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

revert-16555-model_data_cryption_link_all_lib
dengkaipeng 6 years ago
commit 193185b840

@ -24,6 +24,8 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
"${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION}")
message(STATUS "C compiler: ${CMAKE_C_COMPILER}, version: "
"${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
message(STATUS "AR tools: ${CMAKE_AR}")
if(WIN32)
set(CMAKE_SUPPRESS_REGENERATION ON)
set(CMAKE_STATIC_LIBRARY_PREFIX lib)
@ -62,6 +64,7 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(WITH_PSLIB "Compile with pslib support" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" 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(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)
@ -188,7 +191,14 @@ include(configure) # add paddle env configuration
if(WITH_GPU)
include(cuda)
include(tensorrt)
include(anakin_subgraph)
endif()
if(WITH_GPU AND NOT WIN32)
message(STATUS "add dgc lib.")
include(external/dgc)
endif()
if(WITH_MKL OR WITH_MKLML)
include(external/anakin)
elseif()

@ -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:
- verbose level 1: [framework](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework)
- verbose level 3: [operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/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 7: [math](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/legacy/math)
- 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/fluid/operators)
- 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/fluid/operators/math/)

@ -75,8 +75,9 @@ RUN curl -s -q https://glide.sh/get | sh
# and its size is only one-third of the official one.
# 2. Manually add ~IPluginFactory() in IPluginFactory class of NvInfer.h, otherwise, it couldn't work in paddle.
# See https://github.com/PaddlePaddle/Paddle/issues/10129 for details.
RUN wget -qO- http://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.0.3.Ubuntu-16.04.4.x86_64-gnu.cuda-8.0.cudnn7.0.tar.gz | \
tar -xz -C /usr/local && \
RUN wget -q https://paddlepaddledeps.cdn.bcebos.com/TensorRT-4.0.1.6-ubuntu14.04.x86_64-gnu.cuda.8.0.cudnn7.0.tar.gz --no-check-certificate && \
tar -zxf TensorRT-4.0.1.6-ubuntu14.04.x86_64-gnu.cuda.8.0.cudnn7.0.tar.gz -C /usr/local && \
cp -rf /usr/local/TensorRT/include /usr && \
cp -rf /usr/local/TensorRT/lib /usr

@ -179,7 +179,6 @@ def train_parallel(train_args, test_args, args, train_prog, test_prog,
else:
build_strategy.reduce_strategy = fluid.BuildStrategy(
).ReduceStrategy.AllReduce
build_strategy.fuse_broadcast_op = args.fuse_broadcast_op
avg_loss = train_args[0]

@ -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()

@ -24,7 +24,7 @@ set(BOOST_PROJECT "extern_boost")
# So we use 1.41.0 here.
set(BOOST_VER "1.41.0")
set(BOOST_TAR "boost_1_41_0" CACHE STRING "" FORCE)
set(BOOST_URL "http://paddlepaddledeps.cdn.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
set(BOOST_URL "http://paddlepaddledeps.bj.bcebos.com/${BOOST_TAR}.tar.gz" CACHE STRING "" FORCE)
MESSAGE(STATUS "BOOST_TAR: ${BOOST_TAR}, BOOST_URL: ${BOOST_URL}")

@ -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)

@ -44,7 +44,7 @@ ExternalProject_Add(
# 3. keep only zlib, cares, protobuf, boringssl under "third_party",
# checkout and clean other dirs under third_party
# 4. remove .git, and package the directory.
URL "http://paddlepaddledeps.cdn.bcebos.com/grpc-v1.10.x.tar.gz"
URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.10.x.tar.gz"
URL_MD5 "1f268a2aff6759839dccd256adcc91cf"
PREFIX ${GRPC_SOURCES_DIR}
UPDATE_COMMAND ""

@ -31,9 +31,17 @@ IF(APPLE)
return()
ENDIF()
MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
# Introduce variables:
# * CMAKE_INSTALL_LIBDIR
INCLUDE(GNUInstallDirs)
SET(LIBDIR "lib")
if(CMAKE_INSTALL_LIBDIR MATCHES ".*lib64$")
SET(LIBDIR "lib64")
endif()
MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/l${LIBDIR} to runtime path")
SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/${LIBDIR}")
INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
@ -58,7 +66,7 @@ ExternalProject_Add(
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_DEPENDS}
GIT_REPOSITORY "https://github.com/intel/mkl-dnn.git"
GIT_TAG "830a10059a018cd2634d94195140cf2d8790a75a"
GIT_TAG "863ff6e7042cec7d2e29897fe9f0872e0888b0fc"
PREFIX ${MKLDNN_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
@ -79,9 +87,9 @@ ExternalProject_Add(
-DMKLROOT:PATH=${MKLML_ROOT}
)
if(WIN32)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/${LIBDIR}/mkldnn.lib" CACHE FILEPATH "mkldnn library." FORCE)
else(WIN32)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/${LIBDIR}/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
endif(WIN32)
ADD_LIBRARY(shared_mkldnn SHARED IMPORTED GLOBAL)
@ -101,7 +109,7 @@ ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
# copy the real so.0 lib to install dir
# it can be directly contained in wheel or capi
if(WIN32)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/lib/mkldnn.dll)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/bin/mkldnn.dll)
else(WIN32)
SET(MKLDNN_SHARED_LIB ${MKLDNN_INSTALL_DIR}/libmkldnn.so.0)
ADD_CUSTOM_COMMAND(OUTPUT ${MKLDNN_SHARED_LIB}

@ -34,7 +34,7 @@ SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLML_ROOT}/lib")
SET(TIME_VERSION "2019.0.1.20181227")
IF(WIN32)
SET(MKLML_VER "mklml_win_${TIME_VERSION}" CACHE STRING "" FORCE)
SET(MKLML_URL "https://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
SET(MKLML_URL "https://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.zip" CACHE STRING "" FORCE)
SET(MKLML_LIB ${MKLML_LIB_DIR}/mklml.lib)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5md.lib)
SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/mklml.dll)
@ -43,7 +43,7 @@ ELSE()
#TODO(intel-huying):
# Now enable Erf function in mklml library temporarily, it will be updated as offical version later.
SET(MKLML_VER "Glibc225_vsErf_mklml_lnx_${TIME_VERSION}" CACHE STRING "" FORCE)
SET(MKLML_URL "http://paddlepaddledeps.cdn.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
SET(MKLML_URL "http://paddlepaddledeps.bj.bcebos.com/${MKLML_VER}.tgz" CACHE STRING "" FORCE)
SET(MKLML_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)
SET(MKLML_IOMP_LIB ${MKLML_LIB_DIR}/libiomp5.so)
SET(MKLML_SHARED_LIB ${MKLML_LIB_DIR}/libmklml_intel.so)

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

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

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

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

@ -5,13 +5,13 @@ Kexin Zhao <zhaokexin01@baidu.com>
## Introduction
Deep learning is usually a two-stage work: training and inference. The training stage estimates model parameters (weights) from data. The inference stage loads the weights and uses them to interpret inputs. Typically, weights are 32-bit float values (float32). Some new devices, including NVIDIA Volta GPUs, support higher speed computation using 16-bit float values (float16).
This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16.
This article explains our efforts with PaddlePaddle to train using float32 and to inference using float16. We describe a [*transpiler*](https://github.com/PaddlePaddle/Paddle/blob/a4d3de0071e1f3912230c3ab3f9ac74cf06b093a/doc/fluid/design/motivation/fluid_compiler.md), which converts a PaddlePaddle Fluid model, which, to be precise, should be called a [Fluid *program*](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/program.md), into the inference program, and converts the weights from float32 into float16.
## What is float16?
float16 (or FP16) is a half-precision floating-point format that uses 16 bits in memory to represent a value. The advantage over 32-bit single-precision floating-point format (commonly known as float or float32 data type) is that it requires half the storage and bandwidth at the expense of precision and range. Fortunately, DNN inference has a high tolerance for the loss of precision and range when using float16 to represent the weights, and the inference accuracy will only be minimally affected in most cases, which gives us the opportunity to use float16 data type to speed up the inference.
Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type.
Interested readers can refer to our [design doc](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/data_type/float16.md) and [code](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/platform/float16.h) for more details on how we implement the float16 data type.
## Why float16?
The trend in today's deep learning community is to use bigger and deeper model, which translates to larger memory footprint, higher computation demands, and as a result higher energy consumption on computing devices. The advantages of float16 over float32 are correspondingly three-fold:
@ -24,12 +24,12 @@ The trend in today's deep learning community is to use bigger and deeper model,
## Fluid implementation of float16 inference
### Overview
Fluid use [Program](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block.
Fluid use [Program](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#program) instead of computation graph to describe a neural network model and the optimization procedure. Fluid program is a python wrapper around a protobuf message called [ProgramDesc](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/program.md). Similar to programming languages, the basic structure of a Fluid program is some nested [blocks](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#block), where each block consists of some [variable](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#variable) definitions and a sequence of [operators](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/modules/python_api.md#operator). An [executor](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/concepts/executor.md) will run a given program by sequentially executing the operators in the entrance block.
### Basic requirement
When an executor runs an operator, it uses a kernel to perform computations on tensors contained in the input variables, and then writes the results to the tensors in the output variables. Each operator has multiple kernels for different combinations of data types, devices, and library types, respectively. The operator will select the appropriate kernel to run based on, among other things, the data type of the input tensors. By default, every Fluid operator has a kernel for float data type that takes float inputs and generates float outputs.
If we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will make the program run in float mode and gives us a final output of float data type.
If we provide float input to the first operator in a program, then each operator will use float kernel to compute float output and send it as input to the next operator to trigger its float kernel. This chain effect will make the program run in float mode and gives us a final output of float data type.
The same principle applies if we want a program to run in float16 mode. We provide input variable of the float16 data type to the first operator, and every subsequent operator will invoke the float16 kernel until we get the final output in float16. So the preliminary requirements for float16 inference are to add float16 kernels to operators that are needed in a specific kind of neural networks. Our current focus is on Convolutional Neural Networks (CNN) and hence we have added float16 kernels to the following operators: convolution, pooling, GEMM, elementwise addition, batch norm, dropout, various activations including relu and tanh, and softmax.
@ -75,7 +75,7 @@ In this scenario, we already have a float32 inference program and some associate
We can then run various inference experiments in float16 mode and save the float16 program and weights on disk for future deployment. To enhance the code usability, we maintain a consistent API so that user can use the same float32 input data to run inference program in either float32 and float16 mode and obtain output data both of float32 data type. Consequently, we need to add cast operators in the float16 inference program for conversions between the float16 tensor and float32 tensor.
The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference).
The float16 transpiler is implemented to fulfill the requirements mentioned above. The details of the float16 transpiler can be found [here](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/design/data_type/float16.md#float16-inference).
### Experiment results
Simply running the following commands to reproduce the experiment results presented in this section:
@ -113,7 +113,7 @@ We repeat the test ten times and get the following results:
| #10 | 62.53% | 62.48% |
| average| 62.63% | 62.62% |
We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over ten tests.
We can see that the accuracy of float16 inference is very close to that of float32 inference in every experiment (within 0.05% difference) and is overall 0.01% better than its float32 counterpart averaged over ten tests.
#### Performance benchmark
Currently, Fluid only supports float16 inference on NVIDIA GPUs. There is no motivation to support float16 inference on non-ARM CPUs where float16 is not natively supported, and float16 calculation will only be slower than its float32 counterpart.
@ -132,7 +132,7 @@ Average inference time for one mini-batch on Vgg16 model tested on ImageNet data
|float16| 3.32 | 4.11 | 5.88 | 9.41 | 16.54 | 30.47 | 60.23 |
|Speedup| 4.22 | 2.36  | 3.91 | 3.00 | 3.26  | 2.77 | 2.97 |
We can see that float16 inference provides **2x ~ 4x** speedup on different batch sizes.
We can see that float16 inference provides **2x ~ 4x** speedup on different batch sizes.
Convolution operation is ususally the computational bottleneck of CNN, so we also check the average time spent on the Fluid convolution operators for one mini-batch as follows:
@ -162,7 +162,7 @@ We find that the speedup provided by float16 inference starts relatively small a
We also did the same benchmark on a single NVIDIA GeForce GTX 1080 Ti GPU that does not support Tensor Core. The results show that for Vgg16, float16 inference provides consistent small speedup (around 1.15x) for all mini-batch sizes, while for Resnet50, float16 inference is slower than its float32 counterpart in small batch sizes (mb = 1 and 2) and then delivers around 1.15x speedup for all larger batch sizes. By comparing the benchmarks on 1080 Ti and V100, we find that Tensor Core, which is specialized for float16 computations, is a critical component of high performance float16 inference.
Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/contrib/float16/float16_benchmark.md) for complete benchmark results.
Please refer to [here](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/contrib/float16/float16_benchmark.md) for complete benchmark results.
### Summary
1. Fluid is now able to run inference in float16 mode via a float16 transpiler. We currently support CNN programs, including Vgg and Resnet, to run in float16 inference mode.

File diff suppressed because it is too large Load Diff

@ -38,10 +38,10 @@ if(WITH_GPU)
nv_library(tensor SRCS tensor.cc .tensor_util.cu DEPS place memory data_type device_context)
add_dependencies(tensor tensor_util)
else()
nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context )
nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context profiler)
endif(WIN32)
else()
cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context )
cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context profiler)
endif()
cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@ -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)
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_test(reader_test SRCS reader_test.cc DEPS reader)
@ -164,6 +164,8 @@ else()
set(NGRAPH_EXE_DEPS)
endif()
cc_library(executor_gc_helper SRCS executor_gc_helper.cc DEPS scope proto_desc operator garbage_collector)
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_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)
endif()
target_link_libraries(executor garbage_collector)
target_link_libraries(executor while_op_helper executor_gc_helper)
cc_library(parallel_executor SRCS parallel_executor.cc DEPS
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
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_library(selected_rows SRCS selected_rows.cc DEPS tensor)
cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)

@ -5,10 +5,12 @@ 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(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(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_graph_print_pass SRCS multi_devices_graph_print_pass.cc DEPS multi_devices_helper)
cc_library(multi_devices_graph_check_pass SRCS multi_devices_graph_check_pass.cc DEPS multi_devices_helper)
cc_library(alloc_continuous_space_for_grad_pass SRCS alloc_continuous_space_for_grad_pass.cc DEPS graph graph_helper)
cc_library(variable_visitor SRCS variable_visitor.cc DEPS lod_tensor selected_rows)
@ -21,6 +23,8 @@ endif()
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 dgc)
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)
if(WITH_DISTRIBUTE)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
@ -35,6 +39,8 @@ if(WITH_GPU)
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(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
variable_visitor)
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_rpc)
@ -46,9 +52,7 @@ else()
cc_library(fused_broadcast_op_handle SRCS fused_broadcast_op_handle.cc DEPS broadcast_op_handle)
endif()
cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_base scope lod_tensor)
cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
if(WITH_GPU)
cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
@ -61,14 +65,17 @@ cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_
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)
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(while_op_eager_deletion_pass SRCS while_op_eager_deletion_pass.cc DEPS while_op_helper graph_helper pass computation_op_handle)
cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass while_op_eager_deletion_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)
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)
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)
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 memory_optimize_pass inplace_op_pass)
if (WITH_GPU)
@ -97,5 +104,5 @@ cc_library(build_strategy SRCS build_strategy.cc DEPS
graph_viz_pass multi_devices_graph_pass
multi_devices_graph_print_pass multi_devices_graph_check_pass
fuse_elewise_add_act_pass multi_batch_merge_pass
fuse_relu_depthwise_conv_pass
memory_optimize_pass lock_free_optimize_pass)
fuse_relu_depthwise_conv_pass
memory_optimize_pass lock_free_optimize_pass alloc_continuous_space_for_grad_pass fuse_all_reduce_op_pass)

@ -13,6 +13,7 @@
// limitations under the License.
#include <algorithm>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
@ -52,13 +53,28 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
// Note that must assert topology sort is stable
auto& ops = graph->Get<const std::vector<OpDesc*>>(kStaleProgramOpDescs);
for (auto* op_desc : ops) {
auto outputs = op_desc->Outputs();
for (auto& o_it : outputs) {
for (auto& v : o_it.second) { // values
vars[v] = order;
try {
bool is_bk_op =
static_cast<bool>(boost::get<int>(op_desc->GetAttr(
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;
@ -70,7 +86,8 @@ 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,
OpHandleBase* op2) {
@ -83,6 +100,10 @@ std::unique_ptr<ir::Graph> AllReduceDepsPass::ApplyImpl(
auto l_it = vars.find(i0->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) {

File diff suppressed because it is too large Load Diff

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

@ -57,7 +57,7 @@ struct BroadcastOpHandle : public OpHandleBase {
std::string Name() const override;
bool IsMultiDeviceTransfer() override { return false; };
bool IsMultiDeviceTransfer() override { return true; };
protected:
void RunImpl() override;

@ -16,6 +16,7 @@ limitations under the License. */
#include <glog/logging.h>
#include <memory>
#include <utility>
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/details/multi_devices_graph_pass.h"
@ -45,12 +46,27 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
public:
explicit ParallelExecutorPassBuilder(const BuildStrategy &strategy)
: ir::PassBuilder(), strategy_(strategy) {
// Add a graph viz pass to record a graph.
if (!strategy_.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass");
const std::string graph_path = string::Sprintf(
"%s%s", strategy_.debug_graphviz_path_.c_str(), "_original_graph");
viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
}
if (strategy_.enable_sequential_execution_) {
VLOG(10) << "Add sequential_execution_pass";
AppendPass("sequential_execution_pass");
}
// Add op fusion.
if (strategy.sync_batch_norm_) {
AppendPass("sync_batch_norm_pass");
}
// Add op fusion.
if (strategy.fuse_relu_depthwise_conv_) {
VLOG(10) << "Add fuse_relu_depthwise_conv_pass";
AppendPass("fuse_relu_depthwise_conv_pass");
}
@ -62,29 +78,30 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// Add automatically inplace.
if (strategy_.enable_inplace_) {
VLOG(10) << "Add inplace_pass";
AppendPass("inplace_pass");
}
if (strategy.fuse_elewise_add_act_ops_) {
VLOG(10) << "Add fuse_elewise_add_act_pass";
AppendPass("fuse_elewise_add_act_pass");
}
// for single card training, fuse_all_reduce_ops is unnecessary.
// alloc_continuous_space_for_grad_pass should be before of MultiDevPass.
if (strategy.fuse_all_reduce_ops_) {
VLOG(10) << "Add alloc_continuous_space_for_grad_pass";
AppendPass("alloc_continuous_space_for_grad_pass");
}
// Add a graph viz pass to record a graph.
if (!strategy_.debug_graphviz_path_.empty()) {
if (!strategy.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass");
const std::string graph_path = string::Sprintf(
"%s%s", strategy_.debug_graphviz_path_.c_str(), "_original_graph");
"%s%s", strategy.debug_graphviz_path_.c_str(), "_fused_graph");
viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
}
if (strategy.fuse_elewise_add_act_ops_) {
auto fuse_elewise_add_act_pass = AppendPass("fuse_elewise_add_act_pass");
// Add a graph viz pass to record a graph.
if (!strategy.debug_graphviz_path_.empty()) {
auto viz_pass = AppendPass("graph_viz_pass");
const std::string graph_path = string::Sprintf(
"%s%s", strategy.debug_graphviz_path_.c_str(), "_fused_graph");
viz_pass->Set<std::string>("graph_viz_path",
new std::string(graph_path));
}
}
CollectiveContext *context = CollectiveContext::GetInstance();
context->endpoints_ = strategy_.trainers_endpoints_;
context->trainer_id_ = strategy_.trainer_id_;
@ -102,11 +119,19 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// A side-effect of that, memory optimize cannot forsee the fetched vars
// , so fetchlist should be set persistable before call the Run interface.
if (strategy.memory_optimize_) {
auto memory_optimize_pass = AppendPass("memory_optimize_pass");
VLOG(10) << "Add memory_optimize_pass";
AppendPass("memory_optimize_pass");
}
AppendMultiDevPass(strategy);
if (strategy.fuse_all_reduce_ops_) {
// NOTE: fuse_all_reduce_ops will count the number of all_reduce operator
// first, if the number is zero, fuse_all_reduce_ops will do nothing.
VLOG(10) << "Add fuse_all_reduce_op_pass";
AppendPass("fuse_all_reduce_op_pass");
}
// Add a graph print pass to record a graph with device info.
if (!strategy_.debug_graphviz_path_.empty()) {
auto multi_devices_print_pass = AppendPass("multi_devices_print_pass");
@ -122,28 +147,34 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
// Verify that the graph is correct for multi-device executor.
AppendPass("multi_devices_check_pass");
if (VLOG_IS_ON(2)) {
AppendPass("all_reduce_deps_pass");
}
if (SeqOnlyAllReduceOps(strategy)) {
VLOG(10) << "Add all_reduce_deps_pass";
AppendPass("all_reduce_deps_pass");
}
if (strategy_.remove_unnecessary_lock_) {
VLOG(10) << "Add modify_op_lock_and_record_event_pass";
AppendPass("modify_op_lock_and_record_event_pass");
}
}
// Convert graph to run on multi-devices.
void AppendMultiDevPass(const BuildStrategy &strategy) {
ir::Pass *multi_devices_pass;
ir::Pass *multi_devices_pass = nullptr;
if (strategy_.is_distribution_) {
VLOG(3) << "multi device parameter server mode";
VLOG(10) << "Add dist_multi_devices_pass";
multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
} else {
if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
VLOG(3) << "multi devices collective mode with allreduce";
VLOG(10) << "Add all_reduce_mode_multi_devices_pass";
multi_devices_pass =
AppendPass("allreduce_mode_multi_devices_pass").get();
AppendPass("all_reduce_mode_multi_devices_pass").get();
} else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
VLOG(3) << "multi deivces collective mode with reduce";
VLOG(10) << "Add reduce_mode_multi_devices_pass";
multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
} else {
PADDLE_THROW("Unknown reduce strategy.");
@ -200,9 +231,26 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
pass->Erase("nccl_ctxs");
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
pass->Erase(kNCCLCtxs);
pass->SetNotOwned<platform::NCCLContextMap>(kNCCLCtxs, nctx);
#endif
} else if (pass->Type() == "fuse_all_reduce_op_pass") {
pass->Erase(kPlaces);
pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
pass->Erase(kLocalScopes);
pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
&local_scopes);
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
pass->Erase(kNCCLCtxs);
pass->SetNotOwned<platform::NCCLContextMap>(kNCCLCtxs, nctx);
#endif
} else if (pass->Type() == "alloc_continuous_space_for_grad_pass") {
pass->Erase(kPlaces);
pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
pass->Erase(kLocalScopes);
pass->SetNotOwned<const std::vector<Scope *>>(kLocalScopes,
&local_scopes);
} else if (pass->Type() == "sequential_execution_pass") {
LOG(INFO) << "set enable_sequential_execution:"
<< enable_sequential_execution_;
@ -227,12 +275,13 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
} // namespace framework
} // namespace paddle
USE_PASS(sync_batch_norm_pass);
USE_PASS(fuse_relu_depthwise_conv_pass);
USE_PASS(fuse_elewise_add_act_pass);
USE_PASS(graph_viz_pass);
USE_PASS(multi_batch_merge_pass);
USE_PASS(reduce_mode_multi_devices_pass);
USE_PASS(allreduce_mode_multi_devices_pass);
USE_PASS(all_reduce_mode_multi_devices_pass);
USE_PASS(dist_multi_devices_pass);
USE_PASS(multi_devices_check_pass);
USE_PASS(multi_devices_print_pass);
@ -242,4 +291,6 @@ USE_PASS(all_reduce_deps_pass);
USE_PASS(modify_op_lock_and_record_event_pass);
USE_PASS(inplace_pass);
USE_PASS(lock_free_optimize_pass);
USE_PASS(alloc_continuous_space_for_grad_pass);
USE_PASS(graph_to_program_pass);
USE_PASS(fuse_all_reduce_op_pass);

@ -16,6 +16,7 @@
#include <memory>
#include <string>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/ir/pass_builder.h"
@ -75,8 +76,12 @@ struct BuildStrategy {
bool fuse_elewise_add_act_ops_{false};
bool fuse_all_reduce_ops_{false};
bool fuse_relu_depthwise_conv_{false};
bool sync_batch_norm_{false};
bool memory_optimize_{true};
// TODO(dzhwinter):
// make enable_inplace, memory_optimize_

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

Loading…
Cancel
Save