Skip to content

P
PaddleDetection

s920243400 / PaddleDetection
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleDetection

通知 2
Star 0
Fork 0
P
PaddleDetection
提交 2b47fb3d 编写于 作者: T typhoonzero
浏览文件
操作

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

上级 9508c726 0a8addf8
展开全部 隐藏空白更改
内联 并排
Showing with 7985 addition and 3423 deletion
.gitignore
浏览文件 @ 2b47fb3d
......@@ -28,3 +28,4 @@ cmake_install.cmake
paddle/.timestamp
python/paddlepaddle.egg-info/
paddle/pybind/pybind.h
python/paddle/version.py
CMakeLists.txt
浏览文件 @ 2b47fb3d
......@@ -22,6 +22,8 @@ SET(CMAKE_C_FLAGS_RELWITHDEBINFO "-O3 -g -DNDEBUG")
include(system)
project(paddle CXX C Go)
message(STATUS "CXX compiler: " ${CMAKE_CXX_COMPILER} ", version: " ${CMAKE_CXX_COMPILER_VERSION})
message(STATUS "C compiler: " ${CMAKE_C_COMPILER} ", version: " ${CMAKE_C_COMPILER_VERSION})
find_package(Sphinx)
if(NOT CMAKE_CROSSCOMPILING)
......@@ -58,6 +60,7 @@ option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF)
option(WITH_DISTRIBUTE "Compile with grpc distributed support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
# CMAKE_BUILD_TYPE
if(NOT CMAKE_BUILD_TYPE)
......
RELEASE.cn.md
浏览文件 @ 2b47fb3d
# v0.11.0版本
## PaddlePaddle Fluid
- PaddlePaddle发布版本v0.11.0包含一个新的特性*PaddlePaddle Fluid*. Fluid 是设计用来让用户像Pytorch和Tensorflow Eager Execution一样执行程序。在这些系统中，不再有*模型*这个概念，应用也不再包含一个用于描述Operator图或者一系列层的符号描述，而是像通用程序那样描述训练或者预测的过程。而Fluid与PyTorch或Eager Execution的区别在于Fluid不依赖Python提供的控制流，例如 if-else-then或者for，而是提供了基于C++实现的控制流并暴露了对应的用with语法实现的Python接口。例如：
https://github.com/PaddlePaddle/Paddle/blob/3df78ed2a98d37f7ae6725894cc7514effd5664b/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44
- 在v0.11.0版本中，我们提供了一个C++类`Executor`用于运行一个Fluid程序。Executor类似一个解释器。在未来的版本中，我们将提升和优化Executor成为一个调试器，就像GDB。并可能提供一些编译器，这个编译器会读取一个上文所描述的应用然后编译成一个等价的
源代码，这个源代码可以被nvcc编译成可以使用CUDA的二进制，或者被icc编译成可以充分利用Intel CPU的二进制。
## 新特点
* 发布 `PaddlePaddle Fluid`。
* 增加了用于模型预测的C-API。
* 用Fluid API实现了一个简单的GAN的例子。
* 增加了关于性能调优的文档。
* 为`paddle.v2.dataset`下载数据集提供了重试机制.
* C++中使用protobuf-lite替换protobuf减少了二进制的大小。
* 发布了新特性 [Elastic Deep Learning (EDL)](https://github.com/PaddlePaddle/cloud/tree/develop/doc/autoscale/experiment).
* 基于Bazel API利用cmake实现了一个的新的构建系统函数库。
* 当使用编译选项`WITH_MKL=ON`时自动下载和编译Intel® [MKLML](https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz) 函数库.
* [Intel® MKL-DNN on PaddlePaddle](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn):
- 完成了 11个 MKL-DNN 层: Convolution, Fully connectivity, Pooling, ReLU, Tanh, ELU, Softmax, BatchNorm, AddTo, Concat, LRN。
- 完成了 3个 MKL-DNN 网络: VGG-19, ResNet-50, GoogleNet
- 基于Intel Skylake 6148 CPU的[性能测试](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md) : 相对于MKLML有2~3倍的训练加速。
* 增加 [softsign activation](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/activation.html#softsign)
* 增加 [dot product layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#dot-prod)
* 增加 [L2 distance layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#l2-distance)
* 增加 [sub-nested sequence layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#sub-nested-seq)
* 增加 [kmax sequence score layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#kmax-sequence-score)
* 增加 [sequence slice layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#seq-slice)
* 增加 [row convolution layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#row-conv)
* 增加移动端友好的网页
## 改进
* 使用一个Python`whl`包即可安装.
* [V2 API可以实现用户定制化评估](https://github.com/PaddlePaddle/models/tree/develop/ltr#训练过程中输出自定义评估指标)。
* 将 `PADDLE_ONLY_CPU` 改为 `PADDLE_WITH_GPU`, 因为我们会支持多种设备。
* 删除了有一些bug的BarrierStat。
* 清理和删除了paddle::Parameter中未使用的函数。
* 删除了ProtoDataProvider。
* Huber loss同时支持回归和分类。
* 为sequence pooling 层增加`stride`参数。
* v2 API自动使用cudnn batch normalization。
* 可以使用一个固定的参数名共享BN层的参数。
* 2D convolution operation支持variable-dimension input特性。
* 重构cmake中关于CUDA的部分并实现自动检测GPU架构的功能。
* 优化网页导航。
## 错误修复
* 修复ROI pooling的Bug. cc9a761
* 修复当label是dense vector是AUC变成0的问题. #5274
* 修复WarpCTC 层的Bug.
# v0.10.0版本
我们非常高兴发布了PaddlePaddle V0.10.0版，并开发了新的[Python API](http://research.baidu.com/paddlepaddles-new-api-simplifies-deep-learning-programs/)。
......
RELEASE.md
浏览文件 @ 2b47fb3d
# Release v0.11.0
## PaddlePaddle Fluid
- Release 0.11.0 includes a new feature *PaddlePaddle Fluid*. Fluid is
designed to allow users to program like PyTorch and TensorFlow Eager Execution.
In these systems, there is no longer the concept *model* and applications
do not include a symbolic description of a graph of operators nor a sequence
of layers. Instead, applications look exactly like a usual program that
describes a process of training or inference. The difference between
Fluid and PyTorch or Eager Execution is that Fluid doesn't rely on Python's
control-flow, `if-then-else` nor `for`. Instead, Fluid provides its
C++ implementations and their Python binding using the `with` statement. For an example
https://github.com/PaddlePaddle/Paddle/blob/3df78ed2a98d37f7ae6725894cc7514effd5664b/python/paddle/v2/fluid/tests/test_while_op.py#L36-L44
- In 0.11.0, we provides a C++ class `Executor` to run a Fluid program.
Executor works like an interpreter. In future version, we will improve
`Executor` into a debugger like GDB, and we might provide some compilers,
which, for example, takes an application like the above one, and outputs
an equivalent C++ source program, which can be compiled using
[`nvcc`](http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html)
to generate binaries that use CUDA, or using
[`icc`](https://software.intel.com/en-us/c-compilers) to generate binaries
that make full use of Intel CPUs.
## New Features
* Release `PaddlePaddle Fluid`.
* Add C-API for model inference
* Use fluid API to create a simple GAN demo.
* Add develop guide about performance tunning.
* Add retry when download `paddle.v2.dataset`.
* Linking protobuf-lite not protobuf in C++. Reduce the binary size.
* Feature [Elastic Deep Learning (EDL)](https://github.com/PaddlePaddle/cloud/tree/develop/doc/autoscale/experiment) released.
* A new style cmake functions for Paddle. It is based on Bazel API.
* Automatically download and compile with Intel® [MKLML](https://github.com/01org/mkl-dnn/releases/download/v0.11/mklml_lnx_2018.0.1.20171007.tgz) library as CBLAS when build `WITH_MKL=ON`.
* [Intel® MKL-DNN on PaddlePaddle](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/design/mkldnn):
- Complete 11 MKL-DNN layers: Convolution, Fully connectivity, Pooling, ReLU, Tanh, ELU, Softmax, BatchNorm, AddTo, Concat, LRN.
- Complete 3 MKL-DNN networks: VGG-19, ResNet-50, GoogleNet
- [Benchmark](https://github.com/PaddlePaddle/Paddle/blob/develop/benchmark/IntelOptimizedPaddle.md) on Intel Skylake 6148 CPU: 2~3x training speedup compared with MKLML.
* Add the [`softsign` activation](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/activation.html#softsign).
* Add the [dot product layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#dot-prod).
* Add the [L2 distance layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#l2-distance).
* Add the [sub-nested sequence layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#sub-nested-seq).
* Add the [kmax sequence score layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#kmax-sequence-score).
* Add the [sequence slice layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#seq-slice).
* Add the [row convolution layer](http://www.paddlepaddle.org/docs/develop/documentation/zh/api/v2/config/layer.html#row-conv)
* Add mobile friendly webpages.
## Improvements
* Build and install using a single `whl` package.
* [Custom evaluating in V2 API](https://github.com/PaddlePaddle/models/tree/develop/ltr#训练过程中输出自定义评估指标).
* Change `PADDLE_ONLY_CPU` to `PADDLE_WITH_GPU`, since we will support many kinds of devices.
* Remove buggy BarrierStat.
* Clean and remove unused functions in paddle::Parameter.
* Remove ProtoDataProvider.
* Huber loss supports both regression and classification.
* Add the `stride` parameter for sequence pooling layers.
* Enable v2 API use cudnn batch normalization automatically.
* The BN layer's parameter can be shared by a fixed the parameter name.
* Support variable-dimension input feature for 2D convolution operation.
* Refine cmake about CUDA to automatically detect GPU architecture.
* Improved website navigation.
## Bug Fixes
* Fix bug in ROI pooling. cc9a761
* Fix AUC is zero when label is dense vector. #5274
* Fix bug in WarpCTC layer.
# Release v0.10.0
We are glad to release version 0.10.0. In this version, we are happy to release the new
......
benchmark/IntelOptimizedPaddle.md
浏览文件 @ 2b47fb3d
......@@ -2,27 +2,25 @@
Machine:
- Server
- Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz, 2 Sockets, 20 Cores per socket
- Laptop
- DELL XPS15-9560-R1745: i7-7700HQ 8G 256GSSD
- i5 MacBook Pro (Retina, 13-inch, Early 2015)
- Desktop
- i7-6700k
- 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:latest (for MKLML and MKL-DNN), paddlepaddle/paddle:latest-openblas (for OpenBLAS)
- MKL-DNN tag v0.11
- MKLML 2018.0.1.20171007
- OpenBLAS v0.2.20
(TODO: will rerun after 0.11.0)
PaddlePaddle: (TODO: will rerun after 0.11.0)
- paddlepaddle/paddle:latest (for MKLML and MKL-DNN)
- MKL-DNN tag v0.11
- MKLML 2018.0.1.20171007
- paddlepaddle/paddle:latest-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
Input image size - 3 * 224 * 224, Time: images/second
......@@ -35,9 +33,7 @@ Input image size - 3 * 224 * 224, Time: images/second
| MKLML | 12.12 | 13.70 | 16.18 |
| MKL-DNN | 28.46 | 29.83 | 30.44 |
chart on batch size 128
TBD
<img src="figs/vgg-cpu-train.png" width="500">
- ResNet-50
......@@ -47,9 +43,7 @@ TBD
| MKLML | 32.52 | 31.89 | 33.12 |
| MKL-DNN | 81.69 | 82.35 | 84.08 |
chart on batch size 128
TBD
<img src="figs/resnet-cpu-train.png" width="500">
- GoogLeNet
......@@ -59,10 +53,35 @@ TBD
| MKLML | 128.46| 137.89| 158.63 |
| MKL-DNN     | 250.46| 264.83| 269.50 |
chart on batch size 128
TBD
<img src="figs/googlenet-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 | 1.07 | 1.08 | 1.06 | 0.88 | 0.65 |
| MKLML | 5.58 | 9.80 | 15.15 | 21.21 | 28.67 |
| MKL-DNN | 75.07 | 88.64 | 82.58 | 92.29 | 96.75 |
- ResNet-50
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|-------|--------|--------|--------|--------|
| OpenBLAS | 3.35 | 3.19 | 3.09 | 2.55 | 1.96 |
| MKLML | 6.33 | 12.02 | 22.88 | 40.53 | 63.09 |
| MKL-DNN | 107.83| 148.84 | 177.78 | 189.35 | 217.69 |
- GoogLeNet
| BatchSize | 1 | 2 | 4 | 8 | 16 |
|-----------|--------|--------|--------|--------|--------|
| OpenBLAS | 12.04 | 11.31 | 10.00 | 9.07 | 4.34 |
| MKLML | 22.74 | 41.56 | 81.22 | 133.47 | 210.53 |
| MKL-DNN | 175.10 | 272.92 | 450.70 | 512.00 | 600.94 |
### Laptop
TBD
### Desktop
TBD
cmake/cblas.cmake
浏览文件 @ 2b47fb3d
......@@ -17,7 +17,7 @@ if(WITH_MKLML AND MKLML_INC_DIR AND MKLML_LIB)
set(CBLAS_INC_DIR ${MKLML_INC_DIR})
set(CBLAS_LIBRARIES ${MKLML_LIB})
add_definitions(-DPADDLE_USE_MKLML)
add_definitions(-DPADDLE_WITH_MKLML)
add_definitions(-DLAPACK_FOUND)
message(STATUS "Found cblas and lapack in MKLML "
......
cmake/configure.cmake
浏览文件 @ 2b47fb3d
......@@ -24,6 +24,11 @@ if(WITH_DOUBLE)
add_definitions(-DPADDLE_TYPE_DOUBLE)
endif(WITH_DOUBLE)
if(WITH_ARM_FP16)
add_definitions(-DPADDLE_ARM_FP16)
add_definitions("-march=armv8.2-a+fp16+simd")
endif(WITH_ARM_FP16)
if(WITH_TESTING)
add_definitions(-DPADDLE_WITH_TESTING)
endif(WITH_TESTING)
......
cmake/external/cares.cmake
浏览文件 @ 2b47fb3d
......@@ -33,7 +33,7 @@ ExternalProject_Add(
UPDATE_COMMAND ""
CONFIGURE_COMMAND ./buildconf && ./configure --disable-shared --prefix=${CARES_INSTALL_DIR}
BUILD_IN_SOURCE 1
BUILD_COMMAND make
BUILD_COMMAND make -j8
INSTALL_COMMAND make install
)
......
cmake/external/mkldnn.cmake
浏览文件 @ 2b47fb3d
......@@ -67,5 +67,5 @@ ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
MESSAGE(STATUS "MKLDNN library: ${MKLDNN_LIB}")
add_definitions(-DPADDLE_USE_MKLDNN)
add_definitions(-DPADDLE_WITH_MKLDNN)
LIST(APPEND external_project_dependencies mkldnn)
cmake/external/openblas.cmake
浏览文件 @ 2b47fb3d
......@@ -114,11 +114,7 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
# linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
ADD_LIBRARY(cblas SHARED ${dummyfile})
ELSE()
ADD_LIBRARY(cblas STATIC ${dummyfile})
ENDIF()
ADD_LIBRARY(cblas STATIC ${dummyfile})
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
IF(NOT ${CBLAS_FOUND})
......
doc/api/v2/config/activation.rst
浏览文件 @ 2b47fb3d
......@@ -99,3 +99,10 @@ STanh
.. automodule:: paddle.v2.activation
:members: STanh
:noindex:
SoftSign
========
.. automodule:: paddle.v2.activation
:members: SoftSign
:noindex:
doc/design/refactor/distributed_architecture.md
浏览文件 @ 2b47fb3d
......@@ -53,7 +53,7 @@ The IR for PaddlePaddle after refactoring is called a `Block`, it specifies the
The user can not directly specify the parameter update rule for the parameter server in the Python module, since the parameter server does not use the same computation definition as the trainer. Instead, the update rule is baked inside the parameter server. The user can not specify the update rule explicitly.
This could be fixed by making the parameter server run the same computation definition as the trainer (the user's Python module). For a detailed explanation, refer to this document -
[Design Doc: Operation Graph Based Parameter Server](./dist_train.md)
[Design Doc: Operation Graph Based Parameter Server](./parameter_server.md)
## Distributed Training Architecture
......
doc/design/releasing_process.md
浏览文件 @ 2b47fb3d
......@@ -5,8 +5,9 @@ PaddlePaddle使用git-flow branching model做分支管理，使用[Semantic Vers
PaddlePaddle每次发新的版本，遵循以下流程:
1. 从`develop`分支派生出新的分支，分支名为`release/版本号`。例如，`release/0.10.0`
2. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
3. 对这个版本的提交，做如下几个操作:
1. 将新分支的版本打上tag，tag为`版本号rc.Patch号`。第一个tag为`0.10.0rc1`，第二个为`0.10.0rc2`，依次类推。
1. 对这个版本的提交，做如下几个操作:
* 修改`python/setup.py.in`中的版本信息,并将`istaged`字段设为`True`。
* 编译这个版本的Docker发行镜像，发布到dockerhub。如果失败，修复Docker编译镜像问题，Patch号加一，返回第二步
* 编译这个版本的Ubuntu Deb包。如果失败，修复Ubuntu Deb包编译问题，Patch号加一，返回第二步。
* 使用Regression Test List作为检查列表，测试Docker镜像/ubuntu安装包的功能正确性
......@@ -20,9 +21,9 @@ PaddlePaddle每次发新的版本，遵循以下流程:
pip install twine
twine upload dist/[package to upload]
```
4. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
5. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面
6. 协同完成Release Note的书写
1. 第三步完成后，将`release/版本号`分支合入master分支，并删除`release/版本号`分支。将master分支的合入commit打上tag，tag为`版本号`。同时再将`master`分支合入`develop`分支。最后删除`release/版本号`分支。
1. 编译master分支的Docker发行镜像，发布到dockerhub。编译ubuntu的deb包，发布到github release页面
1. 协同完成Release Note的书写
需要注意的是:
......@@ -30,7 +31,7 @@ PaddlePaddle每次发新的版本，遵循以下流程:
* `release/版本号`分支一旦建立，一般不允许再从`develop`分支合入`release/版本号`。这样保证`release/版本号`分支功能的封闭，方便测试人员测试PaddlePaddle的行为。
* 在`release/版本号`分支存在的时候，如果有bugfix的行为，需要将bugfix的分支同时merge到`master`, `develop`和`release/版本号`这三个分支。
# PaddlePaddle 分支规范
## PaddlePaddle 分支规范
PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-branching-model/)分支规范，并适应github的特性做了一些区别。
......@@ -47,11 +48,11 @@ PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-
* BugFix分支也是在开发者自己的fork版本库维护，与功能分支不同的是，BugFix分支需要分别给主版本库的`master`、`develop`与可能有的`release/版本号`分支，同时提起`Pull Request`。
# PaddlePaddle回归测试列表
## PaddlePaddle回归测试列表
本列表说明PaddlePaddle发版之前需要测试的功能点。
## PaddlePaddle Book中所有章节
### PaddlePaddle Book中所有章节
PaddlePaddle每次发版本首先要保证PaddlePaddle Book中所有章节功能的正确性。功能的正确性包括验证PaddlePaddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。
......
doc/getstarted/build_and_install/build_from_source_cn.rst
浏览文件 @ 2b47fb3d
......@@ -19,7 +19,7 @@ PaddlePaddle主要使用 `CMake <https://cmake.org>`_ 以及GCC, G++作为编译
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
# 如果使用Docker编译环境，执行下面的命令编译CPU-Only的二进制
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/scripts/docker/build.sh
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh
# 如果不使用Docker编译环境，执行下面的命令
mkdir build
cd build
......@@ -30,7 +30,7 @@ PaddlePaddle主要使用 `CMake <https://cmake.org>`_ 以及GCC, G++作为编译
.. code-block:: bash
pip install python/dist/*.whl
pip install build/python/dist/*.whl
.. _run_test:
......@@ -45,7 +45,7 @@ PaddlePaddle主要使用 `CMake <https://cmake.org>`_ 以及GCC, G++作为编译
.. code-block:: bash
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/scripts/docker/build.sh
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh
如果不使用Docker，可以执行ctest命令即可：
......
doc/getstarted/build_and_install/build_from_source_en.rst
浏览文件 @ 2b47fb3d
......@@ -21,7 +21,7 @@ Then run:
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
# run the following command to build a CPU-Only binaries if you are using docker
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/scripts/docker/build.sh
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x /paddle/paddle/scripts/docker/build.sh
# else run these commands
mkdir build
cd build
......@@ -34,7 +34,7 @@ machine or copy it to the target machine.
.. code-block:: bash
pip install python/dist/*.whl
pip install build/python/dist/*.whl
.. _run_test:
......@@ -49,7 +49,7 @@ Set :code:`WITH_GPU=ON` Can also run tests on GPU.
.. code-block:: bash
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/scripts/docker/build.sh
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=ON" -e "RUN_TEST=ON" paddlepaddle/paddle_manylinux_devel:cuda8.0_cudnn5 bash -x paddle/paddle/scripts/docker/build.sh
If you don't use Docker, just run ctest will start the tests:
......@@ -117,7 +117,7 @@ You can add :code:`-D` argument to pass such options, like:
"WITH_PYTHON", "Build with integrated Python interpreter", "ON"
"WITH_STYLE_CHECK", "Check code style when building", "ON"
"WITH_TESTING", "Build unit tests", "ON"
"WITH_DOC", "Build documentaions", "OFF"
"WITH_DOC", "Build documentations", "OFF"
"WITH_SWIG_PY", "Build Python SWIG interface for V2 API", "Auto"
"WITH_GOLANG", "Build fault-tolerant parameter server written in go", "ON"
"WITH_MKL", "Use MKL as BLAS library, else use OpenBLAS", "ON"
......
doc/getstarted/build_and_install/index_cn.rst
浏览文件 @ 2b47fb3d
......@@ -13,7 +13,7 @@ PaddlePaddle提供pip和Docker的安装方式：
pip_install_cn.rst
docker_install_cn.rst
../../howto/dev/build_cn.md
编译流程
++++++++
......
doc/getstarted/build_and_install/index_en.rst
浏览文件 @ 2b47fb3d
......@@ -13,6 +13,7 @@ You can choose either pip or Docker to complete your install:
pip_install_en.rst
docker_install_en.rst
../../howto/dev/build_en.md
Build from Source
......
doc/howto/dev/build_cn.md
浏览文件 @ 2b47fb3d
# 编译PaddlePaddle和运行单元测试
# 用Docker编译和测试PaddlePaddle
## 需要的软硬件
......
doc/howto/dev/build_en.md
浏览文件 @ 2b47fb3d
# Build PaddlePaddle from Source Code and Run Unit Test
# Build using Docker
## What Developers Need
......
doc/howto/dev/contribute_to_paddle_cn.md
浏览文件 @ 2b47fb3d
......@@ -76,18 +76,18 @@ no changes added to commit (use "git add" and/or "git commit -a")
## 构建和测试
编译 PaddlePaddle 的源码以及生成文档需要多种开发工具。为了方便大家，我们的标准开发流程是把这些工具都装进一个Docker image，称为*开发镜像*，通常名字是 `paddle:dev`。然后所有用 `cmake && make` 的地方（比如IDE配置里）都用 `docker run paddle:dev`来代替。
编译 PaddlePaddle 的源码以及生成文档需要多种开发工具。为了方便大家，我们的标准开发流程是把这些工具都装进一个Docker image，称为*开发镜像*，通常名字是 `paddle:latest-dev` 或者 `paddle:[version tag]-dev` 如 `paddle:0.11.0-dev`。然后所有用 `cmake && make` 的地方（比如IDE配置里）都用 `docker run paddle:latest-dev`来代替。
如要build这个开发镜像，在源码目录树的根目录中运行：
```bash
➜ docker build -t paddle:dev .
➜ docker build -t paddle:latest-dev .
```
随后可以用这个开发镜像开始build PaddlePaddle的源码。比如如果要build一个不依赖GPU，但是支持AVX指令集，并且包括unit tests的PaddlePaddle，可以：
```bash
➜ docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TEST=ON" paddle:dev
➜ docker run -v $(pwd):/paddle -e "WITH_GPU=OFF" -e "WITH_AVX=ON" -e "WITH_TESTING=ON" paddle:latest-dev
```
这个过程除了编译PaddlePaddle为 `./build/libpaddle.so`，并且输出一个 `./build/paddle.deb`文件之外，还会输出一个 `build/Dockerfile`。我们只需要运行下面命令把编译好的PaddlePaddle打包成一个*生产镜像*（`paddle:prod`）：
......@@ -99,7 +99,7 @@ no changes added to commit (use "git add" and/or "git commit -a")
如果要运行所有的单元测试，可以用如下命令：
```bash
➜ docker run -it -v $(pwd):/paddle paddle:dev bash -c "cd /paddle/build && ctest"
➜ docker run -it -v $(pwd):/paddle paddle:latest-dev bash -c "cd /paddle/build && ctest"
```
关于构建和测试的更多信息，请参见[这篇文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/getstarted/build_and_install/docker_install_cn.rst)。
......
doc/howto/dev/new_op_cn.md
浏览文件 @ 2b47fb3d
# 如何写新的Operator
- [概念简介](#概念简介)
- [实现C++类](#实现C++类)
- [定义ProtoMaker类](#定义ProtoMaker类)
- [定义Operator类](#定义Operator类)
- [定义OpKernel类](#定义OpKernel类)
- [注册Operator](#注册Operator)
- [实现C++类](#实现c类)
- [定义ProtoMaker类](#定义protomaker类)
- [定义Operator类](#定义operator类)
- [定义OpKernel类](#定义opkernel类)
- [注册Operator](#注册operator)
- [编译](#编译)
- [绑定Python](#绑定Python)
- [绑定Python](#绑定python)
- [实现单元测试](#实现单元测试)
- [前向Operator单测](#前向Operator单测)
- [反向Operator单测](#反向Operator单测)
- [前向Operator单测](#前向operator单测)
- [反向Operator单测](#反向operator单测)
- [编译和执行](#编译和执行)
- [注意事项](#注意事项)
## 概念简介
......@@ -30,8 +31,8 @@
-------------- | :----------------------
OpProtoMake定义 | `.cc`文件，Backward Op不需要定义OpProtoMake
Op定义 | `.cc`文件
Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，GPU 实现在`.cu`文件中。
注册Op | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，GPU实现在`.cu`文件中
Kernel实现 | CPU、CUDA共享Kernel实现在`.h`文件中，否则，CPU 实现在`.cc`文件中，CUDA 实现在`.cu`文件中。
注册Op | Op注册实现在`.cc`文件；Kernel注册CPU实现在`.cc`文件中，CUDA实现在`.cu`文件中
实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下，文件命名以`*_op.h`（如有） 、 `*_op.cc` 、`*_op.cu`（如有）结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
......@@ -43,7 +44,7 @@ Kernel实现 | CPU、GPU共享Kernel实现在`.h`文件中，否则，CPU
## 实现C++类
### 1. 定义ProtoMaker类
### 定义ProtoMaker类
矩阵乘法的公式：$Out = X * Y$, 可见该计算由两个输入，一个输出组成。
......@@ -100,7 +101,7 @@ The equation is: Out = scale*X
- `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。
### 2. 定义Operator类
### 定义Operator类
下面的点实现了MulOp的定义：
......@@ -149,11 +150,11 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
通常`OpProtoMaker`和`Op`类的定义写在`.cc`文件中，和下面将要介绍的注册函数一起放在`.cc`中
### 3. 定义OpKernel类
### 定义OpKernel类
`MulKernel`继承自`framework::OpKernel`，带有下面两个模板参数:
- `typename Place`: 表示设备类型，不同设备(CPU、GPU)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
- `typename DeviceContext`: 表示设备类型，不同设备(CPU、CUDA)共享同一个Kernel时，需加该模板参数，不共享则不加，一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
- `typename T` : 表示数据类型，如`float`, `double`等。
......@@ -165,7 +166,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
下面是 `MulKernel` `Compute`的实现：
```cpp
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class MulKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
......@@ -173,33 +174,32 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
auto* Y = context.Input<Tensor>("Y");
auto* Z = context.Output<Tensor>("Out");
Z->mutable_data<T>(context.GetPlace());
auto* device_context =
const_cast<platform::DeviceContext*>(context.device_context_);
math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
auto& device_context = context.template device_context<DeviceContext>();
math::matmul<DeviceContext, T>(*X, false, *Y, false, 1, Z, 0, device_context);
}
};
```
需要注意：**不同设备(CPU、GPU)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
需要注意：**不同设备(CPU、CUDA)共享一个Op定义，是否则共享同一个`OpKernel`，取决于`Compute`调用的函数是否支持不同设备。**
`MulOp`的CPU、GPU实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
`MulOp`的CPU、CUDA实现共享同一个`Kernel`。`OpKernel`不共享的例子可以参考：[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)。
为了使`OpKernel`的计算过程书写更加简单，并且CPU、GPU的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。
为了使`OpKernel`的计算过程书写更加简单，并且CPU、CUDA的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md)。
到此，前向Op实现完成。接下来，需要在`.cc`文件中注册该op和kernel。
反向Op类的定义，反向OpKernel的定义与前向Op类似，这里不再赘述。**但需注意反向Op没有`ProtoMaker`**。
### 4. 注册Operator
### 注册Operator
- 在`.cc`文件中注册前向、反向Op类，注册CPU Kernel。
```cpp
namespace ops = paddle::operators;
REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::CPUPlace, float>);
ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>);
```
在上面的代码中：
......@@ -209,20 +209,20 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
- `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulGradKernel`类。
- 在 `.cu`文件中注册GPU Kernel。
- 请注意，如果GPU Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下：
- 在 `.cu`文件中注册CUDA Kernel。
- 请注意，如果CUDA Kernel的实现基于Eigen unsupported模块，那么在 `.cu`的开始请加上宏定义 `#define EIGEN_USE_GPU`，代码示例如下：
```cpp
// if use Eigen unsupported module before include head files
// #define EIGEN_USE_GPU
#define EIGEN_USE_GPU
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel<paddle::platform::CUDADeviceContext, float>);
REGISTER_OP_CUDA_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
```
### 5. 编译
### 编译
运行下面命令可以进行编译：
......@@ -236,71 +236,57 @@ make mul_op
## 实现单元测试
单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
单测包括对比前向Op不同设备(CPU、CUDA)的实现、对比反向OP不同设备(CPU、CUDA)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单元测试](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)。
### 前向Operator单元测试
### 前向Operator单测
前向Op单元测试继承自`unittest.TestCase`，并定义元类`__metaclass__ = OpTestMeta`。各项更加具体的单元测试在`OpTestMeta`里完成。测试前向Operator，需要：
Op单元测试继承自`OpTest`。各项更加具体的单元测试在`TestMulOp`里完成。测试Operator，需要：
1. 在`setUp`函数定义输入、输出，以及相关的属性参数。
2. 生成随机的输入数据。
3. 在Python脚本中实现与前向operator相同的计算逻辑，得到输出值，与operator前向计算的输出进行对比。
4. 反向计算已经自动集成进测试框架，直接调用相应接口即可。
```python
import unittest
import numpy as np
from gradient_checker import GradientChecker, create_op
from op_test_util import OpTestMeta
from op_test import OpTest
class TestMulOp(unittest.TestCase):
__metaclass__ = OpTestMeta
class TestMulOp(OpTest):
def setUp(self):
self.type = "mul"
self.op_type = "mul"
self.inputs = {
'X': np.random.random((32, 84)).astype("float32"),
'Y': np.random.random((84, 100)).astype("float32")
}
self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
```
上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
- `self.type = "mul" ` : 定义类型，与operator注册时注册的类型一致。
- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
def test_check_output(self):
self.check_output()
### 反向Operator单元测试
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
反向Op单元测试继承自`GradientChecker`，而`GradientChecker`继承自`unittest.TestCase`，因此，**反向单元测试函数需要以`test_`开头**。
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
```python
class TestMulGradOp(GradientChecker):
def setUp(self):
self.op = create_op("mul")
self.inputs = {
'X': np.random.random((32, 84)).astype("float32"),
'Y': np.random.random((84, 100)).astype("float32")
}
def test_check_grad_normal(self):
# mul op will enlarge the relative error
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
```
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
上面的代码首先导入依赖的包，下面是对`setUp`函数中操作的重要变量的详细解释：
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
```
- `self.op_type = "mul" ` : 定义类型，与operator注册时注册的类型一致。
- `self.inputs` : 定义输入，类型为`numpy.array`，并初始化。
- `self.outputs` : 定义输出，并在Python脚本中完成与operator同样的计算逻辑，返回Python端的计算结果。
下面解释代码中一些关键的地方:
### 反向operator单测
- 调用`create_op("mul")`创建反向Op对应的前向Op。
而反向测试中：
- `test_check_grad_normal`中调用`check_grad`使用数值法检测梯度正确性和稳定性。
- 第一个参数`["X", "Y"]` : 指定对输入变量`X`、`Y`做梯度检测。
- 第二个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`。
......@@ -308,7 +294,7 @@ class TestMulGradOp(GradientChecker):
- `test_check_grad_ingore_x`和`test_check_grad_ingore_y`分支用来测试只需要计算一个输入梯度的情况。
### 编译和执行单元测试
### 编译和执行
`python/paddle/v2/framework/tests` 目录下新增的 `test_*.py` 单元测试会被自动加入工程进行编译。
......@@ -328,5 +314,5 @@ ctest -R test_mul_op
- 为每个Op创建单独的`*_op.h`（如有）、`*_op.cc`和`*_op.cu`（如有）。不允许一个文件中包含多个Op，这将会导致编译出错。
- 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OP(B, ...)`等，这将会导致单元测试出错。
- 如果Op没有实现GPU Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
- 如果Op没有实现CUDA Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。
- 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。
doc/howto/dev/new_op_en.md
浏览文件 @ 2b47fb3d
# How to write a new operator
- [Background](#background)
- [Implementing C++ Types](#implementing-c++-types)
- [Defining ProtoMaker](#defining-protoMaker)
- [Implementing C++ Types](#implementing-c-types)
- [Defining ProtoMaker](#defining-protomaker)
- [Defining Operator](#defining-operator)
- [Registering Operator](#registering-operator)
- [Compilation](#compilation)
......@@ -28,8 +28,8 @@ An operator can be differentiated by whether in has kernel methods. An operator
-------------- | :----------------------
OpProtoMake definition | `.cc`files, Backward Op does not need an OpProtoMake interface.
Op definition | `.cc` files
Kernel implementation | The kernel methods shared between CPU and GPU are defined in `.h` files. CPU-specific kernels live in `.cc` files, while GPU-specific kernels are implemented in `.cu`files.
Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the GPU implementation.
Kernel implementation | The kernel methods shared between CPU and CUDA are defined in `.h` files. CPU-specific kernels live in `.cc` files, while CUDA-specific kernels are implemented in `.cu`files.
Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation.
New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions. **
......@@ -41,7 +41,7 @@ Let's take matrix multiplication operator, [MulOp](https://github.com/PaddlePadd
## Implementing C++ Types
### 1. Defining Class ProtoMaker
### Defining ProtoMaker
Matrix Multiplication can be written as $Out = X * Y$, meaning that the operation consists of two inputs and pne output.
......@@ -98,7 +98,7 @@ There are two changes in this example:
- `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` adds `scale`constant as an attribute, and sets the default value to 1.0.
### 2. Defining Operator
### Defining Operator
The following code defines the interface for MulOp:
......@@ -147,11 +147,11 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, which also include the registration methods introduced later.
### 3. Defining OpKernel
### Defining OpKernel
`MulKernel` inherits `framework::OpKernel`, which includes the following templates:
- `typename Place` denotes device type. When different devices, namely the CPU and the GPU, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
- `typename DeviceContext` denotes device context type. When different devices, namely the CPUDeviceContext and the CUDADeviceContext, share the same kernel, this template needs to be added. If they don't share kernels, this must not be added. An example of a non-sharing kernel is [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
- `typename T` denotes data type, such as `float` or `double`.
......@@ -163,7 +163,7 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
`MulKernel`'s implementation of `Compute` is as follows:
```cpp
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class MulKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
......@@ -171,16 +171,15 @@ Usually `OpProtoMaker` and `Op`'s type definitions are written in `.cc` files, w
auto* Y = context.Input<Tensor>("Y");
auto* Z = context.Output<Tensor>("Out");
Z->mutable_data<T>(context.GetPlace());
auto* device_context =
const_cast<platform::DeviceContext*>(context.device_context_);
math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
auto& device_context = context.template device_context<DeviceContext>();
math::matmul<DeviceContext, T>(*X, false, *Y, false, 1, Z, 0, device_context);
}
};
```
Note that **different devices (CPU, GPU)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.**
Note that **different devices (CPU, CUDA)share an Op definition; whether or not they share the same `OpKernel` depends on whether `Compute` calls functions that support both devices.**
`MulOp`'s CPU and GPU share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
`MulOp`'s CPU and CUDA share the same `Kernel`. A non-sharing `OpKernel` example can be seen in [`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43).
To ease the writing of `OpKernel` compute, and for reusing code cross-device, [`Eigen-unsupported Tensor`](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md?fileviewer=file-view-default) module is used to implement `Compute` interface. To learn about how the Eigen library is used in PaddlePaddle, please see [usage document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/use_eigen_cn.md).
......@@ -189,16 +188,16 @@ This concludes the forward implementation of an operator. Next its operation and
The definition of its corresponding backward operator, if applicable, is similar to that of an forward operator. **Note that a backward operator does not include a `ProtoMaker`**.
### 4. Registering Operator
### Registering Operator
- In `.cc` files, register forward and backward operator classes and the CPU kernel.
```cpp
namespace ops = paddle::operators;
REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::CPUPlace, float>);
ops::MulGradKernel<paddle::platform::CPUDeviceContext, float>);
```
In that code block,
......@@ -208,20 +207,20 @@ The definition of its corresponding backward operator, if applicable, is similar
- `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`.
- Registering GPU Kernel in `.cu` files
- Note that if GPU Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as
- Registering CUDA Kernel in `.cu` files
- Note that if CUDA Kernel is implemented using the `Eigen unsupported` module, then on top of `.cu`, a macro definition `#define EIGEN_USE_GPU` is needed, such as
```cpp
// if use Eigen unsupported module before include head files
#define EIGEN_USE_GPU
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel<paddle::platform::CUDADeviceContext, float>);
REGISTER_OP_CUDA_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::CUDADeviceContext, float>);
```
### 5. Compilation
### Compilation
Run the following commands to compile.
......@@ -253,62 +252,51 @@ A forward operator unit test inherits `unittest.TestCase` and defines metaclass
2. Generating random input data.
3. Implementing the same computation logic in a Python script:
3. Implementing the same computation logic in a Python script.
4. Call check gradient function to check the backward operator.
```python
import unittest
import numpy as np
from gradient_checker import GradientChecker, create_op
from op_test_util import OpTestMeta
from op_test import OpTest
class TestMulOp(unittest.TestCase):
__metaclass__ = OpTestMeta
class TestMulOp(OpTest):
def setUp(self):
self.type = "mul"
self.op_type = "mul"
self.inputs = {
'X': np.random.random((32, 84)).astype("float32"),
'Y': np.random.random((84, 100)).astype("float32")
}
self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
```
Get its output, and compare it with the forward operator's own output.
The code above first loads required packages. In addition, we have
- `self.type = "mul" ` defines the type that is identical to what the operator's registered type.
- `self.op_type = "mul" ` defines the type that is identical to what the operator's registered type.
- `self.inputs` defines input, with type `numpy.array` and initializes it.
- `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script.
### Testing Backward Operators
A backward operator unit test inherits `GradientChecker`, which inherits `unittest.TestCase`. As a result, **a backward operator unit test needs to be have the prefix `test_`**.
```python
class TestMulGradOp(GradientChecker):
def setUp(self):
self.op = create_op("mul")
self.inputs = {
'X': np.random.random((32, 84)).astype("float32"),
'Y': np.random.random((84, 100)).astype("float32")
}
def test_check_grad_normal(self):
# mul op will enlarge the relative error
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
def test_check_grad_ingore_x(self):
self.check_grad(
['Y'], 'Out', max_relative_error=0.5, no_grad_set=set("X"))
def test_check_grad_ingore_y(self):
self.check_grad(
['X'], 'Out', max_relative_error=0.5, no_grad_set=set('Y'))
```
Some key points in the code above include:
Some key points in checking gradient above include:
- `create_op("mul")` creates the backward operator's corresponding forward operator.
- `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods.
- The first variable `["X", "Y"]` appoints `X` and `Y` to be scale tested.
- The second variable `"Out"` points to the network's final output target `Out`.
......@@ -338,5 +326,5 @@ ctest -R test_mul_op
- Every `*_op.h` (if applicable), `*_op.cc`, and `*_op.cu` (if applicable) must be created for a unique Op. Compiling will fail if multiple operators are included per file.
- The type with which an operator is registered needs to be identical to the Op's name. Registering `REGISTER_OP(B, ...)` in `A_op.cc` will cause unit testing failures.
- If the operator does not implement a GPU kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail.
- If the operator does not implement a CUDA kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail.
- If multiple operators rely on some shared methods, a file NOT named `*_op.*` can be created to store them, such as `gather.h`.
doc/howto/read_source.md 0 → 100644
浏览文件 @ 2b47fb3d
# PaddlePaddle Fluid Source Code Overview
Examples: https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/v2/fluid/tests/book
Core: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework
Operator: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators
Optimizer: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/optimizer
Memory: https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory
# Compile Time
The following **defines** the NN. The definition goes into this [protocol buffer](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto).
```python
x = fluid.layers.data(name='x', shape=[13], dtype='float32')
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
y_predict = fluid.layers.fc(input=x, size=1, act=None)
cost = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_cost = fluid.layers.mean(x=cost)
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
sgd_optimizer.minimize(avg_cost)
```
- Variables: `x`, `y`, `y_predict`, `cost` and `avg_cost`. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/framework.py#L93)
- Layers: `fluid.layers.data`, `fluid.layers.fc` and `fluid.layers.mean` are layers. [Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/layers.py)
- Every Layer has one or more operators and variables/parameters
- All the operators are defined at [`paddle/operators/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators). Other worth-looking files:
- Base class: [`paddle/framework/operator.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h)
- Operator Registration: [`paddle/framework/op_registry.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h)
- Operator Lookup: [`paddle/framework/op_info.h`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_info.h)
- Optimizer: `fluid.optimizer.SGD`. It does the following
- Add backward operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/backward.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/backward.cc)]
- Add optimizer operators. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/optimizer.py), [C++](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/optimizer)]
# Run Time
The following **evaluates** the NN. Instantiates all the variables, operators.
```python
place = fluid.CPUPlace()
feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe = fluid.Executor(place)
# Allocate memory. Initialize Parameter.
exe.run(fluid.default_startup_program())
# Allocate memory. Do computation.
exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[avg_cost])
```
- Place: `place`. one of CPU, GPU or FPGA. [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h)
- The device handle are at [paddle/platform/device_context.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/device_context.h)
- Executor: `fluid.Executor(place)`. [[Python](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/executor.py), [C++](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.cc)]
- Feeds the data: `feed=feeder.feed(data)`
- Evaluates all the operators
- Fetches the result: `fetch_list=[avg_cost]`
- Other worth looking files:
- Scope: [paddle/framework/scope.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/scope.h). Where all the variables live
- Variable: [paddle/framework/variable.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h). Where all the data (most likely tensors) live
- Tensor: [paddle/framework/tensor.h](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/tensor.h). Where we allocate memory through [`paddle/memory/`](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/memory)
paddle/api/CMakeLists.txt
浏览文件 @ 2b47fb3d
......@@ -25,8 +25,18 @@ FILE(GLOB PY_PADDLE_PYTHON_FILES ${PADDLE_SOURCE_DIR}/paddle/py_paddle/*.py)
SET_SOURCE_FILES_PROPERTIES(Paddle.i PROPERTIES CPLUSPLUS ON)
SET(SWIG_NEED_FLAGS
-ftls-model=global-dynamic
-Wno-parentheses-equality
-Wno-self-assign
-Wno-maybe-uninitialized
-Wno-missing-field-initializers)
FOREACH(flag ${SWIG_NEED_FLAGS})
safe_set_cxxflag(SWIG_CXX_FLAGS ${flag})
ENDFOREACH()
SET(CMAKE_SWIG_OUTDIR ${CMAKE_CURRENT_BINARY_DIR})
SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-parentheses-equality -Wno-missing-field-initializers -Wno-self-assign -ftls-model=global-dynamic")
SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SWIG_CXX_FLAGS}")
SET(SWIG_MODULE_swig_paddle_EXTRA_DEPS
paddle_parameter
......
paddle/capi/Main.cpp
浏览文件 @ 2b47fb3d
......@@ -43,4 +43,11 @@ paddle_error paddle_init(int argc, char** argv) {
isInit = true;
return kPD_NO_ERROR;
}
paddle_error paddle_init_thread() {
if (FLAGS_use_gpu) {
hl_init(FLAGS_gpu_id);
}
return kPD_NO_ERROR;
}
}
paddle/capi/Matrix.cpp
浏览文件 @ 2b47fb3d
......@@ -40,7 +40,7 @@ paddle_error paddle_matrix_destroy(paddle_matrix mat) {
paddle_error paddle_matrix_set_row(paddle_matrix mat,
uint64_t rowID,
paddle_real* rowArray) {
if (mat == nullptr) return kPD_NULLPTR;
if (mat == nullptr || rowArray == nullptr) return kPD_NULLPTR;
auto ptr = cast(mat);
if (ptr->mat == nullptr) return kPD_NULLPTR;
if (rowID >= ptr->mat->getHeight()) return kPD_OUT_OF_RANGE;
......
paddle/capi/error.cpp 0 → 100644
浏览文件 @ 2b47fb3d
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "error.h"
const char* paddle_error_string(paddle_error err) {
switch (err) {
case kPD_NULLPTR:
return "nullptr error";
case kPD_OUT_OF_RANGE:
return "out of range error";
case kPD_PROTOBUF_ERROR:
return "protobuf error";
case kPD_NOT_SUPPORTED:
return "not supported error";
case kPD_UNDEFINED_ERROR:
return "undefined error";
default:
return "";
}
}
paddle/capi/error.h
浏览文件 @ 2b47fb3d
......@@ -15,6 +15,8 @@ limitations under the License. */
#ifndef __PADDLE_CAPI_ERROR_H__
#define __PADDLE_CAPI_ERROR_H__
#include "config.h"
/**
* Error Type for Paddle API.
*/
......@@ -27,4 +29,9 @@ typedef enum {
kPD_UNDEFINED_ERROR = -1,
} paddle_error;
/**
* Error string for Paddle API.
*/
PD_API const char* paddle_error_string(paddle_error err);
#endif
paddle/capi/examples/model_inference/multi_thread/CMakeLists.txt
浏览文件 @ 2b47fb3d
project(multi_thread)
cmake_minimum_required(VERSION 2.8)
aux_source_directory(. SRC_LIST)
add_executable(${PROJECT_NAME} ${SRC_LIST})
find_package (Threads)
if(NOT PADDLE_ROOT)
set(PADDLE_ROOT $ENV{PADDLE_ROOT} CACHE PATH "Paddle Path")
endif()
if(PADDLE_ROOT)
include_directories(${PADDLE_ROOT}/include)
link_directories(${PADDLE_ROOT}/lib)
endif()
set(CPU_SRCS main.c)
add_executable(${PROJECT_NAME} ${CPU_SRCS})
set_property(TARGET ${PROJECT_NAME} PROPERTY C_STANDARD 99)
target_link_libraries(${PROJECT_NAME} -lpaddle_capi_shared
${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${PROJECT_NAME}
-lpaddle_capi_shared
${CMAKE_THREAD_LIBS_INIT})
find_package(CUDA QUIET)
if(CUDA_FOUND)
set(GPU_SRCS main_gpu.c)
cuda_add_executable(${PROJECT_NAME}_gpu ${GPU_SRCS})
set_property(TARGET ${PROJECT_NAME}_gpu PROPERTY C_STANDARD 99)
target_link_libraries(${PROJECT_NAME}_gpu
-lpaddle_capi_shared
${CMAKE_THREAD_LIBS_INIT})
endif(CUDA_FOUND)
paddle/capi/examples/model_inference/multi_thread/main_gpu.c 0 → 100644
浏览文件 @ 2b47fb3d
#include <paddle/capi.h>
#include <pthread.h>
#include <time.h>
#include "../common/common.h"
#define CONFIG_BIN "./trainer_config.bin"
#define NUM_THREAD 4
#define NUM_ITER 1000
pthread_mutex_t mutex;
/*
* @brief It is an simple inference example that runs multi-threads on a GPU.
* Each thread holds it own local gradient_machine but shares the same
* parameters.
* If you want to run on different GPUs, you need to launch
* multi-processes or set trainer_count > 1.
*/
void* thread_main(void* gm_ptr) {
// Initialize the thread environment of Paddle.
CHECK(paddle_init_thread());
paddle_gradient_machine machine = (paddle_gradient_machine)(gm_ptr);
// Create input arguments.
paddle_arguments in_args = paddle_arguments_create_none();
// Create input matrix.
paddle_matrix mat = paddle_matrix_create(/* sample_num */ 1,
/* size */ 784,
/* useGPU */ true);
// Create output arguments.
paddle_arguments out_args = paddle_arguments_create_none();
// Create output matrix.
paddle_matrix prob = paddle_matrix_create_none();
// CPU buffer to cache the input and output.
paddle_real* cpu_input = (paddle_real*)malloc(784 * sizeof(paddle_real));
paddle_real* cpu_output = (paddle_real*)malloc(10 * sizeof(paddle_real));
for (int iter = 0; iter < NUM_ITER; ++iter) {
// There is only one input layer of this network.
CHECK(paddle_arguments_resize(in_args, 1));
CHECK(paddle_arguments_set_value(in_args, 0, mat));
for (int i = 0; i < 784; ++i) {
cpu_input[i] = rand() / ((float)RAND_MAX);
}
CHECK(paddle_matrix_set_value(mat, cpu_input));
CHECK(paddle_gradient_machine_forward(machine,
in_args,
out_args,
/* isTrain */ false));
CHECK(paddle_arguments_get_value(out_args, 0, prob));
CHECK(paddle_matrix_get_value(prob, cpu_output));
pthread_mutex_lock(&mutex);
printf("Prob: ");
for (int i = 0; i < 10; ++i) {
printf("%.2f ", cpu_output[i]);
}
printf("\n");
pthread_mutex_unlock(&mutex);
}
CHECK(paddle_matrix_destroy(prob));
CHECK(paddle_arguments_destroy(out_args));
CHECK(paddle_matrix_destroy(mat));
CHECK(paddle_arguments_destroy(in_args));
CHECK(paddle_gradient_machine_destroy(machine));
free(cpu_input);
free(cpu_output);
return NULL;
}
int main() {
// Initalize Paddle
char* argv[] = {"--use_gpu=True"};
CHECK(paddle_init(1, (char**)argv));
// Reading config binary file. It is generated by `convert_protobin.sh`
long size;
void* buf = read_config(CONFIG_BIN, &size);
// Create a gradient machine for inference.
paddle_gradient_machine machine;
CHECK(paddle_gradient_machine_create_for_inference(&machine, buf, (int)size));
CHECK(paddle_gradient_machine_randomize_param(machine));
// Loading parameter. Uncomment the following line and change the directory.
// CHECK(paddle_gradient_machine_load_parameter_from_disk(machine,
// "./some_where_to_params"));
srand(time(0));
pthread_mutex_init(&mutex, NULL);
pthread_t threads[NUM_THREAD];
for (int i = 0; i < NUM_THREAD; ++i) {
paddle_gradient_machine thread_local_machine;
CHECK(paddle_gradient_machine_create_shared_param(
machine, buf, size, &thread_local_machine));
pthread_create(&threads[i], NULL, thread_main, thread_local_machine);
}
for (int i = 0; i < NUM_THREAD; ++i) {
pthread_join(threads[i], NULL);
}
pthread_mutex_destroy(&mutex);
return 0;
}
paddle/capi/main.h
浏览文件 @ 2b47fb3d
......@@ -26,6 +26,13 @@ extern "C" {
*/
PD_API paddle_error paddle_init(int argc, char** argv);
/**
* Initialize the thread environment of Paddle.
* @note it is requisite for GPU runs but optional for CPU runs.
* For GPU runs, all threads will run on the same GPU devices.
*/
PD_API paddle_error paddle_init_thread();
#ifdef __cplusplus
}
#endif
......
paddle/cuda/include/hl_cnn.h
浏览文件 @ 2b47fb3d
......@@ -116,6 +116,7 @@ extern void hl_maxpool_backward(const int frameCnt,
* @param[in] paddingW padding width.
* @param[out] tgtData output data.
* @param[in] tgtStride stride between output data samples.
* @param[in] excludeMode whether to consider paddings for size.
*
*/
extern void hl_avgpool_forward(const int frameCnt,
......@@ -132,7 +133,8 @@ extern void hl_avgpool_forward(const int frameCnt,
const int paddingH,
const int paddingW,
real* tgtData,
const int tgtStride);
const int tgtStride,
bool excludeMode);
/**
* @brief Maximum pool backward.
......@@ -154,6 +156,7 @@ extern void hl_avgpool_forward(const int frameCnt,
* @param[in] scaleB scale.
* @param[out] backGrad output grad.
* @param[in] outStride stride between output data samples.
* @param[in] excludeMode whether to consider paddings for size.
*
*/
extern void hl_avgpool_backward(const int frameCnt,
......@@ -172,7 +175,8 @@ extern void hl_avgpool_backward(const int frameCnt,
real scaleA,
real scaleB,
real* backGrad,
const int outStride);
const int outStride,
bool excludeMode);
extern void hl_maxpool3D_forward(const int frameCnt,
const real* inputData,
......
paddle/cuda/include/stub/hl_cnn_stub.h
浏览文件 @ 2b47fb3d
......@@ -68,7 +68,8 @@ inline void hl_avgpool_forward(const int frameCnt,
const int paddingH,
const int paddingW,
real* tgtData,
const int tgtStride) {}
const int tgtStride,
const bool excludeMode) {}
inline void hl_avgpool_backward(const int frameCnt,
const real* outGrad,
......@@ -86,7 +87,8 @@ inline void hl_avgpool_backward(const int frameCnt,
real scaleA,
real scaleB,
real* backGrad,
const int outStride) {}
const int outStride,
const bool excludeMode) {}
inline void hl_maxpool3D_forward(const int frameCnt,
const real* inputData,
......
paddle/cuda/src/hl_cuda_cnn.cu
浏览文件 @ 2b47fb3d
......@@ -210,7 +210,8 @@ __global__ void KeAvgPoolForward(const int nthreads,
const int padH,
const int padW,
real* tgtData,
const int tgtStride) {
const int tgtStride,
const bool excludeMode) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
if (index < nthreads) {
int pw = index % pooledW;
......@@ -224,7 +225,8 @@ __global__ void KeAvgPoolForward(const int nthreads,
int wend = min(wstart + sizeX, width);
hstart = max(hstart, 0);
wstart = max(wstart, 0);
int pool_size = (hend - hstart) * (wend - wstart);
int poolSize =
excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX;
real aveval = 0;
inputData += (frameNum * channels + c) * height * width;
......@@ -235,7 +237,7 @@ __global__ void KeAvgPoolForward(const int nthreads,
}
int tgtIndex =
index % (pooledW * pooledH * channels) + frameNum * tgtStride;
tgtData[tgtIndex] = aveval / pool_size;
tgtData[tgtIndex] = aveval / poolSize;
}
}
......@@ -253,7 +255,8 @@ void hl_avgpool_forward(const int frameCnt,
const int paddingH,
const int paddingW,
real* tgtData,
const int tgtStride) {
const int tgtStride,
const bool excludeMode) {
int num_kernels = pooledH * pooledW * channels * frameCnt;
int blocks = (num_kernels + 1024 - 1) / 1024;
KeAvgPoolForward<<<blocks, 1024, 0, STREAM_DEFAULT>>>(num_kernels,
......@@ -270,7 +273,8 @@ void hl_avgpool_forward(const int frameCnt,
paddingH,
paddingW,
tgtData,
tgtStride);
tgtStride,
excludeMode);
CHECK_SYNC("hl_avgpool_forward failed");
}
......@@ -290,7 +294,8 @@ __global__ void KeAvgPoolBackward(const int nthreads,
real scaleA,
real scaleB,
real* tgtGrad,
const int outStride) {
const int outStride,
const bool excludeMode) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
if (index < nthreads) {
int offsetW = index % width + padW;
......@@ -314,8 +319,9 @@ __global__ void KeAvgPoolBackward(const int nthreads,
int wstart = pw * strideW - padW;
int wend = min(wstart + sizeX, width);
wstart = max(wstart, 0);
int poolsize = (hend - hstart) * (wend - wstart);
gradient += outGrad[ph * pooledW + pw] / poolsize;
int poolSize =
excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX;
gradient += outGrad[ph * pooledW + pw] / poolSize;
}
}
tgtGrad[index] = scaleB * tgtGrad[index] + scaleA * gradient;
......@@ -338,7 +344,8 @@ void hl_avgpool_backward(const int frameCnt,
real scaleA,
real scaleB,
real* backGrad,
const int outStride) {
const int outStride,
const bool excludeMode) {
int num_kernels = height * width * channels * frameCnt;
int blocks = (num_kernels + 1024 - 1) / 1024;
......@@ -358,7 +365,8 @@ void hl_avgpool_backward(const int frameCnt,
scaleA,
scaleB,
backGrad,
outStride);
outStride,
excludeMode);
CHECK_SYNC("hl_avgpool_backward failed");
}
......
paddle/framework/backward.cc
浏览文件 @ 2b47fb3d
......@@ -33,8 +33,8 @@ static std::unordered_set<std::string>* g_ctrl_flow_ops_ = nullptr;
// We should design a better way to backward CtrlFlowOps.
static std::unordered_set<std::string>& CtrlFlowOps() {
if (g_ctrl_flow_ops_ == nullptr) {
g_ctrl_flow_ops_ =
new std::unordered_set<std::string>{"increment", "lod_rank_table"};
g_ctrl_flow_ops_ = new std::unordered_set<std::string>{
"increment", "lod_rank_table", "less_than"};
}
return *g_ctrl_flow_ops_;
}
......@@ -190,8 +190,9 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
// collect all the offset for each alias,
// insert a sum operator to add all aliases to output
insert_position.push_back(
{dup_op.back(), OpRegistry::CreateOp("sum", {{"X", dup_outputs}},
{{"Out", {name}}}, {})});
{dup_op.back(),
OpRegistry::CreateOp("sum", {{"X", dup_outputs}}, {{"Out", {name}}},
AttributeMap{})});
}
// make sure the inserted `sum` ops follow the BFS order.
......@@ -216,7 +217,8 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
// If part of input gradient of that operator is not calculated, fill
// zero variables to that input gradient.
net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}},
{{"Y", {grad_input}}}, {}));
{{"Y", {grad_input}}},
AttributeMap{}));
}
return false;
});
......@@ -392,8 +394,9 @@ std::vector<std::unique_ptr<OpDescBind>> MakeOpGrad(
0, in_name.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
std::string new_name = prefix + kZeroVarSuffix;
desc->Rename(in_name, new_name);
std::unique_ptr<OpDescBind> fill_zeros_op(new OpDescBind(
"fill_zeros_like", {{"X", {prefix}}}, {{"Y", {new_name}}}, {}));
std::unique_ptr<OpDescBind> fill_zeros_op(
new OpDescBind("fill_zeros_like", {{"X", {prefix}}},
{{"Y", {new_name}}}, AttributeMap{}));
pending_fill_zeros_ops.push_back(std::move(fill_zeros_op));
}
}
......@@ -483,8 +486,9 @@ std::vector<std::unique_ptr<OpDescBind>> MakeBlockBackward(
sum_op_inputs.emplace_back(new_name);
next_g_name = sum_op_inputs.back();
}
std::unique_ptr<OpDescBind> sum_op(new OpDescBind(
"sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}}, {}));
std::unique_ptr<OpDescBind> sum_op(
new OpDescBind("sum", {{"X", sum_op_inputs}}, {{"Out", {out_name}}},
AttributeMap{}));
pending_sum_ops.push_back({dup_op.back(), std::move(sum_op)});
}
}
......
paddle/framework/backward_test.cc
浏览文件 @ 2b47fb3d
......@@ -106,15 +106,15 @@ class FcOp : public operators::NetOp {
FcOp(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs, const AttributeMap &attrs)
: NetOp(type, inputs, outputs, attrs) {
AppendOp(OpRegistry::CreateOp("mul",
{{"X", {Input("X")}}, {"Y", {Input("W")}}},
{{"Out", {Output("mul_result")}}}, {}));
AppendOp(OpRegistry::CreateOp(
"mul", {{"X", {Input("X")}}, {"Y", {Input("W")}}},
{{"Out", {Output("mul_result")}}}, AttributeMap{}));
auto input_b = Inputs("b");
std::string before_act = "mul_result";
if (input_b.size() != 0) {
AppendOp(OpRegistry::CreateOp(
"rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}},
{{"Out", {Output("add_result")}}}, {}));
{{"Out", {Output("add_result")}}}, AttributeMap{}));
before_act = "add_result";
} else {
auto out_varname = Output("add_result");
......@@ -124,7 +124,7 @@ class FcOp : public operators::NetOp {
}
AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
{{"Out", {Output("Out")}}}, {}));
{{"Out", {Output("Out")}}}, AttributeMap{}));
CompleteAddOp(false);
}
};
......@@ -278,8 +278,9 @@ REGISTER_OPERATOR(scale, f::NoneOp);
REGISTER_OP_CPU_KERNEL(scale, f::NoneKernel<paddle::platform::CPUPlace, float>);
TEST(Backward, simple_op_not_need_grad) {
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
auto fwd =
f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}},
{{"Out", {"out"}}}, f::AttributeMap{});
ASSERT_NE(fwd, nullptr);
auto gop = f::Backward(*fwd, {"x"});
ASSERT_EQ(gop->Output(f::GradVarName("X")), f::kEmptyVarName);
......@@ -296,9 +297,10 @@ TEST(Backward, net_fc_backward_normal) {
{{"mul_result", {"mul_res"}},
{"add_result", {"add_re"}},
{"Out", {"out"}}},
{});
f::AttributeMap{});
ASSERT_NE(fwd, nullptr);
std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
std::shared_ptr<f::OperatorBase> gop =
f::Backward(*fwd, std::unordered_set<std::string>{});
ASSERT_TRUE(gop->IsNetOp());
auto net = static_cast<ops::NetOp *>(gop.get());
......@@ -322,9 +324,10 @@ TEST(Backward, net_fc_backward_not_have_b) {
{{"mul_result", {"mul_res"}},
{"add_result", {"add_res"}},
{"Out", {"tmp"}}},
{});
f::AttributeMap{});
ASSERT_NE(fwd, nullptr);
std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
std::shared_ptr<f::OperatorBase> gop =
f::Backward(*fwd, std::unordered_set<std::string>{});
ASSERT_TRUE(gop->IsNetOp());
auto net = static_cast<ops::NetOp *>(gop.get());
......@@ -346,13 +349,13 @@ TEST(Backward, net_input_of_network_not_need_grad) {
{{"mul_result", {"mul_tmp_0"}},
{"add_result", {"add_tmp_0"}},
{"Out", {"hidden0"}}},
{}));
f::AttributeMap{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}},
{{"mul_result", {"mul_tmp_1"}},
{"add_result", {"add_tmp_1"}},
{"Out", {"hidden1"}}},
{}));
f::AttributeMap{}));
net.CompleteAddOp();
auto bwd = Backward(net, {"x"}); // x@GRAD is not need.
ASSERT_TRUE(bwd->IsNetOp());
......@@ -381,12 +384,13 @@ TEST(Backward, net_input_of_network_not_need_grad) {
TEST(Backward, net_shared_weight) {
ops::NetOp net;
net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
{{"Out", {"out"}}}, {}));
{{"Out", {"out"}}}, f::AttributeMap{}));
net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
{{"Out", {"FinalOut"}}}, {}));
{{"Out", {"FinalOut"}}},
f::AttributeMap{}));
net.CompleteAddOp();
auto bwd = f::Backward(net, {});
auto bwd = f::Backward(net, std::unordered_set<std::string>{});
ASSERT_TRUE(bwd->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
ASSERT_EQ(3UL, bwd_net->ops_.size());
......@@ -394,8 +398,9 @@ TEST(Backward, net_shared_weight) {
}
TEST(Backward, op_all_input_are_not_need) {
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
auto fwd =
f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}},
{{"Out", {"out"}}}, f::AttributeMap{});
auto backward = f::Backward(*fwd, {"x", "b"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
......@@ -403,8 +408,9 @@ TEST(Backward, op_all_input_are_not_need) {
}
TEST(Backward, op_all_output_are_not_need) {
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
auto fwd =
f::OpRegistry::CreateOp("rowwise_add", {{"X", {"x"}}, {"b", {"b"}}},
{{"Out", {"out"}}}, f::AttributeMap{});
auto backward = f::Backward(*fwd, {"out"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
......@@ -412,8 +418,9 @@ TEST(Backward, op_all_output_are_not_need) {
}
TEST(Backward, op_part_of_output_are_not_need) {
auto fwd = f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}},
{{"y", {"Y"}}, {"z", {"Z"}}}, {});
auto fwd =
f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}},
{{"y", {"Y"}}, {"z", {"Z"}}}, f::AttributeMap{});
auto backward = f::Backward(*fwd, {"Z"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
......@@ -437,7 +444,7 @@ TEST(Backward, op_part_of_output_are_not_need) {
TEST(Backward, op_part_of_input_are_not_need) {
auto fwd = f::OpRegistry::CreateOp("mul", {{"X", {"a"}}, {"Y", {"b"}}},
{{"Out", {"out"}}}, {});
{{"Out", {"out"}}}, f::AttributeMap{});
auto backward = f::Backward(*fwd, {"a"});
auto &grad_mul = *backward;
ASSERT_EQ(grad_mul.Type(), "mul_grad");
......@@ -458,19 +465,19 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
{{"mul_result", {"mul_out1"}},
{"add_result", {"add_out1"}},
{"Out", {"out1"}}},
{}));
f::AttributeMap{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}},
{{"mul_result", {"mul_out2"}},
{"add_result", {"tmp_out2"}},
{"Out", {"out2"}}},
{}));
f::AttributeMap{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}},
{{"mul_result", {"mul_out3"}},
{"add_result", {"tmp_out3"}},
{"Out", {"out3"}}},
{}));
f::AttributeMap{}));
net.CompleteAddOp();
auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"});
......@@ -509,7 +516,8 @@ TEST(Backward, simple_single_op) {
auto target = f::VarDescBind("out");
target.SetShape({1});
auto var_to_grad = AppendBackward(program, target, {});
auto var_to_grad =
AppendBackward(program, target, std::unordered_set<std::string>{});
ASSERT_EQ(block->AllOps().size(), 3UL);
f::OpDescBind *fill_op = block->AllOps()[1];
......@@ -546,7 +554,7 @@ TEST(Backward, default_attribute) {
auto target = f::VarDescBind("out");
target.SetShape({1});
AppendBackward(program, target, {});
AppendBackward(program, target, std::unordered_set<std::string>{});
ASSERT_EQ(block->AllOps().size(), 3UL);
EXPECT_EQ(boost::get<int>(op->GetAttr("x_num_col_dims")), 1);
......@@ -585,7 +593,8 @@ TEST(Backward, simple_mult_op) {
auto target = f::VarDescBind("out3");
target.SetShape({1});
size_t forward_len = block->AllOps().size();
auto var_to_grad = AppendBackward(program, target, {});
auto var_to_grad =
AppendBackward(program, target, std::unordered_set<std::string>{});
ASSERT_EQ(block->AllOps().size(), 6UL + 1);
f::OpDescBind *fill_op = block->AllOps()[forward_len];
......@@ -817,7 +826,8 @@ TEST(Backward, shared_var) {
auto target = f::VarDescBind("out3");
target.SetShape({1});
size_t forward_len = block->AllOps().size();
auto var_to_grad = AppendBackward(program, target, {});
auto var_to_grad =
AppendBackward(program, target, std::unordered_set<std::string>{});
ASSERT_EQ(block->AllOps().size(), 8UL);
f::OpDescBind *fill_op = block->AllOps()[forward_len];
......
paddle/framework/op_desc.cc
浏览文件 @ 2b47fb3d
......@@ -59,7 +59,7 @@ class CompileTimeInferShapeContext : public InferShapeContext {
auto *in_var = block_.FindVarRecursive(Inputs(in)[i]);
auto *out_var = block_.FindVarRecursive(Outputs(out)[j]);
if (in_var->GetType() != VarDesc::LOD_TENSOR) {
VLOG(3) << "input " << in << "is not LodTensor";
VLOG(3) << "input " << in << " is not LodTensor";
return;
}
PADDLE_ENFORCE_EQ(in_var->GetType(), VarDesc::LOD_TENSOR,
......@@ -316,8 +316,8 @@ static void InitInferShapeFuncs() {
for (auto &kern_pair : OperatorWithKernel::AllOpKernels()) {
auto op_type = kern_pair.first;
auto &op_info = info_map.at(op_type);
auto op =
static_cast<OperatorWithKernel *>(op_info.Creator()("", {}, {}, {}));
auto op = static_cast<OperatorWithKernel *>(op_info.Creator()(
"", VariableNameMap{}, VariableNameMap{}, AttributeMap{}));
if (op_info.infer_shape_) { // infer_shape has been registered.
continue;
}
......
paddle/framework/op_registry.h
浏览文件 @ 2b47fb3d
......@@ -181,8 +181,8 @@ class OpKernelRegistrar : public Registrar {
return 0; \
}
#define REGISTER_OP_GPU_KERNEL(op_type, ...) \
REGISTER_OP_KERNEL(op_type, GPU, ::paddle::platform::GPUPlace, __VA_ARGS__)
#define REGISTER_OP_CUDA_KERNEL(op_type, ...) \
REGISTER_OP_KERNEL(op_type, CUDA, ::paddle::platform::GPUPlace, __VA_ARGS__)
#define REGISTER_OP_CPU_KERNEL(op_type, ...) \
REGISTER_OP_KERNEL(op_type, CPU, ::paddle::platform::CPUPlace, __VA_ARGS__)
......@@ -217,7 +217,7 @@ class OpKernelRegistrar : public Registrar {
#else
#define USE_OP_KERNEL(op_type) \
USE_OP_DEVICE_KERNEL(op_type, CPU); \
USE_OP_DEVICE_KERNEL(op_type, GPU)
USE_OP_DEVICE_KERNEL(op_type, CUDA)
#endif
#define USE_NO_KERNEL_OP(op_type) USE_OP_ITSELF(op_type);
......@@ -226,9 +226,9 @@ class OpKernelRegistrar : public Registrar {
USE_OP_ITSELF(op_type); \
USE_OP_DEVICE_KERNEL(op_type, CPU);
#define USE_GPU_ONLY_OP(op_type) \
USE_OP_ITSELF(op_type); \
USE_OP_DEVICE_KERNEL(op_type, GPU)
#define USE_CUDA_ONLY_OP(op_type) \
USE_OP_ITSELF(op_type); \
USE_OP_DEVICE_KERNEL(op_type, CUDA)
#define USE_OP(op_type) \
USE_OP_ITSELF(op_type); \
......
paddle/framework/operator.cc
浏览文件 @ 2b47fb3d
......@@ -22,20 +22,6 @@ limitations under the License. */
namespace paddle {
namespace framework {
template <>
Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
platform::CPUPlace, Eigen::DefaultDevice>() const {
return *device_context_.GetEigenDevice<platform::CPUPlace>();
}
#ifdef PADDLE_WITH_CUDA
template <>
Eigen::GpuDevice&
ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
return *device_context_.GetEigenDevice<platform::GPUPlace>();
}
#endif
std::string OperatorBase::Input(const std::string& name) const {
auto& ins = Inputs(name);
PADDLE_ENFORCE_LE(ins.size(), 1UL,
......@@ -426,13 +412,10 @@ void OperatorWithKernel::Run(const Scope& scope,
}
kernel_iter->second->Compute(ctx);
// throws errors if have.
dev_ctx.Finish();
}
OpKernelType OperatorWithKernel::GetKernelType(
const ExecutionContext& ctx) const {
return OpKernelType(IndicateDataType(ctx), ctx.device_context());
return OpKernelType(IndicateDataType(ctx), ctx.GetPlace());
}
DataType OperatorWithKernel::IndicateDataType(
const ExecutionContext& ctx) const {
......
paddle/framework/operator.h
浏览文件 @ 2b47fb3d
......@@ -276,17 +276,25 @@ class ExecutionContext {
out_tensor->set_lod(in_tensor.lod());
}
template <typename PlaceType,
typename DeviceType = typename platform::EigenDeviceConverter<
PlaceType>::EigenDeviceType>
DeviceType& GetEigenDevice() const;
platform::Place GetPlace() const { return device_context_.GetPlace(); }
template <typename DeviceContextType>
const DeviceContextType& device_context() const {
return *reinterpret_cast<const DeviceContextType*>(&device_context_);
}
const platform::DeviceContext& device_context() const {
return device_context_;
}
#ifdef PADDLE_WITH_CUDA
const inline platform::CUDADeviceContext& cuda_device_context() const {
PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace()));
return *reinterpret_cast<const platform::CUDADeviceContext*>(
&device_context_);
}
#endif
//! Get actual name vector for this input.
const std::vector<std::string>& Inputs(const std::string& name) const {
return op_.Inputs(name);
......@@ -297,14 +305,6 @@ class ExecutionContext {
return op_.Outputs(name);
}
#ifdef PADDLE_WITH_CUDA
const inline platform::CUDADeviceContext& cuda_device_context() const {
PADDLE_ENFORCE(platform::is_gpu_place(device_context_.GetPlace()));
return *reinterpret_cast<const platform::CUDADeviceContext*>(
&device_context_);
}
#endif
private:
const OperatorBase& op_;
const Scope& scope_;
......
paddle/framework/operator_test.cc
浏览文件 @ 2b47fb3d
......@@ -115,7 +115,7 @@ class OpWithKernelTest : public OperatorWithKernel {
protected:
void InferShape(framework::InferShapeContext* ctx) const override {}
OpKernelType GetKernelType(const ExecutionContext& ctx) const override {
return OpKernelType(DataType::FP32, ctx.device_context());
return OpKernelType(DataType::FP32, ctx.GetPlace());
}
};
......@@ -261,7 +261,9 @@ class OperatorClone : public paddle::framework::OperatorBase {
};
TEST(Operator, Clone) {
OperatorClone a("ABC", {}, {}, {});
OperatorClone a("ABC", paddle::framework::VariableNameMap{},
paddle::framework::VariableNameMap{},
paddle::framework::AttributeMap{});
auto b = a.Clone();
ASSERT_EQ(a.Type(), b->Type());
}
paddle/framework/prune_test.cc
浏览文件 @ 2b47fb3d
......@@ -54,7 +54,8 @@ TEST(Prune, one_operator) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, {}, block);
AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, f::AttributeMap{},
block);
f::ProgramDesc *pdesc = program.Proto();
f::ProgramDesc pruned;
......@@ -71,10 +72,14 @@ TEST(Prune, forward) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, {}, block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"c"}}}, {}, block);
AddOp("one_one", {{"input", {"c"}}}, {{"output", {"d"}}}, {}, block);
AddOp("one_one", {{"input", {"d"}}}, {{"output", {"e"}}}, {}, block);
AddOp("one_one", {{"input", {"a"}}}, {{"output", {"b"}}}, f::AttributeMap{},
block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"c"}}}, f::AttributeMap{},
block);
AddOp("one_one", {{"input", {"c"}}}, {{"output", {"d"}}}, f::AttributeMap{},
block);
AddOp("one_one", {{"input", {"d"}}}, {{"output", {"e"}}}, f::AttributeMap{},
block);
f::ProgramDesc *pdesc = program.Proto();
......@@ -90,11 +95,14 @@ TEST(Prune, multi_input_op) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_one", {{"input", {"a0"}}}, {{"output", {"b0"}}}, {}, block);
AddOp("one_one", {{"input", {"a1"}}}, {{"output", {"b1"}}}, {}, block);
AddOp("one_one", {{"input", {"a2"}}}, {{"output", {"b2"}}}, {}, block);
AddOp("three_one", {{"input", {"b0", "b1", "b2"}}}, {{"output", {"c"}}}, {},
AddOp("one_one", {{"input", {"a0"}}}, {{"output", {"b0"}}}, f::AttributeMap{},
block);
AddOp("one_one", {{"input", {"a1"}}}, {{"output", {"b1"}}}, f::AttributeMap{},
block);
AddOp("one_one", {{"input", {"a2"}}}, {{"output", {"b2"}}}, f::AttributeMap{},
block);
AddOp("three_one", {{"input", {"b0", "b1", "b2"}}}, {{"output", {"c"}}},
f::AttributeMap{}, block);
f::ProgramDesc *pdesc = program.Proto();
pdesc->mutable_blocks(0)->mutable_ops(3)->set_is_target(true);
......@@ -108,9 +116,12 @@ TEST(Prune, multi_output_op) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}}, {}, block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, {}, block);
AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, {}, block);
AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}},
f::AttributeMap{}, block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, f::AttributeMap{},
block);
AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, f::AttributeMap{},
block);
f::ProgramDesc *pdesc = program.Proto();
pdesc->mutable_blocks(0)->mutable_ops(2)->set_is_target(true);
......@@ -124,9 +135,12 @@ TEST(Prune, multi_target) {
f::ProgramDescBind program;
f::BlockDescBind *block = program.MutableBlock(0);
AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}}, {}, block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, {}, block);
AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, {}, block);
AddOp("one_two", {{"input", {"a"}}}, {{"output", {"b", "c"}}},
f::AttributeMap{}, block);
AddOp("one_one", {{"input", {"b"}}}, {{"output", {"b1"}}}, f::AttributeMap{},
block);
AddOp("one_one", {{"input", {"c"}}}, {{"output", {"c1"}}}, f::AttributeMap{},
block);
f::ProgramDesc *pdesc = program.Proto();
pdesc->mutable_blocks(0)->mutable_ops(1)->set_is_target(true);
......
paddle/gserver/activations/ActivationFunction.cpp
浏览文件 @ 2b47fb3d
......@@ -24,7 +24,7 @@ limitations under the License. */
#include "paddle/utils/ClassRegistrar.h"
#include "paddle/utils/Logging.h"
#ifdef PADDLE_USE_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
#include "MKLDNNActivation.h"
#endif
......@@ -490,7 +490,7 @@ Error __must_check backward(Argument& act) {
END_DEFINE_ACTIVATION(log)
ActivationFunction* ActivationFunction::create(const std::string& type) {
#ifdef PADDLE_USE_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
if (!type.empty() && type.compare(0, 7, "mkldnn_") == 0) {
return MKLDNNActivation::create(type);
}
......
paddle/gserver/gradientmachines/NeuralNetwork.cpp
浏览文件 @ 2b47fb3d
......@@ -20,7 +20,7 @@ limitations under the License. */
#include "paddle/utils/Logging.h"
#include "paddle/utils/Stat.h"
#ifdef PADDLE_USE_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/gserver/layers/MKLDNNLayer.h"
#endif
......@@ -307,7 +307,7 @@ void NeuralNetwork::backward(const UpdateCallback& callback) {
}
void NeuralNetwork::finish() {
#ifdef PADDLE_USE_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
FOR_EACH_R(layer, layers_) {
MKLDNNLayerPtr dnnLayer = std::dynamic_pointer_cast<MKLDNNLayer>(*layer);
if (dnnLayer) {
......
paddle/gserver/layers/ConvTransProjection.cpp
浏览文件 @ 2b47fb3d
......@@ -24,13 +24,13 @@ size_t ConvTransProjection::calOutputSize() {
if (outputH_ == 0) outputH_ = configOutH_;
if (outputW_ == 0) outputW_ = configOutW_;
imageH_ = imageSize(outputH_,
filterH_,
(filterH_ - 1) * dilationH_ + 1,
paddingH_,
strideH_,
/* caffeMode */ true);
imageW_ = imageSize(outputW_,
filterW_,
(filterW_ - 1) * dilationW_ + 1,
paddingW_,
strideW_,
/* caffeMode */ true);
......
paddle/gserver/layers/MKLDNNLRNLayer.cpp 0 → 100644
浏览文件 @ 2b47fb3d
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "MKLDNNLRNLayer.h"
#include "paddle/utils/Logging.h"
using namespace mkldnn; // NOLINT
typedef memory::format format;
namespace paddle {
REGISTER_LAYER(mkldnn_lrn, MKLDNNLRNLayer);
bool MKLDNNLRNLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
if (!MKLDNNLayer::init(layerMap, parameterMap)) {
return false;
}
/* the size of inputs for norm-layer is 1 */
CHECK_EQ(config_.inputs_size(), 1UL);
const NormConfig& conf = config_.inputs(0).norm_conf();
localSize_ = conf.size();
alpha_ = conf.scale();
beta_ = conf.pow();
ic_ = conf.channels();
oc_ = ic_;
iw_ = conf.img_size();
ow_ = conf.output_x();
ih_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
oh_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
CHECK_EQ(iw_, ow_);
CHECK_EQ(ih_, oh_);
return true;
}
void MKLDNNLRNLayer::reshape(
int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) {
CHECK_EQ(inputLayers_.size(), 1UL);
reshapeInput(bs, ih, iw);
// ic_ and oc can not be changed
CHECK_EQ((size_t)ic,
inputLayers_[0]->getOutputValue()->getElementCnt() / bs / ih / iw)
<< "Input channel can not be changed";
oh = ih;
ow = iw;
reshapeOutput(oh, ow);
resizeOutput(bs, oc * oh * ow);
}
void MKLDNNLRNLayer::resetFwd(std::vector<primitive>& pipeline,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out) {
resetFwdBuffers(inputs[0], out);
resetFwdPD(fwdPD_, inputs[0], out);
resetFwdPipeline(pipeline, fwdPD_, inputs[0], out);
}
void MKLDNNLRNLayer::resetBwd(std::vector<primitive>& pipeline,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out) {
std::shared_ptr<lrn_bwd::primitive_desc> pd;
resetBwdBuffers(inputs[0], out);
resetBwdPD(pd, inputs[0], out);
resetBwdPipeline(pipeline, pd, inputs[0], out);
}
void MKLDNNLRNLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
resetInValue(in);
CHECK(in);
resetOutValue(out, in->getPrimitiveDesc());
}
void MKLDNNLRNLayer::resetFwdPD(std::shared_ptr<lrn_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr in,
MKLDNNMatrixPtr out) {
prop_kind pk = passType_ == PASS_TEST ? prop_kind::forward_scoring
: prop_kind::forward_training;
auto fwdDesc = lrn_fwd::desc(pk,
algorithm::lrn_across_channels,
in->getMemoryDesc(),
localSize_,
alpha_,
beta_,
1.0f);
pd.reset(new lrn_fwd::primitive_desc(fwdDesc, engine_));
// prepare workspace if necessary
workspace_ =
passType_ != PASS_TEST
? std::make_shared<memory>(memory(pd->workspace_primitive_desc()))
: nullptr;
}
void MKLDNNLRNLayer::resetFwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<lrn_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
fwd_ = workspace_
? std::make_shared<lrn_fwd>(lrn_fwd(*pd, *in, *workspace_, *out))
: std::make_shared<lrn_fwd>(lrn_fwd(*pd, *in, *out));
pipeline.push_back(*fwd_);
}
void MKLDNNLRNLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
CHECK(inVals_[0] && outVal_);
resetOutGrad(out, outVal_->getPrimitiveDesc());
resetInGrad(in, inVals_[0]->getPrimitiveDesc());
}
void MKLDNNLRNLayer::resetBwdPD(std::shared_ptr<lrn_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
pd = nullptr;
if (in == nullptr) {
return;
}
CHECK(out);
auto bwdDesc = lrn_bwd::desc(algorithm::lrn_across_channels,
in->getMemoryDesc(),
out->getMemoryDesc(),
localSize_,
alpha_,
beta_,
1.0f);
pd.reset(new lrn_bwd::primitive_desc(bwdDesc, engine_, *fwdPD_));
}
void MKLDNNLRNLayer::resetBwdPipeline(
std::vector<primitive>& pipeline,
std::shared_ptr<lrn_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out) {
if (pd == nullptr) {
return;
}
CHECK(inVals_[0]);
CHECK(workspace_);
bwdData_ = std::make_shared<lrn_bwd>(
lrn_bwd(*pd, *inVals_[0], *out, *workspace_, *in));
pipeline.push_back(*bwdData_);
}
} // namespace paddle
paddle/gserver/layers/MKLDNNLRNLayer.h 0 → 100644
浏览文件 @ 2b47fb3d
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "MKLDNNLayer.h"
#include "mkldnn.hpp"
namespace paddle {
typedef mkldnn::lrn_forward lrn_fwd;
typedef mkldnn::lrn_backward lrn_bwd;
/**
* @brief A subclass of MKLDNNLayer LRN(Local Response Norm) layer.
*
* The config file api is mkldnn_lrn
*/
class MKLDNNLRNLayer : public MKLDNNLayer {
protected:
// save forward primitive_desc, which can be used in backward
std::shared_ptr<lrn_fwd::primitive_desc> fwdPD_;
// according to https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
// test_lrn_backward.cpp, lrn need workspace for backward
std::shared_ptr<mkldnn::memory> workspace_;
int localSize_;
float alpha_, beta_; // scale and pow in paddle
public:
explicit MKLDNNLRNLayer(const LayerConfig& config) : MKLDNNLayer(config) {}
~MKLDNNLRNLayer() {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void reshape(
int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override;
void resetFwd(std::vector<mkldnn::primitive>& pipeline,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out) override;
void resetBwd(std::vector<mkldnn::primitive>& pipeline,
std::vector<MKLDNNMatrixPtr>& inputs,
MKLDNNMatrixPtr& out) override;
protected:
void resetFwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
void resetFwdPD(std::shared_ptr<lrn_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr in,
MKLDNNMatrixPtr out);
void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<lrn_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out);
void resetBwdBuffers(MKLDNNMatrixPtr& in, MKLDNNMatrixPtr& out);
void resetBwdPD(std::shared_ptr<lrn_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out);
void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<lrn_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& out);
};
} // namespace paddle
paddle/gserver/layers/PoolLayer.cpp
浏览文件 @ 2b47fb3d
......@@ -45,6 +45,8 @@ bool PoolLayer::init(const LayerMap& layerMap,
strideY_ = conf.has_stride_y() ? conf.stride_y() : conf.stride();
confPaddingY_ = conf.has_padding_y() ? conf.padding_y() : conf.padding();
outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
excludeMode_ = conf.has_exclude_mode() ? conf.exclude_mode() : true;
return true;
}
......
paddle/gserver/layers/PoolLayer.h
浏览文件 @ 2b47fb3d
......@@ -38,6 +38,8 @@ protected:
std::string poolType_;
bool excludeMode_;
public:
explicit PoolLayer(const LayerConfig& config) : Layer(config) {}
......
paddle/gserver/layers/PoolProjection.cpp
浏览文件 @ 2b47fb3d
......@@ -36,6 +36,8 @@ PoolProjection::PoolProjection(const ProjectionConfig& config,
strideY_ = conf.has_stride_y() ? conf.stride_y() : conf.stride();
confPaddingY_ = conf.has_padding_y() ? conf.padding_y() : conf.padding();
outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
excludeMode_ = conf.has_exclude_mode() ? conf.exclude_mode() : true;
}
size_t PoolProjection::getSize() {
......@@ -141,7 +143,8 @@ void AvgPoolProjection::forward() {
outputY_,
outputX_,
confPaddingY_,
confPadding_);
confPadding_,
excludeMode_);
}
void AvgPoolProjection::backward(const UpdateCallback& callback) {
......@@ -166,6 +169,7 @@ void AvgPoolProjection::backward(const UpdateCallback& callback) {
1,
1,
confPaddingY_,
confPadding_);
confPadding_,
excludeMode_);
}
} // namespace paddle
paddle/gserver/layers/PoolProjection.h
浏览文件 @ 2b47fb3d
......@@ -28,6 +28,7 @@ protected:
int confPaddingY_, confPadding_;
size_t channels_;
std::string poolType_;
bool excludeMode_;
public:
PoolProjection(const ProjectionConfig& config,
......
paddle/gserver/tests/mkldnn_simple_net.conf
浏览文件 @ 2b47fb3d
......@@ -51,6 +51,8 @@ tmp = img_pool_layer(input=tmp,
padding=1,
pool_type=MaxPooling())
tmp = img_cmrnorm_layer(input=tmp, size=5, scale=0.0001, power=0.75)
tmp = fc_layer(input=tmp,
size=channels,
bias_attr=False,
......
paddle/gserver/tests/test_LayerGrad.cpp
浏览文件 @ 2b47fb3d
......@@ -238,9 +238,24 @@ void testProjectionConv(size_t groups, bool isDeconv) {
/* caffeMode */ true);
conv->set_output_x(output_x);
conv->set_output_y(output_y);
LOG(INFO) << "DILATION:" << DILATION << "; output_x: " << output_x
<< "; output_y: " << output_y;
if (isDeconv) {
int deconv_image_x = imageSize(output_x,
(conv->filter_size() - 1) * DILATION + 1,
conv->padding(),
conv->stride(),
/* caffeMode */ true);
int deconv_image_y = imageSize(output_y,
(conv->filter_size_y() - 1) * DILATION + 1,
conv->padding_y(),
conv->stride_y(),
/* caffeMode */ true);
LOG(INFO) << " deconv_image_x: " << deconv_image_x
<< "; deconv_image_y: " << deconv_image_y;
conf.set_input_size(output_x * output_y * CHANNELS);
conf.set_output_size(IMAGE_SIZE * IMAGE_SIZE * NUM_FILTERS);
conf.set_output_size(deconv_image_x * deconv_image_y * NUM_FILTERS);
} else {
conf.set_input_size(IMAGE_SIZE * IMAGE_SIZE * CHANNELS);
conf.set_output_size(output_x * output_y * NUM_FILTERS);
......@@ -1211,7 +1226,10 @@ void setPoolConfig(TestConfig* config,
pool->set_output_y(oh);
}
void testPoolLayer(const string& poolType, bool trans, bool useGpu) {
void testPoolLayer(const string& poolType,
bool trans,
bool useGpu,
bool excludeMode = true) {
TestConfig config;
config.inputDefs.push_back({INPUT_DATA, "layer_0", 3136, 0});
LayerInputConfig* input = config.layerConfig.add_inputs();
......@@ -1219,6 +1237,7 @@ void testPoolLayer(const string& poolType, bool trans, bool useGpu) {
pool->set_img_size(14);
pool->set_img_size_y(14);
pool->set_exclude_mode(excludeMode);
setPoolConfig(&config, pool, poolType);
config.layerConfig.set_size(pool->output_x() * pool->output_y() *
pool->channels());
......@@ -1250,16 +1269,26 @@ void testPoolLayer2(const string& poolType, bool trans, bool useGpu) {
TEST(Layer, PoolLayer) {
testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ false);
testPoolLayer("avg-projection",
/* trans= */ false,
/* useGpu= */ false,
/* excludeMode= */ false);
testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ false);
testPoolLayer("max-pool-with-mask", /* trans= */ false, /* useGpu= */ false);
#ifdef PADDLE_WITH_CUDA
testPoolLayer("avg-projection", /* trans= */ false, /* useGpu= */ true);
testPoolLayer("avg-projection",
/* trans= */ false,
/* useGpu= */ true,
/* excludeMode= */ false);
testPoolLayer("max-projection", /* trans= */ false, /* useGpu= */ true);
testPoolLayer("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true);
testPoolLayer("cudnn-avg-pool", /* trans= */ false, /* useGpu= */ true);
testPoolLayer2("cudnn-max-pool", /* trans= */ false, /* useGpu= */ true);
testPoolLayer2("cudnn-avg-pool", /* trans= */ false, /* useGpu= */ true);
testPoolLayer2(
"cudnn-avg-incl-pad-pool", /* trans= */ false, /* useGpu= */ true);
testPoolLayer("max-pool-with-mask", /* trans= */ false, /* useGpu= */ true);
#endif
}
......
paddle/gserver/tests/test_MKLDNN.cpp
浏览文件 @ 2b47fb3d
......@@ -272,6 +272,51 @@ TEST(MKLDNNLayer, BatchNormLayer) {
testBatchNormLayer({4, 16, 8, 10});
}
struct testLRNDesc {
int bs, ic, ih, iw;
float scale, pow;
int localSize;
};
void getMKLDNNLRNConfig(TestConfig& cfg, const testLRNDesc& pm) {
cfg.layerConfig.set_type("mkldnn_lrn");
cfg.layerConfig.set_active_type("relu");
size_t layerSize = pm.ic * pm.ih * pm.iw;
cfg.inputDefs.push_back({INPUT_DATA, "layer_0", layerSize, 0});
LayerInputConfig* input = cfg.layerConfig.add_inputs();
NormConfig* norm = input->mutable_norm_conf();
norm->set_channels(pm.ic);
norm->set_size(pm.localSize);
norm->set_scale(pm.scale);
norm->set_pow(pm.pow);
norm->set_blocked(0);
norm->set_img_size(pm.iw);
norm->set_img_size_y(pm.ih);
norm->set_output_x(norm->img_size());
norm->set_output_y(norm->img_size_y());
cfg.layerConfig.set_size(layerSize);
cfg.biasSize = 0;
}
void testLRNLayer(const testLRNDesc& pm) {
TestConfig dnnConfig;
getMKLDNNLRNConfig(dnnConfig, pm);
// mkldnn_lrn <==> norm with cmrnorm-projection type
TestConfig refConfig = dnnConfig;
refConfig.layerConfig.set_type("norm");
LayerInputConfig* input = refConfig.layerConfig.mutable_inputs(0);
NormConfig* norm = input->mutable_norm_conf();
norm->set_norm_type("cmrnorm-projection");
norm->set_scale(norm->scale() / norm->size());
RUN_MKLDNN_TEST(dnnConfig, refConfig, pm)
}
TEST(MKLDNNLayer, LRNLayer) {
testLRNLayer({4, 10, 12, 12, 0.001f, 0.75f, 5});
testLRNLayer({2, 32, 6, 6, 0.001f, 0.75f, 5});
testLRNLayer({4, 16, 8, 10, 0.01f, 0.5f, 5});
}
struct testImageDesc {
int bs, ic, ih, iw;
};
......
paddle/math/Allocator.h
浏览文件 @ 2b47fb3d
......@@ -48,7 +48,7 @@ public:
*/
virtual void* alloc(size_t size) {
void* ptr;
#ifdef PADDLE_USE_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
// refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
// memory alignment
CHECK_EQ(posix_memalign(&ptr, 4096ul, size), 0);
......
paddle/math/MathFunctions.cpp
浏览文件 @ 2b47fb3d
......@@ -206,7 +206,7 @@ double dotProduct<double>(const int n, const double* x, const double* y) {
}
#endif
#if defined(PADDLE_USE_MKLML)
#if defined(PADDLE_WITH_MKLML)
template <>
void vExp<float>(const int n, const float* a, float* r) {
......
paddle/math/MathFunctions.h
浏览文件 @ 2b47fb3d
......@@ -15,7 +15,7 @@ limitations under the License. */
#ifndef MATHFUNCTIONS_H_
#define MATHFUNCTIONS_H_
#ifdef PADDLE_USE_MKLML
#ifdef PADDLE_WITH_MKLML
#include <mkl_cblas.h>
#include <mkl_lapacke.h>
#include <mkl_vml_functions.h>
......
paddle/math/Matrix.cpp
浏览文件 @ 2b47fb3d
......@@ -28,6 +28,7 @@ limitations under the License. */
#include "hl_top_k.h"
#include "paddle/utils/Logging.h"
#include "NEONFunctions.h"
#include "paddle/function/GemmFunctor.h"
#include "paddle/utils/ThreadLocal.h"
......@@ -1130,7 +1131,8 @@ void GpuMatrix::avgPoolForward(Matrix& inputMat,
size_t outputH,
size_t outputW,
size_t paddingH,
size_t paddingW) {
size_t paddingW,
bool excludeMode) {
CHECK(inputMat.useGpu_ == true) << "Matrix type are not equal";
real* inputData = inputMat.getData();
......@@ -1153,7 +1155,8 @@ void GpuMatrix::avgPoolForward(Matrix& inputMat,
paddingH,
paddingW,
data_,
getStride());
getStride(),
excludeMode);
}
void GpuMatrix::avgPoolBackward(Matrix& outGrad,
......@@ -1168,7 +1171,8 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad,
real scaleTargets,
real scaleOutput,
size_t paddingH,
size_t paddingW) {
size_t paddingW,
bool excludeMode) {
CHECK(outGrad.useGpu_ == true) << "Matrix type are not equal";
real* outDiff = outGrad.getData();
......@@ -1194,7 +1198,8 @@ void GpuMatrix::avgPoolBackward(Matrix& outGrad,
scaleTargets,
scaleOutput,
data_,
outGrad.getStride());
outGrad.getStride(),
excludeMode);
}
void GpuMatrix::maxPool3DForward(Matrix& inputMat,
......@@ -2136,7 +2141,8 @@ void CpuMatrix::avgPoolForward(Matrix& input,
size_t outputH,
size_t outputW,
size_t paddingH,
size_t paddingW) {
size_t paddingW,
bool excludeMode) {
// The main loop
size_t num = input.getHeight();
size_t inLength = imgSizeH * imgSizeW;
......@@ -2165,7 +2171,8 @@ void CpuMatrix::avgPoolForward(Matrix& input,
tgtData[ph * outputW + pw] += inData[h * imgSizeW + w];
}
}
int poolSize = (hend - hstart) * (wend - wstart);
int poolSize =
excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX;
CHECK(poolSize);
tgtData[ph * outputW + pw] /= poolSize;
}
......@@ -2189,7 +2196,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
real scaleTargets,
real scaleOutput,
size_t paddingH,
size_t paddingW) {
size_t paddingW,
bool excludeMode) {
size_t num = input.getHeight();
size_t channels = input.getWidth() / outputH / outputW;
size_t inLength = imgSizeH * imgSizeW;
......@@ -2211,7 +2219,8 @@ void CpuMatrix::avgPoolBackward(Matrix& input,
int wstart = pw * strideW - paddingW;
int wend = std::min(wstart + sizeX, imgSizeW);
wstart = std::max(wstart, 0);
int poolSize = (hend - hstart) * (wend - wstart);
int poolSize =
excludeMode ? (hend - hstart) * (wend - wstart) : sizeY * sizeX;
CHECK(poolSize);
for (int h = hstart; h < hend; ++h) {
......@@ -4157,16 +4166,36 @@ void CpuMatrix::print(std::ostream& os) const {
void CpuMatrix::paramReluForward(Matrix& data, Matrix& W) {
real* input = data.getData();
real* w = W.getData();
real* output = data_;
size_t numElements = data.getWidth();
size_t numSamples = data.getHeight();
size_t paraSize = W.getHeight() * W.getWidth();
CHECK(!(numElements % paraSize)); // this check from ParameterReluLayer::init
size_t partial_sum = numElements / paraSize;
if (paraSize == numElements) {
for (size_t n = 0; n < numSamples * numElements; ++n) {
output[n] = input[n] > 0 ? input[n] : input[n] * w[n % numElements];
}
return;
}
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
for (size_t n = 0; n < numSamples; ++n) {
for (size_t i = 0; i < paraSize; i++) {
neon::prelu(
input + i * partial_sum, w[i], output + i * partial_sum, partial_sum);
}
input = input + numElements;
output = output + numElements;
}
#else
for (size_t n = 0, k = 0; n < numSamples; ++n) {
for (size_t i = 0; i < numElements; ++i, ++k) {
data_[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
output[k] = input[k] > 0 ? input[k] : input[k] * w[i / partial_sum];
}
}
#endif
}
void CpuMatrix::paramReluBackwardW(Matrix& oGrad, Matrix& data) {
......
paddle/math/Matrix.h
浏览文件 @ 2b47fb3d
......@@ -911,7 +911,8 @@ public:
size_t outputH,
size_t outputW,
size_t paddingH,
size_t paddingW) {
size_t paddingW,
bool excludeMode = true) {
LOG(FATAL) << "Not implemeted";
}
......@@ -927,9 +928,11 @@ public:
real scaleTargets,
real scaleOutput,
size_t paddingH,
size_t paddingW) {
size_t paddingW,
bool excludeMode = true) {
LOG(FATAL) << "Not implemeted";
}
/**
* Pooling 3D forward operation, pick out the largest element
* in the sizeX of value
......@@ -1458,7 +1461,8 @@ public:
size_t outputH,
size_t outputW,
size_t paddingH,
size_t paddingW);
size_t paddingW,
bool excludeMode = true);
void avgPoolBackward(Matrix& input,
size_t imgSizeH,
......@@ -1472,7 +1476,8 @@ public:
real scaleTargets,
real scaleOutput,
size_t paddingH,
size_t paddingW);
size_t paddingW,
bool excludeMode = true);
void maxPool3DForward(Matrix& inputMat,
Matrix& maxPoolIdx,
......@@ -1730,7 +1735,8 @@ public:
size_t outputH,
size_t outputW,
size_t paddingH,
size_t paddingW);
size_t paddingW,
bool excludeMode = true);
void avgPoolBackward(Matrix& input,
size_t imgSizeH,
......@@ -1744,7 +1750,8 @@ public:
real scaleTargets,
real scaleOutput,
size_t paddingH,
size_t paddingW);
size_t paddingW,
bool excludeMode = true);
void maxPool3DForward(Matrix& inputMat,
Matrix& maxPoolIdx,
......
paddle/math/NEONFunctions.cpp
浏览文件 @ 2b47fb3d
......@@ -49,6 +49,46 @@ void relu(const float* a, float* b, int len) {
}
}
// b[i] = a[i] > 0.0f ? a[i] : a[i] * w
void prelu(const float* a, float w, float* b, int len) {
int offset = len % 16;
float32x4_t ma0, ma1, ma2, ma3;
float32x4_t zero = vdupq_n_f32(0.f);
float32x4_t vw = vdupq_n_f32(w);
for (int k = 0; k < len / 16; k++, a += 16, b += 16) {
ma0 = vld1q_f32(a);
ma1 = vld1q_f32(a + 4);
ma2 = vld1q_f32(a + 8);
ma3 = vld1q_f32(a + 12);
uint32x4_t flag0 = vcgtq_f32(ma0, zero);
uint32x4_t flag1 = vcgtq_f32(ma1, zero);
uint32x4_t flag2 = vcgtq_f32(ma2, zero);
uint32x4_t flag3 = vcgtq_f32(ma3, zero);
float32x4_t mul0 = vmulq_f32(ma0, vw);
float32x4_t mul1 = vmulq_f32(ma1, vw);
float32x4_t mul2 = vmulq_f32(ma2, vw);
float32x4_t mul3 = vmulq_f32(ma3, vw);
ma0 = vbslq_f32(flag0, ma0, mul0);
ma1 = vbslq_f32(flag1, ma1, mul1);
ma2 = vbslq_f32(flag2, ma2, mul2);
ma3 = vbslq_f32(flag3, ma3, mul3);
vst1q_f32(b, ma0);
vst1q_f32(b + 4, ma1);
vst1q_f32(b + 8, ma2);
vst1q_f32(b + 12, ma3);
}
for (int i = 0; i < offset; i++) {
b[i] = a[i] > 0.0f ? a[i] : a[i] * w;
}
}
} // namespace neon
} // namespace paddle
......
paddle/math/NEONFunctions.h
浏览文件 @ 2b47fb3d
......@@ -18,6 +18,7 @@ namespace paddle {
namespace neon {
void relu(const float* a, float* b, int len);
void prelu(const float* a, float w, float* b, int len);
} // namespace neon
} // namespace paddle
paddle/math/float16.h 0 → 100644
浏览文件 @ 2b47fb3d
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <stdint.h>
#ifdef PADDLE_WITH_CUDA
#include <cuda.h>
#endif // PADDLE_WITH_CUDA
#include "unsupported/Eigen/CXX11/Tensor"
#include "paddle/platform/hostdevice.h"
#ifdef __GNUC__
#define PADDLE_GNUC_VER (__GNUC__ * 10 + __GNUC_MINOR__)
#else
#define PADDLE_GNUC_VER 0
#endif // __GNUC__
#ifdef __clang__
#define PADDLE_CLANG_VER (__clang_major__ * 10 + __clang_minor__)
#else
#define PADDLE_CLANG_VER 0
#endif // __clang__
#if defined(__CUDACC__) && CUDA_VERSION >= 7050
#define PADDLE_CUDA_FP16
#include <cuda_fp16.h>
#endif
#if defined(__arm__) || defined(__aarch64__)
#define PADDLE_ARM
#endif
#if defined(__ARM_NEON) || defined(__ARM_NEON__)
#define PADDLE_NEON
#include <arm_neon.h>
#endif
#if defined(PADDLE_NEON) && defined(PADDLE_ARM_FP16) && \
(PADDLE_GNUC_VER >= 62 || PADDLE_CLANG_VER >= 37)
#define PADDLE_WITH_NATIVE_FP16
#endif
#ifndef PADDLE_ARM
#include <immintrin.h>
#endif // PADDLE_ARM
#define PADDLE_ALIGN(x) __attribute__((aligned(x)))
namespace paddle {
// Use PADDLE_ALIGNED(2) to ensure that each float16 will be allocated
// and aligned at least on a 2-byte boundary, which leads to efficient
// memory access of float16 struct and also makes float16 compatible
// with CUDA half, ARM float16_t, and Eigen::half data types.
struct PADDLE_ALIGN(2) float16 {
public:
uint16_t x;
// Constructors
HOSTDEVICE inline float16() : x(0) {}
HOSTDEVICE inline float16(const float16& h) : x(h.x) {}
#ifdef PADDLE_CUDA_FP16
HOSTDEVICE inline explicit float16(const half& h) {
#if CUDA_VERSION >= 9000
x = reinterpret_cast<__half_raw*>(&h)->x;
#else
x = h.x;
#endif // CUDA_VERSION >= 9000
}
#endif // PADDLE_CUDA_FP16
HOSTDEVICE inline explicit float16(const Eigen::half& h) : x(h.x) {}
#ifdef PADDLE_WITH_NATIVE_FP16
// __fp16 is a native half precision data type for arm cpu,
// float16_t is an alias for __fp16
HOSTDEVICE inline explicit float16(const float16_t& h) {
x = *reinterpret_cast<const uint16_t*>(&h);
}
#endif
HOSTDEVICE inline explicit float16(float val) {
#if defined(PADDLE_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
half tmp = __float2half(val);
x = *reinterpret_cast<uint16_t*>(&tmp);
#elif defined(PADDLE_WITH_NATIVE_FP16)
float32x4_t tmp = vld1q_dup_f32(&val);
float16_t res = vget_lane_f16(vcvt_f16_f32(tmp), 0);
x = *reinterpret_cast<uint16_t*>(&res);
#elif defined(__F16C__)
x = _cvtss_sh(val, 0);
#else
// Conversion routine adapted from
// http://stackoverflow.com/questions/1659440/32-bit-to-16-bit-floating-point-conversion
Bits v, s;
v.f = val;
uint32_t sign = v.si & sigN;
v.si ^= sign;
sign >>= shiftSign; // logical shift
s.si = mulN;
s.si = s.f * v.f; // correct subnormals
v.si ^= (s.si ^ v.si) & -(minN > v.si);
v.si ^= (infN ^ v.si) & -((infN > v.si) & (v.si > maxN));
v.si ^= (nanN ^ v.si) & -((nanN > v.si) & (v.si > infN));
v.ui >>= shift; // logical shift
v.si ^= ((v.si - maxD) ^ v.si) & -(v.si > maxC);
v.si ^= ((v.si - minD) ^ v.si) & -(v.si > subC);
x = v.ui | sign;
#endif
}
HOSTDEVICE inline explicit float16(bool b) : x(b ? 0x3c00 : 0) {}
template <class T>
HOSTDEVICE inline explicit float16(const T& val)
: x(float16(static_cast<float>(val)).x) {}
HOSTDEVICE inline float16& operator=(const float16& rhs) {
x = rhs.x;
return *this;
}
// Assignment operators
#ifdef PADDLE_CUDA_FP16
HOSTDEVICE inline float16& operator=(const half& rhs) {
#if CUDA_VERSION >= 9000
x = reinterpret_cast<__half_raw*>(&rhs)->x;
#else
x = rhs.x;
#endif
return *this;
}
#endif
HOSTDEVICE inline float16& operator=(const Eigen::half& rhs) {
x = rhs.x;
return *this;
}
#ifdef PADDLE_WITH_NATIVE_FP16
HOSTDEVICE inline float16& operator=(const float16_t& rhs) {
x = *reinterpret_cast<const uint16_t*>(&rhs);
return *this;
}
#endif
HOSTDEVICE inline float16& operator=(bool b) {
x = b ? 0x3c00 : 0;
return *this;
}
HOSTDEVICE inline float16& operator=(int8_t val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(uint8_t val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(int16_t val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(uint16_t val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(int32_t val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(uint32_t val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(int64_t val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(uint64_t val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(float val) {
x = float16(val).x;
return *this;
}
HOSTDEVICE inline float16& operator=(double val) {
x = float16(val).x;
return *this;
}
// Conversion opertors
#ifdef PADDLE_CUDA_FP16
HOSTDEVICE inline explicit operator half() const {
#if CUDA_VERSION >= 9000
__half_raw h;
h.x = x;
return half(h);
#else
half h;
h.x = x;
return h;
#endif // CUDA_VERSION >= 9000
}
#endif // PADDLE_CUDA_FP16
HOSTDEVICE inline explicit operator Eigen::half() const {
Eigen::half h;
h.x = x;
return h;
}
#ifdef PADDLE_WITH_NATIVE_FP16
HOSTDEVICE inline explicit operator float16_t() const {
return *reinterpret_cast<const float16_t*>(this);
}
#endif
HOSTDEVICE inline explicit operator float() const {
#if defined(PADDLE_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
half tmp = *reinterpret_cast<const half*>(this);
return __half2float(tmp);
#elif defined(PADDLE_WITH_NATIVE_FP16)
float16x4_t res = vld1_dup_f16(reinterpret_cast<const float16_t*>(this));
return vgetq_lane_f32(vcvt_f32_f16(res), 0);
#elif defined(__F16C__)
return _cvtsh_ss(this->x);
#else
// Conversion routine adapted from
// http://stackoverflow.com/questions/1659440/32-bit-to-16-bit-floating-point-conversion
Bits v;
v.ui = this->x;
int32_t sign = v.si & sigC;
v.si ^= sign;
sign <<= shiftSign;
v.si ^= ((v.si + minD) ^ v.si) & -(v.si > subC);
v.si ^= ((v.si + maxD) ^ v.si) & -(v.si > maxC);
Bits s;
s.si = mulC;
s.f *= v.si;
int32_t mask = -(norC > v.si);
v.si <<= shift;
v.si ^= (s.si ^ v.si) & mask;
v.si |= sign;
return v.f;
#endif
}
HOSTDEVICE inline explicit operator bool() const { return (x & 0x7fff) != 0; }
HOSTDEVICE inline explicit operator int8_t() const {
return static_cast<int8_t>(float(*this));
}
HOSTDEVICE inline explicit operator uint8_t() const {
return static_cast<uint8_t>(float(*this));
}
HOSTDEVICE inline explicit operator int16_t() const {
return static_cast<int16_t>(float(*this));
}
HOSTDEVICE inline explicit operator uint16_t() const {
return static_cast<uint16_t>(float(*this));
}
HOSTDEVICE inline explicit operator int32_t() const {
return static_cast<int32_t>(float(*this));
}
HOSTDEVICE inline explicit operator uint32_t() const {
return static_cast<uint32_t>(float(*this));
}
HOSTDEVICE inline explicit operator int64_t() const {
return static_cast<int64_t>(float(*this));
}
HOSTDEVICE inline explicit operator uint64_t() const {
return static_cast<uint64_t>(float(*this));
}
HOSTDEVICE inline explicit operator double() const {
return static_cast<double>(float(*this));
}
private:
union Bits {
float f;
int32_t si;
uint32_t ui;
};
static const int shift = 13;
static const int shiftSign = 16;
static const int32_t infN = 0x7F800000;
static const int32_t maxN = 0x477FE000; // max flt16 as flt32
static const int32_t minN = 0x38800000; // min flt16 normal as flt32
static const int32_t sigN = 0x80000000; // sign bit
static constexpr int32_t infC = infN >> shift;
static constexpr int32_t nanN = (infC + 1)
<< shift; // minimum flt16 nan as float32
static constexpr int32_t maxC = maxN >> shift;
static constexpr int32_t minC = minN >> shift;
static constexpr int32_t sigC = sigN >> shiftSign;
static const int32_t mulN = 0x52000000; // (1 << 23) / minN
static const int32_t mulC = 0x33800000; // minN / (1 << (23 - shift))
static const int32_t subC = 0x003FF; // max flt32 subnormal downshifted
static const int32_t norC = 0x00400; // min flt32 normal downshifted
static constexpr int32_t maxD = infC - maxC - 1;
static constexpr int32_t minD = minC - subC - 1;
};
// Arithmetic operators on GPU
// CUDA 9.0 provides built-in arithmetic operators for half while
// CUDA 7.5 and 8.0 do not. The arithmetic operators defined here are
// for users to write similar CUDA code in CUDA 7.5 and 8.0 as in
// CUDA 9.0 regarding the half data type.
#if defined(PADDLE_CUDA_FP16) && CUDA_VERSION < 9000
DEVICE inline half operator+(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hadd(a, b);
#else
float res = float(float16(a)) + float(float16(b));
return half(float16(res));
#endif
}
DEVICE inline half operator-(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hsub(a, b);
#else
float res = float(float16(a)) - float(float16(b));
return half(float16(res));
#endif
}
DEVICE inline half operator*(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hmul(a, b);
#else
float res = float(float16(a)) * float(float16(b));
return half(float16(res));
#endif
}
DEVICE inline half operator/(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
float num = __half2float(a);
float denom = __half2float(b);
return __float2half(num / denom);
#else
float res = float(float16(a)) / float(float16(b));
return half(float16(res));
#endif
}
DEVICE inline half operator-(const half& a) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hneg(a);
#else
float res = -float(float16(a));
return half(float16(res));
#endif
}
DEVICE inline half& operator+=(half& a, const half& b) {
a = a + b;
return a;
}
DEVICE inline half& operator-=(half& a, const half& b) {
a = a - b;
return a;
}
DEVICE inline half& operator*=(half& a, const half& b) {
a = a * b;
return a;
}
DEVICE inline half& operator/=(half& a, const half& b) {
a = a / b;
return a;
}
DEVICE inline bool operator==(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __heq(a, b);
#else
return float(float16(a)) == float(float16(b));
#endif
}
DEVICE inline bool operator!=(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hne(a, b);
#else
return float(float16(a)) != float(float16(b));
#endif
}
DEVICE inline bool operator<(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hlt(a, b);
#else
return float(float16(a)) < float(float16(b));
#endif
}
DEVICE inline bool operator<=(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hle(a, b);
#else
return float(float16(a)) <= float(float16(b));
#endif
}
DEVICE inline bool operator>(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hgt(a, b);
#else
return float(float16(a)) > float(float16(b));
#endif
}
DEVICE inline bool operator>=(const half& a, const half& b) {
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
return __hge(a, b);
#else
return float(float16(a)) >= float(float16(b));
#endif
}
#endif // PADDLE_CUDA_FP16
// Arithmetic operators on ARMv8.2-A CPU
#if defined(PADDLE_WITH_NATIVE_FP16)
HOST inline float16 operator+(const float16& a, const float16& b) {
float16 res;
asm volatile(
"ld1 {v0.h}[0], [%[a_ptr]]\n"
"ld1 {v1.h}[0], [%[b_ptr]]\n"
"fadd h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&(res.x))
: // clobbers
"memory", "v0", "v1");
return res;
}
HOST inline float16 operator-(const float16& a, const float16& b) {
float16 res;
asm volatile(
"ld1 {v0.h}[0], [%[a_ptr]]\n"
"ld1 {v1.h}[0], [%[b_ptr]]\n"
"fsub h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&(res.x))
: // clobbers
"memory", "v0", "v1");
return res;
}
HOST inline float16 operator*(const float16& a, const float16& b) {
float16 res;
asm volatile(
"ld1 {v0.h}[0], [%[a_ptr]]\n"
"ld1 {v1.h}[0], [%[b_ptr]]\n"
"fmul h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&(res.x))
: // clobbers
"memory", "v0", "v1");
return res;
}
HOST inline float16 operator/(const float16& a, const float16& b) {
float16 res;
asm volatile(
"ld1 {v0.h}[0], [%[a_ptr]]\n"
"ld1 {v1.h}[0], [%[b_ptr]]\n"
"fdiv h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&(res.x))
: // clobbers
"memory", "v0", "v1");
return res;
}
HOST inline float16 operator-(const float16& a) {
float16 res;
asm volatile(
"ld1 {v0.h}[0], [%[a_ptr]]\n"
"fneg h0, h0\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[res_ptr] "r"(&(res.x))
: // clobbers
"memory", "v0");
return res;
}
HOST inline float16& operator+=(float16& a, const float16& b) {
a = a + b;
return a;
}
HOST inline float16& operator-=(float16& a, const float16& b) {
a = a - b;
return a;
}
HOST inline float16& operator*=(float16& a, const float16& b) {
a = a * b;
return a;
}
HOST inline float16& operator/=(float16& a, const float16& b) {
a = a / b;
return a;
}
HOST inline bool operator==(const float16& a, const float16& b) {
uint16_t res;
asm volatile(
"ld1 {v0.h}[0], [%[a_ptr]]\n"
"ld1 {v1.h}[0], [%[b_ptr]]\n"
"fcmeq h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&res)
: // clobbers
"memory", "v0", "v1");
return (res & 0xffff) != 0;
}
HOST inline bool operator!=(const float16& a, const float16& b) {
return !(a == b);
}
HOST inline bool operator<(const float16& a, const float16& b) {
uint16_t res;
asm volatile(
"ld1 {v1.h}[0], [%[a_ptr]]\n"
"ld1 {v0.h}[0], [%[b_ptr]]\n"
"fcmgt h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&res)
: // clobbers
"memory", "v0", "v1");
return (res & 0xffff) != 0;
}
HOST inline bool operator<=(const float16& a, const float16& b) {
uint16_t res;
asm volatile(
"ld1 {v1.h}[0], [%[a_ptr]]\n"
"ld1 {v0.h}[0], [%[b_ptr]]\n"
"fcmge h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&res)
: // clobbers
"memory", "v0", "v1");
return (res & 0xffff) != 0;
}
HOST inline bool operator>(const float16& a, const float16& b) {
uint16_t res;
asm volatile(
"ld1 {v0.h}[0], [%[a_ptr]]\n"
"ld1 {v1.h}[0], [%[b_ptr]]\n"
"fcmgt h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&res)
: // clobbers
"memory", "v0", "v1");
return (res & 0xffff) != 0;
}
HOST inline bool operator>=(const float16& a, const float16& b) {
uint16_t res;
asm volatile(
"ld1 {v0.h}[0], [%[a_ptr]]\n"
"ld1 {v1.h}[0], [%[b_ptr]]\n"
"fcmge h0, h0, h1\n"
"st1 {v0.h}[0], [%[res_ptr]]\n"
: // outputs
: // inputs
[a_ptr] "r"(&(a.x)),
[b_ptr] "r"(&(b.x)),
[res_ptr] "r"(&res)
: // clobbers
"memory", "v0", "v1");
return (res & 0xffff) != 0;
}
// Arithmetic operators, software emulated on other CPU
#else
HOSTDEVICE inline float16 operator+(const float16& a, const float16& b) {
return float16(float(a) + float(b));
}
HOSTDEVICE inline float16 operator-(const float16& a, const float16& b) {
return float16(float(a) - float(b));
}
HOSTDEVICE inline float16 operator*(const float16& a, const float16& b) {
return float16(float(a) * float(b));
}
HOSTDEVICE inline float16 operator/(const float16& a, const float16& b) {
return float16(float(a) / float(b));
}
HOSTDEVICE inline float16 operator-(const float16& a) {
float16 res;
res.x = a.x ^ 0x8000;
return res;
}
HOSTDEVICE inline float16& operator+=(float16& a, const float16& b) {
a = float16(float(a) + float(b));
return a;
}
HOSTDEVICE inline float16& operator-=(float16& a, const float16& b) {
a = float16(float(a) - float(b));
return a;
}
HOSTDEVICE inline float16& operator*=(float16& a, const float16& b) {
a = float16(float(a) * float(b));
return a;
}
HOSTDEVICE inline float16& operator/=(float16& a, const float16& b) {
a = float16(float(a) / float(b));
return a;
}
HOSTDEVICE inline bool operator==(const float16& a, const float16& b) {
return float(a) == float(b);
}
HOSTDEVICE inline bool operator!=(const float16& a, const float16& b) {
return float(a) != float(b);
}
HOSTDEVICE inline bool operator<(const float16& a, const float16& b) {
return float(a) < float(b);
}
HOSTDEVICE inline bool operator<=(const float16& a, const float16& b) {
return float(a) <= float(b);
}
HOSTDEVICE inline bool operator>(const float16& a, const float16& b) {
return float(a) > float(b);
}
HOSTDEVICE inline bool operator>=(const float16& a, const float16& b) {
return float(a) >= float(b);
}
#endif
} // namespace paddle
paddle/math/tests/CMakeLists.txt
浏览文件 @ 2b47fb3d
......@@ -22,6 +22,7 @@ if(WITH_GPU)
link_paddle_test(test_Tensor)
CUDA_ADD_EXECUTABLE(test_lazyAssign test_lazyAssign.cu)
link_paddle_test(test_lazyAssign)
nv_test(test_float16_gpu SRCS test_float16.cu)
else()
compile_cu_as_cpp(test_Tensor.cu)
add_unittest(test_Tensor test_Tensor.cu)
......@@ -33,3 +34,4 @@ add_simple_unittest(test_FPException)
add_simple_unittest(test_GpuProfiler)
add_simple_unittest(test_BaseMatrix)
add_simple_unittest(test_Matrix)
add_simple_unittest(test_float16)
paddle/math/tests/test_float16.cpp 0 → 100644
浏览文件 @ 2b47fb3d
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/math/float16.h"
#include <gtest/gtest.h>
namespace paddle {
TEST(float16, conversion_cpu) {
// Explicit conversion from Eigen::half
EXPECT_EQ(float16(Eigen::half(1.0f)).x, 0x3c00);
EXPECT_EQ(float16(Eigen::half(0.5f)).x, 0x3800);
EXPECT_EQ(float16(Eigen::half(0.33333f)).x, 0x3555);
EXPECT_EQ(float16(Eigen::half(0.0f)).x, 0x0000);
EXPECT_EQ(float16(Eigen::half(-0.0f)).x, 0x8000);
EXPECT_EQ(float16(Eigen::half(65504.0f)).x, 0x7bff);
EXPECT_EQ(float16(Eigen::half(65536.0f)).x, 0x7c00);
// Conversion from float
EXPECT_EQ(float16(1.0f).x, 0x3c00);
EXPECT_EQ(float16(0.5f).x, 0x3800);
EXPECT_EQ(float16(0.33333f).x, 0x3555);
EXPECT_EQ(float16(0.0f).x, 0x0000);
EXPECT_EQ(float16(-0.0f).x, 0x8000);
EXPECT_EQ(float16(65504.0f).x, 0x7bff);
EXPECT_EQ(float16(65536.0f).x, 0x7c00);
// Conversion from double
EXPECT_EQ(float16(1.0).x, 0x3c00);
EXPECT_EQ(float16(0.5).x, 0x3800);
EXPECT_EQ(float16(0.33333).x, 0x3555);
EXPECT_EQ(float16(0.0).x, 0x0000);
EXPECT_EQ(float16(-0.0).x, 0x8000);
EXPECT_EQ(float16(65504.0).x, 0x7bff);
EXPECT_EQ(float16(65536.0).x, 0x7c00);
// Conversion from int
EXPECT_EQ(float16(-1).x, 0xbc00);
EXPECT_EQ(float16(0).x, 0x0000);
EXPECT_EQ(float16(1).x, 0x3c00);
EXPECT_EQ(float16(2).x, 0x4000);
EXPECT_EQ(float16(3).x, 0x4200);
// Conversion from bool
EXPECT_EQ(float16(true).x, 0x3c00);
EXPECT_EQ(float16(false).x, 0x0000);
// Default constructor
float16 v_def;
EXPECT_EQ(v_def.x, 0x0000);
// Assignment operator
float16 v_assign;
v_assign = v_def;
EXPECT_EQ(v_assign.x, 0x0000);
v_assign = Eigen::half(1.0f);
EXPECT_EQ(v_assign.x, 0x3c00);
v_assign = 0.5f;
EXPECT_EQ(v_assign.x, 0x3800);
v_assign = 0.33333;
EXPECT_EQ(v_assign.x, 0x3555);
v_assign = -1;
EXPECT_EQ(v_assign.x, 0xbc00);
v_assign = true;
EXPECT_EQ(v_assign.x, 0x3c00);
// Conversion operator
EXPECT_EQ(Eigen::half(float16(1.0f)).x, 0x3c00);
EXPECT_EQ(float(float16(0.5f)), 0.5f);
EXPECT_NEAR(double(float16(0.33333)), 0.33333, 0.0001);
EXPECT_EQ(int(float16(-1)), -1);
EXPECT_EQ(bool(float16(true)), true);
}
TEST(float16, arithmetic_cpu) {
EXPECT_EQ(float(float16(1) + float16(1)), 2);
EXPECT_EQ(float(float16(5) + float16(-5)), 0);
EXPECT_NEAR(float(float16(0.33333f) + float16(0.66667f)), 1.0f, 0.001);
EXPECT_EQ(float(float16(3) - float16(5)), -2);
EXPECT_NEAR(float(float16(0.66667f) - float16(0.33333f)), 0.33334f, 0.001);
EXPECT_NEAR(float(float16(3.3f) * float16(2.0f)), 6.6f, 0.01);
EXPECT_NEAR(float(float16(-2.1f) * float16(-3.0f)), 6.3f, 0.01);
EXPECT_NEAR(float(float16(2.0f) / float16(3.0f)), 0.66667f, 0.001);
EXPECT_EQ(float(float16(1.0f) / float16(2.0f)), 0.5f);
EXPECT_EQ(float(-float16(512.0f)), -512.0f);
EXPECT_EQ(float(-float16(-512.0f)), 512.0f);
}
TEST(float16, comparison_cpu) {
EXPECT_TRUE(float16(1.0f) == float16(1.0f));
EXPECT_FALSE(float16(-1.0f) == float16(-0.5f));
EXPECT_TRUE(float16(1.0f) != float16(0.5f));
EXPECT_FALSE(float16(-1.0f) != float16(-1.0f));
EXPECT_TRUE(float16(1.0f) < float16(2.0f));
EXPECT_FALSE(float16(-1.0f) < float16(-1.0f));
EXPECT_TRUE(float16(1.0f) <= float16(1.0f));
EXPECT_TRUE(float16(2.0f) > float16(1.0f));
EXPECT_FALSE(float16(-2.0f) > float16(-2.0f));
EXPECT_TRUE(float16(2.0f) >= float16(2.0f));
EXPECT_TRUE(float16(0.0f) == float16(-0.0f));
EXPECT_TRUE(float16(0.0f) <= float16(-0.0f));
EXPECT_TRUE(float16(0.0f) >= float16(-0.0f));
EXPECT_FALSE(float16(0.0f) < float16(-0.0f));
EXPECT_FALSE(float16(-0.0f) < float16(0.0f));
EXPECT_FALSE(float16(0.0f) > float16(-0.0f));
EXPECT_FALSE(float16(-0.0f) > float16(0.0f));
}
} // namespace paddle
paddle/math/tests/test_float16.cu 0 → 100644
浏览文件 @ 2b47fb3d
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/math/float16.h"
#include <gtest/gtest.h>
#include "paddle/utils/Logging.h"
#define ARITHMETIC_KERNEL(op_type, sign) \
__global__ void op_type(const half* in1, const half* in2, half* out) { \
out[0] = in1[0] sign in2[0]; \
}
#define COMPOUND_KERNEL(op_type, sign) \
__global__ void op_type(half* in1, const half* in2) { in1[0] sign in2[0]; }
#define COMPARISON_KERNEL(op_type, sign) \
__global__ void op_type(const half* in1, const half* in2, bool* out) { \
out[0] = in1[0] sign in2[0]; \
}
#define ARITHMETIC_KERNEL_LAUNCH(op_type) \
void Test##op_type(float v_in1, float v_in2, float v_out) { \
LOG(INFO) << "Test " << #op_type << " on GPU!"; \
half *in1, *in2, *out; \
half *d_in1, *d_in2, *d_out; \
int size = sizeof(half); \
cudaMalloc((void**)&d_in1, size); \
cudaMalloc((void**)&d_in2, size); \
cudaMalloc((void**)&d_out, size); \
in1 = (half*)malloc(size); \
in2 = (half*)malloc(size); \
out = (half*)malloc(size); \
in1[0] = half(float16(v_in1)); \
in2[0] = half(float16(v_in2)); \
cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); \
cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice); \
op_type<<<1, 1>>>(d_in1, d_in2, d_out); \
cudaMemcpy(out, d_out, size, cudaMemcpyDeviceToHost); \
EXPECT_EQ(float(float16(out[0])), v_out); \
free(in1); \
free(in2); \
free(out); \
cudaFree(d_in1); \
cudaFree(d_in2); \
cudaFree(d_out); \
}
#define COMPOUND_KERNEL_LAUNCH(op_type) \
void Test##op_type(float v_in1, float v_in2, float v_out) { \
LOG(INFO) << "Test " << #op_type << " on GPU!"; \
half *in1, *in2; \
half *d_in1, *d_in2; \
int size = sizeof(half); \
cudaMalloc((void**)&d_in1, size); \
cudaMalloc((void**)&d_in2, size); \
in1 = (half*)malloc(size); \
in2 = (half*)malloc(size); \
in1[0] = half(float16(v_in1)); \
in2[0] = half(float16(v_in2)); \
cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); \
cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice); \
op_type<<<1, 1>>>(d_in1, d_in2); \
cudaMemcpy(in1, d_in1, size, cudaMemcpyDeviceToHost); \
EXPECT_EQ(float(float16(in1[0])), v_out); \
free(in1); \
free(in2); \
cudaFree(d_in1); \
cudaFree(d_in2); \
}
#define COMPARISON_KERNEL_LAUNCH(op_type) \
void Test##op_type(float v_in1, float v_in2, bool v_out) { \
LOG(INFO) << "Test " << #op_type << " on GPU!"; \
half *in1, *in2; \
half *d_in1, *d_in2; \
bool *out, *d_out; \
int size = sizeof(half); \
cudaMalloc((void**)&d_in1, size); \
cudaMalloc((void**)&d_in2, size); \
cudaMalloc((void**)&d_out, 1); \
in1 = (half*)malloc(size); \
in2 = (half*)malloc(size); \
out = (bool*)malloc(1); \
in1[0] = half(float16(v_in1)); \
in2[0] = half(float16(v_in2)); \
cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); \
cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice); \
op_type<<<1, 1>>>(d_in1, d_in2, d_out); \
cudaMemcpy(out, d_out, 1, cudaMemcpyDeviceToHost); \
EXPECT_EQ(out[0], v_out); \
free(in1); \
free(in2); \
free(out); \
cudaFree(d_in1); \
cudaFree(d_in2); \
cudaFree(d_out); \
}
#ifdef PADDLE_CUDA_FP16
namespace paddle {
#if CUDA_VERSION < 9000
ARITHMETIC_KERNEL(Add, +)
ARITHMETIC_KERNEL(Sub, -)
ARITHMETIC_KERNEL(Mul, *)
ARITHMETIC_KERNEL(Div, /)
ARITHMETIC_KERNEL_LAUNCH(Add)
ARITHMETIC_KERNEL_LAUNCH(Sub)
ARITHMETIC_KERNEL_LAUNCH(Mul)
ARITHMETIC_KERNEL_LAUNCH(Div)
// Negative sign kernel
__global__ void Neg(half* in) { in[0] = -in[0]; }
void TestNeg(float v_in, float v_out) {
LOG(INFO) << "Test Neg on GPU!";
half *in, *d_in;
int size = sizeof(half);
cudaMalloc((void**)&d_in, size);
in = (half*)malloc(size);
in[0] = half(float16(v_in));
cudaMemcpy(d_in, in, size, cudaMemcpyHostToDevice);
Neg<<<1, 1>>>(d_in);
cudaMemcpy(in, d_in, size, cudaMemcpyDeviceToHost);
EXPECT_EQ(float(float16(in[0])), v_out);
free(in);
cudaFree(d_in);
}
COMPOUND_KERNEL(AddAssign, +=)
COMPOUND_KERNEL(SubAssign, -=)
COMPOUND_KERNEL(MulAssign, *=)
COMPOUND_KERNEL(DivAssign, /=)
COMPOUND_KERNEL_LAUNCH(AddAssign)
COMPOUND_KERNEL_LAUNCH(SubAssign)
COMPOUND_KERNEL_LAUNCH(MulAssign)
COMPOUND_KERNEL_LAUNCH(DivAssign)
COMPARISON_KERNEL(Equal, ==)
COMPARISON_KERNEL(NotEqual, !=)
COMPARISON_KERNEL(Less, <)
COMPARISON_KERNEL(LessEqual, <=)
COMPARISON_KERNEL(Greater, >)
COMPARISON_KERNEL(GreaterEqual, >=)
COMPARISON_KERNEL_LAUNCH(Equal)
COMPARISON_KERNEL_LAUNCH(NotEqual)
COMPARISON_KERNEL_LAUNCH(Less)
COMPARISON_KERNEL_LAUNCH(LessEqual)
COMPARISON_KERNEL_LAUNCH(Greater)
COMPARISON_KERNEL_LAUNCH(GreaterEqual)
TEST(float16, arithmetic_on_gpu) {
TestAdd(1, 2, 3);
TestSub(2, 1, 1);
TestMul(2, 3, 6);
TestDiv(6, 2, 3);
TestNeg(1, -1);
}
TEST(float16, compound_on_gpu) {
TestAddAssign(1, 2, 3);
TestSubAssign(2, 1, 1);
TestMulAssign(2, 3, 6);
TestDivAssign(6, 2, 3);
}
TEST(float16, comparision_on_gpu) {
TestEqual(1, 1, true);
TestEqual(1, 2, false);
TestNotEqual(2, 3, true);
TestNotEqual(2, 2, false);
TestLess(3, 4, true);
TestLess(3, 3, false);
TestLessEqual(3, 3, true);
TestLessEqual(3, 2, false);
TestGreater(4, 3, true);
TestGreater(4, 4, false);
TestGreaterEqual(4, 4, true);
TestGreaterEqual(4, 5, false);
}
#endif // CUDA_VERSION
TEST(float16, conversion_on_gpu) {
// Explicit conversion to and from cuda half
EXPECT_EQ(float16(half(float16(1.0f))).x, 0x3c00);
EXPECT_EQ(float16(half(float16(0.5f))).x, 0x3800);
EXPECT_EQ(float16(half(float16(0.33333f))).x, 0x3555);
EXPECT_EQ(float16(half(float16(0.0f))).x, 0x0000);
EXPECT_EQ(float16(half(float16(-0.0f))).x, 0x8000);
EXPECT_EQ(float16(half(float16(65504.0f))).x, 0x7bff);
EXPECT_EQ(float16(half(float16(65536.0f))).x, 0x7c00);
// Assignment operator
float16 v_assign;
v_assign = half(float16(1.0f));
EXPECT_EQ(v_assign.x, 0x3c00);
}
} // namespace paddle
#endif // PADDLE_CUDA_FP16
paddle/memory/detail/system_allocator.cc
浏览文件 @ 2b47fb3d
......@@ -43,7 +43,7 @@ void* CPUAllocator::Alloc(size_t& index, size_t size) {
void* p;
#ifdef PADDLE_USE_MKLDNN
#ifdef PADDLE_WITH_MKLDNN
// refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
// memory alignment
PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0);
......@@ -83,7 +83,7 @@ void* GPUAllocator::Alloc(size_t& index, size_t size) {
paddle::platform::GpuMemoryUsage(available, capacity);
// Reserve memory for page tables, etc.
size_t reserving = capacity - paddle::platform::GpuMaxAllocSize();
size_t reserving = 0.05 * capacity + paddle::platform::GpuMinChunkSize();
size_t usable = available > reserving ? available - reserving : 0;
// If remaining size no less than expected size, using general
......
paddle/memory/memory.cc
浏览文件 @ 2b47fb3d
......@@ -64,19 +64,21 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
int gpu_num = platform::GetCUDADeviceCount();
as = new BuddyAllocator*[gpu_num];
for (int gpu = 0; gpu < gpu_num; gpu++) {
platform::SetDeviceId(gpu);
as[gpu] = new BuddyAllocator(new detail::GPUAllocator,
platform::GpuMinChunkSize(),
platform::GpuMaxChunkSize());
as[gpu] = nullptr;
}
}
platform::SetDeviceId(gpu_id);
if (!as[gpu_id]) {
as[gpu_id] = new BuddyAllocator(new detail::GPUAllocator,
platform::GpuMinChunkSize(),
platform::GpuMaxChunkSize());
VLOG(10) << "\n\nNOTE: each GPU device use "
<< FLAGS_fraction_of_gpu_memory_to_use * 100
<< "% of GPU memory.\n"
<< "You can set environment variable '"
<< platform::kEnvFractionGpuMemoryToUse
<< "You can set GFlags environment variable '"
<< "FLAGS_fraction_of_gpu_memory_to_use"
<< "' to change the fraction of GPU usage.\n\n";
}
platform::SetDeviceId(gpu_id);
return as[gpu_id];
}
......
paddle/operators/CMakeLists.txt
浏览文件 @ 2b47fb3d
......@@ -138,7 +138,7 @@ function(op_library TARGET)
if ("${TARGET}" STREQUAL "nccl_op")
set(pybind_flag 1)
# It's enough to just adding one operator to pybind
file(APPEND ${pybind_file} "USE_GPU_ONLY_OP(ncclAllReduce);\n")
file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n")
endif()
# reduce_op contains several operators
......
paddle/operators/accuracy_op.cc
浏览文件 @ 2b47fb3d
......@@ -57,7 +57,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
const framework::ExecutionContext &ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Input<Tensor>("Out")->type()),
ctx.device_context());
ctx.GetPlace());
}
};
......
paddle/operators/accuracy_op.cu
浏览文件 @ 2b47fb3d
......@@ -104,5 +104,6 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
// FIXME(typhoonzero): types of T is for inference data.
// label data is always int64
REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel<float>,
paddle::operators::AccuracyOpCUDAKernel<double>);
REGISTER_OP_CUDA_KERNEL(accuracy,
paddle::operators::AccuracyOpCUDAKernel<float>,
paddle::operators::AccuracyOpCUDAKernel<double>);
paddle/operators/accuracy_op.h
浏览文件 @ 2b47fb3d
......@@ -21,7 +21,7 @@ namespace operators {
using Tensor = framework::Tensor;
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class AccuracyKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
......
paddle/operators/activation_op.cc
浏览文件 @ 2b47fb3d
......@@ -44,9 +44,9 @@ class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput("X", "Input of Sigmoid operator");
AddOutput("Y", "Output of Sigmoid operator");
AddComment(R"DOC(
Sigmoid Activation Operator.
Sigmoid Activation Operator
$y = 1 / (1 + e^{-x})$
$$y = \frac{1}{1 + e^{-x}}$$
)DOC");
}
......@@ -60,9 +60,9 @@ class LogSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput("X", "Input of LogSigmoid operator");
AddOutput("Y", "Output of LogSigmoid operator");
AddComment(R"DOC(
Logsigmoid Activation Operator.
Logsigmoid Activation Operator
$y = \log(1 / (1 + e^{-x}))$
$$y = \log \frac{1}{1 + e^{-x}}$$
)DOC");
}
......@@ -506,6 +506,22 @@ It is recommended to use the defaults for this activation.
}
};
class SwishOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SwishOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "Input of Swish operator");
AddOutput("Y", "Output of Swish operator");
AddAttr<float>("beta", "Constant beta of swish operator").SetDefault(1.0f);
AddComment(R"DOC(
Swish Activation Operator.
$$y = \frac{x}{1 + e^{- \beta x}}$$
)DOC");
}
};
} // namespace operators
} // namespace paddle
......@@ -592,16 +608,20 @@ REGISTER_OP(thresholded_relu, ops::ActivationOp, ops::ThresholdedReluOpMaker,
REGISTER_OP(hard_sigmoid, ops::ActivationOp, ops::HardSigmoidOpMaker,
hard_sigmoid_grad, ops::ActivationOpGrad);
#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor) \
REGISTER_OP_CPU_KERNEL( \
act_type, \
ops::ActivationKernel<paddle::platform::CPUPlace, ops::functor<float>>, \
ops::ActivationKernel<paddle::platform::CPUPlace, \
ops::functor<double>>); \
REGISTER_OP_CPU_KERNEL( \
act_type##_grad, ops::ActivationGradKernel<paddle::platform::CPUPlace, \
ops::grad_functor<float>>, \
ops::ActivationGradKernel<paddle::platform::CPUPlace, \
REGISTER_OP(swish, ops::ActivationOp, ops::SwishOpMaker, swish_grad,
ops::ActivationOpGrad);
#define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor) \
REGISTER_OP_CPU_KERNEL( \
act_type, ops::ActivationKernel<paddle::platform::CPUDeviceContext, \
ops::functor<float>>, \
ops::ActivationKernel<paddle::platform::CPUDeviceContext, \
ops::functor<double>>); \
REGISTER_OP_CPU_KERNEL( \
act_type##_grad, \
ops::ActivationGradKernel<paddle::platform::CPUDeviceContext, \
ops::grad_functor<float>>, \
ops::ActivationGradKernel<paddle::platform::CPUDeviceContext, \
ops::grad_functor<double>>);
FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);
paddle/operators/activation_op.cu
浏览文件 @ 2b47fb3d
......@@ -17,16 +17,17 @@
namespace ops = paddle::operators;
#define REGISTER_ACTIVATION_GPU_KERNEL(act_type, functor, grad_functor) \
REGISTER_OP_GPU_KERNEL( \
act_type, \
ops::ActivationKernel<paddle::platform::GPUPlace, ops::functor<float>>, \
ops::ActivationKernel<paddle::platform::GPUPlace, \
ops::functor<double>>); \
REGISTER_OP_GPU_KERNEL( \
act_type##_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, \
ops::grad_functor<float>>, \
ops::ActivationGradKernel<paddle::platform::GPUPlace, \
#define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, functor, grad_functor) \
REGISTER_OP_CUDA_KERNEL( \
act_type, ops::ActivationKernel<paddle::platform::CUDADeviceContext, \
ops::functor<float>>, \
ops::ActivationKernel<paddle::platform::CUDADeviceContext, \
ops::functor<double>>); \
REGISTER_OP_CUDA_KERNEL( \
act_type##_grad, \
ops::ActivationGradKernel<paddle::platform::CUDADeviceContext, \
ops::grad_functor<float>>, \
ops::ActivationGradKernel<paddle::platform::CUDADeviceContext, \
ops::grad_functor<double>>);
FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_GPU_KERNEL);
FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);
paddle/operators/activation_op.h
浏览文件 @ 2b47fb3d
......@@ -19,7 +19,7 @@
namespace paddle {
namespace operators {
template <typename Place, typename Functor>
template <typename DeviceContext, typename Functor>
class ActivationKernel
: public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
public:
......@@ -32,18 +32,19 @@ class ActivationKernel
auto x = framework::EigenVector<T>::Flatten(*X);
auto y = framework::EigenVector<T>::Flatten(*Y);
auto place = context.GetEigenDevice<Place>();
auto* place =
context.template device_context<DeviceContext>().eigen_device();
Functor functor;
auto attrs = functor.GetAttrs();
for (auto& attr : attrs) {
*attr.second = context.Attr<float>(attr.first);
}
functor(place, x, y);
functor(*place, x, y);
}
};
template <typename Place, typename Functor>
template <typename DeviceContext, typename Functor>
class ActivationGradKernel
: public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
public:
......@@ -59,13 +60,14 @@ class ActivationGradKernel
auto x = framework::EigenVector<T>::Flatten(*X);
auto y = framework::EigenVector<T>::Flatten(*Y);
auto dx = framework::EigenVector<T>::Flatten(*dX);
auto place = context.GetEigenDevice<Place>();
auto* place =
context.template device_context<DeviceContext>().eigen_device();
Functor functor;
auto attrs = functor.GetAttrs();
for (auto& attr : attrs) {
*attr.second = context.Attr<float>(attr.first);
}
functor(place, x, y, dy, dx);
functor(*place, x, y, dy, dx);
}
};
......@@ -700,6 +702,35 @@ struct HardSigmoidGradFunctor : public BaseActivationFunctor<T> {
}
};
template <typename T>
struct SwishFunctor : public BaseActivationFunctor<T> {
float beta;
typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
return {{"beta", &beta}};
}
template <typename Device, typename X, typename Y>
void operator()(Device d, X x, Y y) const {
y.device(d) = x / (static_cast<T>(1) + (static_cast<T>(-beta) * x).exp());
}
};
template <typename T>
struct SwishGradFunctor : public BaseActivationFunctor<T> {
float beta;
typename BaseActivationFunctor<T>::AttrPair GetAttrs() {
return {{"beta", &beta}};
}
template <typename Device, typename X, typename Y, typename dY, typename dX>
void operator()(Device d, X x, Y y, dY dy, dX dx) const {
auto temp1 = static_cast<T>(1) /
(static_cast<T>(1) + (static_cast<T>(-beta) * x).exp());
auto temp2 = temp1 * (static_cast<T>(1) - (beta * y));
dx.device(d) = dy * ((beta * y) + temp2);
}
};
} // namespace operators
} // namespace paddle
......@@ -730,4 +761,5 @@ struct HardSigmoidGradFunctor : public BaseActivationFunctor<T> {
__macro(elu, ELUFunctor, ELUGradFunctor); \
__macro(hard_shrink, HardShrinkFunctor, HardShrinkGradFunctor); \
__macro(hard_sigmoid, HardSigmoidFunctor, HardSigmoidGradFunctor); \
__macro(swish, SwishFunctor, SwishGradFunctor); \
__macro(thresholded_relu, ThresholdedReluFunctor, ThresholdedReluGradFunctor);
paddle/operators/adadelta_op.cc
浏览文件 @ 2b47fb3d
......@@ -92,12 +92,12 @@ for gradient descent.
Adadelta updates are as follows:
$$avgSquaredGradOut = \rho * avgSquaredGrad + (1 - \rho) * grad * grad \break
paramUpdate = - $\sqrt{((avgSquaredUpdate + \epsilon) /
(avgSquaredGrad_out + \epsilon))}$ * grad \break
avgSquaredUpdateOut = \rho * avgSquaredUpdate + (1 - \rho) *
{(paramUpdate)}^2 \break
paramOut = param + paramUpdate$$
$$
avg\_squared\_grad\_out = \rho * avg\_squared\_grad + (1 - \rho) * grad * grad \\
param\_update = - \sqrt{\frac{avg\_squared\_update + \epsilon}{avg\_squared\_grad\_out + \epsilon}} * grad \\
avg\_squared\_update\_out = \rho * avg\_squared\_update + (1 - \rho) * {param\_update}^2 \\
param\_out = param + param\_update
$$
)DOC");
}
......@@ -109,5 +109,5 @@ paramOut = param + paramUpdate$$
namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(adadelta, ops::AdadeltaOp, ops::AdadeltaOpMaker);
REGISTER_OP_CPU_KERNEL(
adadelta, ops::AdadeltaOpKernel<paddle::platform::CPUPlace, float>,
ops::AdadeltaOpKernel<paddle::platform::CPUPlace, double>);
adadelta, ops::AdadeltaOpKernel<paddle::platform::CPUDeviceContext, float>,
ops::AdadeltaOpKernel<paddle::platform::CPUDeviceContext, double>);
paddle/operators/adadelta_op.cu
浏览文件 @ 2b47fb3d
......@@ -16,6 +16,6 @@
#include "paddle/operators/adadelta_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
adadelta, ops::AdadeltaOpKernel<paddle::platform::GPUPlace, float>,
ops::AdadeltaOpKernel<paddle::platform::GPUPlace, double>);
REGISTER_OP_CUDA_KERNEL(
adadelta, ops::AdadeltaOpKernel<paddle::platform::CUDADeviceContext, float>,
ops::AdadeltaOpKernel<paddle::platform::CUDADeviceContext, double>);
paddle/operators/adadelta_op.h
浏览文件 @ 2b47fb3d
......@@ -19,7 +19,7 @@ limitations under the License. */
namespace paddle {
namespace operators {
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class AdadeltaOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
......@@ -51,7 +51,7 @@ class AdadeltaOpKernel : public framework::OpKernel<T> {
framework::EigenVector<T>::Flatten(*avg_squared_grad_out_tensor);
auto avg_squared_update_out =
framework::EigenVector<T>::Flatten(*avg_squared_update_out_tensor);
auto place = ctx.GetEigenDevice<Place>();
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
avg_squared_grad_out.device(place) =
rho * avg_squared_grad + (1 - rho) * grad.square();
......
paddle/operators/adagrad_op.cc
浏览文件 @ 2b47fb3d
......@@ -80,8 +80,8 @@ Adaptive Gradient Algorithm (Adagrad).
The update is done as follows:
$$momentOut = moment + grad * grad \break
paramOut = param - learningRate * grad / ($\sqrt{momentOut}$ + \epsilon) \break
$$moment\_out = moment + grad * grad \\
param\_out = param - \frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}
$$
The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
......@@ -100,8 +100,8 @@ size_t FindPos(const std::vector<int64_t>& rows, int64_t value) {
} // namespace
template <typename T>
struct SparseAdagradFunctor<platform::CPUPlace, T> {
void operator()(const platform::DeviceContext& context,
struct SparseAdagradFunctor<platform::CPUDeviceContext, T> {
void operator()(const platform::CPUDeviceContext& context,
const framework::SelectedRows& grad,
const framework::Tensor& learning_rate, T epsilon,
framework::Tensor* moment, framework::Tensor* param) {
......@@ -120,7 +120,7 @@ struct SparseAdagradFunctor<platform::CPUPlace, T> {
{static_cast<int64_t>(merge_rows.size()), grad_width}),
context.GetPlace());
math::SetConstant<platform::CPUPlace, T> constant_functor;
math::SetConstant<platform::CPUDeviceContext, T> constant_functor;
constant_functor(context, grad_merge->mutable_value(), 0.0);
auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
......@@ -144,9 +144,9 @@ struct SparseAdagradFunctor<platform::CPUPlace, T> {
auto gs =
framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
gs.device(*context.GetEigenDevice<platform::CPUPlace>()) = gm * gm;
gs.device(*context.eigen_device()) = gm * gm;
math::SelectedRowsAddToTensor<platform::CPUPlace, T> functor;
math::SelectedRowsAddToTensor<platform::CPUDeviceContext, T> functor;
functor(context, *grad_square, moment);
// 3. update parameter
......@@ -164,13 +164,13 @@ struct SparseAdagradFunctor<platform::CPUPlace, T> {
}
};
template struct SparseAdagradFunctor<platform::CPUPlace, float>;
template struct SparseAdagradFunctor<platform::CPUPlace, double>;
template struct SparseAdagradFunctor<platform::CPUDeviceContext, float>;
template struct SparseAdagradFunctor<platform::CPUDeviceContext, double>;
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(adagrad, ops::AdagradOp, ops::AdagradOpMaker);
REGISTER_OP_CPU_KERNEL(
adagrad, ops::AdagradOpKernel<paddle::platform::CPUPlace, float>,
ops::AdagradOpKernel<paddle::platform::CPUPlace, double>);
adagrad, ops::AdagradOpKernel<paddle::platform::CPUDeviceContext, float>,
ops::AdagradOpKernel<paddle::platform::CPUDeviceContext, double>);
paddle/operators/adagrad_op.cu
浏览文件 @ 2b47fb3d
......@@ -72,8 +72,8 @@ __global__ void SparseAdagradFunctorKernel(const T* grad, const int64_t* rows,
} // namespace
template <typename T>
struct SparseAdagradFunctor<platform::GPUPlace, T> {
void operator()(const platform::DeviceContext& context,
struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
void operator()(const platform::CUDADeviceContext& context,
const framework::SelectedRows& grad,
const framework::Tensor& learning_rate, T epsilon,
framework::Tensor* moment, framework::Tensor* param) {
......@@ -92,7 +92,7 @@ struct SparseAdagradFunctor<platform::GPUPlace, T> {
{static_cast<int64_t>(merge_rows.size()), grad_width}),
context.GetPlace());
math::SetConstant<platform::GPUPlace, T> constant_functor;
math::SetConstant<platform::CUDADeviceContext, T> constant_functor;
constant_functor(context, grad_merge->mutable_value(), 0.0);
auto* grad_merge_data = grad_merge->mutable_value()->data<T>();
......@@ -119,9 +119,9 @@ struct SparseAdagradFunctor<platform::GPUPlace, T> {
auto gs =
framework::EigenVector<T>::Flatten(*(grad_square->mutable_value()));
auto gm = framework::EigenVector<T>::Flatten(grad_merge->value());
gs.device(*context.GetEigenDevice<platform::GPUPlace>()) = gm * gm;
gs.device(*context.eigen_device()) = gm * gm;
math::SelectedRowsAddToTensor<platform::GPUPlace, T> functor;
math::SelectedRowsAddToTensor<platform::CUDADeviceContext, T> functor;
functor(context, *grad_square, moment);
// 3. update parameter
......@@ -139,13 +139,13 @@ struct SparseAdagradFunctor<platform::GPUPlace, T> {
}
};
template struct SparseAdagradFunctor<platform::GPUPlace, float>;
template struct SparseAdagradFunctor<platform::GPUPlace, double>;
template struct SparseAdagradFunctor<platform::CUDADeviceContext, float>;
template struct SparseAdagradFunctor<platform::CUDADeviceContext, double>;
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
adagrad, ops::AdagradOpKernel<paddle::platform::GPUPlace, float>,
ops::AdagradOpKernel<paddle::platform::GPUPlace, double>);
REGISTER_OP_CUDA_KERNEL(
adagrad, ops::AdagradOpKernel<paddle::platform::CUDADeviceContext, float>,
ops::AdagradOpKernel<paddle::platform::CUDADeviceContext, double>);
paddle/operators/adagrad_op.h
浏览文件 @ 2b47fb3d
......@@ -19,15 +19,15 @@ limitations under the License. */
namespace paddle {
namespace operators {
template <typename Place, typename T>
template <typename DeviceContext, typename T>
struct SparseAdagradFunctor {
void operator()(const platform::DeviceContext& context,
void operator()(const DeviceContext& context,
const framework::SelectedRows& grad,
const framework::Tensor& learning_rate, T epsilon,
framework::Tensor* moment, framework::Tensor* param);
};
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class AdagradOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
......@@ -52,11 +52,11 @@ class AdagradOpKernel : public framework::OpKernel<T> {
auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
auto place = ctx.GetEigenDevice<Place>();
auto* place = ctx.template device_context<DeviceContext>().eigen_device();
moment_out.device(place) = moment + grad * grad;
moment_out.device(*place) = moment + grad * grad;
Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
param_out.device(place) =
param_out.device(*place) =
param - lr.broadcast(m_dsize) * grad / (moment_out.sqrt() + epsilon);
} else if (grad_var->IsType<framework::SelectedRows>()) {
auto* param_tensor = ctx.Input<framework::Tensor>("Param");
......@@ -65,8 +65,9 @@ class AdagradOpKernel : public framework::OpKernel<T> {
auto* moment_tensor = ctx.Input<framework::Tensor>("Moment");
PADDLE_ENFORCE_EQ(moment_tensor, moment_out_tensor);
SparseAdagradFunctor<Place, T> functor;
functor(ctx.device_context(), *ctx.Input<framework::SelectedRows>("Grad"),
SparseAdagradFunctor<DeviceContext, T> functor;
functor(ctx.template device_context<DeviceContext>(),
*ctx.Input<framework::SelectedRows>("Grad"),
*ctx.Input<framework::Tensor>("LearningRate"), epsilon,
moment_out_tensor, param_out_tensor);
} else {
......
paddle/operators/adam_op.cc
浏览文件 @ 2b47fb3d
......@@ -112,11 +112,13 @@ adaptive estimates of lower-order moments.
Adam updates:
$$moment_1_{out} = \beta_1 * moment_1 + (1 - \beta_1) * grad \break
moment_2_{out} = \beta_2 * moment_2 + (1 - \beta_2) * grad * grad \break
learningRate = learningRate *
$\sqrt{(1 - \beta_2_{pow})}$ / (1 - \beta_1_{pow}) \break
paramOut = param - learningRate * moment_1/ ($\sqrt{(moment_2)} + \epsilon)$$
$$
moment\_1\_out = \beta_1 * moment\_1 + (1 - \beta_1) * grad \\
moment\_2_\out = \beta_2 * moment\_2 + (1 - \beta_2) * grad * grad \\
learning\_rate = learning\_rate *
\frac{\sqrt{1 - \beta_{2\_pow}}}{1 - \beta_{1\_pow}} \\
param\_out = param - learning\_rate * \frac{moment\_1}{\sqrt{moment\_2} + \epsilon}
$$
)DOC");
}
......@@ -126,6 +128,6 @@ paramOut = param - learningRate * moment_1/ ($\sqrt{(moment_2)} + \epsilon)$$
namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(adam, ops::AdamOp, ops::AdamOpMaker);
REGISTER_OP_CPU_KERNEL(adam,
ops::AdamOpKernel<paddle::platform::CPUPlace, float>,
ops::AdamOpKernel<paddle::platform::CPUPlace, double>);
REGISTER_OP_CPU_KERNEL(
adam, ops::AdamOpKernel<paddle::platform::CPUDeviceContext, float>,
ops::AdamOpKernel<paddle::platform::CPUDeviceContext, double>);
paddle/operators/adam_op.cu
浏览文件 @ 2b47fb3d
......@@ -16,6 +16,6 @@
#include "paddle/operators/adam_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(adam,
ops::AdamOpKernel<paddle::platform::GPUPlace, float>,
ops::AdamOpKernel<paddle::platform::GPUPlace, double>);
REGISTER_OP_CUDA_KERNEL(
adam, ops::AdamOpKernel<paddle::platform::CUDADeviceContext, float>,
ops::AdamOpKernel<paddle::platform::CUDADeviceContext, double>);
paddle/operators/adam_op.h
浏览文件 @ 2b47fb3d
......@@ -19,7 +19,7 @@ limitations under the License. */
namespace paddle {
namespace operators {
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class AdamOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
......@@ -52,17 +52,17 @@ class AdamOpKernel : public framework::OpKernel<T> {
auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
auto moment1_out = framework::EigenVector<T>::Flatten(*moment1_out_tensor);
auto moment2_out = framework::EigenVector<T>::Flatten(*moment2_out_tensor);
auto place = ctx.GetEigenDevice<Place>();
auto* place = ctx.template device_context<DeviceContext>().eigen_device();
moment1_out.device(place) = beta1 * moment1 + (1 - beta1) * grad;
moment2_out.device(place) = beta2 * moment2 + (1 - beta2) * grad.square();
moment1_out.device(*place) = beta1 * moment1 + (1 - beta1) * grad;
moment2_out.device(*place) = beta2 * moment2 + (1 - beta2) * grad.square();
// All of these are tensors of 1 element
auto lr_t = lr * (1 - beta2_pow).sqrt() / (1 - beta1_pow);
// Eigen does not support automatic broadcast
// Get dimensions of moment vector to broadcast lr_t
Eigen::DSizes<int, 1> m_dsize(moment1_out_tensor->numel());
param_out.device(place) =
param_out.device(*place) =
param -
lr_t.broadcast(m_dsize) *
(moment1_out / (moment2_out.sqrt() + epsilon));
......
paddle/operators/adamax_op.cc
浏览文件 @ 2b47fb3d
......@@ -107,10 +107,12 @@ Adam algorithm based on the infinity norm.
Adamax updates:
$$momentOut = \beta_1 * moment + (1 - \beta_1) * grad \break
infNormOut = max(\beta_2 * infNorm + \epsilon, |grad|) \break
learningRate = learningRate /(1 - \beta_1_{pow}) \break
paramOut = param - learningRate * momentPut / infNormOut$$
$$
moment\_out = \beta_1 * moment + (1 - \beta_1) * grad \\
inf\_norm\_out = max(\beta_2 * inf\_norm + \epsilon, |grad|) \\
learning\_rate = \frac{learning\_rate}{1 - \beta_{1\_pow}} \\
param\_out = param - learning\_rate * \frac{moment\_out}{inf\_norm\_out}
$$
The original paper does not have an epsilon attribute.
However, it is added here for numerical stability to prevent the
......@@ -125,6 +127,6 @@ division by 0 error.
namespace ops = paddle::operators;
REGISTER_OP_WITHOUT_GRADIENT(adamax, ops::AdamaxOp, ops::AdamaxOpMaker);
REGISTER_OP_CPU_KERNEL(adamax,
ops::AdamaxOpKernel<paddle::platform::CPUPlace, float>,
ops::AdamaxOpKernel<paddle::platform::CPUPlace, double>);
REGISTER_OP_CPU_KERNEL(
adamax, ops::AdamaxOpKernel<paddle::platform::CPUDeviceContext, float>,
ops::AdamaxOpKernel<paddle::platform::CPUDeviceContext, double>);
paddle/operators/adamax_op.cu
浏览文件 @ 2b47fb3d
......@@ -16,6 +16,6 @@
#include "paddle/operators/adamax_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(adamax,
ops::AdamaxOpKernel<paddle::platform::GPUPlace, float>,
ops::AdamaxOpKernel<paddle::platform::GPUPlace, double>);
REGISTER_OP_CUDA_KERNEL(
adamax, ops::AdamaxOpKernel<paddle::platform::CUDADeviceContext, float>,
ops::AdamaxOpKernel<paddle::platform::CUDADeviceContext, double>);
paddle/operators/adamax_op.h
浏览文件 @ 2b47fb3d
......@@ -19,7 +19,7 @@ limitations under the License. */
namespace paddle {
namespace operators {
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class AdamaxOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
......@@ -51,14 +51,14 @@ class AdamaxOpKernel : public framework::OpKernel<T> {
auto moment_out = framework::EigenVector<T>::Flatten(*moment_out_tensor);
auto inf_norm_out =
framework::EigenVector<T>::Flatten(*inf_norm_out_tensor);
auto place = ctx.GetEigenDevice<Place>();
auto* place = ctx.template device_context<DeviceContext>().eigen_device();
moment_out.device(place) = beta1 * moment + (1 - beta1) * grad;
inf_norm_out.device(place) =
moment_out.device(*place) = beta1 * moment + (1 - beta1) * grad;
inf_norm_out.device(*place) =
grad.abs().cwiseMax((beta2 * inf_norm) + epsilon);
auto lr_t = lr / (1 - beta1_pow);
Eigen::DSizes<int, 1> m_dsize(moment_out_tensor->numel());
param_out.device(place) =
param_out.device(*place) =
param - lr_t.broadcast(m_dsize) * (moment_out / inf_norm_out);
}
};
......
paddle/operators/auc_op.h
浏览文件 @ 2b47fb3d
......@@ -25,7 +25,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class AucKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
......
paddle/operators/batch_norm_op.cc
浏览文件 @ 2b47fb3d
......@@ -135,7 +135,8 @@ The required data format for this layer is one of the following:
};
template <typename T>
class BatchNormKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
class BatchNormKernel<platform::CPUDeviceContext, T>
: public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
const float epsilon = ctx.Attr<float>("epsilon");
......@@ -318,12 +319,12 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
PADDLE_THROW("can't find Y@GRAD");
}
return framework::OpKernelType(framework::ToDataType(t->type()),
ctx.device_context());
ctx.GetPlace());
}
};
template <typename T>
class BatchNormGradKernel<platform::CPUPlace, T>
class BatchNormGradKernel<platform::CPUDeviceContext, T>
: public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
......@@ -436,8 +437,9 @@ class BatchNormGradKernel<platform::CPUPlace, T>
namespace ops = paddle::operators;
REGISTER_OP(batch_norm, ops::BatchNormOp, ops::BatchNormOpMaker,
batch_norm_grad, ops::BatchNormGradOp);
REGISTER_OP_CPU_KERNEL(batch_norm,
ops::BatchNormKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
batch_norm,
ops::BatchNormKernel<paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(
batch_norm_grad,
ops::BatchNormGradKernel<paddle::platform::CPUPlace, float>);
ops::BatchNormGradKernel<paddle::platform::CPUDeviceContext, float>);
paddle/operators/batch_norm_op.cu.cc
浏览文件 @ 2b47fb3d
......@@ -47,7 +47,8 @@ void ExtractNCWHD(const framework::DDim &dims,
}
template <typename T>
class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
class BatchNormKernel<platform::CUDADeviceContext, T>
: public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
......@@ -121,11 +122,12 @@ class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
saved_mean->mutable_data<T>(ctx.GetPlace());
saved_variance->mutable_data<T>(ctx.GetPlace());
math::SetConstant<platform::GPUPlace, T> functor;
functor(ctx.device_context(), saved_mean, 0);
functor(ctx.device_context(), saved_variance, 0);
auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
math::SetConstant<platform::CUDADeviceContext, T> functor;
functor(dev_ctx, saved_mean, 0);
functor(dev_ctx, saved_variance, 0);
auto handle = ctx.cuda_device_context().cudnn_handle();
auto handle = dev_ctx.cudnn_handle();
// Now, depending on whether we are running test or not, we have two paths.
if (is_test) {
......@@ -171,7 +173,7 @@ class BatchNormKernel<platform::GPUPlace, T> : public framework::OpKernel<T> {
};
template <typename T>
class BatchNormGradKernel<platform::GPUPlace, T>
class BatchNormGradKernel<platform::CUDADeviceContext, T>
: public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
......@@ -244,11 +246,12 @@ class BatchNormGradKernel<platform::GPUPlace, T>
const void *saved_mean_data = saved_mean->template data<T>();
const void *saved_var_data = saved_var->template data<T>();
auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward(
ctx.cuda_device_context().cudnn_handle(), mode_,
CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(),
CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(), data_desc_,
x->template data<T>(), data_desc_, d_y->template data<T>(), data_desc_,
dev_ctx.cudnn_handle(), mode_, CudnnDataType<T>::kOne(),
CudnnDataType<T>::kZero(), CudnnDataType<T>::kOne(),
CudnnDataType<T>::kZero(), data_desc_, x->template data<T>(),
data_desc_, d_y->template data<T>(), data_desc_,
d_x->template mutable_data<T>(ctx.GetPlace()), bn_param_desc_,
scale->template data<T>(),
d_scale->template mutable_data<T>(ctx.GetPlace()),
......@@ -266,8 +269,9 @@ class BatchNormGradKernel<platform::GPUPlace, T>
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(batch_norm,
ops::BatchNormKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
REGISTER_OP_CUDA_KERNEL(
batch_norm,
ops::BatchNormKernel<paddle::platform::CUDADeviceContext, float>);
REGISTER_OP_CUDA_KERNEL(
batch_norm_grad,
ops::BatchNormGradKernel<paddle::platform::GPUPlace, float>);
ops::BatchNormGradKernel<paddle::platform::CUDADeviceContext, float>);
paddle/operators/batch_norm_op.h
浏览文件 @ 2b47fb3d
......@@ -34,13 +34,13 @@ inline TensorFormat StringToTensorFormat(const std::string& str) {
}
}
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class BatchNormKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override;
};
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class BatchNormGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override;
......
paddle/operators/bilinear_tensor_product_op.cc
浏览文件 @ 2b47fb3d
......@@ -159,9 +159,12 @@ REGISTER_OP(bilinear_tensor_product, ops::BilinearTensorProductOp,
ops::BilinearTensorProductOpGrad);
REGISTER_OP_CPU_KERNEL(
bilinear_tensor_product,
ops::BilinearTensorProductKernel<paddle::platform::CPUPlace, float>,
ops::BilinearTensorProductKernel<paddle::platform::CPUPlace, double>);
ops::BilinearTensorProductKernel<paddle::platform::CPUDeviceContext, float>,
ops::BilinearTensorProductKernel<paddle::platform::CPUDeviceContext,
double>);
REGISTER_OP_CPU_KERNEL(
bilinear_tensor_product_grad,
ops::BilinearTensorProductGradKernel<paddle::platform::CPUPlace, float>,
ops::BilinearTensorProductGradKernel<paddle::platform::CPUPlace, double>);
ops::BilinearTensorProductGradKernel<paddle::platform::CPUDeviceContext,
float>,
ops::BilinearTensorProductGradKernel<paddle::platform::CPUDeviceContext,
double>);
paddle/operators/bilinear_tensor_product_op.cu
浏览文件 @ 2b47fb3d
......@@ -16,11 +16,15 @@ limitations under the License. */
#include "paddle/operators/bilinear_tensor_product_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
REGISTER_OP_CUDA_KERNEL(
bilinear_tensor_product,
ops::BilinearTensorProductKernel<paddle::platform::GPUPlace, float>,
ops::BilinearTensorProductKernel<paddle::platform::GPUPlace, double>);
REGISTER_OP_GPU_KERNEL(
ops::BilinearTensorProductKernel<paddle::platform::CUDADeviceContext,
float>,
ops::BilinearTensorProductKernel<paddle::platform::CUDADeviceContext,
double>);
REGISTER_OP_CUDA_KERNEL(
bilinear_tensor_product_grad,
ops::BilinearTensorProductGradKernel<paddle::platform::GPUPlace, float>,
ops::BilinearTensorProductGradKernel<paddle::platform::GPUPlace, double>);
ops::BilinearTensorProductGradKernel<paddle::platform::CUDADeviceContext,
float>,
ops::BilinearTensorProductGradKernel<paddle::platform::CUDADeviceContext,
double>);
paddle/operators/bilinear_tensor_product_op.h
浏览文件 @ 2b47fb3d
......@@ -27,7 +27,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class BilinearTensorProductKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
......@@ -46,7 +46,8 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
int out_dim = weight_dims[0];
auto x_dim = weight_dims[1];
auto y_dim = weight_dims[2];
auto place = ctx.GetEigenDevice<Place>();
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
// Create the intermediate variable to caculate the result of
// Input(X) multiplied by Input(Weight_i), the formula is:
......@@ -60,9 +61,9 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
auto output_col_vec = output_mat.chip(i, 1);
Tensor weight_mat =
weight->Slice(i, i + 1).Resize(framework::make_ddim({x_dim, y_dim}));
math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasNoTrans,
batch_size, y_dim, x_dim, 1, x->data<T>(),
weight_mat.data<T>(), 0, left_mul.data<T>());
math::gemm<DeviceContext, T>(dev_ctx, CblasNoTrans, CblasNoTrans,
batch_size, y_dim, x_dim, 1, x->data<T>(),
weight_mat.data<T>(), 0, left_mul.data<T>());
output_col_vec.device(place) =
(left_mul_mat * y_mat).sum(Eigen::DSizes<int, 1>(1));
}
......@@ -74,7 +75,7 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
}
};
template <typename Place, typename T>
template <typename DeviceContext, typename T>
class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
......@@ -96,8 +97,8 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
auto x_mat = EigenMatrix<T>::From(*x);
auto y_mat = EigenMatrix<T>::From(*y);
auto d_out_mat = EigenMatrix<T>::From(*d_out);
auto place = ctx.GetEigenDevice<Place>();
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
// Create the intermediate variable to caculate the Output(Y@Grad).
Tensor x_scale;
x_scale.mutable_data<T>(framework::make_ddim({batch_size, x_dim}),
......@@ -110,18 +111,18 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
ctx.GetPlace());
auto y_scale_mat = EigenMatrix<T>::From(y_scale);
math::SetConstant<Place, T> set_zero;
math::SetConstant<DeviceContext, T> set_zero;
// Set Output(X@Grad) be zero.
if (d_x) {
d_x->mutable_data<T>(ctx.GetPlace());
set_zero(ctx.device_context(), d_x, static_cast<T>(0));
set_zero(dev_ctx, d_x, static_cast<T>(0));
}
// Set Output(Y@Grad) be zero.
if (d_y) {
d_y->mutable_data<T>(ctx.GetPlace());
set_zero(ctx.device_context(), d_y, static_cast<T>(0));
set_zero(dev_ctx, d_y, static_cast<T>(0));
}
// Caculate the Output(X@Grad) and Output(Y@Grad).
......@@ -137,18 +138,18 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_x) *
y_mat;
math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasTrans,
batch_size, x_dim, y_dim, 1, y_scale.data<T>(),
weight_i.data<T>(), 1, d_x->data<T>());
math::gemm<DeviceContext, T>(
dev_ctx, CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1,
y_scale.data<T>(), weight_i.data<T>(), 1, d_x->data<T>());
}
if (d_y) {
x_scale_mat.device(place) =
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_y) *
x_mat;
math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasNoTrans,
batch_size, y_dim, x_dim, 1, x_scale.data<T>(),
weight_i.data<T>(), 1, d_y->data<T>());
math::gemm<DeviceContext, T>(
dev_ctx, CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1,
x_scale.data<T>(), weight_i.data<T>(), 1, d_y->data<T>());
}
}
}
......@@ -165,9 +166,9 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_weight) *
x_mat;
math::gemm<Place, T>(ctx.device_context(), CblasTrans, CblasNoTrans,
x_dim, y_dim, batch_size, 1, x_scale.data<T>(),
y->data<T>(), 0, d_weight_i.data<T>());
math::gemm<DeviceContext, T>(dev_ctx, CblasTrans, CblasNoTrans, x_dim,
y_dim, batch_size, 1, x_scale.data<T>(),
y->data<T>(), 0, d_weight_i.data<T>());
}
}
......
paddle/operators/cast_op.cc
浏览文件 @ 2b47fb3d
......@@ -68,10 +68,11 @@ class CastOpGradMaker : public framework::SingleGradOpDescMaker {
} // namespace paddle
namespace ops = paddle::operators;
using CPU = paddle::platform::CPUPlace;
using CPU = paddle::platform::CPUDeviceContext;
REGISTER_OP_WITH_KERNEL(cast, ops::CastOpGradMaker, ops::CastOpInferShape,
ops::CastOpProtoMaker);
REGISTER_OP_CPU_KERNEL(cast, ops::CastOpKernel<CPU, float>,
ops::CastOpKernel<CPU, double>,
ops::CastOpKernel<CPU, int>,
ops::CastOpKernel<CPU, int64_t>);
ops::CastOpKernel<CPU, int64_t>,
ops::CastOpKernel<CPU, bool>);
paddle/operators/cast_op.cu
浏览文件 @ 2b47fb3d
......@@ -16,7 +16,8 @@
template <typename T>
using CastOpKernel =
paddle::operators::CastOpKernel<paddle::platform::GPUPlace, T>;
paddle::operators::CastOpKernel<paddle::platform::CUDADeviceContext, T>;
REGISTER_OP_GPU_KERNEL(cast, CastOpKernel<float>, CastOpKernel<double>,
CastOpKernel<int>, CastOpKernel<int64_t>);
REGISTER_OP_CUDA_KERNEL(cast, CastOpKernel<float>, CastOpKernel<double>,
CastOpKernel<int>, CastOpKernel<int64_t>,
CastOpKernel<bool>);
paddle/operators/cast_op.h
浏览文件 @ 2b47fb3d
......@@ -27,13 +27,13 @@ struct CastOpTransformFunctor {
HOSTDEVICE OutT operator()(InT in) const { return static_cast<OutT>(in); }
};
template <typename Place, typename InT>
template <typename DeviceContext, typename InT>
struct CastOpFunctor {
const framework::Tensor* in_;
framework::Tensor* out_;
const platform::DeviceContext& ctx_;
const DeviceContext& ctx_;
CastOpFunctor(const framework::Tensor* in, framework::Tensor* out,
const platform::DeviceContext& ctx)
const DeviceContext& ctx)
: in_(in), out_(out), ctx_(ctx) {}
template <typename OutT>
......@@ -42,13 +42,13 @@ struct CastOpFunctor {
auto numel = in_->numel();
auto* in_end = in_begin + numel;
auto* out_begin = out_->mutable_data<OutT>(ctx_.GetPlace());
platform::Transform<Place> trans;
platform::Transform<DeviceContext> trans;
trans(ctx_, in_begin, in_end, out_begin,
CastOpTransformFunctor<InT, OutT>());
}
};
template <typename Place, typename InT>
template <typename DeviceContext, typename InT>
class CastOpKernel : public framework::OpKernel<InT> {
public:
void Compute(const framework::ExecutionContext& context) const override {
......@@ -56,7 +56,8 @@ class CastOpKernel : public framework::OpKernel<InT> {
auto* out = context.Output<framework::Tensor>("Out");
framework::VisitDataType(
static_cast<framework::DataType>(context.Attr<int>("out_dtype")),
CastOpFunctor<Place, InT>(in, out, context.device_context()));
CastOpFunctor<DeviceContext, InT>(
in, out, context.template device_context<DeviceContext>()));
}
};
......
paddle/operators/chunk_eval_op.cc
浏览文件 @ 2b47fb3d
......@@ -58,9 +58,10 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Inference",
"(Tensor, default: Tensor<int>). Predictions from the network.");
"(Tensor, default: Tensor<int64_t>). "
"Predictions from the network.");
AddInput("Label",
"(Tensor, default: Tensor<int>). The true tag sequences.");
"(Tensor, default: Tensor<int64_t>). The true tag sequences.");
AddOutput("Precision",
"(float). The evaluated precision (called positive predictive "
"value) of chunks on the given mini-batch.");
......@@ -84,7 +85,7 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
.SetDefault(std::vector<int>{});
AddComment(R"DOC(
For some basics of chunking, please refer to
‘Chunking with Support Vector Mechines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>’.
‘Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>’.
CheckEvalOp computes the precision, recall, and F1-score of chunk detection,
......@@ -97,7 +98,7 @@ Here is a NER example of labeling for these tagging schemes:
IOE: I-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O E-LOC
IOBES: B-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O S-LOC
There are three chunk types(named entity types) including PER(person), ORG(orgnazation)
There are three chunk types(named entity types) including PER(person), ORG(organization)
and LOC(LOCATION), and we can see that the labels have the form <tag type>-<chunk type>.
Since the calculations actually use label ids rather than labels, extra attention
......
paddle/operators/chunk_eval_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/clip_by_norm_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/clip_by_norm_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/clip_by_norm_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/clip_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/clip_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/clip_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/compare_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/compare_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/concat_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/concat_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/concat_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/conv_cudnn_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/conv_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/conv_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/conv_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/conv_shift_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/conv_shift_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/cos_sim_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/cos_sim_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/cos_sim_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/crf_decoding_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/crf_decoding_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/crop_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/crop_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/crop_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/cross_entropy_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/dropout_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/dropout_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/dropout_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/expand_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/expand_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/expand_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/fill_op.cc 0 → 100644
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/ftrl_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/ftrl_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/ftrl_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gather.cu.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gather_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gather_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gru_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gru_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gru_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gru_unit_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gru_unit_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/gru_unit_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/hinge_loss_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/hinge_loss_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/hinge_loss_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/huber_loss_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/huber_loss_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/huber_loss_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/l1_norm_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/l1_norm_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/l1_norm_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lod_reset_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lod_reset_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/log_loss_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/log_loss_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/log_loss_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/logical_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/logical_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/logical_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lookup_table_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lrn_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lrn_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lrn_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lstm_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lstm_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lstm_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lstm_unit_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/lstm_unit_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/gru_compute.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/im2col.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/im2col.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/im2col.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/matmul.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/maxouting.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/maxouting.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/maxouting.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/pooling.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/pooling.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/pooling.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/softmax.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/softmax.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/softmax.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/unpooling.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/unpooling.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/unpooling.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/vol2col.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/vol2col.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/math/vol2col.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/matmul_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/matmul_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/matmul_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/maxout_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/maxout_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/maxout_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/mean_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/mean_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/mean_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/minus_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/minus_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/minus_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/momentum_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/momentum_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/momentum_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/mul_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/mul_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/mul_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/multiplex_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/multiplex_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/multiplex_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/nccl_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/nccl_op_test.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/nce_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/nce_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/net_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/net_op_test.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/pad_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/pad_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/pad_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/pool_cudnn_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/pool_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/pool_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/pool_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/prelu_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/prelu_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/prelu_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/proximal_gd_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/proximal_gd_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/proximal_gd_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/rank_loss_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/rank_loss_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/rank_loss_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/recurrent_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/reduce_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/reduce_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/reduce_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/reshape_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/reshape_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/reshape_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/rmsprop_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/rmsprop_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/rmsprop_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/roi_pool_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/roi_pool_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/roi_pool_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/row_conv_op.cc 0 → 100644
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/row_conv_op.cu 0 → 100644
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/row_conv_op.h 0 → 100644
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/scale_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/scale_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/scale_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/scatter_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/scatter_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/seq_expand_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/seq_expand_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/seq_expand_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sequence_conv_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sequence_pool_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sgd_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sgd_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sgd_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sign_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sign_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sign_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/softmax_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/softmax_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/softmax_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/split_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/split_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sum_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sum_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/sum_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/top_k_op.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/top_k_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/transpose_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/transpose_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/transpose_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/unpool_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/unpool_op.cu.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/unpool_op.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/operators/while_op.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/platform/device_context.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/platform/device_context.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/platform/dynload/cudnn.cc
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/platform/dynload/cudnn.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/platform/transform.h
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/platform/transform_test.cu
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/scripts/check_env.sh 0 → 100755
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/scripts/docker/README.md
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/scripts/docker/build.sh
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/scripts/submit_local.sh.in
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/scripts/travis/build_doc.sh
浏览文件 @ 2b47fb3d
此差异已折叠。
paddle/utils/Flags.cpp
浏览文件 @ 2b47fb3d
此差异已折叠。
proto/ModelConfig.proto
浏览文件 @ 2b47fb3d
此差异已折叠。
python/CMakeLists.txt
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/__init__.py
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/dataset/common.py
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/fluid/framework.py
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/fluid/layers.py
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/fluid/nets.py
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/fluid/optimizer.py
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/fluid/registry.py 0 → 100644
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/fluid/tests/test_fill_op.py 0 → 100644
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/fluid/tests/test_registry.py 0 → 100644
浏览文件 @ 2b47fb3d
此差异已折叠。
python/paddle/v2/fluid/tests/test_row_conv_op.py 0 → 100644
浏览文件 @ 2b47fb3d
此差异已折叠。
python/setup.py.in
浏览文件 @ 2b47fb3d
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