape View File

@ -22,14 +22,9 @@
#include "paddle/fluid/framework/details/computation_op_handle.h" #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_op_handle.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h" #include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/garbage_collector.h"
#include "paddle/fluid/framework/ir/graph_helper.h" #include "paddle/fluid/framework/ir/graph_helper.h"
DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
"Fraction of eager deletion. If less than 1.0, all variables in "
"the program would be sorted according to its memory size, and "
"only the FLAGS_memory_fraction_of_eager_deletion of the largest "
"variables would be deleted.");
namespace paddle { namespace paddle {
namespace framework { namespace framework {
namespace details { namespace details {
@ -175,12 +170,10 @@ static OpToVarNameSetMap ShrinkGCVars(
class EagerDeletionPass : public ir::Pass { class EagerDeletionPass : public ir::Pass {
protected: protected:
std::unique_ptr<ir::Graph> ApplyImpl( void ApplyImpl(ir::Graph *graph) const override;
std::unique_ptr<ir::Graph> graph) const override;
}; };
std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl( void EagerDeletionPass::ApplyImpl(ir::Graph *graph) const {
std::unique_ptr<ir::Graph> graph) const {
auto &ref_cnts = auto &ref_cnts =
Get<std::vector<AtomicReferenceCountMap>>(kRuntimeReferenceCount); Get<std::vector<AtomicReferenceCountMap>>(kRuntimeReferenceCount);
PADDLE_ENFORCE(ref_cnts.empty(), PADDLE_ENFORCE(ref_cnts.empty(),
@ -206,8 +199,9 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
} }
} }
op_vars_map = ShrinkGCVars(op_vars_map, vars, places, double memory_fraction = framework::GetEagerDeletionMemoryFraction();
FLAGS_memory_fraction_of_eager_deletion);
op_vars_map = ShrinkGCVars(op_vars_map, vars, places, memory_fraction);
for (auto &pair : op_vars_map) { for (auto &pair : op_vars_map) {
auto *op = pair.first; auto *op = pair.first;
@ -239,13 +233,12 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
eager_deletion_op->AddOutput(dummy_leaf); eager_deletion_op->AddOutput(dummy_leaf);
} }
VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = " VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = " << memory_fraction;
<< FLAGS_memory_fraction_of_eager_deletion;
VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)"; VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
auto while_op_eager_deletion_pass = auto while_op_eager_deletion_pass =
ir::PassRegistry::Instance().Get("while_op_eager_deletion_pass"); ir::PassRegistry::Instance().Get("while_op_eager_deletion_pass");
return while_op_eager_deletion_pass->Apply(std::move(graph)); while_op_eager_deletion_pass->Apply(graph);
} }
} // namespace details } // namespace details

140
paddle/fluid/framework/details/early_delete_op_handle.h
Unescape View File

@ -1,140 +0,0 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/details/var_handle.h"
#include "paddle/fluid/framework/garbage_collector.h"
#include "paddle/fluid/framework/lod_tensor_array.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 {
class EarlyDeleteOpHandle : public OpHandleBase {
public:
EarlyDeleteOpHandle(ir::Node* node, const Scope* scope,
const platform::Place& place,
const std::vector<std::string>& names,
GarbageCollector* gc)
: OpHandleBase(node),
scope_(scope),
place_(place),
names_(names),
gc_(gc) {
#ifdef PADDLE_WITH_CUDA
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(place);
PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
}
#endif
}
~EarlyDeleteOpHandle() {
#ifdef PADDLE_WITH_CUDA
if (IsStreamGarabageCollector()) {
auto gpu_place = boost::get<platform::CUDAPlace>(dev_ctx_->GetPlace());
PADDLE_ENFORCE(cudaSetDevice(gpu_place.device));
PADDLE_ENFORCE(cudaEventDestroy(event_));
}
#endif
}
std::string Name() const override { return "early_delete"; }
protected:
void RunImpl() override {
std::vector<std::shared_ptr<memory::Allocation>> tensors;
auto* local_scope = scope_->FindVar(kLocalExecScopeName)->Get<Scope*>();
for (auto& var_name : names_) {
auto* var = local_scope->FindVar(var_name);
PADDLE_ENFORCE(var != nullptr,
string::Sprintf("Local Scope not has var %s", var_name));
if (var->IsType<LoDTensor>()) {
tensors.emplace_back(var->GetMutable<LoDTensor>()->MoveMemoryHolder());
} else if (var->IsType<SelectedRows>()) {
tensors.emplace_back(var->GetMutable<SelectedRows>()
->mutable_value()
->MoveMemoryHolder());
} else if (var->IsType<LoDTensorArray>()) {
LoDTensorArray* tensor_array = var->GetMutable<LoDTensorArray>();
for (auto& tensor : *tensor_array) {
tensors.emplace_back(tensor.MoveMemoryHolder());
}
}
}
if (!tensors.empty()) {
ClearTensors(tensors);
}
}
private:
void ClearTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
if (platform::is_cpu_place(place_)) {
ClearCPUTensors(tensors);
} else {
ClearGPUTensors(tensors);
}
}
void ClearCPUTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
auto* gc = dynamic_cast<CPUGarbageCollector*>(gc_);
if (gc != nullptr) {
gc->Add(tensors);
}
}
void ClearGPUTensors(
const std::vector<std::shared_ptr<memory::Allocation>>& tensors) {
#ifdef PADDLE_WITH_CUDA
auto* gc = dynamic_cast<StreamGarbageCollector*>(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*>(gc_) != nullptr;
#endif
}
const Scope* scope_;
const platform::Place place_;
std::vector<std::string> names_;
GarbageCollector* gc_;
#ifdef PADDLE_WITH_CUDA
platform::CUDADeviceContext* dev_ctx_;
cudaEvent_t event_;
#endif
};
} // namespace details
} // namespace framework
} // namespace paddle

66
paddle/fluid/framework/details/fetch_barrier_op_handle.cc
Unescape View File

@ -0,0 +1,66 @@
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/fetch_barrier_op_handle.h"
#include <string>
namespace paddle {
namespace framework {
namespace details {
FetchBarrierOpHandle::FetchBarrierOpHandle(
ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places)
// fetch_barrier op always run on place0, but output on all places.
: OpHandleBase(node),
op_(framework::OpRegistry::CreateOp(*node->Op())),
local_scopes_(local_scopes),
places_(places),
run_scope_(local_scopes[0]),
place_(places[0]) {
for (auto &p : places) {
this->SetDeviceContext(p, platform::DeviceContextPool::Instance().Get(p));
}
}
bool FetchBarrierOpHandle::IsMultiDeviceTransfer() {
// override IsMultiDeviceTransfer to return true
return true;
}
void FetchBarrierOpHandle::RunImpl() {
WaitInputVarGenerated(place_);
auto run_func = [this]() {
op_->Run(*run_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>(), place_);
};
if (is_lock_and_record_event_free_) {
run_func();
} else {
this->RunAndRecordEvent(run_func);
}
}
bool FetchBarrierOpHandle::NeedWait(VarHandleBase *in_var) {
bool need_wait =
in_var && in_var->GeneratedOp() &&
in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_);
return need_wait;
}
std::string FetchBarrierOpHandle::Name() const { return op_->Type(); }
} // namespace details
} // namespace framework
} // namespace paddle

61
paddle/fluid/framework/details/fetch_barrier_op_handle.h
Unescape View File

@ -0,0 +1,61 @@
// 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 <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
namespace framework {
namespace details {
// **NOTE**: fetch_barrier op is special it outputs all recved variables on
// all places if there are multiple places, must init with
// multiple dev_ctxes_ !!!!
struct FetchBarrierOpHandle : public OpHandleBase {
public:
FetchBarrierOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places);
bool IsMultiDeviceTransfer() override;
std::string Name() const override;
protected:
void RunImpl() override;
bool NeedWait(VarHandleBase *in_var) override;
private:
std::unique_ptr<OperatorBase> op_;
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
Scope *run_scope_;
platform::Place place_;
bool is_lock_and_record_event_free_{false};
};
} // namespace details
} // namespace framework
} // namespace paddle

6
paddle/fluid/framework/details/fuse_all_reduce_op_pass.cc
Unescape View File

@ -28,8 +28,7 @@ namespace details {
class FuseAllReduceOpPass : public ir::Pass { class FuseAllReduceOpPass : public ir::Pass {
protected: protected:
std::unique_ptr<ir::Graph> ApplyImpl( void ApplyImpl(ir::Graph *graph) const override {
std::unique_ptr<ir::Graph> graph) const override {
ir::Graph &result = *graph; ir::Graph &result = *graph;
auto &places = Get<const std::vector<platform::Place>>(kPlaces); auto &places = Get<const std::vector<platform::Place>>(kPlaces);
@ -71,7 +70,7 @@ class FuseAllReduceOpPass : public ir::Pass {
VLOG(10) << "Find all_reduce_ops: " << all_reduce_ops.size(); VLOG(10) << "Find all_reduce_ops: " << all_reduce_ops.size();
if (all_reduce_ops.size() == 0) { if (all_reduce_ops.size() == 0) {
return std::move(graph); return;
} }
PADDLE_ENFORCE_EQ(all_reduce_ops.size(), grads.size(), PADDLE_ENFORCE_EQ(all_reduce_ops.size(), grads.size(),
@ -99,7 +98,6 @@ class FuseAllReduceOpPass : public ir::Pass {
group_all_reduce_ops, &result); group_all_reduce_ops, &result);
#endif #endif
} }
return std::move(graph);
} }
void InsertFusedAllReduce(const std::vector<platform::Place> &places, void InsertFusedAllReduce(const std::vector<platform::Place> &places,

179
paddle/fluid/framework/details/inplace_op_pass.cc
Unescape View File

@ -17,6 +17,8 @@
#include <deque> #include <deque>
#include <iterator> #include <iterator>
#include <memory> #include <memory>
#include <queue>
#include <sstream>
#include <stack> #include <stack>
#include <string> #include <string>
#include <unordered_map> #include <unordered_map>
@ -142,20 +144,19 @@ void InplacePass::InitSSAGraphNodes() const {
} }
} }
std::unique_ptr<ir::Graph> InplacePass::ApplyImpl( void InplacePass::ApplyImpl(ir::Graph* graph) const {
std::unique_ptr<ir::Graph> graph) const {
var_nodes_.clear(); var_nodes_.clear();
view_.Build(graph.get()); view_.Build(graph);
InitSSAGraphNodes(); InitSSAGraphNodes();
auto cnt = 0;
for (auto* op : view_.AllOps()) { for (auto* op : view_.AllOps()) {
VLOG(4) << "Handle op " << cnt++ << ": " << op->Name();
if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name())) if (FLAGS_enable_inplace_whitelist && !whitelist_.count(op->Name()))
continue; continue;
TryInplaceOpInputOutput(op, graph.get()); TryInplaceOpInputOutput(op, graph);
} }
graph->ResolveHazard(var_nodes_); // graph->ResolveHazard(var_nodes_);
return graph;
} }
void InplacePass::InplaceModifyDesc(const std::string& var, void InplacePass::InplaceModifyDesc(const std::string& var,
@ -264,13 +265,10 @@ void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
void InplacePass::TryInplaceOpInputOutput(ir::Node* op, void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
ir::Graph* graph) const { ir::Graph* graph) const {
VLOG(4) << "Try to inplace op " << op->Name(); VLOG(4) << "Try to inplace op " << op->Name();
// FIXME(liuwei1031): Graph is not aware of the existence of BlockDescs and // PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
// ProgramDescs. // "op_desc is nullptr");
// The operations related to BlockDesc or ProgramDesc should perform on Graph
// or Node directly!
PADDLE_ENFORCE(op->Op() != nullptr && op->Op()->Block() != nullptr,
"op_desc is nullptr");
// some pre-requirments need to meet if the op want to inplaced. // some pre-requirments need to meet if the op want to inplaced.
PADDLE_ENFORCE(op->Op() != nullptr, "op_desc is nullptr");
auto* op_desc = op->Op(); auto* op_desc = op->Op();
auto& infer_inplace = auto& infer_inplace =
@ -281,21 +279,58 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
PADDLE_ENFORCE(static_cast<bool>(infer_inplace), PADDLE_ENFORCE(static_cast<bool>(infer_inplace),
"%s's infer_inplace has not been registered", op_desc->Type()); "%s's infer_inplace has not been registered", op_desc->Type());
auto* block = op_desc->Block(); auto in_to_outs = infer_inplace(*op_desc);
auto in_to_outs = infer_inplace(*op_desc, block);
auto& all_ops = view_.AllOps(); auto& all_ops = view_.AllOps();
auto cursor = std::find(all_ops.begin(), all_ops.end(), op); auto cursor = std::find(all_ops.begin(), all_ops.end(), op);
size_t idx = std::distance(all_ops.begin(), cursor); size_t idx = std::distance(all_ops.begin(), cursor);
for (auto& pair : in_to_outs) { for (auto& pair : in_to_outs) {
auto& in_var_name = pair.first; auto& in_para_name = pair.first;
auto& out_var_name = pair.second; auto& out_para_name = pair.second;
auto input_vars = op->Op()->Input(in_para_name);
if (!input_vars.size()) {
VLOG(4) << "Parameter " << in_para_name << " is empty skip "
<< in_para_name << " => " << out_para_name << " pair";
continue;
}
auto output_vars = op->Op()->Output(out_para_name);
if (!output_vars.size()) {
VLOG(4) << "Parameter " << out_para_name << " is empty skip "
<< in_para_name << " => " << out_para_name << " pair";
continue;
}
auto in_var_name = input_vars.at(0);
auto out_var_name = output_vars.at(0);
auto* in_node = view_.GetNodeByName(in_var_name, op->inputs); auto* in_node = view_.GetNodeByName(in_var_name, op->inputs);
auto* out_node = view_.GetNodeByName(out_var_name, op->outputs); auto* out_node = view_.GetNodeByName(out_var_name, op->outputs);
VLOG(4) << "Try to inplace " << in_var_name << " with " << out_var_name;
bool can_replace = true;
if (in_var_name == out_var_name) {
can_replace = false;
VLOG(4) << "SKIP: Input variable " << in_var_name << " & Output variable "
<< out_var_name << " are the same";
} else if (!NodeCanReused(in_node)) {
can_replace = false;
VLOG(4) << "SKIP: Input varialbe " << in_var_name << "cannot be reused";
} else if (!NodeCanReused(out_node)) {
can_replace = false;
VLOG(4) << "SKIP: Output variable " << out_var_name
<< " cannot be reused";
} else if (details::NodeSize(*in_node->Var()) !=
details::NodeSize(*out_node->Var())) {
can_replace = false;
VLOG(4) << "SKIP: Input and Output varialbe size not match";
}
if (!can_replace) continue;
// 2. there is no external pending op on the input node // 2. there is no external pending op on the input node
if (view_.PendingOpsOnVar(in_node).size() > 1) { // if (view_.PendingOpsOnVar(in_node).size() > 1) {
if (in_node->outputs.size() > 1 && !view_.CheckDeps(in_node, op)) {
VLOG(4) << string::Sprintf( VLOG(4) << string::Sprintf(
"Skiped pair %s => %s. %s input has external dependency." "Skiped pair %s => %s. %s input has external dependency."
"inplace such pair will overwrite the memory.", "inplace such pair will overwrite the memory.",
@ -342,6 +377,97 @@ void InplacePass::TryInplaceOpInputOutput(ir::Node* op,
} }
} }
void GraphView::TopoSort(ir::Graph* graph) {
//
ops_.clear();
auto deps_num = [](ir::Node* op) {
auto cnt = 0;
for (auto& var : op->inputs)
if (var->inputs.size() > 0) ++cnt;
return cnt;
};
std::queue<std::pair<ir::Node*, uint32_t>> ready_ops;
int level = 0;
auto nodes = graph->Nodes();
std::unordered_map<ir::Node*, uint32_t> deps_map;
for (auto& node : nodes) {
if (node->IsOp() && node->Op() != nullptr) {
deps_map[node] = deps_num(node);
if (0 == deps_map[node]) {
ready_ops.push({node, level});
}
}
}
while (!ready_ops.empty()) {
auto item = ready_ops.front();
ready_ops.pop();
ops_.emplace_back(item.first);
// record level when pop from queue
op_level_[item.first] = item.second;
for (auto node : item.first->outputs) {
for (auto op : node->outputs) {
--deps_map[op];
if (deps_map[op] == 0) ready_ops.push({op, item.second + 1});
}
}
}
bool all_ops_checked = true;
for (auto& node : nodes) {
if (node->IsOp() && node->Op() != nullptr && deps_map[node] > 0) {
all_ops_checked = false;
break;
}
}
PADDLE_ENFORCE(all_ops_checked, "All ops deps should be 0 after analysis");
}
// return true if current op node depeneds on all other op that use the same
// variable node
bool GraphView::CheckDeps(ir::Node* var, ir::Node* current_op) const {
// get op list that rely on the same variable
auto op_list = var->outputs;
for (auto& op : op_list) {
if (op == current_op) continue;
VLOG(4) << " GraphView::CheckDeps : " << op->Name() << " & "
<< current_op->Name();
if (!CheckOpDeps(op, current_op)) return false;
VLOG(4) << "";
}
return true;
}
// check if op2 depends on op1's output
bool GraphView::CheckOpDeps(ir::Node* op1, ir::Node* op2) const {
auto print_op = [&](ir::Node* op, const char* name) {
std::ostringstream os;
os << " " << name << " : " << op->Name() << " ";
os << "Input args : ";
for (auto& arg : op->inputs) os << arg->Name() << " ";
os << "Output args : ";
for (auto& arg : op->outputs) os << arg->Name() << " ";
os << "Level : " << op_level_.at(op);
VLOG(4) << os.str();
};
print_op(op1, "OP1");
print_op(op2, "OP2");
if (op1 == op2) return true;
if (op_level_.at(op1) >= op_level_.at(op2)) return false;
for (auto& var : op2->inputs)
if (var->inputs.size() > 0 && CheckOpDeps(op1, var->inputs[0])) return true;
return false;
}
ir::Node* GraphView::GetNodeByName(const std::string& name, ir::Node* GraphView::GetNodeByName(const std::string& name,
const std::vector<ir::Node*>& nodes) const { const std::vector<ir::Node*>& nodes) const {
// nodes should be op->inputs/outputs // nodes should be op->inputs/outputs
@ -387,22 +513,7 @@ void GraphView::Build(ir::Graph* g) {
// Because we insert some new created node. Which may have data race between // Because we insert some new created node. Which may have data race between
// nodes. // nodes.
// resolve data harzards depends on the var nodes in right order. // resolve data harzards depends on the var nodes in right order.
ops_ = SortOpLikeDescOrder(*g); TopoSort(g);
// 1. track the nodes which reused previous node in Python memory optimize.
// these node can not be inplaced, otherwise may generate a circle in graph.
std::unordered_set<std::string> all_vars;
for (auto& node : g->Nodes()) {
if (node->IsVar()) continue;
for (auto& out : node->outputs) {
if (out->IsCtrlVar() || out->Var() == nullptr) continue;
if (all_vars.count(out->Name())) {
dup_nodes_.emplace(out->Name());
} else {
all_vars.emplace(out->Name());
}
}
}
// 2. track the nodes which used by parameter server. // 2. track the nodes which used by parameter server.
// these node can not be inplaced, otherwise trainer // these node can not be inplaced, otherwise trainer

9
paddle/fluid/framework/details/inplace_op_pass.h
Unescape View File

@ -14,6 +14,7 @@
#pragma once #pragma once
#include <map> #include <map>
#include <memory>
#include <string> #include <string>
#include <unordered_map> #include <unordered_map>
#include <unordered_set> #include <unordered_set>
@ -50,10 +51,15 @@ class GraphView {
// map the parameter and gradient, must be skipped. // map the parameter and gradient, must be skipped.
bool InSkipSet(const std::string& var) const; bool InSkipSet(const std::string& var) const;
bool CheckDeps(ir::Node* var, ir::Node* current_op) const;
bool CheckOpDeps(ir::Node* op1, ir::Node* op2) const;
void TopoSort(ir::Graph* g);
private: private:
std::vector<ir::Node*> ops_; std::vector<ir::Node*> ops_;
std::unordered_set<std::string> dup_nodes_; // mem opt affect nodes std::unordered_set<std::string> dup_nodes_; // mem opt affect nodes
std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_; std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_;
std::unordered_map<ir::Node*, uint32_t> op_level_;
}; };
// swap pairs in sequence // swap pairs in sequence
@ -63,8 +69,7 @@ class InplacePass : public ir::Pass {
InplacePass(); InplacePass();
protected: protected:
std::unique_ptr<ir::Graph> ApplyImpl( void ApplyImpl(ir::Graph* graph) const override;
std::unique_ptr<ir::Graph> graph) const override;
void InitSSAGraphNodes() const; void InitSSAGraphNodes() const;

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save