Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2299
Star 20931
Fork 5422
P
Paddle
提交 b8491bfd 编写于 作者: Q Qiao Longfei
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add-communicator

上级 8bda4ab2 3ccd8964
展开全部 隐藏空白更改
内联 并排
Showing with 9573 addition and 6370 deletion
AUTHORS.md
浏览文件 @ b8491bfd
......@@ -44,6 +44,7 @@
| qingqing01 | Qing-Qing Dang |
| reyoung | Yang Yu |
| Sand3r- | Michal Gallus |
| sfraczek | Sylwester Fraczek |
| Superjom | Chun-Wei Yan |
| tensor-tang | Jian Tang |
| tianbingsz | Tian-Bing Xu |
......@@ -54,6 +55,7 @@
| wangyang59 | Yang Wang |
| wangzhen-nlp | Zhen Wang |
| wen-bo-yang | Wen-Bo Yang |
| wojtuss | Wojciech Uss |
| wwhu | Wei-Wei Hu |
| xinghai-sun | Xing-Hai Sun |
| Xreki | Yi-Qun Liu |
......
CMakeLists.txt
浏览文件 @ b8491bfd
......@@ -54,23 +54,12 @@ option(WITH_NGRAPH "Compile PaddlePaddle with nGraph support." OFF)
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" OFF)
option(WITH_PYTHON "Compile PaddlePaddle with python interpreter" ON)
option(WITH_DOUBLE "Compile PaddlePaddle with double precision" OFF)
option(WITH_RDMA "Compile PaddlePaddle with RDMA support" OFF)
option(WITH_TIMER "Compile PaddlePaddle with stats timer" OFF)
option(WITH_PROFILER "Compile PaddlePaddle with GPU profiler and gperftools" OFF)
option(WITH_JEMALLOC "Compile PaddlePaddle with jemalloc" OFF)
option(WITH_DOC "Compile PaddlePaddle with documentation" OFF)
option(WITH_COVERAGE "Compile PaddlePaddle with code coverage" OFF)
option(COVERALLS_UPLOAD "Package code coverage data to coveralls" OFF)
option(ON_TRAVIS "Exclude special unit test on Travis CI" OFF)
option(WITH_FLUID_ONLY "Compile PaddlePaddle fluid only" OFF)
option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF)
option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(WITH_PSLIB "Compile with pslib support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
option(WITH_ANAKIN "Compile with Anakin library" OFF)
......@@ -105,8 +94,6 @@ endif()
if (WIN32)
set(WITH_DISTRIBUTE OFF CACHE STRING
"Disable DISTRIBUTE when compiling for Windows" FORCE)
set(WITH_FLUID_ONLY ON CACHE STRING
"Enable FLUID_ONLY when compiling for Windows" FORCE)
endif()
set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
......@@ -148,7 +135,6 @@ include(external/openblas) # download, build, install openblas
include(external/mkldnn) # download, build, install mkldnn
include(external/ngraph) # download, build, install nGraph
include(external/boost) # download boost
include(external/any) # download libn::any
include(external/eigen) # download eigen3
include(external/pybind11) # download pybind11
include(external/cares)
......@@ -225,7 +211,6 @@ include(generic) # simplify cmake module
include(package) # set paddle packages
include(ccache) # set ccache for compilation
include(util) # set unittest and link libs
include(rdma) # set rdma libraries
include(version) # set PADDLE_VERSION
include(coveralls) # set code coverage
include(inference_lib) # add paddle fluid inference libraries
......@@ -233,38 +218,11 @@ include(inference_lib) # add paddle fluid inference libraries
include_directories("${PADDLE_SOURCE_DIR}")
set(EXTERNAL_LIBS
gflags
glog
${CBLAS_LIBRARIES}
protobuf
zlib
${PYTHON_LIBRARIES}
)
if(WITH_PSLIB)
list(APPEND EXTERNAL_LIBS pslib)
list(APPEND EXTERNAL_LIBS pslib_brpc)
list(APPEND EXTERNAL_LIBS libmct)
endif(WITH_PSLIB)
if(WITH_AMD_GPU)
find_package(HIP)
include(hip)
endif(WITH_AMD_GPU)
if(WITH_MKLML)
list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
endif()
if(WITH_LIBXSMM)
list(APPEND EXTERNAL_LIBS ${LIBXSMM_LIBS})
endif()
if(WITH_MKLDNN)
list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
endif()
set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
......
README.md
浏览文件 @ b8491bfd
......@@ -3,8 +3,8 @@
English | [简体中文](./README_cn.md)
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......@@ -18,7 +18,7 @@ learning to many products at Baidu.
Our vision is to enable deep learning for everyone via PaddlePaddle.
Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
### Latest PaddlePaddle Release: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### Latest PaddlePaddle Release: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3)
### Install Latest Stable Release:
```
# Linux CPU
......@@ -26,9 +26,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.2.0.post87
pip install paddlepaddle-gpu==1.3.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.2.0.post85
pip install paddlepaddle-gpu==1.3.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
......@@ -75,26 +75,26 @@ pip install paddlepaddle-gpu==1.2.0.post85
## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html) on our website.
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) on our website.
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) documentation.
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) documentation.
- [Deep Learning 101](https://github.com/PaddlePaddle/book)
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
- [Distributed Training](http://paddlepaddle.org/documentation/docs/en/1.3/user_guides/howto/training/multi_node_en.html)
You can run distributed training jobs on MPI clusters.
- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
- [Python API](http://paddlepaddle.org/documentation/docs/en/1.3/api/index_en.html)
Our new API enables much shorter programs.
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
- [How to Contribute](http://paddlepaddle.org/documentation/docs/en/1.3/advanced_usage/development/contribute_to_paddle/index_en.html)
We appreciate your contributions!
......
README_cn.md
浏览文件 @ b8491bfd
......@@ -3,8 +3,8 @@
[English](./README.md) | 简体中文
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......@@ -16,7 +16,7 @@ PaddlePaddle (PArallel Distributed Deep LEarning) 是一个简单易用、高效
跟进PaddlePaddle最新特性请参考我们的[版本说明](https://github.com/PaddlePaddle/Paddle/releases)
### PaddlePaddle最新版本: [Fluid 1.2.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.2)
### PaddlePaddle最新版本: [Fluid 1.3.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.3)
### 安装最新稳定版本:
```
# Linux CPU
......@@ -24,9 +24,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.2.0.post87
pip install paddlepaddle-gpu==1.3.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.2.0.post85
pip install paddlepaddle-gpu==1.3.0.post85
# 其他平台上的安装指引请参考 http://paddlepaddle.org/
```
......@@ -57,26 +57,26 @@ pip install paddlepaddle-gpu==1.2.0.post85
## 安装
推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html)
推荐阅读官网上的[安装说明](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/install/index_cn.html)
## 文档
我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.2/getstarted/index_en.html)和
[中文](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/index.html) 文档
我们提供[英文](http://paddlepaddle.org/documentation/docs/en/1.3/beginners_guide/index_en.html)和
[中文](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/index.html) 文档
- [深度学习101](https://github.com/PaddlePaddle/book)
或许您想从这个在线交互式书籍开始，可以在Jupyter Notebook中运行
- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.2/user_guides/howto/training/cluster_howto.html)
- [分布式训练](http://paddlepaddle.org/documentation/docs/zh/1.3/user_guides/howto/training/multi_node.html)
可以在MPI集群上运行分布式训练任务
- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.2/api_cn/index_cn.html)
- [Python API](http://paddlepaddle.org/documentation/docs/zh/1.3/api_cn/index_cn.html)
新的API支持代码更少更简洁的程序
- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.2/advanced_usage/development/contribute_to_paddle/index_cn.html)
- [贡献方式](http://paddlepaddle.org/documentation/docs/zh/1.3/advanced_usage/development/contribute_to_paddle/index_cn.html)
欢迎您的贡献!
......
benchmark/IntelOptimizedPaddle.md 已删除 100644 → 0
浏览文件 @ 8bda4ab2
# Benchmark
Machine:
- Server: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 2 Sockets, 20 Cores per socket
- Laptop: TBD
System: CentOS release 6.3 (Final), Docker 1.12.1.
PaddlePaddle:
- paddlepaddle/paddle:0.11.0 (for MKLML and MKL-DNN)
- MKL-DNN tag v0.11
- MKLML 2018.0.1.20171007
- paddlepaddle/paddle:0.11.0-openblas (for OpenBLAS)
- OpenBLAS v0.2.20
On each machine, we will test and compare the performance of training on single node using MKL-DNN / MKLML / OpenBLAS respectively.
## Benchmark Model
### Server
#### Training
Test on batch size 64, 128, 256 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
Pay attetion that the speed below includes forward, backward and parameter update time. So we can not directly compare the data with the benchmark of caffe `time` [command](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/caffe/image/run.sh#L9), which only contain forward and backward. The updating time of parameter would become very heavy when the weight size are large, especially on alexnet.
Input image size - 3 * 224 * 224, Time: images/second
- VGG-19
| BatchSize | 64 | 128 | 256 |
|--------------|-------| -----| --------|
| OpenBLAS | 7.80 | 9.00 | 10.80 |
| MKLML | 12.12 | 13.70 | 16.18 |
| MKL-DNN | 28.46 | 29.83 | 30.44 |
<img src="figs/vgg-cpu-train.png" width="500">
- ResNet-50
| BatchSize | 64 | 128 | 256 |
|--------------|-------| ------| -------|
| OpenBLAS | 25.22 | 25.68 | 27.12 |
| MKLML | 32.52 | 31.89 | 33.12 |
| MKL-DNN | 81.69 | 82.35 | 84.08 |
<img src="figs/resnet-cpu-train.png" width="500">
- GoogLeNet
| BatchSize | 64 | 128 | 256 |
|--------------|-------| ------| -------|
| OpenBLAS | 89.52 | 96.97 | 108.25 |
| MKLML | 128.46| 137.89| 158.63 |
| MKL-DNN     | 250.46| 264.83| 269.50 |
<img src="figs/googlenet-cpu-train.png" width="500">
- AlexNet
| BatchSize | 64 | 128 | 256 |
|--------------|--------| ------ | -------|
| OpenBLAS | 45.62 | 72.79 | 107.22 |
| MKLML | 66.37 | 105.60 | 144.04 |
| MKL-DNN | 399.00 | 498.94 | 626.53 |
<img src="figs/alexnet-cpu-train.png" width="500">
#### Inference
Test on batch size 1, 2, 4, 8, 16 on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
- VGG-19
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|-------|-------|-------|-------|-------|
| OpenBLAS | 1.10 | 1.96 | 3.62 | 3.63 | 2.25 |
| MKLML | 5.58 | 9.80 | 15.15 | 21.21 | 28.67 |
| MKL-DNN | 75.07 | 88.64 | 82.58 | 92.29 | 96.75 |
<img src="figs/vgg-cpu-infer.png" width="500">
- ResNet-50
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|-------|--------|--------|--------|--------|
| OpenBLAS | 3.31 | 6.72 | 11.59 | 13.17 | 9.27 |
| MKLML | 6.33 | 12.02 | 22.88 | 40.53 | 63.09 |
| MKL-DNN | 107.83| 148.84 | 177.78 | 189.35 | 217.69 |
<img src="figs/resnet-cpu-infer.png" width="500">
- GoogLeNet
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|--------|--------|--------|--------|--------|
| OpenBLAS | 12.06 | 23.56 | 34.48 | 36.45 | 23.12 |
| MKLML | 22.74 | 41.56 | 81.22 | 133.47 | 210.53 |
| MKL-DNN | 175.10 | 272.92 | 450.70 | 512.00 | 600.94 |
<img src="figs/googlenet-cpu-infer.png" width="500">
- AlexNet
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|--------|--------|--------|--------|--------|
| OpenBLAS | 3.53 | 6.23 | 15.04 | 26.06 | 31.62 |
| MKLML | 21.32 | 36.55 | 73.06 | 131.15 | 192.77 |
| MKL-DNN | 442.91 | 656.41 | 719.10 | 847.68 | 850.51 |
<img src="figs/alexnet-cpu-infer.png" width="500">
### Laptop
TBD
benchmark/README.md 已删除 100644 → 0
浏览文件 @ 8bda4ab2
# Benchmark
Machine:
- CPU: 12-core Intel(R) Xeon(R) CPU E5-2620 v2 @2.10GHz
- GPU: Tesla K40m
- cuDNN: v5.1
- system: Docker 1.12.1, all platforms are tested in docker environment.
Platforms:
- PaddlePaddle: paddledev/paddle:gpu-devel-v0.9.0a0
- Tensorflow: gcr.io/tensorflow/tensorflow:0.11.0rc0-gpu
- Caffe: kaixhin/cuda-caffe
Several convolutional neural networks and recurrent neural networks are used to test.
## Image
### Benchmark Model
AlexNet, GoogleNet and a small network used in Caffe.
- [AlexNet](https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet): but the group size is one.
- [GoogleNet](https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet): but remove loss1 and loss2 when testing benchmark.
- [SmallNet](https://github.com/BVLC/caffe/blob/master/examples/cifar10/cifar10\_quick\_train\_test.prototxt)
### Single-GPU
- AlexNet: input - 3 * 227 * 227, Time: ms/batch
| BatchSize | 64 | 128 | 256 | 512 |
|--------------|-----| -----| ------| -----|
| PaddlePaddle | 195 | 334 | 602 | 1629 |
| TensorFlow | 223 | 364 | 645 | 1235 |
| Caffe | 324 | 627 | 1232 | 2513 |
**Notation**
All platforms use cuDNN-v5.1. We see that caffe is slower in this experiment, because its workspace limit size of cuDNN-conv interface is 8 * 1024 * 1024, which is smaller in PaddlePaddle and TensorFlow. Note that Caffe will be faster if increasing the workspace limit size.
- GoogletNet: input - 3 * 224 * 224, Time: ms/batch
| BatchSize | 64 | 128 | 256 |
|--------------|-------| -------| --------|
| PaddlePaddle | 613 | 1149 | 2348 |
| TensorFlow | 644 | 1176 | 2219 |
| Caffe | 694 | 1364 | out of memory |
- SmallNet: input - 3 * 32 * 32, Time ms/batch
| BatchSize | 64 | 128 | 256 | 512 |
|--------------|--------| -------- | --------|---------|
| PaddlePaddle | 10.463 | 18.184 | 33.113 | 63.039 |
| TensorFlow | 9 | 15 | 28 | 59 |
| Caffe | 9.373 | 16.6606 | 31.4797 | 59.719 |
**Notation**
All the single-GPU experiments in caffe use `caffe time` to calculate elapsed time, which does not include parameter updating time. However, both PaddlePaddle and TensorFlow experiments contain the parameter updating time. As compared with the total time, this part is relatively little on single machine, we can ignore it.
In Tensorflow, they implement algorithm searching method instead of using the algorithm searching interface in cuDNN.
### Multi-GPU: 4 GPUs
- AlexNet, ms / batch
| total-BatchSize | 128 * 4 | 256 * 4 |
|------------------|----------| -----------|
| PaddlePaddle | 347 | 622 |
| TensorFlow | 377 | 675 |
| Caffe | 1229 | 2435 |
For example, if `total-BatchSize = 128 * 4`, the speedup ratio is calculated by
```
time_at_1gpu_batch_128 * 4 / time_at_4gpu_total_batch_512
= (334 * 4)/347
= 3.85
```
<img src="figs/alexnet-4gpu.png" width="420">
- GoogleNet, ms / batch
| total-BatchSize | 128 * 4 | 256 * 4 |
|-------------------|--------------| ----------- |
| PaddlePaddle | 1178 | 2367 |
| TensorFlow | 1210 | 2292 |
| Caffe | 2007 | out of memory |
<img src="figs/googlenet-4gpu.png" width="420">
## RNN
We use lstm network for text classfication to test benchmark.
### Dataset
- [IMDB](http://www.iro.umontreal.ca/~lisa/deep/data/imdb.pkl)
- Sequence length is 100. In fact, PaddlePaddle supports training with variable-length sequence, but TensorFlow needs to pad. Thus, we also pad sequence length to 100 in PaddlePaddle in order to compare.
- Dictionary size=30000
- Peephole connection is used in `lstmemory` by default in PaddlePaddle. It is also configured in TensorFlow.
### Single-GPU
#### LSTM in Text Classification
Testing `2 lstm layer + fc` network with different hidden size and batch size.
- Batch size = 64, ms / batch
| hidden_size | 256 | 512 | 1280 |
|--------------|-------| -------| --------|
| PaddlePaddle | 83 | 184 | 641 |
| TensorFlow | 175 | 280 | 818 |
- Batch size = 128, ms / batch
| hidden_size | 256 | 512 | 1280 |
|--------------|------- | -------| --------|
| PaddlePaddle | 110 | 261 | 1007 |
| TensorFlow | 181 | 361 | 1237 |
- Batch size = 256, ms / batch
| hidden_size | 256 | 512 | 1280 |
|--------------|-------| -------| --------|
| PaddlePaddle | 170 | 414 | 1655 |
| TensorFlow | 238 | 536 | 1905 |
<img src="figs/rnn_lstm_cls.png" width="600">
#### Seq2Seq
The benchmark of sequence-to-sequence network will be added later.
### Multi GPU: 4 GPUs
#### LSTM in Text Classification
- hidden_size = 256, ms / batch
| batch_size | 256 | 512 |
|--------------| -------| --------|
| PaddlePaddle | 90 | 118 |
| TensorFlow | 226 | 118 |
- hidden_size = 512, ms / batch
| batch_size | 256 | 512 |
|--------------| -------| --------|
| PaddlePaddle | 189 | 268 |
| TensorFlow | 297 | 383 |
<img src="figs/rnn_lstm_4gpus.png" width="420">
#### Seq2Seq
The benchmark of sequence-to-sequence network will be added later.
benchmark/fluid/Dockerfile
浏览文件 @ b8491bfd
......@@ -15,9 +15,6 @@ RUN ln -s /usr/lib/x86_64-linux-gnu/libcudnn.so.7 /usr/lib/libcudnn.so && ln -s
RUN pip install -U pip
RUN pip install -U kubernetes paddlepaddle
RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.cifar.train10()\npaddle.dataset.flowers.fetch()" | python'
RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.mnist.train()\npaddle.dataset.mnist.test()\npaddle.dataset.imdb.fetch()" | python'
RUN sh -c 'echo "import paddle.v2 as paddle\npaddle.dataset.imikolov.fetch()" | python'
RUN pip uninstall -y paddlepaddle && mkdir /workspace
ADD https://raw.githubusercontent.com/PaddlePaddle/cloud/develop/docker/paddle_k8s /usr/bin
......
benchmark/paddle/image/alexnet.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle.trainer_config_helpers import *
height = 227
width = 227
num_class = 1000
batch_size = get_config_arg('batch_size', int, 128)
gp = get_config_arg('layer_num', int, 1)
is_infer = get_config_arg("is_infer", bool, False)
num_samples = get_config_arg('num_samples', int, 2560)
args = {
'height': height,
'width': width,
'color': True,
'num_class': num_class,
'is_infer': is_infer,
'num_samples': num_samples
}
define_py_data_sources2(
"train.list" if not is_infer else None,
"test.list" if is_infer else None,
module="provider",
obj="process",
args=args)
settings(
batch_size=batch_size,
learning_rate=0.01 / batch_size,
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * batch_size))
# conv1
net = data_layer('data', size=height * width * 3)
net = img_conv_layer(
input=net,
filter_size=11,
num_channels=3,
num_filters=96,
stride=4,
padding=1)
net = img_cmrnorm_layer(input=net, size=5, scale=0.0001, power=0.75)
net = img_pool_layer(input=net, pool_size=3, stride=2)
# conv2
net = img_conv_layer(
input=net, filter_size=5, num_filters=256, stride=1, padding=2, groups=gp)
net = img_cmrnorm_layer(input=net, size=5, scale=0.0001, power=0.75)
net = img_pool_layer(input=net, pool_size=3, stride=2)
# conv3
net = img_conv_layer(
input=net, filter_size=3, num_filters=384, stride=1, padding=1)
# conv4
net = img_conv_layer(
input=net, filter_size=3, num_filters=384, stride=1, padding=1, groups=gp)
# conv5
net = img_conv_layer(
input=net, filter_size=3, num_filters=256, stride=1, padding=1, groups=gp)
net = img_pool_layer(input=net, pool_size=3, stride=2)
net = fc_layer(
input=net,
size=4096,
act=ReluActivation(),
layer_attr=ExtraAttr(drop_rate=0.5))
net = fc_layer(
input=net,
size=4096,
act=ReluActivation(),
layer_attr=ExtraAttr(drop_rate=0.5))
net = fc_layer(input=net, size=1000, act=SoftmaxActivation())
if is_infer:
outputs(net)
else:
lab = data_layer('label', num_class)
loss = cross_entropy(input=net, label=lab)
outputs(loss)
benchmark/paddle/image/googlenet.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
#!/usr/bin/env python
from paddle.trainer_config_helpers import *
height = 224
width = 224
num_class = 1000
batch_size = get_config_arg('batch_size', int, 128)
use_gpu = get_config_arg('use_gpu', bool, True)
is_infer = get_config_arg("is_infer", bool, False)
num_samples = get_config_arg('num_samples', int, 2560)
args = {
'height': height,
'width': width,
'color': True,
'num_class': num_class,
'is_infer': is_infer,
'num_samples': num_samples
}
define_py_data_sources2(
"train.list" if not is_infer else None,
"test.list" if is_infer else None,
module="provider",
obj="process",
args=args)
settings(
batch_size=batch_size,
learning_rate=0.01 / batch_size,
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * batch_size))
conv_projection = conv_projection if use_gpu else img_conv_layer
def inception2(name, input, channels, \
filter1,
filter3R, filter3,
filter5R, filter5,
proj):
conv1 = name + '_1'
conv3r = name + '_3r'
conv3 = name + '_3'
conv5r = name + '_5r'
conv5 = name + '_5'
maxpool = name + '_max'
convproj = name + '_proj'
cov1 = img_conv_layer(
name=conv1,
input=input,
filter_size=1,
num_channels=channels,
num_filters=filter1,
stride=1,
padding=0)
cov3r = img_conv_layer(
name=conv3r,
input=input,
filter_size=1,
num_channels=channels,
num_filters=filter3R,
stride=1,
padding=0)
cov3 = img_conv_layer(
name=conv3,
input=cov3r,
filter_size=3,
num_filters=filter3,
stride=1,
padding=1)
cov5r = img_conv_layer(
name=conv5r,
input=input,
filter_size=1,
num_channels=channels,
num_filters=filter5R,
stride=1,
padding=0)
cov5 = img_conv_layer(
name=conv5,
input=cov5r,
filter_size=5,
num_filters=filter5,
stride=1,
padding=2)
pool1 = img_pool_layer(
name=maxpool,
input=input,
pool_size=3,
num_channels=channels,
stride=1,
padding=1)
covprj = img_conv_layer(
name=convproj,
input=pool1,
filter_size=1,
num_filters=proj,
stride=1,
padding=0)
cat = concat_layer(name=name, input=[cov1, cov3, cov5, covprj])
return cat
def inception(name, input, channels, \
filter1,
filter3R, filter3,
filter5R, filter5,
proj):
cov1 = conv_projection(
input=input,
filter_size=1,
num_channels=channels,
num_filters=filter1,
stride=1,
padding=0)
cov3r = img_conv_layer(
name=name + '_3r',
input=input,
filter_size=1,
num_channels=channels,
num_filters=filter3R,
stride=1,
padding=0)
cov3 = conv_projection(
input=cov3r, filter_size=3, num_filters=filter3, stride=1, padding=1)
cov5r = img_conv_layer(
name=name + '_5r',
input=input,
filter_size=1,
num_channels=channels,
num_filters=filter5R,
stride=1,
padding=0)
cov5 = conv_projection(
input=cov5r, filter_size=5, num_filters=filter5, stride=1, padding=2)
pool1 = img_pool_layer(
name=name + '_max',
input=input,
pool_size=3,
num_channels=channels,
stride=1,
padding=1)
covprj = conv_projection(
input=pool1, filter_size=1, num_filters=proj, stride=1, padding=0)
cat = concat_layer(
name=name,
input=[cov1, cov3, cov5, covprj],
bias_attr=True if use_gpu else False,
act=ReluActivation())
return cat
data = data_layer(name="input", size=3 * height * width)
# stage 1
conv1 = img_conv_layer(
name="conv1",
input=data,
filter_size=7,
num_channels=3,
num_filters=64,
stride=2,
padding=3)
pool1 = img_pool_layer(
name="pool1", input=conv1, pool_size=3, num_channels=64, stride=2)
# stage 2
conv2_1 = img_conv_layer(
name="conv2_1",
input=pool1,
filter_size=1,
num_filters=64,
stride=1,
padding=0)
conv2_2 = img_conv_layer(
name="conv2_2",
input=conv2_1,
filter_size=3,
num_filters=192,
stride=1,
padding=1)
pool2 = img_pool_layer(
name="pool2", input=conv2_2, pool_size=3, num_channels=192, stride=2)
# stage 3
ince3a = inception("ince3a", pool2, 192, 64, 96, 128, 16, 32, 32)
ince3b = inception("ince3b", ince3a, 256, 128, 128, 192, 32, 96, 64)
pool3 = img_pool_layer(
name="pool3", input=ince3b, num_channels=480, pool_size=3, stride=2)
# stage 4
ince4a = inception("ince4a", pool3, 480, 192, 96, 208, 16, 48, 64)
ince4b = inception("ince4b", ince4a, 512, 160, 112, 224, 24, 64, 64)
ince4c = inception("ince4c", ince4b, 512, 128, 128, 256, 24, 64, 64)
ince4d = inception("ince4d", ince4c, 512, 112, 144, 288, 32, 64, 64)
ince4e = inception("ince4e", ince4d, 528, 256, 160, 320, 32, 128, 128)
pool4 = img_pool_layer(
name="pool4", input=ince4e, num_channels=832, pool_size=3, stride=2)
# stage 5
ince5a = inception("ince5a", pool4, 832, 256, 160, 320, 32, 128, 128)
ince5b = inception("ince5b", ince5a, 832, 384, 192, 384, 48, 128, 128)
pool5 = img_pool_layer(
name="pool5",
input=ince5b,
num_channels=1024,
pool_size=7,
stride=7,
pool_type=AvgPooling())
# We remove loss1 and loss2 for all system when testing benchmark
# output 1
# pool_o1 = img_pool_layer(name="pool_o1", input=ince4a, num_channels=512, pool_size=5, stride=3, pool_type=AvgPooling())
# conv_o1 = img_conv_layer(name="conv_o1", input=pool_o1, filter_size=1, num_filters=128, stride=1, padding=0)
# fc_o1 = fc_layer(name="fc_o1", input=conv_o1, size=1024, layer_attr=ExtraAttr(drop_rate=0.7), act=ReluActivation())
# out1 = fc_layer(name="output1", input=fc_o1, size=1000, act=SoftmaxActivation())
# loss1 = cross_entropy(name='loss1', input=out1, label=lab, coeff=0.3)
# output 2
#pool_o2 = img_pool_layer(name="pool_o2", input=ince4d, num_channels=528, pool_size=5, stride=3, pool_type=AvgPooling())
#conv_o2 = img_conv_layer(name="conv_o2", input=pool_o2, filter_size=1, num_filters=128, stride=1, padding=0)
#fc_o2 = fc_layer(name="fc_o2", input=conv_o2, size=1024, layer_attr=ExtraAttr(drop_rate=0.7), act=ReluActivation())
#out2 = fc_layer(name="output2", input=fc_o2, size=1000, act=SoftmaxActivation())
#loss2 = cross_entropy(name='loss2', input=out2, label=lab, coeff=0.3)
# output 3
dropout = dropout_layer(name="dropout", input=pool5, dropout_rate=0.4)
out3 = fc_layer(
name="output3", input=dropout, size=1000, act=SoftmaxActivation())
if is_infer:
outputs(out3)
else:
lab = data_layer(name="label", size=num_class)
loss3 = cross_entropy(name='loss3', input=out3, label=lab)
outputs(loss3)
benchmark/paddle/image/plotlog.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
import argparse
import matplotlib.pyplot as plt
def parse_args():
parser = argparse.ArgumentParser('Parse Log')
parser.add_argument(
'--file_path', '-f', type=str, help='the path of the log file')
parser.add_argument(
'--sample_rate',
'-s',
type=float,
default=1.0,
help='the rate to take samples from log')
parser.add_argument(
'--log_period', '-p', type=int, default=1, help='the period of log')
args = parser.parse_args()
return args
def parse_file(file_name):
loss = []
error = []
with open(file_name) as f:
for i, line in enumerate(f):
line = line.strip()
if not line.startswith('pass'):
continue
line_split = line.split(' ')
if len(line_split) != 5:
continue
loss_str = line_split[2][:-1]
cur_loss = float(loss_str.split('=')[-1])
loss.append(cur_loss)
err_str = line_split[3][:-1]
cur_err = float(err_str.split('=')[-1])
error.append(cur_err)
accuracy = [1.0 - err for err in error]
return loss, accuracy
def sample(metric, sample_rate):
interval = int(1.0 / sample_rate)
if interval > len(metric):
return metric[:1]
num = len(metric) / interval
idx = [interval * i for i in range(num)]
metric_sample = [metric[id] for id in idx]
return metric_sample
def plot_metric(metric,
batch_id,
graph_title,
line_style='b-',
line_label='y',
line_num=1):
plt.figure()
plt.title(graph_title)
if line_num == 1:
plt.plot(batch_id, metric, line_style, label=line_label)
else:
for i in range(line_num):
plt.plot(batch_id, metric[i], line_style[i], label=line_label[i])
plt.xlabel('batch')
plt.ylabel(graph_title)
plt.legend()
plt.savefig(graph_title + '.jpg')
plt.close()
def main():
args = parse_args()
assert args.sample_rate > 0. and args.sample_rate <= 1.0, "The sample rate should in the range (0, 1]."
loss, accuracy = parse_file(args.file_path)
batch = [args.log_period * i for i in range(len(loss))]
batch_sample = sample(batch, args.sample_rate)
loss_sample = sample(loss, args.sample_rate)
accuracy_sample = sample(accuracy, args.sample_rate)
plot_metric(loss_sample, batch_sample, 'loss', line_label='loss')
plot_metric(
accuracy_sample,
batch_sample,
'accuracy',
line_style='g-',
line_label='accuracy')
if __name__ == '__main__':
main()
benchmark/paddle/image/resnet.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
#!/usr/bin/env python
from paddle.trainer_config_helpers import *
height = 224
width = 224
num_class = 1000
batch_size = get_config_arg('batch_size', int, 64)
layer_num = get_config_arg("layer_num", int, 50)
is_infer = get_config_arg("is_infer", bool, False)
num_samples = get_config_arg('num_samples', int, 2560)
args = {
'height': height,
'width': width,
'color': True,
'num_class': num_class,
'is_infer': is_infer,
'num_samples': num_samples
}
define_py_data_sources2(
"train.list" if not is_infer else None,
"test.list" if is_infer else None,
module="provider",
obj="process",
args=args)
settings(
batch_size=batch_size,
learning_rate=0.01 / batch_size,
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * batch_size))
#######################Network Configuration #############
def conv_bn_layer(name,
input,
filter_size,
num_filters,
stride,
padding,
channels=None,
active_type=ReluActivation()):
"""
A wrapper for conv layer with batch normalization layers.
Note:
conv layer has no activation.
"""
tmp = img_conv_layer(
name=name + "_conv",
input=input,
filter_size=filter_size,
num_channels=channels,
num_filters=num_filters,
stride=stride,
padding=padding,
act=LinearActivation(),
bias_attr=False)
return batch_norm_layer(
name=name + "_bn",
input=tmp,
act=active_type,
use_global_stats=is_infer)
def bottleneck_block(name, input, num_filters1, num_filters2):
"""
A wrapper for bottlenect building block in ResNet.
Last conv_bn_layer has no activation.
Addto layer has activation of relu.
"""
last_name = conv_bn_layer(
name=name + '_branch2a',
input=input,
filter_size=1,
num_filters=num_filters1,
stride=1,
padding=0)
last_name = conv_bn_layer(
name=name + '_branch2b',
input=last_name,
filter_size=3,
num_filters=num_filters1,
stride=1,
padding=1)
last_name = conv_bn_layer(
name=name + '_branch2c',
input=last_name,
filter_size=1,
num_filters=num_filters2,
stride=1,
padding=0,
active_type=LinearActivation())
return addto_layer(
name=name + "_addto", input=[input, last_name], act=ReluActivation())
def mid_projection(name, input, num_filters1, num_filters2, stride=2):
"""
A wrapper for middile projection in ResNet.
projection shortcuts are used for increasing dimensions,
and other shortcuts are identity
branch1: projection shortcuts are used for increasing
dimensions, has no activation.
branch2x: bottleneck building block, shortcuts are identity.
"""
# stride = 2
branch1 = conv_bn_layer(
name=name + '_branch1',
input=input,
filter_size=1,
num_filters=num_filters2,
stride=stride,
padding=0,
active_type=LinearActivation())
last_name = conv_bn_layer(
name=name + '_branch2a',
input=input,
filter_size=1,
num_filters=num_filters1,
stride=stride,
padding=0)
last_name = conv_bn_layer(
name=name + '_branch2b',
input=last_name,
filter_size=3,
num_filters=num_filters1,
stride=1,
padding=1)
last_name = conv_bn_layer(
name=name + '_branch2c',
input=last_name,
filter_size=1,
num_filters=num_filters2,
stride=1,
padding=0,
active_type=LinearActivation())
return addto_layer(
name=name + "_addto", input=[branch1, last_name], act=ReluActivation())
img = data_layer(name='image', size=height * width * 3)
def deep_res_net(res2_num=3, res3_num=4, res4_num=6, res5_num=3):
"""
A wrapper for 50,101,152 layers of ResNet.
res2_num: number of blocks stacked in conv2_x
res3_num: number of blocks stacked in conv3_x
res4_num: number of blocks stacked in conv4_x
res5_num: number of blocks stacked in conv5_x
"""
# For ImageNet
# conv1: 112x112
tmp = conv_bn_layer(
"conv1",
input=img,
filter_size=7,
channels=3,
num_filters=64,
stride=2,
padding=3)
tmp = img_pool_layer(name="pool1", input=tmp, pool_size=3, stride=2)
# conv2_x: 56x56
tmp = mid_projection(
name="res2_1", input=tmp, num_filters1=64, num_filters2=256, stride=1)
for i in xrange(2, res2_num + 1, 1):
tmp = bottleneck_block(
name="res2_" + str(i), input=tmp, num_filters1=64, num_filters2=256)
# conv3_x: 28x28
tmp = mid_projection(
name="res3_1", input=tmp, num_filters1=128, num_filters2=512)
for i in xrange(2, res3_num + 1, 1):
tmp = bottleneck_block(
name="res3_" + str(i),
input=tmp,
num_filters1=128,
num_filters2=512)
# conv4_x: 14x14
tmp = mid_projection(
name="res4_1", input=tmp, num_filters1=256, num_filters2=1024)
for i in xrange(2, res4_num + 1, 1):
tmp = bottleneck_block(
name="res4_" + str(i),
input=tmp,
num_filters1=256,
num_filters2=1024)
# conv5_x: 7x7
tmp = mid_projection(
name="res5_1", input=tmp, num_filters1=512, num_filters2=2048)
for i in xrange(2, res5_num + 1, 1):
tmp = bottleneck_block(
name="res5_" + str(i),
input=tmp,
num_filters1=512,
num_filters2=2048)
tmp = img_pool_layer(
name='avgpool',
input=tmp,
pool_size=7,
stride=1,
pool_type=AvgPooling())
return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation())
if layer_num == 50:
resnet = deep_res_net(3, 4, 6, 3)
elif layer_num == 101:
resnet = deep_res_net(3, 4, 23, 3)
elif layer_num == 152:
resnet = deep_res_net(3, 8, 36, 3)
else:
print("Wrong layer number.")
if is_infer:
outputs(resnet)
else:
lbl = data_layer(name="label", size=num_class)
loss = cross_entropy(name='loss', input=resnet, label=lbl)
outputs(loss)
benchmark/paddle/image/run.sh 已删除 100755 → 0
浏览文件 @ 8bda4ab2
#!/bin/bash
set -e
function train() {
cfg=$1
thread=$2
bz=$3
args="batch_size=$3"
prefix=$4
paddle train --job=time \
--config=$cfg \
--use_gpu=True \
--trainer_count=$thread \
--log_period=10 \
--test_period=100 \
--config_args=$args \
> logs/$prefix-${thread}gpu-$bz.log 2>&1
}
if [ ! -d "train.list" ]; then
echo " " > train.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
#========single-gpu=========#
# alexnet
train alexnet.py 1 64 alexnet
train alexnet.py 1 128 alexnet
train alexnet.py 1 256 alexnet
train alexnet.py 1 512 alexnet
# googlenet
train googlenet.py 1 64 googlenet
train googlenet.py 1 128 googlenet
train googlenet.py 1 256 googlenet
# smallnet
train smallnet_mnist_cifar.py 1 64 smallnet
train smallnet_mnist_cifar.py 1 128 smallnet
train smallnet_mnist_cifar.py 1 256 smallnet
train smallnet_mnist_cifar.py 1 512 smallnet
############################
#========multi-gpus=========#
train alexnet.py 4 512 alexnet
train alexnet.py 4 1024 alexnet
train googlenet.py 4 512 googlenet
train googlenet.py 4 1024 googlenet
benchmark/paddle/image/run_mkl_infer.sh 已删除 100755 → 0
浏览文件 @ 8bda4ab2
#!/bin/bash
set -e
function clock_to_seconds() {
hours=`echo $1 | awk -F ':' '{print $1}'`
mins=`echo $1 | awk -F ':' '{print $2}'`
secs=`echo $1 | awk -F ':' '{print $3}'`
echo `awk 'BEGIN{printf "%.2f",('$secs' + '$mins' * 60 + '$hours' * 3600)}'`
}
function infer() {
unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
topology=$1
layer_num=$2
bs=$3
use_mkldnn=$4
if [ $4 == "True" ]; then
thread=1
log="logs/infer-${topology}-${layer_num}-mkldnn-${bs}.log"
elif [ $4 == "False" ]; then
thread=`nproc`
if [ $thread -gt $bs ]; then
thread=$bs
fi
log="logs/infer-${topology}-${layer_num}-${thread}mklml-${bs}.log"
else
echo "Wrong input $4, use True or False."
exit 0
fi
models_in="models/${topology}-${layer_num}/pass-00000/"
if [ ! -d $models_in ]; then
echo "Training model ${topology}_${layer_num}"
paddle train --job=train \
--config="${topology}.py" \
--use_mkldnn=True \
--use_gpu=False \
--trainer_count=1 \
--num_passes=1 \
--save_dir="models/${topology}-${layer_num}" \
--config_args="batch_size=128,layer_num=${layer_num},num_samples=256" \
> /dev/null 2>&1
echo "Done"
fi
log_period=$((256 / bs))
paddle train --job=test \
--config="${topology}.py" \
--use_mkldnn=$use_mkldnn \
--use_gpu=False \
--trainer_count=$thread \
--log_period=$log_period \
--config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True" \
--init_model_path=$models_in \
2>&1 | tee ${log}
# calculate the last 5 logs period time of 1280 samples,
# the time before are burning time.
start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
start_sec=`clock_to_seconds $start`
end_sec=`clock_to_seconds $end`
fps=`awk 'BEGIN{printf "%.2f",(1280 / ('$end_sec' - '$start_sec'))}'`
echo "Last 1280 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
}
if [ ! -f "train.list" ]; then
echo " " > train.list
fi
if [ ! -f "test.list" ]; then
echo " " > test.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
if [ ! -d "models" ]; then
mkdir -p models
fi
# inference benchmark
for use_mkldnn in True False; do
for batchsize in 1 2 4 8 16; do
infer vgg 19 $batchsize $use_mkldnn
infer resnet 50 $batchsize $use_mkldnn
infer googlenet v1 $batchsize $use_mkldnn
infer alexnet 2 $batchsize $use_mkldnn
done
done
benchmark/paddle/image/run_mkl_train.sh 已删除 100755 → 0
浏览文件 @ 8bda4ab2
#!/bin/bash
set -e
function train() {
unset OMP_NUM_THREADS MKL_NUM_THREADS OMP_DYNAMIC KMP_AFFINITY
topology=$1
layer_num=$2
bs=$3
use_mkldnn=$4
if [ $4 == "True" ]; then
thread=1
log="logs/train-${topology}-${layer_num}-mkldnn-${bs}.log"
elif [ $4 == "False" ]; then
thread=`nproc`
# each trainer_count use only 1 core to avoid conflict
log="logs/train-${topology}-${layer_num}-${thread}mklml-${bs}.log"
else
echo "Wrong input $4, use True or False."
exit 0
fi
args="batch_size=${bs},layer_num=${layer_num}"
config="${topology}.py"
paddle train --job=time \
--config=$config \
--use_mkldnn=$use_mkldnn \
--use_gpu=False \
--trainer_count=$thread \
--log_period=10 \
--test_period=100 \
--config_args=$args \
2>&1 | tee ${log}
avg_time=`tail ${log} -n 1 | awk -F ' ' '{print $8}' | sed 's/avg=//'`
fps=`awk 'BEGIN{printf "%.2f",('$bs' / '$avg_time' * 1000)}'`
echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
}
if [ ! -f "train.list" ]; then
echo " " > train.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
# training benchmark
for use_mkldnn in True False; do
for batchsize in 64 128 256; do
train vgg 19 $batchsize $use_mkldnn
train resnet 50 $batchsize $use_mkldnn
train googlenet v1 $batchsize $use_mkldnn
train alexnet 2 $batchsize $use_mkldnn
done
done
benchmark/paddle/image/run_openblas_infer.sh 已删除 100755 → 0
浏览文件 @ 8bda4ab2
#!/bin/bash
set -e
function clock_to_seconds() {
hours=`echo $1 | awk -F ':' '{print $1}'`
mins=`echo $1 | awk -F ':' '{print $2}'`
secs=`echo $1 | awk -F ':' '{print $3}'`
echo `awk 'BEGIN{printf "%.2f",('$secs' + '$mins' * 60 + '$hours' * 3600)}'`
}
function infer() {
export OPENBLAS_MAIN_FREE=1
topology=$1
layer_num=$2
bs=$3
trainers=`nproc`
if [ $trainers -gt $bs ]; then
trainers=$bs
fi
log="logs/infer-${topology}-${layer_num}-${trainers}openblas-${bs}.log"
threads=$((`nproc` / trainers))
if [ $threads -eq 0 ]; then
threads=1
fi
export OPENBLAS_NUM_THREADS=$threads
models_in="models/${topology}-${layer_num}/pass-00000/"
if [ ! -d $models_in ]; then
echo "./run_mkl_infer.sh to save the model first"
exit 0
fi
log_period=$((32 / bs))
paddle train --job=test \
--config="${topology}.py" \
--use_mkldnn=False \
--use_gpu=False \
--trainer_count=$trainers \
--log_period=$log_period \
--config_args="batch_size=${bs},layer_num=${layer_num},is_infer=True,num_samples=256" \
--init_model_path=$models_in \
2>&1 | tee ${log}
# calculate the last 5 logs period time of 160(=32*5) samples,
# the time before are burning time.
start=`tail ${log} -n 7 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
end=`tail ${log} -n 2 | head -n 1 | awk -F ' ' '{print $2}' | xargs`
start_sec=`clock_to_seconds $start`
end_sec=`clock_to_seconds $end`
fps=`awk 'BEGIN{printf "%.2f",(160 / ('$end_sec' - '$start_sec'))}'`
echo "Last 160 samples start: ${start}(${start_sec} sec), end: ${end}(${end_sec} sec;" >> ${log}
echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
}
if [ ! -f "train.list" ]; then
echo " " > train.list
fi
if [ ! -f "test.list" ]; then
echo " " > test.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
# inference benchmark
for batchsize in 1 2 4 8 16; do
infer vgg 19 $batchsize
infer resnet 50 $batchsize
infer googlenet v1 $batchsize
infer alexnet 2 $batchsize
done
benchmark/paddle/image/run_openblas_train.sh 已删除 100755 → 0
浏览文件 @ 8bda4ab2
#!/bin/bash
set -e
function train() {
export OPENBLAS_NUM_THREADS=1
topology=$1
layer_num=$2
bs=$3
thread=`nproc`
# each trainer_count use only 1 core to avoid conflict
log="logs/train-${topology}-${layer_num}-${thread}openblas-${bs}.log"
args="batch_size=${bs},layer_num=${layer_num}"
config="${topology}.py"
paddle train --job=time \
--config=$config \
--use_mkldnn=False \
--use_gpu=False \
--trainer_count=$thread \
--log_period=3 \
--test_period=30 \
--config_args=$args \
2>&1 | tee ${log}
avg_time=`tail ${log} -n 1 | awk -F ' ' '{print $8}' | sed 's/avg=//'`
fps=`awk 'BEGIN{printf "%.2f",('$bs' / '$avg_time' * 1000)}'`
echo "FPS: $fps images/sec" 2>&1 | tee -a ${log}
}
if [ ! -f "train.list" ]; then
echo " " > train.list
fi
if [ ! -d "logs" ]; then
mkdir logs
fi
# training benchmark
for batchsize in 64 128 256; do
train vgg 19 $batchsize
train resnet 50 $batchsize
train googlenet v1 $batchsize
train alexnet 2 $batchsize
done
benchmark/paddle/image/smallnet_mnist_cifar.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
#!/usr/bin/env python
from paddle.trainer_config_helpers import *
height = 32
width = 32
num_class = 10
batch_size = get_config_arg('batch_size', int, 128)
args = {'height': height, 'width': width, 'color': True, 'num_class': num_class}
define_py_data_sources2(
"train.list", None, module="provider", obj="process", args=args)
settings(
batch_size=batch_size,
learning_rate=0.01 / batch_size,
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * batch_size))
# conv1
net = data_layer('data', size=height * width * 3)
net = img_conv_layer(
input=net,
filter_size=5,
num_channels=3,
num_filters=32,
stride=1,
padding=2)
net = img_pool_layer(input=net, pool_size=3, stride=2, padding=1)
# conv2
net = img_conv_layer(
input=net, filter_size=5, num_filters=32, stride=1, padding=2)
net = img_pool_layer(
input=net, pool_size=3, stride=2, padding=1, pool_type=AvgPooling())
# conv3
net = img_conv_layer(
input=net, filter_size=3, num_filters=64, stride=1, padding=1)
net = img_pool_layer(
input=net, pool_size=3, stride=2, padding=1, pool_type=AvgPooling())
net = fc_layer(input=net, size=64, act=ReluActivation())
net = fc_layer(input=net, size=10, act=SoftmaxActivation())
lab = data_layer('label', num_class)
loss = classification_cost(input=net, label=lab)
outputs(loss)
benchmark/paddle/image/vgg.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
#!/usr/bin/env python
from paddle.trainer_config_helpers import *
height = 224
width = 224
num_class = 1000
batch_size = get_config_arg('batch_size', int, 64)
layer_num = get_config_arg('layer_num', int, 19)
is_infer = get_config_arg("is_infer", bool, False)
num_samples = get_config_arg('num_samples', int, 2560)
args = {
'height': height,
'width': width,
'color': True,
'num_class': num_class,
'is_infer': is_infer,
'num_samples': num_samples
}
define_py_data_sources2(
"train.list" if not is_infer else None,
"test.list" if is_infer else None,
module="provider",
obj="process",
args=args)
settings(
batch_size=batch_size,
learning_rate=0.001 / batch_size,
learning_method=MomentumOptimizer(0.9),
regularization=L2Regularization(0.0005 * batch_size))
img = data_layer(name='image', size=height * width * 3)
def vgg_network(vgg_num=3):
tmp = img_conv_group(
input=img,
num_channels=3,
conv_padding=1,
conv_num_filter=[64, 64],
conv_filter_size=3,
conv_act=ReluActivation(),
pool_size=2,
pool_stride=2,
pool_type=MaxPooling())
tmp = img_conv_group(
input=tmp,
conv_num_filter=[128, 128],
conv_padding=1,
conv_filter_size=3,
conv_act=ReluActivation(),
pool_stride=2,
pool_type=MaxPooling(),
pool_size=2)
channels = []
for i in range(vgg_num):
channels.append(256)
tmp = img_conv_group(
input=tmp,
conv_num_filter=channels,
conv_padding=1,
conv_filter_size=3,
conv_act=ReluActivation(),
pool_stride=2,
pool_type=MaxPooling(),
pool_size=2)
channels = []
for i in range(vgg_num):
channels.append(512)
tmp = img_conv_group(
input=tmp,
conv_num_filter=channels,
conv_padding=1,
conv_filter_size=3,
conv_act=ReluActivation(),
pool_stride=2,
pool_type=MaxPooling(),
pool_size=2)
tmp = img_conv_group(
input=tmp,
conv_num_filter=channels,
conv_padding=1,
conv_filter_size=3,
conv_act=ReluActivation(),
pool_stride=2,
pool_type=MaxPooling(),
pool_size=2)
tmp = fc_layer(
input=tmp,
size=4096,
act=ReluActivation(),
layer_attr=ExtraAttr(drop_rate=0.5))
tmp = fc_layer(
input=tmp,
size=4096,
act=ReluActivation(),
layer_attr=ExtraAttr(drop_rate=0.5))
return fc_layer(input=tmp, size=num_class, act=SoftmaxActivation())
if layer_num == 16:
vgg = vgg_network(3)
elif layer_num == 19:
vgg = vgg_network(4)
else:
print("Wrong layer number.")
if is_infer:
outputs(vgg)
else:
lab = data_layer('label', num_class)
loss = cross_entropy(input=vgg, label=lab)
outputs(loss)
benchmark/paddle/rnn/provider.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import io, os
import random
import numpy as np
import six.moves.cPickle as pickle
from paddle.trainer.PyDataProvider2 import *
def remove_unk(x, n_words):
return [[1 if w >= n_words else w for w in sen] for sen in x]
# ==============================================================
# tensorflow uses fixed length, but PaddlePaddle can process
# variable-length. Padding is used in benchmark in order to
# compare with other platform.
# ==============================================================
def pad_sequences(sequences,
maxlen=None,
dtype='int32',
padding='post',
truncating='post',
value=0.):
lengths = [len(s) for s in sequences]
nb_samples = len(sequences)
if maxlen is None:
maxlen = np.max(lengths)
x = (np.ones((nb_samples, maxlen)) * value).astype(dtype)
for idx, s in enumerate(sequences):
if len(s) == 0:
continue # empty list was found
if truncating == 'pre':
trunc = s[-maxlen:]
elif truncating == 'post':
trunc = s[:maxlen]
else:
raise ValueError("Truncating type '%s' not understood" % padding)
if padding == 'post':
x[idx, :len(trunc)] = trunc
elif padding == 'pre':
x[idx, -len(trunc):] = trunc
else:
raise ValueError("Padding type '%s' not understood" % padding)
return x
def initHook(settings, vocab_size, pad_seq, maxlen, **kwargs):
settings.vocab_size = vocab_size
settings.pad_seq = pad_seq
settings.maxlen = maxlen
settings.input_types = [
integer_value_sequence(vocab_size), integer_value(2)
]
@provider(
init_hook=initHook, min_pool_size=-1, cache=CacheType.CACHE_PASS_IN_MEM)
def process(settings, file):
f = open(file, 'rb')
train_set = pickle.load(f)
f.close()
x, y = train_set
# remove unk, namely remove the words out of dictionary
x = remove_unk(x, settings.vocab_size)
if settings.pad_seq:
x = pad_sequences(x, maxlen=settings.maxlen, value=0.)
for i in range(len(y)):
yield map(int, x[i]), int(y[i])
benchmark/paddle/rnn/rnn.py 已删除 100755 → 0
浏览文件 @ 8bda4ab2
#!/usr/bin/env python
from paddle.trainer_config_helpers import *
import imdb
num_class = 2
vocab_size = 30000
fixedlen = 100
batch_size = get_config_arg('batch_size', int, 128)
lstm_num = get_config_arg('lstm_num', int, 1)
hidden_size = get_config_arg('hidden_size', int, 128)
# whether to pad sequence into fixed length
pad_seq = get_config_arg('pad_seq', bool, True)
imdb.create_data('imdb.pkl')
args = {'vocab_size': vocab_size, 'pad_seq': pad_seq, 'maxlen': fixedlen}
define_py_data_sources2(
"train.list", None, module="provider", obj="process", args=args)
settings(
batch_size=batch_size,
learning_rate=2e-3,
learning_method=AdamOptimizer(),
regularization=L2Regularization(8e-4),
gradient_clipping_threshold=25)
net = data_layer('data', size=vocab_size)
net = embedding_layer(input=net, size=128)
for i in xrange(lstm_num):
net = simple_lstm(input=net, size=hidden_size)
net = last_seq(input=net)
net = fc_layer(input=net, size=2, act=SoftmaxActivation())
lab = data_layer('label', num_class)
loss = classification_cost(input=net, label=lab)
outputs(loss)
benchmark/paddle/rnn/run.sh 已删除 100755 → 0
浏览文件 @ 8bda4ab2
#!/bin/bash
set -e
function train() {
cfg=$1
thread=$2
args="lstm_num=${3},seq_pad=${4},hidden_size=${5},batch_size=${6}"
paddle train --job=time \
--config=$cfg \
--use_gpu=1 \
--trainer_count=$thread \
--log_period=10 \
--test_period=100 \
--num_passes=1 \
--feed_data=1 \
--config_args=$args \
>logs/rnn-pad${4}-${thread}gpu-lstm${3}-batch${6}-hid${5}.log 2>&1
}
if [ ! -d "logs" ]; then
mkdir logs
fi
## padding, single gpu
#-----config--gpu--lstm_num--padding--hidden_size--batch_size
## lstm_num=2, batch_size=64
train rnn.py 1 2 1 256 64
train rnn.py 1 2 1 512 64
train rnn.py 1 2 1 1280 64
## lstm_num=2, batch_size=128
train rnn.py 1 2 1 256 128
train rnn.py 1 2 1 512 128
train rnn.py 1 2 1 1280 128
## lstm_num=4, batch_size=256
train rnn.py 1 2 1 256 256
train rnn.py 1 2 1 512 256
train rnn.py 1 2 1 1280 256
#==================multi gpus=====================#
# hidden_size=256, lstm_num=2, different batch size
train rnn.py 4 2 1 256 128
train rnn.py 4 2 1 256 256
train rnn.py 4 2 1 256 512
# hidden_size=512, lstm_num=4, different batch size
train rnn.py 4 2 1 512 128
train rnn.py 4 2 1 512 256
train rnn.py 4 2 1 512 512
benchmark/tensorflow/machine_translation.py
浏览文件 @ b8491bfd
......@@ -35,8 +35,6 @@ import os
import argparse
import time
import paddle.v2 as paddle
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--embedding_dim",
......
benchmark/tensorflow/mnist.py
浏览文件 @ b8491bfd
......@@ -21,7 +21,6 @@ import time
import numpy as np
import tensorflow as tf
import paddle.v2 as paddle
DTYPE = tf.float32
......
benchmark/tensorflow/resnet.py
浏览文件 @ b8491bfd
......@@ -27,7 +27,6 @@ import argparse
import time
import numpy as np
import paddle.v2 as paddle
import tensorflow as tf
DTYPE = tf.float32
......
benchmark/tensorflow/stacked_dynamic_lstm.py
浏览文件 @ b8491bfd
......@@ -21,8 +21,6 @@ import argparse
import time
import tensorflow as tf
import paddle.v2 as paddle
def parse_args():
parser = argparse.ArgumentParser("LSTM model benchmark.")
......
benchmark/tensorflow/vgg.py
浏览文件 @ b8491bfd
......@@ -13,7 +13,6 @@
# limitations under the License.
"""VGG16 benchmark in TensorFlow"""
import tensorflow as tf
import paddle.v2 as paddle
import numpy as np
import argparse
import time
......
cmake/configure.cmake
浏览文件 @ b8491bfd
......@@ -20,31 +20,10 @@ if(WITH_DSO)
add_definitions(-DPADDLE_USE_DSO)
endif(WITH_DSO)
if(WITH_DOUBLE)
add_definitions(-DPADDLE_TYPE_DOUBLE)
endif(WITH_DOUBLE)
if(WITH_ARM_FP16)
add_definitions(-DPADDLE_ARM_FP16)
add_definitions("-march=armv8.2-a+fp16+simd")
endif(WITH_ARM_FP16)
if(WITH_TESTING)
add_definitions(-DPADDLE_WITH_TESTING)
endif(WITH_TESTING)
if(NOT WITH_TIMER)
add_definitions(-DPADDLE_DISABLE_TIMER)
endif(NOT WITH_TIMER)
if(USE_EIGEN_FOR_BLAS)
add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
endif(USE_EIGEN_FOR_BLAS)
if(EIGEN_USE_THREADS)
add_definitions(-DEIGEN_USE_THREADS)
endif(EIGEN_USE_THREADS)
if(NOT WITH_PROFILER)
add_definitions(-DPADDLE_DISABLE_PROFILER)
endif(NOT WITH_PROFILER)
......@@ -78,10 +57,6 @@ if(WIN32)
endif(NOT MSVC)
endif(WIN32)
if(NOT WITH_GOLANG)
add_definitions(-DPADDLE_WITHOUT_GOLANG)
endif(NOT WITH_GOLANG)
if(WITH_PSLIB)
add_definitions(-DPADDLE_WITH_PSLIB)
endif()
......@@ -171,55 +146,6 @@ if(WITH_DISTRIBUTE)
add_definitions(-DPADDLE_WITH_DISTRIBUTE)
endif()
if(WITH_GOLANG)
# we need to symlink Paddle directory into GOPATH. If we
# don't do it and we have code that depends on Paddle, go
# get ./... will download a new Paddle repo from Github,
# without the changes in our current Paddle repo that we
# want to build.
set(GOPATH "${CMAKE_CURRENT_BINARY_DIR}/go")
file(MAKE_DIRECTORY ${GOPATH})
set(PADDLE_IN_GOPATH "${GOPATH}/src/github.com/PaddlePaddle/Paddle")
file(MAKE_DIRECTORY "${PADDLE_IN_GOPATH}")
set(PADDLE_GO_PATH "${CMAKE_SOURCE_DIR}/go")
add_custom_target(go_path)
add_custom_command(TARGET go_path
# Symlink Paddle directory into GOPATH
COMMAND mkdir -p ${PADDLE_IN_GOPATH}
COMMAND rm -rf ${PADDLE_IN_GOPATH}
COMMAND ln -sf ${CMAKE_SOURCE_DIR} ${PADDLE_IN_GOPATH}
# Automatically get all dependencies specified in the source code
# We can't run `go get -d ./...` for every target, because
# multiple `go get` can not run concurrently, but make need to be
# able to run with multiple jobs.
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
)
if (GLIDE_INSTALL)
if(EXISTS $ENV{GOPATH}/bin/glide)
set(GLIDE "$ENV{GOPATH}/bin/glide")
else()
message(FATAL_ERROR "no glide executeble found: $ENV{GOPATH}/bin/glide")
endif()
# this command will only run when the file it depends is missing
# or has changed, or the output is missing.
add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide
COMMAND env GOPATH=${GOPATH} ${GLIDE} install
COMMAND touch ${CMAKE_BINARY_DIR}/glide
DEPENDS ${PADDLE_SOURCE_DIR}/go/glide.lock
WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go"
)
# depends on the custom command which outputs
# ${CMAKE_BINARY_DIR}/glide, the custom command does not need to
# run every time this target is built.
add_custom_target(go_vendor DEPENDS ${CMAKE_BINARY_DIR}/glide go_path)
endif()
endif(WITH_GOLANG)
if(WITH_GRPC)
add_definitions(-DPADDLE_WITH_GRPC)
endif(WITH_GRPC)
......
cmake/cuda.cmake
浏览文件 @ b8491bfd
......@@ -168,10 +168,7 @@ elseif (${CUDA_VERSION} LESS 11.0) # CUDA 10.x
endif()
include_directories(${CUDA_INCLUDE_DIRS})
list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
if(NOT WITH_DSO)
# TODO(panyx0718): CUPTI only allows DSO?
list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUPTI_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY} ${NCCL_LIBRARY})
if(WIN32)
set_property(GLOBAL PROPERTY CUDA_MODULES ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
endif(WIN32)
......
cmake/external/anakin.cmake
浏览文件 @ b8491bfd
......@@ -74,5 +74,3 @@ add_dependencies(anakin_shared extern_anakin)
add_library(anakin_saber SHARED IMPORTED GLOBAL)
set_property(TARGET anakin_saber PROPERTY IMPORTED_LOCATION ${ANAKIN_SABER_LIB})
add_dependencies(anakin_saber extern_anakin)
list(APPEND external_project_dependencies anakin_shared anakin_saber)
cmake/external/any.cmake 已删除 100644 → 0
浏览文件 @ 8bda4ab2
INCLUDE(ExternalProject)
SET(ANY_SOURCE_DIR ${THIRD_PARTY_PATH}/any)
INCLUDE_DIRECTORIES(${ANY_SOURCE_DIR}/src/extern_lib_any)
ExternalProject_Add(
extern_lib_any
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/PaddlePaddle/any.git"
GIT_TAG "15595d8324be9e8a9a80d9ae442fdd12bd66df5d"
PREFIX ${ANY_SOURCE_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
INSTALL_COMMAND ""
TEST_COMMAND ""
)
if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/lib_any_dummy.c)
file(WRITE ${dummyfile} "const char * dummy_any = \"${dummyfile}\";")
add_library(lib_any STATIC ${dummyfile})
else()
add_library(lib_any INTERFACE)
endif()
add_dependencies(lib_any extern_lib_any)
add_definitions(-DANY_IMPL_ANY_CAST_MOVEABLE)
LIST(APPEND external_project_dependencies lib_any)
cmake/external/boost.cmake
浏览文件 @ b8491bfd
......@@ -57,5 +57,4 @@ else()
endif()
add_dependencies(boost ${BOOST_PROJECT})
list(APPEND external_project_dependencies boost)
set(Boost_INCLUDE_DIR ${BOOST_INCLUDE_DIR})
cmake/external/brpc.cmake
浏览文件 @ b8491bfd
......@@ -69,5 +69,3 @@ SET_PROPERTY(TARGET brpc PROPERTY IMPORTED_LOCATION ${BRPC_LIBRARIES})
ADD_DEPENDENCIES(brpc extern_brpc)
add_definitions(-DBRPC_WITH_GLOG)
LIST(APPEND external_project_dependencies brpc)
cmake/external/cub.cmake
浏览文件 @ b8491bfd
......@@ -31,5 +31,3 @@ else()
endif()
add_dependencies(cub extern_cub)
LIST(APPEND external_project_dependencies cub)
cmake/external/dlpack.cmake
浏览文件 @ b8491bfd
......@@ -27,5 +27,3 @@ else()
endif()
add_dependencies(dlpack extern_dlpack)
LIST(APPEND external_project_dependencies dlpack)
cmake/external/eigen.cmake
浏览文件 @ b8491bfd
......@@ -52,5 +52,3 @@ else()
endif()
add_dependencies(eigen3 extern_eigen3)
LIST(APPEND external_project_dependencies eigen3)
cmake/external/gflags.cmake
浏览文件 @ b8491bfd
......@@ -61,8 +61,6 @@ ADD_LIBRARY(gflags STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET gflags PROPERTY IMPORTED_LOCATION ${GFLAGS_LIBRARIES})
ADD_DEPENDENCIES(gflags extern_gflags)
LIST(APPEND external_project_dependencies gflags)
# On Windows (including MinGW), the Shlwapi library is used by gflags if available.
if (WIN32)
include(CheckIncludeFileCXX)
......
cmake/external/glog.cmake
浏览文件 @ b8491bfd
......@@ -72,5 +72,3 @@ ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES})
ADD_DEPENDENCIES(glog extern_glog gflags)
LINK_LIBRARIES(glog gflags)
LIST(APPEND external_project_dependencies glog)
cmake/external/gtest.cmake
浏览文件 @ b8491bfd
......@@ -79,5 +79,4 @@ IF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
SET_PROPERTY(TARGET gtest_main PROPERTY IMPORTED_LOCATION ${GTEST_MAIN_LIBRARIES})
ADD_DEPENDENCIES(gtest_main extern_gtest)
LIST(APPEND external_project_dependencies gtest gtest_main)
ENDIF(WITH_TESTING OR (WITH_DISTRIBUTE AND NOT WITH_GRPC))
cmake/external/leveldb.cmake
浏览文件 @ b8491bfd
......@@ -39,6 +39,3 @@ ADD_DEPENDENCIES(extern_leveldb snappy)
ADD_LIBRARY(leveldb STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET leveldb PROPERTY IMPORTED_LOCATION ${LEVELDB_LIBRARIES})
ADD_DEPENDENCIES(leveldb extern_leveldb)
LIST(APPEND external_project_dependencies leveldb)
cmake/external/libmct.cmake
浏览文件 @ b8491bfd
......@@ -72,7 +72,4 @@ else()
add_library(libmct INTERFACE)
endif()
#ADD_LIBRARY(libmct SHARED IMPORTED GLOBAL)
ADD_DEPENDENCIES(libmct ${LIBMCT_PROJECT})
LIST(APPEND external_project_dependencies libmct)
cmake/external/libxsmm.cmake
浏览文件 @ b8491bfd
......@@ -53,5 +53,3 @@ MESSAGE(STATUS "Libxsmm library: ${LIBXSMM_LIBS}")
include_directories(${LIBXSMM_INCLUDE_DIR})
ADD_DEFINITIONS(-DPADDLE_WITH_LIBXSMM)
ADD_DEPENDENCIES(libxsmm extern_libxsmm)
LIST(APPEND external_project_dependencies libxsmm)
cmake/external/mkldnn.cmake
浏览文件 @ b8491bfd
......@@ -89,7 +89,6 @@ SET_PROPERTY(TARGET shared_mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
ADD_DEPENDENCIES(shared_mkldnn ${MKLDNN_PROJECT})
MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
add_definitions(-DPADDLE_WITH_MKLDNN)
LIST(APPEND external_project_dependencies shared_mkldnn)
# generate a static dummy target to track mkldnn dependencies
# for cc_library(xxx SRCS xxx.c DEPS mkldnn)
......
cmake/external/mklml.cmake
浏览文件 @ b8491bfd
......@@ -73,4 +73,3 @@ INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
ADD_LIBRARY(mklml SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET mklml PROPERTY IMPORTED_LOCATION ${MKLML_LIB})
ADD_DEPENDENCIES(mklml ${MKLML_PROJECT})
LIST(APPEND external_project_dependencies mklml)
cmake/external/ngraph.cmake
浏览文件 @ b8491bfd
......@@ -77,4 +77,3 @@ add_dependencies(ngraph ${NGRAPH_PROJECT})
target_compile_definitions(ngraph INTERFACE -DPADDLE_WITH_NGRAPH)
target_include_directories(ngraph INTERFACE ${NGRAPH_INC_DIR})
target_link_libraries(ngraph INTERFACE ${NGRAPH_SHARED_LIB})
LIST(APPEND external_project_dependencies ngraph)
cmake/external/openblas.cmake
浏览文件 @ b8491bfd
......@@ -11,11 +11,6 @@
# 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.
IF(USE_EIGEN_FOR_BLAS)
return()
ENDIF(USE_EIGEN_FOR_BLAS)
INCLUDE(cblas)
IF(NOT ${CBLAS_FOUND})
......@@ -91,7 +86,6 @@ ENDIF()
IF(NOT ${CBLAS_FOUND})
ADD_DEPENDENCIES(cblas extern_openblas)
LIST(APPEND external_project_dependencies cblas)
ELSE()
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
ADD_DEPENDENCIES(cblas mklml)
......
cmake/external/protobuf.cmake
浏览文件 @ b8491bfd
......@@ -129,7 +129,6 @@ macro(PROMPT_PROTOBUF_LIB)
ADD_DEPENDENCIES(protoc ${dep})
ENDFOREACH()
LIST(APPEND external_project_dependencies protobuf)
RETURN()
endmacro()
macro(SET_PROTOBUF_VERSION)
......@@ -231,7 +230,7 @@ FUNCTION(build_protobuf TARGET_NAME BUILD_FOR_HOST)
)
ENDFUNCTION()
SET(PROTOBUF_VERSION 3.1)
SET(PROTOBUF_VERSION 3.1.0)
IF(NOT PROTOBUF_FOUND)
build_protobuf(extern_protobuf FALSE)
......
cmake/external/pslib.cmake
浏览文件 @ b8491bfd
......@@ -70,4 +70,3 @@ ExternalProject_Add(
ADD_LIBRARY(pslib SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET pslib PROPERTY IMPORTED_LOCATION ${PSLIB_LIB})
ADD_DEPENDENCIES(pslib ${PSLIB_PROJECT})
LIST(APPEND external_project_dependencies pslib)
cmake/external/pslib_brpc.cmake
浏览文件 @ b8491bfd
......@@ -70,4 +70,3 @@ ExternalProject_Add(
ADD_LIBRARY(pslib_brpc SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET pslib_brpc PROPERTY IMPORTED_LOCATION ${PSLIB_BRPC_LIB})
ADD_DEPENDENCIES(pslib_brpc ${PSLIB_BRPC_PROJECT})
LIST(APPEND external_project_dependencies pslib_brpc)
cmake/external/threadpool.cmake
浏览文件 @ b8491bfd
......@@ -26,5 +26,3 @@ else()
endif()
add_dependencies(simple_threadpool extern_threadpool)
LIST(APPEND external_project_dependencies simple_threadpool)
cmake/external/warpctc.cmake
浏览文件 @ b8491bfd
......@@ -83,5 +83,3 @@ INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include wa
ADD_LIBRARY(warpctc SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET warpctc PROPERTY IMPORTED_LOCATION ${WARPCTC_LIBRARIES})
ADD_DEPENDENCIES(warpctc extern_warpctc)
LIST(APPEND external_project_dependencies warpctc)
cmake/external/xbyak.cmake
浏览文件 @ b8491bfd
......@@ -55,4 +55,3 @@ else()
endif()
add_dependencies(xbyak ${XBYAK_PROJECT})
list(APPEND external_project_dependencies xbyak)
cmake/external/xxhash.cmake
浏览文件 @ b8491bfd
......@@ -71,5 +71,3 @@ add_library(xxhash STATIC IMPORTED GLOBAL)
set_property(TARGET xxhash PROPERTY IMPORTED_LOCATION ${XXHASH_LIBRARIES})
include_directories(${XXHASH_INCLUDE_DIR})
add_dependencies(xxhash extern_xxhash)
LIST(APPEND external_project_dependencies xxhash)
cmake/external/zlib.cmake
浏览文件 @ b8491bfd
......@@ -57,5 +57,3 @@ ENDIF(WIN32)
ADD_LIBRARY(zlib STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET zlib PROPERTY IMPORTED_LOCATION ${ZLIB_LIBRARIES})
ADD_DEPENDENCIES(zlib extern_zlib)
LIST(APPEND external_project_dependencies zlib)
cmake/flags.cmake
浏览文件 @ b8491bfd
......@@ -21,7 +21,7 @@ function(CheckCompilerCXX11Flag)
if (${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 3.3)
message(FATAL_ERROR "Unsupported Clang version. Clang >= 3.3 required.")
endif()
endif()
endif()
endif()
endfunction()
......@@ -147,6 +147,7 @@ set(GPU_COMMON_FLAGS
-Wno-error=unused-function # Warnings in Numpy Header.
-Wno-error=array-bounds # Warnings in Eigen::array
)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m64")
endif(NOT WIN32)
if (APPLE)
......
cmake/hip.cmake
浏览文件 @ b8491bfd
......@@ -11,8 +11,6 @@ include_directories("/opt/rocm/rocrand/include")
include_directories("/opt/rocm/rccl/include")
include_directories("/opt/rocm/thrust")
list(APPEND EXTERNAL_LIBS "-L/opt/rocm/lib/ -lhip_hcc")
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -fPIC -DPADDLE_WITH_HIP -std=c++11" )
if(WITH_DSO)
......@@ -31,22 +29,12 @@ if(WITH_GRPC)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_GRPC")
endif(WITH_GRPC)
if(NOT WITH_GOLANG)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITHOUT_GOLANG")
endif(NOT WITH_GOLANG)
if(WITH_MKLDNN)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_WITH_MKLDNN")
endif(WITH_MKLDNN)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DANY_IMPL_ANY_CAST_MOVEABLE")
if(NOT WITH_RDMA)
set(HIP_HCC_FLAGS "${HIP_HCC_FLAGS} -DPADDLE_DISABLE_RDMA")
endif(NOT WITH_RDMA)
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
list(APPEND HIP_HCC_FLAGS ${CMAKE_CXX_FLAGS_DEBUG})
elseif(CMAKE_BUILD_TYPE STREQUAL "RelWithDebInfo")
......
cmake/rdma.cmake 已删除 100644 → 0
浏览文件 @ 8bda4ab2
# user should download rdma first from subversion repository
# execute following instruction to download svn mannally
# svn co https://svn.baidu.com/sys/ip/trunk/rdma/sockrdmav1 rdma/
# svn co https://svn.baidu.com/sys/ip/trunk/rdma/thirdparty rdma/
# we use static output in svn repositories to avoid implict bugs from not standard runtime env.
if(WITH_RDMA)
set(RDMA_ROOT $ENV{RDMA_ROOT} CACHE PATH "Folder contains RDMA sock library and thirdparty library")
function(generate_rdma_links)
#redirect to current DIR to isolate the pollution from system runtime environment
#it can benifits unified control for different gcc environment.
#e.g, by default gcc48 did not refer /usr/lib64 which could contain low version
#runtime libraries that will crash process while loading it. That redirect trick
#can fix it.
execute_process(
COMMAND mkdir -p librdma
COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so.1
COMMAND ln -s -f /usr/lib64/libibverbs.so.1.0.0 librdma/libibverbs.so
COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so.1
COMMAND ln -s -f /usr/lib64/librdmacm.so.1.0.0 librdma/librdmacm.so
COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so.1
COMMAND ln -s -f /lib64/libnl.so.1.1.4 librdma/libnl.so
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
)
endfunction(generate_rdma_links)
#check and set headers
find_path(RDMA_INC_SXISOCK sxi_sock.h PATHS ${RDMA_ROOT}/sockrdmav1/output/include)
find_path(RDMA_INC_XIO libxio.h PATHS ${RDMA_ROOT}/thirdparty/output/accelio)
find_path(RDMA_INC_EVENT event2 PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_path(RDMA_INC_NUMA numa.h PATHS ${RDMA_ROOT}/thirdparty/output/libnuma)
#check and set libs
find_library(RDMA_LIB_SXISOCK NAMES sxisock PATHS ${RDMA_ROOT}/sockrdmav1/output)
find_library(RDMA_LIB_XIO NAMES xio PATHS ${RDMA_ROOT}/thirdparty/output/accelio)
find_library(RDMA_LIB_EVENT NAMES event PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_library(RDMA_LIB_EVENT_CORE NAMES event_core PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_library(RDMA_LIB_EVENT_EXTRA NAMES event_extra PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_library(RDMA_LIB_EVENT_PTHREADS NAMES event_pthreads PATHS ${RDMA_ROOT}/thirdparty/output/libevent)
find_library(RDMA_LIB_NUMA NAMES numa PATHS ${RDMA_ROOT}/thirdparty/output/libnuma)
if(
RDMA_INC_SXISOCK AND
RDMA_INC_XIO AND
RDMA_INC_EVENT AND
RDMA_INC_NUMA AND
RDMA_LIB_SXISOCK AND
RDMA_LIB_XIO AND
RDMA_LIB_EVENT AND
RDMA_LIB_EVENT_CORE AND
RDMA_LIB_EVENT_EXTRA AND
RDMA_LIB_EVENT_PTHREADS AND
RDMA_LIB_NUMA
)
set(RDMA_INC_DIR
${RDMA_INC_SXISOCK}
${RDMA_INC_XIO}
${RDMA_INC_EVENT}
${RDMA_INC_NUMA})
set(RDMA_LIBS
${RDMA_LIB_SXISOCK}
${RDMA_LIB_XIO}
${RDMA_LIB_EVENT}
${RDMA_LIB_EVENT_CORE}
${RDMA_LIB_EVENT_EXTRA}
${RDMA_LIB_EVENT_PTHREADS}
${RDMA_LIB_NUMA}
)
set(RDMA_LD_FLAGS "-L./librdma -libverbs -lrdmacm -Xlinker -rpath ./librdma")
include_directories("${RDMA_INC_DIR}")
else()
#if this module is not called, RDMA_INC_DIR RDMA_LIBS will be null, so top module always refer this variable
message(FATAL_ERROR, "RDMA libraries are not found, try to set RDMA_ROOT or check all related libraries.")
endif()
else(WITH_RDMA)
set(RDMA_LIBS "")
set(RDMA_LD_FLAGS "")
add_definitions(-DPADDLE_DISABLE_RDMA)
endif(WITH_RDMA)
cmake/tensorrt.cmake
浏览文件 @ b8491bfd
......@@ -33,6 +33,5 @@ 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})
list(APPEND EXTERNAL_LIBS ${TENSORRT_LIBRARY})
add_definitions(-DPADDLE_WITH_TENSORRT)
endif()
paddle/contrib/float16/run_float16_demo.sh
浏览文件 @ b8491bfd
......@@ -14,9 +14,7 @@ cmake .. -DWITH_AVX=OFF \
-DWITH_MKL=OFF \
-DWITH_GPU=ON \
-DWITH_TESTING=ON \
-DWITH_TIMER=ON \
-DWITH_PROFILER=ON \
-DWITH_FLUID_ONLY=ON
make -j `nproc`
pip install -U "$WHEEL_PATH/$(ls $WHEEL_PATH)"
......
paddle/fluid/API.spec
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ b8491bfd
......@@ -158,18 +158,19 @@ cc_library(variable_helper SRCS variable_helper.cc DEPS lod_tensor)
cc_library(naive_executor SRCS naive_executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
if(WITH_DISTRIBUTE)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog
lod_rank_table feed_fetch_method sendrecvop_rpc ${GLOB_DISTRIBUTE_DEPS} graph_to_program_pass variable_helper)
if(WITH_NGRAPH)
set(NGRAPH_EXE_DEPS ngraph_engine)
else()
set(NGRAPH_EXE_DEPS)
endif()
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
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})
set(DISTRIBUTE_COMPILE_FLAGS "-Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
set_source_files_properties(executor.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
else()
if (WITH_NGRAPH)
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper ngraph_engine)
else ()
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper)
endif()
cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method graph_to_program_pass variable_helper ${NGRAPH_EXE_DEPS})
cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
endif()
......
paddle/fluid/framework/async_executor.cc
浏览文件 @ b8491bfd
......@@ -244,6 +244,7 @@ void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
auto& block = main_program.Block(0);
for (auto var_name : fetch_var_names) {
auto var_desc = block.FindVar(var_name);
PADDLE_ENFORCE_NOT_NULL(var_desc, "%s is not found.", var_name);
auto shapes = var_desc->GetShape();
PADDLE_ENFORCE(shapes[shapes.size() - 1] == 1,
"var %s: Fetched var has wrong shape, "
......
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @ b8491bfd
......@@ -134,11 +134,6 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
out_layout =
out_layout == DataLayout::kAnyLayout ? DataLayout::kNCHW : out_layout;
auto& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx = dynamic_cast<platform::MKLDNNDeviceContext*>(
pool.Get(expected_kernel_type.place_));
auto& cpu_engine = dev_ctx->GetEngine();
std::vector<int> in_tz = paddle::framework::vectorize2int(in.dims());
std::vector<int> out_tz = in_tz;
......@@ -147,29 +142,25 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
"Input tensor type is not supported: %s", in.type());
memory::data_type out_type = in_type;
auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
auto out_format =
platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
// output tensor has the same dims as input. Reorder don't change dims
out->Resize(in.dims());
if (in_format != out_format) {
// tempory mem pd fr out , to make reorder
auto out_mem_pd = paddle::platform::create_prim_desc_from_dims(
paddle::framework::vectorize2int(out->dims()),
mkldnn::memory::format::blocked, out_type);
if (in.get_mkldnn_prim_desc() != out_mem_pd) {
void* in_data = GetDataFromTensor(in, in_type);
auto out_data = out->mutable_data(expected_kernel_type.place_, in.type());
auto in_memory =
memory({{{in_tz}, in_type, in_format}, cpu_engine}, in_data);
auto out_memory =
memory({{{out_tz}, out_type, out_format}, cpu_engine}, out_data);
auto in_memory = memory(in.get_mkldnn_prim_desc(), in_data);
auto out_memory = memory(out_mem_pd, out_data);
platform::Reorder(in_memory, out_memory);
} else {
out->ShareDataWith(in);
}
out->set_layout(out_layout);
// reset format since the out tensor will be feed to non-MKLDNN OPkernel
out->set_format(memory::format::format_undef);
#endif
}
......
paddle/fluid/framework/data_transform.cc
浏览文件 @ b8491bfd
......@@ -51,13 +51,31 @@ void TransformData(const OpKernelType &expected_kernel_type,
#ifdef PADDLE_WITH_MKLDNN
// Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
// Just set layout/format. No real transform occur
auto out_format = platform::MKLDNNFormatForSize(in.dims().size(),
ToMKLDNNFormat(lin));
out.ShareDataWith(input_tensor);
out.set_layout(DataLayout::kMKLDNN);
out.set_format(out_format);
// TODO(jczaja): Remove that once all mkldnn ops
// are modified to work with mkldnn_blocked
auto mkldnn_fmt = [&](int rank) {
switch (rank) {
case 5:
return mkldnn::memory::format::ncdhw;
case 4:
return mkldnn::memory::format::nchw;
case 3:
return mkldnn::memory::format::ncw;
case 2:
return mkldnn::memory::format::nc;
case 1:
return mkldnn::memory::format::x;
default:
return mkldnn::memory::format::blocked;
}
};
auto out_mem_pd = paddle::platform::create_prim_desc_from_dims(
paddle::framework::vectorize2int(out.dims()),
mkldnn_fmt(out.dims().size()));
out.set_mkldnn_prim_desc(out_mem_pd);
#endif
} else {
// Case2 - transfrom from MKLDNN OPKernel to Non-MKLDNN OPKernel
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ b8491bfd
......@@ -50,12 +50,15 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_
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)
cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper)
if(WITH_GPU)
cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
else()
cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info)
endif()
cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass)
cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info)
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(memory_early_delete_pass SRCS memory_early_delete_pass.cc DEPS memory_optimize_pass computation_op_handle scale_loss_grad_op_handle rpc_op_handle
all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle graph graph_helper pass)
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)
......@@ -67,13 +70,11 @@ cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_he
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle fused_broadcast_op_handle)
set(SSA_GRAPH_EXECUTOR_DEPS graph framework_proto sequential_execution_pass modify_op_lock_and_record_event_pass all_reduce_deps_pass reference_count_pass eager_deletion_pass memory_optimize_pass memory_early_delete_pass inplace_op_pass)
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)
list(APPEND SSA_GRAPH_EXECUTOR_DEPS reference_count_pass)
endif()
cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph)
cc_test(memory_optimize_pass_test SRCS memory_optimize_pass_test.cc memory_optimize_pass.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry pass)
cc_test(memory_optimize_helper_test SRCS memory_optimize_helper_test.cc memory_optimize_helper.cc DEPS framework_proto graph graph_helper op_registry)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ${SSA_GRAPH_EXECUTOR_DEPS})
cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
......
paddle/fluid/framework/details/all_reduce_deps_pass.cc
浏览文件 @ b8491bfd
......@@ -30,8 +30,6 @@ namespace paddle {
namespace framework {
namespace details {
static constexpr char kAllOpDescs[] = "all_op_descs";
VarHandle* GetValidInput(const OpHandleBase* a) {
for (auto p : a->Inputs()) {
VarHandle* b = dynamic_cast<VarHandle*>(p);
......
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ b8491bfd
......@@ -53,7 +53,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
#endif
void AllReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
......
paddle/fluid/framework/details/broadcast_op_handle.cc
浏览文件 @ b8491bfd
......@@ -22,7 +22,7 @@ namespace framework {
namespace details {
void BroadcastOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
platform::RecordEvent record_event(Name());
if (places_.size() == 1) return;
......@@ -30,7 +30,7 @@ void BroadcastOpHandle::RunImpl() {
VarHandle *in_var_handle;
{
auto in_var_handles = DynamicCast<VarHandle>(inputs_);
PADDLE_ENFORCE_EQ(in_var_handles.size(), 1,
PADDLE_ENFORCE_EQ(in_var_handles.size(), 1UL,
"The number of input should be one.");
in_var_handle = in_var_handles[0];
}
......
paddle/fluid/framework/details/build_strategy.cc
浏览文件 @ b8491bfd
......@@ -34,9 +34,11 @@ namespace details {
static inline bool SeqOnlyAllReduceOps(const BuildStrategy &strategy) {
// Should fix the allreduce op order if scheduling
// them in multiple threads or processes to avoid hang.
// NOTE: ParallelGraph would execute this pass on each graph, so
// don't need to append it here.
return (!strategy.enable_sequential_execution_ &&
strategy.num_trainers_ > 1) ||
strategy.enable_parallel_graph_;
strategy.num_trainers_ > 1) &&
!strategy.enable_parallel_graph_;
}
class ParallelExecutorPassBuilder : public ir::PassBuilder {
......@@ -133,15 +135,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
void AppendMultiDevPass(const BuildStrategy &strategy) {
ir::Pass *multi_devices_pass;
if (strategy_.is_distribution_) {
VLOG(3) << "dist train mode";
VLOG(3) << "multi device parameter server mode";
multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
} else {
if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
VLOG(3) << "allreduce mode";
VLOG(3) << "multi devices collective mode with allreduce";
multi_devices_pass =
AppendPass("allreduce_mode_multi_devices_pass").get();
} else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
VLOG(3) << "reduce mode";
VLOG(3) << "multi deivces collective mode with reduce";
multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
} else {
PADDLE_THROW("Unknown reduce strategy.");
......@@ -211,8 +213,6 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
new std::vector<OpDesc *>(main_program.Block(0).AllOps());
graph->Set<const std::vector<OpDesc *>>(kAllOpDescs,
all_op_descs); // take ownership
graph->Set<GraphNodePool>(kGraphNodePool,
new GraphNodePool); // take ownership
pass->Erase(kAllOpDescs);
pass->SetNotOwned<const std::vector<OpDesc *>>(kAllOpDescs, all_op_descs);
......@@ -247,7 +247,9 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
continue;
}
}
VLOG(3) << "Start Apply Pass " << pass->Type();
graph = pass->Apply(std::move(graph));
VLOG(3) << "Finish Apply Pass " << pass->Type();
}
return graph;
}
......
paddle/fluid/framework/details/build_strategy.h
浏览文件 @ b8491bfd
......@@ -77,9 +77,6 @@ struct BuildStrategy {
bool fuse_relu_depthwise_conv_{false};
bool memory_optimize_{false};
bool memory_early_delete_{false};
// TODO(dzhwinter):
// make enable_inplace, memory_optimize_
// memory_early_delete_ true by default
......
paddle/fluid/framework/details/computation_op_handle.h
浏览文件 @ b8491bfd
......@@ -26,7 +26,7 @@
namespace paddle {
namespace framework {
namespace details {
struct ComputationOpHandle : public OpHandleBase {
class ComputationOpHandle : public OpHandleBase {
public:
ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,
size_t scope_idx);
......
paddle/fluid/framework/details/data_balance_op_handle.cc
浏览文件 @ b8491bfd
......@@ -86,7 +86,7 @@ std::vector<std::array<int, 3>> DataBalanceOpHandle::GetBalancePlan(
}
void DataBalanceOpHandle::RunImpl() {
PADDLE_ENFORCE_GT(places_.size(), 1,
PADDLE_ENFORCE_GT(places_.size(), 1UL,
"Data balance can only be enabled when the number of "
"places to run larger than 1.");
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
......
paddle/fluid/framework/details/fuse_vars_op_handle.cc
浏览文件 @ b8491bfd
......@@ -23,7 +23,7 @@ void FuseVarsOpHandle::RunImpl() {
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(in_var_handles.size(), 0);
PADDLE_ENFORCE_EQ(in_var_handles.size(), 0UL);
PADDLE_ENFORCE_EQ(out_var_handles.size() - 1, inputs_numel_.size(), "");
auto scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
......
paddle/fluid/framework/details/fused_broadcast_op_handle.cc
浏览文件 @ b8491bfd
......@@ -22,7 +22,7 @@ namespace framework {
namespace details {
void FusedBroadcastOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
platform::RecordEvent record_event(Name());
if (places_.size() == 1UL) return;
......
paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc
浏览文件 @ b8491bfd
......@@ -34,8 +34,8 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
->Var(details::kLocalExecScopeName)
->GetMutable<Scope*>() = &local_scope;
for (size_t j = 0; j < input_scope_idxes.size(); ++j) {
local_scope.Var("out_var" + j);
if (i == j) local_scope.Var("in_var" + j);
local_scope.Var("out_var" + std::to_string(j));
if (i == j) local_scope.Var("in_var" + std::to_string(j));
}
param_scopes_.emplace_back(&local_scope);
}
......@@ -62,20 +62,21 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
// add input var handle
nodes_.emplace_back(
ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable));
VarHandle* in_var_handle =
new VarHandle(nodes_.back().get(), 1, input_scope_idxes[i],
"in_var" + i, place_list_[input_scope_idxes[i]]);
nodes_.emplace_back(ir::CreateNodeForTest("in_node" + std::to_string(i),
ir::Node::Type::kVariable));
VarHandle* in_var_handle = new VarHandle(
nodes_.back().get(), 1, input_scope_idxes[i],
"in_var" + std::to_string(i), place_list_[input_scope_idxes[i]]);
vars_.emplace_back(in_var_handle);
op_handle_->AddInput(in_var_handle);
// add output var handle
for (size_t j = 0; j < place_list_.size(); ++j) {
nodes_.emplace_back(
ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable));
VarHandle* out_var_handle = new VarHandle(
nodes_.back().get(), 2, j, "out_var" + i, place_list_[j]);
nodes_.emplace_back(ir::CreateNodeForTest(
"out_node" + std::to_string(i), ir::Node::Type::kVariable));
VarHandle* out_var_handle =
new VarHandle(nodes_.back().get(), 2, j,
"out_var" + std::to_string(i), place_list_[j]);
vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle);
}
......@@ -86,7 +87,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
std::vector<std::vector<float>> send_vec;
f::LoD lod{{0, 10, 20}};
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string varname("in_var" + i);
const std::string varname("in_var" + std::to_string(i));
float val_scalar = static_cast<float>(i);
send_vec.push_back(
InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar));
......@@ -96,7 +97,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
WaitAll();
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string& varname("out_var" + i);
const std::string& varname("out_var" + std::to_string(i));
for (size_t j = 0; j < place_list_.size(); ++j) {
LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]);
}
......@@ -109,7 +110,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
2, 4, 6, 3, 1, 1, 1, 1, 3, 7};
int height = static_cast<int>(kDims[0] * 2);
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string varname("in_var" + i);
const std::string varname("in_var" + std::to_string(i));
float val_scalar = static_cast<float>(i);
send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i],
rows, height, val_scalar));
......@@ -119,7 +120,7 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
WaitAll();
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string& varname("out_var" + i);
const std::string& varname("out_var" + std::to_string(i));
for (size_t j = 0; j < place_list_.size(); ++j) {
SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows,
height);
......
paddle/fluid/framework/details/inplace_op_pass.cc
浏览文件 @ b8491bfd
......@@ -49,7 +49,7 @@ DEFINE_bool(
"If this option turns on, only these op in whitelist can be inplaced."
"If it turns off, all of the running op can be candidate of inplaced op."
"Such as scale, elementwise_add"
"By default, it's turned on");
"By default, it's turned off");
DECLARE_string(memory_optimize_debug);
......@@ -171,16 +171,15 @@ void InplacePass::InplaceModifyDesc(const std::string& var,
}
}
const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var,
const std::string& cache_var,
const size_t& idx,
ir::Graph* graph) const {
const NodeSwapQueue InplacePass::TryInplaceModifyVar(
const std::string& var, const std::string& cache_var, const size_t& idx,
ir::Graph* graph) const {
PADDLE_ENFORCE(var_nodes_[var].size() >= 1 &&
var_nodes_[var].at(0)->Var() != nullptr);
std::unique_ptr<VarDesc> var_desc(new VarDesc(*var_nodes_[var].at(0)->Var()));
var_desc->SetName(cache_var);
SSANodePair swap_nodes;
NodeSwapQueue swap_nodes;
for (size_t i = idx; i < view_.AllOps().size(); ++i) {
auto* op = view_.AllOps()[i];
......@@ -230,7 +229,7 @@ const SSANodePair InplacePass::TryInplaceModifyVar(const std::string& var,
return swap_nodes;
}
void InplacePass::CommitModify(const SSANodePair& swap_nodes,
void InplacePass::CommitModify(const NodeSwapQueue& swap_nodes,
ir::Graph* graph) const {
for (auto& pair : swap_nodes) {
auto *node = pair.first, *cache_node = pair.second;
......@@ -245,7 +244,7 @@ void InplacePass::CommitModify(const SSANodePair& swap_nodes,
}
}
void InplacePass::WithdrawModify(const SSANodePair& nodes,
void InplacePass::WithdrawModify(const NodeSwapQueue& nodes,
ir::Graph* graph) const {
for (auto& pair : nodes) {
auto *node = pair.first, *cache_node = pair.second;
......
paddle/fluid/framework/details/inplace_op_pass.h
浏览文件 @ b8491bfd
......@@ -56,7 +56,8 @@ class GraphView {
std::map<ir::Node*, std::unordered_set<ir::Node*>> adj_list_;
};
typedef std::vector<std::pair<ir::Node*, ir::Node*>> SSANodePair;
// swap pairs in sequence
typedef std::vector<std::pair<ir::Node*, ir::Node*>> NodeSwapQueue;
class InplacePass : public ir::Pass {
public:
InplacePass();
......@@ -68,14 +69,14 @@ class InplacePass : public ir::Pass {
void InitSSAGraphNodes() const;
private:
const SSANodePair TryInplaceModifyVar(const std::string& var,
const std::string& cache_var,
const size_t& idx,
ir::Graph* graph) const;
const NodeSwapQueue TryInplaceModifyVar(const std::string& var,
const std::string& cache_var,
const size_t& idx,
ir::Graph* graph) const;
void CommitModify(const SSANodePair&, ir::Graph* graph) const;
void CommitModify(const NodeSwapQueue&, ir::Graph* graph) const;
void WithdrawModify(const SSANodePair& nodes, ir::Graph* graph) const;
void WithdrawModify(const NodeSwapQueue& nodes, ir::Graph* graph) const;
void InplaceModifyDesc(const std::string& in_var, const std::string& out_var,
const size_t& idx) const;
......
paddle/fluid/framework/details/memory_early_delete_pass.cc 已删除 100644 → 0
浏览文件 @ 8bda4ab2
// 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/memory_early_delete_pass.h"
#include <queue>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/reference_count_pass_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
static ComputationOpHandle* FindNextComputationOpHandle(VarHandle* var_in) {
std::queue<VarHandleBase*> queue;
queue.push(var_in);
do {
auto* var = queue.front();
queue.pop();
for (auto* op : var->PendingOps()) {
auto* compute_op = dynamic_cast<ComputationOpHandle*>(op);
if (compute_op != nullptr && compute_op->GetPlace() == var_in->place()) {
return compute_op;
}
for (auto* out_var : op->Outputs()) {
queue.push(out_var);
}
}
} while (!queue.empty());
return nullptr;
}
std::unique_ptr<ir::Graph> MemoryEarlyDeletePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto& graph_pool = Get<GraphNodePool>(kGraphNodePool);
auto& gcs = Get<GarbageCollectorMap>(kGarbageCollector);
std::unordered_map<std::string, std::unordered_set<OpDesc*>> unlived_vars;
unlived_vars.reserve(graph_pool.size());
for (auto& pair : graph_pool) {
unlived_vars.insert(std::make_pair(pair.first, pair.second));
}
auto compare_and_insert_early_delete_op = [&](
OpHandleBase* op, const std::vector<VarHandleBase*>& vars) {
if (unlived_vars.empty()) return;
// unlived vars can be deleted after the last used op has finished.
auto* compute_op = dynamic_cast<ComputationOpHandle*>(op);
const auto& places = Get<std::vector<platform::Place>>(kAllPlaces);
for (auto& var : vars) {
auto* var_handle = dynamic_cast<VarHandle*>(var);
auto var_name = var->Node()->Name();
auto& var_place = var_handle->place();
if (unlived_vars.count(var_name) == 0) continue;
if (!unlived_vars[var_name].empty()) {
if (compute_op != nullptr &&
unlived_vars[var_name].count(compute_op->Node()->Op()) != 0) {
unlived_vars[var_name].erase(compute_op->Node()->Op());
}
continue;
}
if (var_handle == nullptr || !var_handle->Node()->IsVar() ||
var_handle->Node()->IsCtrlVar())
continue;
// shameless copyed from reference count pass.
if (compute_op == nullptr) {
// use next computation op scope
compute_op = FindNextComputationOpHandle(var_handle);
}
auto* early_delete_node =
graph->CreateEmptyNode("early_delete", ir::Node::Type::kOperation);
GarbageCollector* gc = gcs.at(places[compute_op->GetScopeIdx()]).get();
auto* early_delete_handle = new EarlyDeleteOpHandle(
early_delete_node, compute_op->GetScope(), var_place, {var_name}, gc);
if (compute_op->Outputs().empty()) {
auto* dep_var = new DummyVarHandle(graph->CreateControlDepVar());
compute_op->AddOutput(dep_var);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
}
early_delete_handle->AddInput(compute_op->Outputs().front());
VLOG(5) << "Add early delete op " << var_name << " to Operator"
<< compute_op->Name();
}
};
auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto& op : all_ops) {
compare_and_insert_early_delete_op(op, op->Inputs());
compare_and_insert_early_delete_op(op, op->Outputs());
}
return graph;
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(memory_early_delete_pass,
paddle::framework::details::MemoryEarlyDeletePass)
.RequireGraphAttr(paddle::framework::details::kGraphNodePool)
.RequireGraphAttr(paddle::framework::details::kGarbageCollector);
paddle/fluid/framework/details/memory_early_delete_pass.h 已删除 100644 → 0
浏览文件 @ 8bda4ab2
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/fluid/framework/details/early_delete_op_handle.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace details {
class MemoryEarlyDeletePass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/memory_optimize_helper.cc
浏览文件 @ b8491bfd
......@@ -13,17 +13,114 @@
// limitations under the License.
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include <algorithm>
#include <deque>
#include <functional>
#include <iostream>
#include <iterator>
#include <numeric>
#include <sstream>
#include <string>
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/platform/cpu_info.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/gpu_info.h"
#endif // PADDLE_WITH_CUDA
namespace paddle {
namespace framework {
namespace details {
using paddle::framework::VarDesc;
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph) {
PADDLE_ENFORCE(graph.Has(kAllOpDescs),
"Graph has no attribute of kAllOpDescs.");
// 1. get op desc order
auto& op_descs = graph.Get<const std::vector<OpDesc*>>(kAllOpDescs);
// 2. topology sort order
auto nodes = graph.Nodes();
std::deque<ir::Node*> ops;
FilterVariables(nodes, [&](ir::Node* op) {
if (op->IsOp() && op->Op() != nullptr) {
ops.emplace_back(op);
}
});
std::unordered_map<ir::Node*, size_t> op_deps;
std::list<ir::Node*> ready_ops;
std::unordered_map<ir::Node*, std::unordered_set<ir::Node*>> pending_ops;
for (auto* op : ops) {
std::unordered_set<ir::Node*> preceding_op;
for (auto* in : op->inputs) {
if (in->inputs.empty()) continue;
PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp());
preceding_op.emplace(in->inputs[0]);
pending_ops[in->inputs[0]].emplace(op);
}
op_deps[op] = preceding_op.size();
if (preceding_op.empty()) {
ready_ops.emplace_back(op);
}
}
// 3. generated op list based desc order and the topology order
std::vector<ir::Node*> ret;
std::list<OpDesc*> op_descs_list(op_descs.begin(), op_descs.end());
auto update_by_found_node = [&](ir::Node* found_node) {
for (auto* pending_op : pending_ops[found_node]) {
if (--op_deps[pending_op] == 0) {
ready_ops.emplace_back(pending_op);
}
}
ready_ops.remove(found_node);
ret.emplace_back(found_node);
};
while (!ready_ops.empty()) {
bool all_of_ready_op_unmatched = true;
for (auto it = op_descs_list.begin(); it != op_descs_list.end();) {
auto op_desc = *it;
ir::Node* found_node = nullptr;
for (auto* op : ready_ops) {
if (IsSameDesc(op->Op(), op_desc)) {
found_node = op;
break;
}
}
// 3.1 op desc deleted by other pass
if (found_node == nullptr) {
++it;
continue;
} else {
all_of_ready_op_unmatched = false;
it = op_descs_list.erase(it);
}
update_by_found_node(found_node);
}
// 3.2 op descs are added by other pass
// preceding op non empty means some new op descs are
// created, but not contained in return node list.
// these new op desc may depend on each other.
std::list<ir::Node*> prev_ready_ops(ready_ops);
if (all_of_ready_op_unmatched) {
for (auto op : prev_ready_ops) {
update_by_found_node(op);
}
}
}
PADDLE_ENFORCE(std::all_of(
op_deps.begin(), op_deps.end(),
[&](const std::pair<ir::Node*, size_t>& p) { return p.second == 0; }));
return ret;
}
size_t NodeSizeInBytes(const VarDesc& node) {
size_t NodeSize(const VarDesc& node) {
auto shape = node.GetShape();
int size =
std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
......@@ -31,9 +128,15 @@ size_t NodeSizeInBytes(const VarDesc& node) {
return type_size * std::abs(size);
}
size_t NodeSizeInBytes(ir::Node* n) {
auto* desc = FindVarDescInBlock(n);
return NodeSizeInBytes(*desc);
size_t NodeSize(ir::Node* n) {
VarDesc* desc = nullptr;
// some op do not have block pointer
if (n->inputs[0]->Op() != nullptr) {
desc = FindVarDescInBlock(n);
} else {
desc = n->Var();
}
return NodeSize(*desc);
}
std::string DebugStringImpl(VarDesc* var) {
......@@ -59,7 +162,6 @@ std::string DebugStringImpl(VarDesc* var) {
std::string DebugString(ir::Node* var) {
return DebugStringImpl(FindVarDescInBlock(var));
}
// return DebugString(var->Var()); }
// NOTE(dzh): based ir node, if a large node has been reused
// by a small size node, then next time it appear in pool, it will
......@@ -76,22 +178,26 @@ struct NodeComparator {
bool operator()(ir::Node* lhs, ir::Node* rhs) const {
auto* lhs_desc = FindVarDescInBlock(lhs);
auto* rhs_desc = FindVarDescInBlock(rhs);
// match data type
if (lhs_desc->GetDataType() != rhs_desc->GetDataType()) {
return false;
}
// match shape
auto lhs_shape = lhs_desc->GetShape();
auto rhs_shape = rhs_desc->GetShape();
if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) ||
(lhs_shape[0] != -1 && rhs_shape[0] != -1)) {
return NodeSizeInBytes(lhs) <= NodeSizeInBytes(rhs);
return NodeSize(lhs) <= NodeSize(rhs);
} else {
return false;
}
}
};
void OrderedNodeList::Insert(ir::Node* var, ir::Node* op) {
void OrderedSet::Insert(ir::Node* var) {
PADDLE_ENFORCE(var->IsVar() && !var->IsCtrlVar());
PADDLE_ENFORCE(op->IsOp());
if (mark_table_.count(var->Name()) != 0) {
mark_table_[var->Name()]->second.insert(op);
mark_table_[var->Name()]->emplace_back(var);
return;
}
......@@ -99,14 +205,15 @@ void OrderedNodeList::Insert(ir::Node* var, ir::Node* op) {
auto var_shape = var_desc->GetShape();
int batch_size = static_cast<int>(var_shape[0]);
NodeComparator compare_node;
NodeComparator functor;
Iter it = nodes_.begin();
while (it != nodes_.end()) {
auto* cache_desc = FindVarDescInBlock(it->first);
auto& prev = it->front();
auto* cache_desc = FindVarDescInBlock(prev);
int cache_batch_size = cache_desc->GetShape()[0];
if ((cache_batch_size == -1 && batch_size == -1) ||
(cache_batch_size != -1 && batch_size != -1)) {
if (compare_node(it->first, var)) {
if (functor(prev, var)) {
++it;
} else {
break;
......@@ -118,62 +225,127 @@ void OrderedNodeList::Insert(ir::Node* var, ir::Node* op) {
}
}
it =
nodes_.insert(it, std::make_pair(var, std::unordered_set<ir::Node*>{op}));
it = nodes_.insert(it, {var});
mark_table_[var->Name()] = it;
}
int OrderedNodeList::GetIndex(ir::Node* var) {
int OrderedSet::GetNodeIndexInPool(ir::Node* var) {
return std::distance(nodes_.begin(), mark_table_[var->Name()]);
}
ir::Node* OrderedNodeList::NodeMatch(ir::Node* var) const {
ir::Node* OrderedSet::FindBestFitNode(ir::Node* var) const {
ir::Node* found_node = nullptr;
NodeComparator compare_node;
NodeComparator functor;
for (auto it = nodes_.begin(); it != nodes_.end(); ++it) {
if (compare_node(var, it->first)) {
found_node = it->first;
auto& candidate = it->front();
if (functor(var, candidate)) {
found_node = candidate;
break;
}
}
return found_node;
}
void OrderedNodeList::Erase(ir::Node* var) { Erase(var->Name()); }
ir::Node* OrderedSet::FindNextBestFitNode(ir::Node* var, ir::Node* prev) const {
ir::Node* found_node = nullptr;
NodeComparator functor;
auto it =
std::find_if(nodes_.begin(), nodes_.end(), [&](const NodeVector& v) {
if (v.front() == prev)
return true;
else
return false;
});
PADDLE_ENFORCE(it != nodes_.end(), "Not found previous in node list!");
for (it = std::next(it); it != nodes_.end(); ++it) {
auto& candidate = it->front();
if (functor(var, candidate)) {
found_node = candidate;
break;
}
}
return found_node;
}
void OrderedNodeList::Erase(const std::string& var) {
bool OrderedSet::Has(ir::Node* var) const {
if (mark_table_.count(var->Name())) {
auto& node_in_samename = mark_table_.at(var->Name());
auto iter =
std::find_if(node_in_samename->begin(), node_in_samename->end(),
[&](ir::Node* n) { return n->Name() == var->Name(); });
return iter != node_in_samename->end();
}
return false;
}
void OrderedSet::Erase(const std::string& var) {
PADDLE_ENFORCE(mark_table_.count(var));
nodes_.erase(mark_table_[var]);
mark_table_.erase(var);
}
std::string OrderedNodeList::ToString() const {
void OrderedSet::Erase(ir::Node* var) {
PADDLE_ENFORCE(var != nullptr);
Erase(var->Name());
}
std::string OrderedSet::ToString() const {
std::stringstream ss;
for (auto it = nodes_.begin(); it != nodes_.end(); ++it) {
ss << DebugString(it->first) << " ";
for (auto& node : *it) {
ss << DebugString(node) << " ";
}
}
return ss.str();
}
bool NodeCanReused(ir::Node* node) {
// valid the node is a var node
if (node == nullptr || !node->IsVar() || node->IsCtrlVar()) return false;
// auto* desc = node->Var();
bool flag = NodeCanReused(*node->Var());
bool flag = true;
// op output force generated in cpu, can not be reused.
for (auto* op : node->inputs) {
if (op->Op()->HasAttr("force_cpu")) {
// op output force generated in cpu, can not be reused.
flag &= framework::AttrReader(op->Op()->GetAttrMap())
.Get<bool>("force_cpu") == 0;
}
}
// var desc validation.
flag &= NodeCanReused(*node->Var());
return flag;
}
int MinChunkSize() {
int size{0};
#ifdef PADDLE_WITH_CUDA
size = platform::GpuMinChunkSize();
#else
size = platform::CpuMinChunkSize();
#endif // PADDLE_WITH_CUDA
return size;
}
bool NodeCanReused(const VarDesc& node) {
auto type = node.GetType();
if (node.Persistable() || type != proto::VarType::LOD_TENSOR ||
node.GetShape().empty()) {
// only these types holds bulk of gpu memory
if (!(type == proto::VarType::LOD_TENSOR ||
type == proto::VarType::SELECTED_ROWS ||
type == proto::VarType::LOD_TENSOR_ARRAY)) {
return false;
}
// persistable variable is parameter
if (node.Persistable()) {
return false;
}
// shape < min_chunk_size is meaningless.
// further more, fetched loss always has size = 1
// which should not be reused.
auto shape = node.GetShape();
int size = std::abs(
std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>()));
if (shape.empty() || size < MinChunkSize()) {
return false;
}
// vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
......@@ -193,6 +365,176 @@ bool OpHasSubBlock(OpDesc* desc) {
return false;
}
ControlFlowGraph::ControlFlowGraph(const ir::Graph& graph) {
ops_ = SortOpLikeDescOrder(graph);
ConnectNodes();
}
void ControlFlowGraph::BuildCFGGraph() {
// FIXME(dzh): same effect with ConnectNodes, but use the control
// link to build dependency graph, it goes wrong in transformer.
for (ir::Node* op : ops_) {
for (auto& input_var : op->inputs) {
if (!input_var->inputs.empty()) {
PADDLE_ENFORCE(
input_var->inputs.size() == 1 && input_var->inputs[0]->IsOp(),
"Preceding Op Node of Var Node must be unique");
auto* pred_op = input_var->inputs[0];
if (pred_op->Op() != nullptr) {
predecessors_[op].insert(pred_op);
successors_[pred_op].insert(op);
}
}
if (input_var->IsVar() && !input_var->IsCtrlVar()) {
uses_[op].insert(input_var->Name());
}
}
for (auto& output_var : op->outputs) {
// output var may be used by many op
for (auto* succ_op : output_var->outputs) {
if (succ_op->Op() != nullptr) {
successors_[op].insert(succ_op);
predecessors_[succ_op].insert(op);
}
}
if (output_var->IsVar() && !output_var->IsCtrlVar()) {
defs_[op].insert(output_var->Name());
}
}
}
}
void ControlFlowGraph::ConnectNodes() {
for (size_t i = 0; i < ops_.size(); ++i) {
auto& op = ops_[i];
try {
auto& next_op = ops_.at(i + 1);
successors_[op].insert(next_op);
predecessors_[next_op].insert(op);
} catch (...) {
// do nothing
}
FilterVariables(op->inputs,
[&](ir::Node* var) { uses_[op].emplace(var->Name()); });
FilterVariables(op->outputs,
[&](ir::Node* var) { defs_[op].emplace(var->Name()); });
}
}
void ControlFlowGraph::LiveVariableAnalysis() {
// NOTE(dzh): variable liveless analysis (a.k.a reversed_ops algorithm)
// compute the liveness of for each variable though reversed_ops algorithm.
// It iterates the operators from end to begin, compute the live in/live out
// variable set for each op, then the diff between in/out will be used for
// the variable reuse. For detail refer to
// http://www.cs.cornell.edu/courses/cs4120/2013fa/lectures/lec26-fa13.pdf
std::list<ir::Node*> work_list(ops_.rbegin(), ops_.rend());
while (!work_list.empty()) {
ir::Node* op = work_list.front();
work_list.pop_front();
// get the live_in calculated before. Empty if first.
auto prev_live_in = std::move(live_in_[op]);
for (auto& s : successors_[op]) {
for (auto& var : live_in_[s]) {
live_out_[op].insert(var);
}
}
for (auto& var : uses_[op]) {
live_in_[op].insert(var);
}
for (auto& var : live_out_[op]) {
live_in_[op].insert(var);
}
for (auto& var : defs_[op]) {
live_in_[op].erase(var);
}
// If the live_in is not changed, then the liveness analysis of
// predecessors is completed.
//
// Otherwise, recalculate the predecessors liveness
if (live_in_[op] != prev_live_in) {
for (auto& pre : predecessors_[op]) {
work_list.push_back(pre);
}
}
}
}
void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node,
const std::string& new_node,
int begin_idx) {
// update graph from begin idx to the end
for (size_t i = begin_idx; i != ops_.size(); ++i) {
auto* op = ops_[i];
if (uses_[op].find(old_node) != uses_[op].end()) {
uses_[op].erase(old_node);
uses_[op].insert(new_node);
}
if (defs_[op].find(old_node) != defs_[op].end()) {
defs_[op].erase(old_node);
defs_[op].insert(new_node);
}
if (live_in_[op].find(old_node) != live_in_[op].end()) {
live_in_[op].erase(old_node);
live_in_[op].insert(new_node);
}
if (live_out_[op].find(old_node) != live_out_[op].end()) {
live_out_[op].erase(old_node);
live_out_[op].insert(new_node);
}
}
}
const std::set<std::string> ControlFlowGraph::LiveIn(ir::Node* op) const {
auto it = live_in_.find(op);
PADDLE_ENFORCE(
it != live_in_.end(),
string::Sprintf("Expect %s in live_in, but Not Found.", op->Name()));
return it->second;
}
const std::set<std::string> ControlFlowGraph::LiveOut(ir::Node* op) const {
auto it = live_out_.find(op);
PADDLE_ENFORCE(
it != live_out_.end(),
string::Sprintf("Expect %s in live_out, but Not Found.", op->Name()));
return it->second;
}
const std::set<std::string> ControlFlowGraph::Use(ir::Node* op) const {
auto it = uses_.find(op);
PADDLE_ENFORCE(
it != uses_.end(),
string::Sprintf("Expect %s in live_out, but Not Found.", op->Name()));
return it->second;
}
const std::vector<ir::Node*> ControlFlowGraph::Ops() const { return ops_; }
std::vector<ir::Node*>& ControlFlowGraph::Ops() { return ops_; }
ir::Node* ControlFlowGraph::GetNodeByName(const std::string& name,
ir::Node* op) const {
// in ssa-graph, different version nodes have same name,
// this function get the latest version var before target op
// It may return nullptr, such as data node.
ir::Node* found_node = nullptr;
for (auto* node : ops_) {
if (node == op) break;
for (auto& output : node->outputs) {
PADDLE_ENFORCE((output != nullptr && output->IsVar()),
"Output is empty!");
if (output->Var() && output->Name() == name) {
found_node = output;
}
}
}
return found_node;
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/memory_optimize_helper.h
浏览文件 @ b8491bfd
......@@ -17,6 +17,8 @@
#include <iostream>
#include <iterator>
#include <list>
#include <map>
#include <set>
#include <string>
#include <utility>
#include <vector>
......@@ -27,41 +29,41 @@ namespace paddle {
namespace framework {
namespace details {
constexpr char kFetchedVars[] = "fetched_vars";
constexpr char kGraphNodePool[] = "graph_node_pool";
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
// NOTE(dzh): Variable and the operators use the var.
// for early delete pass.
// Because analysis var pass build base on ir::Node, which maybe released
// or modified between passes, so we use OpDesc* to mark ops.
using GraphNodePool = std::vector<
std::pair<std::string /*var node*/, std::unordered_set<OpDesc*> /* ops */>>;
// NOTE(dzh): by default, it sort node in ascend order(by node bytes size).
// in fluid, -1 means the batch_size is determined in runtime.
// the node batch_size equal -1 always ranking in the front than the node not.
// NOTE(dzh): A ordered set for node reuse in memory optimize.
// the orderedset sort node in ascend order(by node bytes size).
// in fluid, -1 means the batch_size, which is determined in runtime.
// So the reuse happens between nodes who's batch_size both are -1
// simultaneously or not.
//
// sort rule:
// rule 0 : smaller node ranking in front.
// rule 1 : batch_size equal -1 ranking in the front than the node not.
//
// For example,
// node0[-1, 1] node1[-1, 1, 1], node2[1,1], node3[1,1024], ..
// O(1) insert, delete
class OrderedNodeList {
public:
using NodePair = std::pair<ir::Node*, std::unordered_set<ir::Node*>>;
using Iter = typename std::list<NodePair>::iterator;
using ConstIter = typename std::list<NodePair>::const_iterator;
void Insert(ir::Node* var, ir::Node* op);
class OrderedSet {
public:
// nodes with same name exists in pool.
using NodeVector = std::vector<ir::Node*>;
using Iter = typename std::list<NodeVector>::iterator;
using ConstIter = typename std::list<NodeVector>::const_iterator;
void Insert(ir::Node* var);
void Erase(ir::Node* var);
void Erase(const std::string& var);
bool Has(ir::Node* var) { return mark_table_.count(var->Name()); }
bool Has(const std::string& var) { return mark_table_.count(var); }
ir::Node* NodeMatch(ir::Node* var) const;
bool Has(ir::Node* var) const;
void Clear() {
mark_table_.clear();
nodes_.clear();
}
// find the bestfit shape node block with var.
ir::Node* FindBestFitNode(ir::Node* var) const;
ir::Node* FindNextBestFitNode(ir::Node* var, ir::Node* prev) const;
// map store non-const iterator, can not promise const
int GetIndex(ir::Node* var);
int GetNodeIndexInPool(ir::Node* var);
// pool all node to string
std::string ToString() const;
......@@ -69,18 +71,54 @@ class OrderedNodeList {
Iter end() { return nodes_.end(); }
ConstIter begin() const { return nodes_.begin(); }
ConstIter end() const { return nodes_.end(); }
size_t size() const { return nodes_.size(); }
void Clear() {
mark_table_.clear();
nodes_.clear();
}
size_t size() const { return nodes_.size(); }
private:
// for searching.
std::unordered_map<std::string, Iter> mark_table_;
// node swap pairs. var -> ops dep var
std::list<NodePair> nodes_;
// node pool
std::list<NodeVector> nodes_;
};
class ControlFlowGraph {
public:
ControlFlowGraph() = default;
// IR Graph
explicit ControlFlowGraph(const ir::Graph& graph);
void LiveVariableAnalysis();
void RenameVarInCFGGraph(const std::string& old_node,
const std::string& new_node, int begin_idx);
const std::set<std::string> LiveIn(ir::Node* op) const;
const std::set<std::string> LiveOut(ir::Node* op) const;
const std::set<std::string> Use(ir::Node* op) const;
const std::vector<ir::Node*> Ops() const;
std::vector<ir::Node*>& Ops();
// for ssa-graph nodes
ir::Node* GetNodeByName(const std::string& name, ir::Node* op) const;
private:
void BuildCFGGraph();
void ConnectNodes();
using NodeListMap = std::unordered_map<ir::Node*, std::set<ir::Node*>>;
using VarSetMap = std::map<ir::Node*, std::set<std::string>>;
// successors ops use the output variables.
NodeListMap successors_;
// predecessors ops generated input variables.
NodeListMap predecessors_;
// variables lived before run current op.
VarSetMap live_in_;
// variables lived after run current op.
VarSetMap live_out_;
VarSetMap uses_; // op inputs
VarSetMap defs_; // op outputs
std::vector<ir::Node*> ops_; // op sequence by topology sort
};
// valid a tensor can be reuse or not
......@@ -93,15 +131,24 @@ bool NodeCanReused(const VarDesc& node);
bool OpHasSubBlock(OpDesc* desc);
// node memory size in bytes
size_t NodeSizeInBytes(ir::Node* n);
size_t NodeSize(ir::Node* n);
// node memory size in bytes
size_t NodeSizeInBytes(const VarDesc&);
size_t NodeSize(const VarDesc&);
std::string DebugString(ir::Node* var);
// NOTE(dzhwinter)
// after node reuse, the replaced node shape is
// different with its VarDesc. So need to find the
// correct VarDesc in Block.
VarDesc* FindVarDescInBlock(ir::Node* n);
static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) {
return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() &&
op1->Outputs() == op2->Outputs();
}
template <typename Container, typename Callback>
class FilterVariableImpl {
public:
......
paddle/fluid/framework/details/memory_optimize_helper_test.cc
浏览文件 @ b8491bfd
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/details/memory_optimize_helper.h"
#include <algorithm>
#include <iostream>
#include <iterator>
#include <memory>
#include <sstream>
#include <string>
......@@ -22,13 +23,19 @@
#include <vector>
#include "glog/logging.h"
#include "gtest/gtest.h"
#include "paddle/fluid/framework/details/graph_test_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
namespace paddle {
namespace framework {
namespace details {
TEST(OrderedNodeList, Normal) {
OrderedNodeList pool;
TEST(OrderedSet, Normal) {
OrderedSet pool;
std::vector<std::unique_ptr<ir::Node>> nodes;
// clang-format off
......@@ -56,8 +63,15 @@ TEST(OrderedNodeList, Normal) {
nodes.emplace_back(std::move(node));
}
// Insert
for (auto& node : nodes) {
pool.Insert(node.get(), op.get());
pool.Insert(node.get());
}
// Has/size
ASSERT_EQ(pool.size(), shapes.size());
for (auto& node : nodes) {
ASSERT_TRUE(pool.Has(node.get()));
}
// assert its order and interface.
......@@ -66,14 +80,14 @@ TEST(OrderedNodeList, Normal) {
std::cout << pool.ToString() << std::endl;
ASSERT_EQ(pool.size(), static_cast<size_t>(COUNT - 1));
ASSERT_EQ(pool.GetIndex(nodes.back().get()), 0);
ASSERT_EQ(pool.GetNodeIndexInPool(nodes.back().get()), 0);
{
auto v1 = block_desc->Var("11");
v1->SetShape({-1, 256, 56, 56});
std::unique_ptr<ir::Node> node1 = ir::CreateNodeForTest(v1);
node1->inputs.emplace_back(op.get());
auto* cache = pool.NodeMatch(node1.get());
auto* cache = pool.FindBestFitNode(node1.get());
ASSERT_EQ(cache, nullptr);
}
{
......@@ -81,16 +95,447 @@ TEST(OrderedNodeList, Normal) {
v2->SetShape({-1, 2, 5});
std::unique_ptr<ir::Node> node1 = ir::CreateNodeForTest(v2);
node1->inputs.emplace_back(op.get());
auto* cache = pool.NodeMatch(node1.get());
ASSERT_EQ(pool.GetIndex(cache), 2); // match 6:[-1,2,5]
auto* cache = pool.FindBestFitNode(node1.get());
ASSERT_EQ(pool.GetNodeIndexInPool(cache), 2); // match 6:[-1,2,5]
}
{
auto v3 = block_desc->Var("13");
v3->SetShape({2, 5});
std::unique_ptr<ir::Node> node1 = ir::CreateNodeForTest(v3);
node1->inputs.emplace_back(op.get());
auto* cache = pool.NodeMatch(node1.get());
ASSERT_EQ(pool.GetIndex(cache), 5); // match 4:[5,2]
auto* cache = pool.FindBestFitNode(node1.get());
ASSERT_EQ(pool.GetNodeIndexInPool(cache), 5); // match 4:[5,2]
}
}
TEST(OrderedSet, FindBestFitNode) {
OrderedSet pool;
std::vector<std::unique_ptr<ir::Node>> nodes;
ProgramDesc prog;
BlockDesc* block_desc = prog.MutableBlock(0);
auto* op_desc = block_desc->AppendOp();
op_desc->SetType("dummy");
std::unique_ptr<ir::Node> op = ir::CreateNodeForTest(op_desc);
{
auto desc = block_desc->Var("a");
desc->SetShape({128, 128});
std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
node->inputs.emplace_back(op.get());
nodes.emplace_back(std::move(node));
}
{
auto desc = block_desc->Var("b");
desc->SetShape({128, 129});
std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
node->inputs.emplace_back(op.get());
nodes.emplace_back(std::move(node));
}
{
auto desc = block_desc->Var("c");
desc->SetShape({128, 128});
std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
node->inputs.emplace_back(op.get());
nodes.emplace_back(std::move(node));
}
for (auto& node : nodes) {
pool.Insert(node.get());
}
// FindNextBestFitNode
auto* n = nodes[0].get();
auto* cache = pool.FindBestFitNode(n);
PADDLE_ENFORCE(cache->Name() == "a");
cache = pool.FindNextBestFitNode(n, cache);
PADDLE_ENFORCE(cache->Name() == "c");
cache = pool.FindNextBestFitNode(n, cache);
PADDLE_ENFORCE(cache->Name() == "b");
}
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_OPERATOR(sum, paddle::framework::DummyOp,
paddle::framework::SumOpMaker,
paddle::framework::DummyVarTypeInference);
REGISTER_OPERATOR(assign, paddle::framework::DummyOp,
paddle::framework::AssignOpMaker,
paddle::framework::DummyVarTypeInference);
REGISTER_OPERATOR(dummy, paddle::framework::DummyOp,
paddle::framework::SumOpMaker,
paddle::framework::DummyVarTypeInference);
/*
https://en.wikipedia.org/wiki/Live_variable_analysis
Create a customed classical dependency graph, left row is the instruction
number.
1. a = 1
2. b = a
3. c = a
4. d = b + c
5. e = d
a--------+
| |
b c
| |
d--------+
|
e
Then analysis these variable's liveness range
*/
namespace paddle {
namespace framework {
namespace details {
inline static ProgramDesc FillProgramDesc() {
ProgramDesc prog;
prog.MutableBlock(0)->Var("a")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("b")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("c")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("d")->SetType(proto::VarType::LOD_TENSOR);
prog.MutableBlock(0)->Var("e")->SetType(proto::VarType::LOD_TENSOR);
{
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("assign");
op->SetInput("X", {"a"});
op->SetOutput("Out", {"b"});
}
{
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("assign");
op->SetInput("X", {"a"});
op->SetOutput("Out", {"c"});
}
{
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("sum");
op->SetInput("X", {"b", "c"});
op->SetOutput("Out", {"d"});
}
{
auto* op = prog.MutableBlock(0)->AppendOp();
op->SetType("assign");
op->SetInput("X", {"d"});
op->SetOutput("Out", {"e"});
}
return prog;
}
TEST(CFGGraph, IRGraph) {
// prepare ir graph
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
ControlFlowGraph cfg(graph);
cfg.LiveVariableAnalysis();
// test assign op
ASSERT_TRUE((std::set<std::string>{"a"} == cfg.LiveIn(cfg.Ops()[0])));
ASSERT_TRUE((std::set<std::string>{"a", "b"} == cfg.LiveOut(cfg.Ops()[0])));
// test assign op
ASSERT_TRUE((std::set<std::string>{"a", "b"} == cfg.LiveIn(cfg.Ops()[1])));
ASSERT_TRUE((std::set<std::string>{"b", "c"} == cfg.LiveOut(cfg.Ops()[1])));
// test sum op
ASSERT_TRUE((std::set<std::string>{"b", "c"} == cfg.LiveIn(cfg.Ops()[2])));
ASSERT_TRUE((std::set<std::string>{"d"} == cfg.LiveOut(cfg.Ops()[2])));
// test assign op
ASSERT_TRUE((std::set<std::string>{"d"} == cfg.LiveIn(cfg.Ops()[3])));
ASSERT_TRUE((std::set<std::string>{} == cfg.LiveOut(cfg.Ops()[3])));
}
// 1. normal test
TEST(SortOpLikeDescOrder, NormalTest) {
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto nodes = SortOpLikeDescOrder(graph);
auto op_descs = prog.Block(0).AllOps();
for (size_t i = 0; i < nodes.size(); ++i) {
auto node = nodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
// 2. remove some op_desc
TEST(SortOpLikeDescOrder, RemoveOpDesc) {
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto nodes = graph.Nodes();
auto op_descs = prog.Block(0).AllOps();
ir::Node* found_node = nullptr;
for (auto node : nodes) {
if (node->IsOp() && node->outputs.back()->Name() == "e") {
found_node = node;
break;
}
}
PADDLE_ENFORCE(found_node != nullptr);
for (auto it = op_descs.begin(); it != op_descs.end();) {
if (IsSameDesc(*it, found_node->Op())) {
it = op_descs.erase(it);
} else {
++it;
}
}
auto find_node_in_graph = [&](std::string s) {
ir::Node* ret = nullptr;
for (auto n : graph.Nodes()) {
if (n->Name() == s) {
ret = n;
break;
}
}
PADDLE_ENFORCE(ret != nullptr);
return ret;
};
ir::Node* e = find_node_in_graph("e");
ir::Node* d = find_node_in_graph("d");
std::remove(d->outputs.begin(), d->outputs.end(), found_node);
graph.RemoveNode(found_node);
graph.RemoveNode(e);
// other node keeps the same order
auto remain_nodes = SortOpLikeDescOrder(graph);
for (size_t i = 0; i < remain_nodes.size(); ++i) {
auto node = remain_nodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
// 3. add some op_desc
TEST(SortOpLikeDescOrder, AddOpDesc) {
auto prog = FillProgramDesc();
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
ir::Graph graph(prog);
auto find_node_in_graph = [&](std::string s) {
ir::Node* ret = nullptr;
for (auto n : graph.Nodes()) {
if (n->Name() == s) {
ret = n;
break;
}
}
PADDLE_ENFORCE(ret != nullptr);
return ret;
};
// cached desc different with real one
// mimic the intermidiete pass modify the programdesc.
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto op_descs = prog.Block(0).AllOps();
auto op = prog.MutableBlock(0)->AppendOp();
prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR);
op->SetType("sum");
op->SetInput("X", {"b", "c"});
op->SetOutput("Out", {"d1"});
ir::Node* node = graph.CreateOpNode(op);
ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1"));
ir::Node* b = find_node_in_graph("b");
ir::Node* c = find_node_in_graph("c");
node->outputs.emplace_back(d1);
node->inputs.emplace_back(b);
node->inputs.emplace_back(c);
d1->inputs.emplace_back(node);
b->outputs.emplace_back(node);
c->outputs.emplace_back(node);
op_descs.insert(op_descs.begin() + 4, op);
auto nodes = SortOpLikeDescOrder(graph);
for (size_t i = 0; i < nodes.size(); ++i) {
auto node = nodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
// 4. add and delete some op_desc
TEST(SortOpLikeDescOrder, AddAndDeleteOpDesc) {
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto find_node_in_graph = [&](std::string s) {
ir::Node* ret = nullptr;
for (auto n : graph.Nodes()) {
if (n->Name() == s) {
ret = n;
break;
}
}
PADDLE_ENFORCE(ret != nullptr);
return ret;
};
// remove sum node
auto op_descs = prog.Block(0).AllOps();
ir::Node* found_node = nullptr;
auto nodes = graph.Nodes();
for (auto node : nodes) {
if (node->Name() == "sum") {
found_node = node;
break;
}
}
PADDLE_ENFORCE(found_node != nullptr);
for (auto it = op_descs.begin(); it != op_descs.end();) {
if (IsSameDesc(*it, found_node->Op())) {
it = op_descs.erase(it);
} else {
++it;
}
}
{
ir::Node* d = find_node_in_graph("d");
ir::Node* c = find_node_in_graph("c");
ir::Node* e = find_node_in_graph("e");
std::remove(d->outputs.begin(), d->outputs.end(), found_node);
std::remove(c->outputs.begin(), c->outputs.end(), found_node);
ir::Node* pending_op = found_node->outputs[0]->outputs[0];
graph.RemoveNode(e);
graph.RemoveNode(pending_op);
graph.RemoveNode(found_node);
}
// add node
auto op = prog.MutableBlock(0)->AppendOp();
prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR);
op->SetType("sum");
op->SetInput("X", {"b", "c"});
op->SetOutput("Out", {"d1"});
{
ir::Node* node = graph.CreateOpNode(op);
ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1"));
ir::Node* b = find_node_in_graph("b");
ir::Node* c = find_node_in_graph("c");
node->outputs.emplace_back(d1);
node->inputs.emplace_back(b);
node->inputs.emplace_back(c);
b->outputs.emplace_back(node);
c->outputs.emplace_back(node);
}
op_descs.insert(op_descs.begin() + 2, op);
// check the order
auto mynodes = SortOpLikeDescOrder(graph);
for (size_t i = 0; i < mynodes.size(); ++i) {
auto node = mynodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
// 5. add and replace some op_desc inplace.
TEST(SortOpLikeDescOrder, AddAndReplaceOpDescInplace) {
auto prog = FillProgramDesc();
ir::Graph graph(prog);
const std::vector<OpDesc*>* all_op_descs =
new std::vector<OpDesc*>(prog.Block(0).AllOps());
graph.Set(details::kAllOpDescs, all_op_descs); // take ownership
auto find_node_in_graph = [&](std::string s) {
ir::Node* ret = nullptr;
for (auto n : graph.Nodes()) {
if (n->Name() == s) {
ret = n;
break;
}
}
PADDLE_ENFORCE(ret != nullptr);
return ret;
};
auto op_descs = prog.Block(0).AllOps();
// add node
auto op = prog.MutableBlock(0)->AppendOp();
prog.MutableBlock(0)->Var("d1")->SetType(proto::VarType::LOD_TENSOR);
op->SetType("sum");
op->SetInput("X", {"b", "c"});
op->SetOutput("Out", {"d1"});
{
ir::Node* node = graph.CreateOpNode(op);
ir::Node* d1 = graph.CreateVarNode(prog.MutableBlock(0)->Var("d1"));
ir::Node* b = find_node_in_graph("b");
ir::Node* c = find_node_in_graph("c");
node->outputs.emplace_back(d1);
node->inputs.emplace_back(b);
node->inputs.emplace_back(c);
d1->inputs.emplace_back(node);
b->outputs.emplace_back(node);
c->outputs.emplace_back(node);
}
op_descs.emplace_back(op);
// replace op_desc inplace
auto nodes = graph.Nodes();
ir::Node* found_node = nullptr;
for (auto node : nodes) {
if (node->IsOp() && node->Op() && node->Name() == "assign") {
if (node->outputs.size() == 1 && node->outputs[0]->Name() == "e") {
found_node = node;
break;
}
}
}
{
ir::Node* d = find_node_in_graph("d");
ir::Node* e = find_node_in_graph("e");
std::remove(d->outputs.begin(), d->outputs.end(), found_node);
std::remove(e->inputs.begin(), e->inputs.end(), found_node);
graph.RemoveNode(found_node);
}
op_descs.erase(op_descs.begin() + 3);
auto replace_op = prog.MutableBlock(0)->AppendOp();
replace_op->SetType("sum");
replace_op->SetInput("X", {"d", "d1"});
replace_op->SetOutput("Out", {"e"});
{
ir::Node* sum2 = graph.CreateOpNode(replace_op);
ir::Node* e = find_node_in_graph("e");
ir::Node* d = find_node_in_graph("d");
ir::Node* d1 = find_node_in_graph("d1");
sum2->inputs.emplace_back(d);
sum2->inputs.emplace_back(d1);
sum2->outputs.emplace_back(e);
e->inputs.emplace_back(sum2);
d->outputs.emplace_back(sum2);
d1->outputs.emplace_back(sum2);
}
op_descs.emplace_back(replace_op);
// compare op order
auto graph_nodes = SortOpLikeDescOrder(graph);
for (size_t i = 0; i < graph_nodes.size(); ++i) {
auto node = graph_nodes[i];
auto op_desc = op_descs[i];
ASSERT_TRUE(IsSameDesc(node->Op(), op_desc));
}
}
......
paddle/fluid/framework/details/memory_optimize_pass.cc
浏览文件 @ b8491bfd
......@@ -43,11 +43,6 @@ namespace paddle {
namespace framework {
namespace details {
static inline bool IsSameDesc(OpDesc* op1, OpDesc* op2) {
return op1->Type() == op2->Type() && op1->Inputs() == op2->Inputs() &&
op1->Outputs() == op2->Outputs();
}
std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
auto nodes = graph->Nodes();
......@@ -74,82 +69,67 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
}
for (auto& var : op->outputs) {
if (!NodeCanReused(var) || cfg_->Use(op).count(var->Name()) == 0 ||
skip_set_.count(var->Name()))
if (var->IsVar() && !var->IsCtrlVar() && skip_set_.count(var->Name())) {
VLOG(3) << "Skip set contains variable of " << var->Name()
<< "disable reuse on it. skipped";
continue;
ir::Node* cache = pool_.NodeMatch(var);
if (var->Name() == FLAGS_memory_optimize_debug) {
VLOG(3) << "start match var " << DebugString(var) << " of op "
<< op->Name();
VLOG(3) << pool_.ToString();
VLOG(3) << "matched in pool : "
<< ((cache == nullptr) ? "False" : "True");
}
if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) {
ir::Node* cache = pool_.FindBestFitNode(var);
while (cache != nullptr && var->Name() == cache->Name()) {
VLOG(3) << "The same cache variable is cascade reused. "
<< cache->Name() << " is re-filled to the pool after "
<< "the reused op is finished. Current op can not "
<< "replace it again. Skip this candidate.";
cache = pool_.FindNextBestFitNode(var, cache);
}
if (var->Name() == FLAGS_memory_optimize_debug) {
VLOG(3) << "start match var " << DebugString(var) << " of op "
<< op->Name();
VLOG(3) << pool_.ToString();
VLOG(3) << "matched in pool : "
<< ((cache == nullptr) ? "False" : "True");
}
if (cache == nullptr) continue;
if (var->Name() == cache->Name()) {
VLOG(3) << "The same cache variable is cascade reused." << var->Name()
<< " is re-filled to the pool after"
<< "the reused op is finished. Current op can not "
<< "replace it again. Skip this candidate.";
continue;
int node_idx_in_pool = pool_.GetIndex(cache);
VLOG(3) << string::Sprintf(
"!!! %s, %s => %s, cache idx %d, pool size %d",
std::to_string(reuse_id++), DebugString(var), DebugString(cache),
node_idx_in_pool, static_cast<int>(pool_.size()));
// update CFG Graph on the fly.
// reused var maybe re-fill into the pool
cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx);
// NOTE(dzhwinter): we need to both update the ProgramDesc
// and IR Graph. because op_desc/var_desc is used in CreateOp,
// CreateVar when running happens. But IR Graph
// define the dependence relationship between nodes.
RenameVarInGraphDesc(var->Name(), cache->Name(), idx);
RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get());
pool_.Erase(cache);
}
// fill the pool
std::unordered_set<std::string> unlived_vars;
for (auto var : cfg_->LiveIn(op)) {
if (cfg_->LiveOut(op).count(var) == 0) {
unlived_vars.emplace(var);
if (cache != nullptr) {
int node_idx_in_pool = pool_.GetNodeIndexInPool(cache);
VLOG(3) << string::Sprintf(
"!!! %s, %s => %s, cache idx %d, pool size %d",
std::to_string(reuse_id++), DebugString(var), DebugString(cache),
node_idx_in_pool, static_cast<int>(pool_.size()));
// NOTE(dzhwinter): update the ProgramDesc/IR Graph
// and the CFG Graph on the fly.
//
// IR Graph define the dependence relationship between nodes.
//
// ProgramDesc defines the input/output vars. Its used in
// CreateOp, CreateVar when running happens.
//
// CFG Graph store the liveness information, when reuse happens
// we also need to update the variable liveness.
const std::string var_name = var->Name();
const std::string cache_name = cache->Name();
cfg_->RenameVarInCFGGraph(var_name, cache_name, idx);
RenameVarInGraphDesc(var_name, cache_name, idx);
RenameVarInGraphNode(var_name, cache_name, idx, graph.get());
pool_.Erase(cache_name);
}
}
for (auto var : unlived_vars) {
ir::Node* var_node = cfg_->GetNodeFromVarName(var, op);
}
// fill the pool
for (auto var : cfg_->LiveIn(op)) {
if (cfg_->LiveOut(op).count(var) == 0) {
ir::Node* var_node = cfg_->GetNodeByName(var, op);
if (var_node == nullptr || var_node->IsCtrlVar()) continue;
if (NodeCanReused(var_node) && !pool_.Has(var_node)) {
pool_.Insert(var_node, op);
pool_.Insert(var_node);
}
}
}
}
graph->ResolveHazard(var_nodes_);
// For early delete pass. use GraphNodePool load the unlived vars.
// 1. find all deps op for each unlived var in memory pool.
for (auto& op : graph->Nodes()) {
for (auto& var : op->inputs) {
if (pool_.Has(var)) {
pool_.Insert(var, op);
}
}
}
// 2. convert ir node based memory pool to graph node
// because Node* maybe released bettwen passes.
auto& graph_pool = graph->Get<GraphNodePool>(kGraphNodePool);
for (auto it = pool_.begin(); it != pool_.end(); ++it) {
std::unordered_set<OpDesc*> descs;
for (auto& op : it->second) {
PADDLE_ENFORCE(op->IsOp());
descs.insert(op->Op());
}
graph_pool.push_back(std::make_pair(it->first->Name(), descs));
}
return graph;
}
......@@ -198,12 +178,12 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
PADDLE_ENFORCE(sub_op != nullptr);
for (auto* var : sub_op->outputs) {
if (NodeCanReused(var)) {
ir::Node* cache = pool_.NodeMatch(var);
ir::Node* cache = pool_.FindBestFitNode(var);
if (cache != nullptr) {
if (var->Var()->GetDataType() != cache->Var()->GetDataType()) {
continue;
}
int node_idx_in_pool = pool_.GetIndex(cache);
int node_idx_in_pool = pool_.GetNodeIndexInPool(cache);
VLOG(3) << string::Sprintf(
"!!! %s, %s => %s, cache idx %d, pool size %d",
std::to_string(sub_reuse_id++), DebugString(var),
......@@ -214,7 +194,8 @@ void MemoryOptimizePass::SubGraphOptimize(OpDesc* op_desc) const {
// effect. Because it is a single op in graph. No need to
// update the ir nodes.
sub_op_desc->Rename(var->Name(), cache->Name());
if (sub_op_desc->Block()->HasVar(var->Name())) {
if (sub_op_desc->Block() != nullptr &&
sub_op_desc->Block()->HasVar(var->Name())) {
sub_op_desc->Block()->RemoveVar(var->Name());
}
}
......@@ -255,7 +236,13 @@ void MemoryOptimizePass::RenameVarInGraphDesc(const std::string& var,
auto* op_desc = op->Op();
op_desc->RenameInput(var, cache_var);
op_desc->RenameOutput(var, cache_var);
if (op_desc->Block()->HasVar(var)) op_desc->Block()->RemoveVar(var);
if (op_desc->Block() != nullptr) {
op_desc->Block()->RemoveVar(var);
} else {
LOG(WARNING) << "op " << op->Name() << " not know its block."
<< "Is the op_desc created without block pointer? "
<< "Can not find " << var << " in Block(0)";
}
op_desc->Flush();
}
}
......@@ -297,8 +284,7 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
// redirect the input to the latest version of cache_var
for (auto* node : op->inputs) {
if (node->Name() == var) {
ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
var_nodes_[cache_var].emplace_back(cache_node);
ir::Node* cache_node = var_nodes_[cache_var].back();
// swap node to cache_node
cache_node->outputs.insert(cache_node->outputs.end(),
......@@ -307,11 +293,15 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
auto* prev_op = node->inputs[0];
std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node,
cache_node);
cache_node->inputs.emplace_back(prev_op);
for (auto* next_op : node->outputs) {
std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
cache_node);
}
// erase unused node
auto& nodes = var_nodes_.at(var);
nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
graph->RemoveNode(node);
}
}
......@@ -331,271 +321,14 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
cache_node);
}
}
}
}
// release node of unused var in graph
for (auto* node : var_nodes_[var]) {
graph->RemoveNode(node);
}
var_nodes_.at(var).clear();
}
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph) {
PADDLE_ENFORCE(graph.Has(kAllOpDescs),
"Graph has no attribute of kAllOpDescs.");
// 1. get op desc order
auto& op_descs = graph.Get<const std::vector<OpDesc*>>(kAllOpDescs);
// 2. topology sort order
auto nodes = graph.Nodes();
std::deque<ir::Node*> ops;
FilterVariables(nodes, [&](ir::Node* op) {
if (op->IsOp() && op->Op() != nullptr) {
ops.emplace_back(op);
}
});
std::unordered_map<ir::Node*, size_t> op_deps;
std::list<ir::Node*> ready_ops;
std::unordered_map<ir::Node*, std::unordered_set<ir::Node*>> pending_ops;
for (auto* op : ops) {
std::unordered_set<ir::Node*> preceding_op;
for (auto* in : op->inputs) {
if (in->inputs.empty()) continue;
PADDLE_ENFORCE(in->inputs.size() == 1 && in->inputs[0]->IsOp());
preceding_op.emplace(in->inputs[0]);
pending_ops[in->inputs[0]].emplace(op);
}
op_deps[op] = preceding_op.size();
if (preceding_op.empty()) {
ready_ops.emplace_back(op);
}
}
// 3. generated op list based desc order and the topology order
std::vector<ir::Node*> ret;
std::list<OpDesc*> op_descs_list(op_descs.begin(), op_descs.end());
auto update_by_found_node = [&](ir::Node* found_node) {
for (auto* pending_op : pending_ops[found_node]) {
if (--op_deps[pending_op] == 0) {
ready_ops.emplace_back(pending_op);
}
}
ready_ops.remove(found_node);
ret.emplace_back(found_node);
};
while (!ready_ops.empty()) {
bool all_of_ready_op_unmatched = true;
for (auto it = op_descs_list.begin(); it != op_descs_list.end();) {
auto op_desc = *it;
ir::Node* found_node = nullptr;
for (auto* op : ready_ops) {
if (IsSameDesc(op->Op(), op_desc)) {
found_node = op;
break;
}
}
// 3.1 op desc deleted by other pass
if (found_node == nullptr) {
++it;
continue;
} else {
all_of_ready_op_unmatched = false;
it = op_descs_list.erase(it);
}
update_by_found_node(found_node);
}
// 3.2 op descs are added by other pass
// preceding op non empty means some new op descs are
// created, but not contained in return node list.
// these new op desc may depend on each other.
std::list<ir::Node*> prev_ready_ops(ready_ops);
if (all_of_ready_op_unmatched) {
for (auto op : prev_ready_ops) {
update_by_found_node(op);
}
}
}
PADDLE_ENFORCE(std::all_of(
op_deps.begin(), op_deps.end(),
[&](const std::pair<ir::Node*, size_t>& p) { return p.second == 0; }));
return ret;
}
ControlFlowGraph::ControlFlowGraph(const ir::Graph& graph) {
ops_ = SortOpLikeDescOrder(graph);
ConnectNodes();
}
void ControlFlowGraph::BuildCFGGraph() {
// FIXME(dzh): same effect with ConnectNodes, but use the control
// link to build dependency graph, it goes wrong in transformer.
for (ir::Node* op : ops_) {
for (auto& input_var : op->inputs) {
if (!input_var->inputs.empty()) {
PADDLE_ENFORCE(
input_var->inputs.size() == 1 && input_var->inputs[0]->IsOp(),
"Preceding Op Node of Var Node must be unique");
auto* pred_op = input_var->inputs[0];
if (pred_op->Op() != nullptr) {
predecessors_[op].insert(pred_op);
successors_[pred_op].insert(op);
}
}
if (input_var->IsVar() && !input_var->IsCtrlVar()) {
uses_[op].insert(input_var->Name());
}
}
for (auto& output_var : op->outputs) {
// output var may be used by many op
for (auto* succ_op : output_var->outputs) {
if (succ_op->Op() != nullptr) {
successors_[op].insert(succ_op);
predecessors_[succ_op].insert(op);
}
}
if (output_var->IsVar() && !output_var->IsCtrlVar()) {
defs_[op].insert(output_var->Name());
}
}
}
}
void ControlFlowGraph::ConnectNodes() {
for (size_t i = 0; i < ops_.size(); ++i) {
auto& op = ops_[i];
try {
auto& next_op = ops_.at(i + 1);
successors_[op].insert(next_op);
predecessors_[next_op].insert(op);
} catch (...) {
// do nothing
}
FilterVariables(op->inputs,
[&](ir::Node* var) { uses_[op].emplace(var->Name()); });
FilterVariables(op->outputs,
[&](ir::Node* var) { defs_[op].emplace(var->Name()); });
}
}
void ControlFlowGraph::LiveVariableAnalysis() {
// NOTE(dzh): variable liveless analysis (a.k.a reversed_ops algorithm)
// compute the liveness of for each variable though reversed_ops algorithm.
// It iterates the operators from end to begin, compute the live in/live out
// variable set for each op, then the diff between in/out will be used for
// the variable reuse. For detail refer to
// http://www.cs.cornell.edu/courses/cs4120/2013fa/lectures/lec26-fa13.pdf
std::list<ir::Node*> work_list(ops_.rbegin(), ops_.rend());
while (!work_list.empty()) {
ir::Node* op = work_list.front();
work_list.pop_front();
// get the live_in calculated before. Empty if first.
auto prev_live_in = std::move(live_in_[op]);
for (auto& s : successors_[op]) {
for (auto& var : live_in_[s]) {
live_out_[op].insert(var);
}
}
for (auto& var : uses_[op]) {
live_in_[op].insert(var);
}
for (auto& var : live_out_[op]) {
live_in_[op].insert(var);
}
for (auto& var : defs_[op]) {
live_in_[op].erase(var);
}
// If the live_in is not changed, then the liveness analysis of
// predecessors is completed.
//
// Otherwise, recalculate the predecessors liveness
if (live_in_[op] != prev_live_in) {
for (auto& pre : predecessors_[op]) {
work_list.push_back(pre);
}
}
}
}
void ControlFlowGraph::RenameVarInCFGGraph(const std::string& old_node,
const std::string& new_node,
int begin_idx) {
// update graph from begin idx to the end
for (size_t i = begin_idx; i != ops_.size(); ++i) {
auto* op = ops_[i];
if (uses_[op].find(old_node) != uses_[op].end()) {
uses_[op].erase(old_node);
uses_[op].insert(new_node);
}
if (defs_[op].find(old_node) != defs_[op].end()) {
defs_[op].erase(old_node);
defs_[op].insert(new_node);
}
if (live_in_[op].find(old_node) != live_in_[op].end()) {
live_in_[op].erase(old_node);
live_in_[op].insert(new_node);
}
if (live_out_[op].find(old_node) != live_out_[op].end()) {
live_out_[op].erase(old_node);
live_out_[op].insert(new_node);
}
}
}
const std::set<std::string> ControlFlowGraph::LiveIn(ir::Node* op) const {
auto it = live_in_.find(op);
PADDLE_ENFORCE(
it != live_in_.end(),
string::Sprintf("Expect %s in live_in, but Not Found.", op->Name()));
return it->second;
}
const std::set<std::string> ControlFlowGraph::LiveOut(ir::Node* op) const {
auto it = live_out_.find(op);
PADDLE_ENFORCE(
it != live_out_.end(),
string::Sprintf("Expect %s in live_out, but Not Found.", op->Name()));
return it->second;
}
const std::set<std::string> ControlFlowGraph::Use(ir::Node* op) const {
auto it = uses_.find(op);
PADDLE_ENFORCE(
it != uses_.end(),
string::Sprintf("Expect %s in live_out, but Not Found.", op->Name()));
return it->second;
}
const std::vector<ir::Node*> ControlFlowGraph::Ops() const { return ops_; }
std::vector<ir::Node*>& ControlFlowGraph::Ops() { return ops_; }
ir::Node* ControlFlowGraph::GetNodeFromVarName(const std::string& name,
ir::Node* op) const {
// in ssa-graph, different version nodes have same name,
// this function get the latest version var before target op
// It may return nullptr, such as data node.
ir::Node* found_node = nullptr;
for (auto* node : ops_) {
if (node == op) break;
for (auto& output : node->outputs) {
if (output->Name() == name) {
found_node = output;
// erase unused node
auto& nodes = var_nodes_.at(var);
nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
graph->RemoveNode(node);
}
}
}
return found_node;
}
} // namespace details
......@@ -604,5 +337,4 @@ ir::Node* ControlFlowGraph::GetNodeFromVarName(const std::string& name,
REGISTER_PASS(memory_optimize_pass,
paddle::framework::details::MemoryOptimizePass)
.RequireGraphAttr(paddle::framework::details::kGraphNodePool)
.RequireGraphAttr(paddle::framework::details::kAllOpDescs);
paddle/fluid/framework/details/memory_optimize_pass.h
浏览文件 @ b8491bfd
......@@ -32,20 +32,15 @@
namespace paddle {
namespace framework {
namespace details {
constexpr char kAllOpDescs[] = "all_op_descs";
std::vector<ir::Node*> SortOpLikeDescOrder(const ir::Graph& graph);
class ControlFlowGraph;
class MemoryOptimizePass : public ir::Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
private:
// fill the variable map(var_nodes) by version.
void InitSSAGraphNodes() const;
private:
// update program descs
void RenameVarInGraphDesc(const std::string& var,
const std::string& cache_var, size_t idx) const;
......@@ -62,7 +57,7 @@ class MemoryOptimizePass : public ir::Pass {
private:
// Reuse Node Pool, Owned.
mutable OrderedNodeList pool_;
mutable OrderedSet pool_;
// controlflow Graph
mutable std::unique_ptr<ControlFlowGraph> cfg_;
// skip set
......@@ -71,45 +66,6 @@ class MemoryOptimizePass : public ir::Pass {
mutable std::map<std::string, std::vector<ir::Node*>> var_nodes_;
};
class ControlFlowGraph {
public:
ControlFlowGraph() = default;
// For IR Graph in parallelexecutor
explicit ControlFlowGraph(const ir::Graph& graph);
void LiveVariableAnalysis();
void RenameVarInCFGGraph(const std::string& old_node,
const std::string& new_node, int begin_idx);
const std::set<std::string> LiveIn(ir::Node* op) const;
const std::set<std::string> LiveOut(ir::Node* op) const;
const std::set<std::string> Use(ir::Node* op) const;
const std::vector<ir::Node*> Ops() const;
std::vector<ir::Node*>& Ops();
// for ssa-graph nodes
ir::Node* GetNodeFromVarName(const std::string& name, ir::Node* op) const;
private:
void BuildCFGGraph();
void ConnectNodes();
using NodeListMap = std::unordered_map<ir::Node*, std::set<ir::Node*>>;
using VarSetMap = std::map<ir::Node*, std::set<std::string>>;
// successors ops use the output variables.
NodeListMap successors_;
// predecessors ops generated input variables.
NodeListMap predecessors_;
// variables lived before run current op.
VarSetMap live_in_;
// variables lived after run current op.
VarSetMap live_out_;
VarSetMap uses_; // op inputs
VarSetMap defs_; // op outputs
std::vector<ir::Node*> ops_; // op sequence by topology sort
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ b8491bfd
......@@ -392,20 +392,32 @@ void MultiDevSSAGraphBuilderBase::CreateComputationalOp(ir::Graph *result,
void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
ir::Graph *result, const std::string &og) const {
OpHandleBase *op_handle = nullptr;
auto append_allreduce_op = [&](
const std::vector<Scope *> &scopes,
const std::vector<platform::Place> &places) -> OpHandleBase * {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_));
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
scopes, places, nccl_ctxs_));
#else
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
local_scopes_, places_));
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
scopes, places));
#endif
auto *op_handle = result->Get<GraphOps>(kGraphOps).back();
return result->Get<GraphOps>(kGraphOps).back();
};
if (!strategy_.enable_parallel_graph_)
op_handle = append_allreduce_op(local_scopes_, places_);
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
if (strategy_.enable_parallel_graph_) {
op_handle = append_allreduce_op({local_scopes_[i]}, {places_[i]});
}
SetCommunicationContext(op_handle, places_[i]);
auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
PADDLE_ENFORCE(!vars.empty());
auto &prev_grad = vars.back();
......@@ -413,7 +425,7 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(
auto var =
new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable),
vars.size(), i, og, p);
vars.size(), i, og, places_[i]);
vars.emplace_back(var);
op_handle->AddOutput(var);
}
......@@ -925,7 +937,21 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
}
void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
if (UseGPU() && strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
// broad cast received parameters when training in parameter server mode.
if (need_broadcast_var_) {
// There are 4 conditions:
// 1. GPU && Reduce: Reduce gradient then broadcast gradient to other GPUS.
// Need to broadcast received parameters to other GPU.
// 2. GPU && AllReduce: AllReduce all graident to each GPU. Need to
// broadcast received parameters to other GPU.
// 3. CPU && AllReduce: AllReduce all gradient to each thread. Need to
// broadcast received parameters to other scope.
// 4. CPU && Reduce: because all parameters share the same memory, did not
// broadcast received parameters.
if (!UseGPU() &&
strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
return;
}
if (strategy_.fuse_broadcast_op_) {
CreateFusedBroadcastOp(result, bcast_var_name_set_);
} else {
......
paddle/fluid/framework/details/multi_devices_helper.h
浏览文件 @ b8491bfd
......@@ -36,13 +36,14 @@ namespace details {
// map from variable name to variables. The variables, who have the same name,
// will have a differsent version. The offset in the
// `std::vector<VarHandle*>` is the version of varaibles.
typedef std::vector<std::unordered_map<std::string, std::vector<VarHandle*>>>
typedef std::vector<std::unordered_map<std::string, std::vector<VarHandle *>>>
GraphVars;
const char kGraphVars[] = "vars";
// aux variables to represent dependency. Useful to resolve data hazard.
typedef std::unordered_set<VarHandleBase*> GraphDepVars;
typedef std::unordered_set<VarHandleBase *> GraphDepVars;
const char kGraphDepVars[] = "dep_vars";
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/op_handle_base.h
浏览文件 @ b8491bfd
......@@ -70,6 +70,9 @@ class OpHandleBase {
auto it = dev_ctxes_.find(place);
return it != dev_ctxes_.end() ? it->second : nullptr;
}
const std::map<platform::Place, platform::DeviceContext *> &DeviceContext() {
return dev_ctxes_;
}
void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) {
dev_ctxes_[place] = ctx_;
......
paddle/fluid/framework/details/parallel_ssa_graph_executor.cc
浏览文件 @ b8491bfd
......@@ -13,22 +13,92 @@
// limitations under the License.
#include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
std::vector<std::unique_ptr<ir::Graph>>
ParallelSSAGraphExecutor::SeparateMultiDevicesGraph(
std::unique_ptr<ir::Graph> &&graph) {
std::vector<std::unique_ptr<ir::Graph>> graphs;
graphs.reserve(places_.size());
for (size_t i = 0; i < places_.size(); ++i) {
ProgramDesc empty;
graphs.emplace_back(std::unique_ptr<ir::Graph>(new ir::Graph(empty)));
auto &g = graphs.back();
g->Set(kGraphVars, new GraphVars(1UL));
g->Set(kGraphDepVars, new GraphDepVars);
}
auto op_handles = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
for (auto &op : op_handles) {
auto &dev_ctx = op->DeviceContext();
auto &p = dev_ctx.begin()->first;
int dev_id = boost::get<platform::CUDAPlace>(p).device;
auto &dev_dummys = graphs[dev_id]->Get<GraphDepVars>(kGraphDepVars);
graphs[dev_id]->AddNode(graph->RemoveNode(op->Node()).release());
for (auto &var : op->Inputs()) {
auto dummy_ptr = dynamic_cast<DummyVarHandle *>(var);
if (dummy_ptr) {
dev_dummys.insert(var);
if (graph->Nodes().count(var->Node()))
graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release());
}
}
for (auto &var : op->Outputs()) {
auto dummy_ptr = dynamic_cast<DummyVarHandle *>(var);
if (dummy_ptr) {
dev_dummys.insert(var);
if (graph->Nodes().count(var->Node()))
graphs[dev_id]->AddNode(graph->RemoveNode(var->Node()).release());
}
}
}
for (size_t dev_id = 0; dev_id < places_.size(); ++dev_id) {
auto &dev_vars = graphs[dev_id]->Get<GraphVars>(kGraphVars)[0];
auto &origin_vars = graph->Get<GraphVars>(kGraphVars)[dev_id];
for (auto &name_pair : origin_vars) {
dev_vars.emplace(name_pair.first, name_pair.second);
for (auto &version_pair : name_pair.second) {
if (graph->Nodes().count(version_pair->Node())) {
graphs[dev_id]->AddNode(
graph->RemoveNode(version_pair->Node()).release());
}
}
}
}
return graphs;
}
ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::vector<std::unique_ptr<ir::Graph>> &&graphs)
const framework::ProgramDesc &main_prog, std::unique_ptr<ir::Graph> &&graph)
: strategy_(std::move(strategy)),
local_scopes_(std::move(local_scopes)),
pool_(places.size() >= 2 ? new ::ThreadPool(places.size()) : nullptr),
places_(std::move(places)),
graphs_(std::move(graphs)) {
main_prog_(main_prog),
// TODO(Yancey1989): Copying graphs is not safely since it deleted the
// attrs.
graphs_(SeparateMultiDevicesGraph(std::move(graph))) {
PADDLE_ENFORCE_EQ(places_.size(), local_scopes_.size());
auto seq_allreduce_pass =
ir::PassRegistry::Instance().Get("all_reduce_deps_pass");
seq_allreduce_pass->Erase(details::kAllOpDescs);
seq_allreduce_pass->Set<const std::vector<OpDesc *>>(
details::kAllOpDescs,
new std::vector<OpDesc *>(main_prog_.Block(0).AllOps()));
for (size_t i = 0; i < graphs_.size(); ++i) {
graphs_[i] = seq_allreduce_pass->Apply(std::move(graphs_[i]));
}
// set the correct size of thread pool to each device.
strategy_.num_threads_ = strategy_.num_threads_ < places_.size()
? 1UL
......@@ -37,7 +107,7 @@ ParallelSSAGraphExecutor::ParallelSSAGraphExecutor(
<< " to run the operators of the graph on each device.";
for (size_t i = 0; i < places.size(); ++i) {
executors_.emplace_back(new details::ThreadedSSAGraphExecutor(
strategy_, {local_scopes_[i]}, {places_[i]}, std::move(graphs_[i])));
strategy_, local_scopes_, {places_[i]}, std::move(graphs_.at(i))));
}
}
......@@ -65,7 +135,7 @@ FeedFetchList ParallelSSAGraphExecutor::Run(
if (pool_) {
run_futures.emplace_back(pool_->enqueue(std::move(call)));
} else {
fetch_data.emplace_back(std::move(call()));
fetch_data.emplace_back(call());
}
}
......@@ -74,7 +144,7 @@ FeedFetchList ParallelSSAGraphExecutor::Run(
if (exception_holder_.IsCaught()) {
f.wait();
} else {
fetch_data.emplace_back(std::move(f.get()));
fetch_data.emplace_back(f.get());
}
}
}
......
paddle/fluid/framework/details/parallel_ssa_graph_executor.h
浏览文件 @ b8491bfd
......@@ -18,7 +18,9 @@
#include <vector>
#include "ThreadPool.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
#include "paddle/fluid/framework/ir/graph.h"
namespace paddle {
namespace framework {
......@@ -29,17 +31,23 @@ class ParallelSSAGraphExecutor : public SSAGraphExecutor {
ParallelSSAGraphExecutor(const ExecutionStrategy &strategy,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::vector<std::unique_ptr<ir::Graph>> &&graphs);
const framework::ProgramDesc &main_prog,
std::unique_ptr<ir::Graph> &&graph);
~ParallelSSAGraphExecutor() final = default;
const ir::Graph &Graph() const override { return *graphs_[0]; }
FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
private:
std::vector<std::unique_ptr<ir::Graph>> SeparateMultiDevicesGraph(
std::unique_ptr<ir::Graph> &&graph);
ExecutionStrategy strategy_;
std::vector<Scope *> local_scopes_;
std::unique_ptr<::ThreadPool> pool_{nullptr};
std::vector<platform::Place> places_;
framework::ProgramDesc main_prog_;
std::vector<std::unique_ptr<ir::Graph>> graphs_;
std::vector<std::unique_ptr<details::ThreadedSSAGraphExecutor>> executors_;
......
paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc
浏览文件 @ b8491bfd
......@@ -63,7 +63,7 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
eptr = std::current_exception();
}
platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr);
platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun");
++drop_scope_counter_;
bool stream_end = false;
......
paddle/fluid/framework/ir/graph.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/framework/ir/graph.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/framework/operator.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/framework/operator.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/framework/operator_kernel_configs.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/framework/scope.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/framework/tensor.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/imperative/layer.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/imperative/layer.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/imperative/tracer.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/inference/api/api.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/benchmark/CMakeLists.txt 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/benchmark/op_tester.cc 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/benchmark/op_tester.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/benchmark/op_tester_config.cc 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/benchmark/op_tester_config.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/conv_op.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/expand_op.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/expand_op.cu
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/fc_op.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/jit/gen/embseqpool.cc 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/jit/gen/embseqpool.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/jit/helper.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/jit/test.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/lstm_op.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/lstmp_op.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/lstmp_op.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/math/sample_prob.cu 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/math/sample_prob.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/ngraph/ops/CMakeLists.txt 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/ngraph/ops/activation_op.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/ngraph/ops/batch_norm_op.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/ngraph/ops/cross_entropy_op.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/ngraph/ops/momentum_op.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/ngraph/ops/op_bridge.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/ngraph/ops/sum_op.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/pool_op.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/sample_logits_op.cc 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/sample_logits_op.cu 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/sample_logits_op.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/operators/shape_op.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/platform/enforce.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/platform/event.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/platform/init.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/platform/mkldnn_utils.h 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/platform/place.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/platform/profiler.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/platform/profiler.cu 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/platform/profiler.h
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/pybind/ir.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ b8491bfd
此差异已折叠。
paddle/fluid/train/demo/README.md
浏览文件 @ b8491bfd
此差异已折叠。
paddle/scripts/README.md
浏览文件 @ b8491bfd
此差异已折叠。
paddle/scripts/fast_install.sh
浏览文件 @ b8491bfd
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ b8491bfd
此差异已折叠。
paddle/scripts/submit_local.sh.in
浏览文件 @ b8491bfd
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/__init__.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/distributed/__init__.py 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/compiler.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/contrib/slim/tests/CMakeLists.txt 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/contrib/slim/tests/test_graph.py 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/initializer.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/layer_helper.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/layers/ops.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/profiler.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/tests/unittests/test_base_layer.py 0 → 100644
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/fluid/unique_name.py
浏览文件 @ b8491bfd
此差异已折叠。
python/paddle/utils/dump_v2_config.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
此差异已折叠。
python/paddle/utils/image_multiproc.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
此差异已折叠。
python/paddle/utils/make_model_diagram.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
此差异已折叠。
python/paddle/utils/merge_model.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
此差异已折叠。
python/paddle/utils/predefined_net.py 已删除 100644 → 0
浏览文件 @ 8bda4ab2
此差异已折叠。
python/requirements.txt
浏览文件 @ b8491bfd
此差异已折叠。
python/setup.py.in
浏览文件 @ b8491bfd
此差异已折叠。
tools/manylinux1/Dockerfile.x64
浏览文件 @ b8491bfd
此差异已折叠。
tools/manylinux1/build_all.sh
浏览文件 @ b8491bfd
此差异已折叠。
tools/timeline.py
浏览文件 @ b8491bfd
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或