Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 ae5923e2 编写于 作者: C chengduoZH
浏览文件
操作

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

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/add_gather_and_BCast_op_handle
上级 e7684911 f605f647
展开全部 隐藏空白更改
内联 并排
Showing with 3054 addition and 934 deletion
benchmark/fluid/mnist.py
浏览文件 @ ae5923e2
......@@ -139,9 +139,6 @@ def run_benchmark(model, args):
# inference program
inference_program = fluid.default_main_program().clone()
with fluid.program_guard(inference_program):
inference_program = fluid.io.get_inference_program(
target_vars=[batch_acc, batch_size_tensor])
# Optimization
opt = fluid.optimizer.AdamOptimizer(
......@@ -161,7 +158,7 @@ def run_benchmark(model, args):
train_reader = paddle.batch(
paddle.dataset.mnist.train(), batch_size=args.batch_size)
accuracy = fluid.average.WeightedAverage()
accuracy = fluid.metrics.Accuracy()
iters, num_samples, start_time = 0, 0, time.time()
for pass_id in range(args.pass_num):
accuracy.reset()
......@@ -184,7 +181,7 @@ def run_benchmark(model, args):
"label": y_data},
fetch_list=[avg_cost, batch_acc, batch_size_tensor]
) # The accuracy is the accumulation of batches, but not the current batch.
accuracy.add(value=outs[1], weight=outs[2])
accuracy.update(value=outs[1], weight=outs[2])
iters += 1
num_samples += len(y_data)
loss = np.array(outs[0])
......
cmake/external/mkldnn.cmake
浏览文件 @ ae5923e2
......@@ -36,7 +36,8 @@ MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR})
INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR}) # For MKLDNN code to include internal headers.
INCLUDE_DIRECTORIES(${THIRD_PARTY_PATH}/install) # For Paddle code to include mkldnn.h
IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
SET(MKLDNN_DEPENDS ${MKLML_PROJECT})
......
cmake/external/nccl.cmake 已删除 100644 → 0
浏览文件 @ e7684911
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
if(NOT WITH_GPU)
return()
endif()
include(ExternalProject)
set(NCCL_SOURCE_DIR ${THIRD_PARTY_PATH}/nccl)
include_directories(${NCCL_SOURCE_DIR}/src/extern_nccl/src)
if(WITH_DSO)
# If we use DSO, we do not build nccl, just download the dependencies
set(NCCL_BUILD_COMMAND "")
set(NCCL_INSTALL_COMMAND "")
set(NCCL_INSTALL_DIR "")
else()
# otherwise, we build nccl and link it.
set(NCCL_INSTALL_DIR ${THIRD_PARTY_PATH}/install/nccl)
# Note: cuda 8.0 is needed to make nccl
# When cuda is not installed on the system directory, need to set CUDA_HOME to your cuda root
set(NCCL_BUILD_COMMAND "make -j 8")
set(NCCL_INSTALL_COMMAND "make install PREFIX=${NCCL_INSTALL_DIR}")
endif()
ExternalProject_Add(
extern_nccl
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/NVIDIA/nccl.git"
GIT_TAG "v1.3.4-1"
PREFIX "${NCCL_SOURCE_DIR}"
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_COMMAND "${NCCL_BUILD_COMMAND}"
INSTALL_COMMAND "${NCCL_INSTALL_COMMAND}"
INSTALL_DIR "${NCCL_INSTALL_DIR}"
TEST_COMMAND ""
)
if(WITH_DSO)
if(${CMAKE_VERSION} VERSION_LESS "3.3.0")
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/lib_nccl_dummy.c)
file(WRITE ${dummyfile} "const char * dummy_nccl = \"${dummyfile}\";")
add_library(nccl STATIC ${dummyfile})
else()
add_library(nccl INTERFACE)
endif()
else()
add_library(nccl STATIC IMPORTED GLOBAL)
set_property(TARGET nccl PROPERTY IMPORTED_LOCATION
${NCCL_INSTALL_DIR}/lib/libnccl_static.a)
endif()
add_dependencies(nccl extern_nccl)
doc/fluid/design/concepts/index_cn.rst
浏览文件 @ ae5923e2
......@@ -16,3 +16,4 @@
block.md
scope.md
executor.md
parallel_executor.md
doc/fluid/design/concepts/index_en.rst
浏览文件 @ ae5923e2
......@@ -16,3 +16,4 @@ Core Concepts
block.md
scope.md
executor.md
parallel_executor.md
doc/fluid/design/muti_devices/kernel_hint_design.md
浏览文件 @ ae5923e2
# Problem
# Kernel Hint Design
## Problem
In PaddlePaddle's [Design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/switch_kernel.md), one Operator may have multiple kernels. Users may have some personal preference to choose a certain type of kernel for an operator, such as `force_cpu` to choose a CPU kernel, `use_cudnn` to choose a CUDNN kernel, we need to provide a way for users to do this.
In the current design, we use KernelType to describe one kernel.
......
doc/fluid/design/muti_devices/kernel_selection.md
浏览文件 @ ae5923e2
# Background
# Kernel Selection
## Background
Every operator has many kernels because there are multiple data types, places, data layout, library type that Fluid supports. We use the `OpKernelType ` to describe kernel types that operators can hold.
The `OpKernelType ` is as follows:
......
doc/v2/build_and_install/index_en.rst
浏览文件 @ ae5923e2
Install and Build
=================
install and Compile
==========
.. _install_steps:
Install Steps
++++++++
PaddlePaddle provides various methods of installation for many different users
You can choose either pip or Docker to complete your install:
Focus on Deep Learning Model Development
-----------------
PaddlePaddle provides lots of packages of python wheel , that pip can install:
.. toctree::
:maxdepth: 1
:maxdepth: 1
pip_install_en.rst
docker_install_en.rst
pip_install_en.rst
Build from Source
-----------------
This is the most convenient way of installation. Please choose the right installation package with machine configure and system.
Follow the Bottom Frame
----------
PaddlePaddle also supports installation using Docker. Please refer to the tutorial below:
.. toctree::
:maxdepth: 1
docker_install_en.rst
.. warning::
We recommend running PaddlePaddle in Docker. This method has the following advantages：
We recommend to directly install via above installation steps, you'll only need to build PaddlePaddle from source when you need a modifed binary.
- Does not require installation of third-party dependencies.
- Easy to share runtime environment.
.. toctree::
Lastly, users can also compile and install PaddlePaddle from source code. The instructions are below:
.. toctree::
:maxdepth: 1
build_from_source_en.md
build_from_source_en.rst
.. warning::
One caveat with this approach is that developers will have to download, compile and install all third-party dependencies. Thus this process of installation is more time consuming.
FAQ
++++++++++
-----------
For any problems during installation, please refer to the page below for answers:
:ref:`常见问题解答 <install_faq>`
If the problem still persists, you are welcome to seek assistance from the PaddlePaddle community：
`FAQ <http://www.paddlepaddle.org/docs/develop/documentation/zh/faq/build_and_install/index_en.html>`_
`创建issue <https://github.com/PaddlePaddle/Paddle/issues/new>`_
doc/v2/dev/write_docs_cn.rst
浏览文件 @ ae5923e2
......@@ -65,39 +65,55 @@ PaddlePaddle.org工具可以配合Docker使用，需要在系统里先安装好D
不使用PaddlePaddle.org工具
--------------------------
使用Docker构建PaddlePaddle的文档，需要在系统里先安装好Docker工具包。Docker安装请参考 `Docker的官网 <https://docs.docker.com/>`_ 。安装好Docker之后可以使用源码目录下的脚本构建文档，即
使用Docker构建PaddlePaddle的文档，需要在系统里先安装好Docker工具包。Docker安装请参考 `Docker的官网 <https://docs.docker.com/>`_ 。该方法与 `从源码编译PaddlePaddle <http://paddlepaddle.org/docs/develop/documentation/zh/build_and_install/build_from_source_cn.html>`_ 相似，通过从源码中构建可用于编译PaddlePaddle文档的Docker镜像并运行，在进入Docker容器后使用源码中的脚本构建PaddlePaddle文档，具体步骤如下：
[TBD]
.. code-block:: bash
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
# 从源码中构建可用于编译PaddlePaddle文档的Docker镜像
docker build -t paddle:dev .
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" -e "WITH_DOC=ON" paddle:dev /bin/bash
# 进入Docker容器后使用build.sh脚本构建PaddlePaddle文档
bash -x /paddle/paddle/scripts/docker/build.sh
注：上述命令把当前目录（源码根目录）映射为 container 里的 :code:`/paddle` 目录。
编译完成后，会产生 ``doc/v2`` 和 ``doc/fluid`` 两个目录，在这两个目录下分别都生成 ``cn/html/`` 、 ``en/html`` 、 ``api/en/html`` 共三个子目录，分别进入这些目录下，执行以下命令：
.. code-block:: bash
python -m SimpleHTTPServer 8088
在浏览器中输入 http://localhost:8088 就可以看到编译生成的 ``v2`` 和 ``fluid`` 两种版本的中/英文的文档页面和英文的API页面。
如果不想使用Docker，也可以使用以下命令直接构建PaddlePaddle文档，即
.. code-block:: bash
mkdir paddle
cd paddle
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
mkdir -p build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
# 如果只需要构建使用文档，则执行以下命令
make -j $processors gen_proto_py
make -j $processors paddle_docs paddle_docs_cn
make -j $processors paddle_docs
# 如果只需要构建API，则执行以下命令
make -j $processors gen_proto_py framework_py_proto
make -j $processors copy_paddle_pybind
make -j $processors paddle_api_docs
make -j $processors paddle_apis
其中$processors代表启动和CPU核一样多的进程来并行编译，可以根据本机的CPU核数设置相应的值。
编译完成后，进入 ``doc/v2`` 目录，如果选择构建文档则会在该目录下生成 ``cn/html/`` 、 ``en/html`` 两个子目录，选择构建API则会生成 ``api/en/html`` 目录，分别进入这些目录下，执行以下命令：
编译完成后，同样会产生 ``doc/v2`` 和 ``doc/fluid`` 两个目录，如果选择构建文档则会在这两个目录下分别都生成 ``cn/html/`` 、 ``en/html`` 两个子目录，选择构建API则会在这两个目录下分别生成 ``api/en/html`` 目录，分别进入这些子目录下，执行以下命令：
.. code-block:: bash
python -m SimpleHTTPServer 8088
在浏览器中输入 http://localhost:8088 就可以看到编译生成的中/英文的文档页面和英文的API页面,下图为生成的英文文档首页示例。注意，示例中由于使用了sphinx的原始主题，所以页面的风格与官网并不一致，但这并不影响开发者进行调试。
在浏览器中输入 http://localhost:8088 就可以看到编译生成的 ``v2`` 和 ``fluid`` 两种版本的中/英文的文档页面和英文的API页面。下图为生成的 ``v2`` 英文文档首页示例。注意，示例中由于使用了sphinx的原始主题，所以页面的风格与官网并不一致，但这并不影响开发者进行调试。
.. image:: src/doc_en.png
:align: center
......
doc/v2/dev/write_docs_en.rst
浏览文件 @ ae5923e2
......@@ -68,39 +68,56 @@ Please `click here <https://github.com/PaddlePaddle/PaddlePaddle.org/blob/develo
Manually Building the Documentation
-------------------------------------
Build PaddlePaddle's documentation with Docker，you need to install Docker first. Please refer to `Docker's official website <https://docs.docker.com/>`_ on how to install Docker. After Docker is installed, you could use the scripts in the source directory to build the documentation.
Build PaddlePaddle's documentation with Docker，you need to install Docker first. Please refer to `Docker's official website <https://docs.docker.com/>`_ on how to install Docker. This method is quite similar to ` Build From Sources <http://paddlepaddle.org/docs/develop/documentation/en/build_and_install/build_from_source_en.html>`_ , by constructing, from source code, a docker image that can be used to build PaddlePaddle documentation. Enter the Docker container and use the script ``build.sh`` in the source directory to build the PaddlePaddle documentation. The specific steps are as follows:
[TBD]
.. code-block:: bash
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
# Construct a docker image from source code
docker build -t paddle:dev .
docker run -it -v $PWD:/paddle -e "WITH_GPU=OFF" -e "WITH_TESTING=OFF" -e "WITH_DOC=ON" paddle:dev /bin/bash
# Use build.sh to build PaddlePaddle documentation
bash -x /paddle/paddle/scripts/docker/build.sh
Note: The above commands maps the current directory (source root directory) to the :code:`/paddle` directory in the container.
After compiling, there should be two generated directories: ``doc/v2`` and ``doc/fluid``, where three subdirectories ``cn/html/``, ``en/html`` and ``api/en/html`` are generated. Please enter these directories respectively and execute the following commands:
.. code-block:: bash
python -m SimpleHTTPServer 8088
Use a web browser and navigate to http://localhost:8000, you could see the compiled ``v2`` 's and ``fluid`` 's Chinese/English documents page and English APIs page.
If you do not wish to use Docker, you can also use the following commands to directly build the PaddlePaddle documentation.
.. code-block:: bash
mkdir paddle
cd paddle
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
mkdir -p build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DWITH_GPU=OFF -DWITH_MKL=OFF -DWITH_DOC=ON
# If you only need to build documents, use the following commands
make -j $processors gen_proto_py
make -j $processors paddle_docs paddle_docs_cn
make -j $processors paddle_docs
# If you only need to build APIs, use the following commands
make -j $processors gen_proto_py framework_py_proto
make -j $processors copy_paddle_pybind
make -j $processors paddle_api_docs
make -j $processors paddle_apis
$processors indicates that as many processes as the CPU cores are started to compile in parallel. It should be set according to the number of CPU cores of your machine.
After the compilation is complete, enter the ``doc/v2`` directory. If you chose to build documents, it will generate ``cn/html/`` and ``en/html`` subdirectories under this directory. If you chose to build APIs，it will generate``api/en/html`` subdirectory. Please enter these directories respectively and execute the following commands:
After compiling, there also should be two generated directories: ``doc/v2`` and ``doc/fluid`` . If you chose to build documents, two subdirectories ``cn/html/`` and ``en/html`` will be generated in both two directories. If you chose to build APIs，a subdirectory ``api/en/html`` will be generated. Please enter these directories respectively and execute the following commands:
.. code-block:: bash
python -m SimpleHTTPServer 8088
Use a web browser and navigate to http://localhost:8000, you could see the compiled Chinese/English documents page and the English APIs page. The following figure is an example of the built English documents home page. Note that due to the sphinx's original theme used in the example, the style of the page is not consistent with the official website, but this does not affect the developer's debugging.
Use a web browser and navigate to http://localhost:8000, you could see the compiled ``v2`` 's and ``fluid`` 's Chinese/English documents page and English APIs page. The following figure is an example of the built ``v2`` 's English documents home page. Note that due to the sphinx's original theme used in the example, the style of the page is not consistent with the official website, but this does not affect the developer's debugging.
.. image:: src/doc_en.png
:align: center
......
doc/v2/howto/capi/compile_paddle_lib_en.md
浏览文件 @ ae5923e2
## Install and Build
TBD
### Download & Install
Download the latest C-API development package from CI system and install. You can find the required version in the table below:
<table>
<thead>
<tr>
<th>Version Tips</th>
<th>C-API</th>
</tr>
</thead>
<tbody>
<tr>
<td>cpu_avx_mkl</td>
<td><a href="https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuAvxCp27cp27mu/.lastSuccessful/paddle.tgz" rel="nofollow">paddle.tgz</a></td>
</tr>
<tr>
<td>cpu_avx_openblas</td>
<td>-</td>
</tr>
<tr>
<td>cpu_noavx_openblas</td>
<td><a href="https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_CpuNoavxOpenblas/.lastSuccessful/paddle.tgz" rel="nofollow">paddle.tgz</a></td>
</tr>
<tr>
<td>cuda7.5_cudnn5_avx_mkl</td>
<td><a href="https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda75cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz" rel="nofollow">paddle.tgz</a></td>
</tr>
<tr>
<td>cuda8.0_cudnn5_avx_mkl</td>
<td><a href="https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda80cudnn5cp27cp27mu/.lastSuccessful/paddle.tgz" rel="nofollow">paddle.tgz</a></td>
</tr>
<tr>
<td>cuda8.0_cudnn7_avx_mkl</td>
<td><a href="https://guest:@paddleci.ngrok.io/repository/download/Manylinux1_Cuda8cudnn7cp27cp27mu/.lastSuccessful/paddle.tgz" rel="nofollow">paddle.tgz</a></td>
</tr></tbody></table>
### From source
Users can also compile the C-API library from PaddlePaddle source code by compiling with the following compilation options:
<table>
<thead>
<tr>
<th>Options</th>
<th>Value</th>
</tr>
</thead>
<tbody>
<tr>
<td>WITH_C_API</td>
<td>ON</td>
</tr>
<tr>
<td>WITH_PYTHON</td>
<td>OFF（recommended）</td>
</tr>
<tr>
<td>WITH_SWIG_PY</td>
<td>OFF（recommended）</td>
</tr>
<tr>
<td>WITH_GOLANG</td>
<td>OFF（recommended）</td>
</tr>
<tr>
<td>WITH_GPU</td>
<td>ON/OFF</td>
</tr>
<tr>
<td>WITH_MKL</td>
<td>ON/OFF</td>
</tr></tbody></table>
It is best to set up with recommended values to avoid linking with unnecessary libraries. Set other compilation options as you need.
Pull the latest following code snippet from github, and configure compilation options(replace PADDLE_ROOT with the installation path of the PaddlePaddle C-API inference library):
```shell
PADDLE_ROOT=/path/of/capi
git clone https://github.com/PaddlePaddle/Paddle.git
cd Paddle
mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX=$PADDLE_ROOT \
-DCMAKE_BUILD_TYPE=Release \
-DWITH_C_API=ON \
-DWITH_SWIG_PY=OFF \
-DWITH_GOLANG=OFF \
-DWITH_PYTHON=OFF \
-DWITH_MKL=OFF \
-DWITH_GPU=OFF \
..
```
After running the above code to generate Makefile , run: `make && make install`. After successful compilation, the dependencies required by C-API(includes: (1)PaddlePaddle inference library and header files; (2) Third-party libraries and header files) will be stored in the `PADDLE_ROOT` directory.
If the compilation is successful, see the following directory structure under `PADDLE_ROOT`(includes PaddlePaddle header files and libraries, and third-party libraries and header files(determined by the link methods if necessary)):
```text
├── include
│   └── paddle
│   ├── arguments.h
│   ├── capi.h
│   ├── capi_private.h
│   ├── config.h
│   ├── error.h
│   ├── gradient_machine.h
│   ├── main.h
│   ├── matrix.h
│   ├── paddle_capi.map
│   └── vector.h
├── lib
│   ├── libpaddle_capi_engine.a
│   ├── libpaddle_capi_layers.a
│   ├── libpaddle_capi_shared.so
│   └── libpaddle_capi_whole.a
└── third_party
├── gflags
│   ├── include
│   │   └── gflags
│   │   ├── gflags_completions.h
│   │   ├── gflags_declare.h
│   │   ...
│   └── lib
│   └── libgflags.a
├── glog
│   ├── include
│   │   └── glog
│   │   ├── config.h
│   │   ...
│   └── lib
│   └── libglog.a
├── openblas
│   ├── include
│   │   ├── cblas.h
│   │   ...
│   └── lib
│   ...
├── protobuf
│   ├── include
│   │   └── google
│   │   └── protobuf
│   │   ...
│   └── lib
│   └── libprotobuf-lite.a
└── zlib
├── include
│   ...
└── lib
...
```
### Linking Description:
There are three kinds of linking methods:
1. Linking with dynamic library `libpaddle_capi_shared.so`（This way is much more convenient and easier, **Without special requirements, it is recommended**）, refer to the following：
1. Compiling with CPU version and using `OpenBLAS`; only need to link one library named `libpaddle_capi_shared.so` to develop prediction program through C-API.
1. Compiling with CPU version and using `MKL` lib, you need to link MKL library directly to develop prediction program through PaddlePaddle C-API, due to `MKL` has its own dynamic library.
1. Compiling with GPU version, CUDA library will be loaded dynamically on prediction program run-time, and also set CUDA library to  `LD_LIBRARY_PATH` environment variable.
2. Linking with static library `libpaddle_capi_whole.a`，refer to the following：
1. Specify `-Wl,--whole-archive` linking options.
1. Explicitly link third-party libraries such as `gflags`、`glog`、`libz`、`protobuf` .etc, you can find them under `PADDLE_ROOT/third_party` directory.
1. Use OpenBLAS library if compiling C-API，must explicitly link `libopenblas.a`.
1. Use MKL when compiling C-API, must explicitly link MKL dynamic library.
3. Linking with static library `libpaddle_capi_layers.a` and `libpaddle_capi_engine.a`，refer to the following：
1. This linking methods is mainly used for mobile prediction.
1. Split `libpaddle_capi_whole.a` into two static linking library at least to reduce the size of linking libraries.
1. Specify `-Wl,--whole-archive -lpaddle_capi_layers`  and `-Wl,--no-whole-archive -lpaddle_capi_engine` for linking.
1. The third-party dependencies need explicitly link same as method 2 above.
doc/v2/howto/cluster/multi_cluster/k8s_distributed_en.md
浏览文件 @ ae5923e2
# Kubernetes Distributed
# Distributed Training on Kubernetes
TBD
We introduced how to create a PaddlePaddle Job with a single node on Kuberentes in the
previous document.
In this article, we will introduce how to create a PaddlePaddle job with multiple nodes
on Kubernetes cluster.
## Overall Architecture
Before creating a training job, the users need to slice the training data and deploy
the Python scripts along with it into the distributed file system
(We can use the different type of Kuberentes Volumes to mount different distributed
file systems). Before training starts, The program will copy the training data into the
Container and also save the models at the same path during training. The global architecture
is as follows:
![PaddlePaddle on Kubernetes Architecture](src/k8s-paddle-arch.png)
The above figure describes a distributed training architecture which contains 3 nodes, each
Pod mounts a folder of the distributed file system to save training data and models
by Kubernetes Volume. Kubernetes created 3 Pods for this training phase and scheduled these on
3 nodes, each Pod has a PaddlePaddle container. After the containers car created,
PaddlePaddle starts up the communication between PServer and Trainer and read training
data for this training job.
As the description above, we can start up a PaddlePaddle distributed training job on a
Kubernetes ready cluster with the following steps:
1. [Build PaddlePaddle Docker Image](#Build a Docker Image)
1. [Split training data and upload to the distributed file system](#Upload Training Data)
1. [Edit a YAML file and create a Kubernetes Job](#Create a Job)
1. [Check the output](#Check The Output)
We will introduce these steps as follows:
### Build a Docker Image
Training docker image needs to package the paddle pserver and paddle trainer runtimes, as well as two more processes before we can kick off the training:
- Copying the training data into container.
- Generating the initialization arguments for `Paddle PServer` and `Paddle Training` processes.
Since the paddlepaddle official docker image already has the runtimes we need, we'll take it as the base image and pack some additional scripts for the processes mentioned above to build our training image. for more detail, please find from the following link:
- https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/src/k8s_train/Dockerfile
```bash
$ cd doc/howto/usage/k8s/src/k8s_train
$ docker build -t [YOUR_REPO]/paddle:mypaddle .
```
And then upload the new Docker Image to a Docker hub:
```bash
docker push [YOUR_REPO]/paddle:mypaddle
```
**[NOTE]**, in the above command arguments, `[YOUR_REPO]` represents your Docker repository,
you need to use your repository instead of it. We will replace it with your respository name to
represent the Docker Image which built in this step.
### Prepare Training Data
We can download and split the training job by creating a Kubernetes Job, or custom your image
by editing [k8s_train](./src/k8s_train/).
Before creating a Job, we need to bind a [persistenVolumeClaim](https://kubernetes.io/docs/user-guide/persistent-volumes) by the different type of
the different file system, the generated dataset would be saved on this volume.
```yaml
apiVersion: batch/v1
kind: Job
metadata:
name: paddle-data
spec:
template:
metadata:
name: pi
spec:
hostNetwork: true
containers:
- name: paddle-data
image: paddlepaddle/paddle-tutorial:k8s_data
imagePullPolicy: Always
volumeMounts:
- mountPath: "/mnt"
name: nfs
env:
- name: OUT_DIR
value: /home/work/mfs/paddle-cluster-job
- name: SPLIT_COUNT
value: "3"
volumes:
- name: nfs
persistentVolumeClaim:
claimName: mfs
restartPolicy: Never
```
Create the Job with the following command:
```bash
> kubectl create -f xxx.yaml
```
If created successfully, you can see some information like this:
```base
[root@paddle-kubernetes-node0 nfsdir]$ tree -d
.
`-- paddle-cluster-job
|-- 0
| `-- data
|-- 1
| `-- data
|-- 2
| `-- data
|-- output
|-- quick_start
```
The `paddle-cluster-job` above is the job name for this training job; we need 3
PaddlePaddle training nodes and save the split training data in `paddle-cluster-job` path,
the folder `0`, `1` and `2` represents the `training_id` on each node, `quick_start` folder is used to store training data, `output` folder is used to store the models and logs.
### Create a Job
Kubernetes allow users to create objects with YAML files, and we can use a command-line tool
to create it.
The Job YAML file describes that which Docker Image would be used in this training job, how much nodes would be created, what's the startup arguments of `Paddle PServer/Trainer` process and what's the type of Volumes. You can find the details of the YAML filed in
[Kubernetes Job API](http://kubernetes.io/docs/api-reference/batch/v1/definitions/#_v1_job).
The following is an example for this training job:
```yaml
apiVersion: batch/v1
kind: Job
metadata:
name: paddle-cluster-job
spec:
parallelism: 3
completions: 3
template:
metadata:
name: paddle-cluster-job
spec:
volumes:
- name: jobpath
hostPath:
path: /home/work/mfs
containers:
- name: trainer
image: [YOUR_REPO]/paddle:mypaddle
command: ["bin/bash", "-c", "/root/start.sh"]
env:
- name: JOB_NAME
value: paddle-cluster-job
- name: JOB_PATH
value: /home/jobpath
- name: JOB_NAMESPACE
value: default
- name: TRAIN_CONFIG_DIR
value: recommendation
- name: CONF_PADDLE_NIC
value: eth0
- name: CONF_PADDLE_PORT
value: "7164"
- name: CONF_PADDLE_PORTS_NUM
value: "2"
- name: CONF_PADDLE_PORTS_NUM_SPARSE
value: "2"
- name: CONF_PADDLE_GRADIENT_NUM
value: "3"
volumeMounts:
- name: jobpath
mountPath: /home/jobpath
restartPolicy: Never
```
In the above YAML file:
- `metadata.name`, The job name.
- `parallelism`, Whether the Kubernetes Job would create `parallelism` Pods at the same time.
- `completions`, The Job would become the success status only when the number of successful Pod(the exit code is 0)
is equal to `completions`.
- `volumeMounts`, the name field `jobpath` is a key, the `mountPath` field represents
the path in the container, and we can define the `jobpath` in `volumes` filed, use `hostPath`
to configure the host path we want to mount.
- `env`, the environment variables in the Container, we pass some startup arguments by
this approach, some details are as following:
- JOB_PATH：the mount path in the container
- JOB_NAME：the job name
- TRAIN_CONFIG_DIR：the job path in the container, we can find the training data path by
combine with JOB_NAME.
- CONF_PADDLE_NIC: the argument `--nics` of `Paddle PServer` process, the network
device name.
- CONF_PADDLE_PORT: the argument `--port` of `Paddle PServer` process.
- CONF_PADDLE_PORTS_NUM: the argument `--ports_num` of `Paddle PServer`, the port number
for dense prameter update.
- CONF_PADDLE_PORTS_NUM_SPARSE：the argument `--ports_num_for_sparse` of `Paddle PServer`,
the port number for sparse parameter update.
- CONF_PADDLE_GRADIENT_NUM：the number of training node, the argument
`--num_gradient_servers` of `Paddle PServer` and `Paddle Trainer`.
You can find some details information at [here]
(http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cmd_parameter/detail_introduction_cn.html)。
We can use the command-line tool of Kubernetes to create a Job when we finish the YAML file:
```bash
kubectl create -f job.yaml
```
Upon successful creation, Kubernetes would create 3 Pods as PaddlePaddle training node,
pull the Docker image and begin to train.
### Checkout the Output
At the process of training, we can check the logs and the output models which is stored in
the `output` folder.
**NOTE**, `node_0`, `node_1` and `node_2` represent the
`trainer_id` of the PaddlePaddle training job rather than the node id of Kubernetes.
```bash
[root@paddle-kubernetes-node0 output]# tree -d
.
├── node_0
│   ├── server.log
│   └── train.log
├── node_1
│   ├── server.log
│   └── train.log
├── node_2
......
├── pass-00002
│   ├── done
│   ├── ___embedding_0__.w0
│   ├── ___embedding_1__.w0
......
```
We can checkout the status of each training Pod by viewing the logs:
```bash
[root@paddle-kubernetes-node0 node_0]# cat train.log
I1116 09:10:17.123121 50 Util.cpp:155] commandline:
/usr/local/bin/../opt/paddle/bin/paddle_trainer
--nics=eth0 --port=7164
--ports_num=2 --comment=paddle_process_by_paddle
--pservers=192.168.129.66,192.168.223.143,192.168.129.71
--ports_num_for_sparse=2 --config=./trainer_config.py
--trainer_count=4 --num_passes=10 --use_gpu=0
--log_period=50 --dot_period=10 --saving_period=1
--local=0 --trainer_id=0
--save_dir=/home/jobpath/paddle-cluster-job/output
I1116 09:10:17.123440 50 Util.cpp:130] Calling runInitFunctions
I1116 09:10:17.123764 50 Util.cpp:143] Call runInitFunctions done.
[WARNING 2016-11-16 09:10:17,227 default_decorators.py:40] please use keyword arguments in paddle config.
[INFO 2016-11-16 09:10:17,239 networks.py:1282] The input order is [movie_id, title, genres, user_id, gender, age, occupation, rating]
[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__square_error_cost_0__]
I1116 09:10:17.392917 50 Trainer.cpp:170] trainer mode: Normal
I1116 09:10:17.613910 50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process
I1116 09:10:17.680917 50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process
I1116 09:10:17.681543 50 GradientMachine.cpp:134] Initing parameters..
I1116 09:10:18.012390 50 GradientMachine.cpp:141] Init parameters done.
I1116 09:10:18.018641 50 ParameterClient2.cpp:122] pserver 0 192.168.129.66:7164
I1116 09:10:18.018950 50 ParameterClient2.cpp:122] pserver 1 192.168.129.66:7165
I1116 09:10:18.019069 50 ParameterClient2.cpp:122] pserver 2 192.168.223.143:7164
I1116 09:10:18.019492 50 ParameterClient2.cpp:122] pserver 3 192.168.223.143:7165
I1116 09:10:18.019716 50 ParameterClient2.cpp:122] pserver 4 192.168.129.71:7164
I1116 09:10:18.019836 50 ParameterClient2.cpp:122] pserver 5 192.168.129.71:7165
```
## Some Additional Details
### Using Environment Variables
Usually we use the environment varialbes to configurate the PaddlePaddle Job which runs in
Kubernetes, `start_paddle.py` provides a start up script to convert the environment variable
to the start up arguments of PaddlePaddle process:
```bash
API = "/api/v1/namespaces/"
JOBSELECTOR = "labelSelector=job-name="
JOB_PATH = os.getenv("JOB_PATH") + "/" + os.getenv("JOB_NAME")
JOB_PATH_OUTPUT = JOB_PATH + "/output"
JOBNAME = os.getenv("JOB_NAME")
NAMESPACE = os.getenv("JOB_NAMESPACE")
PADDLE_NIC = os.getenv("CONF_PADDLE_NIC")
PADDLE_PORT = os.getenv("CONF_PADDLE_PORT")
PADDLE_PORTS_NUM = os.getenv("CONF_PADDLE_PORTS_NUM")
PADDLE_PORTS_NUM_SPARSE = os.getenv("CONF_PADDLE_PORTS_NUM_SPARSE")
PADDLE_SERVER_NUM = os.getenv("CONF_PADDLE_GRADIENT_NUM")
```
### Communication between Pods
At the begin of `start_paddle.py`, it would initializes and parses the arguments.
```python
parser = argparse.ArgumentParser(prog="start_paddle.py",
description='simple tool for k8s')
args, train_args_list = parser.parse_known_args()
train_args = refine_unknown_args(train_args_list)
train_args_dict = dict(zip(train_args[:-1:2], train_args[1::2]))
podlist = getPodList()
```
And then query the status of all the other Pods of this Job by the function `getPodList()`, and fetch `triner_id` by the function `getIdMap(podlist)` if all the Pods status is `RUNNING`.
```python
podlist = getPodList()
# need to wait until all pods are running
while not isPodAllRunning(podlist):
time.sleep(10)
podlist = getPodList()
idMap = getIdMap(podlist)
```
**NOTE**: `getPodList()` would prefetch all the Pods in the current namespace, if some
Pods are alreay running, it may cause some error. We will use [statfulesets](https://kubernetes.io/docs/concepts/abstractions/controllers/statefulsets) instead of
Kubernetes Pod or Replicaset in the future.
The function `getIdMap(podlist)` fetches IPs addresses of `podlist` and then sort them
to generate `trainer_id`.
```python
def getIdMap(podlist):
'''
generate tainer_id by ip
'''
ips = []
for pod in podlist["items"]:
ips.append(pod["status"]["podIP"])
ips.sort()
idMap = {}
for i in range(len(ips)):
idMap[ips[i]] = i
return idMap
```
After getting the `idMap`, we can generate the arguments of `Paddle PServer` and `Paddle Trainer`
so that we can start up them by `startPaddle(idMap, train_args_dict)`.
### Create Job
The main goal of `startPaddle` is generating the arguments of `Paddle PServer` and
`Paddle Trainer` processes. Take `Paddle Trainer` as an example, we parse the
environment variable and then get `PADDLE_NIC`, `PADDLE_PORT`, `PADDLE_PORTS_NUM` and etc...,
finally find `trainerId` from `idMap` according to its IP address.
```python
program = 'paddle train'
args = " --nics=" + PADDLE_NIC
args += " --port=" + str(PADDLE_PORT)
args += " --ports_num=" + str(PADDLE_PORTS_NUM)
args += " --comment=" + "paddle_process_by_paddle"
ip_string = ""
for ip in idMap.keys():
ip_string += (ip + ",")
ip_string = ip_string.rstrip(",")
args += " --pservers=" + ip_string
args_ext = ""
for key, value in train_args_dict.items():
args_ext += (' --' + key + '=' + value)
localIP = socket.gethostbyname(socket.gethostname())
trainerId = idMap[localIP]
args += " " + args_ext + " --trainer_id=" + \
str(trainerId) + " --save_dir=" + JOB_PATH_OUTPUT
```
paddle/fluid/framework/block_desc.cc
浏览文件 @ ae5923e2
......@@ -13,11 +13,10 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/block_desc.h"
#include <queue>
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/program_desc.h"
#include <queue>
namespace paddle {
namespace framework {
......@@ -147,52 +146,7 @@ void BlockDesc::RemoveOp(size_t s, size_t e) {
if (ops_.begin() + s == ops_.end() || ops_.begin() + e == ops_.end()) {
return;
}
auto get_vars = [](std::deque<std::unique_ptr<OpDesc>>::iterator &op,
std::vector<std::string> &v) {
auto in_names = (*op)->InputArgumentNames();
v.insert(v.end(), in_names.begin(), in_names.end());
auto out_names = (*op)->OutputArgumentNames();
v.insert(v.end(), out_names.begin(), out_names.end());
std::sort(v.begin(), v.end());
auto last = std::unique(v.begin(), v.end());
v.erase(last, v.end());
};
need_update_ = true;
for (size_t i = s; i < e; i++) {
// since remove op one by one, every time remove the first op.
auto op = ops_.begin() + s;
// collect input and output variables from current delete op
std::vector<std::string> cur_vars;
get_vars(op, cur_vars);
// remove current op
ops_.erase(ops_.begin() + s);
// collect input and output variables from other ops
std::vector<std::string> other_vars;
for (auto it = ops_.begin(); it != ops_.end(); it++) {
get_vars(it, other_vars);
}
// variables should be deleted
std::vector<std::string> delete_vars;
// delete_vars = cur_vars - cur_vars ^ other_input_vars
std::set_difference(cur_vars.begin(), cur_vars.end(), other_vars.begin(),
other_vars.end(),
std::inserter(delete_vars, delete_vars.end()));
// remove variables
for (size_t i = 0; i < delete_vars.size(); i++) {
auto name = delete_vars[i];
auto it = vars_.find(name);
PADDLE_ENFORCE(it != vars_.end(),
"%s is not in variable list, it should not be deleted",
name);
vars_.erase(it);
VLOG(3) << "deleting variable " << name;
}
}
ops_.erase(ops_.begin() + s, ops_.begin() + e);
}
std::vector<OpDesc *> BlockDesc::AllOps() const {
......
paddle/fluid/framework/data_device_transform_test.cu
浏览文件 @ ae5923e2
......@@ -105,7 +105,7 @@ static void BuildVar(const std::string& param_name,
TEST(Operator, CPUtoGPU) {
using namespace paddle::framework;
using namespace paddle::platform;
InitDevices();
InitDevices(true);
paddle::framework::Scope scope;
paddle::platform::CPUPlace cpu_place;
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ ae5923e2
......@@ -2,21 +2,25 @@ cc_library(var_handle SRCS var_handle.cc DEPS place)
cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context)
cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
if(WITH_GPU)
nv_library(nccl_all_reduce_op_handle SRCS nccl_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
nv_library(nccl_all_reduce_op_handle SRCS nccl_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda)
cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(send_op_handle SRCS send_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base)
cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph)
if(WITH_GPU)
set(multi_devices_graph_builder_deps nccl_all_reduce_op_handle)
else()
set(multi_devices_graph_builder_deps)
endif()
cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base)
cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph)
cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
scale_loss_grad_op_handle ${multi_devices_graph_builder_deps})
scale_loss_grad_op_handle send_op_handle ${multi_devices_graph_builder_deps})
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph framework_proto)
cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
simple_threadpool device_context)
cc_library(broadcast_op_handle SRCS broadcast_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
......
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @ ae5923e2
......@@ -15,6 +15,7 @@
#include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
#include "paddle/fluid/framework/details/send_op_handle.h"
#include "paddle/fluid/framework/scope.h"
#ifdef PADDLE_WITH_CUDA
......@@ -54,12 +55,37 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
}
}
void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result, OpDesc *op,
const platform::Place &p,
const size_t &i) const {
auto *op_handle = result->ops_.back().get();
op_handle->dev_ctxes_[p] = const_cast<platform::DeviceContext *>(
platform::DeviceContextPool::Instance().Get(p));
auto var_names = op->InputArgumentNames();
for (auto &each_var_name : var_names) {
VarHandle *var = CreateOrGetLatestVarHandle(result, each_var_name, p, i);
op_handle->AddInput(var);
}
var_names = op->OutputArgumentNames();
for (auto &each_var_name : var_names) {
CreateOpOutput(result, op_handle, each_var_name, p, i);
}
}
std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
const ProgramDesc &program) const {
auto graph = new SSAGraph();
SSAGraph &result = *graph;
std::unordered_set<std::string> og_has_been_broadcast;
result.vars_.resize(places_.size());
// We cannot invoke resize. It is a bug of GCC 4.8
result.vars_ = std::vector<
std::unordered_map<std::string, std::vector<std::unique_ptr<VarHandle>>>>(
places_.size());
bool is_forwarding = true;
for (auto *op : program.Block(0).AllOps()) {
......@@ -72,27 +98,28 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
}
}
// append send op if program is distributed trainer main program.
// always use the first device
if (!is_forwarding && op->Type() == "send") {
auto &p = places_[0];
auto *s = local_scopes_[0];
// FIXME(wuyi): send op always copy from GPU 0
result.ops_.emplace_back(new SendOpHandle(*op, s, p));
// Create inputs for output on original place and no ssa output
// is created for send op.
CreateOpHandleIOs(&result, op, p, 0);
continue;
}
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
auto *s = local_scopes_[i];
result.ops_.emplace_back(new ComputationOpHandle(*op, s, p));
auto *op_handle = result.ops_.back().get();
op_handle->dev_ctxes_[p] = const_cast<platform::DeviceContext *>(
platform::DeviceContextPool::Instance().Get(p));
auto var_names = op->InputArgumentNames();
CreateOpHandleIOs(&result, op, p, i);
for (auto &each_var_name : var_names) {
VarHandle *var =
CreateOrGetLatestVarHandle(&result, each_var_name, p, i);
op_handle->AddInput(var);
}
var_names = op->OutputArgumentNames();
for (auto &each_var_name : var_names) {
CreateOpOutput(&result, op_handle, each_var_name, p, i);
}
auto var_names = op->OutputArgumentNames();
if (is_forwarding) {
if (var_names.size() == 1 && var_names[0] == loss_var_name_) {
......@@ -147,15 +174,16 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
if (vars.empty()) { // This device has no data. continue.
continue;
}
auto *prev_grad = &vars[vars.size() - 1];
op_handle->AddInput(prev_grad);
auto &prev_grad = vars[vars.size() - 1];
op_handle->AddInput(prev_grad.get());
auto &var = vars[vars.size()];
var.place_ = p;
var.name_ = og;
var.version_ = vars.size() - 1;
vars.emplace_back(new VarHandle);
auto &var = vars.back();
var->place_ = p;
var->name_ = og;
var->version_ = vars.size() - 1;
op_handle->AddOutput(&var);
op_handle->AddOutput(var.get());
}
#else
PADDLE_ENFORCE("Not implemented");
......
paddle/fluid/framework/details/multi_devices_graph_builder.h
浏览文件 @ ae5923e2
......@@ -14,6 +14,9 @@
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/ssa_graph_builder.h"
namespace paddle {
......@@ -41,6 +44,10 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
private:
void CreateOpHandleIOs(SSAGraph *result, OpDesc *op, const platform::Place &p,
const size_t &i) const;
private:
std::string loss_var_name_;
const std::vector<platform::Place> &places_;
......
paddle/fluid/framework/details/send_op_handle.cc 0 → 100644
浏览文件 @ ae5923e2
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/send_op_handle.h"
namespace paddle {
namespace framework {
namespace details {
SendOpHandle::SendOpHandle(const framework::OpDesc &op_desc,
const Scope *local_scope,
const platform::Place &place)
: op_(framework::OpRegistry::CreateOp(op_desc)),
local_scope_(local_scope),
place_(place) {}
void SendOpHandle::RunImpl() {
// Wait input done
for (auto *in : inputs_) {
auto &p = static_cast<VarHandle *>(in)->place_;
if (in->DebugString() == "dummy") { // HACK
continue;
}
in->generated_op_->Wait(dev_ctxes_[p]);
}
op_->Run(*local_scope_, place_);
}
std::string SendOpHandle::Name() const { return "send"; }
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/send_op_handle.h 0 → 100644
浏览文件 @ ae5923e2
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/scope.h"
namespace paddle {
namespace framework {
namespace details {
struct SendOpHandle : public OpHandleBase {
std::unique_ptr<OperatorBase> op_;
const Scope* local_scope_;
const platform::Place& place_;
SendOpHandle(const framework::OpDesc& op_desc, const Scope* local_scope,
const platform::Place& place);
std::string Name() const override;
// Delay and buffer nccl_all_reduce together can significantly increase
// performance. Disable this feature by returning false.
bool IsMultiDeviceTransfer() override { return false; };
protected:
void RunImpl() override;
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/ssa_graph.h
浏览文件 @ ae5923e2
......@@ -16,6 +16,8 @@
#include <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/details/var_handle.h"
......@@ -24,7 +26,9 @@ namespace framework {
namespace details {
struct SSAGraph {
std::vector<std::unordered_map<std::string, std::map<int, VarHandle>>> vars_;
std::vector<
std::unordered_map<std::string, std::vector<std::unique_ptr<VarHandle>>>>
vars_;
// aux variables to represent dependency. Useful to resolve data hazard.
std::unordered_set<std::unique_ptr<VarHandleBase>> dep_vars_;
std::vector<std::unique_ptr<OpHandleBase>> ops_;
......
paddle/fluid/framework/details/ssa_graph_builder.cc
浏览文件 @ ae5923e2
......@@ -27,8 +27,8 @@ void SSAGraphBuilder::PolishGraphToSupportDataHazards(SSAGraph *graph) {
auto it_old = name_pair.second.rbegin();
++it_old;
for (; it_old != name_pair.second.rend(); it_new = it_old, ++it_old) {
auto *write_op = it_new->second.generated_op_;
auto &read_ops = it_old->second.pending_ops_;
auto *write_op = (*it_new)->generated_op_;
auto &read_ops = (*it_old)->pending_ops_;
for (auto *read_op : read_ops) {
// Manually add a dependency var from read_op to write_op;
......@@ -54,14 +54,15 @@ VarHandle *SSAGraphBuilder::CreateOrGetLatestVarHandle(
auto &var_holder = var_holders[each_var_name];
VarHandle *var = nullptr;
if (var_holder.empty()) {
var_holder.emplace_back(new VarHandle);
auto &init_var = var_holder[0];
init_var.place_ = place;
init_var.name_ = each_var_name;
init_var.generated_op_ = nullptr;
init_var.version_ = 0;
var = &init_var;
init_var->place_ = place;
init_var->name_ = each_var_name;
init_var->generated_op_ = nullptr;
init_var->version_ = 0;
var = init_var.get();
} else {
var = &var_holder.rbegin()->second;
var = var_holder.rbegin()->get();
}
return var;
}
......@@ -72,11 +73,12 @@ void SSAGraphBuilder::CreateOpOutput(SSAGraph *graph, OpHandleBase *op_handle,
size_t place_offset) {
auto &vars = graph->vars_[place_offset][each_var_name];
size_t version = vars.size();
auto &var = vars[version];
var.version_ = version;
var.name_ = each_var_name;
var.place_ = place;
op_handle->AddOutput(&var);
vars.emplace_back(new VarHandle());
auto &var = vars.back();
var->version_ = version;
var->name_ = each_var_name;
var->place_ = place;
op_handle->AddOutput(var.get());
}
template <typename Callback>
......@@ -84,7 +86,7 @@ void IterAllVar(const SSAGraph &graph, Callback callback) {
for (auto &each : graph.vars_) {
for (auto &pair1 : each) {
for (auto &pair2 : pair1.second) {
callback(pair2.second);
callback(*pair2);
}
}
}
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @ ae5923e2
......@@ -69,7 +69,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
for (auto &var_map : graph_->vars_) {
for (auto &name_pair : var_map) {
for (auto &version_pair : name_pair.second) {
InsertPendingVar(version_pair.second);
InsertPendingVar(*version_pair);
}
}
}
......@@ -95,7 +95,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
for (auto &var_map : graph_->vars_) {
auto it = var_map.find(fetch_var_name);
if (it != var_map.end()) {
fetched_vars[fetch_var_name].push_back(&it->second.rbegin()->second);
fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get());
}
}
}
......
paddle/fluid/framework/init.cc
浏览文件 @ ae5923e2
......@@ -64,7 +64,7 @@ void InitP2P(int count) {
#endif
}
void InitDevices() {
void InitDevices(bool init_p2p) {
/*Init all avaiable devices by default */
std::vector<platform::Place> places;
......@@ -85,7 +85,9 @@ void InitDevices() {
for (int i = 0; i < count; ++i) {
places.emplace_back(platform::CUDAPlace(i));
}
InitP2P(count);
if (init_p2p) {
InitP2P(count);
}
platform::DeviceContextPool::Init(places);
}
......
paddle/fluid/framework/init.h
浏览文件 @ ae5923e2
......@@ -24,7 +24,7 @@ void InitGflags(std::vector<std::string> &argv);
void InitGLOG(const std::string &prog_name);
void InitDevices();
void InitDevices(bool init_p2p);
} // namespace framework
} // namespace paddle
paddle/fluid/framework/init_test.cc
浏览文件 @ ae5923e2
......@@ -21,7 +21,7 @@ TEST(InitDevices, CPU) {
using paddle::platform::DeviceContextPool;
#ifndef PADDLE_WITH_CUDA
InitDevices();
InitDevices(true);
DeviceContextPool& pool = DeviceContextPool::Instance();
ASSERT_EQ(pool.size(), 1U);
#endif
......@@ -33,7 +33,7 @@ TEST(InitDevices, CUDA) {
#ifdef PADDLE_WITH_CUDA
int count = paddle::platform::GetCUDADeviceCount();
InitDevices();
InitDevices(true);
DeviceContextPool& pool = DeviceContextPool::Instance();
ASSERT_EQ(pool.size(), 1U + static_cast<unsigned>(count));
#endif
......
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ ae5923e2
......@@ -12,9 +12,14 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/lod_tensor.h"
#include <stdint.h>
#include <string.h>
#include <algorithm>
#include <iterator>
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/memory/memory.h"
......@@ -22,11 +27,6 @@ limitations under the License. */
#include "paddle/fluid/recordio/scanner.h"
#include "paddle/fluid/recordio/writer.h"
#include <stdint.h>
#include <string.h>
#include <algorithm>
#include <iterator>
namespace paddle {
namespace framework {
......@@ -294,7 +294,7 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
TensorFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
}
void WriteToRecordIO(recordio::Writer &writer,
void WriteToRecordIO(recordio::Writer *writer,
const std::vector<LoDTensor> &tensor,
const platform::DeviceContext &dev_ctx) {
std::stringstream buffer;
......@@ -303,18 +303,20 @@ void WriteToRecordIO(recordio::Writer &writer,
for (auto &each : tensor) {
SerializeToStream(buffer, each, dev_ctx);
}
writer.Write(buffer.str());
writer->Write(buffer.str());
}
std::vector<LoDTensor> ReadFromRecordIO(
recordio::Scanner &scanner, const platform::DeviceContext &dev_ctx) {
std::istringstream sin(scanner.Next());
uint32_t sz;
sin.read(reinterpret_cast<char *>(&sz), sizeof(uint32_t));
recordio::Scanner *scanner, const platform::DeviceContext &dev_ctx) {
std::vector<LoDTensor> result;
result.resize(sz);
for (uint32_t i = 0; i < sz; ++i) {
DeserializeFromStream(sin, &result[i], dev_ctx);
if (scanner->HasNext()) {
std::istringstream sin(scanner->Next());
uint32_t sz;
sin.read(reinterpret_cast<char *>(&sz), sizeof(uint32_t));
result.resize(sz);
for (uint32_t i = 0; i < sz; ++i) {
DeserializeFromStream(sin, &result[i], dev_ctx);
}
}
return result;
}
......
paddle/fluid/framework/lod_tensor.h
浏览文件 @ ae5923e2
......@@ -15,6 +15,9 @@ limitations under the License. */
#pragma once
#include <memory>
#include <string>
#include <utility>
#include <vector>
#ifdef PADDLE_WITH_CUDA
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
......@@ -216,12 +219,12 @@ void SerializeToStream(std::ostream& os, const LoDTensor& tensor,
void DeserializeFromStream(std::istream& is, LoDTensor* tensor,
const platform::DeviceContext& dev_ctx);
extern void WriteToRecordIO(recordio::Writer& writer,
extern void WriteToRecordIO(recordio::Writer* writer,
const std::vector<LoDTensor>& tensor,
const platform::DeviceContext& dev_ctx);
extern std::vector<LoDTensor> ReadFromRecordIO(
recordio::Scanner& scanner, const platform::DeviceContext& dev_ctx);
recordio::Scanner* scanner, const platform::DeviceContext& dev_ctx);
} // namespace framework
} // namespace paddle
paddle/fluid/framework/lod_tensor_test.cc
浏览文件 @ ae5923e2
......@@ -12,17 +12,17 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/recordio/scanner.h"
#include "paddle/fluid/recordio/writer.h"
#include <glog/logging.h>
#include <gtest/gtest.h>
#include <algorithm>
#include <memory>
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/recordio/scanner.h"
#include "paddle/fluid/recordio/writer.h"
namespace paddle {
namespace framework {
......@@ -240,8 +240,8 @@ TEST(LoDTensor, RecordIO) {
*platform::DeviceContextPool::Instance().Get(platform::CPUPlace());
{
recordio::Writer writer(stream, recordio::Compressor::kSnappy);
WriteToRecordIO(writer, {tensor, tensor}, ctx);
WriteToRecordIO(writer, {tensor, tensor}, ctx);
WriteToRecordIO(&writer, {tensor, tensor}, ctx);
WriteToRecordIO(&writer, {tensor, tensor}, ctx);
writer.Flush();
}
......@@ -254,11 +254,11 @@ TEST(LoDTensor, RecordIO) {
{
std::unique_ptr<std::istream> stream_ptr(stream);
recordio::Scanner scanner(std::move(stream_ptr));
auto tensors = ReadFromRecordIO(scanner, ctx);
auto tensors = ReadFromRecordIO(&scanner, ctx);
ASSERT_EQ(tensors.size(), 2);
assert_tensor_ok(tensors[0]);
assert_tensor_ok(tensors[1]);
tensors = ReadFromRecordIO(scanner, ctx);
tensors = ReadFromRecordIO(&scanner, ctx);
ASSERT_EQ(tensors.size(), 2);
assert_tensor_ok(tensors[0]);
assert_tensor_ok(tensors[1]);
......
paddle/fluid/framework/lod_tensor_test.cu
浏览文件 @ ae5923e2
......@@ -30,7 +30,7 @@ __global__ void test(size_t* a, int size) {
}
TEST(LoD, data) {
paddle::framework::InitDevices();
paddle::framework::InitDevices(true);
paddle::framework::LoD lod{{0, 1, 2}};
lod.push_back({0, 2, 4, 5});
......@@ -46,7 +46,7 @@ TEST(LoD, data) {
}
TEST(LoDTensor, LoDInGPU) {
paddle::framework::InitDevices();
paddle::framework::InitDevices(true);
paddle::framework::LoDTensor lod_tensor;
paddle::platform::CUDAPlace place(0);
......
paddle/fluid/framework/operator_test.cc
浏览文件 @ ae5923e2
......@@ -72,7 +72,7 @@ REGISTER_OP_WITHOUT_GRADIENT(test_operator,
paddle::framework::OpWithoutKernelCheckerMaker);
TEST(OperatorBase, all) {
paddle::framework::InitDevices();
paddle::framework::InitDevices(true);
paddle::framework::proto::OpDesc op_desc;
op_desc.set_type("test_operator");
BuildVar("input", {"IN1"}, op_desc.add_inputs());
......@@ -198,7 +198,7 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel,
// test with single input
TEST(OpKernel, all) {
paddle::framework::InitDevices();
paddle::framework::InitDevices(true);
paddle::framework::proto::OpDesc op_desc;
op_desc.set_type("op_with_kernel");
BuildVar("x", {"IN1"}, op_desc.add_inputs());
......@@ -228,7 +228,7 @@ REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
TEST(OpKernel, multi_inputs) {
using namespace paddle::framework;
paddle::framework::InitDevices();
paddle::framework::InitDevices(true);
proto::OpDesc op_desc;
op_desc.set_type("op_multi_inputs_with_kernel");
......@@ -269,7 +269,7 @@ class OperatorClone : public paddle::framework::OperatorBase {
};
TEST(Operator, Clone) {
paddle::framework::InitDevices();
paddle::framework::InitDevices(true);
OperatorClone a("ABC", paddle::framework::VariableNameMap{},
paddle::framework::VariableNameMap{},
paddle::framework::AttributeMap{});
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ ae5923e2
......@@ -115,14 +115,12 @@ void ParallelExecutor::BCastParamsToGPUs(
for (auto &var : vars) {
auto *main_var = main_scope->FindVar(var);
if (!main_var->IsType<LoDTensor>()) {
if (main_var == nullptr || !main_var->IsType<LoDTensor>()) {
continue;
}
auto &main_tensor = main_var->Get<LoDTensor>();
auto &dims = main_tensor.dims();
if (paddle::platform::is_gpu_place(main_tensor.place())) {
size_t numel = main_tensor.numel();
ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type());
......@@ -174,12 +172,17 @@ void ParallelExecutor::SplitTensorToPlaces(
const std::unordered_map<std::string, LoDTensor> &feed_tensors) {
for (auto it : feed_tensors) {
auto lod_tensors = it.second.SplitLoDTensor(member_->places_);
PADDLE_ENFORCE_EQ(
member_->places_.size(), lod_tensors.size(),
"The number of samples of current batch is less than the count of "
"devices, currently, it is not allowed. (%d vs %d)",
member_->places_.size(), lod_tensors.size());
for (size_t j = 0; j < member_->places_.size(); ++j) {
// TODO(panxy0718): Do I need to delete this var?
member_->local_scopes_[j]
->Var(it.first)
->GetMutable<LoDTensor>()
->ShareDataWith(lod_tensors[j]);
auto t =
member_->local_scopes_[j]->Var(it.first)->GetMutable<LoDTensor>();
t->ShareDataWith(lod_tensors[j]);
t->set_lod(lod_tensors[j].lod());
}
}
}
......
paddle/fluid/framework/parallel_executor.h
浏览文件 @ ae5923e2
......@@ -48,13 +48,13 @@ class ParallelExecutor {
const std::string& fetched_var_name,
const std::unordered_map<std::string, LoDTensor>& feed_tensors);
void BCastParamsToGPUs(const std::unordered_set<std::string>& vars) const;
private:
void SplitTensorToPlaces(
const std::unordered_map<std::string, LoDTensor>& feed_tensors);
ParallelExecutorPrivate* member_;
void BCastParamsToGPUs(const std::unordered_set<std::string>& vars) const;
};
} // namespace framework
......
paddle/fluid/framework/program_desc.cc
浏览文件 @ ae5923e2
......@@ -85,9 +85,9 @@ ProgramDesc::ProgramDesc(const std::string &binary_str) {
}
const std::vector<std::string> ProgramDesc::GetFeedTargetNames() {
BlockDesc *global_block = blocks_[0].get();
auto &global_block = Block(0);
std::vector<std::string> feed_target_names;
for (auto *op : global_block->AllOps()) {
for (auto *op : global_block.AllOps()) {
if (op->Type() == kFeedOpType) {
feed_target_names.insert(feed_target_names.begin(), op->Output("Out")[0]);
}
......@@ -96,9 +96,9 @@ const std::vector<std::string> ProgramDesc::GetFeedTargetNames() {
}
const std::vector<std::string> ProgramDesc::GetFetchTargetNames() {
BlockDesc *global_block = blocks_[0].get();
auto &global_block = Block(0);
std::vector<std::string> fetch_target_names;
for (auto *op : global_block->AllOps()) {
for (auto *op : global_block.AllOps()) {
if (op->Type() == kFetchOpType) {
fetch_target_names.push_back(op->Input("X")[0]);
}
......@@ -106,5 +106,43 @@ const std::vector<std::string> ProgramDesc::GetFetchTargetNames() {
return fetch_target_names;
}
void ProgramDesc::SetFeedHolderName(const std::string &feed_holder_name) {
auto *global_block = MutableBlock(0);
int index = 0;
for (auto *op : global_block->AllOps()) {
if (op->Type() == kFeedOpType) {
// Unify the input's name of all feed_ops to feed_holder_name
global_block->RemoveVar(op->Input("X")[0]);
op->SetInput("X", {feed_holder_name});
op->SetAttr("col", {index});
op->CheckAttrs();
index++;
}
}
auto *feed_holder = global_block->Var(feed_holder_name);
feed_holder->SetType(proto::VarType::FEED_MINIBATCH);
feed_holder->SetPersistable(true);
}
void ProgramDesc::SetFetchHolderName(const std::string &fetch_holder_name) {
auto *global_block = MutableBlock(0);
int index = 0;
for (auto *op : global_block->AllOps()) {
if (op->Type() == kFetchOpType) {
// Unify the output's name of all fetch_ops to fetch_holder_name
global_block->RemoveVar(op->Output("Out")[0]);
op->SetOutput("Out", {fetch_holder_name});
op->SetAttr("col", {index});
op->CheckAttrs();
index++;
}
}
auto *fetch_holder = global_block->Var(fetch_holder_name);
fetch_holder->SetType(proto::VarType::FETCH_LIST);
fetch_holder->SetPersistable(true);
}
} // namespace framework
} // namespace paddle
paddle/fluid/framework/program_desc.h
浏览文件 @ ae5923e2
......@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/framework.pb.h"
......@@ -52,9 +53,26 @@ class ProgramDesc {
proto::ProgramDesc *Proto();
// The output variable of feed_op is referenced as feed_target.
// This function is used to collect the output variable's name of all
// feed_ops.
const std::vector<std::string> GetFeedTargetNames();
// The input variable of fetch_op is referenced as fetch_target.
// This function is used to collect the input variable's name of all
// fetch_ops.
const std::vector<std::string> GetFetchTargetNames();
// The input variable of feed_op that holds input Tensor provided by users is
// referenced as feed_holder.
// This function is used to change or unify the feed_holder variables' name.
void SetFeedHolderName(const std::string &feed_holder_name);
// The output variable of fetch_op that holds output Tensor needed by users is
// referenced as fetch_holder.
// This function is used to change or unify the fetch_holder variables' name.
void SetFetchHolderName(const std::string &fetch_holder_name);
private:
proto::ProgramDesc desc_;
......
paddle/fluid/framework/reader.cc
浏览文件 @ ae5923e2
......@@ -22,7 +22,9 @@ FileReader::FileReader(const std::vector<DDim> &dims) : dims_(dims) {}
void FileReader::ReadNext(std::vector<LoDTensor> *out) {
ReadNextImpl(out);
PADDLE_ENFORCE_EQ(out->size(), dims_.size());
if (out->empty()) {
return;
}
for (size_t i = 0; i < dims_.size(); ++i) {
auto &actual = out->at(i).dims();
auto &expect = dims_[i];
......
paddle/fluid/framework/reader.h
浏览文件 @ ae5923e2
......@@ -14,14 +14,13 @@
#pragma once
#include <memory>
#include <vector>
#include "paddle/fluid/framework/ddim.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/platform/place.h"
#include <memory>
#include <thread>
#include <vector>
namespace paddle {
namespace framework {
......@@ -31,8 +30,6 @@ class ReaderBase {
virtual void ReInit() = 0;
virtual bool HasNext() const = 0;
virtual ~ReaderBase();
};
......@@ -44,8 +41,6 @@ class DecoratedReader : public ReaderBase {
void ReInit() override { reader_->ReInit(); }
bool HasNext() const override { return reader_->HasNext(); }
protected:
ReaderBase* reader_;
};
......@@ -80,8 +75,6 @@ class ReaderHolder {
reader_->ReInit();
}
bool HasNext() const { return reader_->HasNext(); }
private:
std::unique_ptr<ReaderBase> reader_;
};
......
paddle/fluid/inference/tests/book/CMakeLists.txt
浏览文件 @ ae5923e2
......@@ -24,7 +24,8 @@ function(inference_test TARGET_NAME)
endforeach()
endfunction(inference_test)
inference_test(fit_a_line)
# This unittest is buggy!
#inference_test(fit_a_line)
inference_test(image_classification ARGS vgg resnet)
inference_test(label_semantic_roles)
inference_test(recognize_digits ARGS mlp conv)
......
paddle/fluid/inference/tests/book/test_inference_fit_a_line.cc
浏览文件 @ ae5923e2
......@@ -12,6 +12,7 @@ limitations under the License. */
#include "gflags/gflags.h"
#include "gtest/gtest.h"
#include "paddle/fluid/inference/tests/test_helper.h"
#include "paddle/fluid/inference/tests/test_multi_thread_helper.h"
DEFINE_string(dirname, "", "Directory of the inference model.");
......@@ -26,32 +27,63 @@ TEST(inference, fit_a_line) {
// 0. Call `paddle::framework::InitDevices()` initialize all the devices
// In unittests, this is done in paddle/testing/paddle_gtest_main.cc
paddle::framework::LoDTensor input;
// The second dim of the input tensor should be 13
// The input data should be >= 0
int64_t batch_size = 10;
SetupTensor<float>(&input, {batch_size, 13}, static_cast<float>(0),
static_cast<float>(10));
std::vector<paddle::framework::LoDTensor*> cpu_feeds;
cpu_feeds.push_back(&input);
for (int num_threads : {1, 2}) {
std::vector<std::vector<paddle::framework::LoDTensor*>> cpu_feeds;
cpu_feeds.resize(num_threads);
for (int i = 0; i < num_threads; ++i) {
auto* input = new paddle::framework::LoDTensor();
// The second dim of the input tensor should be 13
// The input data should be >= 0
int64_t batch_size = 10;
SetupTensor<float>(input, {batch_size, 13}, static_cast<float>(0),
static_cast<float>(10));
cpu_feeds[i].push_back(input);
}
paddle::framework::LoDTensor output1;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
cpu_fetchs1.push_back(&output1);
std::vector<std::vector<paddle::framework::LoDTensor*>> cpu_fetchs1;
cpu_fetchs1.resize(num_threads);
for (int i = 0; i < num_threads; ++i) {
auto* output = new paddle::framework::LoDTensor();
cpu_fetchs1[i].push_back(output);
}
// Run inference on CPU
TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds, cpu_fetchs1);
LOG(INFO) << output1.dims();
// Run inference on CPU
LOG(INFO) << "--- CPU Runs (num_threads: " << num_threads << "): ---";
if (num_threads == 1) {
TestInference<paddle::platform::CPUPlace>(dirname, cpu_feeds[0],
cpu_fetchs1[0]);
} else {
TestMultiThreadInference<paddle::platform::CPUPlace>(
dirname, cpu_feeds, cpu_fetchs1, num_threads);
}
#ifdef PADDLE_WITH_CUDA
paddle::framework::LoDTensor output2;
std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
cpu_fetchs2.push_back(&output2);
std::vector<std::vector<paddle::framework::LoDTensor*>> cpu_fetchs2;
cpu_fetchs2.resize(num_threads);
for (int i = 0; i < num_threads; ++i) {
auto* output = new paddle::framework::LoDTensor();
cpu_fetchs2[i].push_back(output);
}
// Run inference on CUDA GPU
TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds, cpu_fetchs2);
LOG(INFO) << output2.dims();
// Run inference on CUDA GPU
LOG(INFO) << "--- GPU Runs (num_threads: " << num_threads << "): ---";
if (num_threads == 1) {
TestInference<paddle::platform::CUDAPlace>(dirname, cpu_feeds[0],
cpu_fetchs2[0]);
} else {
TestMultiThreadInference<paddle::platform::CUDAPlace>(
dirname, cpu_feeds, cpu_fetchs2, num_threads);
}
CheckError<float>(output1, output2);
for (int i = 0; i < num_threads; ++i) {
CheckError<float>(*cpu_fetchs1[i][0], *cpu_fetchs2[i][0]);
delete cpu_fetchs2[i][0];
}
#endif
for (int i = 0; i < num_threads; ++i) {
delete cpu_feeds[i][0];
delete cpu_fetchs1[i][0];
}
} // num_threads-loop
}
paddle/fluid/inference/tests/test_helper.h
浏览文件 @ ae5923e2
......@@ -25,7 +25,8 @@ limitations under the License. */
template <typename T>
void SetupTensor(paddle::framework::LoDTensor* input,
paddle::framework::DDim dims, T lower, T upper) {
std::mt19937 rng(100); // An arbitrarily chosen but fixed seed.
static unsigned int seed = 100;
std::mt19937 rng(seed++);
std::uniform_real_distribution<double> uniform_dist(0, 1);
T* input_ptr = input->mutable_data<T>(dims, paddle::platform::CPUPlace());
......
paddle/fluid/inference/tests/test_multi_thread_helper.h 0 → 100644
浏览文件 @ ae5923e2
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <map>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/inference/io.h"
void ThreadedRunInference(
const std::unique_ptr<paddle::framework::ProgramDesc>& inference_program,
paddle::framework::Executor* executor, paddle::framework::Scope* scope,
const int thread_id,
const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
const std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
auto copy_program = std::unique_ptr<paddle::framework::ProgramDesc>(
new paddle::framework::ProgramDesc(*inference_program));
std::string feed_holder_name = "feed_" + paddle::string::to_string(thread_id);
std::string fetch_holder_name =
"fetch_" + paddle::string::to_string(thread_id);
copy_program->SetFeedHolderName(feed_holder_name);
copy_program->SetFetchHolderName(fetch_holder_name);
// 3. Get the feed_target_names and fetch_target_names
const std::vector<std::string>& feed_target_names =
copy_program->GetFeedTargetNames();
const std::vector<std::string>& fetch_target_names =
copy_program->GetFetchTargetNames();
// 4. Prepare inputs: set up maps for feed targets
std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
for (size_t i = 0; i < feed_target_names.size(); ++i) {
// Please make sure that cpu_feeds[i] is right for feed_target_names[i]
feed_targets[feed_target_names[i]] = cpu_feeds[i];
}
// 5. Define Tensor to get the outputs: set up maps for fetch targets
std::map<std::string, paddle::framework::LoDTensor*> fetch_targets;
for (size_t i = 0; i < fetch_target_names.size(); ++i) {
fetch_targets[fetch_target_names[i]] = cpu_fetchs[i];
}
// 6. Run the inference program
executor->Run(*copy_program, scope, feed_targets, fetch_targets, true,
feed_holder_name, fetch_holder_name);
}
template <typename Place>
void TestMultiThreadInference(
const std::string& dirname,
const std::vector<std::vector<paddle::framework::LoDTensor*>>& cpu_feeds,
const std::vector<std::vector<paddle::framework::LoDTensor*>>& cpu_fetchs,
const int num_threads) {
// 1. Define place, executor, scope
auto place = Place();
auto executor = paddle::framework::Executor(place);
auto* scope = new paddle::framework::Scope();
// 2. Initialize the inference_program and load parameters
std::unique_ptr<paddle::framework::ProgramDesc> inference_program =
paddle::inference::Load(executor, *scope, dirname);
std::vector<std::thread*> threads;
for (int i = 0; i < num_threads; ++i) {
threads.push_back(new std::thread(
ThreadedRunInference, std::ref(inference_program), &executor, scope, i,
std::ref(cpu_feeds[i]), std::ref(cpu_fetchs[i])));
}
for (int i = 0; i < num_threads; ++i) {
threads[i]->join();
delete threads[i];
}
delete scope;
}
paddle/fluid/operators/activation_op.cc
浏览文件 @ ae5923e2
......@@ -662,14 +662,3 @@ REGISTER_OP(swish, ops::ActivationOp, ops::SwishOpMaker, swish_grad,
ops::grad_functor<double>>);
FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);
REGISTER_OP_CPU_KERNEL(relu,
ops::ActivationKernel<paddle::platform::CPUDeviceContext,
ops::ReluFunctor<float>>,
ops::ActivationKernel<paddle::platform::CPUDeviceContext,
ops::ReluFunctor<double>>);
REGISTER_OP_CPU_KERNEL(
relu_grad, ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
ops::ReluGradFunctor<float>>,
ops::ActivationGradKernel<paddle::platform::CPUDeviceContext,
ops::ReluGradFunctor<double>>);
paddle/fluid/operators/activation_op.cu
浏览文件 @ ae5923e2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
......@@ -17,31 +14,19 @@ limitations under the License. */
#include "paddle/fluid/platform/float16.h"
namespace ops = paddle::operators;
#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, \
namespace plat = paddle::platform;
#define REGISTER_ACTIVATION_CUDA_KERNEL(act_type, functor, grad_functor) \
REGISTER_OP_CUDA_KERNEL( \
act_type, \
ops::ActivationKernel<plat::CUDADeviceContext, ops::functor<float>>, \
ops::ActivationKernel<plat::CUDADeviceContext, ops::functor<double>>, \
ops::ActivationKernel<plat::CUDADeviceContext, \
ops::functor<plat::float16>>); \
REGISTER_OP_CUDA_KERNEL( \
act_type##_grad, ops::ActivationGradKernel<plat::CUDADeviceContext, \
ops::grad_functor<float>>, \
ops::ActivationGradKernel<plat::CUDADeviceContext, \
ops::grad_functor<double>>);
FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);
REGISTER_OP_CUDA_KERNEL(
relu, ops::ActivationKernel<paddle::platform::CUDADeviceContext,
ops::ReluFunctor<float>>,
ops::ActivationKernel<paddle::platform::CUDADeviceContext,
ops::ReluFunctor<double>>,
ops::ActivationKernel<paddle::platform::CUDADeviceContext,
ops::ReluFunctor<paddle::platform::float16>>);
REGISTER_OP_CUDA_KERNEL(
relu_grad, ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,
ops::ReluGradFunctor<float>>,
ops::ActivationGradKernel<paddle::platform::CUDADeviceContext,
ops::ReluGradFunctor<double>>);
paddle/fluid/operators/activation_op.h
浏览文件 @ ae5923e2
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
......@@ -13,9 +10,13 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <utility>
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
#include "paddle/fluid/platform/float16.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
......@@ -336,11 +337,25 @@ struct Sine {
HOSTDEVICE T operator()(const T& val) const { return sin(val); }
};
template <>
struct Sine<platform::float16> {
HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
return platform::float16(sin(static_cast<float>(val)));
}
};
template <typename T>
struct Cosine {
HOSTDEVICE T operator()(const T& val) const { return cos(val); }
};
template <>
struct Cosine<platform::float16> {
HOSTDEVICE platform::float16 operator()(const platform::float16& val) const {
return platform::float16(cos(static_cast<float>(val)));
}
};
// cosine'(x) = -sin(x)
template <typename T>
struct CosGradFunctor : public BaseActivationFunctor<T> {
......@@ -824,6 +839,7 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
__macro(sigmoid, SigmoidFunctor, SigmoidGradFunctor); \
__macro(logsigmoid, LogSigmoidFunctor, LogSigmoidGradFunctor); \
__macro(exp, ExpFunctor, ExpGradFunctor); \
__macro(relu, ReluFunctor, ReluGradFunctor); \
__macro(tanh, TanhFunctor, TanhGradFunctor); \
__macro(softshrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \
__macro(sqrt, SqrtFunctor, SqrtGradFunctor); \
......
paddle/fluid/operators/adagrad_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/adagrad_op.h"
#include <vector>
#include <cmath>
......
paddle/fluid/operators/array_operator.h
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/device_context.h"
......
paddle/fluid/operators/assign_value_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,8 @@
// limitations under the License.
#include "paddle/fluid/operators/assign_value_op.h"
#include <string>
#include <vector>
namespace paddle {
namespace operators {
......
paddle/fluid/operators/assign_value_op.h
浏览文件 @ ae5923e2
......@@ -14,6 +14,7 @@
#pragma once
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/enforce.h"
......
paddle/fluid/operators/auc_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/auc_op.h"
#include <string>
namespace paddle {
namespace operators {
......
paddle/fluid/operators/auc_op.h
浏览文件 @ ae5923e2
......@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
......@@ -40,7 +42,7 @@ class AucKernel : public framework::OpKernel<T> {
std::vector<float> thresholds_list;
thresholds_list.reserve(num_thresholds);
for (int i = 1; i < num_thresholds - 1; i++) {
thresholds_list[i] = (float)i / (num_thresholds - 1);
thresholds_list[i] = static_cast<float>(i) / (num_thresholds - 1);
}
const float kEpsilon = 1e-7;
thresholds_list[0] = 0.0f - kEpsilon;
......@@ -105,11 +107,12 @@ class AucKernel : public framework::OpKernel<T> {
float* fp_rate_data = fp_rate.mutable_data<float>(ctx.GetPlace());
float* rec_rate_data = rec_rate.mutable_data<float>(ctx.GetPlace());
for (int i = 0; i < num_thresholds; i++) {
tp_rate_data[i] =
((float)tp_data[i] + epsilon) / (tp_data[i] + fn_data[i] + epsilon);
fp_rate_data[i] = (float)fp_data[i] / (fp_data[i] + tn_data[i] + epsilon);
rec_rate_data[i] =
((float)tp_data[i] + epsilon) / (tp_data[i] + fp_data[i] + epsilon);
tp_rate_data[i] = (static_cast<float>(tp_data[i]) + epsilon) /
(tp_data[i] + fn_data[i] + epsilon);
fp_rate_data[i] =
static_cast<float>(fp_data[i]) / (fp_data[i] + tn_data[i] + epsilon);
rec_rate_data[i] = (static_cast<float>(tp_data[i]) + epsilon) /
(tp_data[i] + fp_data[i] + epsilon);
}
*auc_data = 0.0f;
if (curve == "ROC") {
......
paddle/fluid/operators/average_accumulates_op.cc
浏览文件 @ ae5923e2
......@@ -19,15 +19,15 @@ namespace operators {
template <>
void GetAccumulators<paddle::platform::CPUDeviceContext>(
const framework::ExecutionContext& ctx, int64_t& num_updates_,
int64_t& num_accumulates_, int64_t& old_num_accumulates_) {
const framework::ExecutionContext& ctx, int64_t* num_updates_,
int64_t* num_accumulates_, int64_t* old_num_accumulates_) {
auto* in_old_num_accumulates = ctx.Input<Tensor>("in_old_num_accumulates");
auto* in_num_accumulates = ctx.Input<Tensor>("in_num_accumulates");
auto* in_num_updates = ctx.Input<Tensor>("in_num_updates");
old_num_accumulates_ = in_old_num_accumulates->data<int64_t>()[0];
num_accumulates_ = in_num_accumulates->data<int64_t>()[0];
num_updates_ = in_num_updates->data<int64_t>()[0];
*old_num_accumulates_ = in_old_num_accumulates->data<int64_t>()[0];
*num_accumulates_ = in_num_accumulates->data<int64_t>()[0];
*num_updates_ = in_num_updates->data<int64_t>()[0];
}
template <>
......
paddle/fluid/operators/average_accumulates_op.cu
浏览文件 @ ae5923e2
......@@ -19,18 +19,18 @@ namespace paddle {
namespace operators {
template <>
void GetAccumulators<paddle::platform::CUDADeviceContext>(
const framework::ExecutionContext& ctx, int64_t& num_updates_,
int64_t& num_accumulates_, int64_t& old_num_accumulates_) {
const framework::ExecutionContext& ctx, int64_t* num_updates_,
int64_t* num_accumulates_, int64_t* old_num_accumulates_) {
auto* in_old_num_accumulates = ctx.Input<Tensor>("in_old_num_accumulates");
auto* in_num_accumulates = ctx.Input<Tensor>("in_num_accumulates");
auto* in_num_updates = ctx.Input<Tensor>("in_num_updates");
auto stream = ctx.cuda_device_context().stream();
memory::Copy(platform::CPUPlace(), &old_num_accumulates_,
memory::Copy(platform::CPUPlace(), old_num_accumulates_,
platform::CUDAPlace(), in_old_num_accumulates->data<int64_t>(),
sizeof(int64_t), stream);
memory::Copy(platform::CPUPlace(), &num_accumulates_, platform::CUDAPlace(),
memory::Copy(platform::CPUPlace(), num_accumulates_, platform::CUDAPlace(),
in_num_accumulates->data<int64_t>(), sizeof(int64_t), stream);
memory::Copy(platform::CPUPlace(), &num_updates_, platform::CUDAPlace(),
memory::Copy(platform::CPUPlace(), num_updates_, platform::CUDAPlace(),
in_num_updates->data<int64_t>(), sizeof(int64_t), stream);
}
......
paddle/fluid/operators/average_accumulates_op.h
浏览文件 @ ae5923e2
......@@ -29,8 +29,8 @@ using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename DeviceContext>
void GetAccumulators(const framework::ExecutionContext& ctx,
int64_t& num_updates, int64_t& num_accumulates,
int64_t& old_num_accumulates);
int64_t* num_updates, int64_t* num_accumulates,
int64_t* old_num_accumulates);
template <typename DeviceContext>
void SetAccumulators(const framework::ExecutionContext& ctx,
......@@ -47,8 +47,8 @@ class AverageAccumulatesKernel : public framework::OpKernel<T> {
int64_t num_updates = 0;
int64_t num_accumulates = 0;
int64_t old_num_accumulates = 0;
GetAccumulators<DeviceContext>(ctx, num_updates, num_accumulates,
old_num_accumulates);
GetAccumulators<DeviceContext>(ctx, &num_updates, &num_accumulates,
&old_num_accumulates);
// Get attrs
float average_window = ctx.Attr<float>("average_window");
......
paddle/fluid/operators/batch_norm_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/batch_norm_op.h"
#include <string>
#include "paddle/fluid/framework/data_layout.h"
namespace paddle {
......
paddle/fluid/operators/batch_norm_op.cu.cc
浏览文件 @ ae5923e2
......@@ -13,9 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/batch_norm_op.h"
#include "paddle/fluid/framework/data_layout.h"
#include <cfloat>
#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/cudnn_helper.h"
#include "paddle/fluid/platform/float16.h"
......
paddle/fluid/operators/batch_size_like.h
浏览文件 @ ae5923e2
......@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <algorithm>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
......
paddle/fluid/operators/box_coder_op.h
浏览文件 @ ae5923e2
......@@ -10,6 +10,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
......
paddle/fluid/operators/compare_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/compare_op.h"
#include <string>
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
......
paddle/fluid/operators/concat_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/concat_op.h"
#include <string>
#include <vector>
namespace paddle {
......
paddle/fluid/operators/cond_op.h
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include <vector>
#include "glog/logging.h"
#include "paddle/fluid/framework/ddim.h"
......
paddle/fluid/operators/conv_transpose_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/conv_transpose_op.h"
#include <string>
#include <vector>
namespace paddle {
namespace operators {
......
paddle/fluid/operators/conv_transpose_op.h
浏览文件 @ ae5923e2
......@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/im2col.h"
......
paddle/fluid/operators/crf_decoding_op.h
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <limits>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
......
paddle/fluid/operators/crop_op.h
浏览文件 @ ae5923e2
......@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <utility>
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/strided_memcpy.h"
......
paddle/fluid/operators/detail/grpc_client.cc
浏览文件 @ ae5923e2
......@@ -65,9 +65,8 @@ bool RPCClient::AsyncSendVariable(const std::string& ep,
}
void ProcGetResponse(const VarHandle& var_h,
// const sendrecv::VariableMessage& ret_msg) {
const ::grpc::ByteBuffer& ret_msg) {
framework::Variable* outvar = NULL;
framework::Variable* outvar = nullptr;
DeserializeFromByteBuffer(ret_msg, *var_h.ctx, var_h.scope, &outvar);
}
......
paddle/fluid/operators/detail/serde_test.cc
浏览文件 @ ae5923e2
......@@ -107,7 +107,7 @@ void RunSerdeTestSelectedRows(platform::Place place) {
for (int i = 0; i < tensor_numel; ++i) {
EXPECT_FLOAT_EQ(tensor_data2[i], 32.7);
}
for (int64_t i = 0; i < rows2->size(); ++i) {
for (size_t i = 0; i < rows2->size(); ++i) {
EXPECT_EQ(rows_data2[i], i);
}
EXPECT_EQ(slr2->height(), 1000);
......
paddle/fluid/operators/elementwise_op_function.h
浏览文件 @ ae5923e2
......@@ -13,14 +13,15 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <algorithm>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/transform.h"
#ifdef __NVCC__
#include <cuda.h>
#include <thrust/iterator/iterator_adaptor.h>
#include "paddle/fluid/platform/cuda_helper.h"
constexpr int ELEMWISE_MAX_BLOCK_DIM = 1024;
#endif
......@@ -43,35 +44,35 @@ namespace operators {
*/
inline void get_mid_dims(const framework::DDim& x_dims,
const framework::DDim& y_dims, const int axis,
int& pre, int& n, int& post) {
pre = 1;
n = 1;
post = 1;
int* pre, int* n, int* post) {
*pre = 1;
*n = 1;
*post = 1;
for (int i = 0; i < axis; ++i) {
pre *= x_dims[i];
(*pre) *= x_dims[i];
}
for (int i = 0; i < y_dims.size(); ++i) {
PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
"Broadcast dimension mismatch.");
n *= y_dims[i];
(*n) *= y_dims[i];
}
for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
post *= x_dims[i];
(*post) *= x_dims[i];
}
}
inline void trim_trailing_singular_dims(framework::DDim& dims) {
inline void trim_trailing_singular_dims(framework::DDim* dims) {
// Remove trailing dimensions of size 1 for y
auto actual_dims_size = dims.size();
auto actual_dims_size = dims->size();
for (; actual_dims_size != 0; --actual_dims_size) {
if (dims[actual_dims_size - 1] != 1) break;
if ((*dims)[actual_dims_size - 1] != 1) break;
}
if (actual_dims_size != dims.size()) {
auto actual_dims = framework::vectorize(dims);
if (actual_dims_size != dims->size()) {
auto actual_dims = framework::vectorize(*dims);
actual_dims.resize(actual_dims_size);
dims = framework::make_ddim(actual_dims);
*dims = framework::make_ddim(actual_dims);
}
}
......@@ -159,7 +160,7 @@ class RowwiseTransformIterator<T, platform::CUDADeviceContext>
RowwiseTransformIterator<T, platform::CUDADeviceContext>, const T*>
super_t;
HOSTDEVICE RowwiseTransformIterator(const T* x, int n)
: super_t(x), begin_(x), n_(n){};
: super_t(x), begin_(x), n_(n) {}
friend class thrust::iterator_core_access;
private:
......@@ -179,7 +180,7 @@ class MidWiseTransformIterator<T, platform::CUDADeviceContext>
MidWiseTransformIterator<T, platform::CUDADeviceContext>, const T*>
super_t;
HOSTDEVICE MidWiseTransformIterator(const T* x, int n, int post)
: super_t(x), begin_(x), n_(n), post_(post){};
: super_t(x), begin_(x), n_(n), post_(post) {}
friend class thrust::iterator_core_access;
private:
......@@ -333,6 +334,55 @@ static void ElemwiseGradBroadcast1CPU(const T* x, const T* y, const T* out,
}
}
#ifdef __NVCC__
// __shfl_down has been deprecated as of CUDA 9.0.
#if CUDA_VERSION < 9000
template <typename T>
__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) {
return __shfl_down(val, delta);
}
#define CREATE_SHFL_MASK(mask, predicate) mask = 0u;
#else
#define FULL_WARP_MASK 0xFFFFFFFF
#define CREATE_SHFL_MASK(mask, predicate) \
mask = __ballot_sync(FULL_WARP_MASK, (predicate))
#endif
template <typename T>
__device__ T reduceSum(T val, int tid, int len) {
// TODO(zcd): The warp size should be taken from the
// parameters of the GPU but not specified as 32 simply.
// To make the reduceSum more efficiently,
// I use Warp-Level Parallelism and assume the Warp size
// is 32 which may be different for different GPU,
// but most card's warp size is 32.
__shared__ T shm[32];
const int warpSize = 32;
unsigned mask = 0u;
CREATE_SHFL_MASK(mask, tid < len);
for (int offset = warpSize / 2; offset > 0; offset /= 2)
val += __shfl_down_sync(mask, val, offset);
if (tid < warpSize) shm[tid] = 0;
__syncthreads();
if (tid % warpSize == 0) {
shm[tid / warpSize] = val;
}
CREATE_SHFL_MASK(mask, tid < warpSize);
if (tid < warpSize) {
val = shm[tid];
for (int offset = warpSize / 2; offset > 0; offset /= 2)
val += __shfl_down_sync(mask, val, offset);
}
return val;
}
template <typename T, typename DX_OP, typename DY_OP>
static __global__ void ElemwiseGradBroadcast1CUDAKernel(
const T* x, const T* y, const T* out, const T* dout, int h, int w,
......@@ -355,7 +405,7 @@ static __global__ void ElemwiseGradBroadcast1CUDAKernel(
if (dy) {
h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
val = platform::reduceSum(val, tid, h);
val = reduceSum(val, tid, h);
if (threadIdx.x == 0) {
dy[j] = val;
}
......@@ -432,7 +482,7 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel(
if (dy) {
int h = pre * post;
h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
val = platform::reduceSum(val, tid, h);
val = reduceSum(val, tid, h);
if (threadIdx.x == 0) {
dy[j] = val;
}
......@@ -472,11 +522,11 @@ void ElemwiseGradCompute(const framework::ExecutionContext& ctx,
auto y_dim = y.dims();
axis = (axis == -1 ? x_dim.size() - y_dim.size() : axis);
trim_trailing_singular_dims(y_dim);
trim_trailing_singular_dims(&y_dim);
axis = (y_dim.size() == 0) ? x_dim.size() : axis;
int pre, n, post;
get_mid_dims(x_dim, y_dim, axis, pre, n, post);
get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post);
if (post == 1) {
int h = pre;
int w = n;
......@@ -514,7 +564,7 @@ void ElemwiseGradCompute(const framework::ExecutionContext& ctx,
}
}
}
};
}
template <typename DeviceContext, typename T, typename functor,
typename broadcastfunctor, typename broadcast2functor>
......@@ -543,11 +593,11 @@ void ElementwiseGradCompute(const framework::ExecutionContext& ctx,
}
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
trim_trailing_singular_dims(y_dims);
trim_trailing_singular_dims(&y_dims);
axis = (y_dims.size() == 0) ? x_dims.size() : axis;
int pre, n, post;
get_mid_dims(x_dims, y_dims, axis, pre, n, post);
get_mid_dims(x_dims, y_dims, axis, &pre, &n, &post);
if (post == 1) {
broadcastfunctor f;
......@@ -582,11 +632,11 @@ void ElementwiseComputeEx(const framework::ExecutionContext& ctx,
axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
"Axis should be in range [0, x_dims)");
trim_trailing_singular_dims(y_dims);
trim_trailing_singular_dims(&y_dims);
axis = (y_dims.size() == 0) ? x_dims.size() : axis;
int pre, n, post;
get_mid_dims(x_dims, y_dims, axis, pre, n, post);
get_mid_dims(x_dims, y_dims, axis, &pre, &n, &post);
if (post == 1) {
functor.RunRowWise(n, pre);
return;
......
paddle/fluid/operators/math/math_function.cu
浏览文件 @ ae5923e2
......@@ -39,13 +39,14 @@ void gemm<platform::CUDADeviceContext, float16>(
cublasOperation_t cuTransB =
(transB == CblasNoTrans) ? CUBLAS_OP_N : CUBLAS_OP_T;
float h_alpha = static_cast<float>(alpha);
float h_beta = static_cast<float>(beta);
// TODO(kexinzhao): add processing code for compute capability < 53 case
PADDLE_ENFORCE_GE(context.GetComputeCapability(), 53,
"cublas fp16 gemm requires GPU compute capability >= 53");
#if CUDA_VERSION >= 8000
float h_alpha = static_cast<float>(alpha);
float h_beta = static_cast<float>(beta);
cublasGemmAlgo_t algo = CUBLAS_GEMM_DFALT;
#if CUDA_VERSION >= 9000
if (context.GetComputeCapability() >= 70) {
......@@ -56,7 +57,7 @@ void gemm<platform::CUDADeviceContext, float16>(
PADDLE_ENFORCE(platform::dynload::cublasSetMathMode(context.cublas_handle(),
CUBLAS_DEFAULT_MATH));
}
#endif
#endif // CUDA_VERSION >= 9000
// cublasHgemm does true FP16 computation which is slow for non-Volta
// GPUs. So use cublasGemmEx instead which does pesudo FP16 computation:
......@@ -66,6 +67,18 @@ void gemm<platform::CUDADeviceContext, float16>(
context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, B,
CUDA_R_16F, ldb, A, CUDA_R_16F, lda, &h_beta, C, CUDA_R_16F, N,
CUDA_R_32F, algo));
#else
// CUDA 7.5 does not support cublasGemmEx, hence we fall back to use hgemm
const half h_alpha = static_cast<const half>(alpha);
const half h_beta = static_cast<const half>(beta);
const half* h_A = reinterpret_cast<const half*>(A);
const half* h_B = reinterpret_cast<const half*>(B);
half* h_C = reinterpret_cast<half*>(C);
PADDLE_ENFORCE(platform::dynload::cublasHgemm(
context.cublas_handle(), cuTransB, cuTransA, N, M, K, &h_alpha, h_B, ldb,
h_A, lda, &h_beta, h_C, N));
#endif // CUDA_VERSION >= 8000
}
template <>
......
paddle/fluid/operators/read_op.cc
浏览文件 @ ae5923e2
......@@ -66,13 +66,7 @@ class ReadOp : public framework::OperatorBase {
std::vector<std::string> out_arg_names = Outputs("Out");
std::vector<framework::LoDTensor> ins;
reader->ReadNext(&ins);
if (ins.empty()) {
reader->ReInit();
reader->ReadNext(&ins);
PADDLE_ENFORCE(
!ins.empty(),
"Reader can not read the next data even it has been re-initialized.");
}
PADDLE_ENFORCE(!ins.empty(), "There is no next data.");
PADDLE_ENFORCE_EQ(ins.size(), out_arg_names.size());
for (size_t i = 0; i < ins.size(); ++i) {
auto* out =
......
paddle/fluid/operators/reader/CMakeLists.txt
浏览文件 @ ae5923e2
......@@ -22,5 +22,6 @@ reader_library(create_batch_reader_op SRCS create_batch_reader_op.cc)
reader_library(create_recordio_file_reader_op SRCS create_recordio_file_reader_op.cc)
reader_library(create_double_buffer_reader_op SRCS create_double_buffer_reader_op.cc)
reader_library(create_multi_pass_reader_op SRCS create_multi_pass_reader_op.cc)
reader_library(create_threaded_reader_op SRCS create_threaded_reader_op.cc)
# Export local libraries to parent
set(READER_LIBRARY ${LOCAL_READER_LIBS} PARENT_SCOPE)
paddle/fluid/operators/reader/create_double_buffer_reader_op.cc
浏览文件 @ ae5923e2
......@@ -63,13 +63,14 @@ class DoubleBufferReader : public framework::DecoratedReader {
StartPrefetcher();
}
bool HasNext() const override;
void ReadNext(std::vector<framework::LoDTensor>* out) override;
void ReInit() override;
~DoubleBufferReader() { EndPrefetcher(); }
private:
bool HasNext() const;
void StartPrefetcher() {
channel_ = framework::MakeChannel<Item>(kChannelSize);
prefetcher_ = std::thread([this] { PrefetchThreadFunc(); });
......@@ -109,7 +110,9 @@ class CreateDoubleBufferReaderOp : public framework::OperatorBase {
auto place_str = Attr<std::string>("place");
platform::Place place;
if (place_str == "CPU") {
if (place_str == "AUTO") {
place = dev_place;
} else if (place_str == "CPU") {
place = platform::CPUPlace();
} else {
std::istringstream sin(place_str);
......@@ -140,28 +143,22 @@ class CreateDoubleBufferReaderOpMaker : public DecoratedReaderMakerBase {
enum_range.insert(string::Sprintf("CUDA:%d", i));
}
enum_range.insert("CPU");
AddAttr<std::string>("place", "The double buffer place, default is CPU")
.SetDefault("CPU")
enum_range.insert("AUTO");
AddAttr<std::string>("place", "The double buffer place")
.SetDefault("AUTO")
.InEnum({enum_range});
}
};
bool DoubleBufferReader::HasNext() const {
while (!channel_->IsClosed() && !channel_->CanReceive()) {
}
return channel_->CanReceive();
}
void DoubleBufferReader::ReadNext(std::vector<framework::LoDTensor>* out) {
if (!HasNext()) {
PADDLE_THROW("There is no next data!");
}
Item batch;
channel_->Receive(&batch);
*out = batch.payloads_;
if (batch.ctx_) {
batch.ctx_->Wait();
out->clear();
if (HasNext()) {
Item batch;
channel_->Receive(&batch);
*out = batch.payloads_;
if (batch.ctx_) {
batch.ctx_->Wait();
}
}
}
......@@ -171,16 +168,26 @@ void DoubleBufferReader::ReInit() {
StartPrefetcher();
}
bool DoubleBufferReader::HasNext() const {
while (!channel_->IsClosed() && !channel_->CanReceive()) {
}
return channel_->CanReceive();
}
void DoubleBufferReader::PrefetchThreadFunc() {
VLOG(5) << "A new prefetch thread starts.";
std::vector<std::vector<framework::LoDTensor>> cpu_tensor_cache(kCacheSize);
std::vector<std::vector<framework::LoDTensor>> gpu_tensor_cache(kCacheSize);
size_t cached_tensor_id = 0;
while (reader_->HasNext()) {
while (true) {
Item batch;
auto& cpu_batch = cpu_tensor_cache[cached_tensor_id];
reader_->ReadNext(&cpu_batch);
if (cpu_batch.empty()) {
// The underlying reader have no next data.
break;
}
if (platform::is_gpu_place(place_)) {
auto& gpu_batch = gpu_tensor_cache[cached_tensor_id];
auto* gpu_ctx = ctxs_[cached_tensor_id].get();
......
paddle/fluid/operators/reader/create_multi_pass_reader_op.cc
浏览文件 @ ae5923e2
......@@ -25,22 +25,12 @@ class MultiPassReader : public framework::DecoratedReader {
: DecoratedReader(reader), pass_num_(pass_num), pass_count_(0) {}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
if (!HasNext()) {
PADDLE_THROW("There is no next data!");
}
reader_->ReadNext(out);
}
bool HasNext() const override {
if (reader_->HasNext()) {
return true;
} else {
if (out->empty()) {
++pass_count_;
if (pass_count_ >= pass_num_) {
return false;
} else {
if (pass_count_ < pass_num_) {
reader_->ReInit();
return true;
reader_->ReadNext(out);
}
}
}
......
paddle/fluid/operators/reader/create_random_data_generator_op.cc
浏览文件 @ ae5923e2
......@@ -52,8 +52,6 @@ class RandomDataGenerator : public framework::ReaderBase {
void ReInit() override { return; }
bool HasNext() const override { return true; }
private:
float min_;
float max_;
......@@ -74,7 +72,7 @@ class CreateRandomDataGeneratorOp : public framework::OperatorBase {
const auto& ranks = Attr<std::vector<int>>("ranks");
PADDLE_ENFORCE(!shape_concat.empty() && !ranks.empty());
PADDLE_ENFORCE_EQ(std::accumulate(ranks.begin(), ranks.end(), 0),
int(shape_concat.size()),
static_cast<int>(shape_concat.size()),
"The accumulate of all ranks should be equal to the "
"shape concat's length.");
std::vector<framework::DDim> shapes = RestoreShapes(shape_concat, ranks);
......
paddle/fluid/operators/reader/create_recordio_file_reader_op.cc
浏览文件 @ ae5923e2
......@@ -12,8 +12,6 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <mutex>
#include <thread>
#include "paddle/fluid/operators/reader/reader_op_registry.h"
#include "paddle/fluid/recordio/scanner.h"
......@@ -35,17 +33,15 @@ class RecordIOFileReader : public framework::FileReader {
LOG(INFO) << "Creating file reader" << filename;
}
bool HasNext() const override { return scanner_.HasNext(); }
void ReInit() override { scanner_.Reset(); }
protected:
void ReadNextImpl(std::vector<framework::LoDTensor>* out) override {
if (ThreadSafe) {
std::lock_guard<std::mutex> guard(*mutex_);
*out = framework::ReadFromRecordIO(scanner_, dev_ctx_);
*out = framework::ReadFromRecordIO(&scanner_, dev_ctx_);
} else {
*out = framework::ReadFromRecordIO(scanner_, dev_ctx_);
*out = framework::ReadFromRecordIO(&scanner_, dev_ctx_);
}
}
......@@ -66,7 +62,7 @@ class CreateRecordIOReaderOp : public framework::OperatorBase {
const auto& ranks = Attr<std::vector<int>>("ranks");
PADDLE_ENFORCE(!shape_concat.empty() && !ranks.empty());
PADDLE_ENFORCE_EQ(std::accumulate(ranks.begin(), ranks.end(), 0),
int(shape_concat.size()),
static_cast<int>(shape_concat.size()),
"The accumulate of all ranks should be equal to the "
"shape concat's length.");
std::string filename = Attr<std::string>("filename");
......
paddle/fluid/operators/reader/create_shuffle_reader_op.cc
浏览文件 @ ae5923e2
......@@ -30,35 +30,33 @@ class ShuffleReader : public framework::DecoratedReader {
std::random_device device;
seed_ = device();
}
ReadIntoBuffers();
ReloadBuffer();
}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
if (!HasNext()) {
PADDLE_THROW("There is no next data!");
}
out->clear();
if (iteration_pos_ >= buffer_.size()) {
VLOG(10) << "Resetting shuffle buffer";
ReadIntoBuffers();
ReloadBuffer();
if (buffer_.empty()) {
return;
}
}
*out = buffer_[iteration_pos_++];
}
bool HasNext() const override {
return iteration_pos_ < buffer_.size() || reader_->HasNext();
}
private:
void ReadIntoBuffers() {
void ReloadBuffer() {
buffer_.clear();
buffer_.reserve(buffer_size_);
iteration_pos_ = 0;
for (size_t i = 0; i < buffer_size_; ++i) {
if (!reader_->HasNext()) {
std::vector<framework::LoDTensor> ins;
reader_->ReadNext(&ins);
if (ins.empty()) {
break;
}
buffer_.emplace_back();
reader_->ReadNext(&buffer_.back());
buffer_.emplace_back(ins);
}
std::mt19937 g(seed_);
std::shuffle(buffer_.begin(), buffer_.end(), g);
......
paddle/fluid/operators/reader/create_threaded_reader_op.cc 0 → 100644
浏览文件 @ ae5923e2
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/operators/detail/safe_ref.h"
#include "paddle/fluid/operators/reader/reader_op_registry.h"
namespace paddle {
namespace operators {
namespace reader {
class ThreadedReader : public framework::DecoratedReader {
public:
ThreadedReader(ReaderBase* reader, bool safe_mode)
: DecoratedReader(reader), safe_mode_(safe_mode) {}
void ReadNext(std::vector<framework::LoDTensor>* out) override {
std::lock_guard<std::mutex> lock(mutex_);
reader_->ReadNext(out);
}
void ReInit() override {
if (safe_mode_) {
PADDLE_THROW(
"ThreadedReader::ReInit() is disabled when 'safe_mode' is true.");
}
VLOG(5) << "ThreadedReader::ReInit() is invoked! It might be buggy in "
"multi-thread environment.";
reader_->ReInit();
}
private:
bool safe_mode_;
std::mutex mutex_;
};
class CreateThreadedReaderOp : public framework::OperatorBase {
public:
using framework::OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope& scope,
const platform::Place& dev_place) const override {
auto* out = detail::Ref(scope.FindVar(Output("Out")))
.GetMutable<framework::ReaderHolder>();
if (out->Get() != nullptr) {
return;
}
const auto& underlying_reader = scope.FindVar(Input("UnderlyingReader"))
->Get<framework::ReaderHolder>();
bool safe_mode = Attr<bool>("safe_mode");
out->Reset(new ThreadedReader(underlying_reader.Get(), safe_mode));
}
};
class CreateThreadedReaderOpMaker : public DecoratedReaderMakerBase {
public:
CreateThreadedReaderOpMaker(OpProto* op_proto, OpAttrChecker* op_checker)
: DecoratedReaderMakerBase(op_proto, op_checker) {
AddAttr<bool>("safe_mode",
"When 'safe_mode' is true, 'ReInit()' is disabled to avoid "
"unexpected bugs in multi-thread environment.")
.SetDefault(true);
AddComment(R"DOC(
CreateThreadedReader Operator
This operator creates a threaded reader. A threaded reader's
'ReadNext()' can be invoked by several threads at the same
time.
When the attribute 'safe_mode' is true, the threaded reader's
'ReInit()' is disabled to avoid unexpected bugs in multi-thread
environment.
)DOC");
}
};
} // namespace reader
} // namespace operators
} // namespace paddle
namespace reader = paddle::operators::reader;
REGISTER_DECORATED_READER_OPERATOR(create_threaded_reader,
reader::CreateThreadedReaderOp,
reader::CreateThreadedReaderOpMaker);
paddle/fluid/operators/reader/open_files_op.cc
浏览文件 @ ae5923e2
......@@ -12,6 +12,8 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <thread> // NOLINT
#include "paddle/fluid/framework/channel.h"
#include "paddle/fluid/operators/reader/reader_op_registry.h"
......@@ -19,38 +21,23 @@ namespace paddle {
namespace operators {
namespace reader {
class MultipleReader : public framework::ReaderBase {
class MultiFileReader : public framework::ReaderBase {
public:
class ThreadBufferMap {
public:
std::vector<framework::LoDTensor>& operator[](
const std::thread::id& thread_id) {
std::lock_guard<std::mutex> lock(mutex_);
return buffer_[thread_id];
}
void Clear() { buffer_.clear(); }
private:
std::mutex mutex_;
std::unordered_map<std::thread::id, std::vector<framework::LoDTensor>>
buffer_;
};
MultipleReader(const std::vector<std::string>& file_names,
const std::vector<framework::DDim>& dims, size_t thread_num)
: file_names_(file_names), dims_(dims) {
MultiFileReader(const std::vector<std::string>& file_names,
const std::vector<framework::DDim>& dims, size_t thread_num,
size_t buffer_size)
: file_names_(file_names), dims_(dims), buffer_size_(buffer_size) {
prefetchers_.resize(thread_num);
StartNewScheduler();
}
void ReadNext(std::vector<framework::LoDTensor>* out) override;
bool HasNext() const override;
void ReInit() override;
~MultipleReader() { EndScheduler(); }
~MultiFileReader() { EndScheduler(); }
private:
bool HasNext();
void StartNewScheduler();
void EndScheduler();
void ScheduleThreadFunc();
......@@ -60,39 +47,36 @@ class MultipleReader : public framework::ReaderBase {
std::vector<framework::DDim> dims_;
std::thread scheduler_;
std::vector<std::thread> prefetchers_;
size_t buffer_size_;
framework::Channel<size_t>* waiting_file_idx_;
framework::Channel<size_t>* available_thread_idx_;
framework::Channel<std::vector<framework::LoDTensor>>* buffer_;
mutable ThreadBufferMap thread_buffer_map_;
};
void MultipleReader::ReadNext(std::vector<framework::LoDTensor>* out) {
if (!HasNext()) {
PADDLE_THROW("There is no next data!");
void MultiFileReader::ReadNext(std::vector<framework::LoDTensor>* out) {
out->clear();
if (HasNext()) {
buffer_->Receive(out);
}
auto& thread_local_buffer = thread_buffer_map_[std::this_thread::get_id()];
*out = thread_local_buffer;
thread_local_buffer.clear();
}
bool MultipleReader::HasNext() const {
auto& thread_local_buffer = thread_buffer_map_[std::this_thread::get_id()];
return thread_local_buffer.empty() ? buffer_->Receive(&thread_local_buffer)
: true;
}
void MultipleReader::ReInit() {
void MultiFileReader::ReInit() {
EndScheduler();
thread_buffer_map_.Clear();
StartNewScheduler();
}
void MultipleReader::StartNewScheduler() {
bool MultiFileReader::HasNext() {
while (!buffer_->IsClosed() && !buffer_->CanReceive()) {
}
return buffer_->CanReceive();
}
void MultiFileReader::StartNewScheduler() {
size_t thread_num = prefetchers_.size();
waiting_file_idx_ = framework::MakeChannel<size_t>(file_names_.size());
available_thread_idx_ = framework::MakeChannel<size_t>(thread_num);
buffer_ =
framework::MakeChannel<std::vector<framework::LoDTensor>>(thread_num);
framework::MakeChannel<std::vector<framework::LoDTensor>>(buffer_size_);
for (size_t i = 0; i < file_names_.size(); ++i) {
waiting_file_idx_->Send(&i);
......@@ -105,7 +89,7 @@ void MultipleReader::StartNewScheduler() {
scheduler_ = std::thread([this] { ScheduleThreadFunc(); });
}
void MultipleReader::EndScheduler() {
void MultiFileReader::EndScheduler() {
available_thread_idx_->Close();
buffer_->Close();
waiting_file_idx_->Close();
......@@ -117,8 +101,8 @@ void MultipleReader::EndScheduler() {
delete waiting_file_idx_;
}
void MultipleReader::ScheduleThreadFunc() {
VLOG(5) << "MultipleReader schedule thread starts.";
void MultiFileReader::ScheduleThreadFunc() {
VLOG(5) << "MultiFileReader schedule thread starts.";
size_t completed_thread_num = 0;
size_t thread_idx;
while (available_thread_idx_->Receive(&thread_idx)) {
......@@ -150,17 +134,20 @@ void MultipleReader::ScheduleThreadFunc() {
p.join();
}
}
VLOG(5) << "MultipleReader schedule thread terminates.";
VLOG(5) << "MultiFileReader schedule thread terminates.";
}
void MultipleReader::PrefetchThreadFunc(std::string file_name,
size_t thread_idx) {
void MultiFileReader::PrefetchThreadFunc(std::string file_name,
size_t thread_idx) {
VLOG(5) << "The prefetch thread of file '" << file_name << "' starts.";
std::unique_ptr<framework::ReaderBase> reader =
CreateReaderByFileName(file_name, dims_);
while (reader->HasNext()) {
while (true) {
std::vector<framework::LoDTensor> ins;
reader->ReadNext(&ins);
if (ins.empty()) {
break;
}
try {
buffer_->Send(&ins);
} catch (paddle::platform::EnforceNotMet e) {
......@@ -197,11 +184,13 @@ class OpenFilesOp : public framework::OperatorBase {
const auto& file_names = Attr<std::vector<std::string>>("file_names");
PADDLE_ENFORCE(!file_names.empty(), "No file to be read!");
const size_t thread_num = Attr<int>("thread_num");
const size_t buffer_size = Attr<int>("buffer_size");
auto* out = scope.FindVar(Output("Out"))
->template GetMutable<framework::ReaderHolder>();
out->Reset(new MultipleReader(
file_names, RestoreShapes(shape_concat, ranks), thread_num));
out->Reset(new MultiFileReader(file_names,
RestoreShapes(shape_concat, ranks),
thread_num, buffer_size));
}
};
......@@ -212,11 +201,12 @@ class OpenFilesOpMaker : public FileReaderMakerBase {
AddAttr<std::vector<std::string>>("file_names", "Files to be read.");
AddAttr<int>("thread_num", "The maximal concurrent prefetch thread number.")
.GreaterThan(0);
AddAttr<int>("buffer_size", "The size of prefetch buffer.").GreaterThan(0);
AddComment(R"DOC(
OpenFiles Operator
An OpenFilesOp creates a MultipleReader, which is able to
An OpenFilesOp creates a MultiFileReader, which is able to
read data multi-threaded from multiple files.
)DOC");
}
......
paddle/fluid/operators/spp_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/spp_op.h"
#include <string>
#include <vector>
namespace paddle {
namespace operators {
......
paddle/fluid/operators/spp_op.h
浏览文件 @ ae5923e2
......@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/pooling.h"
......
paddle/fluid/operators/sum_op.cc
浏览文件 @ ae5923e2
......@@ -10,6 +10,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/sum_op.h"
#include <algorithm>
#include <string>
#include <vector>
#include "paddle/fluid/framework/var_type_inference.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
......
paddle/fluid/operators/sum_op.h
浏览文件 @ ae5923e2
......@@ -10,6 +10,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
......
paddle/fluid/operators/top_k_op.h
浏览文件 @ ae5923e2
......@@ -15,6 +15,8 @@ limitations under the License. */
#pragma once
#include <algorithm>
#include <iostream>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
......
paddle/fluid/operators/transpose_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/transpose_op.h"
#include <vector>
namespace paddle {
namespace operators {
......
paddle/fluid/operators/transpose_op.h
浏览文件 @ ae5923e2
......@@ -14,6 +14,7 @@ limitations under the License. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
......
paddle/fluid/operators/unpool_op.cc
浏览文件 @ ae5923e2
......@@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/unpool_op.h"
#include <string>
#include <vector>
namespace paddle {
namespace operators {
......
paddle/fluid/operators/unpool_op.h
浏览文件 @ ae5923e2
......@@ -14,6 +14,8 @@ limitations under the License. */
#pragma once
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/unpooling.h"
......
paddle/fluid/operators/warpctc_op.h
浏览文件 @ ae5923e2
......@@ -14,6 +14,7 @@ limitations under the License. */
#pragma once
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/sequence_padding.h"
......
paddle/fluid/platform/cuda_helper.h
浏览文件 @ ae5923e2
......@@ -33,22 +33,26 @@ constexpr int PADDLE_CUDA_NUM_THREADS = 512;
USE_CUDA_ATOMIC(Add, float);
USE_CUDA_ATOMIC(Add, int);
USE_CUDA_ATOMIC(Add, unsigned int);
USE_CUDA_ATOMIC(Add, unsigned long long int);
// CUDA API uses unsigned long long int, we cannot use uint64_t here.
// It because unsigned long long int is not necessarily uint64_t
USE_CUDA_ATOMIC(Add, unsigned long long int); // NOLINT
CUDA_ATOMIC_WRAPPER(Add, int64_t) {
static_assert(sizeof(int64_t) == sizeof(long long int),
// Here, we check long long int must be int64_t.
static_assert(sizeof(int64_t) == sizeof(long long int), // NOLINT
"long long should be int64");
return CudaAtomicAdd(reinterpret_cast<unsigned long long int*>(address),
static_cast<unsigned long long int>(val));
return CudaAtomicAdd(
reinterpret_cast<unsigned long long int*>(address), // NOLINT
static_cast<unsigned long long int>(val)); // NOLINT
}
#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600
USE_CUDA_ATOMIC(Add, double);
#else
CUDA_ATOMIC_WRAPPER(Add, double) {
unsigned long long int* address_as_ull =
reinterpret_cast<unsigned long long int*>(address);
unsigned long long int old = *address_as_ull, assumed;
unsigned long long int* address_as_ull = // NOLINT
reinterpret_cast<unsigned long long int*>(address); // NOLINT
unsigned long long int old = *address_as_ull, assumed; // NOLINT
do {
assumed = old;
......@@ -62,53 +66,5 @@ CUDA_ATOMIC_WRAPPER(Add, double) {
}
#endif
// __shfl_down has been deprecated as of CUDA 9.0.
#if CUDA_VERSION < 9000
template <typename T>
__forceinline__ __device__ T __shfl_down_sync(unsigned, T val, int delta) {
return __shfl_down(val, delta);
}
#define CREATE_SHFL_MASK(mask, predicate) mask = 0u;
#else
#define FULL_WARP_MASK 0xFFFFFFFF
#define CREATE_SHFL_MASK(mask, predicate) \
mask = __ballot_sync(FULL_WARP_MASK, (predicate))
#endif
template <typename T>
__device__ T reduceSum(T val, int tid, int len) {
// TODO(zcd): The warp size should be taken from the
// parameters of the GPU but not specified as 32 simply.
// To make the reduceSum more efficiently,
// I use Warp-Level Parallelism and assume the Warp size
// is 32 which may be different for different GPU,
// but most card's warp size is 32.
__shared__ T shm[32];
const int warpSize = 32;
unsigned mask = 0u;
CREATE_SHFL_MASK(mask, tid < len);
for (int offset = warpSize / 2; offset > 0; offset /= 2)
val += __shfl_down_sync(mask, val, offset);
if (tid < warpSize) shm[tid] = 0;
__syncthreads();
if (tid % warpSize == 0) {
shm[tid / warpSize] = val;
}
CREATE_SHFL_MASK(mask, tid < warpSize);
if (tid < warpSize) {
val = shm[tid];
for (int offset = warpSize / 2; offset > 0; offset /= 2)
val += __shfl_down_sync(mask, val, offset);
}
return val;
}
} // namespace platform
} // namespace paddle
paddle/fluid/platform/device_context.cc
浏览文件 @ ae5923e2
......@@ -8,10 +8,14 @@ distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/platform/device_context.h"
#include <string>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/memory/memory.h"
namespace paddle {
namespace platform {
......
paddle/fluid/platform/device_context.h
浏览文件 @ ae5923e2
......@@ -8,13 +8,13 @@ distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <memory>
#include <string>
#include <unordered_map>
#include <vector>
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/dynload/cublas.h"
#include "paddle/fluid/platform/dynload/cudnn.h"
......
paddle/fluid/platform/device_context_test.cu
浏览文件 @ ae5923e2
......@@ -11,11 +11,12 @@ 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 "gtest/gtest.h"
#include "paddle/fluid/platform/device_context.h"
#include <vector>
#include "glog/logging.h"
#include "gtest/gtest.h"
TEST(Device, Init) {
using paddle::platform::DeviceContext;
......
paddle/fluid/platform/device_tracer.cc
浏览文件 @ ae5923e2
......@@ -11,15 +11,19 @@ distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/platform/device_tracer.h"
#include <google/protobuf/text_format.h>
#include <deque>
#include <fstream>
#include <map>
#include <mutex>
#include <mutex> // NOLINT
#include <numeric>
#include <thread>
#include <string>
#include <thread> // NOLINT
#include <vector>
#include "glog/logging.h"
#include "google/protobuf/text_format.h"
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/string/printf.h"
......@@ -123,7 +127,7 @@ void DisableActivity() {
void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size,
size_t *maxNumRecords) {
uint8_t *buf = (uint8_t *)malloc(kBufSize + kAlignSize);
uint8_t *buf = reinterpret_cast<uint8_t *>(malloc(kBufSize + kAlignSize));
*size = kBufSize;
*buffer = ALIGN_BUFFER(buf, kAlignSize);
*maxNumRecords = 0;
......
paddle/fluid/platform/device_tracer.h
浏览文件 @ ae5923e2
......@@ -11,8 +11,10 @@ distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/platform/dynload/cupti.h"
#include "paddle/fluid/platform/profiler.pb.h"
......
paddle/fluid/platform/dynload/cublas.cc
浏览文件 @ ae5923e2
......@@ -28,6 +28,10 @@ CUBLAS_BLAS_ROUTINE_EACH(DEFINE_WRAP);
CUBLAS_BLAS_ROUTINE_EACH_R2(DEFINE_WRAP);
#endif
#ifdef CUBLAS_BLAS_ROUTINE_EACH_R3
CUBLAS_BLAS_ROUTINE_EACH_R3(DEFINE_WRAP);
#endif
} // namespace dynload
} // namespace platform
} // namespace paddle
paddle/fluid/platform/dynload/cublas.h
浏览文件 @ ae5923e2
......@@ -71,7 +71,6 @@ extern void *cublas_dso_handle;
__macro(cublasDgemm_v2); \
__macro(cublasHgemm); \
__macro(cublasSgemmEx); \
__macro(cublasGemmEx); \
__macro(cublasSgeam_v2); \
__macro(cublasDgeam_v2); \
__macro(cublasCreate_v2); \
......@@ -83,11 +82,6 @@ extern void *cublas_dso_handle;
__macro(cublasDgemmBatched); \
__macro(cublasCgemmBatched); \
__macro(cublasZgemmBatched); \
__macro(cublasSgemmStridedBatched); \
__macro(cublasDgemmStridedBatched); \
__macro(cublasCgemmStridedBatched); \
__macro(cublasZgemmStridedBatched); \
__macro(cublasHgemmStridedBatched); \
__macro(cublasSgetrfBatched); \
__macro(cublasSgetriBatched); \
__macro(cublasDgetrfBatched); \
......@@ -95,10 +89,24 @@ extern void *cublas_dso_handle;
CUBLAS_BLAS_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
// APIs available after CUDA 8.0
#if CUDA_VERSION >= 8000
#define CUBLAS_BLAS_ROUTINE_EACH_R2(__macro) \
__macro(cublasGemmEx); \
__macro(cublasSgemmStridedBatched); \
__macro(cublasDgemmStridedBatched); \
__macro(cublasCgemmStridedBatched); \
__macro(cublasZgemmStridedBatched); \
__macro(cublasHgemmStridedBatched);
CUBLAS_BLAS_ROUTINE_EACH_R2(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
#endif
// APIs available after CUDA 9.0
#if CUDA_VERSION >= 9000
#define CUBLAS_BLAS_ROUTINE_EACH_R2(__macro) __macro(cublasSetMathMode);
CUBLAS_BLAS_ROUTINE_EACH_R2(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
#define CUBLAS_BLAS_ROUTINE_EACH_R3(__macro) __macro(cublasSetMathMode);
CUBLAS_BLAS_ROUTINE_EACH_R3(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
#endif
#undef DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
......
paddle/fluid/platform/float16.h
浏览文件 @ ae5923e2
......@@ -1003,6 +1003,46 @@ HOSTDEVICE inline float16 exp(const float16& a) {
return float16(::expf(static_cast<float>(a)));
}
template <>
HOSTDEVICE inline float16 log(const float16& a) {
return float16(::logf(static_cast<float>(a)));
}
template <>
HOSTDEVICE inline float16 tanh(const float16& a) {
return float16(::tanhf(static_cast<float>(a)));
}
template <>
HOSTDEVICE inline float16 sqrt(const float16& a) {
return float16(::sqrtf(static_cast<float>(a)));
}
template <>
HOSTDEVICE inline float16 ceil(const float16& a) {
return float16(::ceilf(static_cast<float>(a)));
}
template <>
HOSTDEVICE inline float16 floor(const float16& a) {
return float16(::floorf(static_cast<float>(a)));
}
template <>
HOSTDEVICE inline float16 round(const float16& a) {
return float16(::roundf(static_cast<float>(a)));
}
template <>
HOSTDEVICE inline float16 pow(const float16& a, const float16& b) {
return float16(::powf(static_cast<float>(a), static_cast<float>(b)));
}
template <>
HOSTDEVICE inline float16 abs(const float16& a) {
return float16(::fabs(static_cast<float>(a)));
}
} // namespace numext
} // namespace Eigen
paddle/fluid/platform/mkldnn_helper.h
浏览文件 @ ae5923e2
......@@ -11,11 +11,11 @@ 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 <mkldnn.hpp>
#include <vector>
#include "mkldnn/include/mkldnn.hpp"
#include "paddle/fluid/framework/operator.h"
namespace paddle {
......
paddle/fluid/platform/nccl_helper.h
浏览文件 @ ae5923e2
......@@ -14,8 +14,9 @@
#pragma once
#include <thread>
#include <thread> // NOLINT
#include <typeindex>
#include <vector>
#include "paddle/fluid/platform/dynload/nccl.h"
#include "paddle/fluid/platform/enforce.h"
......@@ -29,6 +30,8 @@ inline ncclDataType_t ToNCCLDataType(std::type_index type) {
return ncclDouble;
} else if (type == typeid(int)) { // NOLINT
return ncclInt;
} else if (type == typeid(int64_t)) { // NOLINT
return ncclInt64;
} else {
PADDLE_THROW("Not supported");
}
......@@ -58,7 +61,7 @@ struct NCCLContext {
ncclComm_t comm_;
explicit NCCLContext(int dev_id)
: ctx_(new CUDADeviceContext(CUDAPlace(dev_id))) {}
: ctx_(new CUDADeviceContext(CUDAPlace(dev_id))), comm_{nullptr} {}
cudaStream_t stream() const { return ctx_->stream(); }
......@@ -66,23 +69,23 @@ struct NCCLContext {
return boost::get<platform::CUDAPlace>(ctx_->GetPlace()).device;
}
static void InitNCCLContext(std::unordered_map<int, NCCLContext> &contexts,
static void InitNCCLContext(std::unordered_map<int, NCCLContext> *contexts,
const std::vector<platform::Place> &places) {
std::vector<ncclComm_t> comms;
std::vector<int> devs;
comms.resize(contexts.size());
devs.reserve(contexts.size());
comms.resize(contexts->size());
devs.reserve(contexts->size());
for (auto &p : places) {
devs.push_back(boost::get<platform::CUDAPlace>(p).device);
}
PADDLE_ENFORCE(platform::dynload::ncclCommInitAll(
&comms[0], static_cast<int>(contexts.size()), &devs[0]));
&comms[0], static_cast<int>(contexts->size()), &devs[0]));
int i = 0;
for (auto &dev_id : devs) {
contexts.at(dev_id).comm_ = comms[i++];
contexts->at(dev_id).comm_ = comms[i++];
}
}
};
......@@ -91,7 +94,8 @@ struct NCCLContextMap {
std::unordered_map<int, NCCLContext> contexts_;
std::vector<int> order_;
NCCLContextMap(const std::vector<platform::Place> &places) {
explicit NCCLContextMap(const std::vector<platform::Place> &places) {
PADDLE_ENFORCE(!places.empty());
order_.reserve(places.size());
for (auto &p : places) {
int dev_id = boost::get<CUDAPlace>(p).device;
......@@ -102,15 +106,17 @@ struct NCCLContextMap {
order_.size(), contexts_.size(),
"NCCL Context Map does not support contain two or more same device");
std::vector<ncclComm_t> comms;
comms.resize(order_.size());
if (places.size() > 1) {
std::vector<ncclComm_t> comms;
comms.resize(order_.size());
PADDLE_ENFORCE(platform::dynload::ncclCommInitAll(
&comms[0], static_cast<int>(order_.size()), &order_[0]));
PADDLE_ENFORCE(platform::dynload::ncclCommInitAll(
&comms[0], static_cast<int>(order_.size()), &order_[0]));
int i = 0;
for (auto &dev_id : order_) {
contexts_.at(dev_id).comm_ = comms[i++];
int i = 0;
for (auto &dev_id : order_) {
contexts_.at(dev_id).comm_ = comms[i++];
}
}
}
......
paddle/fluid/platform/profiler.cc
浏览文件 @ ae5923e2
此差异已折叠。
paddle/fluid/platform/profiler.h
浏览文件 @ ae5923e2
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ ae5923e2
此差异已折叠。
paddle/fluid/pybind/recordio.cc
浏览文件 @ ae5923e2
此差异已折叠。
paddle/testing/paddle_gtest_main.cc
浏览文件 @ ae5923e2
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ ae5923e2
此差异已折叠。
python/paddle/fluid/average.py
浏览文件 @ ae5923e2
此差异已折叠。
python/paddle/fluid/evaluator.py
浏览文件 @ ae5923e2
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ ae5923e2
此差异已折叠。
python/paddle/fluid/initializer.py
浏览文件 @ ae5923e2
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ ae5923e2
此差异已折叠。
python/paddle/fluid/metrics.py 0 → 100644
浏览文件 @ ae5923e2
此差异已折叠。
python/setup.py.in
浏览文件 @ ae5923e2
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