Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 332
Star 4
Fork 1
P
Paddle-Lite
未验证 提交 c67b701f 编写于 作者: H haohan 提交者： GitHub
浏览文件
操作

Merge pull request #1 from PaddlePaddle/develop 

Merge newest version
上级 0474c028 fb8a569e
隐藏空白更改
内联 并排
Showing with 6927 addition and 1794 deletion
.gitignore
浏览文件 @ c67b701f
......@@ -69,3 +69,10 @@ build
# clion building directories
cmake-build-debug
cmake-build-release
#ios demo
demo/ios/PaddleMobileDemo/PaddleMobileDemo/googlenet_combine/
demo/ios/PaddleMobileDemo/PaddleMobileDemo/*.jpg
demo/ios/PaddleMobileDemo/PaddleMobileDemo/PaddleMobile/*.a
*.xcuserstate
\ No newline at end of file
CMakeLists.txt
浏览文件 @ c67b701f
......@@ -6,12 +6,25 @@ option(USE_OPENMP "openmp support" OFF)
option(USE_EXCEPTION "use std exception" ON)
option(LOG_PROFILE "log profile" ON)
# select the platform to build
option(CPU "cpu" OFF)
option(MALI_GPU "mali gpu" ON)
option(CPU "armv7 with neon" ON)
option(MALI_GPU "mali gpu" OFF)
option(FPGA "fpga" OFF)
set(DEBUGING ON)
if (ARM_LINUX)
include("${CMAKE_CURRENT_LIST_DIR}/tools/arm-platform.cmake")
endif ()
file(GLOB_RECURSE PADDLE_MOBILE_CC src/*.cc src/*.cpp src/*.c src/*.mm)
file(GLOB_RECURSE PADDLE_MOBILE_H src/*.h)
if (CPU)
add_definitions(-DPADDLE_MOBILE_CPU)
add_definitions(-DPADDLE_MOBILE_CPU)
else()
list(REMOVE_ITEM PADDLE_MOBILE_CC ./src/operators/kernel/arm/*.h)
list(REMOVE_ITEM PADDLE_MOBILE_CC ./src/operators/kernel/arm/*.cc)
list(REMOVE_ITEM PADDLE_MOBILE_CC ./src/operators/kernel/arm/*.cpp)
endif()
if (MALI_GPU)
......@@ -27,32 +40,42 @@ if (MALI_GPU)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -L${ACL_ROOT}/build/opencl-1.2-stubs")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -lOpenCL")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_ACL=1")
else()
file(GLOB_RECURSE _tmp_list src/operators/kernel/mali/*.cpp src/operators/kernel/mali/*.cc)
foreach(f ${_tmp_list})
list(REMOVE_ITEM PADDLE_MOBILE_CC ${f})
endforeach()
file(GLOB_RECURSE _tmp_list_h src/operators/kernel/mali/*.h)
foreach(f ${_tmp_list_h})
list(REMOVE_ITEM PADDLE_MOBILE_H ${f})
endforeach()
endif()
if(FPGA)
add_definitions(-DPADDLE_MOBILE_FPGA)
else()
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/operators/kernel/fpga/*.h)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/operators/kernel/fpga/*.cc)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/operators/kernel/fpga/*.cpp)
endif()
set(CMAKE_CXX_FLAGS "-std=c++14 -O3 -s ${CMAKE_CXX_FLAGS}")
if (DEBUGING)
set(CMAKE_BUILD_TYPE Debug)
set(CMAKE_CXX_FLAGS_DEBUG "-O3 -DNDEBUG")
else()
set(CMAKE_BUILD_TYPE Release)
set(CMAKE_CXX_FLAGS_RELEASE "-O3 -DNDEBUG")
endif ()
if(DEBUGING)
message(STATUS "debuging")
add_definitions(-DPADDLE_MOBILE_DEBUG)
if(ANDROID)
message(STATUS "debug")
set(CMAKE_BUILD_TYPE Debug)
set(CMAKE_CXX_FLAGS_DEBUG "-g -DNDEBUG")
add_definitions(-DPADDLE_MOBILE_DEBUG)
if (ANDROID_NDK_TOOLCHAIN_INCLUDED)
add_definitions(-DARMV7)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -llog")
endif()
else()
message(STATUS "releasing")
add_definitions(-fvisibility=hidden -fvisibility-inlines-hidden)
endif()
endif ()
else ()
set(CMAKE_BUILD_TYPE Release)
set(CMAKE_CXX_FLAGS_RELEASE "-DNDEBUG")
add_definitions(-fvisibility=hidden -fvisibility-inlines-hidden)
endif ()
if (USE_EXCEPTION)
message(STATUS "use exception")
......@@ -66,110 +89,47 @@ if (LOG_PROFILE)
add_definitions(-DPADDLE_MOBILE_PROFILE)
endif()
if(IS_MAC)
add_definitions(-DX86)
elseif(IS_IOS)
add_definitions(-DIOS)
elseif(V7)
add_definitions(-DARMV7)
elseif(V8)
add_definitions(-DARMV8)
else ()
add_definitions(-DX86)
if(USE_OPENMP)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp")
add_definitions(-DPADDLE_MOBILE_USE_OPENMP)
endif()
set(CMAKE_VERBOSE_MAKEFILE ON)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY build)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY build)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY build)
file(GLOB_RECURSE PADDLE_MOBILE_CC src/*.cc src/*.cpp src/*.c)
file(GLOB_RECURSE PADDLE_MOBILE_H src/*.h)
if (NOT ANDROID)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/jni/*.cpp)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/jni/*.h)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/operators/math/math_func_neon.h)
if (NOT ANDROID_NDK_TOOLCHAIN_INCLUDED)
list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/jni/paddle_mobile_jni.h)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/jni/paddle_mobile_jni.cpp)
list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/operators/math/math_func_neon.h)
endif ()
include_directories(src/)
if(USE_OPENMP)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp")
add_definitions(-DPADDLE_MOBILE_USE_OPENMP)
endif()
set(CMAKE_VERBOSE_MAKEFILE ON)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY build)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY build)
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY build)
if (googlenet)
add_definitions(-DCONCAT_OP)
add_definitions(-DCONV_OP)
add_definitions(-DLRN_OP)
add_definitions(-DMUL_OP)
add_definitions(-DELEMENTWISEADD_OP)
add_definitions(-DFUSION_FC_OP)
add_definitions(-DPOOL_OP)
add_definitions(-DRELU_OP)
add_definitions(-DFUSION_CONVADD_OP)
add_definitions(-DFUSION_CONVADD_RELU_OP)
elseif (mobilenet)
add_definitions(-DCONV_OP)
add_definitions(-DELEMENTWISEADD_OP)
add_definitions(-DRELU_OP)
add_definitions(-DSOFTMAX_OP)
add_definitions(-DSOFTMAX_OP)
add_definitions(-DDEPTHWISECONV_OP)
add_definitions(-DBATCHNORM_OP)
add_definitions(-DPOOL_OP)
add_definitions(-DRESHAPE_OP)
elseif (yolo)
add_definitions(-DBATCHNORM_OP)
add_definitions(-DCONV_OP)
add_definitions(-DRELU_OP)
add_definitions(-DELEMENTWISEADD_OP)
elseif (squeezenet)
add_definitions(-DCONCAT_OP)
add_definitions(-DCONV_OP)
add_definitions(-DRELU_OP)
add_definitions(-DELEMENTWISEADD_OP)
add_definitions(-DPOOL_OP)
add_definitions(-DRESHAPE_OP)
add_definitions(-DSOFTMAX_OP)
elseif(resnet)
add_definitions(-DCONV_OP)
add_definitions(-DBATCHNORM_OP)
add_definitions(-DELEMENTWISEADD_OP)
add_definitions(-DSOFTMAX_OP)
add_definitions(-DMUL_OP)
add_definitions(-DPOOL_OP)
add_definitions(-DRELU_OP)
else ()
add_definitions(-DBATCHNORM_OP)
add_definitions(-DBOXCODER_OP)
add_definitions(-DCONCAT_OP)
add_definitions(-DCONV_OP)
add_definitions(-DDEPTHWISECONV_OP)
add_definitions(-DELEMENTWISEADD_OP)
add_definitions(-DFUSION_CONVADD_OP)
add_definitions(-DCONVADDRELU_OP)
add_definitions(-DFUSION_FC_OP)
add_definitions(-DLRN_OP)
add_definitions(-DMUL_OP)
add_definitions(-DMULTICLASSNMS_OP)
add_definitions(-DPOOL_OP)
add_definitions(-DPRIORBOX_OP)
add_definitions(-DRELU_OP)
add_definitions(-DRESHAPE_OP)
add_definitions(-DSIGMOID_OP)
add_definitions(-DSOFTMAX_OP)
add_definitions(-DTRANSPOSE_OP)
add_definitions(-DFUSION_CONVADD_RELU_OP)
endif()
# NET default
set(NET "defult" CACHE STRING "select net type")
set_property(CACHE NET PROPERTY STRINGS "defult" "googlenet" "mobilenet" "yolo" "squeezenet")
include("${CMAKE_CURRENT_LIST_DIR}/tools/op.cmake")
if (IS_IOS)
add_library(paddle-mobile STATIC ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
elseif(ANDROID)
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
else()
list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/PaddleMobile.h)
list(REMOVE_ITEM PADDLE_MOBILE_CC ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/PaddleMobile.mm)
list(REMOVE_ITEM PADDLE_MOBILE_H ${CMAKE_CURRENT_SOURCE_DIR}/src/ios_io/op_symbols.h)
endif ()
if (ANDROID_NDK_TOOLCHAIN_INCLUDED)
list(REMOVE_DUPLICATES CMAKE_CXX_FLAGS)
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
elseif(IS_IOS)
else ()
add_library(paddle-mobile SHARED ${PADDLE_MOBILE_CC} ${PADDLE_MOBILE_H})
endif ()
......@@ -177,4 +137,3 @@ if(DEBUGING)
add_subdirectory(test)
endif()
CONTRIBUTING.md
浏览文件 @ c67b701f
......@@ -183,6 +183,9 @@ upstream
接下来等待 review，如果有需要修改的地方，参照上述步骤更新 origin 中的对应分支即可。
![](http://otkwwi4x8.bkt.clouddn.com/2018-06-20-15294877166787.jpg)
之后就可以提交代码了
## 删除远程分支
在 PR 被 merge 进主仓库后，我们可以在 PR 的页面删除远程仓库的分支。
......@@ -219,7 +222,7 @@ upstream
- 原因：如果仅仅修改一个文件但提交了十几个commit，每个commit只做了少量的修改，这会给评审人带来很大困扰。评审人需要逐一查看每个commit才能知道做了哪些修改，且不排除commit之间的修改存在相互覆盖的情况。
- 建议：每次提交时，保持尽量少的commit，可以通过`git commit --amend`补充上次的commit。对已经Push到远程仓库的多个commit，可以参考[squash commits after push](http://stackoverflow.com/questions/5667884/how-to-squash-commits-in-git-after-they-have-been-pushed)。
- 请注意每个commit的名称：应能反映当前commit的内容，不能太随意。
3. 如果解决了某个Issue的问题，请在该Pull Request的**第一个**评论框中加上：`fix #issue_number`，这样当该PUll Request被合并后，会自动关闭对应的Issue。关键词包括：close, closes, closed, fix, fixes, fixed, resolve, resolves, resolved，请选择合适的词汇。详细可参考[Closing issues via commit messages](https://help.github.com/articles/closing-issues-via-commit-messages)。
3. 如果解决了某个Issue的问题，请在该Pull Request的**第一个**评论框中加上：`fix #issue_number`，这样当该Pull Request被合并后，会自动关闭对应的Issue。关键词包括：close, closes, closed, fix, fixes, fixed, resolve, resolves, resolved，请选择合适的词汇。详细可参考[Closing issues via commit messages](https://help.github.com/articles/closing-issues-via-commit-messages)。
此外，在回复评审人意见时，请您遵守以下约定：
......
Dockerfile
浏览文件 @ c67b701f
FROM ubuntu:18.04
FROM ubuntu:16.04
RUN echo '\
deb <mirror> bionic main restricted universe multiverse\n\
deb <mirror> bionic-updates main restricted universe multiverse\n\
deb <mirror> bionic-backports main restricted universe multiverse\n\
deb <mirror> bionic-security main restricted universe multiverse\n'\
deb <mirror> <version> main restricted universe multiverse\n\
deb <mirror> <version>-updates main restricted universe multiverse\n\
deb <mirror> <version>-backports main restricted universe multiverse\n\
deb <mirror> <version>-security main restricted universe multiverse\n'\
> /etc/apt/sources.list
RUN sed -ie 's|<mirror>|http://mirrors.tuna.tsinghua.edu.cn/ubuntu/|' /etc/apt/sources.list
RUN sed -ie 's|<version>|xenial|' /etc/apt/sources.list
RUN apt-get update && apt-get upgrade -y
RUN apt-get install -y --no-install-recommends \
curl \
unzip \
git \
cmake \
make \
cmake-curses-gui \
python \
python-pip \
python-setuptools \
clang-format-5.0 \
graphviz
graphviz \
g++-arm-linux-gnueabi \
gcc-arm-linux-gnueabi
RUN apt-get autoremove -y && apt-get clean
RUN pip install wheel pre-commit
RUN pre-commit autoupdate
RUN ln -s clang-format-5.0 /usr/bin/clang-format
RUN pip install -i https://pypi.tuna.tsinghua.edu.cn/simple --upgrade pip
RUN pip install -i https://pypi.tuna.tsinghua.edu.cn/simple wheel
RUN pip install -i https://pypi.tuna.tsinghua.edu.cn/simple pre-commit
RUN cd /tmp && curl -O http://mirrors.neusoft.edu.cn/android/repository/android-ndk-r17b-linux-x86_64.zip
RUN cd /opt && unzip /tmp/android-ndk-r17b-linux-x86_64.zip
\ No newline at end of file
RUN curl -O https://mms-res.cdn.bcebos.com/cmake-3.10.3-Linux-x86_64.tar.gz && \
tar xzf cmake-3.10.3-Linux-x86_64.tar.gz && \
mv cmake-3.10.3-Linux-x86_64 /opt/cmake-3.10 && \
mv /usr/bin/cmake /usr/bin/cmake.bak && ln -s /opt/cmake-3.10/bin/cmake /usr/bin/cmake
RUN cd /opt && unzip /tmp/android-ndk-r17b-linux-x86_64.zip
ENV NDK_ROOT /opt/android-ndk-r17b
README.md
浏览文件 @ c67b701f
# Paddle-Mobile
# Paddle-Mobile
[![Build Status](https://travis-ci.org/PaddlePaddle/paddle-mobile.svg?branch=develop&longCache=true&style=flat-square)](https://travis-ci.org/PaddlePaddle/paddle-mobile)
[![License](https://img.shields.io/badge/license-Apache%202-brightgreen.svg)](LICENSE)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](https://github.com/PaddlePaddle/paddle-mobile/tree/develop/doc)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
<!--[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle-Mobile.svg)](https://github.com/PaddlePaddle/Paddle-Mobile/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)-->
欢迎来到 Paddle-Mobile GitHub 项目。
Paddle-Moible是PaddlePaddle组织下的项目，是一个致力于嵌入式平台的深度学习的框架。Paddle-Moible设计思想和PaddlePaddle的最新版fluid版本保持了高度一致，同时针对嵌入式做了大量优化。设计之初就对嵌入式的性能、体积、能耗、硬件平台覆盖等方面做了考虑。
## Features
- **ARM CPU**
arm cpu是paddle-mobile的主要支持方向，cpu的通用性一直是其优势。嵌入式深度学习，需要大量的cpu汇编实现。我们正在紧锣密鼓的编码，为的是能充分硬件的每一点加速能力。
arm cpu的优化工作还在进行中，现在使用了常规的cpu优化。在arm a73上paddle-mobile现在单核运行一次mobilenet 1.0是160+ms，显然这不是我们的最终目标，我们正在用大量的汇编改写，后续性能仍会有巨大提升空间。
- **Mali GPU**
Mali GPU是百度和ARM合作开发的，双方团队近期都在致力于将paddle的op能无缝运行在ACL(arm compute library)。目前已经支持squeezenet，googlenet，resnet等几个网络模型，后续会继续加大力度。使全部移动端paddle op能高效运行在mali gpu上。
- **苹果设备的GPU Metal实现**
基于Metal实现的苹果设备的GPU预测库，也已经在实现中，近期也会有相应可运行版本。
- **FPGA**
FPGA实现正在进行中，是基于Xilinx的ZU5目标开发板。
- **灵活性**
* paddle-mobile cpu版不依赖任何第三库, 可进行快速集成。
* 使用泛型特化进行平台切换, 可灵活切换 cpu、gpu 和其他协处理器。
* 可根据特定的常见网络, 进行编译特定的 op, 降低编译时间, 减小包大小。
* 使用 docker 编译, 提供统一的编译环境。
* 高可拓展性, 方便拓展其他协处理器, 提供高性能 arm 算子实现, 方便其他协处理器开发者集成开发。
* 直接兼容 paddle-fluid 模型, 不需要额外的转换操作。
- **体积**
paddle-mobile从设计之初就深入考虑到移动端的包体积的问题，cpu实现中没有外部依赖。在编译过程中，如果该网络不需要的op是完全不会被打入的。同时编译选项优化也为体积压缩提供了帮助。
除了二进制体积，我们对代码体积极力避免过大。整个仓库的代码体积也非常小。
## 文档
### 设计文档
关于paddle-mobile设计文档在下面链接中，如果想了解更多内容。[issue](https://github.com/PaddlePaddle/paddle-mobile/issues)中会有很多早期的设计和讨论过程。
[设计文档链接](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/design_doc.md)
### 开发文档
开发文档主要是关于编译、运行等问题。做为开发者，它可以和贡献文档共同结合使用。
[开发文档]()https://github.com/PaddlePaddle/paddle-mobile/blob/develop/doc/development_doc.md
### 贡献文档
- [贡献文档链接](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/CONTRIBUTING.md)
- 上面文档中涵盖了主要的贡献代码流程，如果在实践中您还遇到了其他问题，可以发[issue](https://github.com/PaddlePaddle/paddle-mobile/issues)。我们看到后会尽快处理。
This project is used to develop the next version deep learning freamwork for mobile device.
# Development
## 模型获得
目前Paddle-Mobile仅支持Paddle fluid训练的模型。如果你手中的模型是不同种类的模型，需要进行模型转换才可以运行。
### 1. 直接使用Paddle Fluid训练
该方式最为可靠，推荐方式
### 2. caffe转为Paddle Fluid模型
[链接](https://github.com/PaddlePaddle/models/tree/develop/fluid/image_classification/caffe2fluid)
### 3. ONNX
ONNX全称为“Open Neural Network Exchange”，即“开放的神经网络切换”。该项目的目的是让不同的神经网络开发框架做到互通互用。
[Used model in development](https://mms-mis.cdn.bcebos.com/paddle-mobile/models.zip)
除直接使用PaddlePaddle训练fluid版本的模型外，还可以通过onnx转换得到个别Paddle fluid模型。
## cross-compilation to android
目前，百度也在做onnx支持工作。相关转换项目在这里：[paddle-onnx](https://github.com/PaddlePaddle/paddle-onnx)。
* NDK is required
* ANDROID_NDK environment variable is required
![](http://otkwwi4x8.bkt.clouddn.com/2018-07-03-15305875853057.jpg)
```bash
sh build.sh android
```
### 4. 部分测试模型下载
[下载链接](https://mms-mis.cdn.bcebos.com/paddle-mobile/models.zip)
## build for x86
paddle-mobile is to run on arm platform. x86 only used to test not arm assembly code. So do not recommend compiling x86.
## 问题解决
Now only support osx.
欢迎提出或解决我们的问题，有疑问可以发issue. [Github Issues](https://github.com/PaddlePaddle/paddle-mobile/issues).
```
sh build.sh mac
```
## Copyright and License
Paddle-Mobile 提供相对宽松的Apache-2.0开源协议 [Apache-2.0 license](LICENSE).
## Old Version of Mobile-Deep-Learning
The old version of MDL was I moved to here [Mobile-Deep-Learning](https://github.com/allonli/mobile-deep-learning)
## 旧版 Mobile-Deep-Learning
原MDL(Mobile-Deep-Learning)工程被迁移到了这里 [Mobile-Deep-Learning](https://github.com/allonli/mobile-deep-learning)
demo/ios/PaddleMobileDemo/PaddleMobileDemo.xcodeproj/project.pbxproj 0 → 100644
浏览文件 @ c67b701f
// !$*UTF8*$!
{
archiveVersion = 1;
classes = {
};
objectVersion = 50;
objects = {
/* Begin PBXBuildFile section */
FC12E93320EB6B2800807EF4 /* AppDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = FC12E93220EB6B2800807EF4 /* AppDelegate.m */; };
FC12E93620EB6B2800807EF4 /* ViewController.m in Sources */ = {isa = PBXBuildFile; fileRef = FC12E93520EB6B2800807EF4 /* ViewController.m */; };
FC12E93920EB6B2800807EF4 /* Main.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = FC12E93720EB6B2800807EF4 /* Main.storyboard */; };
FC12E93B20EB6B2900807EF4 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = FC12E93A20EB6B2900807EF4 /* Assets.xcassets */; };
FC12E93E20EB6B2900807EF4 /* LaunchScreen.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = FC12E93C20EB6B2900807EF4 /* LaunchScreen.storyboard */; };
FC12E94120EB6B2900807EF4 /* main.m in Sources */ = {isa = PBXBuildFile; fileRef = FC12E94020EB6B2900807EF4 /* main.m */; };
FC12E94A20EB6B6800807EF4 /* libpaddle-mobile.a in Frameworks */ = {isa = PBXBuildFile; fileRef = FC12E94820EB6B6800807EF4 /* libpaddle-mobile.a */; };
FC12E94D20EB6BBB00807EF4 /* libstdc++.tbd in Frameworks */ = {isa = PBXBuildFile; fileRef = FC12E94C20EB6BBB00807EF4 /* libstdc++.tbd */; };
FC12E95120EB6BED00807EF4 /* params in Resources */ = {isa = PBXBuildFile; fileRef = FC12E94F20EB6BED00807EF4 /* params */; };
FC12E95220EB6BED00807EF4 /* model in Resources */ = {isa = PBXBuildFile; fileRef = FC12E95020EB6BED00807EF4 /* model */; };
FC12E95420EB6C0D00807EF4 /* apple.jpg in Resources */ = {isa = PBXBuildFile; fileRef = FC12E95320EB6C0D00807EF4 /* apple.jpg */; };
/* End PBXBuildFile section */
/* Begin PBXFileReference section */
FC12E92E20EB6B2800807EF4 /* PaddleMobileDemo.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = PaddleMobileDemo.app; sourceTree = BUILT_PRODUCTS_DIR; };
FC12E93120EB6B2800807EF4 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
FC12E93220EB6B2800807EF4 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
FC12E93420EB6B2800807EF4 /* ViewController.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = ViewController.h; sourceTree = "<group>"; };
FC12E93520EB6B2800807EF4 /* ViewController.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = ViewController.m; sourceTree = "<group>"; };
FC12E93820EB6B2800807EF4 /* Base */ = {isa = PBXFileReference; lastKnownFileType = file.storyboard; name = Base; path = Base.lproj/Main.storyboard; sourceTree = "<group>"; };
FC12E93A20EB6B2900807EF4 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
FC12E93D20EB6B2900807EF4 /* Base */ = {isa = PBXFileReference; lastKnownFileType = file.storyboard; name = Base; path = Base.lproj/LaunchScreen.storyboard; sourceTree = "<group>"; };
FC12E93F20EB6B2900807EF4 /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
FC12E94020EB6B2900807EF4 /* main.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = main.m; sourceTree = "<group>"; };
FC12E94820EB6B6800807EF4 /* libpaddle-mobile.a */ = {isa = PBXFileReference; lastKnownFileType = archive.ar; path = "libpaddle-mobile.a"; sourceTree = "<group>"; };
FC12E94920EB6B6800807EF4 /* PaddleMobile.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = PaddleMobile.h; sourceTree = "<group>"; };
FC12E94C20EB6BBB00807EF4 /* libstdc++.tbd */ = {isa = PBXFileReference; lastKnownFileType = "sourcecode.text-based-dylib-definition"; name = "libstdc++.tbd"; path = "usr/lib/libstdc++.tbd"; sourceTree = SDKROOT; };
FC12E94F20EB6BED00807EF4 /* params */ = {isa = PBXFileReference; lastKnownFileType = file; path = params; sourceTree = "<group>"; };
FC12E95020EB6BED00807EF4 /* model */ = {isa = PBXFileReference; lastKnownFileType = file; path = model; sourceTree = "<group>"; };
FC12E95320EB6C0D00807EF4 /* apple.jpg */ = {isa = PBXFileReference; lastKnownFileType = image.jpeg; path = apple.jpg; sourceTree = "<group>"; };
/* End PBXFileReference section */
/* Begin PBXFrameworksBuildPhase section */
FC12E92B20EB6B2800807EF4 /* Frameworks */ = {
isa = PBXFrameworksBuildPhase;
buildActionMask = 2147483647;
files = (
FC12E94D20EB6BBB00807EF4 /* libstdc++.tbd in Frameworks */,
FC12E94A20EB6B6800807EF4 /* libpaddle-mobile.a in Frameworks */,
);
runOnlyForDeploymentPostprocessing = 0;
};
/* End PBXFrameworksBuildPhase section */
/* Begin PBXGroup section */
FC12E92520EB6B2800807EF4 = {
isa = PBXGroup;
children = (
FC12E93020EB6B2800807EF4 /* PaddleMobileDemo */,
FC12E92F20EB6B2800807EF4 /* Products */,
FC12E94B20EB6BBB00807EF4 /* Frameworks */,
);
sourceTree = "<group>";
};
FC12E92F20EB6B2800807EF4 /* Products */ = {
isa = PBXGroup;
children = (
FC12E92E20EB6B2800807EF4 /* PaddleMobileDemo.app */,
);
name = Products;
sourceTree = "<group>";
};
FC12E93020EB6B2800807EF4 /* PaddleMobileDemo */ = {
isa = PBXGroup;
children = (
FC12E95320EB6C0D00807EF4 /* apple.jpg */,
FC12E94E20EB6BED00807EF4 /* googlenet_combine */,
FC12E94720EB6B6800807EF4 /* PaddleMobile */,
FC12E93120EB6B2800807EF4 /* AppDelegate.h */,
FC12E93220EB6B2800807EF4 /* AppDelegate.m */,
FC12E93420EB6B2800807EF4 /* ViewController.h */,
FC12E93520EB6B2800807EF4 /* ViewController.m */,
FC12E93720EB6B2800807EF4 /* Main.storyboard */,
FC12E93A20EB6B2900807EF4 /* Assets.xcassets */,
FC12E93C20EB6B2900807EF4 /* LaunchScreen.storyboard */,
FC12E93F20EB6B2900807EF4 /* Info.plist */,
FC12E94020EB6B2900807EF4 /* main.m */,
);
path = PaddleMobileDemo;
sourceTree = "<group>";
};
FC12E94720EB6B6800807EF4 /* PaddleMobile */ = {
isa = PBXGroup;
children = (
FC12E94820EB6B6800807EF4 /* libpaddle-mobile.a */,
FC12E94920EB6B6800807EF4 /* PaddleMobile.h */,
);
path = PaddleMobile;
sourceTree = "<group>";
};
FC12E94B20EB6BBB00807EF4 /* Frameworks */ = {
isa = PBXGroup;
children = (
FC12E94C20EB6BBB00807EF4 /* libstdc++.tbd */,
);
name = Frameworks;
sourceTree = "<group>";
};
FC12E94E20EB6BED00807EF4 /* googlenet_combine */ = {
isa = PBXGroup;
children = (
FC12E94F20EB6BED00807EF4 /* params */,
FC12E95020EB6BED00807EF4 /* model */,
);
path = googlenet_combine;
sourceTree = "<group>";
};
/* End PBXGroup section */
/* Begin PBXNativeTarget section */
FC12E92D20EB6B2800807EF4 /* PaddleMobileDemo */ = {
isa = PBXNativeTarget;
buildConfigurationList = FC12E94420EB6B2900807EF4 /* Build configuration list for PBXNativeTarget "PaddleMobileDemo" */;
buildPhases = (
FC12E92A20EB6B2800807EF4 /* Sources */,
FC12E92B20EB6B2800807EF4 /* Frameworks */,
FC12E92C20EB6B2800807EF4 /* Resources */,
);
buildRules = (
);
dependencies = (
);
name = PaddleMobileDemo;
productName = PaddleMobileDemo;
productReference = FC12E92E20EB6B2800807EF4 /* PaddleMobileDemo.app */;
productType = "com.apple.product-type.application";
};
/* End PBXNativeTarget section */
/* Begin PBXProject section */
FC12E92620EB6B2800807EF4 /* Project object */ = {
isa = PBXProject;
attributes = {
LastUpgradeCheck = 0940;
ORGANIZATIONNAME = orange;
TargetAttributes = {
FC12E92D20EB6B2800807EF4 = {
CreatedOnToolsVersion = 9.4.1;
};
};
};
buildConfigurationList = FC12E92920EB6B2800807EF4 /* Build configuration list for PBXProject "PaddleMobileDemo" */;
compatibilityVersion = "Xcode 9.3";
developmentRegion = en;
hasScannedForEncodings = 0;
knownRegions = (
en,
Base,
);
mainGroup = FC12E92520EB6B2800807EF4;
productRefGroup = FC12E92F20EB6B2800807EF4 /* Products */;
projectDirPath = "";
projectRoot = "";
targets = (
FC12E92D20EB6B2800807EF4 /* PaddleMobileDemo */,
);
};
/* End PBXProject section */
/* Begin PBXResourcesBuildPhase section */
FC12E92C20EB6B2800807EF4 /* Resources */ = {
isa = PBXResourcesBuildPhase;
buildActionMask = 2147483647;
files = (
FC12E93E20EB6B2900807EF4 /* LaunchScreen.storyboard in Resources */,
FC12E95220EB6BED00807EF4 /* model in Resources */,
FC12E93B20EB6B2900807EF4 /* Assets.xcassets in Resources */,
FC12E95120EB6BED00807EF4 /* params in Resources */,
FC12E95420EB6C0D00807EF4 /* apple.jpg in Resources */,
FC12E93920EB6B2800807EF4 /* Main.storyboard in Resources */,
);
runOnlyForDeploymentPostprocessing = 0;
};
/* End PBXResourcesBuildPhase section */
/* Begin PBXSourcesBuildPhase section */
FC12E92A20EB6B2800807EF4 /* Sources */ = {
isa = PBXSourcesBuildPhase;
buildActionMask = 2147483647;
files = (
FC12E93620EB6B2800807EF4 /* ViewController.m in Sources */,
FC12E94120EB6B2900807EF4 /* main.m in Sources */,
FC12E93320EB6B2800807EF4 /* AppDelegate.m in Sources */,
);
runOnlyForDeploymentPostprocessing = 0;
};
/* End PBXSourcesBuildPhase section */
/* Begin PBXVariantGroup section */
FC12E93720EB6B2800807EF4 /* Main.storyboard */ = {
isa = PBXVariantGroup;
children = (
FC12E93820EB6B2800807EF4 /* Base */,
);
name = Main.storyboard;
sourceTree = "<group>";
};
FC12E93C20EB6B2900807EF4 /* LaunchScreen.storyboard */ = {
isa = PBXVariantGroup;
children = (
FC12E93D20EB6B2900807EF4 /* Base */,
);
name = LaunchScreen.storyboard;
sourceTree = "<group>";
};
/* End PBXVariantGroup section */
/* Begin XCBuildConfiguration section */
FC12E94220EB6B2900807EF4 /* Debug */ = {
isa = XCBuildConfiguration;
buildSettings = {
ALWAYS_SEARCH_USER_PATHS = NO;
CLANG_ANALYZER_NONNULL = YES;
CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
CLANG_CXX_LANGUAGE_STANDARD = "gnu++14";
CLANG_CXX_LIBRARY = "libc++";
CLANG_ENABLE_MODULES = YES;
CLANG_ENABLE_OBJC_ARC = YES;
CLANG_ENABLE_OBJC_WEAK = YES;
CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
CLANG_WARN_BOOL_CONVERSION = YES;
CLANG_WARN_COMMA = YES;
CLANG_WARN_CONSTANT_CONVERSION = YES;
CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
CLANG_WARN_EMPTY_BODY = YES;
CLANG_WARN_ENUM_CONVERSION = YES;
CLANG_WARN_INFINITE_RECURSION = YES;
CLANG_WARN_INT_CONVERSION = YES;
CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
CLANG_WARN_STRICT_PROTOTYPES = YES;
CLANG_WARN_SUSPICIOUS_MOVE = YES;
CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
CLANG_WARN_UNREACHABLE_CODE = YES;
CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
CODE_SIGN_IDENTITY = "iPhone Developer";
COPY_PHASE_STRIP = NO;
DEBUG_INFORMATION_FORMAT = dwarf;
ENABLE_STRICT_OBJC_MSGSEND = YES;
ENABLE_TESTABILITY = YES;
GCC_C_LANGUAGE_STANDARD = gnu11;
GCC_DYNAMIC_NO_PIC = NO;
GCC_NO_COMMON_BLOCKS = YES;
GCC_OPTIMIZATION_LEVEL = 0;
GCC_PREPROCESSOR_DEFINITIONS = (
"DEBUG=1",
"$(inherited)",
);
GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
GCC_WARN_UNDECLARED_SELECTOR = YES;
GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
GCC_WARN_UNUSED_FUNCTION = YES;
GCC_WARN_UNUSED_VARIABLE = YES;
IPHONEOS_DEPLOYMENT_TARGET = 11.4;
MTL_ENABLE_DEBUG_INFO = YES;
ONLY_ACTIVE_ARCH = YES;
SDKROOT = iphoneos;
};
name = Debug;
};
FC12E94320EB6B2900807EF4 /* Release */ = {
isa = XCBuildConfiguration;
buildSettings = {
ALWAYS_SEARCH_USER_PATHS = NO;
CLANG_ANALYZER_NONNULL = YES;
CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE;
CLANG_CXX_LANGUAGE_STANDARD = "gnu++14";
CLANG_CXX_LIBRARY = "libc++";
CLANG_ENABLE_MODULES = YES;
CLANG_ENABLE_OBJC_ARC = YES;
CLANG_ENABLE_OBJC_WEAK = YES;
CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES;
CLANG_WARN_BOOL_CONVERSION = YES;
CLANG_WARN_COMMA = YES;
CLANG_WARN_CONSTANT_CONVERSION = YES;
CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES;
CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR;
CLANG_WARN_DOCUMENTATION_COMMENTS = YES;
CLANG_WARN_EMPTY_BODY = YES;
CLANG_WARN_ENUM_CONVERSION = YES;
CLANG_WARN_INFINITE_RECURSION = YES;
CLANG_WARN_INT_CONVERSION = YES;
CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES;
CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES;
CLANG_WARN_OBJC_LITERAL_CONVERSION = YES;
CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR;
CLANG_WARN_RANGE_LOOP_ANALYSIS = YES;
CLANG_WARN_STRICT_PROTOTYPES = YES;
CLANG_WARN_SUSPICIOUS_MOVE = YES;
CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE;
CLANG_WARN_UNREACHABLE_CODE = YES;
CLANG_WARN__DUPLICATE_METHOD_MATCH = YES;
CODE_SIGN_IDENTITY = "iPhone Developer";
COPY_PHASE_STRIP = NO;
DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
ENABLE_NS_ASSERTIONS = NO;
ENABLE_STRICT_OBJC_MSGSEND = YES;
GCC_C_LANGUAGE_STANDARD = gnu11;
GCC_NO_COMMON_BLOCKS = YES;
GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR;
GCC_WARN_UNDECLARED_SELECTOR = YES;
GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
GCC_WARN_UNUSED_FUNCTION = YES;
GCC_WARN_UNUSED_VARIABLE = YES;
IPHONEOS_DEPLOYMENT_TARGET = 11.4;
MTL_ENABLE_DEBUG_INFO = NO;
SDKROOT = iphoneos;
VALIDATE_PRODUCT = YES;
};
name = Release;
};
FC12E94520EB6B2900807EF4 /* Debug */ = {
isa = XCBuildConfiguration;
buildSettings = {
ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
CODE_SIGN_STYLE = Automatic;
DEVELOPMENT_TEAM = Z5M2UUN5YV;
ENABLE_BITCODE = NO;
INFOPLIST_FILE = PaddleMobileDemo/Info.plist;
IPHONEOS_DEPLOYMENT_TARGET = 10.0;
LD_RUNPATH_SEARCH_PATHS = (
"$(inherited)",
"@executable_path/Frameworks",
);
LIBRARY_SEARCH_PATHS = (
"$(inherited)",
"$(PROJECT_DIR)/PaddleMobileDemo/PaddleMobile",
);
PRODUCT_BUNDLE_IDENTIFIER = orange.PaddleMobileDemo;
PRODUCT_NAME = "$(TARGET_NAME)";
TARGETED_DEVICE_FAMILY = "1,2";
};
name = Debug;
};
FC12E94620EB6B2900807EF4 /* Release */ = {
isa = XCBuildConfiguration;
buildSettings = {
ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
CODE_SIGN_STYLE = Automatic;
DEVELOPMENT_TEAM = Z5M2UUN5YV;
ENABLE_BITCODE = NO;
INFOPLIST_FILE = PaddleMobileDemo/Info.plist;
IPHONEOS_DEPLOYMENT_TARGET = 10.0;
LD_RUNPATH_SEARCH_PATHS = (
"$(inherited)",
"@executable_path/Frameworks",
);
LIBRARY_SEARCH_PATHS = (
"$(inherited)",
"$(PROJECT_DIR)/PaddleMobileDemo/PaddleMobile",
);
PRODUCT_BUNDLE_IDENTIFIER = orange.PaddleMobileDemo;
PRODUCT_NAME = "$(TARGET_NAME)";
TARGETED_DEVICE_FAMILY = "1,2";
};
name = Release;
};
/* End XCBuildConfiguration section */
/* Begin XCConfigurationList section */
FC12E92920EB6B2800807EF4 /* Build configuration list for PBXProject "PaddleMobileDemo" */ = {
isa = XCConfigurationList;
buildConfigurations = (
FC12E94220EB6B2900807EF4 /* Debug */,
FC12E94320EB6B2900807EF4 /* Release */,
);
defaultConfigurationIsVisible = 0;
defaultConfigurationName = Release;
};
FC12E94420EB6B2900807EF4 /* Build configuration list for PBXNativeTarget "PaddleMobileDemo" */ = {
isa = XCConfigurationList;
buildConfigurations = (
FC12E94520EB6B2900807EF4 /* Debug */,
FC12E94620EB6B2900807EF4 /* Release */,
);
defaultConfigurationIsVisible = 0;
defaultConfigurationName = Release;
};
/* End XCConfigurationList section */
};
rootObject = FC12E92620EB6B2800807EF4 /* Project object */;
}
demo/ios/PaddleMobileDemo/PaddleMobileDemo.xcodeproj/project.xcworkspace/contents.xcworkspacedata 0 → 100644
浏览文件 @ c67b701f
<?xml version="1.0" encoding="UTF-8"?>
<Workspace
version = "1.0">
<FileRef
location = "self:PaddleMobileDemo.xcodeproj">
</FileRef>
</Workspace>
demo/ios/PaddleMobileDemo/PaddleMobileDemo.xcodeproj/project.xcworkspace/xcshareddata/IDEWorkspaceChecks.plist 0 → 100644
浏览文件 @ c67b701f
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>IDEDidComputeMac32BitWarning</key>
<true/>
</dict>
</plist>
demo/ios/PaddleMobileDemo/PaddleMobileDemo.xcodeproj/xcuserdata/liuruilong.xcuserdatad/xcschemes/xcschememanagement.plist 0 → 100644
浏览文件 @ c67b701f
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>SchemeUserState</key>
<dict>
<key>PaddleMobileDemo.xcscheme</key>
<dict>
<key>orderHint</key>
<integer>0</integer>
</dict>
</dict>
</dict>
</plist>
demo/ios/PaddleMobileDemo/PaddleMobileDemo/AppDelegate.h 0 → 100644
浏览文件 @ c67b701f
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#import <UIKit/UIKit.h>
@interface AppDelegate : UIResponder <UIApplicationDelegate>
@property (strong, nonatomic) UIWindow *window;
@end
demo/ios/PaddleMobileDemo/PaddleMobileDemo/AppDelegate.m 0 → 100644
浏览文件 @ c67b701f
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#import "AppDelegate.h"
@interface AppDelegate ()
@end
@implementation AppDelegate
- (BOOL)application:(UIApplication *)application didFinishLaunchingWithOptions:(NSDictionary *)launchOptions {
// Override point for customization after application launch.
return YES;
}
- (void)applicationWillResignActive:(UIApplication *)application {
// Sent when the application is about to move from active to inactive state. This can occur for certain types of temporary interruptions (such as an incoming phone call or SMS message) or when the user quits the application and it begins the transition to the background state.
// Use this method to pause ongoing tasks, disable timers, and invalidate graphics rendering callbacks. Games should use this method to pause the game.
}
- (void)applicationDidEnterBackground:(UIApplication *)application {
// Use this method to release shared resources, save user data, invalidate timers, and store enough application state information to restore your application to its current state in case it is terminated later.
// If your application supports background execution, this method is called instead of applicationWillTerminate: when the user quits.
}
- (void)applicationWillEnterForeground:(UIApplication *)application {
// Called as part of the transition from the background to the active state; here you can undo many of the changes made on entering the background.
}
- (void)applicationDidBecomeActive:(UIApplication *)application {
// Restart any tasks that were paused (or not yet started) while the application was inactive. If the application was previously in the background, optionally refresh the user interface.
}
- (void)applicationWillTerminate:(UIApplication *)application {
// Called when the application is about to terminate. Save data if appropriate. See also applicationDidEnterBackground:.
}
@end
demo/ios/PaddleMobileDemo/PaddleMobileDemo/Assets.xcassets/AppIcon.appiconset/Contents.json 0 → 100644
浏览文件 @ c67b701f
{
"images" : [
{
"idiom" : "iphone",
"size" : "20x20",
"scale" : "2x"
},
{
"idiom" : "iphone",
"size" : "20x20",
"scale" : "3x"
},
{
"idiom" : "iphone",
"size" : "29x29",
"scale" : "2x"
},
{
"idiom" : "iphone",
"size" : "29x29",
"scale" : "3x"
},
{
"idiom" : "iphone",
"size" : "40x40",
"scale" : "2x"
},
{
"idiom" : "iphone",
"size" : "40x40",
"scale" : "3x"
},
{
"idiom" : "iphone",
"size" : "60x60",
"scale" : "2x"
},
{
"idiom" : "iphone",
"size" : "60x60",
"scale" : "3x"
},
{
"idiom" : "ipad",
"size" : "20x20",
"scale" : "1x"
},
{
"idiom" : "ipad",
"size" : "20x20",
"scale" : "2x"
},
{
"idiom" : "ipad",
"size" : "29x29",
"scale" : "1x"
},
{
"idiom" : "ipad",
"size" : "29x29",
"scale" : "2x"
},
{
"idiom" : "ipad",
"size" : "40x40",
"scale" : "1x"
},
{
"idiom" : "ipad",
"size" : "40x40",
"scale" : "2x"
},
{
"idiom" : "ipad",
"size" : "76x76",
"scale" : "1x"
},
{
"idiom" : "ipad",
"size" : "76x76",
"scale" : "2x"
},
{
"idiom" : "ipad",
"size" : "83.5x83.5",
"scale" : "2x"
},
{
"idiom" : "ios-marketing",
"size" : "1024x1024",
"scale" : "1x"
}
],
"info" : {
"version" : 1,
"author" : "xcode"
}
}
\ No newline at end of file
demo/ios/PaddleMobileDemo/PaddleMobileDemo/Assets.xcassets/Contents.json 0 → 100644
浏览文件 @ c67b701f
{
"info" : {
"version" : 1,
"author" : "xcode"
}
}
\ No newline at end of file
demo/ios/PaddleMobileDemo/PaddleMobileDemo/Base.lproj/LaunchScreen.storyboard 0 → 100644
浏览文件 @ c67b701f
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<document type="com.apple.InterfaceBuilder3.CocoaTouch.Storyboard.XIB" version="3.0" toolsVersion="13122.16" systemVersion="17A277" targetRuntime="iOS.CocoaTouch" propertyAccessControl="none" useAutolayout="YES" launchScreen="YES" useTraitCollections="YES" useSafeAreas="YES" colorMatched="YES" initialViewController="01J-lp-oVM">
<dependencies>
<plugIn identifier="com.apple.InterfaceBuilder.IBCocoaTouchPlugin" version="13104.12"/>
<capability name="Safe area layout guides" minToolsVersion="9.0"/>
<capability name="documents saved in the Xcode 8 format" minToolsVersion="8.0"/>
</dependencies>
<scenes>
<!--View Controller-->
<scene sceneID="EHf-IW-A2E">
<objects>
<viewController id="01J-lp-oVM" sceneMemberID="viewController">
<view key="view" contentMode="scaleToFill" id="Ze5-6b-2t3">
<rect key="frame" x="0.0" y="0.0" width="375" height="667"/>
<autoresizingMask key="autoresizingMask" widthSizable="YES" heightSizable="YES"/>
<color key="backgroundColor" red="1" green="1" blue="1" alpha="1" colorSpace="custom" customColorSpace="sRGB"/>
<viewLayoutGuide key="safeArea" id="6Tk-OE-BBY"/>
</view>
</viewController>
<placeholder placeholderIdentifier="IBFirstResponder" id="iYj-Kq-Ea1" userLabel="First Responder" sceneMemberID="firstResponder"/>
</objects>
<point key="canvasLocation" x="53" y="375"/>
</scene>
</scenes>
</document>
demo/ios/PaddleMobileDemo/PaddleMobileDemo/Base.lproj/Main.storyboard 0 → 100644
浏览文件 @ c67b701f
<?xml version="1.0" encoding="UTF-8"?>
<document type="com.apple.InterfaceBuilder3.CocoaTouch.Storyboard.XIB" version="3.0" toolsVersion="13122.16" systemVersion="17A277" targetRuntime="iOS.CocoaTouch" propertyAccessControl="none" useAutolayout="YES" useTraitCollections="YES" useSafeAreas="YES" colorMatched="YES" initialViewController="BYZ-38-t0r">
<dependencies>
<plugIn identifier="com.apple.InterfaceBuilder.IBCocoaTouchPlugin" version="13104.12"/>
<capability name="Safe area layout guides" minToolsVersion="9.0"/>
<capability name="documents saved in the Xcode 8 format" minToolsVersion="8.0"/>
</dependencies>
<scenes>
<!--View Controller-->
<scene sceneID="tne-QT-ifu">
<objects>
<viewController id="BYZ-38-t0r" customClass="ViewController" customModuleProvider="" sceneMemberID="viewController">
<view key="view" contentMode="scaleToFill" id="8bC-Xf-vdC">
<rect key="frame" x="0.0" y="0.0" width="375" height="667"/>
<autoresizingMask key="autoresizingMask" widthSizable="YES" heightSizable="YES"/>
<color key="backgroundColor" red="1" green="1" blue="1" alpha="1" colorSpace="custom" customColorSpace="sRGB"/>
<viewLayoutGuide key="safeArea" id="6Tk-OE-BBY"/>
</view>
</viewController>
<placeholder placeholderIdentifier="IBFirstResponder" id="dkx-z0-nzr" sceneMemberID="firstResponder"/>
</objects>
</scene>
</scenes>
</document>
demo/ios/PaddleMobileDemo/PaddleMobileDemo/Info.plist 0 → 100644
浏览文件 @ c67b701f
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
<key>CFBundleDevelopmentRegion</key>
<string>$(DEVELOPMENT_LANGUAGE)</string>
<key>CFBundleExecutable</key>
<string>$(EXECUTABLE_NAME)</string>
<key>CFBundleIdentifier</key>
<string>$(PRODUCT_BUNDLE_IDENTIFIER)</string>
<key>CFBundleInfoDictionaryVersion</key>
<string>6.0</string>
<key>CFBundleName</key>
<string>$(PRODUCT_NAME)</string>
<key>CFBundlePackageType</key>
<string>APPL</string>
<key>CFBundleShortVersionString</key>
<string>1.0</string>
<key>CFBundleVersion</key>
<string>1</string>
<key>LSRequiresIPhoneOS</key>
<true/>
<key>UILaunchStoryboardName</key>
<string>LaunchScreen</string>
<key>UIMainStoryboardFile</key>
<string>Main</string>
<key>UIRequiredDeviceCapabilities</key>
<array>
<string>armv7</string>
</array>
<key>UISupportedInterfaceOrientations</key>
<array>
<string>UIInterfaceOrientationPortrait</string>
<string>UIInterfaceOrientationLandscapeLeft</string>
<string>UIInterfaceOrientationLandscapeRight</string>
</array>
<key>UISupportedInterfaceOrientations~ipad</key>
<array>
<string>UIInterfaceOrientationPortrait</string>
<string>UIInterfaceOrientationPortraitUpsideDown</string>
<string>UIInterfaceOrientationLandscapeLeft</string>
<string>UIInterfaceOrientationLandscapeRight</string>
</array>
</dict>
</plist>
demo/ios/PaddleMobileDemo/PaddleMobileDemo/PaddleMobile/PaddleMobile.h 0 → 100644
浏览文件 @ c67b701f
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#import <CoreImage/CoreImage.h>
#import <Foundation/Foundation.h>
@interface PaddleMobile : NSObject
- (instancetype)init;
- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale;
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim;
- (void)clear;
@end
demo/ios/PaddleMobileDemo/PaddleMobileDemo/ViewController.h 0 → 100644
浏览文件 @ c67b701f
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#import <UIKit/UIKit.h>
@interface ViewController : UIViewController
@end
demo/ios/PaddleMobileDemo/PaddleMobileDemo/ViewController.m 0 → 100644
浏览文件 @ c67b701f
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#import "PaddleMobile.h"
#import "ViewController.h"
@interface ViewController ()
@end
@implementation ViewController
- (void)viewDidLoad {
[super viewDidLoad];
PaddleMobile *pam = [[PaddleMobile alloc] init];
NSString *modelPath = [[NSBundle mainBundle] pathForResource:@"model" ofType:nil];
NSString *paramPath = [[NSBundle mainBundle] pathForResource:@"params" ofType:nil];
if (modelPath.length == 0 || paramPath.length == 0) {
NSLog(@" need model and param");
return;
}
if ([pam load:modelPath andWeightsPath:paramPath]) {
NSLog(@"load success");
UIImage *inputImage = [UIImage imageNamed:@"apple.jpg"];
if (!inputImage) {
NSLog(@" input image is nil");
return;
}
NSDate *beforeDate = [NSDate date];
NSArray *res = [pam predict:inputImage.CGImage dim:@[@1, @3, @224, @224] means:@[@148, @148, @148] scale:1.0];
NSLog(@"res: %@", res);
NSLog(@"elapsed time: %f", [[NSDate date] timeIntervalSinceDate:beforeDate]);
}
}
@end
demo/ios/PaddleMobileDemo/PaddleMobileDemo/main.m 0 → 100644
浏览文件 @ c67b701f
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#import <UIKit/UIKit.h>
#import "AppDelegate.h"
int main(int argc, char * argv[]) {
@autoreleasepool {
return UIApplicationMain(argc, argv, nil, NSStringFromClass([AppDelegate class]));
}
}
doc/build.md 0 → 100644
浏览文件 @ c67b701f
# 环境搭建
## 使用 docker
### 1. 安装 docker
安装 docker 的方式，参考官方文档 [https://docs.docker.com/install/](https://docs.docker.com/install/)
### 2. 使用 docker 搭建构建环境
首先进入 paddle-mobile 的目录下，执行 `docker build`
以 Linux/Mac 为例 (windows 建议在 'Docker Quickstart Terminal' 中执行)
```
$ docker build -t paddle-mobile:dev - < Dockerfile
```
使用 `docker images` 可以看到我们新建的 image
```
$ docker images
REPOSITORY TAG IMAGE ID CREATED SIZE
paddle-mobile dev 33b146787711 45 hours ago 372MB
```
### 3. 使用 docker 构建
进入 paddle-mobile 目录，执行 docker run
```
$ docker run -it --mount type=bind,source=$PWD,target=/paddle-mobile paddle-mobile:dev
root@5affd29d4fc5:/ # cd /paddle-mobile
# 生成构建 android 产出的 Makefile
root@5affd29d4fc5:/ # rm CMakeCache.txt
root@5affd29d4fc5:/ # cmake -DCMAKE_TOOLCHAIN_FILE=tools/toolchains/arm-android-neon.cmake
# 生成构建 linux 产出的 Makefile
root@5affd29d4fc5:/ # rm CMakeCache.txt
root@5affd29d4fc5:/ # cmake -DCMAKE_TOOLCHAIN_FILE=tools/toolchains/arm-linux-gnueabi.cmake
```
### 4. 设置编译选项
可以通过 ccmake 设置编译选项
```
root@5affd29d4fc5:/ # ccmake .
Page 1 of 1
CMAKE_ASM_FLAGS
CMAKE_ASM_FLAGS_DEBUG
CMAKE_ASM_FLAGS_RELEASE
CMAKE_BUILD_TYPE
CMAKE_INSTALL_PREFIX /usr/local
CMAKE_TOOLCHAIN_FILE /paddle-mobile/tools/toolchains/arm-android-neon.cmake
CPU ON
DEBUGING ON
FPGA OFF
LOG_PROFILE ON
MALI_GPU OFF
NET googlenet
USE_EXCEPTION ON
USE_OPENMP OFF
```
修改选项后，按 `c`, `g` 更新 Makefile
### 5. 构建
使用 make 命令进行构建
```
root@5affd29d4fc5:/ # make
```
### 6. 查看构建产出
构架产出可以在 host 机器上查看，在 paddle-mobile 的目录下，build 以及 test/build 下，可以使用 adb 指令或者 scp 传输到 device 上执行
## 不使用 docker
不使用 docker 的方法，可以直接用 cmake 生成 makefile 后构建。使用 ndk 构建 android 应用需要正确设置 NDK_ROOT。构建 linux 应用需要安装 arm-linux-gnueabi-gcc 或者类似的交叉编译工具，可能需要设置 CC，CXX 环境变量，或者在 tools/toolchains/ 中修改 arm-linux-gnueabi.cmake，或者增加自己需要的 toolchain file。
\ No newline at end of file
doc/design_doc.md 0 → 100644
浏览文件 @ c67b701f
# paddle-mobile 设计文档
#### 以下是 paddle-mobile 代码的执行流程图:
![执行流程图](http://otkwwi4x8.bkt.clouddn.com/2018-07-02-15305189473720.png)
#### 主要分为: Loader 模块、 Program 模块、 Executor 模块、 op 模块、 kernel 模块、scope variable Tensor 模块
#### 下面展开说一下各个模块的作用以及设计思路
### 一. Loader
先来看一下模型, 模型分为两种结构:
一种为参数文件是散开的, 如下图, 红框为模型结构的 protobuf 文件, 其余为参数文件
![模型描述](http://otkwwi4x8.bkt.clouddn.com/2018-07-02-15305190629577.png)
另一种为参数文件结合在一起的, 如下图, 红框内为模型结构描述的 protobuf 文件, 另一个文件为结合在一起的参数文件
![模型描述combined](http://otkwwi4x8.bkt.clouddn.com/2018-07-02-15305191057130.png)
loader 模块的作用是将模型结构信息 load 进内存, 将红框内的 protobuf 文件 load 进内存, 并对模型结构进行优化(如将几个细粒度的 op 融合成 粗粒度的 op, 如将 conv、 add、 batchnorm、 relu 融合为 conv\_add\_batchnorm\_relu).
方便进行算法优化.
__那么为什么融合在一起能够做算法优化 ?__
如果未融合的 conv add batchnorm relu 运算是这样的
```
[n]
[conv_res] = conv([n])
for &res in conv_res {
res = add_biase(res)
}
for &res in conv_res {
res = batchnorm(res)
}
for &res in conv_res {
res = relu(res)
}
```
融合后的 conv\_add\_batchnorm\_relu 运算是这样的:
```
[n]
[conv_res] = conv([n])
for &res in conv_res {
res = relu(batchnorm(add_biase(res)))
}
```
由于 conv 可以转换为两个大矩阵相乘, 更进一步可以分为若干个一行一列的小矩阵相乘, 那最终的运算是这样的:
```
[n]
for &res in [res] {
res = relu(batchnorm(add_biase(A * B)))
}
其中 A 和 B 为 1 * k 和 k * 1 矩阵
```
### 二. Program
program 为 loader 模块的结果, 包含了优化前的模型结构对象, 以及优化后的模型结构对象, 此模块基本对应着 paddle 模型的结构, 关于paddle 模型的一些概念的定义, 详细设计可以参考 [program.md](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/program.md), 以下是一个简单的概况:
* programDesc 中包含着若干个(googlenet mobilenet yolo squeezenet resnet 常见的模型只有一个)可以嵌套的 block, blocks中的第一个block中的某个 op 可能会执行 blocks 中后边 block 中的一系列 op 运算(只有多个block才会有此概念)
* block 包含着 ops 和 vars
* ops 为一系列 op 的描述, 描述着每个 op 的类型, 输入输出, 所需参数
* vars 里包含的为所有 op 运算所需的参数描述
### 三. Executor
executor 主要是用于 op 运算的上层调度操作, 主要有两个操作, executor 实例化 和 暴露给上层的 predict 方法
* executor 实例化过程中, 主要进行了这几个操作
1. 根据 loader 产出的 program 初始化 operator 对象
2. 分配所有需要用到的内存, 包括每个op 的输入输出, 权重参数, 目前模型的权重参数文件的内存格式为 NCHW, op 的输入输出中间矩阵参数也是 NCHW 格式
3. 调用每个 op 的 init 方法, init 方法是每个 op 实现者进行参数预处理的地方, 有助于减少 predict 的耗时
* predict, 主要用于拿到外部的输入, 顺序调用 op 的 run 方法进行运算, 并返回最终的结果.
### 四. op
关于 op 模块代码的详细设计可以参考 [operator部分代码设计](https://github.com/PaddlePaddle/paddle-mobile/issues/300), operator主要包含一个kernel用于运算、一个 param 用于存储属性, operator 主要有三个操作, Init、RunImp、InferShape
* Init: Init 函数主要用于参数预处理, 如对 batchNorm 参数进行预处理, 可以将 batchNorm 运算转化为 a * x + b 形式的运算, 这个函数也会调用, kernel 的 Init 函数对 kernel 进行初始化
* RunImp: RunImp 函数会调用自己的kernel 的 compute 方法进行运算
* InferShape: InferShape 函数会根据输入和参数得出输出的形状, 这个函数会在 executor 实例化时, 内存初始化前调用
每个 operator 都需要进行注册才可以被使用, 以 conv 为例, 需在 conv_op.cpp 底部这样写:
```c++
// 三个平台都注册了 conv op
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(conv2d);
REGISTER_OPERATOR_CPU(conv2d, ops::ConvOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(conv2d);
REGISTER_OPERATOR_MALI_GPU(conv2d, ops::ConvOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(conv2d);
REGISTER_OPERATOR_FPGA(conv2d, ops::ConvOp);
#endif
```
__一个关于包大小的优化__:
每个 operator 都由一个宏控制编译, 如 conv_op.h(除了 conv_op.h , conv_op.cpp、conv_kernle.h、conv_kernle.cpp 也都需要加此宏控制)
```c++
#ifdef CONV_OP //这个宏控制着 conv_op 是否被编译, 除了 conv_op.h , conv_op.cpp、conv_kernle.h conv_kernle.cpp 也都需要加此宏控制
#pragma once
#include <string>
#include "framework/operator.h"
#include "operators/kernel/conv_kernel.h"
namespace paddle_mobile {
namespace operators {
using std::string;
template <typename DeviceType, typename T>
class ConvOp
//impl
};
} // namespace operators
} // namespace paddle_mobile
#endif
```
这样做的目的是为了根据不同类型的网络编译特定的op, 在 cmake 中已经配置好不同网络编译的宏, 如果你要进行编译支持 yolo 的模型, 仅需执行:
```sh
cd toools
sh build.sh android yolo
```
这样只会编译 yolo 所包含的四种 op, 极大的减小了包体积和编译时间
### 五. kernel
kernel 为 op 的底层运算实现, 主要有两个函数, Init 和 Compute, 分别用来初始化、预处理 和 运算操作, 值得提出的是, kernel 会根据泛型特化到不同的平台, 如图所示:
![设备特化]![](http://otkwwi4x8.bkt.clouddn.com/2018-07-02-15305191401976.png)
不同平台的 kernel 实现, 为同一个 kernel 类不同泛型的特化实现, 目前有三个平台, arm、mali、fpga, 图中的 central-arm-func\ 目录为 op kernel 的 arm 实现, 它承担了 arm\ 目录下 kernel 的底层实现, 同时 arm 处理器作为中央处理器, central-arm-func\ 也可以作为其他协处理器的底层实现, 如: fpga 的某一个 op kernel 还没有 fpga 协处理器的实现, 就可以直接调用使用这里的 arm 实现.
__如果你有兴趣新增一个协处理器实现, 就可以在次添加一个 kernel 目录, 提供协处理器实现, 如果某个 kernel 你没有实现完, 你也可以直接使用 arm 实现__
### 六. scope variable Tensor
* scope 用来存储管理所需用到的所有 variable(用来存储不同类型的对象, 主要是矩阵Tensor, 也就是说 scpoe 管理着 op 运算过程中所有参数矩阵, 输入输出矩阵), 可以将 scope 理解为一个 map, 这里在 map 上封了一层 scope 的概念是为了方便内存管理
* variable 可以用来存储不同类型的对象, paddle-mobile 里主要用它来存储矩阵 Tensor
* tensor 代表着矩阵, 通过泛型可以用来存储不同类型的矩阵, 但需要注意的是, 存入和取出时的类型必须保持一致, 如果类型不一致, 使用 inline const T \*data() const 获取指针会不能通过类型检查, 通过 inline T \*mutable_data() 获取指针会重新分配内存, 以下是关于 Tensor 的一些小概念:
1. DDim: 用来存储矩阵的维度信息.
2. Slice(): 这个函数用来获取 N 维 (NCHW中的 N) 上切片
3. 当实例化未分配内存时, 调用 inline T *mutable_data() 会分配内存
doc/development_doc.md 0 → 100644
浏览文件 @ c67b701f
### iOS&Android开发文档
# iOS开发文档
## 编译
### 一. 使用 build.sh 编译
```sh
sh build.sh ios
# 如果只想编译某个特定模型的 op, 则需执行以下命令
sh build.sh ios googlenet
# 在这个文件夹下, 你可以拿到生成的 .a 库
cd ../build/release/ios/build
```
### 二. 使用 xcode 编译
我们提供了 ios 开发更为熟悉的 xcode 编译环境:
在 ios/ 目录下打开 PaddleMobile.xcworkspace 即可编译 PaddleMobile 或者 运行 Demo
### 三. 集成
#### 如使用 c++ 接口
将
```
libpaddle-mobile.a
io.h
program.h
types.h
lod_tensor.h
tensor.h
```
拖入工程, io.h 为接口文件, 可在 [github](https://github.com/PaddlePaddle/paddle-mobile/blob/develop/src/io/io.h)上查看接口注释
#### 如使用 oc 接口
将在xcode 编译生成的
```
libPaddleMobile.a
PaddleMobile.h
```
拖入工程, 接口如下:
```
/*
创建单例对象
*/
+ (instancetype)sharedInstance;
/*
load 模型, 开辟内存
*/
- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
/*
进行预测, means 和 scale 为训练模型时的预处理参数, 如训练时没有做这些预处理则直接使用 predict
*/
- (NSArray *)predict:(CGImageRef)image means:(NSArray<NSNumber *> *)means scale:(float)scale;
/*
进行预测
*/
- (NSArray *)predict:(CGImageRef)image;
/*
清理内存
*/
- (void)clear;
```
#Android开发文档
用户可通过如下两种方式，交叉编译Android平台上适用的paddle-mobile库：
- 基于Docker容器编译
- 基于Linux交叉编译
## 基于Docker容器编译
### 1. 安装 docker
安装 docker 的方式，参考官方文档 [https://docs.docker.com/install/](https://docs.docker.com/install/)
### 2. 使用 docker 搭建构建环境
首先进入 paddle-mobile 的目录下，执行 `docker build`
以 Linux/Mac 为例 (windows 建议在 'Docker Quickstart Terminal' 中执行)
```
$ docker build -t paddle-mobile:dev - < Dockerfile
```
使用 `docker images` 可以看到我们新建的 image
```
$ docker images
REPOSITORY TAG IMAGE ID CREATED SIZE
paddle-mobile dev 33b146787711 45 hours ago 372MB
```
### 3. 使用 docker 构建
进入 paddle-mobile 目录，执行 docker run
```
$ docker run -it --mount type=bind,source=$PWD,target=/paddle-mobile paddle-mobile:dev
root@5affd29d4fc5:/ # cd /paddle-mobile
# 生成构建 android 产出的 Makefile
root@5affd29d4fc5:/ # rm CMakeCache.txt
root@5affd29d4fc5:/ # cmake -DCMAKE_TOOLCHAIN_FILE=tools/toolchains/arm-android-neon.cmake
# 生成构建 linux 产出的 Makefile
root@5affd29d4fc5:/ # rm CMakeCache.txt
root@5affd29d4fc5:/ # cmake -DCMAKE_TOOLCHAIN_FILE=tools/toolchains/arm-linux-gnueabi.cmake
```
### 4. 设置编译选项
可以通过 ccmake 设置编译选项
```
root@5affd29d4fc5:/ # ccmake .
Page 1 of 1
CMAKE_ASM_FLAGS
CMAKE_ASM_FLAGS_DEBUG
CMAKE_ASM_FLAGS_RELEASE
CMAKE_BUILD_TYPE
CMAKE_INSTALL_PREFIX /usr/local
CMAKE_TOOLCHAIN_FILE /paddle-mobile/tools/toolchains/arm-android-neon.cmake
CPU ON
DEBUGING ON
FPGA OFF
LOG_PROFILE ON
MALI_GPU OFF
NET googlenet
USE_EXCEPTION ON
USE_OPENMP OFF
```
修改选项后，按 `c`, `g` 更新 Makefile
### 5. 构建
使用 make 命令进行构建
```
root@5affd29d4fc5:/ # make
```
### 6. 查看构建产出
构架产出可以在 host 机器上查看，在 paddle-mobile 的目录下，build 以及 test/build 下，可以使用 adb 指令或者 scp 传输到 device 上执行
## 基于Linux交叉编译
### 交叉编译环境准备
##### 下载Android NDK
从源码交叉编译paddle-mobile,用户需要提前准备好交叉编译环境。Android平台使用的C/C++交叉编译工具链是[Android NDK](https://developer.android.com/ndk/)，用户可以自行前往下载，也可以通过以下命令获取：
```
wget https://dl.google.com/android/repository/android-ndk-r17b-darwin-x86_64.zip
unzip android-ndk-r17b-darwin-x86_64.zip
```
##### 设置环境变量
工程中自带的独立工具链会根据环境变量NDK_ROOT查找NDK，因此需要配置环境变量：
```
export NDK_ROOT = "path to ndk"
```
### 执行编译
在paddle-mobile根目录中，执行以下命令：
```
cd tools
sh build.sh android
```
执行完毕后，生成的so位于build目录中，单测可执行文件位于test/build目录中。
##### Tips:
如果想要获得体积更小的库，可选择编译支持指定模型结构的库。
如执行如下命令：
```
sh build.sh android googlenet
```
会得到一个支持googlnet的体积更小的库。
##测试
在编译完成后，我们提供了自动化的测试脚本，帮助用户将运行单测文件所需要的模型及库文件push到Android设备中，执行以下命令：
```
cd tools/android-debug-script
sh run_on_android.sh (npm) 可选参数npm,用于选择是否传输模型文件到手机上
```
出现如下提示：
```
**** choose OP or NET to test ****
which to test :
```
输入名称即可运行对应的测试文件。
##部署
Android应用可通过JNI接口调用底层C/C++，paddle-mobile对外提供的JNI接口如下：
##### 1 load接口 加载模型参数
```
/*
*@param modelPath 模型文件路径
*@return jboolean
*/
JNIEXPORT jboolean JNICALL Java_com_baidu_paddle_PML_load(JNIEnv *env,
jclass thiz,
jstring modelPath);
```
##### 2 predict接口 执行预测
```
/**
*@param buf 输入数据
*@return 输出数据
JNIEXPORT jfloatArray JNICALL Java_com_baidu_paddle_PML_predict(
JNIEnv *env, jclass thiz, jfloatArray buf);
```
##### 3 clear接口 销毁实例、清理内存操作
```
JNIEXPORT void JNICALL Java_com_baidu_paddle_PMLL_clear(JNIEnv *env,
jclass thiz);
```
src/common/common.h 0 → 100644
浏览文件 @ c67b701f
/* 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 <chrono>
using Time = decltype(std::chrono::high_resolution_clock::now());
inline Time time() { return std::chrono::high_resolution_clock::now(); }
inline double time_diff(Time t1, Time t2) {
typedef std::chrono::microseconds ms;
auto diff = t2 - t1;
ms counter = std::chrono::duration_cast<ms>(diff);
return counter.count() / 1000.0;
}
src/common/log.h
浏览文件 @ c67b701f
......@@ -16,6 +16,7 @@ limitations under the License. */
#include <vector>
#ifdef PADDLE_MOBILE_DEBUG
#include <cstring>
#include <iostream>
#include <sstream>
#include <string>
......@@ -115,26 +116,29 @@ struct ToLog {
Print printer_;
};
#define LOG(level) \
if (level > paddle_mobile::log_level) { \
} else \
paddle_mobile::ToLog( \
level, \
(std::stringstream() \
<< "[file: " \
<< (strrchr(__FILE__, '/') ? (strrchr(__FILE__, '/') + 1) : __FILE__) \
<< "] [line: " << __LINE__ << "] ") \
.str())
#define DLOG \
if (paddle_mobile::kLOG_DEBUG > paddle_mobile::log_level) { \
} else \
paddle_mobile::ToLog( \
paddle_mobile::kLOG_DEBUG, \
(std::stringstream() \
<< "[file: " \
<< (strrchr(__FILE__, '/') ? (strrchr(__FILE__, '/') + 1) : __FILE__) \
<< "] [line: " << __LINE__ << "] ") \
#define LOG(level) \
if (level > paddle_mobile::log_level) { \
} else \
paddle_mobile::ToLog( \
level, static_cast<const std::stringstream &>( \
std::stringstream() \
<< "[file: " \
<< (strrchr(__FILE__, '/') ? (strrchr(__FILE__, '/') + 1) \
: __FILE__) \
<< "] [line: " << __LINE__ << "] ") \
.str())
#define DLOG \
if (paddle_mobile::kLOG_DEBUG > paddle_mobile::log_level) { \
} else \
paddle_mobile::ToLog( \
paddle_mobile::kLOG_DEBUG, \
static_cast<const std::stringstream &>( \
std::stringstream() \
<< "[file: " \
<< (strrchr(__FILE__, '/') ? (strrchr(__FILE__, '/') + 1) \
: __FILE__) \
<< "] [line: " << __LINE__ << "] ") \
.str())
#define LOGF(level, format, ...) \
......@@ -170,7 +174,10 @@ struct ToLog;
struct Print {
friend struct ToLog;
template <typename T>
Print &operator<<(T const &value) {}
Print &operator<<(T const &value) {
Print p = Print();
return p;
}
private:
};
......
src/common/type_define.h
浏览文件 @ c67b701f
......@@ -12,8 +12,9 @@ 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;
#pragma once
#include <functional>
#include <map>
#include <string>
#include <vector>
......
src/common/types.cpp
浏览文件 @ c67b701f
......@@ -23,8 +23,9 @@ const std::string G_OP_TYPE_BOX_CODER = "box_coder";
const std::string G_OP_TYPE_CONCAT = "concat";
const std::string G_OP_TYPE_ELEMENTWISE_ADD = "elementwise_add";
const std::string G_OP_TYPE_FUSION_CONV_ADD_RELU = "fusion_conv_add_relu";
const std::string G_OP_TYPE_FC = "fc";
const std::string G_OP_TYPE_CONV_ADD = "conv_add";
const std::string G_OP_TYPE_FUSION_CONV_ADD_BN_RELU = "fusion_conv_add_bn_relu";
const std::string G_OP_TYPE_FC = "fusion_fc";
const std::string G_OP_TYPE_FUSION_CONV_ADD = "fusion_conv_add";
const std::string G_OP_TYPE_LRN = "lrn";
const std::string G_OP_TYPE_MUL = "mul";
const std::string G_OP_TYPE_MULTICLASS_NMS = "multiclass_nms";
......@@ -39,12 +40,14 @@ const std::string G_OP_TYPE_SPLIT = "split";
const std::string G_OP_TYPE_FEED = "feed";
const std::string G_OP_TYPE_FETCH = "fetch";
const std::string G_OP_TYPE_DEPTHWISE_CONV = "depthwise_conv2d";
const std::string G_OP_TYPE_IM2SEQUENCE = "im2sequence";
const std::string G_OP_TYPE_DROPOUT = "dropout";
std::unordered_map<
std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
op_input_output_key = {
{G_OP_TYPE_CONV, {{"Input"}, {"Output"}}},
{G_OP_TYPE_CONV_ADD, {{"Input"}, {"Out"}}},
{G_OP_TYPE_FUSION_CONV_ADD, {{"Input"}, {"Out"}}},
{G_OP_TYPE_RELU, {{"X"}, {"Out"}}},
{G_OP_TYPE_SOFTMAX, {{"X"}, {"Out"}}},
{G_OP_TYPE_MUL, {{"X"}, {"Out"}}},
......@@ -59,11 +62,14 @@ std::unordered_map<
{G_OP_TYPE_TRANSPOSE, {{"X"}, {"Out"}}},
{G_OP_TYPE_BOX_CODER,
{{"PriorBox", "PriorBoxVar", "TargetBox"}, {"OutputBox"}}},
{G_OP_TYPE_FUSION_CONV_ADD_BN_RELU, {{"Input"}, {"Out"}}},
{G_OP_TYPE_PRIOR_BOX, {{"Image", "Input"}, {"Boxes", "Variances"}}},
{G_OP_TYPE_MULTICLASS_NMS, {{"BBoxes", "Scores"}, {"Out"}}},
{G_OP_TYPE_FC, {{"X", "Y", "Z"}, {"Out"}}},
{G_OP_TYPE_RESHAPE, {{"X"}, {"Out"}}},
{G_OP_TYPE_DEPTHWISE_CONV, {{"Input"}, {"Output"}}},
{G_OP_TYPE_FUSION_CONV_ADD_RELU, {{"Input"}, {"Out"}}}};
{G_OP_TYPE_FUSION_CONV_ADD_RELU, {{"Input"}, {"Out"}}},
{G_OP_TYPE_IM2SEQUENCE, {{"X"}, {"Out"}}},
{G_OP_TYPE_DROPOUT, {{"X"}, {"Out"}}}};
} // namespace paddle_mobile
src/common/types.h
浏览文件 @ c67b701f
......@@ -12,10 +12,11 @@ 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;
#pragma once
#include <string>
#include <unordered_map>
#include <vector>
namespace paddle_mobile {
enum class Precision : int { FP32 = 0 };
......@@ -78,7 +79,9 @@ extern const std::string G_OP_TYPE_CONCAT;
extern const std::string G_OP_TYPE_ELEMENTWISE_ADD;
extern const std::string G_OP_TYPE_FUSION_CONV_ADD_RELU;
extern const std::string G_OP_TYPE_FC;
extern const std::string G_OP_TYPE_CONV_ADD;
extern const std::string G_OP_TYPE_FUSION_CONV_ADD;
extern const std::string G_OP_TYPE_FUSION_CONV_ADD_BN_RELU;
extern const std::string G_OP_TYPE_LRN;
extern const std::string G_OP_TYPE_MUL;
extern const std::string G_OP_TYPE_MULTICLASS_NMS;
......@@ -93,6 +96,8 @@ extern const std::string G_OP_TYPE_SPLIT;
extern const std::string G_OP_TYPE_FEED;
extern const std::string G_OP_TYPE_FETCH;
extern const std::string G_OP_TYPE_DEPTHWISE_CONV;
extern const std::string G_OP_TYPE_IM2SEQUENCE;
extern const std::string G_OP_TYPE_DROPOUT;
extern std::unordered_map<
std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
......
src/common/variant.h
浏览文件 @ c67b701f
......@@ -83,6 +83,7 @@ struct Variant {
return *const_cast<T *>(reinterpret_cast<const T *>(&data));
} else {
PADDLE_MOBILE_THROW_EXCEPTION(" bad cast in variant ");
exit(0);
}
}
......
src/framework/attribute.h
浏览文件 @ c67b701f
......@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include <string>
#include <typeinfo>
#include <unordered_map>
#include <vector>
......@@ -128,6 +129,7 @@ class Attribute {
return vistor(attr.variant_.Get<int64_t>());
} else {
PADDLE_MOBILE_THROW_EXCEPTION("type not support");
exit(0);
}
}
......
src/framework/data_layout.h
浏览文件 @ c67b701f
......@@ -40,6 +40,7 @@ inline DataLayout StringToDataLayout(const std::string &str) {
return DataLayout::kAnyLayout;
} else {
PADDLE_MOBILE_THROW_EXCEPTION("Unknown storage order string: %s", s.c_str())
exit(0);
}
}
......@@ -52,6 +53,8 @@ inline std::string DataLayoutToString(const DataLayout &data_layout) {
case DataLayout::kAnyLayout:
return "ANY_LAYOUT";
default:
PADDLE_MOBILE_THROW_EXCEPTION("Unknown storage order string ")
exit(0);
break;
}
}
......
src/framework/ddim.h
浏览文件 @ c67b701f
......@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include <initializer_list>
#include <typeinfo>
#include <vector>
#include "common/enforce.h"
#include "common/variant.h"
......@@ -57,7 +58,8 @@ struct DDim {
} else if (d.var.TypeId() == typeid(Dim<9>).hash_code()) {
return vistor(d.var.Get<Dim<9>>());
} else {
DLOG << " dim not support";
PADDLE_MOBILE_ENFORCE(false, " dim not support");
exit(0);
}
}
......
src/framework/dim.h
浏览文件 @ c67b701f
......@@ -129,6 +129,7 @@ int64_t &indexer(Dim<D> &dim, int idx) {
template <>
int64_t &indexer<0>(Dim<0> &dim, int idx) {
PADDLE_MOBILE_THROW_EXCEPTION("Invalid index")
exit(0);
}
template <int D>
......@@ -145,6 +146,7 @@ int64_t indexer(const Dim<D> &dim, int idx) {
template <>
int64_t indexer<0>(const Dim<0> &dim, int idx) {
PADDLE_MOBILE_THROW_EXCEPTION("Invalid index")
exit(0);
}
} // namespace
......
src/framework/lod_tensor.cpp
浏览文件 @ c67b701f
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "lod_tensor.h"
#include <algorithm>
namespace paddle_mobile {
namespace framework {
......
src/framework/op_registry.h
浏览文件 @ c67b701f
......@@ -45,6 +45,7 @@ struct OperatorRegistrar : public Registrar {
<< "OperatorRegistrar should be invoked at least by OpClass";
return;
}
printf(" regis ting %s \n", op_type.c_str());
OpInfo<Dtype> info;
OperatorRegistrarRecursive<Dtype, 0, false, ARGS...>(op_type, &info);
OpInfoMap<Dtype>::Instance()->Insert(op_type, info);
......@@ -107,6 +108,7 @@ class OpRegistry {
__op_registrar_##op_type##_##device_name(#op_type); \
int TouchOpRegistrar_##op_type##_##device_name() { \
__op_registrar_##op_type##_##device_name.Touch(); \
printf(" registering !! \n"); \
return 0; \
}
......
src/framework/operator.h
浏览文件 @ c67b701f
......@@ -63,7 +63,7 @@ class OperatorBase {
std::vector<string> GetOutKeys() const;
virtual void RunImpl() const = 0;
virtual void Init() const = 0;
virtual void Init() = 0;
/*
* @b op 运算所需的输入, 如上一层的输出结果、卷积核
* */
......@@ -117,8 +117,8 @@ class OperatorWithKernel : public OperatorBase<Dtype> {
virtual void InferShape() const = 0;
void Init() const {
PADDLE_MOBILE_ENFORCE(kernel_.Init(param_), " %s kernel init failed",
void Init() {
PADDLE_MOBILE_ENFORCE(kernel_.Init(&param_), " %s kernel init failed",
this->type_.c_str());
}
......@@ -146,7 +146,7 @@ class OpKernelBase {
}
#endif
virtual void Compute(const P &para) const = 0;
virtual bool Init(const P &para) const { return true; };
virtual bool Init(P *para) { return true; };
virtual ~OpKernelBase() = default;
private:
......
src/framework/program/block_desc.cpp
浏览文件 @ c67b701f
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "block_desc.h"
#include <algorithm>
namespace paddle_mobile {
namespace framework {
......
src/framework/program/program-optimize/fusion_op_register.h
浏览文件 @ c67b701f
......@@ -34,6 +34,10 @@ class FusionOpRegister {
}
void regist(FusionOpMatcher* matcher) {
if (matchers_.find(matcher->Type()) != matchers_.end()) {
return;
}
std::shared_ptr<FusionOpMatcher> shared_matcher(matcher);
matchers_[matcher->Type()] = shared_matcher;
}
......
src/framework/program/program-optimize/node.cpp
浏览文件 @ c67b701f
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "framework/program/program-optimize/node.h"
#include <algorithm>
#include "framework/operator.h"
namespace paddle_mobile {
......@@ -92,7 +93,8 @@ int Node::Depth(int begin) {
Node &Node::Folder(
int size, std::string type,
std::map<std::string, std::pair<std::string, std::string>> change,
std::map<std::string, std::vector<std::pair<std::string, std::string>>>
change,
std::vector<std::shared_ptr<Node>> *removed_nodes) {
std::shared_ptr<framework::OpDesc> op_desc =
std::make_shared<framework::OpDesc>();
......@@ -109,12 +111,15 @@ Node &Node::Folder(
void Node::Folder(
std::shared_ptr<framework::OpDesc> op_desc,
std::vector<std::shared_ptr<Node>> *outputs, int index,
std::map<std::string, std::pair<std::string, std::string>> *change,
std::map<std::string, std::vector<std::pair<std::string, std::string>>>
*change,
Node *begin_node, std::vector<std::shared_ptr<Node>> *removed_nodes) {
if (change->find(this->type_) != change->end()) {
auto change_pair = (*change)[this->type_];
op_desc->GetInputs()[change_pair.second] =
this->op_desc_->GetInputs()[change_pair.first];
auto change_pairs = (*change)[this->type_];
for (const auto &change_pair : change_pairs) {
op_desc->GetInputs()[change_pair.second] =
this->op_desc_->GetInputs()[change_pair.first];
}
}
for (auto &attr_pair : this->op_desc_->attrs_) {
......
src/framework/program/program-optimize/node.h
浏览文件 @ c67b701f
......@@ -17,6 +17,7 @@ limitations under the License. */
#include <cinttypes>
#include <map>
#include <string>
#include <utility>
#include <vector>
#include "common/log.h"
#include "framework/program/op_desc.h"
......@@ -43,7 +44,8 @@ class Node {
int Depth(int begin = 0);
Node &Folder(
int size, std::string type,
std::map<std::string, std::pair<std::string, std::string>> change_map,
std::map<std::string, std::vector<std::pair<std::string, std::string>>>
change,
std::vector<std::shared_ptr<Node>> *removed_nodes);
std::vector<std::shared_ptr<framework::OpDesc>> OpDescs(int size);
std::shared_ptr<framework::OpDesc> OpDescOfNode() { return op_desc_; }
......@@ -56,7 +58,8 @@ class Node {
void Folder(
std::shared_ptr<framework::OpDesc> op_desc,
std::vector<std::shared_ptr<Node>> *outputs, int index,
std::map<std::string, std::pair<std::string, std::string>> *change,
std::map<std::string, std::vector<std::pair<std::string, std::string>>>
*change,
Node *begin_node, std::vector<std::shared_ptr<Node>> *removed_nodes);
std::shared_ptr<framework::OpDesc> op_desc_;
#ifdef PADDLE_MOBILE_DEBUG
......
src/framework/program/program-optimize/program_optimize.cpp
浏览文件 @ c67b701f
......@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "framework/program/program-optimize/program_optimize.h"
#include <algorithm>
#include "framework/program/program-optimize/fusion_op_register.h"
namespace paddle_mobile {
......
src/framework/program/program.h
浏览文件 @ c67b701f
......@@ -29,7 +29,7 @@ class Program {
std::shared_ptr<Scope> scope;
std::string model_path;
std::string para_path;
bool is_commbine = false;
bool combined = false;
private:
};
......
src/framework/scope.cpp
浏览文件 @ c67b701f
......@@ -14,6 +14,7 @@ limitations under the License. */
#include "framework/scope.h"
#include <algorithm>
#include <set>
#include <string>
#include <vector>
......
src/framework/scope.h
浏览文件 @ c67b701f
......@@ -23,7 +23,17 @@ namespace framework {
class Scope {
public:
Scope() = default;
~Scope() = default;
~Scope() {
for (auto &var : vars_) {
delete var.second;
}
vars_.clear();
for (auto kid : kids_) {
delete kid;
}
kids_.clear();
}
Scope &NewScope() const;
......
src/framework/tensor.h
浏览文件 @ c67b701f
......@@ -22,6 +22,7 @@ limitations under the License. */
#include <vector>
#include "common/enforce.h"
#include <fstream>
#include "common/enforce.h"
#include "framework/data_layout.h"
#include "framework/ddim.h"
......@@ -131,6 +132,22 @@ class Tensor {
return reinterpret_cast<T *>(mutable_data(typeid(T)));
}
#ifdef PADDLE_MOBILE_DEBUG
template <typename T>
inline void dump(std::string filename) const {
const T *dataptr = data<T>();
std::ofstream out(filename.c_str());
for (int i = 0; i < numel(); ++i) {
out << dataptr[i] << " ";
}
out << "形状：";
for (int j = 0; j < dims_.size(); ++j) {
out << dims_[j] << " ";
}
out.close();
}
#endif
inline void *mutable_data(std::type_index type) {
if (holder_ != nullptr) {
holder_->set_type(type);
......
src/io/io.cpp → src/io/executor.cpp
浏览文件 @ c67b701f
......@@ -12,7 +12,8 @@ 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 "io/io.h"
#include "io/executor.h"
#include <algorithm>
#include <vector>
#include "common/enforce.h"
#include "common/log.h"
......@@ -25,7 +26,6 @@ limitations under the License. */
#include "framework/scope.h"
#include "framework/tensor.h"
#ifdef PADDLE_EXECUTOR_MULTITHREAD
#include <algorithm>
#include <queue>
#include <utility>
#include "common/threadpool.h"
......@@ -39,7 +39,7 @@ char *Get_binary_data(std::string filename) {
PADDLE_MOBILE_ENFORCE(file != nullptr, "can't open file: %s ",
filename.c_str());
fseek(file, 0, SEEK_END);
long size = ftell(file);
int64_t size = ftell(file);
PADDLE_MOBILE_ENFORCE(size > 0, "size is too small");
rewind(file);
char *data = new char[size];
......@@ -50,116 +50,6 @@ char *Get_binary_data(std::string filename) {
return data;
}
static size_t ReadBuffer(const char *file_name, uint8_t **out) {
printf("%s \n", file_name);
FILE *fp;
fp = fopen(file_name, "rb");
PADDLE_MOBILE_ENFORCE(fp != NULL, " %s open failed !", file_name);
fseek(fp, 0, SEEK_END);
size_t size = ftell(fp);
rewind(fp);
DLOG << "model size: " << size;
*out = reinterpret_cast<uint8_t *>(malloc(size));
size_t cur_len = 0;
size_t nread;
while ((nread = fread(*out + cur_len, 1, size - cur_len, fp)) != 0) {
cur_len += nread;
}
fclose(fp);
return cur_len;
}
template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::Load(
const std::string &dirname, bool optimize, bool can_add_split) {
auto program =
this->LoadProgram(dirname + "/__model__", optimize, can_add_split);
program.model_path = dirname;
return program;
}
template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::Load(
const std::string &model_path, const std::string &para_path,
bool optimize) {
auto program = this->LoadProgram(model_path, optimize);
program.para_path = para_path;
program.is_commbine = true;
return program;
}
template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::LoadProgram(
const std::string &model_path, bool optimize, bool can_add_split) {
std::string model_filename = model_path;
PaddleMobile__Framework__Proto__ProgramDesc *c_program;
uint8_t *buf = NULL;
size_t read_size = ReadBuffer(model_filename.c_str(), &buf);
PADDLE_MOBILE_ENFORCE(buf != NULL, "read from __model__ is null");
c_program = paddle_mobile__framework__proto__program_desc__unpack(
NULL, read_size, buf);
//
PADDLE_MOBILE_ENFORCE(c_program != NULL, "program is null");
//
DLOG << "n_ops: " << (*c_program->blocks)->n_ops;
//
auto originProgramDesc = std::make_shared<framework::ProgramDesc>(c_program);
framework::Program<Dtype, P> program;
program.originProgram = originProgramDesc;
auto scope = std::make_shared<framework::Scope>();
program.scope = scope;
for (const auto &block : originProgramDesc->Blocks()) {
for (auto var_desc : block->Vars()) {
auto var = scope->Var(var_desc->Name());
if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
if (var_desc->Persistable() &&
var_desc->Type() != framework::VARTYPE_TYPE_FEED_MINIBATCH &&
var_desc->Type() != framework::VARTYPE_TYPE_FETCH_LIST) {
auto dim = var_desc->Tensor_desc().Dims();
auto tensor = var->GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim(dim));
} else {
auto dim = var_desc->Tensor_desc().Dims();
PADDLE_MOBILE_ENFORCE(dim.size() > 0, "dim size is 0");
dim[0] = 1;
auto tensor = var->GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim(dim));
}
} else {
// TODO(codeWorm): some.
}
}
}
if (optimize) {
framework::ProgramOptimize program_optimize;
program.optimizeProgram =
program_optimize.FusionOptimize(originProgramDesc, can_add_split);
}
if (optimize) {
program.optimizeProgram->Description("optimize: ");
} else {
originProgramDesc->Description("program: ");
}
paddle_mobile__framework__proto__program_desc__free_unpacked(c_program, NULL);
return program;
}
template class Loader<CPU, Precision::FP32>;
template class Loader<FPGA, Precision::FP32>;
template class Loader<GPU_MALI, Precision::FP32>;
#pragma mark - executor
template <typename Dtype, Precision P>
Executor<Dtype, P>::Executor(const framework::Program<Dtype> p, int batch_size,
......@@ -193,7 +83,7 @@ Executor<Dtype, P>::Executor(const framework::Program<Dtype> p, int batch_size,
#endif
}
}
if (program_.is_commbine) {
if (program_.combined) {
InitCombineMemory();
} else {
InitMemory();
......@@ -209,30 +99,30 @@ Executor<Dtype, P>::Executor(const framework::Program<Dtype> p, int batch_size,
template <typename Dtype, Precision P>
void Executor<Dtype, P>::LoadMemory(const framework::VarDesc var_desc,
framework::LoDTensor *tensor, char *&data) {
framework::LoDTensor *tensor, char **data) {
// 1. version
uint32_t version = *(uint32_t *)data;
data += sizeof(uint32_t);
uint32_t version = *reinterpret_cast<uint32_t *>(*data);
(*data) += sizeof(uint32_t);
// 2 Lod information
uint64_t *lod_level_ptr = new uint64_t();
memcpy(lod_level_ptr, data, sizeof(uint64_t));
memcpy(lod_level_ptr, (*data), sizeof(uint64_t));
uint64_t lod_level = *lod_level_ptr;
delete lod_level_ptr;
data += sizeof(uint64_t);
(*data) += sizeof(uint64_t);
auto &lod = *tensor->mutable_lod();
lod.resize(lod_level);
for (uint64_t i = 0; i < lod_level; ++i) {
uint64_t size = *(uint64_t *)data;
data += sizeof(uint64_t);
uint64_t size = *reinterpret_cast<uint64_t *>(*data);
(*data) += sizeof(uint64_t);
DLOG << "lod size: " << i << size;
std::vector<size_t> tmp(size / sizeof(size_t));
for (int k = 0; k < tmp.size(); ++k) {
tmp[k] = *(size_t *)data;
DLOG << "tmp[k]: " << k << *(size_t *)data;
data += sizeof(size_t);
tmp[k] = *reinterpret_cast<size_t *>(*data);
(*data) += sizeof(size_t);
}
for (auto j : tmp) {
......@@ -242,18 +132,18 @@ void Executor<Dtype, P>::LoadMemory(const framework::VarDesc var_desc,
}
// 3. tensor version
uint32_t tensor_version = *(uint32_t *)data;
data += sizeof(uint32_t);
uint32_t tensor_version = *reinterpret_cast<uint32_t *>(*data);
(*data) += sizeof(uint32_t);
// 4. tensor desc
int32_t size = *(int32_t *)data;
data += sizeof(int32_t);
int32_t size = *reinterpret_cast<int32_t *>(*data);
(*data) += sizeof(int32_t);
std::unique_ptr<char[]> buf(new char[size]);
for (int m = 0; m < size; ++m) {
buf.get()[m] = data[m];
buf.get()[m] = (*data)[m];
}
data += (sizeof(char) * size);
(*data) += (sizeof(char) * size);
const framework::TensorDesc &desc = var_desc.Tensor_desc();
int memory_size = 1;
......@@ -290,9 +180,9 @@ void Executor<Dtype, P>::LoadMemory(const framework::VarDesc var_desc,
}
for (int n = 0; n < memory_size * type_size; ++n) {
static_cast<char *>(memory)[n] = data[n];
static_cast<char *>(memory)[n] = (*data)[n];
}
data += (sizeof(char) * memory_size * type_size);
(*data) += (sizeof(char) * memory_size * type_size);
}
template <typename Dtype, Precision P>
......@@ -309,7 +199,7 @@ void Executor<Dtype, P>::InitMemory() {
char *origin_data =
Get_binary_data(program_.model_path + "/" + var_desc->Name());
char *data = origin_data;
LoadMemory(*var_desc, tensor, data);
LoadMemory(*var_desc, tensor, &data);
delete origin_data;
} else {
if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
......@@ -335,7 +225,7 @@ void Executor<Dtype, P>::InitCombineMemory() {
if (var_desc->Name() == "feed" || var_desc->Name() == "fetch") {
continue;
}
LoadMemory(*var_desc, tensor, data);
LoadMemory(*var_desc, tensor, &data);
} else {
if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
auto tensor = var->template GetMutable<framework::LoDTensor>();
......@@ -442,7 +332,8 @@ std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
*(program_.scope));
#ifdef PADDLE_MOBILE_PROFILE
#ifdef PADDLE_EXECUTOR_MULTITHREAD
// TODO expose profile info as an interface, user can get them to analysis
// TODO(haipeng): expose profile info as an interface, user can get them to
// analysis
// the performance of their deepnet.
FILE *df = fopen("net.dot", "w");
fprintf(df, "digraph {\n");
......@@ -457,16 +348,19 @@ std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
fprintf(df, "}\n");
fclose(df);
#endif
FILE *pf = fopen("profile.out", "w");
// FILE *pf = fopen("profile.out", "w");
std::unordered_map<std::string, uint64_t> _tp;
for (int i = 0; i < profile.size(); i++) {
const auto &pInfo = profile[i];
uint64_t timeCost = pInfo.runEnd - pInfo.runBegin;
_tp[ops[i]->Type()] += timeCost;
fprintf(pf, "%d\t%s\t%d\t%llu\t%llu\t%llu\n", i, ops[i]->Type().c_str(),
pInfo.tid, pInfo.runBegin, pInfo.runEnd, timeCost);
// fprintf(pf, "%d\t%s\t%d\t%llu\t%llu\t%llu\n", i,
// ops[i]->Type().c_str(),
// pInfo.tid, pInfo.runBegin, pInfo.runEnd, timeCost);
}
fclose(pf);
// fclose(pf);
printf("====================[ profile ]======================\n");
using prof_t = std::pair<std::string, uint64_t>;
std::vector<prof_t> _tv(_tp.begin(), _tp.end());
......@@ -480,8 +374,9 @@ std::shared_ptr<framework::Tensor> Executor<Dtype, P>::Predict(
std::sort(_tv.begin(), _tv.end(), compf);
_tv.push_back(std::make_pair("total", _ptotal));
for (auto const &p : _tv) {
printf("%-16s\t%-10.0f\t%-2.4f\n", p.first.c_str(), (float)p.second,
(float)p.second / _ptotal * 100.0);
printf("%-16s\t%-10.0f\t%-2.4f\n", p.first.c_str(),
static_cast<float>(p.second),
static_cast<float>(p.second) / _ptotal * 100.0);
}
printf("====================[---------]======================\n");
#endif
......
src/io/io.h → src/io/executor.h
浏览文件 @ c67b701f
......@@ -18,6 +18,7 @@ limitations under the License. */
#include <memory>
#include <string>
#include <vector>
#include "common/types.h"
#include "framework/lod_tensor.h"
#include "framework/operator.h"
......@@ -27,36 +28,11 @@ limitations under the License. */
#include <condition_variable>
#include <mutex>
#include <thread>
#include "common/depCore.h"
#include "common/dep_core.h"
#endif
namespace paddle_mobile {
template <typename Dtype = CPU, Precision P = Precision::FP32>
class Loader {
public:
/*
* @b load separate format fluid model
* @b 加载分开形式的 fluid 模型
* */
const framework::Program<Dtype, P> Load(const std::string &dirname,
bool optimize = false,
bool can_add_split = false);
/*
* @b load combine format fluid mode
* @b 加载结合在一起格式的模型
* */
const framework::Program<Dtype, P> Load(const std::string &model_path,
const std::string &para_path,
bool optimize = false);
private:
const framework::Program<Dtype, P> LoadProgram(const std::string &model_path,
bool optimize = false,
bool can_add_split = false);
};
template <typename Dtype = CPU, Precision P = Precision::FP32>
class Executor {
public:
......@@ -86,7 +62,7 @@ class Executor {
Executor() = default;
void InitMemory();
void LoadMemory(const framework::VarDesc var_desc,
framework::LoDTensor *tensor, char *&data);
framework::LoDTensor *tensor, char **data);
void InitCombineMemory();
framework::Program<Dtype> program_;
int batch_size_ = 1;
......
src/io/loader.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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 "io/loader.h"
#include "framework/lod_tensor.h"
#include "framework/program/program-optimize/program_optimize.h"
namespace paddle_mobile {
using framework::Variable;
static size_t ReadBuffer(const char *file_name, uint8_t **out) {
printf("%s \n", file_name);
FILE *fp;
fp = fopen(file_name, "rb");
PADDLE_MOBILE_ENFORCE(fp != NULL, " %s open failed !", file_name);
fseek(fp, 0, SEEK_END);
size_t size = ftell(fp);
rewind(fp);
DLOG << "model size: " << size;
*out = reinterpret_cast<uint8_t *>(malloc(size));
size_t cur_len = 0;
size_t nread;
while ((nread = fread(*out + cur_len, 1, size - cur_len, fp)) != 0) {
cur_len += nread;
}
fclose(fp);
return cur_len;
}
template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::Load(
const std::string &dirname, bool optimize, bool can_add_split) {
auto program =
this->LoadProgram(dirname + "/__model__", optimize, can_add_split);
program.model_path = dirname;
return program;
}
template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::Load(
const std::string &model_path, const std::string &para_path,
bool optimize) {
auto program = this->LoadProgram(model_path, optimize);
program.para_path = para_path;
program.combined = true;
return program;
}
template <typename Dtype, Precision P>
const framework::Program<Dtype, P> Loader<Dtype, P>::LoadProgram(
const std::string &model_path, bool optimize, bool can_add_split) {
std::string model_filename = model_path;
PaddleMobile__Framework__Proto__ProgramDesc *c_program;
uint8_t *buf = NULL;
size_t read_size = ReadBuffer(model_filename.c_str(), &buf);
PADDLE_MOBILE_ENFORCE(buf != NULL, "read from __model__ is null");
c_program = paddle_mobile__framework__proto__program_desc__unpack(
NULL, read_size, buf);
//
PADDLE_MOBILE_ENFORCE(c_program != NULL, "program is null");
//
DLOG << "n_ops: " << (*c_program->blocks)->n_ops;
//
auto originProgramDesc = std::make_shared<framework::ProgramDesc>(c_program);
framework::Program<Dtype, P> program;
program.originProgram = originProgramDesc;
auto scope = std::make_shared<framework::Scope>();
program.scope = scope;
for (const auto &block : originProgramDesc->Blocks()) {
for (auto var_desc : block->Vars()) {
auto var = scope->Var(var_desc->Name());
if (var_desc->Type() == framework::VARTYPE_TYPE_LOD_TENSOR) {
if (var_desc->Persistable() &&
var_desc->Type() != framework::VARTYPE_TYPE_FEED_MINIBATCH &&
var_desc->Type() != framework::VARTYPE_TYPE_FETCH_LIST) {
auto dim = var_desc->Tensor_desc().Dims();
auto tensor = var->GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim(dim));
} else {
auto dim = var_desc->Tensor_desc().Dims();
PADDLE_MOBILE_ENFORCE(dim.size() > 0, "dim size is 0");
dim[0] = 1;
auto tensor = var->GetMutable<framework::LoDTensor>();
tensor->Resize(framework::make_ddim(dim));
}
} else {
// TODO(codeWorm): some.
}
}
}
if (optimize) {
framework::ProgramOptimize program_optimize;
program.optimizeProgram =
program_optimize.FusionOptimize(originProgramDesc, can_add_split);
}
if (optimize) {
program.optimizeProgram->Description("optimize: ");
} else {
originProgramDesc->Description("program: ");
}
paddle_mobile__framework__proto__program_desc__free_unpacked(c_program, NULL);
return program;
}
template class Loader<CPU, Precision::FP32>;
template class Loader<FPGA, Precision::FP32>;
template class Loader<GPU_MALI, Precision::FP32>;
} // namespace paddle_mobile
src/io/loader.h 0 → 100644
浏览文件 @ c67b701f
/* 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 "common/types.h"
#include "framework/program/program.h"
namespace paddle_mobile {
template <typename Dtype = CPU, Precision P = Precision::FP32>
class Loader {
public:
/*
* @b load separate format fluid model
* @b 加载分开形式的 fluid 模型
* */
const framework::Program<Dtype, P> Load(const std::string &dirname,
bool optimize = false,
bool can_add_split = false);
/*
* @b load combine format fluid mode
* @b 加载结合在一起格式的模型
* */
const framework::Program<Dtype, P> Load(const std::string &model_path,
const std::string &para_path,
bool optimize = false);
private:
const framework::Program<Dtype, P> LoadProgram(const std::string &model_path,
bool optimize = false,
bool can_add_split = false);
};
} // namespace paddle_mobile
src/io/paddle_mobile.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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 "io/paddle_mobile.h"
namespace paddle_mobile {
template <typename Dtype, Precision P>
bool PaddleMobile<Dtype, P>::Load(const std::string &dirname, bool optimize,
int batch_size) {
if (loader_.get() == nullptr) {
loader_ = std::make_shared<Loader<Dtype, P>>();
} else {
LOG(kLOG_INFO) << "loader inited";
}
if (executor_.get() == nullptr) {
executor_ = std::make_shared<Executor<Dtype, P>>(
loader_->Load(dirname, optimize), batch_size, optimize);
} else {
LOG(kLOG_INFO) << "executor inited";
}
return true;
}
template <typename Dtype, Precision P>
bool PaddleMobile<Dtype, P>::Load(const std::string &model_path,
const std::string &para_path, bool optimize,
int batch_size) {
if (loader_.get() == nullptr) {
loader_ = std::make_shared<Loader<Dtype, P>>();
} else {
LOG(kLOG_INFO) << "loader inited";
}
if (executor_.get() == nullptr) {
executor_ = std::make_shared<Executor<Dtype, P>>(
loader_->Load(model_path, para_path, optimize), batch_size, optimize);
} else {
LOG(kLOG_INFO) << "executor inited";
}
return true;
}
template <typename Dtype, Precision P>
std::shared_ptr<framework::Tensor> PaddleMobile<Dtype, P>::Predict(
const framework::Tensor &t) {
return executor_->Predict(t);
}
template <typename Dtype, Precision P>
std::vector<typename PaddleMobile<Dtype, P>::Ptype>
PaddleMobile<Dtype, P>::Predict(const std::vector<Ptype> &input,
const std::vector<int64_t> &dims) {
return executor_->Predict(input, dims);
}
template <typename Dtype, Precision P>
void PaddleMobile<Dtype, P>::Clear() {
executor_ = nullptr;
loader_ = nullptr;
}
template <typename Dtype, Precision P>
PaddleMobile<Dtype, P>::~PaddleMobile() {
executor_ = nullptr;
loader_ = nullptr;
}
template class PaddleMobile<CPU, Precision::FP32>;
template class PaddleMobile<FPGA, Precision::FP32>;
template class PaddleMobile<GPU_MALI, Precision::FP32>;
} // namespace paddle_mobile
src/io/paddle_mobile.h 0 → 100644
浏览文件 @ c67b701f
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <memory>
#include <string>
#include <vector>
#include "common/types.h"
#include "framework/tensor.h"
#include "io/executor.h"
#include "io/loader.h"
namespace paddle_mobile {
template <typename Dtype = CPU, Precision P = Precision::FP32>
class PaddleMobile {
typedef typename PrecisionTrait<P>::ptype Ptype;
public:
PaddleMobile() {}
/*
* @b load separate format fluid model
* @b 加载分开形式的 fluid 模型
* */
bool Load(const std::string &dirname, bool optimize = false,
int batch_size = 1);
/*
* @b load combine format fluid mode
* @b 加载结合在一起格式的模型
* */
bool Load(const std::string &model_path, const std::string &para_path,
bool optimize = false, int batch_size = 1);
/*
* @b to predict
* */
std::shared_ptr<framework::Tensor> Predict(const framework::Tensor &t);
/*
* @b to predict with vector and dim
*
* @b 使用 输入 和 输入的维度信息 进行预测
* */
std::vector<Ptype> Predict(const std::vector<Ptype> &input,
const std::vector<int64_t> &dims);
void Clear();
~PaddleMobile();
private:
std::shared_ptr<Loader<Dtype, P>> loader_;
std::shared_ptr<Executor<Dtype, P>> executor_;
};
} // namespace paddle_mobile
src/ios_io/PaddleMobile.h 0 → 100644
浏览文件 @ c67b701f
/* 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
#import <CoreImage/CoreImage.h>
#import <Foundation/Foundation.h>
@interface PaddleMobile : NSObject
- (instancetype)init;
- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath;
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale;
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim;
- (void)clear;
@end
src/ios_io/PaddleMobile.mm 0 → 100644
浏览文件 @ c67b701f
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#import "PaddleMobile.h"
#import "op_symbols.h"
#import "io/paddle_mobile.h"
#import <memory>
#import <vector>
@interface PaddleMobile()
{
paddle_mobile::PaddleMobile<paddle_mobile::CPU, paddle_mobile::Precision::FP32> *pam_;
BOOL loaded_;
}
@end
@implementation PaddleMobile
static std::mutex shared_mutex;
- (instancetype)init {
if (self = [super init]) {
pam_ = new paddle_mobile::PaddleMobile<paddle_mobile::CPU, paddle_mobile::Precision::FP32>();
}
return self;
}
- (void)dealloc {
if (pam_) {
delete pam_;
}
}
+ (instancetype)sharedInstance{
static dispatch_once_t onceToken;
static id sharedManager = nil;
dispatch_once(&onceToken, ^{
sharedManager = [[[self class] alloc] init];
});
return sharedManager;
}
- (BOOL)load:(NSString *)modelPath andWeightsPath:(NSString *)weighsPath{
std::string model_path_str = std::string([modelPath UTF8String]);
std::string weights_path_str = std::string([weighsPath UTF8String]);
if (loaded_ = pam_->Load(model_path_str, weights_path_str, false)) {
return YES;
} else {
return NO;
}
}
-(void)preprocess:(const UInt8 *)input output:(float *)output imageWidth:(int)imageWidth imageHeight:(int)imageHeight imageChannels:(int)imageChannels means:(NSArray<NSNumber *> *)means scale:(float)scale dim:(std::vector<int64_t>)dim{
if (means == nil) {
means = @[@0, @0, @0];
}
int wanted_input_width = dim[3];
int wanted_input_height = dim[2];
int wanted_input_channels = dim[1];
for (int c = 0; c < wanted_input_channels; ++c) {
float *out_channel = output + c * wanted_input_height * wanted_input_width;
for (int y = 0; y < wanted_input_height; ++y) {
float *out_row = out_channel + y * wanted_input_width;
for (int x = 0; x < wanted_input_width; ++x) {
int in_row = (y * imageHeight) / wanted_input_height;
int in_col = (x * imageWidth) / wanted_input_width;
const UInt8 *in_pixel = input + (in_row * imageWidth * imageChannels) + (in_col * imageChannels);
float *out_pos = out_row + x;
if (c == 0) {
*out_pos = (in_pixel[c] - means[c].floatValue) * scale;
}else if (c == 1){
*out_pos = (in_pixel[c] - means[c].floatValue) * scale;
}else if (c == 2){
*out_pos = (in_pixel[c] - means[c].floatValue) * scale;
}
}
}
}
}
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim means:(NSArray<NSNumber *> *)means scale:(float)scale{
std::lock_guard<std::mutex> lock(shared_mutex);
if (!loaded_) {
printf("PaddleMobile doesn't be loaded yet");
return nil;
}
if (dim.count != 4) {
printf("dim must have 4 elements");
return nil;
}
// dim to c++ vector, get numel
std::vector<int64_t > dim_vec;
int numel = 1;
for (int k = 0; k < dim.count; ++k) {
int d = dim[k].intValue;
numel *= d;
dim_vec.push_back(d);
}
const int sourceRowBytes = CGImageGetBytesPerRow(image);
const int image_width = CGImageGetWidth(image);
const int image_height = CGImageGetHeight(image);
const int image_channels = 4;
CGDataProviderRef provider = CGImageGetDataProvider(image);
CFDataRef cfData = CGDataProviderCopyData(provider);
const UInt8 *input = CFDataGetBytePtr(cfData);
// sample image
float *output = (float *)malloc(numel*sizeof(float));
[self preprocess:input output:output imageWidth:image_width imageHeight:image_height imageChannels:image_channels means:means scale:scale dim:dim_vec];
float *dataPointer = nullptr;
if (nullptr != output) {
dataPointer = output;
} else {
return nil;
}
// input
std::vector<float> predict_input;
for (int j = 0; j < numel; ++j) {
predict_input.push_back(dataPointer[j]);
}
// predict
std::vector<float> cpp_result = pam_->Predict(predict_input, dim_vec);
// result
long count = 0;
count = cpp_result.size();
NSMutableArray *result = [[NSMutableArray alloc] init];
for (int i = 0; i < count; i++) {
[result addObject:[NSNumber numberWithFloat:cpp_result[i]]];
}
free(output);
// 待验证
// if ([UIDevice currentDevice].systemVersion.doubleValue < 11.0) {
CFRelease(cfData);
cfData = NULL;
// }
return result;
}
- (NSArray *)predict:(CGImageRef)image dim:(NSArray<NSNumber *> *)dim {
[self predict:image dim:dim means:nil scale:1];
}
- (void)clear{
pam_->Clear();
}
@end
src/ios_io/op_symbols.h 0 → 100644
浏览文件 @ c67b701f
/* 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 "operators/batchnorm_op.h"
#include "operators/box_coder_op.h"
#include "operators/concat_op.h"
#include "operators/conv_op.h"
#include "operators/depthwise_conv_op.h"
#include "operators/dropout_op.h"
#include "operators/elementwise_add_op.h"
#include "operators/feed_op.h"
#include "operators/fetch_op.h"
#include "operators/fusion_conv_add.h"
#include "operators/fusion_conv_add_bn_relu_op.h"
#include "operators/fusion_fc_op.h"
#include "operators/im2sequence_op.h"
#include "operators/lrn_op.h"
#include "operators/mul_op.h"
#include "operators/multiclass_nms_op.h"
#include "operators/pool_op.h"
#include "operators/prior_box_op.h"
#include "operators/relu_op.h"
#include "operators/reshape_op.h"
#include "operators/sigmoid_op.h"
#include "operators/softmax_op.h"
#include "operators/transpose_op.h"
src/jni/paddle_mobile_jni.cpp
浏览文件 @ c67b701f
......@@ -15,6 +15,10 @@ limitations under the License. */
#ifdef ANDROID
#include "paddle_mobile_jni.h"
#include "common/log.h"
#include "framework/tensor.h"
#include "io/paddle_mobile.h"
#ifdef __cplusplus
extern "C" {
#endif
......@@ -28,17 +32,16 @@ using std::string;
extern const char *ANDROID_LOG_TAG =
"paddle_mobile LOG built on " __DATE__ " " __TIME__;
static Executor<CPU> *shared_executor_instance = nullptr;
static PaddleMobile<CPU> *shared_paddle_mobile_instance = nullptr;
// toDo mutex lock
// static std::mutex shared_mutex;
Executor<CPU> *getExecutorInstance(const Program<CPU> p, int batch_size,
bool use_optimize) {
if (nullptr == shared_executor_instance) {
shared_executor_instance = new Executor<CPU>(p, batch_size, use_optimize);
PaddleMobile<CPU> *getPaddleMobileInstance() {
if (nullptr == shared_paddle_mobile_instance) {
shared_paddle_mobile_instance = new PaddleMobile<CPU>();
}
return shared_executor_instance;
return shared_paddle_mobile_instance;
}
string jstring2cppstring(JNIEnv *env, jstring jstr) {
......@@ -51,15 +54,14 @@ string jstring2cppstring(JNIEnv *env, jstring jstr) {
JNIEXPORT jboolean JNICALL Java_com_baidu_paddle_PML_load(JNIEnv *env,
jclass thiz,
jstring modelPath) {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
ANDROIDLOGI("load invoked");
bool optimize = true;
auto program = loader.Load(jstring2cppstring(env, modelPath), optimize);
shared_executor_instance = getExecutorInstance(program, 1, optimize);
return shared_executor_instance != nullptr ? JNI_TRUE : JNI_FALSE;
return getPaddleMobileInstance()->Load(jstring2cppstring(env, modelPath),
optimize);
}
JNIEXPORT jfloatArray JNICALL Java_com_baidu_paddle_PML_predictImage(
JNIEnv *env, jclass thiz, jfloatArray buf) {
JNIEXPORT jfloatArray JNICALL
Java_com_baidu_paddle_PML_predict(JNIEnv *env, jclass thiz, jfloatArray buf) {
jfloatArray result = NULL;
int count = 0;
float *dataPointer = nullptr;
......@@ -73,15 +75,18 @@ JNIEXPORT jfloatArray JNICALL Java_com_baidu_paddle_PML_predictImage(
for (int i = 0; i < framework::product(ddim); i++) {
input_ptr[i] = dataPointer[i];
}
auto output = shared_executor_instance->Predict(input);
auto output = shared_paddle_mobile_instance->Predict(input);
count = output->numel();
result = env->NewFloatArray(count);
env->SetFloatArrayRegion(result, 0, count, output->data<float>());
ANDROIDLOGI("predict finished");
return result;
}
JNIEXPORT void JNICALL Java_com_baidu_paddle_PML_clear(JNIEnv *env,
jclass thiz) {}
jclass thiz) {
getPaddleMobileInstance()->Clear();
}
} // namespace jni
} // namespace paddle_mobile
......
src/jni/paddle_mobile_jni.h
浏览文件 @ c67b701f
......@@ -15,9 +15,6 @@ limitations under the License. */
#pragma once
#ifdef ANDROID
#include <jni.h>
#include "common/log.h"
#include "framework/tensor.h"
#include "io/io.h"
#ifdef __cplusplus
extern "C" {
......@@ -34,8 +31,8 @@ JNIEXPORT jboolean JNICALL Java_com_baidu_paddle_PML_load(JNIEnv *env,
/**
* object detection for anroid
*/
JNIEXPORT jfloatArray JNICALL Java_com_baidu_paddle_PML_predictImage(
JNIEnv *env, jclass thiz, jfloatArray buf);
JNIEXPORT jfloatArray JNICALL
Java_com_baidu_paddle_PML_predict(JNIEnv *env, jclass thiz, jfloatArray buf);
/**
* clear data of the net when destroy for android
......
src/memory/t_malloc.cpp
浏览文件 @ c67b701f
......@@ -12,8 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "memory/t_malloc.h"
#include <cstdlib>
#include <cstring>
......
src/operators/batchnorm_op.cpp
浏览文件 @ c67b701f
......@@ -32,7 +32,6 @@ template class BatchNormOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(batch_norm);
REGISTER_OPERATOR_CPU(batch_norm, ops::BatchNormOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......
src/operators/batchnorm_op.h
浏览文件 @ c67b701f
......@@ -45,4 +45,13 @@ class BatchNormOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(batch_norm);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(batch_norm);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/box_coder_op.cpp
浏览文件 @ c67b701f
......@@ -53,7 +53,6 @@ template class BoxCoderOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(box_coder);
REGISTER_OPERATOR_CPU(box_coder, ops::BoxCoderOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......
src/operators/box_coder_op.h
浏览文件 @ c67b701f
......@@ -51,4 +51,12 @@ class BoxCoderOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(box_coder);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/concat_op.cpp
浏览文件 @ c67b701f
......@@ -63,11 +63,9 @@ template class ConcatOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(concat);
REGISTER_OPERATOR_CPU(concat, ops::ConcatOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(concat);
REGISTER_OPERATOR_MALI_GPU(concat, ops::ConcatOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/concat_op.h
浏览文件 @ c67b701f
......@@ -46,4 +46,13 @@ class ConcatOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(concat);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(concat);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/conv_op.cpp
浏览文件 @ c67b701f
......@@ -55,15 +55,12 @@ template class ConvOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(conv2d);
REGISTER_OPERATOR_CPU(conv2d, ops::ConvOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(conv2d);
REGISTER_OPERATOR_MALI_GPU(conv2d, ops::ConvOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(conv2d);
REGISTER_OPERATOR_FPGA(conv2d, ops::ConvOp);
#endif
......
src/operators/conv_op.h
浏览文件 @ c67b701f
......@@ -46,4 +46,14 @@ class ConvOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(conv2d);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(conv2d);
#endif
#ifdef PADDLE_MOBILE_FPGA
USE_OP_FPGA(conv2d);
#endif
#endif
src/operators/depthwise_conv_op.cpp
浏览文件 @ c67b701f
......@@ -56,7 +56,6 @@ template class DepthwiseConvOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(depthwise_conv2d);
REGISTER_OPERATOR_CPU(depthwise_conv2d, ops::DepthwiseConvOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......
src/operators/depthwise_conv_op.h
浏览文件 @ c67b701f
......@@ -48,4 +48,12 @@ class DepthwiseConvOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(depthwise_conv2d);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/dropout_op.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef DROPOUT_OP
#include "operators/dropout_op.h"
namespace paddle_mobile {
namespace operators {
template <typename Dtype, typename T>
void DropoutOp<Dtype, T>::InferShape() const {
auto input_dims = this->param_.InputX()->dims();
this->param_.Out()->Resize(input_dims);
}
template class DropoutOp<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(dropout, ops::DropoutOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/dropout_op.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef DROPOUT_OP
#pragma once
#include <string>
#include "framework/operator.h"
#include "operators/kernel/dropout_kernel.h"
#include "operators/op_param.h"
namespace paddle_mobile {
namespace operators {
using paddle_mobile::framework::Tensor;
template <typename DeviceType, typename T>
class DropoutOp
: public framework::OperatorWithKernel<
DeviceType, DropoutParam, operators::DropoutKernel<DeviceType, T>> {
public:
DropoutOp(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs, const framework::AttributeMap attrs,
std::shared_ptr<framework::Scope> scope)
: framework::OperatorWithKernel<DeviceType, DropoutParam,
operators::DropoutKernel<DeviceType, T>>(
type, inputs, outputs, attrs, scope) {}
// using framework::OperatorWithKernel<DeviceType, DropoutParam,
// operators::DropoutKernel<DeviceType,
// T>>;
void InferShape() const override;
protected:
};
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(dropout);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/elementwise_add_op.cpp
浏览文件 @ c67b701f
......@@ -30,11 +30,9 @@ template class ElementwiseAddOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(elementwise_add);
REGISTER_OPERATOR_CPU(elementwise_add, ops::ElementwiseAddOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(elementwise_add);
REGISTER_OPERATOR_MALI_GPU(elementwise_add, ops::ElementwiseAddOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/elementwise_add_op.h
浏览文件 @ c67b701f
......@@ -48,4 +48,13 @@ class ElementwiseAddOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(elementwise_add);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(elementwise_add);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/feed_op.cpp
浏览文件 @ c67b701f
......@@ -19,3 +19,14 @@ namespace operators {
template class FeedOp<CPU, float>;
}
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(feed, ops::FeedOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
REGISTER_OPERATOR_MALI_GPU(feed, ops::FeedOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
src/operators/feed_op.h
浏览文件 @ c67b701f
......@@ -32,7 +32,7 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
param_(inputs, outputs, attrs, *scope) {}
void RunImpl() const { param_.Out()->ShareDataWith(*param_.InputX()); }
void Init() const {}
void Init() {}
void InferShape() const {
auto out_dims = param_.Out()->dims();
......@@ -44,17 +44,14 @@ class FeedOp : public framework::OperatorBase<DeviceType> {
FeedParam param_;
};
namespace ops = paddle_mobile::operators;
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(feed);
REGISTER_OPERATOR_CPU(feed, ops::FeedOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(feed);
REGISTER_OPERATOR_MALI_GPU(feed, ops::FeedOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
src/operators/fetch_op.cpp
浏览文件 @ c67b701f
......@@ -19,3 +19,13 @@ namespace operators {
template class FetchOp<CPU, float>;
}
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fetch, ops::FetchOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
REGISTER_OPERATOR_MALI_GPU(fetch, ops::FetchOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
src/operators/fetch_op.h
浏览文件 @ c67b701f
......@@ -33,7 +33,7 @@ class FetchOp : public framework::OperatorBase<DeviceType> {
param_(inputs, outputs, attrs, *scope) {}
void RunImpl() const { param_.Out()->ShareDataWith(*param_.InputX()); }
void Init() const {}
void Init() {}
void InferShape() const {
auto x_dims = param_.InputX()->dims();
......@@ -44,17 +44,14 @@ class FetchOp : public framework::OperatorBase<DeviceType> {
FetchParam param_;
};
namespace ops = paddle_mobile::operators;
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fetch);
REGISTER_OPERATOR_CPU(fetch, ops::FetchOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(fetch);
REGISTER_OPERATOR_MALI_GPU(fetch, ops::FetchOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
src/operators/fusion_conv_add.cpp
浏览文件 @ c67b701f
......@@ -44,17 +44,38 @@ void FusionConvAddOp<Dtype, T>::InferShape() const {
framework::DDim ddim = framework::make_ddim(output_shape);
this->param_.Output()->Resize(ddim);
}
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_ADD_REGISTER
framework::FusionOpRegistrar convadd_registrar(new FusionConvAddMatcher());
#define CONV_ADD_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef CONV_ADD_REGISTER
static framework::FusionOpRegistrar convadd_registrar(
new FusionConvAddMatcher());
#define CONV_ADD_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
template class FusionConvAddOp<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(conv_add);
REGISTER_OPERATOR_CPU(conv_add, ops::FusionConvAddOp);
REGISTER_OPERATOR_CPU(fusion_conv_add, ops::FusionConvAddOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(conv_add);
REGISTER_OPERATOR_MALI_GPU(conv_add, ops::FusionConvAddOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/fusion_conv_add.h
浏览文件 @ c67b701f
......@@ -11,9 +11,8 @@ 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. */
#define FUSION_CONVADD_OP
#ifdef FUSION_CONVADD_OP
#ifdef FUSION_CONVADD_OP
#pragma once
#include <string>
......@@ -40,10 +39,10 @@ class FusionConvAddMatcher : public framework::FusionOpMatcher {
vector<std::shared_ptr<framework::OpDesc>> origin_descs =
node->OpDescs(node_.Depth());
node->Folder(node_.Depth(), Type(),
{{G_OP_TYPE_ELEMENTWISE_ADD, {"Y", "Y"}}}, removed_nodes);
{{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Y"}}}}, removed_nodes);
}
std::string Type() { return G_OP_TYPE_CONV_ADD; }
std::string Type() { return G_OP_TYPE_FUSION_CONV_ADD; }
};
template <typename DeviceType, typename T>
......@@ -68,11 +67,13 @@ class FusionConvAddOp : public framework::OperatorWithKernel<
};
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_ADD_REGISTER
static framework::FusionOpRegistrar convadd_registrar(
extern framework::FusionOpRegistrar convadd_registrar(
new FusionConvAddMatcher());
#define CONV_ADD_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......@@ -91,4 +92,13 @@ static framework::FusionOpRegistrar convadd_registrar(
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(conv_add);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_conv_add_bn_relu_op.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef FUSION_CONVADDBNRELU_OP
#include "operators/fusion_conv_add_bn_relu_op.h"
#include "operators/math/conv_func.h"
namespace paddle_mobile {
namespace operators {
template <typename Dtype, typename T>
void FusionConvAddBNReluOp<Dtype, T>::InferShape() const {
auto in_dims = this->param_.Input()->dims();
auto filter_dims = this->param_.Filter()->dims();
const std::vector<int> &strides = this->param_.Strides();
std::vector<int> paddings = this->param_.Paddings();
int groups = this->param_.Groups();
std::vector<int> dilations = this->param_.Dilations();
PADDLE_MOBILE_ENFORCE((in_dims.size() == filter_dims.size() &&
dilations.size() == paddings.size() &&
paddings.size() == strides.size()),
"ConvParam is not suitable");
std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
for (size_t i = 0; i < strides.size(); ++i) {
output_shape.push_back(
math::ConvOutputSize(in_dims[i + 2], filter_dims[i + 2], dilations[i],
paddings[i], strides[i]));
}
framework::DDim ddim = framework::make_ddim(output_shape);
this->param_.Output()->Resize(ddim);
}
template class FusionConvAddBNReluOp<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(fusion_conv_add_bn_relu, ops::FusionConvAddBNReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_conv_add_bn_relu_op.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef FUSION_CONVADDBNRELU_OP
#pragma once
#include <string>
#include <vector>
#include "framework/operator.h"
#include "framework/program/program-optimize/fusion_op_register.h"
#include "op_param.h"
#include "operators/kernel/conv_add_bn_relu_kernel.h"
namespace paddle_mobile {
namespace operators {
using std::string;
using std::vector;
class FusionConvAddBNReluMatcher : public framework::FusionOpMatcher {
public:
FusionConvAddBNReluMatcher() {
node_ = framework::Node(G_OP_TYPE_CONV);
node_ > std::make_shared<framework::Node>(G_OP_TYPE_ELEMENTWISE_ADD) >
std::make_shared<framework::Node>(G_OP_TYPE_BATCHNORM) >
std::make_shared<framework::Node>(G_OP_TYPE_RELU);
}
void FolderNodes(
framework::Node *node,
std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
vector<std::shared_ptr<framework::OpDesc>> origin_descs =
node->OpDescs(node_.Depth());
node->Folder(node_.Depth(), Type(),
{{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Y"}}},
{G_OP_TYPE_BATCHNORM,
{{"Scale", "Scale"},
{"Mean", "Mean"},
{"Bias", "Bias"},
{"Variance", "Variance"}}}},
removed_nodes);
}
std::string Type() { return G_OP_TYPE_FUSION_CONV_ADD_BN_RELU; }
};
template <typename DeviceType, typename T>
class FusionConvAddBNReluOp
: public framework::OperatorWithKernel<
DeviceType, FusionConvAddBNReluParam,
operators::ConvAddBNReluKernel<DeviceType, T>> {
public:
FusionConvAddBNReluOp(const string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
std::shared_ptr<framework::Scope> scope)
: framework::OperatorWithKernel<
DeviceType, FusionConvAddBNReluParam,
operators::ConvAddBNReluKernel<DeviceType, T>>(
type, inputs, outputs, attrs, scope) {}
using framework::OperatorWithKernel<
DeviceType, FusionConvAddBNReluParam,
operators::ConvAddBNReluKernel<DeviceType, T>>::OperatorWithKernel;
void InferShape() const override;
protected:
};
#ifdef PADDLE_MOBILE_CPU
//#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
// static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
// new FusionConvAddBNReluMatcher());
//#define FUSION_CONV_ADD_BN_RELU_REGISTER
//#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef FUSION_CONV_ADD_BN_RELU_REGISTER
static framework::FusionOpRegistrar fusion_conv_add_bn_relu_registrar(
new FusionConvAddBNReluMatcher());
#define FUSION_CONV_ADD_BN_RELU_REGISTER
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add_bn_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_conv_add_relu_op.cpp
浏览文件 @ c67b701f
......@@ -49,7 +49,6 @@ void FusionConvAddReluOp<Dtype, T>::InferShape() const {
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add_relu);
REGISTER_OPERATOR_CPU(fusion_conv_add_relu, ops::FusionConvAddReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......
src/operators/fusion_conv_add_relu_op.h
浏览文件 @ c67b701f
......@@ -36,7 +36,7 @@ class FusionConvAddReluOpMatcher : public framework::FusionOpMatcher {
framework::Node *node,
std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
node->Folder(node_.Depth(), Type(),
{{G_OP_TYPE_ELEMENTWISE_ADD, {"Y", "Y"}}}, removed_nodes);
{{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Y"}}}}, removed_nodes);
}
std::string Type() { return G_OP_TYPE_FUSION_CONV_ADD_RELU; }
};
......@@ -65,11 +65,11 @@ class FusionConvAddReluOp : public framework::OperatorWithKernel<
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_ADD_RELU_REGISTER
#define CONV_ADD_RELU_REGISTER
//#ifndef CONV_ADD_RELU_REGISTER
//#define CONV_ADD_RELU_REGISTER
// static framework::FusionOpRegistrar fusion_conv_add_relu_registrar(new
// FusionConvAddReluOpMatcher());
#endif
//#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......@@ -80,4 +80,12 @@ class FusionConvAddReluOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_conv_add_relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/fusion_fc_op.cpp
浏览文件 @ c67b701f
......@@ -49,17 +49,37 @@ void FusionFcOp<Dtype, T>::InferShape() const {
framework::DDim ddim = framework::make_ddim(output_dims);
this->param_.Out()->Resize(ddim);
}
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_CPU_REGISTER
#define CONV_CPU_REGISTER
framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef CONV_CPU_REGISTER
#define CONV_CPU_REGISTER
framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
template class FusionFcOp<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fc);
REGISTER_OPERATOR_CPU(fc, ops::FusionFcOp);
REGISTER_OPERATOR_CPU(fusion_fc, ops::FusionFcOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(fc);
REGISTER_OPERATOR_MALI_GPU(fc, ops::FusionFcOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/fusion_fc_op.h
浏览文件 @ c67b701f
......@@ -38,7 +38,7 @@ class FusionFcMatcher : public framework::FusionOpMatcher {
framework::Node *node,
std::vector<std::shared_ptr<framework::Node>> *removed_nodes) {
node->Folder(node_.Depth(), Type(),
{{G_OP_TYPE_ELEMENTWISE_ADD, {"Y", "Z"}}}, removed_nodes);
{{G_OP_TYPE_ELEMENTWISE_ADD, {{"Y", "Z"}}}}, removed_nodes);
}
std::string Type() { return G_OP_TYPE_FC; }
......@@ -66,17 +66,21 @@ class FusionFcOp
};
#ifdef PADDLE_MOBILE_CPU
#ifndef CONV_CPU_REGISTER
#define CONV_CPU_REGISTER
static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
extern framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#endif
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#ifndef CONV_CPU_REGISTER
#define CONV_CPU_REGISTER
static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
#endif
#endif
#ifdef PADDLE_MOBILE_FPGA
......@@ -85,4 +89,13 @@ static framework::FusionOpRegistrar fc_registrar(new FusionFcMatcher());
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(fusion_fc);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(fc);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/im2sequence_op.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef IM2SEQUENCE_OP
#include "operators/im2sequence_op.h"
namespace paddle_mobile {
namespace operators {
int Im2SequenceOutputSize(int input_size, int kernel, int padding_1,
int padding_2, int stride) {
int output_size =
1 + (padding_1 + padding_2 + input_size - kernel + stride - 1) / stride;
return output_size;
}
template <typename Dtype, typename T>
void Im2SequenceOp<Dtype, T>::InferShape() const {
auto in_x_dims = this->param_.Input()->dims();
const std::vector<int> &kernels = this->param_.Kernels();
const std::vector<int> &strides = this->param_.Strides();
std::vector<int> paddings = this->param_.Paddings();
std::vector<int64_t> output_shape({in_x_dims[0], in_x_dims[1]});
for (size_t i = 0; i < strides.size(); ++i) {
output_shape.push_back(Im2SequenceOutputSize(in_x_dims[i + 2], kernels[i],
paddings[i], paddings[i + 2],
strides[i]));
}
framework::DDim ddim = framework::make_ddim(output_shape);
this->param_.Output()->Resize(ddim);
}
template class Im2SequenceOp<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
REGISTER_OPERATOR_CPU(im2sequence, ops::Im2SequenceOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/im2sequence_op.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef IM2SEQUENCE_OP
#pragma once
#include <operators/op_param.h>
#include "framework/operator.h"
#include "operators/kernel/im2sequence_kernel.h"
namespace paddle_mobile {
namespace operators {
using namespace framework;
template <typename DeviceType, typename T>
class Im2SequenceOp : public framework::OperatorWithKernel<
DeviceType, Im2SequenceParam,
operators::Im2SequenceKernel<DeviceType, T>> {
public:
Im2SequenceOp(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
std::shared_ptr<framework::Scope> scope)
: framework::OperatorWithKernel<
DeviceType, Im2SequenceParam,
operators::Im2SequenceKernel<DeviceType, T>>(type, inputs, outputs,
attrs, scope) {}
// using framework::OperatorWithKernel<
// DeviceType, Im2SequenceParam,
// operators::Im2SequenceKernel<DeviceType, T>>::OperatorWithKernel;
void InferShape() const override;
private:
};
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(im2sequence);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/kernel/arm/batchnorm_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,8 +14,6 @@ limitations under the License. */
#ifdef BATCHNORM_OP
#pragma once
#include "operators/kernel/batchnorm_kernel.h"
#include "operators/kernel/central-arm-func/batchnorm_arm_func.h"
......@@ -23,7 +21,7 @@ namespace paddle_mobile {
namespace operators {
template <>
bool BatchNormKernel<CPU, float>::Init(const BatchNormParam &para) const {
bool BatchNormKernel<CPU, float>::Init(BatchNormParam *param) {
return true;
}
......
src/operators/kernel/arm/box_coder_kernel.cpp
浏览文件 @ c67b701f
......@@ -15,129 +15,21 @@ limitations under the License. */
#ifdef BOXCODER_OP
#include "operators/kernel/box_coder_kernel.h"
#include "operators/kernel/central-arm-func/box_coder_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <typename T>
void EncodeCenterSize(const framework::Tensor& target_box,
const framework::Tensor& prior_box,
const framework::Tensor& prior_box_var, T* output) {
int64_t row = target_box.dims()[0];
int64_t col = prior_box.dims()[0];
int64_t len = prior_box.dims()[1];
auto* target_box_data = target_box.data<T>();
auto* prior_box_data = prior_box.data<T>();
auto* prior_box_var_data = prior_box_var.data<T>();
for (int64_t i = 0; i < row; ++i) {
for (int64_t j = 0; j < col; ++j) {
T prior_box_width = prior_box_data[j * len + 2] - prior_box_data[j * len];
T prior_box_height =
prior_box_data[j * len + 3] - prior_box_data[j * len + 1];
T prior_box_center_x =
(prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
T prior_box_center_y =
(prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
T target_box_center_x =
(target_box_data[i * len + 2] + target_box_data[i * len]) / 2;
T target_box_center_y =
(target_box_data[i * len + 3] + target_box_data[i * len + 1]) / 2;
T target_box_width =
target_box_data[i * len + 2] - target_box_data[i * len];
T target_box_height =
target_box_data[i * len + 3] - target_box_data[i * len + 1];
size_t offset = i * col * len + j * len;
output[offset] = (target_box_center_x - prior_box_center_x) /
prior_box_width / prior_box_var_data[j * len];
output[offset + 1] = (target_box_center_y - prior_box_center_y) /
prior_box_height / prior_box_var_data[j * len + 1];
output[offset + 2] =
std::log(std::fabs(target_box_width / prior_box_width)) /
prior_box_var_data[j * len + 2];
output[offset + 3] =
std::log(std::fabs(target_box_height / prior_box_height)) /
prior_box_var_data[j * len + 3];
}
}
}
template <typename T>
void DecodeCenterSize(const framework::Tensor& target_box,
const framework::Tensor& prior_box,
const framework::Tensor& prior_box_var, T* output) {
int64_t row = target_box.dims()[0];
int64_t col = prior_box.dims()[0];
int64_t len = prior_box.dims()[1];
auto* target_box_data = target_box.data<T>();
auto* prior_box_data = prior_box.data<T>();
auto* prior_box_var_data = prior_box_var.data<T>();
for (int64_t i = 0; i < row; ++i) {
for (int64_t j = 0; j < col; ++j) {
size_t offset = i * col * len + j * len;
T prior_box_width = prior_box_data[j * len + 2] - prior_box_data[j * len];
T prior_box_height =
prior_box_data[j * len + 3] - prior_box_data[j * len + 1];
T prior_box_center_x =
(prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
T prior_box_center_y =
(prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
T target_box_center_x = prior_box_var_data[j * len] *
target_box_data[offset] * prior_box_width +
prior_box_center_x;
T target_box_center_y = prior_box_var_data[j * len + 1] *
target_box_data[offset + 1] *
prior_box_height +
prior_box_center_y;
T target_box_width = std::exp(prior_box_var_data[j * len + 2] *
target_box_data[offset + 2]) *
prior_box_width;
T target_box_height = std::exp(prior_box_var_data[j * len + 3] *
target_box_data[offset + 3]) *
prior_box_height;
output[offset] = target_box_center_x - target_box_width / 2;
output[offset + 1] = target_box_center_y - target_box_height / 2;
output[offset + 2] = target_box_center_x + target_box_width / 2;
output[offset + 3] = target_box_center_y + target_box_height / 2;
}
}
}
template <>
bool BoxCoderKernel<CPU, float>::Init(const BoxCoderParam& para) const {
bool BoxCoderKernel<CPU, float>::Init(BoxCoderParam *param) {
return true;
}
template <>
void BoxCoderKernel<CPU, float>::Compute(const BoxCoderParam& param) const {
const auto* input_priorbox = param.InputPriorBox();
const auto* input_priorboxvar = param.InputPriorBoxVar();
const auto* input_targetbox = param.InputTargetBox();
const auto& code_type = param.CodeType();
auto row = input_targetbox->dims()[0];
auto col = input_priorbox->dims()[0];
auto len = input_priorbox->dims()[1];
Tensor* output_box = param.OutputBox();
auto* output_box_dataptr = output_box->mutable_data<float>({row, col, len});
if (code_type == "encode_center_size") {
EncodeCenterSize<float>(*input_targetbox, *input_priorbox,
*input_priorboxvar, output_box_dataptr);
}
if (code_type == "decode_center_size") {
DecodeCenterSize<float>(*input_targetbox, *input_priorbox,
*input_priorboxvar, output_box_dataptr);
}
void BoxCoderKernel<CPU, float>::Compute(const BoxCoderParam &param) const {
BoxCoderCompute<float>(param);
}
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/arm/concat_kernel.cpp
浏览文件 @ c67b701f
......@@ -15,77 +15,19 @@ limitations under the License. */
#ifdef CONCAT_OP
#include "operators/kernel/concat_kernel.h"
#include "operators/kernel/central-arm-func/concat_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <typename T>
class ConcatFunctor {
public:
void operator()(const std::vector<framework::Tensor> &input, const int axis,
framework::Tensor *output) {
size_t num = input.size();
int rows = 1;
auto dim_0 = input[0].dims();
for (int i = 0; i < axis; ++i) {
rows *= dim_0[i];
}
int out_rows = rows, out_cols = 0;
std::vector<int64_t> input_cols(input.size());
for (int i = 0; i < num; ++i) {
int t_cols = input[i].numel() / rows;
out_cols += t_cols;
input_cols[i] = t_cols;
}
// computation
for (int k = 0; k < out_rows; ++k) {
T *dst_ptr = output->data<T>() + k * out_cols;
int col_idx = 0;
for (int j = 0; j < num; ++j) {
int col_len = input_cols[j];
const T *src_prt = input[j].data<T>() + k * col_len;
memory::Copy(dst_ptr + col_idx, src_prt, sizeof(T) * col_len);
col_idx += col_len;
}
}
}
};
template <>
bool ConcatKernel<CPU, float>::Init(const ConcatParam &para) const {
bool ConcatKernel<CPU, float>::Init(ConcatParam *param) {
return true;
}
template <>
void ConcatKernel<CPU, float>::Compute(const ConcatParam &param) const {
auto inputs = param.Inputs();
auto *out = param.Out();
int64_t axis = param.Axis();
out->mutable_data<float>();
/// Sometimes direct copies will be faster, this maybe need deeply analysis.
if (axis == 0 && inputs.size() < 10) {
size_t output_offset = 0;
for (auto *in : inputs) {
auto in_stride = framework::stride_numel(in->dims());
auto out_stride = framework::stride_numel(out->dims());
auto dst = out->data<float>() + output_offset;
auto src = in->data<float>();
PADDLE_MOBILE_ENFORCE(
in_stride.size() == out_stride.size(),
"src and dst tensor should have the same dims size.");
memory::Copy(dst, src, sizeof(float) * in_stride[0]);
output_offset += in_stride[0];
}
} else {
std::vector<framework::Tensor> inputs_concat(inputs.size());
for (int j = 0; j < inputs.size(); ++j) {
inputs_concat[j] = *inputs[j];
}
ConcatFunctor<float> concat_functor;
concat_functor(inputs_concat, static_cast<int>(axis), out);
}
ConcatCompute<float>(param);
}
} // namespace operators
......
src/operators/kernel/arm/conv_add_bn_relu_kernel.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef FUSION_CONVADDBNRELU_OP
#include "operators/kernel/conv_add_bn_relu_kernel.h"
#include "operators/kernel/central-arm-func/conv_add_bn_relu_func.h"
namespace paddle_mobile {
namespace operators {
template <>
bool ConvAddBNReluKernel<CPU, float>::Init(FusionConvAddBNReluParam *param) {
const Tensor *mean = param->InputMean();
const Tensor *variance = param->InputVariance();
const Tensor *scale = param->InputScale();
const Tensor *bias = param->InputBias();
const float epsilon = param->Epsilon();
auto mean_ptr = mean->data<float>();
auto variance_ptr = variance->data<float>();
auto scale_ptr = scale->data<float>();
auto bias_ptr = bias->data<float>();
const int C = mean->numel();
float inv_std_ptr[C];
for (int i = 0; i < C; i++) {
inv_std_ptr[i] =
1 / static_cast<float>(pow((variance_ptr[i] + epsilon), 0.5));
}
Tensor *new_scale = new Tensor();
Tensor *new_bias = new Tensor();
auto new_scale_ptr = new_scale->mutable_data<float>({C});
auto new_bias_ptr = new_bias->mutable_data<float>({C});
for (int i = 0; i < C; i++) {
new_scale_ptr[i] = inv_std_ptr[i] * scale_ptr[i];
new_bias_ptr[i] = bias_ptr[i] - mean_ptr[i] * inv_std_ptr[i] * scale_ptr[i];
}
param->SetNewScale(new_scale);
param->SetNewBias(new_bias);
return true;
}
template <>
void ConvAddBNReluKernel<CPU, float>::Compute(
const FusionConvAddBNReluParam &param) const {
ConvAddBNReluCompute<float>(param);
}
template class ConvAddBNReluKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/arm/conv_add_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,104 +14,21 @@ limitations under the License. */
#ifdef FUSION_CONVADD_OP
#include "operators/kernel/conv_add_kernel.h"
#include "../central-arm-func/conv_add_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <>
bool ConvAddKernel<CPU, float>::Init(const FusionConvAddParam &para) const {
bool ConvAddKernel<CPU, float>::Init(FusionConvAddParam *param) {
return true;
}
template <>
void ConvAddKernel<CPU, float>::Compute(const FusionConvAddParam &param) const {
const Tensor *input = param.Input();
Tensor filter = *param.Filter();
Tensor bias = *param.Bias();
int axis = param.Axis();
Tensor *output = param.Output();
math::expand_bias(bias, axis, output->dims());
output->ShareDataWith(bias);
int groups = param.Groups();
std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings();
std::vector<int> dilations = param.Dilations();
const int batch_size = static_cast<int>(input->dims()[0]);
std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
size_t data_dim = filter_shape_vec.size() - 2;
std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
col_shape_vec[0] = input->dims()[1] / groups;
for (size_t j = 0; j < data_dim; ++j) {
col_shape_vec[j + 1] = filter_shape_vec[j + 2];
col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
}
framework::DDim col_shape(framework::make_ddim(col_shape_vec));
framework::DDim col_matrix_shape =
framework::flatten_to_2d(col_shape, data_dim + 1);
bool is_expand =
math::IsExpand(filter_shape_vec, strides, paddings, dilations);
Tensor col;
Tensor col_matrix;
if (is_expand) {
col.mutable_data<float>(col_shape);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
filter.Resize(filter_matrix_shape);
framework::DDim output_matrix_shape = {
output->dims()[1],
output->numel() / (output->dims()[0] * output->dims()[1])};
// convolution operator: im2col(or vol2col) + gemm
int in_step = static_cast<int>(input->dims()[1]) / groups;
int out_step = static_cast<int>(output->dims()[1]) / groups;
math::Vol2ColFunctor<CPU, float> vol2col;
math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col;
for (int i = 0; i < batch_size; i++) {
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
for (int g = 0; g < groups; g++) {
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
if (!is_expand) {
col.ShareDataWith(in_slice);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
} else if (data_dim == 2U) {
// im2col
im2col(in_slice, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&col);
} else if (data_dim == 3U) {
// vol2col
vol2col(in_slice, dilations, strides, paddings, &col);
}
// gemm
Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
math::matmul<float>(filter_slice, false, col_matrix, false,
static_cast<float>(1), &out_slice,
static_cast<float>(1));
}
}
ConvAddCompute<float>(param);
}
template class ConvAddKernel<CPU, float>;
} // namespace operators
......
src/operators/kernel/arm/conv_add_relu_kernel.cpp
浏览文件 @ c67b701f
......@@ -21,8 +21,7 @@ namespace paddle_mobile {
namespace operators {
template <>
bool ConvAddReluKernel<CPU, float>::Init(
const FusionConvAddReluParam &para) const {
bool ConvAddReluKernel<CPU, float>::Init(FusionConvAddReluParam *param) {
return true;
}
......
src/operators/kernel/arm/conv_kernel.cpp
浏览文件 @ c67b701f
......@@ -21,7 +21,7 @@ namespace paddle_mobile {
namespace operators {
template <>
bool ConvKernel<CPU, float>::Init(const ConvParam &para) const {
bool ConvKernel<CPU, float>::Init(ConvParam *param) {
return true;
}
......
src/operators/kernel/arm/depthwise_conv_kernel.cpp
浏览文件 @ c67b701f
......@@ -21,7 +21,7 @@ namespace paddle_mobile {
namespace operators {
template <>
bool DepthwiseConvKernel<CPU, float>::Init(const ConvParam &para) const {
bool DepthwiseConvKernel<CPU, float>::Init(ConvParam *param) {
return true;
}
......
src/operators/kernel/arm/dropout_kernel.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef DROPOUT_OP
#pragma once
#include "operators/kernel/dropout_kernel.h"
#include <operators/math/transform.h>
namespace paddle_mobile {
namespace operators {
template <>
bool DropoutKernel<CPU, float>::Init(DropoutParam *para) {
return true;
}
template <typename T>
struct DropoutFunctor {
inline T operator()(T in) const { return in; }
};
template <>
void DropoutKernel<CPU, float>::Compute(const DropoutParam &param) const {
const auto *input_x = param.InputX();
auto *input_x_ptr = input_x->data<float>();
auto *out = param.Out();
auto *out_ptr = out->mutable_data<float>();
DropoutFunctor<float> func_;
math::Transform trans;
trans(input_x_ptr, input_x_ptr + input_x->numel(), out_ptr, func_);
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/arm/elementwise_add_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,38 +14,23 @@ limitations under the License. */
#ifdef ELEMENTWISEADD_OP
#pragma once
#include "operators/kernel/elementwise_add_kernel.h"
#include "operators/kernel/central-arm-func/elementwise_add_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <typename T>
struct AddFunctor {
inline T operator()(T a, T b) const { return a + b; }
};
template <>
bool ElementwiseAddKernel<CPU, float>::Init(
const ElementwiseAddParam &para) const {
bool ElementwiseAddKernel<CPU, float>::Init(ElementwiseAddParam *param) {
return true;
}
template <>
void ElementwiseAddKernel<CPU, float>::Compute(
const ElementwiseAddParam &param) const {
const Tensor *input_x = param.InputX();
const Tensor *input_y = param.InputY();
Tensor *Out = param.Out();
Out->mutable_data<float>();
int axis = param.Axis();
ElementwiseComputeEx<AddFunctor<float>, float>(input_x, input_y, axis,
AddFunctor<float>(), Out);
ElementwiseAddCompute<float>(param);
}
template class ElementwiseAddKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/arm/fusion_fc_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,60 +14,20 @@ limitations under the License. */
#ifdef FUSION_FC_OP
#pragma once
#include "operators/kernel/fusion_fc_kernel.h"
#include "operators/kernel/central-arm-func/fusion_fc_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <>
bool FusionFcKernel<CPU, float>::Init(const FusionFcParam &para) const {
bool FusionFcKernel<CPU, float>::Init(FusionFcParam *param) {
return true;
}
template <>
void FusionFcKernel<CPU, float>::Compute(const FusionFcParam &param) const {
const Tensor *input_x = param.InputX();
const Tensor *input_y = param.InputY();
const Tensor *input_z = param.InputZ();
auto *input_z_data = input_z->data<float>();
int axis = param.Axis();
Tensor *out = param.Out();
auto *out_data = out->mutable_data<float>();
const Tensor x_matrix =
input_x->dims().size() > 2
? framework::ReshapeToMatrix(*input_x, param.XNumColDims())
: *input_x;
const Tensor y_matrix =
input_y->dims().size() > 2
? framework::ReshapeToMatrix(*input_y, param.YNumColDims())
: *input_y;
auto out_dim = out->dims();
if (out_dim.size() != 2) {
out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
}
PADDLE_MOBILE_ENFORCE(out_dim.size() == 2, " out_dim.size must be 2.");
PADDLE_MOBILE_ENFORCE(input_z->dims().size() == 1, "inpu_z size must be 1");
PADDLE_MOBILE_ENFORCE(out_dim[1] == input_z->dims()[0],
" out_dim.size must be 2.");
axis = (axis == -1 ? out_dim.size() - input_z->dims().size() : axis);
PADDLE_MOBILE_ENFORCE(axis == 1, " to fit broadcast, axis = 1. ")
int64_t classes = input_z->numel();
for (int i = 0; i < out_dim[0]; i++) {
memory::Copy(out_data + i * classes, input_z_data, sizeof(float) * classes);
}
for (int i = 0; i < out->numel(); i++) {
DLOG << out_data[i];
}
math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
out, static_cast<float>(1));
PADDLE_MOBILE_ENFORCE(out_dim.size() == 2, " out_dim.size must be 2.");
// if (out_dim.size() != 2) {
// out->Resize(out_dim);
// }
FusionFcCompute<float>(param);
}
} // namespace operators
......
src/operators/kernel/arm/im2sequence_kernel.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef IM2SEQUENCE_OP
#include "operators/kernel/im2sequence_kernel.h"
namespace paddle_mobile {
namespace operators {
template <>
bool Im2SequenceKernel<CPU, float>::Init(Im2SequenceParam *para) {
return true;
}
inline int Im2SeqOutputSize(int input_size, int filter_size, int padding_0,
int padding_1, int stride) {
const int output_size =
(input_size + padding_0 + padding_1 - filter_size) / stride + 1;
return output_size;
}
template <>
void Im2SequenceKernel<CPU, float>::Compute(
const Im2SequenceParam &param) const {
const Tensor *in_x = param.Input();
Tensor *out = param.Output();
out->mutable_data<float>();
std::vector<int> kernels = param.Kernels();
std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings();
auto in_x_dim = in_x->dims();
const int batch_size = static_cast<int>(in_x_dim[0]);
const int img_channels = static_cast<int>(in_x_dim[1]);
const int img_height = static_cast<int>(in_x_dim[2]);
const int img_width = static_cast<int>(in_x_dim[3]);
int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0],
paddings[2], strides[0]);
int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1],
paddings[3], strides[1]);
const std::vector<int> dilations({1, 1});
// TODO: verify
auto out_dims = out->dims();
out->Resize({batch_size, out->numel() / batch_size});
for (int i = 0; i < batch_size; i++) {
const Tensor src =
in_x->Slice(i, i + 1).Resize({img_channels, img_height, img_width});
Tensor dst = out->Slice(i, i + 1).Resize(
{output_height, output_width, img_channels, kernels[0], kernels[1]});
math::Im2ColFunctor<math::ColFormat::kOCF, CPU, float> f;
f(src, dilations, strides, paddings, &dst);
}
out->Resize(out_dims);
}
template class Im2SequenceKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/arm/lrn_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,40 +14,22 @@ limitations under the License. */
#ifdef LRN_OP
#pragma once
#include "operators/kernel/lrn_kernel.h"
#include "operators/kernel/central-arm-func/lrn_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <>
bool LrnKernel<CPU, float>::Init(const LrnParam &para) const {
bool LrnKernel<CPU, float>::Init(LrnParam *param) {
return true;
}
template <>
void LrnKernel<CPU, float>::Compute(const LrnParam &param) const {
const Tensor *input_x = param.InputX();
auto x_dims = input_x->dims();
Tensor *out = param.Out();
out->mutable_data<float>();
/// data_format = NCHW
const int N = x_dims[0];
const int C = x_dims[1];
const int H = x_dims[2];
const int W = x_dims[3];
const int n = param.N();
const float alpha = param.Alpha();
const float beta = param.Beta();
const float k = param.K();
LRNFunctor<float> lrnFunctor;
lrnFunctor(*input_x, out, N, C, H, W, n, k, alpha, beta);
LrnCompute<float>(param);
}
template class LrnKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/arm/mul_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,45 +14,22 @@ limitations under the License. */
#ifdef MUL_OP
#pragma once
#include "operators/kernel/mul_kernel.h"
#include "operators/kernel/central-arm-func/mul_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <>
bool MulKernel<CPU, float>::Init(const MulParam &para) const {
bool MulKernel<CPU, float>::Init(MulParam *param) {
return true;
}
template <>
void MulKernel<CPU, float>::Compute(const MulParam &param) const {
const Tensor *input_x = param.InputX();
const Tensor *input_y = param.InputY();
Tensor *out = param.Out();
out->mutable_data<float>();
const Tensor x_matrix =
input_x->dims().size() > 2
? framework::ReshapeToMatrix(*input_x, param.XNumColDims())
: *input_x;
const Tensor y_matrix =
input_y->dims().size() > 2
? framework::ReshapeToMatrix(*input_y, param.YNumColDims())
: *input_y;
auto out_dim = out->dims();
if (out_dim.size() != 2) {
out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
}
math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
out, static_cast<float>(0));
if (out_dim.size() != 2) {
out->Resize(out_dim);
}
MulCompute<float>(param);
}
template class MulKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/arm/multiclass_nms_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,269 +14,21 @@ limitations under the License. */
#ifdef MULTICLASSNMS_OP
#pragma once
#include "operators/kernel/multiclass_nms_kernel.h"
#include "operators/kernel/central-arm-func/multiclass_nms_arm_func.h"
namespace paddle_mobile {
namespace operators {
constexpr int kOutputDim = 6;
constexpr int kBBoxSize = 4;
template <class T>
bool SortScorePairDescend(const std::pair<float, T>& pair1,
const std::pair<float, T>& pair2) {
return pair1.first > pair2.first;
}
template <class T>
static inline void GetMaxScoreIndex(
const std::vector<T>& scores, const T threshold, int top_k,
std::vector<std::pair<T, int>>* sorted_indices) {
for (size_t i = 0; i < scores.size(); ++i) {
if (scores[i] > threshold) {
sorted_indices->push_back(std::make_pair(scores[i], i));
}
}
// Sort the score pair according to the scores in descending order
std::stable_sort(sorted_indices->begin(), sorted_indices->end(),
SortScorePairDescend<int>);
// Keep top_k scores if needed.
if (top_k > -1 && top_k < static_cast<int>(sorted_indices->size())) {
sorted_indices->resize(top_k);
}
}
template <class T>
static inline T BBoxArea(const T* box, const bool normalized) {
if (box[2] < box[0] || box[3] < box[1]) {
// If coordinate values are is invalid
// (e.g. xmax < xmin or ymax < ymin), return 0.
return static_cast<T>(0.);
} else {
const T w = box[2] - box[0];
const T h = box[3] - box[1];
if (normalized) {
return w * h;
} else {
// If coordinate values are not within range [0, 1].
return (w + 1) * (h + 1);
}
}
}
template <class T>
static inline T JaccardOverlap(const T* box1, const T* box2,
const bool normalized) {
if (box2[0] > box1[2] || box2[2] < box1[0] || box2[1] > box1[3] ||
box2[3] < box1[1]) {
return static_cast<T>(0.);
} else {
const T inter_xmin = std::max(box1[0], box2[0]);
const T inter_ymin = std::max(box1[1], box2[1]);
const T inter_xmax = std::min(box1[2], box2[2]);
const T inter_ymax = std::min(box1[3], box2[3]);
const T inter_w = inter_xmax - inter_xmin;
const T inter_h = inter_ymax - inter_ymin;
const T inter_area = inter_w * inter_h;
const T bbox1_area = BBoxArea<T>(box1, normalized);
const T bbox2_area = BBoxArea<T>(box2, normalized);
return inter_area / (bbox1_area + bbox2_area - inter_area);
}
}
template <typename T>
static inline void NMSFast(const Tensor& bbox, const Tensor& scores,
const T score_threshold, const T nms_threshold,
const T eta, const int64_t top_k,
std::vector<int>* selected_indices) {
// The total boxes for each instance.
int64_t num_boxes = bbox.dims()[0];
// 4: [xmin ymin xmax ymax]
int64_t box_size = bbox.dims()[1];
std::vector<T> scores_data(num_boxes);
std::copy_n(scores.data<T>(), num_boxes, scores_data.begin());
std::vector<std::pair<T, int>> sorted_indices;
GetMaxScoreIndex(scores_data, score_threshold, top_k, &sorted_indices);
selected_indices->clear();
T adaptive_threshold = nms_threshold;
const T* bbox_data = bbox.data<T>();
while (sorted_indices.size() != 0) {
const int idx = sorted_indices.front().second;
bool keep = true;
for (size_t k = 0; k < selected_indices->size(); ++k) {
if (keep) {
const int kept_idx = (*selected_indices)[k];
T overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
bbox_data + kept_idx * box_size, true);
keep = overlap <= adaptive_threshold;
} else {
break;
}
}
if (keep) {
selected_indices->push_back(idx);
}
sorted_indices.erase(sorted_indices.begin());
if (keep && eta < 1 && adaptive_threshold > 0.5) {
adaptive_threshold *= eta;
}
}
}
template <typename T>
void MultiClassNMS(const Tensor& scores, const Tensor& bboxes,
std::map<int, std::vector<int>>* indices, int* num_nmsed_out,
const int& background_label, const int& nms_top_k,
const int& keep_top_k, const T& nms_threshold,
const T& nms_eta, const T& score_threshold) {
int64_t class_num = scores.dims()[0];
int64_t predict_dim = scores.dims()[1];
int num_det = 0;
for (int64_t c = 0; c < class_num; ++c) {
if (c == background_label) continue;
Tensor score = scores.Slice(c, c + 1);
/// [c] is key
NMSFast<float>(bboxes, score, score_threshold, nms_threshold, nms_eta,
nms_top_k, &((*indices)[c]));
num_det += (*indices)[c].size();
}
*num_nmsed_out = num_det;
const T* scores_data = scores.data<T>();
if (keep_top_k > -1 && num_det > keep_top_k) {
std::vector<std::pair<float, std::pair<int, int>>> score_index_pairs;
for (const auto& it : *indices) {
int label = it.first;
const T* sdata = scores_data + label * predict_dim;
const std::vector<int>& label_indices = it.second;
for (size_t j = 0; j < label_indices.size(); ++j) {
int idx = label_indices[j];
// PADDLE_ENFORCE_LT(idx, predict_dim);
score_index_pairs.push_back(
std::make_pair(sdata[idx], std::make_pair(label, idx)));
}
}
// Keep top k results per image.
std::stable_sort(score_index_pairs.begin(), score_index_pairs.end(),
SortScorePairDescend<std::pair<int, int>>);
score_index_pairs.resize(keep_top_k);
// Store the new indices.
std::map<int, std::vector<int>> new_indices;
for (size_t j = 0; j < score_index_pairs.size(); ++j) {
int label = score_index_pairs[j].second.first;
int idx = score_index_pairs[j].second.second;
new_indices[label].push_back(idx);
}
new_indices.swap(*indices);
*num_nmsed_out = keep_top_k;
}
}
template <typename T>
void MultiClassOutput(const Tensor& scores, const Tensor& bboxes,
const std::map<int, std::vector<int>>& selected_indices,
Tensor* outs) {
int predict_dim = scores.dims()[1];
auto* scores_data = scores.data<T>();
auto* bboxes_data = bboxes.data<T>();
auto* odata = outs->data<T>();
int count = 0;
for (const auto& it : selected_indices) {
/// one batch
int label = it.first;
const T* sdata = scores_data + label * predict_dim;
const std::vector<int>& indices = it.second;
for (size_t j = 0; j < indices.size(); ++j) {
int idx = indices[j];
const T* bdata = bboxes_data + idx * kBBoxSize;
odata[count * kOutputDim] = label; // label
odata[count * kOutputDim + 1] = sdata[idx]; // score
// xmin, ymin, xmax, ymax
std::memcpy(odata + count * kOutputDim + 2, bdata, 4 * sizeof(T));
count++;
}
}
}
template <>
bool MultiClassNMSKernel<CPU, float>::Init(
const MultiClassNMSParam& para) const {
bool MultiClassNMSKernel<CPU, float>::Init(MultiClassNMSParam *param) {
return true;
}
template <>
void MultiClassNMSKernel<CPU, float>::Compute(
const MultiClassNMSParam& param) const {
const auto* input_bboxes = param.InputBBoxes();
const auto& input_bboxes_dims = input_bboxes->dims();
const auto* input_scores = param.InputScores();
const auto& input_scores_dims = input_scores->dims();
auto* outs = param.Out();
auto background_label = param.BackGroundLabel();
auto nms_top_k = param.NMSTopK();
auto keep_top_k = param.KeepTopK();
auto nms_threshold = param.NMSThreshold();
auto nms_eta = param.NMSEta();
auto score_threshold = param.ScoreThreshold();
int64_t batch_size = input_scores_dims[0];
int64_t class_num = input_scores_dims[1];
int64_t predict_dim = input_scores_dims[2];
int64_t box_dim = input_bboxes_dims[2];
std::vector<std::map<int, std::vector<int>>> all_indices;
std::vector<size_t> batch_starts = {0};
for (int64_t i = 0; i < batch_size; ++i) {
Tensor ins_score = input_scores->Slice(i, i + 1);
ins_score.Resize({class_num, predict_dim});
Tensor ins_boxes = input_bboxes->Slice(i, i + 1);
ins_boxes.Resize({predict_dim, box_dim});
std::map<int, std::vector<int>> indices;
int num_nmsed_out = 0;
MultiClassNMS<float>(ins_score, ins_boxes, &indices, &num_nmsed_out,
background_label, nms_top_k, keep_top_k, nms_threshold,
nms_eta, score_threshold);
all_indices.push_back(indices);
batch_starts.push_back(batch_starts.back() + num_nmsed_out);
}
int num_kept = batch_starts.back();
if (num_kept == 0) {
float* od = outs->mutable_data<float>({1});
od[0] = -1;
} else {
outs->mutable_data<float>({num_kept, kOutputDim});
for (int64_t i = 0; i < batch_size; ++i) {
Tensor ins_score = input_scores->Slice(i, i + 1);
ins_score.Resize({class_num, predict_dim});
Tensor ins_boxes = input_bboxes->Slice(i, i + 1);
ins_boxes.Resize({predict_dim, box_dim});
int64_t s = batch_starts[i];
int64_t e = batch_starts[i + 1];
if (e > s) {
Tensor out = outs->Slice(s, e);
MultiClassOutput<float>(ins_score, ins_boxes, all_indices[i], &out);
}
}
}
// framework::LoD lod;
// lod.emplace_back(batch_starts);
//
// outs->set_lod(lod);
const MultiClassNMSParam &param) const {
MultiClassNMSCompute<float>(param);
}
} // namespace operators
......
src/operators/kernel/arm/pool_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,70 +14,19 @@ limitations under the License. */
#ifdef POOL_OP
#include <operators/kernel/pool_kernel.h>
#include "common/log.h"
#include "operators/kernel/pool_kernel.h"
#include "../central-arm-func/pool_arm_func.h"
namespace paddle_mobile {
namespace operators {
inline void PoolBasic(std::string pooling_type, std::vector<int> ksize,
std::vector<int> strides, std::vector<int> paddings,
const Tensor *in_x, Tensor *out) {
if (pooling_type == "max") {
math::PoolFunctor<CPU, math::MaxPool<float>, float> pool2d_forward;
math::MaxPool<float> pool_process;
pool2d_forward(*in_x, ksize, strides, paddings, pool_process, out);
} else if (pooling_type == "avg") {
math::PoolFunctor<CPU, math::AvgPool<float>, float> pool2d_forward;
math::AvgPool<float> pool_process;
pool2d_forward(*in_x, ksize, strides, paddings, pool_process, out);
}
}
template <>
bool PoolKernel<CPU, float>::Init(const PoolParam &para) const {
bool PoolKernel<CPU, float>::Init(PoolParam *param) {
return true;
}
template <>
void PoolKernel<CPU, float>::Compute(const PoolParam &param) const {
const Tensor *in_x = param.Input();
Tensor *out = param.Output();
std::string pooling_type = param.PoolingType();
std::vector<int> ksize = param.Ksize();
std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings();
if (ksize.size() != 2) {
LOG(paddle_mobile::LogLevel::kLOG_ERROR)
<< "Pool op only supports 2D and 3D input.";
}
if (param.isGlobalPooling()) {
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
}
} else if (ksize[0] == 3 && ksize[0] == ksize[1]) {
if (pooling_type == "max") {
math::Pool3x3Max(strides, paddings, in_x, out);
} else if (pooling_type == "avg") {
math::Pool3x3Avg(strides, paddings, in_x, out);
}
} else if (ksize[0] == 2 && ksize[0] == ksize[1]) {
if (pooling_type == "max") {
math::Pool2x2Max(strides, paddings, in_x, out);
} else if (pooling_type == "avg") {
math::Pool2x2Avg(strides, paddings, in_x, out);
}
} else {
PoolBasic(pooling_type, ksize, strides, paddings, in_x, out);
}
PoolCompute<float>(param);
}
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/arm/prior_box_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,138 +14,20 @@ limitations under the License. */
#ifdef PRIORBOX_OP
#pragma once
#include "operators/kernel/prior_box_kernel.h"
#include "operators/kernel/central-arm-func/prior_box_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <typename T>
struct ClipFunctor {
inline T operator()(T in) const {
return std::min<T>(std::max<T>(in, 0.), 1.);
}
};
template <>
bool PriorBoxKernel<CPU, float>::Init(const PriorBoxParam &para) const {
bool PriorBoxKernel<CPU, float>::Init(PriorBoxParam *param) {
return true;
}
template <>
void PriorBoxKernel<CPU, float>::Compute(const PriorBoxParam &param) const {
const auto *input_ = param.Input();
const auto &input_dims = input_->dims();
const auto *input_image = param.InputImage();
const auto &input_image_dims = input_image->dims();
const auto &min_sizes = param.MinSizes();
const auto &max_sizes = param.MaxSizes();
const auto &variances = param.Variances();
const auto &input_aspect_ratio = param.AspectRatios();
const bool &flip = param.Flip();
const bool &clip = param.Clip();
const float &step_w = param.StepW();
const float &step_h = param.StepH();
const float &offset = param.Offset();
Tensor *output_boxes = param.OutputBoxes();
auto output_boxes_dataptr = output_boxes->mutable_data<float>();
Tensor *output_variances = param.OutputVariances();
auto output_variances_dataptr = output_variances->mutable_data<float>();
std::vector<float> aspect_ratios;
ExpandAspectRatios(input_aspect_ratio, flip, &aspect_ratios);
auto img_width = input_image_dims[3];
auto img_height = input_image_dims[2];
auto feature_width = input_dims[3];
auto feature_height = input_dims[2];
auto stride0 = output_boxes->dims()[1] * output_boxes->dims()[2] *
output_boxes->dims()[3];
auto stride1 = output_boxes->dims()[2] * output_boxes->dims()[3];
auto stride2 = output_boxes->dims()[3];
float step_width, step_height;
/// 300 / 19
if (step_w == 0 || step_h == 0) {
step_width = static_cast<float>(img_width) / feature_width;
step_height = static_cast<float>(img_height) / feature_height;
} else {
step_width = step_w;
step_height = step_h;
}
int num_priors = aspect_ratios.size() * min_sizes.size();
if (!max_sizes.empty()) {
num_priors += max_sizes.size();
}
for (int h = 0; h < feature_height; ++h) {
for (int w = 0; w < feature_width; ++w) {
/// map origin image
float center_x = (w + offset) * step_width;
float center_y = (h + offset) * step_height;
float box_width, box_height;
int idx = 0;
for (size_t s = 0; s < min_sizes.size(); ++s) {
auto min_size = min_sizes[s];
// priors with different aspect ratios
for (float ar : aspect_ratios) {
box_width = min_size * sqrt(ar) / 2.;
box_height = min_size / sqrt(ar) / 2.;
/// box_width/2 , / img_width 为了得到feature map 相对于
/// 原图的归一化位置的比例。
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 0] =
(center_x - box_width) / img_width;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 1] =
(center_y - box_height) / img_height;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 2] =
(center_x + box_width) / img_width;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 3] =
(center_y + box_height) / img_height;
idx++;
}
if (!max_sizes.empty()) {
auto max_size = max_sizes[s];
// square prior with size sqrt(minSize * maxSize)
box_width = box_height = sqrt(min_size * max_size) / 2.;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 0] =
(center_x - box_width) / img_width;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 1] =
(center_y - box_height) / img_height;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 2] =
(center_x + box_width) / img_width;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 3] =
(center_y + box_height) / img_height;
idx++;
}
}
}
}
if (clip) {
math::Transform trans;
ClipFunctor<float> clip_func;
trans(output_boxes_dataptr, output_boxes_dataptr + output_boxes->numel(),
output_boxes_dataptr, clip_func);
}
if ((variances.size() != 4)) {
LOG(kLOG_ERROR) << " variances.size() must be 4.";
}
int64_t box_num = feature_height * feature_width * num_priors;
for (int i = 0; i < box_num; i++) {
output_variances_dataptr[4 * i] = variances[0];
output_variances_dataptr[4 * i + 1] = variances[1];
output_variances_dataptr[4 * i + 2] = variances[2];
output_variances_dataptr[4 * i + 3] = variances[3];
}
PriorBoxCompute<float>(param);
}
} // namespace operators
......
src/operators/kernel/arm/relu_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,101 +14,22 @@ limitations under the License. */
#ifdef RELU_OP
#pragma once
#include "operators/kernel/relu_kernel.h"
#include <operators/math/transform.h>
#include "operators/kernel/central-arm-func/relu_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <typename T>
struct ReluFunctor {
inline T operator()(T in) const { return in > 0 ? in : 0; }
};
template <>
bool ReluKernel<CPU, float>::Init(const ReluParam &para) const {
bool ReluKernel<CPU, float>::Init(ReluParam *param) {
return true;
}
/*
* @b 特化到具体平台的实现, param 从 op 层传入
* */
template <>
void ReluKernel<CPU, float>::Compute(const ReluParam &param) const {
const auto *input_x = param.InputX();
auto *input_x_ptr = input_x->data<float>();
auto *out = param.Out();
auto *out_ptr = out->mutable_data<float>();
int numel = input_x->numel();
// if (numel > 64) {
// asm volatile(
// "pld [%[input_x_ptr], #0] \n\t"
// "vmov.f32 q8, #0.0 \n\t"
// "subs %[num], %[num], #32 \n\t"
// "blt end_num_%= \n\t"
// "loop_num_%=: \n\t"
// "pld [%[input_x_ptr], #1024] \n\t"
//
// "vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
//
// "vmax.f32 q0, q0, q8 \n\t"
// "vmax.f32 q1, q1, q8 \n\t"
// "vmax.f32 q2, q2, q8 \n\t"
// "vmax.f32 q3, q3, q8 \n\t"
// "vmax.f32 q4, q4, q8 \n\t"
// "vmax.f32 q5, q5, q8 \n\t"
// "vmax.f32 q6, q6, q8 \n\t"
// "vmax.f32 q7, q7, q8 \n\t"
//
// "vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
// "vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
// "vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
// "vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
//
// "subs %[num], %[num], #32 \n\t"
// "bge loop_num_%= \n\t"
// "end_num_%=: \n\t"
// "cmp %[num], #0 \n\t"
// "bge end_%= \n\t"
// "mov r6, #4 \n\t"
// "mul r5, %[num], r6 \n\t"
// "add %[input_x_ptr], %[input_x_ptr], r5 \n\t"
// "vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
// "vmax.f32 q0, q0, q8 \n\t"
// "vmax.f32 q1, q1, q8 \n\t"
// "vmax.f32 q2, q2, q8 \n\t"
// "vmax.f32 q3, q3, q8 \n\t"
// "vmax.f32 q4, q4, q8 \n\t"
// "vmax.f32 q5, q5, q8 \n\t"
// "vmax.f32 q6, q6, q8 \n\t"
// "vmax.f32 q7, q7, q8 \n\t"
// "add %[out_ptr], %[out_ptr], r5 \n\t"
// "vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
// "vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
// "vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
// "vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
// "end_%=: \n\t"
// :
// :
// [out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num]
// "r"(numel) : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6",
// "q7", "q8", "r5",
// "r6");
// } else {
ReluFunctor<float> func_;
math::Transform trans;
trans(input_x_ptr, input_x_ptr + numel, out_ptr, func_);
// }
ReluCompute<float>(param);
}
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/arm/reshape_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,44 +14,20 @@ limitations under the License. */
#ifdef RESHAPE_OP
#pragma once
#include "operators/kernel/reshape_kernel.h"
#include "operators/kernel/central-arm-func/reshape_arm_func.h"
namespace paddle_mobile {
namespace operators {
template <>
bool ReshapeKernel<CPU, float>::Init(const ReshapeParam &para) const {
bool ReshapeKernel<CPU, float>::Init(ReshapeParam *param) {
return true;
}
template <>
void ReshapeKernel<CPU, float>::Compute(const ReshapeParam &param) const {
const auto *input_x = param.InputX();
const auto &input_x_dims = input_x->dims();
auto *out = param.Out();
framework::DDim out_dims = out->dims();
const auto *input_shape = param.InputShape();
if (input_shape) {
auto *shape_data = input_shape->data<int>();
framework::Tensor cpu_shape_tensor;
auto shape =
std::vector<int>(shape_data, shape_data + input_shape->numel());
out_dims = ValidateShape(shape, input_x->dims());
}
bool inplace = param.Inplace();
out->Resize(out_dims);
if (!inplace) {
out->mutable_data<float>();
framework::TensorCopy(*input_x, out);
out->Resize(out_dims);
} else {
out->ShareDataWith(*input_x);
out->Resize(out_dims);
}
ReshapeCompute<float>(param);
}
} // namespace operators
......
src/operators/kernel/arm/sigmoid_kernel.cpp
浏览文件 @ c67b701f
......@@ -15,74 +15,25 @@ limitations under the License. */
#ifdef SIGMOID_OP
#include "../sigmoid_kernel.h"
#include "../central-arm-func/sigmoid_arm_func.h"
#if __ARM_NEON
#include "../../math/math_func_neon.h"
#endif
#include <cmath>
namespace paddle_mobile {
namespace operators {
using framework::DDim;
using framework::Tensor;
void sigmoid(const Tensor *X, Tensor *Y) {
#if __ARM_NEON
const float *input = X->data<float>();
float *output = Y->mutable_data<float>();
const DDim &dDim = X->dims();
int axis_index = 1;
if (dDim.size() < 4) {
axis_index = 0;
}
DDim outer_ddim =
paddle_mobile::framework::slice_ddim(dDim, 0, axis_index + 1);
DDim inner_ddim =
paddle_mobile::framework::slice_ddim(dDim, axis_index + 1, dDim.size());
int out_size = paddle_mobile::framework::product(outer_ddim);
int inner_size = paddle_mobile::framework::product(inner_ddim);
DLOG << "outsize=" << out_size;
DLOG << "innersize=" << inner_size;
#pragma omp parallel for
for (int i = 0; i < out_size; ++i) {
const float *input_outer_ptr = input + i * inner_size;
float *output_outer_ptr = output + i * inner_size;
int nn = inner_size >> 2;
int remain = inner_size - (nn << 2);
float32x4_t _one = vdupq_n_f32(1.f);
for (; nn > 0; nn--) {
float32x4_t data = vld1q_f32(input_outer_ptr);
data = vnegq_f32(data);
data = exp_ps(data);
data = vaddq_f32(data, _one);
float32x4_t out_data = vrecpeq_f32(data);
out_data = vmulq_f32(vrecpsq_f32(data, out_data), out_data);
vst1q_f32(output_outer_ptr, out_data);
input_outer_ptr += 4;
output_outer_ptr += 4;
}
for (; remain > 0; remain--) {
*output_outer_ptr = 1.f / (1.f + exp(-*input_outer_ptr));
output_outer_ptr++;
input_outer_ptr++;
}
}
#endif
}
template <>
bool SigmoidKernel<CPU, float>::Init(const SigmoidParam &para) const {
bool SigmoidKernel<CPU, float>::Init(SigmoidParam *param) {
return true;
}
template <>
void SigmoidKernel<CPU, float>::Compute(const SigmoidParam &param) const {
const Tensor *in_x = param.InputX();
Tensor *out = param.Out();
auto x_dims = in_x->dims();
out->Resize(x_dims);
sigmoid(in_x, out);
SigmoidCompute<float>(param);
}
template class SigmoidKernel<CPU, float>;
......
src/operators/kernel/arm/softmax_kernel.cpp
浏览文件 @ c67b701f
......@@ -15,22 +15,19 @@ limitations under the License. */
#ifdef SOFTMAX_OP
#include "../softmax_kernel.h"
#include "../../math/softmax.h"
#include "../central-arm-func/softmax_arm_func.h"
#include "operators/math/softmax.h"
namespace paddle_mobile {
namespace operators {
template <>
bool SoftmaxKernel<CPU, float>::Init(const SoftmaxParam &para) const {
bool SoftmaxKernel<CPU, float>::Init(SoftmaxParam *param) {
return true;
}
template <>
void SoftmaxKernel<CPU, float>::Compute(const SoftmaxParam &param) const {
const Tensor *in_x = param.InputX();
Tensor *out = param.Out();
auto x_dims = in_x->dims();
out->Resize(x_dims);
math::SoftmaxFuntor<CPU, float>()(in_x, out);
SoftmaxCompute<float>(param);
}
template class SoftmaxKernel<CPU, float>;
......
src/operators/kernel/arm/transpose_kernel.cpp
浏览文件 @ c67b701f
......@@ -14,72 +14,19 @@ limitations under the License. */
#ifdef TRANSPOSE_OP
#include "operators/kernel/transpose_kernel.h"
#include "operators/kernel/central-arm-func/transpose_arm_func.h"
namespace paddle_mobile {
namespace operators {
// vector<int> pos;
// template <typename T>
// void TransposeFunc(const int numel, const T* input, const vector<int> axis,
// const vector<int> old_strides, const vector<int>
// new_strides, T* output) {
// for (int i = 0; i < numel; ++i) {
// int old_idx = 0;
// int idx = i;
// for (int j = 0; j < axis.size(); ++j) {
// int order = axis[j];
// old_idx += (idx / new_strides[j]) * old_strides[order];
// idx %= new_strides[j];
// }
// output[i] = input[old_idx];
// }
// }
template <>
bool TransposeKernel<CPU, float>::Init(const TransposeParam& para) const {
bool TransposeKernel<CPU, float>::Init(TransposeParam *param) {
return true;
}
template <>
void TransposeKernel<CPU, float>::Compute(const TransposeParam& param) const {
const auto* input_x = param.InputX();
const auto input_x_dims = input_x->dims();
auto* out = param.Out();
const auto axis = param.Axis();
const auto* input_x_data = input_x->data<float>();
auto* out_data = out->mutable_data<float>();
size_t ndim = axis.size();
std::vector<int> xdim(ndim);
std::vector<int> xstride(ndim);
std::vector<int> xout(ndim);
for (int i = 0; i < ndim; i++) {
int j = ndim - 1 - i;
xdim[j] = input_x_dims[axis[i]];
xstride[j] = 1;
for (int k = axis[i] + 1; k < ndim; k++) {
xstride[j] *= input_x_dims[k];
}
xout[j] = xstride[j] * xdim[j];
}
auto numel = input_x->numel();
size_t pind = 0;
std::vector<int> ind(ndim);
for (int i = 0; i < numel; i++) {
out_data[i] = input_x_data[pind];
ind[0]++;
pind += xstride[0];
for (int j = 0; j < ndim - 1; j++) {
if (ind[j] == xdim[j]) {
ind[j + 1]++;
ind[j] = 0;
pind += xstride[j + 1];
pind -= xout[j];
} else {
break;
}
}
}
void TransposeKernel<CPU, float>::Compute(const TransposeParam &param) const {
TransposeCompute<float>(param);
}
} // namespace operators
......
src/operators/kernel/batchnorm_kernel.h
浏览文件 @ c67b701f
......@@ -29,7 +29,7 @@ class BatchNormKernel
: public framework::OpKernelBase<DeviceType, BatchNormParam> {
public:
void Compute(const BatchNormParam &param) const;
bool Init(const BatchNormParam &para) const;
bool Init(BatchNormParam *param);
};
} // namespace operators
......
src/operators/kernel/box_coder_kernel.h
浏览文件 @ c67b701f
......@@ -30,7 +30,7 @@ class BoxCoderKernel
: public framework::OpKernelBase<DeviceType, BoxCoderParam> {
public:
void Compute(const BoxCoderParam& param) const;
bool Init(const BoxCoderParam& para) const;
bool Init(BoxCoderParam* param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/central-arm-func/batchnorm_arm_func.h
浏览文件 @ c67b701f
......@@ -16,6 +16,7 @@ limitations under the License. */
#pragma once
#include <cmath>
#include "operators/op_param.h"
namespace paddle_mobile {
......@@ -52,6 +53,8 @@ void BatchnormCompute(const BatchNormParam &param) {
"C must equal to variance.numel()");
int HXW = H * W;
#ifdef ARMV7
if (HXW > 32) {
int NXC = N * C;
float *inv_std_ptr = new float[NXC * 4];
......@@ -226,6 +229,37 @@ void BatchnormCompute(const BatchNormParam &param) {
delete[] inv_std_ptr;
}
#endif
float *inv_std_ptr = new float[C];
for (int i = 0; i < C; i++) {
inv_std_ptr[i] =
1 / static_cast<float>(pow((variance_ptr[i] + epsilon), 0.5));
}
Tensor new_scale;
auto new_scale_ptr = new_scale.mutable_data<float>(framework::make_ddim({C}));
Tensor new_bias;
auto new_bias_ptr = new_bias.mutable_data<float>(framework::make_ddim({C}));
/// ((x - est_mean) * (inv_var) * scale + bias equal to
/// (x * inv_var * scale) + (bias - est_mean * inv_var * scale)
for (int i = 0; i < C; i++) {
new_scale_ptr[i] = inv_std_ptr[i] * scale_ptr[i];
new_bias_ptr[i] = bias_ptr[i] - mean_ptr[i] * inv_std_ptr[i] * scale_ptr[i];
{
for (int n = 0; n < N; n++) {
for (int h = 0; h < H; h++) {
int tmp_index = n * stride0 + i * stride1 + h * stride2;
for (int w = 0; w < W; w++) {
int index = tmp_index + w;
out_ptr[index] =
input_x_ptr[index] * new_scale_ptr[i] + new_bias_ptr[i];
}
}
}
}
}
delete[] inv_std_ptr;
}
} // namespace operators
......
src/operators/kernel/central-arm-func/box_coder_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef BOXCODER_OP
#pragma once
#include <cmath>
namespace paddle_mobile {
namespace operators {
template <typename T>
void EncodeCenterSize(const framework::Tensor& target_box,
const framework::Tensor& prior_box,
const framework::Tensor& prior_box_var, T* output) {
int64_t row = target_box.dims()[0];
int64_t col = prior_box.dims()[0];
int64_t len = prior_box.dims()[1];
auto* target_box_data = target_box.data<T>();
auto* prior_box_data = prior_box.data<T>();
auto* prior_box_var_data = prior_box_var.data<T>();
for (int64_t i = 0; i < row; ++i) {
for (int64_t j = 0; j < col; ++j) {
T prior_box_width = prior_box_data[j * len + 2] - prior_box_data[j * len];
T prior_box_height =
prior_box_data[j * len + 3] - prior_box_data[j * len + 1];
T prior_box_center_x =
(prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
T prior_box_center_y =
(prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
T target_box_center_x =
(target_box_data[i * len + 2] + target_box_data[i * len]) / 2;
T target_box_center_y =
(target_box_data[i * len + 3] + target_box_data[i * len + 1]) / 2;
T target_box_width =
target_box_data[i * len + 2] - target_box_data[i * len];
T target_box_height =
target_box_data[i * len + 3] - target_box_data[i * len + 1];
size_t offset = i * col * len + j * len;
output[offset] = (target_box_center_x - prior_box_center_x) /
prior_box_width / prior_box_var_data[j * len];
output[offset + 1] = (target_box_center_y - prior_box_center_y) /
prior_box_height / prior_box_var_data[j * len + 1];
output[offset + 2] =
std::log(std::fabs(target_box_width / prior_box_width)) /
prior_box_var_data[j * len + 2];
output[offset + 3] =
std::log(std::fabs(target_box_height / prior_box_height)) /
prior_box_var_data[j * len + 3];
}
}
}
template <typename T>
void DecodeCenterSize(const framework::Tensor& target_box,
const framework::Tensor& prior_box,
const framework::Tensor& prior_box_var, T* output) {
int64_t row = target_box.dims()[0];
int64_t col = prior_box.dims()[0];
int64_t len = prior_box.dims()[1];
auto* target_box_data = target_box.data<T>();
auto* prior_box_data = prior_box.data<T>();
auto* prior_box_var_data = prior_box_var.data<T>();
for (int64_t i = 0; i < row; ++i) {
for (int64_t j = 0; j < col; ++j) {
size_t offset = i * col * len + j * len;
T prior_box_width = prior_box_data[j * len + 2] - prior_box_data[j * len];
T prior_box_height =
prior_box_data[j * len + 3] - prior_box_data[j * len + 1];
T prior_box_center_x =
(prior_box_data[j * len + 2] + prior_box_data[j * len]) / 2;
T prior_box_center_y =
(prior_box_data[j * len + 3] + prior_box_data[j * len + 1]) / 2;
T target_box_center_x = prior_box_var_data[j * len] *
target_box_data[offset] * prior_box_width +
prior_box_center_x;
T target_box_center_y = prior_box_var_data[j * len + 1] *
target_box_data[offset + 1] *
prior_box_height +
prior_box_center_y;
T target_box_width = std::exp(prior_box_var_data[j * len + 2] *
target_box_data[offset + 2]) *
prior_box_width;
T target_box_height = std::exp(prior_box_var_data[j * len + 3] *
target_box_data[offset + 3]) *
prior_box_height;
output[offset] = target_box_center_x - target_box_width / 2;
output[offset + 1] = target_box_center_y - target_box_height / 2;
output[offset + 2] = target_box_center_x + target_box_width / 2;
output[offset + 3] = target_box_center_y + target_box_height / 2;
}
}
}
template <typename P>
void BoxCoderCompute(const BoxCoderParam& param) {
const auto* input_priorbox = param.InputPriorBox();
const auto* input_priorboxvar = param.InputPriorBoxVar();
const auto* input_targetbox = param.InputTargetBox();
const auto& code_type = param.CodeType();
auto row = input_targetbox->dims()[0];
auto col = input_priorbox->dims()[0];
auto len = input_priorbox->dims()[1];
Tensor* output_box = param.OutputBox();
auto* output_box_dataptr = output_box->mutable_data<float>({row, col, len});
if (code_type == "encode_center_size") {
EncodeCenterSize<float>(*input_targetbox, *input_priorbox,
*input_priorboxvar, output_box_dataptr);
}
if (code_type == "decode_center_size") {
DecodeCenterSize<float>(*input_targetbox, *input_priorbox,
*input_priorboxvar, output_box_dataptr);
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/concat_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef CONCAT_OP
#pragma once
#include <vector>
namespace paddle_mobile {
namespace operators {
template <typename T>
class ConcatFunctor {
public:
void operator()(const std::vector<framework::Tensor> &input, const int axis,
framework::Tensor *output) {
size_t num = input.size();
int rows = 1;
auto dim_0 = input[0].dims();
for (int i = 0; i < axis; ++i) {
rows *= dim_0[i];
}
int out_rows = rows, out_cols = 0;
std::vector<int64_t> input_cols(input.size());
for (int i = 0; i < num; ++i) {
int t_cols = input[i].numel() / rows;
out_cols += t_cols;
input_cols[i] = t_cols;
}
// computation
for (int k = 0; k < out_rows; ++k) {
T *dst_ptr = output->data<T>() + k * out_cols;
int col_idx = 0;
for (int j = 0; j < num; ++j) {
int col_len = input_cols[j];
const T *src_prt = input[j].data<T>() + k * col_len;
memory::Copy(dst_ptr + col_idx, src_prt, sizeof(T) * col_len);
col_idx += col_len;
}
}
}
};
template <typename P>
void ConcatCompute(const ConcatParam &param) {
auto inputs = param.Inputs();
auto *out = param.Out();
int64_t axis = param.Axis();
out->mutable_data<float>();
/// Sometimes direct copies will be faster, this maybe need deeply analysis.
if (axis == 0 && inputs.size() < 10) {
size_t output_offset = 0;
for (auto *in : inputs) {
auto in_stride = framework::stride_numel(in->dims());
auto out_stride = framework::stride_numel(out->dims());
auto dst = out->data<float>() + output_offset;
auto src = in->data<float>();
PADDLE_MOBILE_ENFORCE(
in_stride.size() == out_stride.size(),
"src and dst tensor should have the same dims size.");
memory::Copy(dst, src, sizeof(float) * in_stride[0]);
output_offset += in_stride[0];
}
} else {
std::vector<framework::Tensor> inputs_concat(inputs.size());
for (int j = 0; j < inputs.size(); ++j) {
inputs_concat[j] = *inputs[j];
}
ConcatFunctor<float> concat_functor;
concat_functor(inputs_concat, static_cast<int>(axis), out);
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/conv_add_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef FUSION_CONVADD_OP
#pragma once
#include <vector>
#include "operators/math/conv_func.h"
#include "operators/math/depthwise_conv_3x3.h"
#include "operators/math/im2col.h"
#include "operators/math/math_function.h"
#include "operators/math/vol2col.h"
#include "operators/op_param.h"
namespace paddle_mobile {
namespace operators {
void ConvAddBasic(const FusionConvAddParam &param) {
const Tensor *input = param.Input();
Tensor filter = *param.Filter();
Tensor bias = *param.Bias();
int axis = param.Axis();
Tensor *output = param.Output();
math::expand_bias(bias, axis, output->dims());
output->ShareDataWith(bias);
int groups = param.Groups();
std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings();
std::vector<int> dilations = param.Dilations();
const int batch_size = static_cast<int>(input->dims()[0]);
std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
size_t data_dim = filter_shape_vec.size() - 2;
std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
col_shape_vec[0] = input->dims()[1] / groups;
for (size_t j = 0; j < data_dim; ++j) {
col_shape_vec[j + 1] = filter_shape_vec[j + 2];
col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
}
framework::DDim col_shape(framework::make_ddim(col_shape_vec));
framework::DDim col_matrix_shape =
framework::flatten_to_2d(col_shape, data_dim + 1);
bool is_expand =
math::IsExpand(filter_shape_vec, strides, paddings, dilations);
Tensor col;
Tensor col_matrix;
if (is_expand) {
col.mutable_data<float>(col_shape);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
filter.Resize(filter_matrix_shape);
framework::DDim output_matrix_shape = {
output->dims()[1],
output->numel() / (output->dims()[0] * output->dims()[1])};
// convolution operator: im2col(or vol2col) + gemm
int in_step = static_cast<int>(input->dims()[1]) / groups;
int out_step = static_cast<int>(output->dims()[1]) / groups;
math::Vol2ColFunctor<CPU, float> vol2col;
math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col;
for (int i = 0; i < batch_size; i++) {
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
for (int g = 0; g < groups; g++) {
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
if (!is_expand) {
col.ShareDataWith(in_slice);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
} else if (data_dim == 2U) {
// im2col
im2col(in_slice, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&col);
} else if (data_dim == 3U) {
// vol2col
vol2col(in_slice, dilations, strides, paddings, &col);
}
// gemm
Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
math::matmul<float>(filter_slice, false, col_matrix, false,
static_cast<float>(1), &out_slice,
static_cast<float>(1));
}
}
}
template <typename P>
void ConvAddCompute(const FusionConvAddParam &param) {
if (param.Groups() == param.Input()->dims()[1] &&
param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
param.Filter()->dims()[2] == 3 && param.Strides()[0] == 1) {
math::DepthwiseConv3x3s1p1(param.Input(), param.Filter(), param.Output(),
param.Bias(), true);
} else if (param.Groups() == param.Input()->dims()[1] &&
param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
param.Filter()->dims()[2] == 3) {
math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
param.Filter(), param.Bias(), param.Output(), true);
} else {
ConvAddBasic(param);
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/conv_add_bn_relu_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef FUSION_CONVADDBNRELU_OP
#pragma once
#include "operators/math/depthwise_conv_3x3.h"
#include "operators/op_param.h"
namespace paddle_mobile {
namespace operators {
void ConvAddBNReluBasic(const FusionConvAddBNReluParam &param) {
const Tensor *input = param.Input();
Tensor filter = *param.Filter();
Tensor bias = *param.Bias();
Tensor new_bias = *param.NewBias();
Tensor new_scale = *param.NewScale();
auto new_bias_ptr = new_bias.data<float>();
auto new_scale_ptr = new_scale.data<float>();
int axis = param.Axis();
Tensor *output = param.Output();
math::expand_bias(bias, axis, output->dims());
output->ShareDataWith(bias);
int groups = param.Groups();
std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings();
std::vector<int> dilations = param.Dilations();
const int batch_size = static_cast<int>(input->dims()[0]);
std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
size_t data_dim = filter_shape_vec.size() - 2;
std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
col_shape_vec[0] = input->dims()[1] / groups;
for (size_t j = 0; j < data_dim; ++j) {
col_shape_vec[j + 1] = filter_shape_vec[j + 2];
col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
}
framework::DDim col_shape(framework::make_ddim(col_shape_vec));
framework::DDim col_matrix_shape =
framework::flatten_to_2d(col_shape, data_dim + 1);
bool is_expand =
math::IsExpand(filter_shape_vec, strides, paddings, dilations);
Tensor col;
Tensor col_matrix;
if (is_expand) {
col.mutable_data<float>(col_shape);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
filter.Resize(filter_matrix_shape);
framework::DDim output_matrix_shape = {
output->dims()[1],
output->numel() / (output->dims()[0] * output->dims()[1])};
// convolution operator: im2col(or vol2col) + gemm
int in_step = static_cast<int>(input->dims()[1]) / groups;
int out_step = static_cast<int>(output->dims()[1]) / groups;
math::Vol2ColFunctor<CPU, float> vol2col;
math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col;
for (int i = 0; i < batch_size; i++) {
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
for (int g = 0; g < groups; g++) {
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
if (!is_expand) {
col.ShareDataWith(in_slice);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
} else if (data_dim == 2U) {
// im2col
im2col(in_slice, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&col);
} else if (data_dim == 3U) {
// vol2col
vol2col(in_slice, dilations, strides, paddings, &col);
}
// gemm
Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
math::matmul<float>(filter_slice, false, col_matrix, false,
static_cast<float>(1), &out_slice,
static_cast<float>(1));
}
}
/// todo : use neon in special case instead of 2for(300ms)
auto output_ptr = output->data<float>();
for (int c = 0; c < output_matrix_shape[0]; c++) {
int start = c * output_matrix_shape[1];
for (int j = 0; j < output_matrix_shape[1]; j++) {
output_ptr[start + j] =
output_ptr[start + j] * new_scale_ptr[c] + new_bias_ptr[c];
output_ptr[start + j] =
output_ptr[start + j] < 0 ? 0 : output_ptr[start + j];
}
}
}
template <typename P>
void ConvAddBNReluCompute(const FusionConvAddBNReluParam &param) {
Tensor Bias;
Bias.mutable_data<float>({param.Groups()});
if (param.Groups() == param.Input()->dims()[1] &&
param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
param.Filter()->dims()[2] == 3 && param.Strides()[0] == 1) {
math::DepthwiseConvAddBNRelu3x3s1p1(
param.Input(), param.Filter(), param.Output(), &Bias, 1,
param.NewScale(), param.NewBias(), 1, 1);
} else if (0 && param.Groups() == param.Input()->dims()[1] &&
param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
param.Filter()->dims()[2] == 3 && param.Strides()[0] == 2) {
math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
param.Filter(), &Bias, param.Output(), false);
} else {
ConvAddBNReluBasic(param);
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/conv_arm_func.h
浏览文件 @ c67b701f
......@@ -17,13 +17,15 @@ limitations under the License. */
#pragma once
#include <vector>
#include "operators/math/conv_func.h"
#include "operators/math/depthwise_conv_3x3.h"
#include "operators/math/im2col.h"
#include "operators/math/math_function.h"
#include "operators/math/vol2col.h"
#include "operators/op_param.h"
namespace paddle_mobile {
namespace operators {
template <typename P>
void ConvCompute(const ConvParam &param) {
inline void ConvBasic(const ConvParam &param) {
const Tensor *input = param.Input();
Tensor filter = *param.Filter();
Tensor *output = param.Output();
......@@ -109,6 +111,25 @@ void ConvCompute(const ConvParam &param) {
}
}
template <typename P>
void ConvCompute(const ConvParam &param) {
if (param.Groups() == param.Input()->dims()[1] &&
param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
param.Filter()->dims()[2] == 3 && param.Strides()[0] == 1) {
math::DepthwiseConv3x3s1p1(param.Input(), param.Filter(), param.Output(),
nullptr, false);
} else if (param.Groups() == param.Input()->dims()[1] &&
param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
param.Filter()->dims()[2] == 3) {
math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
param.Filter(), nullptr, param.Output(), false);
} else {
ConvBasic(param);
}
}
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/central-arm-func/depthwise_conv_arm_func.h
浏览文件 @ c67b701f
......@@ -15,8 +15,10 @@ limitations under the License. */
#ifdef DEPTHWISECONV_OP
#pragma once
#include <operators/math/depthwise_conv_3x3.h>
#include <vector>
#include "operators/math/conv_func.h"
#include "operators/kernel/central-arm-func/conv_arm_func.h"
#include "operators/op_param.h"
namespace paddle_mobile {
......@@ -24,89 +26,21 @@ namespace operators {
template <typename P>
void DepthwiseConvCompute(const ConvParam &param) {
const Tensor *input = param.Input();
Tensor filter = *param.Filter();
Tensor *output = param.Output();
output->mutable_data<float>();
int groups = param.Groups();
std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings();
std::vector<int> dilations = param.Dilations();
// DLOG << " compute end get Attrs " << strides[0];
const int batch_size = static_cast<int>(input->dims()[0]);
std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
size_t data_dim = filter_shape_vec.size() - 2;
std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
col_shape_vec[0] = input->dims()[1] / groups;
for (size_t j = 0; j < data_dim; ++j) {
col_shape_vec[j + 1] = filter_shape_vec[j + 2];
col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
}
framework::DDim col_shape(framework::make_ddim(col_shape_vec));
framework::DDim col_matrix_shape =
framework::flatten_to_2d(col_shape, data_dim + 1);
bool is_expand =
math::IsExpand(filter_shape_vec, strides, paddings, dilations);
Tensor col;
Tensor col_matrix;
if (is_expand) {
col.mutable_data<float>(col_shape);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
filter.Resize(filter_matrix_shape);
framework::DDim output_matrix_shape = {
output->dims()[1],
output->numel() / (output->dims()[0] * output->dims()[1])};
// convolution operator: im2col(or vol2col) + gemm
int in_step = static_cast<int>(input->dims()[1]) / groups;
int out_step = static_cast<int>(output->dims()[1]) / groups;
math::Vol2ColFunctor<CPU, float> vol2col;
math::Im2ColFunctor<math::ColFormat::kCFO, CPU, float> im2col;
for (int i = 0; i < batch_size; i++) {
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
for (int g = 0; g < groups; g++) {
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
if (!is_expand) {
col.ShareDataWith(in_slice);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
} else if (data_dim == 2U) {
// im2col
im2col(in_slice, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&col);
} else if (data_dim == 3U) {
// vol2col
vol2col(in_slice, dilations, strides, paddings, &col);
}
// gemm
Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
math::matmul<float>(filter_slice, false, col_matrix, false,
static_cast<float>(1), &out_slice,
static_cast<float>(0));
}
Tensor Bias;
Bias.mutable_data<float>({param.Groups()});
if (param.Groups() == param.Input()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
param.Filter()->dims()[2] == 3 && param.Strides()[0] == 1) {
math::DepthwiseConv3x3s1p1(param.Input(), param.Filter(), param.Output(),
&Bias, false);
} else if (param.Groups() == param.Input()->dims()[1] &&
param.Input()->dims()[1] == param.Output()->dims()[1] &&
param.Filter()->dims()[2] == param.Filter()->dims()[3] &&
param.Filter()->dims()[2] == 3 && param.Strides()[0] == 2) {
math::DepthwiseConv3x3(param.Input(), param.Strides(), param.Paddings(),
param.Filter(), &Bias, param.Output(), false);
} else {
ConvBasic(param);
}
}
......
src/operators/kernel/central-arm-func/elementwise_add_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef ELEMENTWISEADD_OP
#pragma once
namespace paddle_mobile {
namespace operators {
template <typename T>
struct AddFunctor {
inline T operator()(T a, T b) const { return a + b; }
};
template <typename P>
void ElementwiseAddCompute(const ElementwiseAddParam &param) {
const Tensor *input_x = param.InputX();
const Tensor *input_y = param.InputY();
Tensor *Out = param.Out();
Out->mutable_data<float>();
int axis = param.Axis();
ElementwiseComputeEx<AddFunctor<float>, float>(input_x, input_y, axis,
AddFunctor<float>(), Out);
}
template class ElementwiseAddKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/fusion_fc_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef FUSION_FC_OP
#pragma once
namespace paddle_mobile {
namespace operators {
template <typename P>
void FusionFcCompute(const FusionFcParam &param) {
const Tensor *input_x = param.InputX();
const Tensor *input_y = param.InputY();
const Tensor *input_z = param.InputZ();
auto *input_z_data = input_z->data<float>();
int axis = param.Axis();
Tensor *out = param.Out();
auto *out_data = out->mutable_data<float>();
const Tensor x_matrix =
input_x->dims().size() > 2
? framework::ReshapeToMatrix(*input_x, param.XNumColDims())
: *input_x;
const Tensor y_matrix =
input_y->dims().size() > 2
? framework::ReshapeToMatrix(*input_y, param.YNumColDims())
: *input_y;
auto out_dim = out->dims();
if (out_dim.size() != 2) {
out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
}
PADDLE_MOBILE_ENFORCE(out_dim.size() == 2, " out_dim.size must be 2.");
PADDLE_MOBILE_ENFORCE(input_z->dims().size() == 1, "inpu_z size must be 1");
PADDLE_MOBILE_ENFORCE(out_dim[1] == input_z->dims()[0],
" out_dim.size must be 2.");
axis = (axis == -1 ? out_dim.size() - input_z->dims().size() : axis);
PADDLE_MOBILE_ENFORCE(axis == 1, " to fit broadcast, axis = 1. ")
int64_t classes = input_z->numel();
for (int i = 0; i < out_dim[0]; i++) {
memory::Copy(out_data + i * classes, input_z_data, sizeof(float) * classes);
}
for (int i = 0; i < out->numel(); i++) {
DLOG << out_data[i];
}
math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
out, static_cast<float>(1));
PADDLE_MOBILE_ENFORCE(out_dim.size() == 2, " out_dim.size must be 2.");
// if (out_dim.size() != 2) {
// out->Resize(out_dim);
// }
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/lrn_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef LRN_OP
#pragma once
namespace paddle_mobile {
namespace operators {
template <typename P>
void LrnCompute(const LrnParam &param) {
const Tensor *input_x = param.InputX();
auto x_dims = input_x->dims();
Tensor *out = param.Out();
out->mutable_data<float>();
/// data_format = NCHW
const int N = x_dims[0];
const int C = x_dims[1];
const int H = x_dims[2];
const int W = x_dims[3];
const int n = param.N();
const float alpha = param.Alpha();
const float beta = param.Beta();
const float k = param.K();
LRNFunctor<float> lrnFunctor;
lrnFunctor(*input_x, out, N, C, H, W, n, k, alpha, beta);
}
template class LrnKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/mul_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef MUL_OP
#pragma once
namespace paddle_mobile {
namespace operators {
template <typename P>
void MulCompute(const MulParam &param) {
const Tensor *input_x = param.InputX();
const Tensor *input_y = param.InputY();
Tensor *out = param.Out();
out->mutable_data<float>();
const Tensor x_matrix =
input_x->dims().size() > 2
? framework::ReshapeToMatrix(*input_x, param.XNumColDims())
: *input_x;
const Tensor y_matrix =
input_y->dims().size() > 2
? framework::ReshapeToMatrix(*input_y, param.YNumColDims())
: *input_y;
auto out_dim = out->dims();
if (out_dim.size() != 2) {
out->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
}
math::matmul<float>(x_matrix, false, y_matrix, false, static_cast<float>(1),
out, static_cast<float>(0));
if (out_dim.size() != 2) {
out->Resize(out_dim);
}
}
template class MulKernel<CPU, float>;
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/multiclass_nms_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef MULTICLASSNMS_OP
#pragma once
#include <algorithm>
#include <map>
#include <utility>
#include <vector>
namespace paddle_mobile {
namespace operators {
constexpr int kOutputDim = 6;
constexpr int kBBoxSize = 4;
template <class T>
bool SortScorePairDescend(const std::pair<float, T>& pair1,
const std::pair<float, T>& pair2) {
return pair1.first > pair2.first;
}
template <class T>
static inline void GetMaxScoreIndex(
const std::vector<T>& scores, const T threshold, int top_k,
std::vector<std::pair<T, int>>* sorted_indices) {
for (size_t i = 0; i < scores.size(); ++i) {
if (scores[i] > threshold) {
sorted_indices->push_back(std::make_pair(scores[i], i));
}
}
// Sort the score pair according to the scores in descending order
std::stable_sort(sorted_indices->begin(), sorted_indices->end(),
SortScorePairDescend<int>);
// Keep top_k scores if needed.
if (top_k > -1 && top_k < static_cast<int>(sorted_indices->size())) {
sorted_indices->resize(top_k);
}
}
template <class T>
static inline T BBoxArea(const T* box, const bool normalized) {
if (box[2] < box[0] || box[3] < box[1]) {
// If coordinate values are is invalid
// (e.g. xmax < xmin or ymax < ymin), return 0.
return static_cast<T>(0.);
} else {
const T w = box[2] - box[0];
const T h = box[3] - box[1];
if (normalized) {
return w * h;
} else {
// If coordinate values are not within range [0, 1].
return (w + 1) * (h + 1);
}
}
}
template <class T>
static inline T JaccardOverlap(const T* box1, const T* box2,
const bool normalized) {
if (box2[0] > box1[2] || box2[2] < box1[0] || box2[1] > box1[3] ||
box2[3] < box1[1]) {
return static_cast<T>(0.);
} else {
const T inter_xmin = std::max(box1[0], box2[0]);
const T inter_ymin = std::max(box1[1], box2[1]);
const T inter_xmax = std::min(box1[2], box2[2]);
const T inter_ymax = std::min(box1[3], box2[3]);
const T inter_w = inter_xmax - inter_xmin;
const T inter_h = inter_ymax - inter_ymin;
const T inter_area = inter_w * inter_h;
const T bbox1_area = BBoxArea<T>(box1, normalized);
const T bbox2_area = BBoxArea<T>(box2, normalized);
return inter_area / (bbox1_area + bbox2_area - inter_area);
}
}
template <typename T>
static inline void NMSFast(const Tensor& bbox, const Tensor& scores,
const T score_threshold, const T nms_threshold,
const T eta, const int64_t top_k,
std::vector<int>* selected_indices) {
// The total boxes for each instance.
int64_t num_boxes = bbox.dims()[0];
// 4: [xmin ymin xmax ymax]
int64_t box_size = bbox.dims()[1];
std::vector<T> scores_data(num_boxes);
std::copy_n(scores.data<T>(), num_boxes, scores_data.begin());
std::vector<std::pair<T, int>> sorted_indices;
GetMaxScoreIndex(scores_data, score_threshold, top_k, &sorted_indices);
selected_indices->clear();
T adaptive_threshold = nms_threshold;
const T* bbox_data = bbox.data<T>();
while (sorted_indices.size() != 0) {
const int idx = sorted_indices.front().second;
bool keep = true;
for (size_t k = 0; k < selected_indices->size(); ++k) {
if (keep) {
const int kept_idx = (*selected_indices)[k];
T overlap = JaccardOverlap<T>(bbox_data + idx * box_size,
bbox_data + kept_idx * box_size, true);
keep = overlap <= adaptive_threshold;
} else {
break;
}
}
if (keep) {
selected_indices->push_back(idx);
}
sorted_indices.erase(sorted_indices.begin());
if (keep && eta < 1 && adaptive_threshold > 0.5) {
adaptive_threshold *= eta;
}
}
}
template <typename T>
void MultiClassNMS(const Tensor& scores, const Tensor& bboxes,
std::map<int, std::vector<int>>* indices, int* num_nmsed_out,
const int& background_label, const int& nms_top_k,
const int& keep_top_k, const T& nms_threshold,
const T& nms_eta, const T& score_threshold) {
int64_t class_num = scores.dims()[0];
int64_t predict_dim = scores.dims()[1];
int num_det = 0;
for (int64_t c = 0; c < class_num; ++c) {
if (c == background_label) continue;
Tensor score = scores.Slice(c, c + 1);
/// [c] is key
NMSFast<float>(bboxes, score, score_threshold, nms_threshold, nms_eta,
nms_top_k, &((*indices)[c]));
num_det += (*indices)[c].size();
}
*num_nmsed_out = num_det;
const T* scores_data = scores.data<T>();
if (keep_top_k > -1 && num_det > keep_top_k) {
std::vector<std::pair<float, std::pair<int, int>>> score_index_pairs;
for (const auto& it : *indices) {
int label = it.first;
const T* sdata = scores_data + label * predict_dim;
const std::vector<int>& label_indices = it.second;
for (size_t j = 0; j < label_indices.size(); ++j) {
int idx = label_indices[j];
// PADDLE_ENFORCE_LT(idx, predict_dim);
score_index_pairs.push_back(
std::make_pair(sdata[idx], std::make_pair(label, idx)));
}
}
// Keep top k results per image.
std::stable_sort(score_index_pairs.begin(), score_index_pairs.end(),
SortScorePairDescend<std::pair<int, int>>);
score_index_pairs.resize(keep_top_k);
// Store the new indices.
std::map<int, std::vector<int>> new_indices;
for (size_t j = 0; j < score_index_pairs.size(); ++j) {
int label = score_index_pairs[j].second.first;
int idx = score_index_pairs[j].second.second;
new_indices[label].push_back(idx);
}
new_indices.swap(*indices);
*num_nmsed_out = keep_top_k;
}
}
template <typename T>
void MultiClassOutput(const Tensor& scores, const Tensor& bboxes,
const std::map<int, std::vector<int>>& selected_indices,
Tensor* outs) {
int predict_dim = scores.dims()[1];
auto* scores_data = scores.data<T>();
auto* bboxes_data = bboxes.data<T>();
auto* odata = outs->data<T>();
int count = 0;
for (const auto& it : selected_indices) {
/// one batch
int label = it.first;
const T* sdata = scores_data + label * predict_dim;
const std::vector<int>& indices = it.second;
for (size_t j = 0; j < indices.size(); ++j) {
int idx = indices[j];
const T* bdata = bboxes_data + idx * kBBoxSize;
odata[count * kOutputDim] = label; // label
odata[count * kOutputDim + 1] = sdata[idx]; // score
// xmin, ymin, xmax, ymax
std::memcpy(odata + count * kOutputDim + 2, bdata, 4 * sizeof(T));
count++;
}
}
}
template <typename P>
void MultiClassNMSCompute(const MultiClassNMSParam& param) {
const auto* input_bboxes = param.InputBBoxes();
const auto& input_bboxes_dims = input_bboxes->dims();
const auto* input_scores = param.InputScores();
const auto& input_scores_dims = input_scores->dims();
auto* outs = param.Out();
auto background_label = param.BackGroundLabel();
auto nms_top_k = param.NMSTopK();
auto keep_top_k = param.KeepTopK();
auto nms_threshold = param.NMSThreshold();
auto nms_eta = param.NMSEta();
auto score_threshold = param.ScoreThreshold();
int64_t batch_size = input_scores_dims[0];
int64_t class_num = input_scores_dims[1];
int64_t predict_dim = input_scores_dims[2];
int64_t box_dim = input_bboxes_dims[2];
std::vector<std::map<int, std::vector<int>>> all_indices;
std::vector<size_t> batch_starts = {0};
for (int64_t i = 0; i < batch_size; ++i) {
Tensor ins_score = input_scores->Slice(i, i + 1);
ins_score.Resize({class_num, predict_dim});
Tensor ins_boxes = input_bboxes->Slice(i, i + 1);
ins_boxes.Resize({predict_dim, box_dim});
std::map<int, std::vector<int>> indices;
int num_nmsed_out = 0;
MultiClassNMS<float>(ins_score, ins_boxes, &indices, &num_nmsed_out,
background_label, nms_top_k, keep_top_k, nms_threshold,
nms_eta, score_threshold);
all_indices.push_back(indices);
batch_starts.push_back(batch_starts.back() + num_nmsed_out);
}
int num_kept = batch_starts.back();
if (num_kept == 0) {
float* od = outs->mutable_data<float>({1});
od[0] = -1;
} else {
outs->mutable_data<float>({num_kept, kOutputDim});
for (int64_t i = 0; i < batch_size; ++i) {
Tensor ins_score = input_scores->Slice(i, i + 1);
ins_score.Resize({class_num, predict_dim});
Tensor ins_boxes = input_bboxes->Slice(i, i + 1);
ins_boxes.Resize({predict_dim, box_dim});
int64_t s = batch_starts[i];
int64_t e = batch_starts[i + 1];
if (e > s) {
Tensor out = outs->Slice(s, e);
MultiClassOutput<float>(ins_score, ins_boxes, all_indices[i], &out);
}
}
}
// framework::LoD lod;
// lod.emplace_back(batch_starts);
//
// outs->set_lod(lod);
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/pool_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef POOL_OP
#pragma once
#include <string>
#include <vector>
#include "operators/math/pooling.h"
namespace paddle_mobile {
namespace operators {
using framework::Tensor;
inline void PoolBasic(std::string pooling_type, std::vector<int> ksize,
std::vector<int> strides, std::vector<int> paddings,
const Tensor *in_x, Tensor *out) {
if (pooling_type == "max") {
math::PoolFunctor<CPU, math::MaxPool<float>, float> pool2d_forward;
math::MaxPool<float> pool_process;
pool2d_forward(*in_x, ksize, strides, paddings, pool_process, out);
} else if (pooling_type == "avg") {
math::PoolFunctor<CPU, math::AvgPool<float>, float> pool2d_forward;
math::AvgPool<float> pool_process;
pool2d_forward(*in_x, ksize, strides, paddings, pool_process, out);
}
}
template <typename P>
void PoolCompute(const PoolParam &param) {
const Tensor *in_x = param.Input();
Tensor *out = param.Output();
std::string pooling_type = param.PoolingType();
std::vector<int> ksize = param.Ksize();
std::vector<int> strides = param.Strides();
std::vector<int> paddings = param.Paddings();
if (ksize.size() != 2) {
LOG(paddle_mobile::LogLevel::kLOG_ERROR)
<< "Pool op only supports 2D and 3D input.";
}
if (param.isGlobalPooling()) {
for (size_t i = 0; i < ksize.size(); ++i) {
paddings[i] = 0;
ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
}
} else if (ksize[0] == 3 && ksize[0] == ksize[1]) {
if (pooling_type == "max") {
if (strides[0] == strides[1] && strides[0] == 1 &&
paddings[0] == paddings[1] && paddings[1] == 1) {
math::Pool3x3Maxs1p1(in_x, out);
} else {
math::Pool3x3Max(strides, paddings, in_x, out);
}
} else if (pooling_type == "avg") {
if (strides[0] == strides[1] && strides[0] == 1 &&
paddings[0] == paddings[1] && paddings[1] == 1) {
math::Pool3x3Avgs1p1(in_x, out);
} else {
math::Pool3x3Avg(strides, paddings, in_x, out);
}
}
} else if (ksize[0] == 2 && ksize[0] == ksize[1]) {
#ifndef IOS
if (pooling_type == "max") {
math::Pool2x2Max(strides, paddings, in_x, out);
} else if (pooling_type == "avg") {
math::Pool2x2Avg(strides, paddings, in_x, out);
}
#else
PoolBasic(pooling_type, ksize, strides, paddings, in_x, out);
#endif
} else {
PoolBasic(pooling_type, ksize, strides, paddings, in_x, out);
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/prior_box_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef PRIORBOX_OP
#pragma once
#include <algorithm>
#include <vector>
namespace paddle_mobile {
namespace operators {
template <typename T>
struct ClipFunctor {
inline T operator()(T in) const {
return std::min<T>(std::max<T>(in, 0.), 1.);
}
};
template <typename P>
void PriorBoxCompute(const PriorBoxParam &param) {
const auto *input_ = param.Input();
const auto &input_dims = input_->dims();
const auto *input_image = param.InputImage();
const auto &input_image_dims = input_image->dims();
const auto &min_sizes = param.MinSizes();
const auto &max_sizes = param.MaxSizes();
const auto &variances = param.Variances();
const auto &input_aspect_ratio = param.AspectRatios();
const bool &flip = param.Flip();
const bool &clip = param.Clip();
const float &step_w = param.StepW();
const float &step_h = param.StepH();
const float &offset = param.Offset();
Tensor *output_boxes = param.OutputBoxes();
auto output_boxes_dataptr = output_boxes->mutable_data<float>();
Tensor *output_variances = param.OutputVariances();
auto output_variances_dataptr = output_variances->mutable_data<float>();
std::vector<float> aspect_ratios;
ExpandAspectRatios(input_aspect_ratio, flip, &aspect_ratios);
auto img_width = input_image_dims[3];
auto img_height = input_image_dims[2];
auto feature_width = input_dims[3];
auto feature_height = input_dims[2];
auto stride0 = output_boxes->dims()[1] * output_boxes->dims()[2] *
output_boxes->dims()[3];
auto stride1 = output_boxes->dims()[2] * output_boxes->dims()[3];
auto stride2 = output_boxes->dims()[3];
float step_width, step_height;
/// 300 / 19
if (step_w == 0 || step_h == 0) {
step_width = static_cast<float>(img_width) / feature_width;
step_height = static_cast<float>(img_height) / feature_height;
} else {
step_width = step_w;
step_height = step_h;
}
int num_priors = aspect_ratios.size() * min_sizes.size();
if (!max_sizes.empty()) {
num_priors += max_sizes.size();
}
for (int h = 0; h < feature_height; ++h) {
for (int w = 0; w < feature_width; ++w) {
/// map origin image
float center_x = (w + offset) * step_width;
float center_y = (h + offset) * step_height;
float box_width, box_height;
int idx = 0;
for (size_t s = 0; s < min_sizes.size(); ++s) {
auto min_size = min_sizes[s];
// priors with different aspect ratios
for (float ar : aspect_ratios) {
box_width = min_size * sqrt(ar) / 2.;
box_height = min_size / sqrt(ar) / 2.;
/// box_width/2 , / img_width 为了得到feature map 相对于
/// 原图的归一化位置的比例。
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 0] =
(center_x - box_width) / img_width;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 1] =
(center_y - box_height) / img_height;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 2] =
(center_x + box_width) / img_width;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 3] =
(center_y + box_height) / img_height;
idx++;
}
if (!max_sizes.empty()) {
auto max_size = max_sizes[s];
// square prior with size sqrt(minSize * maxSize)
box_width = box_height = sqrt(min_size * max_size) / 2.;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 0] =
(center_x - box_width) / img_width;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 1] =
(center_y - box_height) / img_height;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 2] =
(center_x + box_width) / img_width;
output_boxes_dataptr[h * stride0 + w * stride1 + idx * stride2 + 3] =
(center_y + box_height) / img_height;
idx++;
}
}
}
}
if (clip) {
math::Transform trans;
ClipFunctor<float> clip_func;
trans(output_boxes_dataptr, output_boxes_dataptr + output_boxes->numel(),
output_boxes_dataptr, clip_func);
}
if ((variances.size() != 4)) {
LOG(kLOG_ERROR) << " variances.size() must be 4.";
}
int64_t box_num = feature_height * feature_width * num_priors;
for (int i = 0; i < box_num; i++) {
output_variances_dataptr[4 * i] = variances[0];
output_variances_dataptr[4 * i + 1] = variances[1];
output_variances_dataptr[4 * i + 2] = variances[2];
output_variances_dataptr[4 * i + 3] = variances[3];
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/relu_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef RELU_OP
#pragma once
#include <operators/math/transform.h>
namespace paddle_mobile {
namespace operators {
template <typename T>
struct ReluFunctor {
inline T operator()(T in) const { return in > 0 ? in : 0; }
};
/*
* @b 特化到具体平台的实现, param 从 op 层传入
* */
template <typename P>
void ReluCompute(const ReluParam &param) {
const auto *input_x = param.InputX();
auto *input_x_ptr = input_x->data<float>();
auto *out = param.Out();
auto *out_ptr = out->mutable_data<float>();
int numel = input_x->numel();
// if (numel > 64) {
// asm volatile(
// "pld [%[input_x_ptr], #0] \n\t"
// "vmov.f32 q8, #0.0 \n\t"
// "subs %[num], %[num], #32 \n\t"
// "blt end_num_%= \n\t"
// "loop_num_%=: \n\t"
// "pld [%[input_x_ptr], #1024] \n\t"
//
// "vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
//
// "vmax.f32 q0, q0, q8 \n\t"
// "vmax.f32 q1, q1, q8 \n\t"
// "vmax.f32 q2, q2, q8 \n\t"
// "vmax.f32 q3, q3, q8 \n\t"
// "vmax.f32 q4, q4, q8 \n\t"
// "vmax.f32 q5, q5, q8 \n\t"
// "vmax.f32 q6, q6, q8 \n\t"
// "vmax.f32 q7, q7, q8 \n\t"
//
// "vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
// "vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
// "vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
// "vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
//
// "subs %[num], %[num], #32 \n\t"
// "bge loop_num_%= \n\t"
// "end_num_%=: \n\t"
// "cmp %[num], #0 \n\t"
// "bge end_%= \n\t"
// "mov r6, #4 \n\t"
// "mul r5, %[num], r6 \n\t"
// "add %[input_x_ptr], %[input_x_ptr], r5 \n\t"
// "vld1.32 {q0, q1}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q2, q3}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q4, q5}, [%[input_x_ptr]]! \n\t"
// "vld1.32 {q6, q7}, [%[input_x_ptr]]! \n\t"
// "vmax.f32 q0, q0, q8 \n\t"
// "vmax.f32 q1, q1, q8 \n\t"
// "vmax.f32 q2, q2, q8 \n\t"
// "vmax.f32 q3, q3, q8 \n\t"
// "vmax.f32 q4, q4, q8 \n\t"
// "vmax.f32 q5, q5, q8 \n\t"
// "vmax.f32 q6, q6, q8 \n\t"
// "vmax.f32 q7, q7, q8 \n\t"
// "add %[out_ptr], %[out_ptr], r5 \n\t"
// "vst1.32 {q0, q1}, [%[out_ptr]]! \n\t"
// "vst1.32 {q2, q3}, [%[out_ptr]]! \n\t"
// "vst1.32 {q4, q5}, [%[out_ptr]]! \n\t"
// "vst1.32 {q6, q7}, [%[out_ptr]]! \n\t"
// "end_%=: \n\t"
// :
// :
// [out_ptr] "r"(out_ptr), [input_x_ptr] "r"(input_x_ptr), [num]
// "r"(numel) : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6",
// "q7", "q8", "r5",
// "r6");
// } else {
ReluFunctor<float> func_;
math::Transform trans;
trans(input_x_ptr, input_x_ptr + numel, out_ptr, func_);
// }
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/reshape_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef RESHAPE_OP
#pragma once
#include <vector>
namespace paddle_mobile {
namespace operators {
template <typename P>
void ReshapeCompute(const ReshapeParam &param) {
const auto *input_x = param.InputX();
const auto &input_x_dims = input_x->dims();
auto *out = param.Out();
framework::DDim out_dims = out->dims();
const auto *input_shape = param.InputShape();
if (input_shape) {
auto *shape_data = input_shape->data<int>();
framework::Tensor cpu_shape_tensor;
auto shape =
std::vector<int>(shape_data, shape_data + input_shape->numel());
out_dims = ValidateShape(shape, input_x->dims());
}
bool inplace = param.Inplace();
out->Resize(out_dims);
if (!inplace) {
out->mutable_data<float>();
framework::TensorCopy(*input_x, out);
out->Resize(out_dims);
} else {
out->ShareDataWith(*input_x);
out->Resize(out_dims);
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/sigmoid_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef SIGMOID_OP
#pragma once
#include "operators/op_param.h"
#if __ARM_NEON
#include <arm_neon.h>
#include "operators/math/math_func_neon.h"
#endif
namespace paddle_mobile {
namespace operators {
using framework::DDim;
void sigmoid(const Tensor *X, Tensor *Y) {
#if __ARM_NEON
const float *input = X->data<float>();
float *output = Y->mutable_data<float>();
const DDim &dDim = X->dims();
int axis_index = 1;
if (dDim.size() < 4) {
axis_index = 0;
}
DDim outer_ddim =
paddle_mobile::framework::slice_ddim(dDim, 0, axis_index + 1);
DDim inner_ddim =
paddle_mobile::framework::slice_ddim(dDim, axis_index + 1, dDim.size());
int out_size = paddle_mobile::framework::product(outer_ddim);
int inner_size = paddle_mobile::framework::product(inner_ddim);
DLOG << "outsize=" << out_size;
DLOG << "innersize=" << inner_size;
#pragma omp parallel for
for (int i = 0; i < out_size; ++i) {
const float *input_outer_ptr = input + i * inner_size;
float *output_outer_ptr = output + i * inner_size;
int nn = inner_size >> 2;
int remain = inner_size - (nn << 2);
float32x4_t _one = vdupq_n_f32(1.f);
for (; nn > 0; nn--) {
float32x4_t data = vld1q_f32(input_outer_ptr);
data = vnegq_f32(data);
data = exp_ps(data);
data = vaddq_f32(data, _one);
float32x4_t out_data = vrecpeq_f32(data);
out_data = vmulq_f32(vrecpsq_f32(data, out_data), out_data);
vst1q_f32(output_outer_ptr, out_data);
input_outer_ptr += 4;
output_outer_ptr += 4;
}
for (; remain > 0; remain--) {
*output_outer_ptr = 1.f / (1.f + exp(-*input_outer_ptr));
output_outer_ptr++;
input_outer_ptr++;
}
}
#endif
}
template <typename P>
void SigmoidCompute(const SigmoidParam &param) {
const Tensor *in_x = param.InputX();
Tensor *out = param.Out();
auto x_dims = in_x->dims();
out->Resize(x_dims);
sigmoid(in_x, out);
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/common/openmp-fix.cpp → src/operators/kernel/central-arm-func/softmax_arm_func.h
浏览文件 @ c67b701f
......@@ -12,16 +12,19 @@ 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. */
#ifdef PADDLE_MOBILE_USE_OPENMP
/**
* android-ndk-r17 has a problem when linking with openmp.
* if paddle-mobile enables -fopenmp, but didn't use those omp_* functions,
* after linking another binary with libpaddle-mobile.so, the omp_get_thread_num
* will not work. see test/common/test_openmp.cc the detailed reason is still
* unclear, but this trick will work. a better solution is hacking the linker,
* try some flags to make it link omp_* functions, but I didn't find out how to
* make it work.
*/
#include <omp.h>
static int _ = omp_get_num_procs();
#ifdef SOFTMAX_OP
#pragma once
#include "../../math/softmax.h"
namespace paddle_mobile {
namespace operators {
template <typename P>
void SoftmaxCompute(const SoftmaxParam &param) {
const Tensor *in_x = param.InputX();
Tensor *out = param.Out();
auto x_dims = in_x->dims();
out->Resize(x_dims);
math::SoftmaxFuntor<CPU, float>()(in_x, out);
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/central-arm-func/transpose_arm_func.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef TRANSPOSE_OP
#pragma once
#include <vector>
namespace paddle_mobile {
namespace operators {
// vector<int> pos;
// template <typename T>
// void TransposeFunc(const int numel, const T* input, const vector<int> axis,
// const vector<int> old_strides, const vector<int>
// new_strides, T* output) {
// for (int i = 0; i < numel; ++i) {
// int old_idx = 0;
// int idx = i;
// for (int j = 0; j < axis.size(); ++j) {
// int order = axis[j];
// old_idx += (idx / new_strides[j]) * old_strides[order];
// idx %= new_strides[j];
// }
// output[i] = input[old_idx];
// }
// }
template <typename P>
void TransposeCompute(const TransposeParam& param) {
const auto* input_x = param.InputX();
const auto input_x_dims = input_x->dims();
auto* out = param.Out();
const auto axis = param.Axis();
const auto* input_x_data = input_x->data<float>();
auto* out_data = out->mutable_data<float>();
size_t ndim = axis.size();
std::vector<int> xdim(ndim);
std::vector<int> xstride(ndim);
std::vector<int> xout(ndim);
for (int i = 0; i < ndim; i++) {
int j = ndim - 1 - i;
xdim[j] = input_x_dims[axis[i]];
xstride[j] = 1;
for (int k = axis[i] + 1; k < ndim; k++) {
xstride[j] *= input_x_dims[k];
}
xout[j] = xstride[j] * xdim[j];
}
auto numel = input_x->numel();
size_t pind = 0;
std::vector<int> ind(ndim);
for (int i = 0; i < numel; i++) {
out_data[i] = input_x_data[pind];
ind[0]++;
pind += xstride[0];
for (int j = 0; j < ndim - 1; j++) {
if (ind[j] == xdim[j]) {
ind[j + 1]++;
ind[j] = 0;
pind += xstride[j + 1];
pind -= xout[j];
} else {
break;
}
}
}
}
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/concat_kernel.h
浏览文件 @ c67b701f
......@@ -27,7 +27,7 @@ template <typename DeviceType, typename T>
class ConcatKernel : public framework::OpKernelBase<DeviceType, ConcatParam> {
public:
void Compute(const ConcatParam &param) const;
bool Init(const ConcatParam &para) const;
bool Init(ConcatParam *param);
};
} // namespace operators
......
src/operators/kernel/conv_add_bn_relu_kernel.h 0 → 100644
浏览文件 @ c67b701f
/* 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
#ifdef FUSION_CONVADDBNRELU_OP
#include <vector>
#include "framework/ddim.h"
#include "framework/operator.h"
#include "operators/math/conv_func.h"
#include "operators/math/im2col.h"
#include "operators/math/math_function.h"
#include "operators/math/vol2col.h"
#include "operators/op_param.h"
namespace paddle_mobile {
namespace operators {
using framework::DDim;
using framework::OpKernelBase;
template <typename DeviceType, typename T>
class ConvAddBNReluKernel
: public OpKernelBase<DeviceType, FusionConvAddBNReluParam> {
public:
void Compute(const FusionConvAddBNReluParam &param) const;
bool Init(FusionConvAddBNReluParam *param);
};
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/conv_add_kernel.h
浏览文件 @ c67b701f
......@@ -20,9 +20,11 @@ limitations under the License. */
#if __ARM_NEON
#include <arm_neon.h>
#endif
#include "common/common.h"
#include "framework/ddim.h"
#include "framework/operator.h"
#include "operators/math/conv_func.h"
#include "operators/math/depthwise_conv_3x3.h"
#include "operators/math/im2col.h"
#include "operators/math/math_function.h"
#include "operators/math/vol2col.h"
......@@ -38,7 +40,7 @@ template <typename DeviceType, typename T>
class ConvAddKernel : public OpKernelBase<DeviceType, FusionConvAddParam> {
public:
void Compute(const FusionConvAddParam &param) const;
bool Init(const FusionConvAddParam &para) const;
bool Init(FusionConvAddParam *param);
};
} // namespace operators
......
src/operators/kernel/conv_add_relu_kernel.h
浏览文件 @ c67b701f
......@@ -36,7 +36,7 @@ class ConvAddReluKernel
: public OpKernelBase<DeviceType, FusionConvAddReluParam> {
public:
void Compute(const FusionConvAddReluParam &param) const;
bool Init(const FusionConvAddReluParam &para) const;
bool Init(FusionConvAddReluParam *param);
};
} // namespace operators
......
src/operators/kernel/conv_kernel.h
浏览文件 @ c67b701f
......@@ -32,7 +32,7 @@ template <typename DeviceType, typename T>
class ConvKernel : public OpKernelBase<DeviceType, ConvParam> {
public:
void Compute(const ConvParam &param) const;
bool Init(const ConvParam &para) const;
bool Init(ConvParam *param);
};
} // namespace operators
......
src/operators/kernel/depthwise_conv_kernel.h
浏览文件 @ c67b701f
......@@ -31,7 +31,7 @@ template <typename DeviceType, typename T>
class DepthwiseConvKernel : public OpKernelBase<DeviceType, ConvParam> {
public:
void Compute(const ConvParam &param) const;
bool Init(const ConvParam &para) const;
bool Init(ConvParam *param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/dropout_kernel.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef DROPOUT_OP
#include "framework/operator.h"
#include "operators/op_param.h"
#pragma once;
namespace paddle_mobile {
namespace operators {
template <typename DeviceType, typename T>
class DropoutKernel : public framework::OpKernelBase<DeviceType, DropoutParam> {
public:
void Compute(const DropoutParam& param) const;
bool Init(DropoutParam* para);
};
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/elementwise_add_kernel.h
浏览文件 @ c67b701f
......@@ -30,7 +30,7 @@ class ElementwiseAddKernel
: public framework::OpKernelBase<DeviceType, ElementwiseAddParam> {
public:
void Compute(const ElementwiseAddParam &param) const;
bool Init(const ElementwiseAddParam &para) const;
bool Init(ElementwiseAddParam *param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/fpga/conv_kernel.cpp
浏览文件 @ c67b701f
......@@ -20,7 +20,7 @@ namespace paddle_mobile {
namespace operators {
template <>
bool ConvKernel<FPGA, float>::Init(const ConvParam &para) const {
bool ConvKernel<FPGA, float>::Init(ConvParam *param) {
return true;
}
......
src/operators/kernel/fusion_fc_kernel.h
浏览文件 @ c67b701f
......@@ -28,7 +28,7 @@ class FusionFcKernel
: public framework::OpKernelBase<DeviceType, FusionFcParam> {
public:
void Compute(const FusionFcParam& param) const;
bool Init(const FusionFcParam& para) const;
bool Init(FusionFcParam* param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/im2sequence_kernel.h 0 → 100644
浏览文件 @ c67b701f
/* 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. */
#ifdef IM2SEQUENCE_OP
#include "framework/operator.h"
#include "operators/math/im2col.h"
#include "operators/math/math_function.h"
#include "operators/math/vol2col.h"
#include "operators/op_param.h"
#pragma once;
namespace paddle_mobile {
namespace operators {
using namespace framework;
template <typename DeviceType, typename T>
class Im2SequenceKernel
: public framework::OpKernelBase<DeviceType, Im2SequenceParam> {
public:
void Compute(const Im2SequenceParam& param) const;
bool Init(Im2SequenceParam* para);
};
} // namespace operators
} // namespace paddle_mobile
#endif
src/operators/kernel/lrn_kernel.h
浏览文件 @ c67b701f
......@@ -14,10 +14,11 @@ limitations under the License. */
#ifdef LRN_OP
#pragma once
#include "framework/operator.h"
#include "operators/op_param.h"
#include <cmath>
#ifdef __ARM_NEON
#include "arm_neon.h"
#include "operators/math/math_func_neon.h"
......@@ -169,7 +170,7 @@ template <typename DeviceType, typename T>
class LrnKernel : public framework::OpKernelBase<DeviceType, LrnParam> {
public:
void Compute(const LrnParam &param) const;
bool Init(const LrnParam &para) const;
bool Init(LrnParam *param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/mali/batchnorm_kernel.cpp
浏览文件 @ c67b701f
......@@ -129,7 +129,7 @@ class AclBatchNormOp : public acl::ACLOperator {
};
template <>
bool BatchNormKernel<GPU_MALI, float>::Init(const BatchNormParam& param) const {
bool BatchNormKernel<GPU_MALI, float>::Init(BatchNormParam* param) {
AclBatchNormOp<GPU_MALI, float>* acl_op =
reinterpret_cast<AclBatchNormOp<GPU_MALI, float>*>(this->GetAclOp());
if (acl_op == nullptr) {
......
src/operators/kernel/mali/conv_kernel.cpp
浏览文件 @ c67b701f
......@@ -196,7 +196,7 @@ class AclConvOp : public acl::ACLOperator {
};
template <>
bool ConvKernel<GPU_MALI, float>::Init(const ConvParam& param) const {
bool ConvKernel<GPU_MALI, float>::Init(ConvParam* param) {
AclConvOp<GPU_MALI, float>* acl_op =
reinterpret_cast<AclConvOp<GPU_MALI, float>*>(this->GetAclOp());
if (acl_op == nullptr) {
......
src/operators/kernel/mali/elementwise_add_kernel.cpp
浏览文件 @ c67b701f
......@@ -27,8 +27,7 @@ struct AddFunctor {
};
template <>
bool ElementwiseAddKernel<GPU_MALI, float>::Init(
const ElementwiseAddParam &para) const {
bool ElementwiseAddKernel<GPU_MALI, float>::Init(ElementwiseAddParam *param) {
return true;
}
......
src/operators/kernel/mali/fushion_fc_kernel.cpp
浏览文件 @ c67b701f
......@@ -22,7 +22,7 @@ namespace paddle_mobile {
namespace operators {
template <>
bool FusionFcKernel<GPU_MALI, float>::Init(const FusionFcParam &para) const {
bool FusionFcKernel<GPU_MALI, float>::Init(FusionFcParam *param) {
return true;
}
......
src/operators/kernel/mali/mul_kernel.cpp
浏览文件 @ c67b701f
......@@ -22,7 +22,7 @@ namespace paddle_mobile {
namespace operators {
template <>
bool MulKernel<GPU_MALI, float>::Init(const MulParam &para) const {
bool MulKernel<GPU_MALI, float>::Init(MulParam *param) {
return true;
}
......
src/operators/kernel/mali/reshape_kernel.cpp
浏览文件 @ c67b701f
......@@ -22,7 +22,7 @@ namespace paddle_mobile {
namespace operators {
template <>
bool ReshapeKernel<GPU_MALI, float>::Init(const ReshapeParam &para) const {
bool ReshapeKernel<GPU_MALI, float>::Init(ReshapeParam *param) {
return true;
}
......
src/operators/kernel/mul_kernel.h
浏览文件 @ c67b701f
......@@ -29,7 +29,7 @@ template <typename DeviceType, typename T>
class MulKernel : public framework::OpKernelBase<DeviceType, MulParam> {
public:
void Compute(const MulParam &param) const;
bool Init(const MulParam &para) const;
bool Init(MulParam *param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/multiclass_nms_kernel.h
浏览文件 @ c67b701f
......@@ -28,7 +28,7 @@ class MultiClassNMSKernel
: public framework::OpKernelBase<DeviceType, MultiClassNMSParam> {
public:
void Compute(const MultiClassNMSParam& param) const;
bool Init(const MultiClassNMSParam& para) const;
bool Init(MultiClassNMSParam* param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/pool_kernel.h
浏览文件 @ c67b701f
......@@ -17,7 +17,6 @@ limitations under the License. */
#pragma once
#include "framework/operator.h"
#include "operators/math/pooling.h"
#include "operators/op_param.h"
namespace paddle_mobile {
......@@ -28,7 +27,7 @@ template <typename DeviceType, typename T>
class PoolKernel : public OpKernelBase<DeviceType, PoolParam> {
public:
void Compute(const PoolParam &param) const override;
bool Init(const PoolParam &para) const;
bool Init(PoolParam *param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/prior_box_kernel.h
浏览文件 @ c67b701f
......@@ -16,8 +16,9 @@ limitations under the License. */
#pragma once
#include <algorithm>
#include <cmath>
#include <vector>
#include "framework/operator.h"
#include "operators/math/transform.h"
#include "operators/op_param.h"
......@@ -54,7 +55,7 @@ class PriorBoxKernel
: public framework::OpKernelBase<DeviceType, PriorBoxParam> {
public:
void Compute(const PriorBoxParam& param) const;
bool Init(const PriorBoxParam& para) const;
bool Init(PriorBoxParam* param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/relu_kernel.h
浏览文件 @ c67b701f
......@@ -27,7 +27,7 @@ template <typename DeviceType, typename T>
class ReluKernel : public framework::OpKernelBase<DeviceType, ReluParam> {
public:
void Compute(const ReluParam& param) const;
bool Init(const ReluParam& para) const;
bool Init(ReluParam* param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/reshape_kernel.h
浏览文件 @ c67b701f
......@@ -71,7 +71,7 @@ template <typename DeviceType, typename T>
class ReshapeKernel : public framework::OpKernelBase<DeviceType, ReshapeParam> {
public:
void Compute(const ReshapeParam& param) const;
bool Init(const ReshapeParam& para) const;
bool Init(ReshapeParam* param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/sigmoid_kernel.h
浏览文件 @ c67b701f
......@@ -26,7 +26,7 @@ template <typename DeviceType, typename T>
class SigmoidKernel : public OpKernelBase<DeviceType, SigmoidParam> {
public:
void Compute(const SigmoidParam& param) const override;
bool Init(const SigmoidParam& para) const;
bool Init(SigmoidParam* param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/softmax_kernel.h
浏览文件 @ c67b701f
......@@ -23,13 +23,11 @@ namespace paddle_mobile {
namespace operators {
using framework::OpKernelBase;
void simoid(Tensor *X, Tensor *Y);
template <typename DeviceType, typename T>
class SoftmaxKernel : public OpKernelBase<DeviceType, SoftmaxParam> {
public:
void Compute(const SoftmaxParam &param) const override;
bool Init(const SoftmaxParam &para) const;
bool Init(SoftmaxParam *param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/kernel/transpose_kernel.h
浏览文件 @ c67b701f
......@@ -29,7 +29,7 @@ class TransposeKernel
: public framework::OpKernelBase<DeviceType, TransposeParam> {
public:
void Compute(const TransposeParam& param) const;
bool Init(const TransposeParam& para) const;
bool Init(TransposeParam* param);
};
} // namespace operators
} // namespace paddle_mobile
......
src/operators/lrn_op.cpp
浏览文件 @ c67b701f
......@@ -30,11 +30,9 @@ template class LrnOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(lrn);
REGISTER_OPERATOR_CPU(lrn, ops::LrnOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(lrn);
REGISTER_OPERATOR_MALI_GPU(lrn, ops::LrnOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/lrn_op.h
浏览文件 @ c67b701f
......@@ -46,4 +46,13 @@ class LrnOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(lrn);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(lrn);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/math/depthwise_conv_3x3.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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 "operators/math/depthwise_conv_3x3.h"
#include <arm_neon.h>
#include <vector>
namespace paddle_mobile {
namespace operators {
namespace math {
void DepthwiseConv3x3(const Tensor *input, vector<int> strides,
vector<int> paddings, const Tensor *filter, Tensor *bias,
Tensor *output, bool if_bias) {
#if __ARM_NEON
const int batch_size = input->dims()[0];
const int input_height = input->dims()[2];
const int input_width = input->dims()[3];
const int output_channels = output->dims()[1];
const int output_height = output->dims()[2];
const int output_width = output->dims()[3];
const int _kernel_size = 3;
const int stride_height = strides[0];
const int stride_width = strides[1];
const int padding_height = paddings[0];
const int padding_width = paddings[1];
const float zero = 0;
const int input_channel_stride = input_height * input_width;
const int output_channel_stride = output_height * output_width;
const int filter_channel_stride = 9;
const float *input_data = input->data<float>();
const float *filter_data = filter->data<float>();
if (if_bias) {
math::expand_bias(*bias, 1, output->dims());
output->ShareDataWith(*bias);
}
float *output_data = output->mutable_data<float>();
const int input_batch_stride = output_channels * input_channel_stride;
const int output_batch_stride = output_channels * output_channel_stride;
const int filter_batch_stride = output_channels * output_channel_stride;
const float *pos1, *pos2, *pos3, *filter1, *filter2, *filter3, *output_ptr;
int hstart, wstart, hend, wend;
float result;
for (int i = 0; i < batch_size; ++i) {
for (int c = 0; c < output_channels; ++c) {
filter1 = filter_data;
filter2 = filter1 + 3;
filter3 = filter2 + 3;
for (int ph = 0; ph < output_height; ph++) {
for (int pw = 0; pw < output_width; pw++) {
hstart = ph * stride_height - padding_height;
wstart = pw * stride_width - padding_width;
hend = min(hstart + _kernel_size, input_height + padding_height);
wend = min(wstart + _kernel_size, input_width + padding_width);
hstart = max(hstart, 0);
wstart = max(wstart, 0);
hend = min(hend, input_height);
wend = min(wend, input_width);
pos1 = input_data + hstart * input_width + wstart;
pos2 = input_data + (hstart + 1) * input_width + wstart;
pos3 = input_data + (hstart + 2) * input_width + wstart;
output_ptr = output_data + ph * output_width + pw;
if (hend - hstart != 3 || wend - wstart != 3) {
result = 0;
float fake_input[9] = {0};
if (hstart == 0 && wstart == 0) {
// 左上角
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 3; ++k) {
if (j >= 3 - hend && k >= 3 - wend) {
fake_input[3 * j + k] =
input_data[(j - (3 - hend)) * input_width + k -
(3 - wend)];
}
}
}
} else if (hstart == 0 && wend == input_width) {
// 右上角
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 3; ++k) {
if (j >= 3 - hend && k <= input_width - wstart - 1) {
fake_input[3 * j + k] =
input_data[(j - (3 - hend)) * input_width + k + wstart];
}
}
}
} else if (hend == input_height && wstart == 0) {
// 左下角
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 3; ++k) {
if (j <= input_height - 1 - hstart && k >= 3 - wend) {
fake_input[3 * j + k] =
input_data[(j + hstart) * input_width + k - (3 - wend)];
}
}
}
} else if (hend == input_height && wend == input_width) {
// 右下角
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 3; ++k) {
if (j <= input_height - hstart - 1 &&
k <= input_width - wstart - 1) {
fake_input[3 * j + k] =
input_data[(j + hstart) * input_width + k + wstart];
}
}
}
} else if (hstart == 0) {
// 顶部
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 3; ++k) {
if (j >= 3 - hend) {
fake_input[3 * j + k] =
input_data[(j - (3 - hend)) * input_width + k + wstart];
}
}
}
} else if (hend == input_height) {
// 底部
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 3; ++k) {
if (j <= input_height - hstart - 1) {
fake_input[3 * j + k] =
input_data[(j + hstart) * input_width + k + wstart];
}
}
}
} else if (wstart == 0) {
// 左侧
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 3; ++k) {
if (k >= 3 - wend) {
fake_input[3 * j + k] =
input_data[(j + hstart) * input_width +
(k - (3 - wend))];
}
}
}
} else if (wend == input_width) {
// 右侧
for (int j = 0; j < 3; ++j) {
for (int k = 0; k < 3; ++k) {
if (k <= input_width - wstart - 1) {
fake_input[3 * j + k] =
input_data[(j + hstart) * input_width + k + wstart];
}
}
}
}
for (int l = 0; l < 9; ++l) {
result += fake_input[l] * filter1[l];
}
if (if_bias) {
output_data[ph * output_width + pw] += result;
} else {
output_data[ph * output_width + pw] = result;
}
} else {
#if defined(ARMV17)
asm volatile(
"vld1.32 {q1}, [%[pos1]] \n\t"
"vld1.32 {q4}, [%[filter1]] \n\t"
"vmov.f32 q0, #0.0 \n\t"
"vld1.32 {q2}, [%[pos2]] \n\t"
"vld1.32 {q5}, [%[filter2]] \n\t"
"vmla.f32 q0, q1, q4 \n\t"
"vld1.32 {q3}, [%[pos3]] \n\t"
"vld1.32 {q6}, [%[filter3]] \n\t"
"vmla.f32 q0, q2, q5 \n\t"
"vmla.f32 q0, q3, q6 \n\t"
"vmov.f32 d1[1], %[zero] \n\t"
"vadd.f32 d4, d0, d1 \n\t"
"vadd.f32 s10, s8, s9 \n\t"
"vst1.32 {d5[0]},[%[output_ptr]] \n\t"
:
: [input_data] "r"(input_data), [pos1] "r"(pos1),
[pos2] "r"(pos2), [pos3] "r"(pos3), [filter1] "r"(filter1),
[filter2] "r"(filter2), [filter3] "r"(filter3),
[output_ptr] "r"(output_ptr), [zero] "r"(zero)
: "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6");
#else
const float32x4_t data1 = vld1q_f32(pos1);
const float32x4_t data2 = vld1q_f32(pos2);
const float32x4_t data3 = vld1q_f32(pos3);
const float32x4_t v_filter1 = vld1q_f32(filter1);
const float32x4_t v_filter2 = vld1q_f32(filter2);
const float32x4_t v_filter3 = vld1q_f32(filter3);
float32x4_t mula = vmulq_f32(data1, v_filter1);
mula = vmlaq_f32(mula, data2, v_filter2);
mula = vmlaq_f32(mula, data3, v_filter3);
float32x2_t res = vpadd_f32(
vget_high_f32(vsetq_lane_f32(0, mula, 3)), vget_low_f32(mula));
res = vpadd_f32(res, res);
if (if_bias) {
output_data[ph * output_width + pw] += vget_lane_f32(res, 0);
} else {
output_data[ph * output_width + pw] = vget_lane_f32(res, 0);
}
#endif
}
}
}
input_data += input_channel_stride;
output_data += output_channel_stride;
filter_data += filter_channel_stride;
}
input_data += input_batch_stride;
output_data += output_batch_stride;
}
#endif
}
void DepthwiseConv3x3s1p1(const Tensor *input, const Tensor *filter,
Tensor *output, Tensor *bias, bool if_bias) {
const float *input_data = input->data<float>();
const float *filter_data = filter->data<float>();
float *output_data = output->data<float>();
const float *bias_data;
if (if_bias) {
bias_data = bias->data<float>();
}
const int h = static_cast<int>(input->dims()[2]);
const int w = static_cast<int>(input->dims()[3]);
const int l = h;
const int batch_size = static_cast<int>(input->dims()[0]);
const int c = static_cast<int>(input->dims()[1]);
const int hxw = h * w;
float32x4_t vbias = vdupq_n_f32(0.0);
for (int b = 0; b < batch_size; ++b) {
const float *filter_data_tmp = filter_data;
for (int j = 0; j < c; ++j) {
if (if_bias) {
vbias = vdupq_n_f32(bias_data[j]);
}
int l_mid = l - 2; // l=1->l_mid=-1,l=2->l_mid=0
float w00 = filter_data_tmp[0];
float w01 = filter_data_tmp[1];
float w02 = filter_data_tmp[2];
float w10 = filter_data_tmp[3];
float w11 = filter_data_tmp[4];
float w12 = filter_data_tmp[5];
float w20 = filter_data_tmp[6];
float w21 = filter_data_tmp[7];
float w22 = filter_data_tmp[8];
output_data[0] = w11 * input_data[0] + w12 * input_data[1] +
w21 * input_data[l] + w22 * input_data[l + 1];
output_data[l - 1] = w10 * input_data[l - 2] + w11 * input_data[l - 1] +
w20 * input_data[2 * l - 2] +
w21 * input_data[2 * l - 1];
output_data[(l - 1) * l] =
w01 * input_data[(l - 2) * l] + w02 * input_data[(l - 2) * l + 1] +
w11 * input_data[(l - 1) * l] + w12 * input_data[(l - 1) * l + 1];
output_data[l * l - 1] = w00 * input_data[(l - 2) * (l + 1)] +
w01 * input_data[(l - 2) * (l + 1) + 1] +
w10 * input_data[l * l - 2] +
w11 * input_data[l * l - 1];
if (if_bias) {
output_data[0] += bias_data[j];
output_data[l - 1] += bias_data[j];
output_data[(l - 1) * l] += bias_data[j];
output_data[l * l - 1] += bias_data[j];
}
for (int i = 1; i < l - 1; ++i) {
output_data[i * l] =
w01 * input_data[i * l - l] + w02 * input_data[i * l - l + 1] +
w11 * input_data[i * l] + w12 * input_data[i * l + 1] +
w21 * input_data[i * l + l] + w22 * input_data[i * l + l + 1];
output_data[i * l + l - 1] = w00 * input_data[i * l + l - 1 - l - 1] +
w01 * input_data[i * l + l - 1 - l] +
w10 * input_data[i * l + l - 1 - 1] +
w11 * input_data[i * l + l - 1] +
w20 * input_data[i * l + l - 1 + l - 1] +
w21 * input_data[i * l + l - 1 + l];
if (if_bias) {
output_data[i * l] += bias_data[j];
output_data[i * l + l - 1] += bias_data[j];
}
}
// top 1 row and bottom 1 row
const float *input_tmp = input_data;
float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
tmp3, tmp4, tmp5, out0;
in0 = vld1q_f32(input_tmp);
in2 = vld1q_f32(input_tmp + l);
const float *input_tmp_end = input_tmp + (l - 2) * l;
in4 = vld1q_f32(input_tmp_end);
in6 = vld1q_f32(input_tmp_end + l);
int c_mid = l_mid;
auto output_ptr = output_data + 1;
for (; c_mid > 3; c_mid -= 4) {
in1 = vld1q_f32(input_tmp + 4);
in3 = vld1q_f32(input_tmp + l + 4);
tmp0 = vextq_f32(in0, in1, 1);
tmp1 = vextq_f32(in0, in1, 2);
tmp2 = vextq_f32(in2, in3, 1);
tmp3 = vextq_f32(in2, in3, 2);
out0 = vmulq_n_f32(in0, w10);
out0 = vmlaq_n_f32(out0, tmp0, w11);
out0 = vmlaq_n_f32(out0, tmp1, w12);
out0 = vmlaq_n_f32(out0, in2, w20);
out0 = vmlaq_n_f32(out0, tmp2, w21);
out0 = vmlaq_n_f32(out0, tmp3, w22);
out0 = vaddq_f32(out0, vbias);
vst1q_f32(output_ptr, out0);
in5 = vld1q_f32(input_tmp_end + 4);
in7 = vld1q_f32(input_tmp_end + l + 4);
tmp0 = vextq_f32(in4, in5, 1);
tmp1 = vextq_f32(in4, in5, 2);
tmp2 = vextq_f32(in6, in7, 1);
tmp3 = vextq_f32(in6, in7, 2);
out0 = vmulq_n_f32(in4, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in6, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vaddq_f32(out0, vbias);
vst1q_f32(output_ptr + (l - 1) * l, out0);
// can optimize to each 8 stride.
input_tmp += 4;
input_tmp_end += 4;
output_ptr += 4;
in0 = in1;
in2 = in3;
in4 = in5;
in6 = in7;
}
// top right pad
float32x4_t pad0 = vdupq_n_f32(input_data[l - 1]);
float32x4_t pad1 = vdupq_n_f32(input_data[2 * l - 1]);
tmp0 = vextq_f32(in0, pad0, 1);
tmp1 = vextq_f32(in0, pad0, 2);
tmp2 = vextq_f32(in2, pad1, 1);
tmp3 = vextq_f32(in2, pad1, 2);
out0 = vmulq_n_f32(in0, w10);
out0 = vmlaq_n_f32(out0, tmp0, w11);
out0 = vmlaq_n_f32(out0, tmp1, w12);
out0 = vmlaq_n_f32(out0, in2, w20);
out0 = vmlaq_n_f32(out0, tmp2, w21);
out0 = vmlaq_n_f32(out0, tmp3, w22);
out0 = vaddq_f32(out0, vbias);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + i, out0, 2);
}
}
// bottom right pad
float32x4_t pad2 = vdupq_n_f32(input_data[l * l - 1 - l]);
float32x4_t pad3 = vdupq_n_f32(input_data[l * l - 1]);
tmp0 = vextq_f32(in4, pad2, 1);
tmp1 = vextq_f32(in4, pad2, 2);
tmp2 = vextq_f32(in6, pad3, 1);
tmp3 = vextq_f32(in6, pad3, 2);
out0 = vmulq_n_f32(in4, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in6, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vaddq_f32(out0, vbias);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 2);
}
}
// mid
for (int i = 0; i < l - 2; ++i) {
auto output_ptr = output_data + (i + 1) * l + 1;
input_tmp = input_data + i * l;
auto in0_tmp = vld1q_f32(input_tmp);
auto in2_tmp = vld1q_f32(input_tmp + l);
auto in4_tmp = vld1q_f32(input_tmp + l + l);
c_mid = l_mid;
for (; c_mid > 3; c_mid -= 4) {
auto in1_tmp = vld1q_f32(input_tmp + 4);
auto in3_tmp = vld1q_f32(input_tmp + l + 4);
auto in5_tmp = vld1q_f32(input_tmp + l + l + 4);
tmp0 = vextq_f32(in0_tmp, in1_tmp, 1);
tmp1 = vextq_f32(in0_tmp, in1_tmp, 2);
tmp2 = vextq_f32(in2_tmp, in3_tmp, 1);
tmp3 = vextq_f32(in2_tmp, in3_tmp, 2);
tmp4 = vextq_f32(in4_tmp, in5_tmp, 1);
tmp5 = vextq_f32(in4_tmp, in5_tmp, 2);
out0 = vmulq_n_f32(in0_tmp, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in2_tmp, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vmlaq_n_f32(out0, in4_tmp, w20);
out0 = vmlaq_n_f32(out0, tmp4, w21);
out0 = vmlaq_n_f32(out0, tmp5, w22);
out0 = vaddq_f32(out0, vbias);
vst1q_f32(output_ptr, out0);
output_ptr += 4;
input_tmp += 4;
in0_tmp = in1_tmp;
in2_tmp = in3_tmp;
in4_tmp = in5_tmp;
}
float32x4_t pad0 = vdupq_n_f32(input_data[i * l + l - 1]);
float32x4_t pad1 = vdupq_n_f32(input_data[i * l + l - 1 + l]);
float32x4_t pad2 = vdupq_n_f32(input_data[i * l + l - 1 + l + l]);
tmp0 = vextq_f32(in0_tmp, pad0, 1);
tmp1 = vextq_f32(in0_tmp, pad0, 2);
tmp2 = vextq_f32(in2_tmp, pad1, 1);
tmp3 = vextq_f32(in2_tmp, pad1, 2);
tmp4 = vextq_f32(in4_tmp, pad2, 1);
tmp5 = vextq_f32(in4_tmp, pad2, 2);
out0 = vmulq_n_f32(in0_tmp, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in2_tmp, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vmlaq_n_f32(out0, in4_tmp, w20);
out0 = vmlaq_n_f32(out0, tmp4, w21);
out0 = vmlaq_n_f32(out0, tmp5, w22);
out0 = vaddq_f32(out0, vbias);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + i, out0, 2);
}
}
}
output_data += hxw;
input_data += hxw;
filter_data_tmp += 9;
}
}
}
void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, const Tensor *filter,
Tensor *output, Tensor *bias, bool if_bias,
const Tensor *new_scale,
const Tensor *new_bias, bool if_bn,
bool if_relu) {
const float *input_data = input->data<float>();
const float *filter_data = filter->data<float>();
float *output_data = output->data<float>();
const float *bias_data = bias->data<float>();
const float *newscale_data = new_scale->data<float>();
const float *newbias_data = new_bias->data<float>();
const int h = static_cast<int>(input->dims()[2]);
const int w = static_cast<int>(input->dims()[3]);
const int l = h;
const int batch_size = static_cast<int>(input->dims()[0]);
const int c = static_cast<int>(input->dims()[1]);
const int hxw = h * w;
float32x4_t vbias = vdupq_n_f32(0.0);
float32x4_t vnewbias = vdupq_n_f32(0.0);
float32x4_t vnewscale = vdupq_n_f32(1.0);
float32x4_t vzero = vdupq_n_f32(0);
for (int b = 0; b < batch_size; ++b) {
const float *filter_data_tmp = filter_data;
for (int j = 0; j < c; ++j) {
if (if_bias) {
vbias = vdupq_n_f32(bias_data[j]);
}
if (if_bn) {
vnewbias = vdupq_n_f32(newbias_data[j]);
vnewscale = vdupq_n_f32(newscale_data[j]);
}
int l_mid = l - 2; // l=1->l_mid=-1,l=2->l_mid=0
float w00 = filter_data_tmp[0];
float w01 = filter_data_tmp[1];
float w02 = filter_data_tmp[2];
float w10 = filter_data_tmp[3];
float w11 = filter_data_tmp[4];
float w12 = filter_data_tmp[5];
float w20 = filter_data_tmp[6];
float w21 = filter_data_tmp[7];
float w22 = filter_data_tmp[8];
output_data[0] = w11 * input_data[0] + w12 * input_data[1] +
w21 * input_data[l] + w22 * input_data[l + 1];
output_data[l - 1] = w10 * input_data[l - 2] + w11 * input_data[l - 1] +
w20 * input_data[2 * l - 2] +
w21 * input_data[2 * l - 1];
output_data[(l - 1) * l] =
w01 * input_data[(l - 2) * l] + w02 * input_data[(l - 2) * l + 1] +
w11 * input_data[(l - 1) * l] + w12 * input_data[(l - 1) * l + 1];
output_data[l * l - 1] = w00 * input_data[(l - 2) * (l + 1)] +
w01 * input_data[(l - 2) * (l + 1) + 1] +
w10 * input_data[l * l - 2] +
w11 * input_data[l * l - 1];
if (if_bias) {
output_data[0] += bias_data[j];
output_data[l - 1] += bias_data[j];
output_data[(l - 1) * l] += bias_data[j];
output_data[l * l - 1] += bias_data[j];
}
if (if_bn) {
output_data[0] = output_data[0] * newscale_data[j] + newbias_data[j];
output_data[l - 1] =
output_data[l - 1] * newscale_data[j] + newbias_data[j];
output_data[(l - 1) * l] =
output_data[(l - 1) * l] * newscale_data[j] + newbias_data[j];
output_data[l * l - 1] =
output_data[l * l - 1] * newscale_data[j] + newbias_data[j];
}
if (if_relu) {
output_data[0] = output_data[0] < 0 ? 0 : output_data[0];
output_data[l - 1] = output_data[l - 1] < 0 ? 0 : output_data[l - 1];
output_data[(l - 1) * l] =
output_data[(l - 1) * l] < 0 ? 0 : output_data[(l - 1) * l];
output_data[l * l - 1] =
output_data[l * l - 1] < 0 ? 0 : output_data[l * l - 1];
}
for (int i = 1; i < l - 1; ++i) {
output_data[i * l] =
w01 * input_data[i * l - l] + w02 * input_data[i * l - l + 1] +
w11 * input_data[i * l] + w12 * input_data[i * l + 1] +
w21 * input_data[i * l + l] + w22 * input_data[i * l + l + 1];
output_data[i * l + l - 1] = w00 * input_data[i * l + l - 1 - l - 1] +
w01 * input_data[i * l + l - 1 - l] +
w10 * input_data[i * l + l - 1 - 1] +
w11 * input_data[i * l + l - 1] +
w20 * input_data[i * l + l - 1 + l - 1] +
w21 * input_data[i * l + l - 1 + l];
if (if_bias) {
output_data[i * l] += bias_data[j];
output_data[i * l + l - 1] += bias_data[j];
}
if (if_bn) {
output_data[i * l] =
output_data[i * l] * newscale_data[j] + newbias_data[j];
output_data[i * l + l - 1] =
output_data[i * l + l - 1] * newscale_data[j] + newbias_data[j];
}
if (if_relu) {
output_data[i * l] = output_data[i * l] < 0 ? 0 : output_data[i * l];
output_data[i * l + l - 1] =
output_data[i * l + l - 1] < 0 ? 0 : output_data[i * l + l - 1];
}
}
// top 1 row and bottom 1 row
const float *input_tmp = input_data;
float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
tmp3, tmp4, tmp5, out0;
in0 = vld1q_f32(input_tmp);
in2 = vld1q_f32(input_tmp + l);
const float *input_tmp_end = input_tmp + (l - 2) * l;
in4 = vld1q_f32(input_tmp_end);
in6 = vld1q_f32(input_tmp_end + l);
int c_mid = l_mid;
auto output_ptr = output_data + 1;
for (; c_mid > 3; c_mid -= 4) {
in1 = vld1q_f32(input_tmp + 4);
in3 = vld1q_f32(input_tmp + l + 4);
tmp0 = vextq_f32(in0, in1, 1);
tmp1 = vextq_f32(in0, in1, 2);
tmp2 = vextq_f32(in2, in3, 1);
tmp3 = vextq_f32(in2, in3, 2);
out0 = vmulq_n_f32(in0, w10);
out0 = vmlaq_n_f32(out0, tmp0, w11);
out0 = vmlaq_n_f32(out0, tmp1, w12);
out0 = vmlaq_n_f32(out0, in2, w20);
out0 = vmlaq_n_f32(out0, tmp2, w21);
out0 = vmlaq_n_f32(out0, tmp3, w22);
out0 = vaddq_f32(out0, vbias);
out0 = vmlaq_f32(vnewbias, vnewscale, out0);
if (if_relu) {
out0 = vmaxq_f32(out0, vzero);
}
vst1q_f32(output_ptr, out0);
in5 = vld1q_f32(input_tmp_end + 4);
in7 = vld1q_f32(input_tmp_end + l + 4);
tmp0 = vextq_f32(in4, in5, 1);
tmp1 = vextq_f32(in4, in5, 2);
tmp2 = vextq_f32(in6, in7, 1);
tmp3 = vextq_f32(in6, in7, 2);
out0 = vmulq_n_f32(in4, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in6, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vaddq_f32(out0, vbias);
out0 = vmlaq_f32(vnewbias, vnewscale, out0);
if (if_relu) {
out0 = vmaxq_f32(out0, vzero);
}
vst1q_f32(output_ptr + (l - 1) * l, out0);
// can optimize to each 8 stride.
input_tmp += 4;
input_tmp_end += 4;
output_ptr += 4;
in0 = in1;
in2 = in3;
in4 = in5;
in6 = in7;
}
// top right pad
float32x4_t pad0 = vdupq_n_f32(input_data[l - 1]);
float32x4_t pad1 = vdupq_n_f32(input_data[2 * l - 1]);
tmp0 = vextq_f32(in0, pad0, 1);
tmp1 = vextq_f32(in0, pad0, 2);
tmp2 = vextq_f32(in2, pad1, 1);
tmp3 = vextq_f32(in2, pad1, 2);
out0 = vmulq_n_f32(in0, w10);
out0 = vmlaq_n_f32(out0, tmp0, w11);
out0 = vmlaq_n_f32(out0, tmp1, w12);
out0 = vmlaq_n_f32(out0, in2, w20);
out0 = vmlaq_n_f32(out0, tmp2, w21);
out0 = vmlaq_n_f32(out0, tmp3, w22);
out0 = vaddq_f32(out0, vbias);
out0 = vmlaq_f32(vnewbias, vnewscale, out0);
if (if_relu) {
out0 = vmaxq_f32(out0, vzero);
}
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + i, out0, 2);
}
}
// bottom right pad
float32x4_t pad2 = vdupq_n_f32(input_data[l * l - 1 - l]);
float32x4_t pad3 = vdupq_n_f32(input_data[l * l - 1]);
tmp0 = vextq_f32(in4, pad2, 1);
tmp1 = vextq_f32(in4, pad2, 2);
tmp2 = vextq_f32(in6, pad3, 1);
tmp3 = vextq_f32(in6, pad3, 2);
out0 = vmulq_n_f32(in4, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in6, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vaddq_f32(out0, vbias);
out0 = vmlaq_f32(vnewbias, vnewscale, out0);
if (if_relu) {
out0 = vmaxq_f32(out0, vzero);
}
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + (l - 1) * l + i, out0, 2);
}
}
// mid
for (int i = 0; i < l - 2; ++i) {
auto output_ptr = output_data + (i + 1) * l + 1;
input_tmp = input_data + i * l;
auto in0_tmp = vld1q_f32(input_tmp);
auto in2_tmp = vld1q_f32(input_tmp + l);
auto in4_tmp = vld1q_f32(input_tmp + l + l);
c_mid = l_mid;
for (; c_mid > 3; c_mid -= 4) {
auto in1_tmp = vld1q_f32(input_tmp + 4);
auto in3_tmp = vld1q_f32(input_tmp + l + 4);
auto in5_tmp = vld1q_f32(input_tmp + l + l + 4);
tmp0 = vextq_f32(in0_tmp, in1_tmp, 1);
tmp1 = vextq_f32(in0_tmp, in1_tmp, 2);
tmp2 = vextq_f32(in2_tmp, in3_tmp, 1);
tmp3 = vextq_f32(in2_tmp, in3_tmp, 2);
tmp4 = vextq_f32(in4_tmp, in5_tmp, 1);
tmp5 = vextq_f32(in4_tmp, in5_tmp, 2);
out0 = vmulq_n_f32(in0_tmp, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in2_tmp, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vmlaq_n_f32(out0, in4_tmp, w20);
out0 = vmlaq_n_f32(out0, tmp4, w21);
out0 = vmlaq_n_f32(out0, tmp5, w22);
out0 = vaddq_f32(out0, vbias);
out0 = vmlaq_f32(vnewbias, vnewscale, out0);
if (if_relu) {
out0 = vmaxq_f32(out0, vzero);
}
vst1q_f32(output_ptr, out0);
output_ptr += 4;
input_tmp += 4;
in0_tmp = in1_tmp;
in2_tmp = in3_tmp;
in4_tmp = in5_tmp;
}
float32x4_t pad0 = vdupq_n_f32(input_data[i * l + l - 1]);
float32x4_t pad1 = vdupq_n_f32(input_data[i * l + l - 1 + l]);
float32x4_t pad2 = vdupq_n_f32(input_data[i * l + l - 1 + l + l]);
tmp0 = vextq_f32(in0_tmp, pad0, 1);
tmp1 = vextq_f32(in0_tmp, pad0, 2);
tmp2 = vextq_f32(in2_tmp, pad1, 1);
tmp3 = vextq_f32(in2_tmp, pad1, 2);
tmp4 = vextq_f32(in4_tmp, pad2, 1);
tmp5 = vextq_f32(in4_tmp, pad2, 2);
out0 = vmulq_n_f32(in0_tmp, w00);
out0 = vmlaq_n_f32(out0, tmp0, w01);
out0 = vmlaq_n_f32(out0, tmp1, w02);
out0 = vmlaq_n_f32(out0, in2_tmp, w10);
out0 = vmlaq_n_f32(out0, tmp2, w11);
out0 = vmlaq_n_f32(out0, tmp3, w12);
out0 = vmlaq_n_f32(out0, in4_tmp, w20);
out0 = vmlaq_n_f32(out0, tmp4, w21);
out0 = vmlaq_n_f32(out0, tmp5, w22);
out0 = vaddq_f32(out0, vbias);
out0 = vmlaq_f32(vnewbias, vnewscale, out0);
if (if_relu) {
out0 = vmaxq_f32(out0, vzero);
}
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + i, out0, 2);
}
}
}
output_data += hxw;
input_data += hxw;
filter_data_tmp += 9;
}
}
}
} // namespace math
} // namespace operators
} // namespace paddle_mobile
src/operators/math/depthwise_conv_3x3.h 0 → 100644
浏览文件 @ c67b701f
/* 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 <algorithm>
#include <vector>
#include "framework/tensor.h"
#include "operators/math/conv_func.h"
namespace paddle_mobile {
namespace operators {
namespace math {
using framework::Tensor;
using std::max;
using std::min;
using std::vector;
void DepthwiseConv3x3(const Tensor *input, vector<int> strides,
vector<int> paddings, const Tensor *filter, Tensor *bias,
Tensor *output, bool if_bias);
void DepthwiseConv3x3s1p1(const Tensor *input, const Tensor *filter,
Tensor *output, Tensor *bias, bool if_bias);
void DepthwiseConvAddBNRelu3x3s1p1(const Tensor *input, const Tensor *filter,
Tensor *output, Tensor *bias, bool if_bias,
const Tensor *new_scale,
const Tensor *new_bias, bool if_bn,
bool if_relu);
} // namespace math
} // namespace operators
} // namespace paddle_mobile
src/operators/math/gemm.cpp
浏览文件 @ c67b701f
......@@ -116,6 +116,8 @@ void PackMatrixB_(int k, int n, int n_tail, const float *B, int ldb,
int i, j;
const float *Bij;
for (j = 0; j < n - n_tail; j += NR) {
#ifdef ARMV7
for (i = 0; i < k; ++i) {
Bij = &B(i, j);
asm volatile(
......@@ -125,6 +127,15 @@ void PackMatrixB_(int k, int n, int n_tail, const float *B, int ldb,
: [Bij] "r"(Bij)
: "memory", "q0");
}
#else
for (i = 0; i < k; ++i) {
Bij = &B(i, j);
*buffer++ = *Bij;
*buffer++ = *(Bij + 1);
*buffer++ = *(Bij + 2);
*buffer++ = *(Bij + 3);
}
#endif
}
if (n_tail != 0) {
for (i = 0; i < k; ++i) {
......@@ -747,10 +758,14 @@ void sgemm(int m, int n, int k, float alpha, const float *A, int lda,
const float *B, int ldb, float beta, float *C, int ldc) {
int i, j, p, mc, nc, kc;
float beta_;
#ifdef ARMV7
if (m == 1) {
VectorKernel(1, n, k, alpha, A, lda, B, ldb, beta, C, ldc);
return;
}
#endif
for (j = 0; j < n; j += NC) {
nc = s_min(n - j, NC);
for (p = 0; p < k; p += KC) {
......@@ -797,6 +812,7 @@ void sgemm_relu(int m, int n, int k, float alpha, const float *A, int lda,
}
}
#ifdef ARMV7
void VectorKernel(int m, int n, int k, float alpha, const float *A, int lda,
const float *B, int ldb, float beta, float *C, int ldc) {
float *bufferC = static_cast<float *>(memory::Alloc(sizeof(float) * n));
......@@ -1010,6 +1026,7 @@ void VectorKernel(int m, int n, int k, float alpha, const float *A, int lda,
}
}
}
#endif
} // namespace math
} // namespace operators
......
src/operators/math/pool_2x2.cpp
浏览文件 @ c67b701f
......@@ -22,6 +22,9 @@ namespace math {
void Pool2x2Max(vector<int> strides, vector<int> paddings, const Tensor *input,
Tensor *output) {
#if __ARM_NEON
#ifdef ARMV7
const int batch_size = input->dims()[0];
const int input_height = input->dims()[2];
......@@ -90,11 +93,15 @@ void Pool2x2Max(vector<int> strides, vector<int> paddings, const Tensor *input,
output_data += output_batch_stride;
}
#endif
#endif
}
void Pool2x2Avg(vector<int> strides, vector<int> paddings, const Tensor *input,
Tensor *output) {
#if __ARM_NEON
#ifdef ARMV7
const int batch_size = input->dims()[0];
const int input_height = input->dims()[2];
......@@ -164,6 +171,12 @@ void Pool2x2Avg(vector<int> strides, vector<int> paddings, const Tensor *input,
input_data += input_batch_stride;
output_data += output_batch_stride;
}
#else
// TODO(): to imp other asm
#endif
#endif
}
......
src/operators/math/pool_3x3.cpp
浏览文件 @ c67b701f
......@@ -13,13 +13,12 @@ See the License for the specific language governing permissions and
limitations under the License. */
#ifdef POOL_OP
#define __ARM_NEON true
#include "pool_3x3.h"
#include "framework/tensor.h"
#if __ARM_NEON
#include <arm_neon.h>
#endif // __ARM_NEON
#include <climits>
namespace paddle_mobile {
namespace operators {
namespace math {
......@@ -27,6 +26,481 @@ using framework::Tensor;
using std::max;
using std::min;
using std::vector;
void Pool3x3Avgs1p1(const Tensor *input, Tensor *output) {
#if __ARM_NEON
const int batch_size = input->dims()[0];
const int h_in = input->dims()[2];
const int w_in = input->dims()[3];
const int output_channels = output->dims()[1];
const int h_out = output->dims()[2];
const int w_out = output->dims()[3];
const int outputdata_channel_stride = h_out * w_out;
const int inputdata_channel_stride = h_in * w_in;
float *out_data = output->data<float>();
const float *input_data = input->data<float>();
const float coef = 1.0 / 9.0;
for (int k = 0; k < batch_size; ++k) {
for (int c = 0; c < output_channels; ++c) {
// four corner point
out_data[0] = (input_data[0] + input_data[1] + input_data[w_in] +
input_data[w_in + 1]) *
coef;
out_data[w_out - 1] =
(input_data[w_in - 2] + input_data[w_in - 1] +
input_data[w_in * 2 - 2] + input_data[2 * w_in - 1]) *
coef;
out_data[(h_out - 1) * w_out] =
(input_data[(h_in - 2) * w_in] + input_data[(h_in - 2) * w_in + 1] +
input_data[(h_in - 1) * w_in] + input_data[(h_in - 1) * w_in + 1]) *
coef;
out_data[h_out * w_out - 1] =
(input_data[h_in * w_in - 1] + input_data[h_in * w_in - 2] +
input_data[(h_in - 1) * w_in - 1] +
input_data[(h_in - 1) * w_in - 2]) *
coef;
// left side & right side
for (int i = 1; i < h_in - 1; ++i) {
out_data[i * w_out] =
(input_data[i * w_in - w_in] + input_data[i * w_in - w_in + 1] +
input_data[i * w_in] + input_data[i * w_in + 1] +
input_data[i * w_in + w_in] + input_data[i * w_in + w_in + 1]) *
coef;
out_data[i * w_out + w_out - 1] =
(input_data[i * w_in - w_in + w_in - 2] +
input_data[i * w_in - w_in + 1 + w_in - 2] +
input_data[i * w_in + w_in - 2] +
input_data[i * w_in + 1 + w_in - 2] +
input_data[i * w_in + w_in + w_in - 2] +
input_data[i * w_in + w_in + 1 + w_in - 2]) *
coef;
}
// top 1 row & bottom 1 row
const float *input_tmp = input_data;
float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
tmp3, tmp4, tmp5, sum, out0;
float32x4_t v_coef = vdupq_n_f32(coef);
in0 = vld1q_f32(input_tmp);
in2 = vld1q_f32(input_tmp + w_in);
const float *input_tmp_end = input_tmp + (h_in - 2) * w_in;
in4 = vld1q_f32(input_tmp_end);
in6 = vld1q_f32(input_tmp_end + w_in);
int c_mid = w_out - 2;
auto output_ptr = out_data + 1;
for (; c_mid > 3; c_mid -= 4) {
in1 = vld1q_f32(input_tmp + 4);
in3 = vld1q_f32(input_tmp + w_in + 4);
tmp0 = vextq_f32(in0, in1, 1);
tmp1 = vextq_f32(in0, in1, 2);
tmp2 = vextq_f32(in2, in3, 1);
tmp3 = vextq_f32(in2, in3, 2);
sum = vaddq_f32(in0, tmp0);
sum = vaddq_f32(sum, tmp1);
sum = vaddq_f32(sum, in2);
sum = vaddq_f32(sum, tmp2);
sum = vaddq_f32(sum, tmp3);
vst1q_f32(output_ptr, vmulq_f32(sum, v_coef));
in5 = vld1q_f32(input_tmp_end + 4);
in7 = vld1q_f32(input_tmp_end + w_in + 4);
tmp0 = vextq_f32(in4, in5, 1);
tmp1 = vextq_f32(in4, in5, 2);
tmp2 = vextq_f32(in6, in7, 1);
tmp3 = vextq_f32(in6, in7, 2);
sum = vaddq_f32(in0, tmp0);
sum = vaddq_f32(sum, tmp1);
sum = vaddq_f32(sum, in2);
sum = vaddq_f32(sum, tmp2);
sum = vaddq_f32(sum, tmp3);
vst1q_f32(output_ptr + (h_out - 1) * w_out, vmulq_f32(sum, v_coef));
// can optimize to each 8 stride.
input_tmp += 4;
input_tmp_end += 4;
output_ptr += 4;
in0 = in1;
in2 = in3;
in4 = in5;
in6 = in7;
}
// top right remain
float32x4_t pad0 = vdupq_n_f32(input_data[w_in - 1]);
float32x4_t pad1 = vdupq_n_f32(input_data[2 * w_in - 1]);
tmp0 = vextq_f32(in0, pad0, 1);
tmp1 = vextq_f32(in0, pad0, 2);
tmp2 = vextq_f32(in2, pad1, 2);
tmp3 = vextq_f32(in2, pad1, 2);
sum = vaddq_f32(in0, tmp0);
sum = vaddq_f32(sum, tmp1);
sum = vaddq_f32(sum, in2);
sum = vaddq_f32(sum, tmp2);
sum = vaddq_f32(sum, tmp3);
out0 = vmulq_f32(sum, v_coef);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + i, out0, 2);
}
}
// bottom_right remain
float32x4_t pad2 = vdupq_n_f32(input_data[(h_in - 1) * w_in - 1]);
float32x4_t pad3 = vdupq_n_f32(input_data[h_in * w_in - 1]);
tmp0 = vextq_f32(in4, pad2, 1);
tmp1 = vextq_f32(in4, pad2, 2);
tmp2 = vextq_f32(in6, pad3, 2);
tmp3 = vextq_f32(in6, pad3, 2);
sum = vaddq_f32(in4, tmp0);
sum = vaddq_f32(sum, tmp1);
sum = vaddq_f32(sum, in6);
sum = vaddq_f32(sum, tmp2);
sum = vaddq_f32(sum, tmp3);
out0 = vmulq_f32(sum, v_coef);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, out0, 2);
}
}
// mid
for (int j = 0; j < h_out - 2; ++j) {
output_ptr = out_data + w_out * (j + 1) + 1;
input_tmp = input_data + j * w_in;
in0 = vld1q_f32(input_tmp);
in2 = vld1q_f32(input_tmp + w_in);
in4 = vld1q_f32(input_tmp + 2 * w_in);
c_mid = w_out - 2;
for (; c_mid > 3; c_mid -= 4) {
in1 = vld1q_f32(input_tmp + 4);
in3 = vld1q_f32(input_tmp + w_in + 4);
in5 = vld1q_f32(input_tmp + 2 * w_in + 4);
tmp0 = vextq_f32(in0, in1, 1);
tmp1 = vextq_f32(in0, in1, 2);
tmp2 = vextq_f32(in2, in3, 1);
tmp3 = vextq_f32(in2, in3, 2);
tmp4 = vextq_f32(in4, in5, 1);
tmp5 = vextq_f32(in4, in5, 2);
sum = vaddq_f32(in0, tmp0);
sum = vaddq_f32(sum, tmp1);
sum = vaddq_f32(sum, in2);
sum = vaddq_f32(sum, tmp2);
sum = vaddq_f32(sum, tmp3);
sum = vaddq_f32(sum, in4);
sum = vaddq_f32(sum, tmp4);
sum = vaddq_f32(sum, tmp5);
out0 = vmulq_f32(sum, v_coef);
vst1q_f32(output_ptr, out0);
output_ptr += 4;
input_tmp += 4;
in0 = in1;
in2 = in3;
in4 = in5;
}
// mid remain
float32x4_t pad0 = vdupq_n_f32(input_data[(j + 1) * w_in - 1]);
float32x4_t pad1 = vdupq_n_f32(input_data[(j + 2) * w_in - 1]);
float32x4_t pad2 = vdupq_n_f32(input_data[(j + 2) * w_in - 1]);
tmp0 = vextq_f32(in0, pad0, 1);
tmp1 = vextq_f32(in0, pad0, 2);
tmp2 = vextq_f32(in2, pad1, 1);
tmp3 = vextq_f32(in2, pad1, 2);
tmp4 = vextq_f32(in4, pad2, 1);
tmp5 = vextq_f32(in4, pad2, 2);
sum = vaddq_f32(in0, tmp0);
sum = vaddq_f32(sum, tmp1);
sum = vaddq_f32(sum, in2);
sum = vaddq_f32(sum, tmp2);
sum = vaddq_f32(sum, tmp3);
sum = vaddq_f32(sum, in4);
sum = vaddq_f32(sum, tmp4);
sum = vaddq_f32(sum, tmp5);
out0 = vmulq_f32(sum, v_coef);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + i, out0, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + i, out0, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + i, out0, 2);
}
}
}
input_data += inputdata_channel_stride;
out_data += outputdata_channel_stride;
}
}
#endif
}
void Pool3x3Maxs1p1(const Tensor *input, Tensor *output) {
#if __ARM_NEON
const int batch_size = input->dims()[0];
const int h_in = input->dims()[2];
const int w_in = input->dims()[3];
const int output_channels = output->dims()[1];
const int h_out = output->dims()[2];
const int w_out = output->dims()[3];
const int outputdata_channel_stride = h_out * w_out;
const int inputdata_channel_stride = h_in * w_in;
float *out_data = output->data<float>();
const float *input_data = input->data<float>();
for (int k = 0; k < batch_size; ++k) {
for (int c = 0; c < output_channels; ++c) {
// four corner point
out_data[0] = std::max(std::max(input_data[0], input_data[1]),
std::max(input_data[w_in], input_data[w_in + 1]));
out_data[w_out - 1] = std::max(
std::max(input_data[w_in - 2], input_data[w_in - 1]),
std::max(input_data[w_in * 2 - 2], input_data[2 * w_in - 1]));
out_data[(h_out - 1) * w_out] =
std::max(std::max(input_data[(h_in - 2) * w_in],
input_data[(h_in - 2) * w_in + 1]),
std::max(input_data[(h_in - 1) * w_in],
input_data[(h_in - 1) * w_in + 1]));
out_data[h_out * w_out - 1] = std::max(
std::max(input_data[(h_in - 1) * w_in - 1],
input_data[(h_in - 1) * w_in - 2]),
std::max(input_data[h_in * w_in - 1], input_data[h_in * w_in - 2]));
// left side & right side
for (int i = 1; i < h_in - 1; ++i) {
float max1 = std::max(input_data[i * w_in - w_in],
input_data[i * w_in - w_in + 1]);
float max2 = std::max(input_data[i * w_in], input_data[i * w_in + 1]);
float max3 = std::max(input_data[i * w_in + w_in],
input_data[i * w_in + w_in + 1]);
out_data[i * w_out] = std::max(std::max(max1, max2), max3);
max1 = std::max(input_data[i * w_in - w_in + w_in - 2],
input_data[i * w_in - w_in + 1 + w_in - 2]);
max2 = std::max(input_data[i * w_in + w_in - 2],
input_data[i * w_in + 1 + w_in - 2]);
max3 = std::max(input_data[i * w_in + w_in + w_in - 2],
input_data[i * w_in + w_in + 1 + w_in - 2]);
out_data[i * w_out + w_out - 1] = std::max(std::max(max1, max2), max3);
}
// top 1 row & bottom 1 row
const float *input_tmp = input_data;
float32x4_t in0, in1, in2, in3, in4, in5, in6, in7, tmp0, tmp1, tmp2,
tmp3, tmp4, tmp5, max;
in0 = vld1q_f32(input_tmp);
in2 = vld1q_f32(input_tmp + w_in);
const float *input_tmp_end = input_tmp + (h_in - 2) * w_in;
in4 = vld1q_f32(input_tmp_end);
in6 = vld1q_f32(input_tmp_end + w_in);
int c_mid = w_out - 2;
auto output_ptr = out_data + 1;
for (; c_mid > 3; c_mid -= 4) {
in1 = vld1q_f32(input_tmp + 4);
in3 = vld1q_f32(input_tmp + w_in + 4);
tmp0 = vextq_f32(in0, in1, 1);
tmp1 = vextq_f32(in0, in1, 2);
tmp2 = vextq_f32(in2, in3, 1);
tmp3 = vextq_f32(in2, in3, 2);
max = vmaxq_f32(in0, tmp0);
max = vmaxq_f32(max, tmp1);
max = vmaxq_f32(max, in2);
max = vmaxq_f32(max, tmp2);
max = vmaxq_f32(max, tmp3);
vst1q_f32(output_ptr, max);
in5 = vld1q_f32(input_tmp_end + 4);
in7 = vld1q_f32(input_tmp_end + w_in + 4);
tmp0 = vextq_f32(in4, in5, 1);
tmp1 = vextq_f32(in4, in5, 2);
tmp2 = vextq_f32(in6, in7, 1);
tmp3 = vextq_f32(in6, in7, 2);
max = vmaxq_f32(in4, tmp0);
max = vmaxq_f32(max, tmp1);
max = vmaxq_f32(max, in6);
max = vmaxq_f32(max, tmp2);
max = vmaxq_f32(max, tmp3);
vst1q_f32(output_ptr + (h_out - 1) * w_out, max);
input_tmp += 4;
input_tmp_end += 4;
output_ptr += 4;
in0 = in1;
in2 = in3;
in4 = in5;
in6 = in7;
}
// top right remain
float32x4_t pad0 = vdupq_n_f32(input_data[w_in - 1]);
float32x4_t pad1 = vdupq_n_f32(input_data[2 * w_in - 1]);
tmp0 = vextq_f32(in0, pad0, 1);
tmp1 = vextq_f32(in0, pad0, 2);
tmp2 = vextq_f32(in2, pad1, 1);
tmp3 = vextq_f32(in2, pad1, 2);
max = vmaxq_f32(in0, tmp0);
max = vmaxq_f32(max, tmp1);
max = vmaxq_f32(max, in2);
max = vmaxq_f32(max, tmp2);
max = vmaxq_f32(max, tmp3);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + i, max, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + i, max, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + i, max, 2);
}
}
// bottom_right remain
float32x4_t pad2 = vdupq_n_f32(input_data[(h_in - 1) * w_in - 1]);
float32x4_t pad3 = vdupq_n_f32(input_data[h_in * w_in - 1]);
tmp0 = vextq_f32(in4, pad2, 1);
tmp1 = vextq_f32(in4, pad2, 2);
tmp2 = vextq_f32(in6, pad3, 1);
tmp3 = vextq_f32(in6, pad3, 2);
max = vmaxq_f32(in4, tmp0);
max = vmaxq_f32(max, tmp1);
max = vmaxq_f32(max, in6);
max = vmaxq_f32(max, tmp2);
max = vmaxq_f32(max, tmp3);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, max, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, max, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + (h_out - 1) * w_out + i, max, 2);
}
}
// mid
for (int j = 0; j < h_out - 2; ++j) {
output_ptr = out_data + (j + 1) * w_out + 1;
input_tmp = input_data + j * w_in;
in0 = vld1q_f32(input_tmp);
in2 = vld1q_f32(input_tmp + w_in);
in4 = vld1q_f32(input_tmp + 2 * w_in);
c_mid = w_out - 2;
for (; c_mid > 3; c_mid -= 4) {
in1 = vld1q_f32(input_tmp + 4);
in3 = vld1q_f32(input_tmp + w_in + 4);
in5 = vld1q_f32(input_tmp + 2 * w_in + 4);
tmp0 = vextq_f32(in0, in1, 1);
tmp1 = vextq_f32(in0, in1, 2);
tmp2 = vextq_f32(in2, in3, 1);
tmp3 = vextq_f32(in2, in3, 2);
tmp4 = vextq_f32(in4, in5, 1);
tmp5 = vextq_f32(in4, in5, 2);
max = vmaxq_f32(in0, tmp0);
max = vmaxq_f32(max, tmp1);
max = vmaxq_f32(max, in2);
max = vmaxq_f32(max, tmp2);
max = vmaxq_f32(max, tmp3);
max = vmaxq_f32(max, in4);
max = vmaxq_f32(max, tmp4);
max = vmaxq_f32(max, tmp5);
vst1q_f32(output_ptr, max);
output_ptr += 4;
input_tmp += 4;
in0 = in1;
in2 = in3;
in4 = in5;
}
// mid remain
float32x4_t pad0 = vdupq_n_f32(input_data[(j + 1) * w_in - 1]);
float32x4_t pad1 = vdupq_n_f32(input_data[(j + 2) * w_in - 1]);
float32x4_t pad2 = vdupq_n_f32(input_data[(j + 3) * w_in - 1]);
tmp0 = vextq_f32(in0, pad0, 1);
tmp1 = vextq_f32(in0, pad0, 2);
tmp2 = vextq_f32(in2, pad1, 1);
tmp3 = vextq_f32(in2, pad1, 2);
tmp4 = vextq_f32(in4, pad2, 1);
tmp5 = vextq_f32(in4, pad2, 2);
max = vmaxq_f32(in0, tmp0);
max = vmaxq_f32(max, tmp1);
max = vmaxq_f32(max, in2);
max = vmaxq_f32(max, tmp2);
max = vmaxq_f32(max, tmp3);
max = vmaxq_f32(max, in4);
max = vmaxq_f32(max, tmp4);
max = vmaxq_f32(max, tmp5);
for (int i = 0; i < c_mid; ++i) {
if (i == 0) {
vst1q_lane_f32(output_ptr + i, max, 0);
}
if (i == 1) {
vst1q_lane_f32(output_ptr + i, max, 1);
}
if (i == 2) {
vst1q_lane_f32(output_ptr + i, max, 2);
}
}
}
input_data += inputdata_channel_stride;
out_data += outputdata_channel_stride;
}
}
#endif
}
void Pool3x3Max(vector<int> strides, vector<int> paddings, const Tensor *input,
Tensor *output) {
......
src/operators/math/pool_3x3.h
浏览文件 @ c67b701f
......@@ -15,7 +15,8 @@ limitations under the License. */
#ifdef POOL_OP
#pragma once
#include <algorithm>
#include <vector>
#include "framework/tensor.h"
#if __ARM_NEON
#include <arm_neon.h>
......@@ -26,7 +27,8 @@ namespace operators {
namespace math {
using framework::Tensor;
using std::vector;
void Pool3x3Avgs1p1(const Tensor *input, Tensor *output);
void Pool3x3Maxs1p1(const Tensor *input, Tensor *output);
void Pool3x3Max(vector<int> strides, vector<int> paddings, const Tensor *input,
Tensor *output);
......
src/operators/math/pooling.cpp
浏览文件 @ c67b701f
......@@ -57,7 +57,7 @@ class PoolFunctor<CPU, PoolProcess, T> {
T *output_data = output->mutable_data<T>();
for (int i = 0; i < batch_size; i++) {
#pragma omp parallel for
// #pragma omp parallel for
for (int c = 0; c < output_channels; ++c) {
for (int ph = 0; ph < output_height; ++ph) {
int hstart = ph * stride_height - padding_height;
......
src/operators/mul_op.cpp
浏览文件 @ c67b701f
......@@ -56,11 +56,9 @@ template class MulOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(mul);
REGISTER_OPERATOR_CPU(mul, ops::MulOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(mul);
REGISTER_OPERATOR_MALI_GPU(mul, ops::MulOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/mul_op.h
浏览文件 @ c67b701f
......@@ -46,4 +46,13 @@ class MulOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(mul);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(mul);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/multiclass_nms_op.cpp
浏览文件 @ c67b701f
......@@ -40,7 +40,6 @@ template class MultiClassNMSOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(multiclass_nms);
REGISTER_OPERATOR_CPU(multiclass_nms, ops::MultiClassNMSOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......
src/operators/multiclass_nms_op.h
浏览文件 @ c67b701f
......@@ -52,4 +52,12 @@ class MultiClassNMSOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(multiclass_nms);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/op_param.h
浏览文件 @ c67b701f
......@@ -195,8 +195,7 @@ class OpParam {
class ConvParam : OpParam {
public:
ConvParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
const AttributeMap &attrs, const Scope &scope) {
filter_ = FilterFrom<LoDTensor>(inputs, scope);
input_ = InputFrom<LoDTensor>(inputs, scope);
output_ = OutputFrom<LoDTensor>(outputs, scope);
......@@ -237,12 +236,11 @@ Print &operator<<(Print &printer, const ConvParam &conv_param);
class ElementwiseAddParam : OpParam {
public:
ElementwiseAddParam(const VariableNameMap &inputs,
const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_x_ = InputXFrom<framework::LoDTensor>(inputs, scope);
input_y_ = InputYFrom<framework::LoDTensor>(inputs, scope);
out_ = OutFrom<framework::LoDTensor>(outputs, scope);
const VariableNameMap &outputs, const AttributeMap &attrs,
const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
input_y_ = InputYFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
axis_ = GetAttr<int>("axis", attrs);
}
......@@ -267,11 +265,10 @@ class ElementwiseAddParam : OpParam {
class MulParam : OpParam {
public:
MulParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_x_ = InputXFrom<framework::LoDTensor>(inputs, scope);
input_y_ = InputYFrom<framework::LoDTensor>(inputs, scope);
out_ = OutFrom<framework::LoDTensor>(outputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
input_y_ = InputYFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
x_num_col_dims_ = GetAttr<int>("x_num_col_dims", attrs);
y_num_col_dims_ = GetAttr<int>("y_num_col_dims", attrs);
}
......@@ -299,10 +296,9 @@ class MulParam : OpParam {
class ConcatParam : public OpParam {
public:
ConcatParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
const AttributeMap &attrs, const Scope &scope) {
inputs_ = InputMultiFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<framework::LoDTensor>(outputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
axis_ = GetAttr<int>("axis", attrs);
}
......@@ -323,11 +319,10 @@ class ConcatParam : public OpParam {
class LrnParam : public OpParam {
public:
LrnParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_x_ = InputXFrom<framework::LoDTensor>(inputs, scope);
out_ = OutFrom<framework::LoDTensor>(outputs, scope);
mid_out_ = MidOutFrom<framework::LoDTensor>(outputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
mid_out_ = MidOutFrom<LoDTensor>(outputs, scope);
n_ = GetAttr<int>("n", attrs);
alpha_ = GetAttr<float>("alpha", attrs);
beta_ = GetAttr<float>("beta", attrs);
......@@ -367,14 +362,13 @@ class LrnParam : public OpParam {
class BatchNormParam : OpParam {
public:
BatchNormParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_x_ = InputXFrom<framework::LoDTensor>(inputs, scope);
output_y_ = OutputYFrom<framework::LoDTensor>(outputs, scope);
input_bias_ = InputBiasFrom<framework::LoDTensor>(inputs, scope);
input_mean_ = InputMeanFrom<framework::LoDTensor>(inputs, scope);
input_scale_ = InputScaleFrom<framework::LoDTensor>(inputs, scope);
input_variance_ = InputVarianceFrom<framework::LoDTensor>(inputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
output_y_ = OutputYFrom<LoDTensor>(outputs, scope);
input_bias_ = InputBiasFrom<LoDTensor>(inputs, scope);
input_mean_ = InputMeanFrom<LoDTensor>(inputs, scope);
input_scale_ = InputScaleFrom<LoDTensor>(inputs, scope);
input_variance_ = InputVarianceFrom<LoDTensor>(inputs, scope);
epsilon_ = GetAttr<float>("epsilon", attrs);
momentum_ = GetAttr<float>("momentum", attrs);
is_test_ = GetAttr<bool>("is_test", attrs);
......@@ -418,11 +412,10 @@ class BatchNormParam : OpParam {
class PoolParam : public OpParam {
public:
PoolParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_ = InputXFrom<framework::LoDTensor>(inputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_ = InputXFrom<LoDTensor>(inputs, scope);
output_ = OutFrom<framework::LoDTensor>(outputs, scope);
output_ = OutFrom<LoDTensor>(outputs, scope);
pooling_type_ = GetAttr<string>("pooling_type", attrs);
ksize_ = GetAttr<vector<int>>("ksize", attrs);
strides_ = GetAttr<vector<int>>("strides", attrs);
......@@ -464,13 +457,11 @@ class PoolParam : public OpParam {
class PriorBoxParam : public OpParam {
public:
PriorBoxParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_ = InputFrom<framework::LoDTensor>(inputs, scope);
input_image_ = InputImageFrom<framework::LoDTensor>(inputs, scope);
output_boxes_ = OutputBoxesFrom<framework::LoDTensor>(outputs, scope);
output_variances_ =
OutputVariancesFrom<framework::LoDTensor>(outputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_ = InputFrom<LoDTensor>(inputs, scope);
input_image_ = InputImageFrom<LoDTensor>(inputs, scope);
output_boxes_ = OutputBoxesFrom<LoDTensor>(outputs, scope);
output_variances_ = OutputVariancesFrom<LoDTensor>(outputs, scope);
min_sizes_ = GetAttr<vector<float>>("min_sizes", attrs);
max_sizes_ = GetAttr<vector<float>>("max_sizes", attrs);
aspect_ratios_ = GetAttr<vector<float>>("aspect_ratios", attrs);
......@@ -528,13 +519,11 @@ class PriorBoxParam : public OpParam {
class BoxCoderParam : public OpParam {
public:
BoxCoderParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_priorbox_ = InputPriorBoxFrom<framework::LoDTensor>(inputs, scope);
input_priorboxvar_ =
InputPriorBoxVarFrom<framework::LoDTensor>(inputs, scope);
input_targetbox_ = InputTargetBoxFrom<framework::LoDTensor>(inputs, scope);
output_box_ = OutputBoxFrom<framework::LoDTensor>(outputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_priorbox_ = InputPriorBoxFrom<LoDTensor>(inputs, scope);
input_priorboxvar_ = InputPriorBoxVarFrom<LoDTensor>(inputs, scope);
input_targetbox_ = InputTargetBoxFrom<LoDTensor>(inputs, scope);
output_box_ = OutputBoxFrom<LoDTensor>(outputs, scope);
code_type_ = GetAttr<std::string>("code_type", attrs);
}
const Tensor *InputPriorBox() const { return input_priorbox_; }
......@@ -560,10 +549,9 @@ class BoxCoderParam : public OpParam {
class SoftmaxParam : public OpParam {
public:
SoftmaxParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_x_ = InputXFrom<framework::LoDTensor>(inputs, scope);
out_ = OutFrom<framework::LoDTensor>(outputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
}
const Tensor *InputX() const { return input_x_; }
Tensor *Out() const { return out_; }
......@@ -578,10 +566,9 @@ class SoftmaxParam : public OpParam {
class SigmoidParam : public OpParam {
public:
SigmoidParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_x_ = InputXFrom<framework::LoDTensor>(inputs, scope);
out_ = OutFrom<framework::LoDTensor>(outputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
}
const Tensor *InputX() const { return input_x_; }
Tensor *Out() const { return out_; }
......@@ -643,9 +630,9 @@ class MultiClassNMSParam : public OpParam {
class FeedParam : public OpParam {
public:
FeedParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs, framework::Scope &scope) {
input_x_ = InputXFrom<framework::LoDTensor>(inputs, scope);
out_ = OutFrom<framework::LoDTensor>(outputs, scope);
const AttributeMap &attrs, Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
auto var = scope.Var("batch_size");
batch_size = var->GetValue<int>();
}
......@@ -662,10 +649,9 @@ class FeedParam : public OpParam {
class FetchParam : public OpParam {
public:
FetchParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const framework::AttributeMap &attrs,
const framework::Scope &scope) {
input_x_ = InputXFrom<framework::LoDTensor>(inputs, scope);
out_ = OutFrom<framework::LoDTensor>(outputs, scope);
const AttributeMap &attrs, const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
}
const Tensor *InputX() const { return input_x_; }
Tensor *Out() const { return out_; }
......@@ -848,5 +834,143 @@ class FusionConvAddReluParam : public FusionConvAddParam {
};
#endif
#ifdef FUSION_CONVADDBNRELU_OP
class FusionConvAddBNReluParam : public OpParam {
public:
FusionConvAddBNReluParam(const VariableNameMap &inputs,
const VariableNameMap &outputs,
const AttributeMap &attrs, const Scope &scope) {
bias_ = InputYFrom<LoDTensor>(inputs, scope);
axis_ = GetAttr<int>("axis", attrs);
filter_ = FilterFrom<LoDTensor>(inputs, scope);
input_ = InputFrom<LoDTensor>(inputs, scope);
output_ = OutFrom<LoDTensor>(outputs, scope);
strides_ = GetAttr<vector<int>>("strides", attrs);
paddings_ = GetAttr<vector<int>>("paddings", attrs);
dilations_ = GetAttr<vector<int>>("dilations", attrs);
groups = GetAttr<int>("groups", attrs);
input_bias_ = InputBiasFrom<LoDTensor>(inputs, scope);
input_mean_ = InputMeanFrom<LoDTensor>(inputs, scope);
input_scale_ = InputScaleFrom<LoDTensor>(inputs, scope);
input_variance_ = InputVarianceFrom<LoDTensor>(inputs, scope);
epsilon_ = GetAttr<float>("epsilon", attrs);
momentum_ = GetAttr<float>("momentum", attrs);
is_test_ = GetAttr<bool>("is_test", attrs);
}
Tensor *Bias() const { return bias_; }
const int &Axis() const { return axis_; }
const Tensor *Input() const { return input_; }
const Tensor *Filter() const { return filter_; }
Tensor *Output() const { return output_; }
const vector<int> &Strides() const { return strides_; }
const vector<int> &Paddings() const { return paddings_; }
const vector<int> &Dilations() const { return dilations_; }
const int &Groups() const { return groups; }
const Tensor *InputBias() const { return input_bias_; }
const Tensor *InputMean() const { return input_mean_; }
const Tensor *InputScale() const { return input_scale_; }
const Tensor *InputVariance() const { return input_variance_; }
const float &Epsilon() const { return epsilon_; }
const float &Momentum() const { return momentum_; }
const bool &IsTest() const { return is_test_; }
void SetNewScale(Tensor *new_scale) { new_scale_ = new_scale; }
void SetNewBias(Tensor *new_bias) { new_bias_ = new_bias; }
const Tensor *NewScale() const { return new_scale_; }
const Tensor *NewBias() const { return new_bias_; }
protected:
Tensor *bias_;
int axis_;
Tensor *input_;
Tensor *output_;
Tensor *filter_;
vector<int> strides_;
vector<int> paddings_;
vector<int> dilations_;
int groups;
Tensor *input_bias_;
Tensor *input_mean_;
Tensor *input_scale_;
Tensor *input_variance_;
float epsilon_;
float momentum_;
bool is_test_;
Tensor *new_bias_;
Tensor *new_scale_;
};
Print &operator<<(Print &printer, const FusionConvAddParam &conv_param);
#endif
#ifdef IM2SEQUENCE_OP
class Im2SequenceParam : public OpParam {
public:
Im2SequenceParam(const VariableNameMap &inputs,
const VariableNameMap &outputs, const AttributeMap &attrs,
const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
kernels_ = GetAttr<vector<int>>("kernels", attrs);
strides_ = GetAttr<vector<int>>("strides", attrs);
paddings_ = GetAttr<vector<int>>("paddings", attrs);
}
const Tensor *Input() const { return input_x_; }
Tensor *Output() const { return out_; }
const vector<int> &Kernels() const { return kernels_; }
const vector<int> &Strides() const { return strides_; }
const vector<int> &Paddings() const { return paddings_; }
private:
Tensor *input_x_;
Tensor *out_;
vector<int> kernels_;
vector<int> strides_;
vector<int> paddings_;
};
#endif
#ifdef DROPOUT_OP
class DropoutParam : public OpParam {
public:
DropoutParam(const VariableNameMap &inputs, const VariableNameMap &outputs,
const AttributeMap &attrs, const Scope &scope) {
input_x_ = InputXFrom<LoDTensor>(inputs, scope);
out_ = OutFrom<LoDTensor>(outputs, scope);
}
const Tensor *InputX() const { return input_x_; }
Tensor *Out() const { return out_; }
private:
Tensor *input_x_;
Tensor *out_;
};
#endif
} // namespace operators
} // namespace paddle_mobile
src/operators/pool_op.cpp
浏览文件 @ c67b701f
......@@ -60,11 +60,9 @@ template class PoolOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(pool2d);
REGISTER_OPERATOR_CPU(pool2d, ops::PoolOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(pool2d);
REGISTER_OPERATOR_MALI_GPU(pool2d, ops::PoolOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/pool_op.h
浏览文件 @ c67b701f
......@@ -48,4 +48,13 @@ class PoolOp : public OperatorWithKernel<DeviceType, PoolParam,
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(pool2d);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(pool2d);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/prior_box_op.cpp
浏览文件 @ c67b701f
......@@ -50,7 +50,6 @@ template class PriorBoxOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(prior_box);
REGISTER_OPERATOR_CPU(prior_box, ops::PriorBoxOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......
src/operators/prior_box_op.h
浏览文件 @ c67b701f
......@@ -51,4 +51,12 @@ class PriorBoxOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(prior_box);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/relu_op.cpp
浏览文件 @ c67b701f
......@@ -34,11 +34,9 @@ template class ReluOp<CPU, float>;
* */
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(relu);
REGISTER_OPERATOR_CPU(relu, ops::ReluOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(relu);
REGISTER_OPERATOR_MALI_GPU(relu, ops::ReluOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/relu_op.h
浏览文件 @ c67b701f
......@@ -53,4 +53,13 @@ class ReluOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(relu);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(relu);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/reshape_op.cpp
浏览文件 @ c67b701f
......@@ -33,11 +33,9 @@ template class ReshapeOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(reshape);
REGISTER_OPERATOR_CPU(reshape, ops::ReshapeOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(reshape);
REGISTER_OPERATOR_MALI_GPU(reshape, ops::ReshapeOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/reshape_op.h
浏览文件 @ c67b701f
......@@ -51,4 +51,14 @@ class ReshapeOp
} // namespace operators
} // namespace paddle_mobile
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(reshape);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(reshape);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/sigmoid_op.cpp
浏览文件 @ c67b701f
......@@ -28,7 +28,6 @@ template class SigmoidOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(sigmoid);
REGISTER_OPERATOR_CPU(sigmoid, ops::SigmoidOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......
src/operators/sigmoid_op.h
浏览文件 @ c67b701f
......@@ -46,4 +46,12 @@ class SigmoidOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(sigmoid);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/softmax_op.cpp
浏览文件 @ c67b701f
......@@ -28,11 +28,9 @@ template class SoftmaxOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(softmax);
REGISTER_OPERATOR_CPU(softmax, ops::SoftmaxOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(softmax);
REGISTER_OPERATOR_MALI_GPU(softmax, ops::SoftmaxOp);
#endif
#ifdef PADDLE_MOBILE_FPGA
......
src/operators/softmax_op.h
浏览文件 @ c67b701f
......@@ -48,4 +48,13 @@ class SoftmaxOp
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(softmax);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
USE_OP_MALI_GPU(softmax);
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
src/operators/transpose_op.cpp
浏览文件 @ c67b701f
......@@ -53,7 +53,6 @@ template class TransposeOp<CPU, float>;
namespace ops = paddle_mobile::operators;
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(transpose);
REGISTER_OPERATOR_CPU(transpose, ops::TransposeOp);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
......
src/operators/transpose_op.h
浏览文件 @ c67b701f
......@@ -50,4 +50,12 @@ class TransposeOp : public framework::OperatorWithKernel<
} // namespace operators
} // namespace paddle_mobile
#ifdef PADDLE_MOBILE_CPU
USE_OP_CPU(transpose);
#endif
#ifdef PADDLE_MOBILE_MALI_GPU
#endif
#ifdef PADDLE_MOBILE_FPGA
#endif
#endif
test/CMakeLists.txt
浏览文件 @ c67b701f
set(dir ${CMAKE_CURRENT_SOURCE_DIR})
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY "${dir}/build")
if (googlenet)
if (NET STREQUAL "googlenet")
# gen test
ADD_EXECUTABLE(test-googlenet net/test_googlenet.cpp test_helper.h test_include.h executor_for_test.h)
target_link_libraries(test-googlenet paddle-mobile)
elseif (mobilenet)
elseif (NET STREQUAL "mobilenet")
# gen test
ADD_EXECUTABLE(test-mobilenet net/test_mobilenet.cpp test_helper.h test_include.h executor_for_test.h)
target_link_libraries(test-mobilenet paddle-mobile)
elseif (yolo)
elseif (NET STREQUAL "yolo")
# gen test
ADD_EXECUTABLE(test-yolo net/test_yolo.cpp test_helper.h test_include.h executor_for_test.h)
target_link_libraries(test-yolo paddle-mobile)
elseif (squeezenet)
elseif (NET STREQUAL "squeezenet")
# gen test
ADD_EXECUTABLE(test-squeezenet net/test_squeezenet.cpp test_helper.h test_include.h executor_for_test.h)
target_link_libraries(test-squeezenet paddle-mobile)
elseif(resnet)
elseif(NET STREQUAL "resnet")
# gen test
ADD_EXECUTABLE(test-resnet net/test_resnet.cpp test_helper.h test_include.h executor_for_test.h)
target_link_libraries(test-resnet paddle-mobile)
......
test/executor_for_test.h
浏览文件 @ c67b701f
......@@ -19,7 +19,7 @@ limitations under the License. */
#include "common/log.h"
#include "framework/op_registry.h"
#include "io/io.h"
#include "io/executor.h"
#include "operators/conv_op.h"
#include "operators/elementwise_add_op.h"
#include "operators/pool_op.h"
......
test/framework/test_load.cpp
浏览文件 @ c67b701f
......@@ -13,13 +13,13 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "../test_helper.h"
#include "io/io.h"
#include "io/loader.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
// ../../../test/models/googlenet
// ../../../test/models/mobilenet
auto program = loader.Load(g_mobilenet_ssd, false, false);
auto program = loader.Load(g_googlenet, true);
// auto program = loader.Load(g_googlenet_combine + "/model",
// g_googlenet_combine +
// "/params", true);
......
test/framework/test_optimize.cpp
浏览文件 @ c67b701f
......@@ -15,7 +15,7 @@ limitations under the License. */
#include "../test_helper.h"
#include "framework/program/program-optimize/node.h"
#include "framework/program/program-optimize/program_optimize.h"
#include "io/io.h"
#include "io/loader.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
......
test/net/test_mobilenet+ssd.cpp
浏览文件 @ c67b701f
......@@ -17,23 +17,23 @@ limitations under the License. */
#include "../test_include.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
auto time1 = time();
auto program = loader.Load(g_mobilenet_ssd, true);
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, true);
if (paddle_mobile.Load(g_mobilenet_ssd, true)) {
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
std::vector<int64_t> dims{1, 3, 300, 300};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 300, 300}, static_cast<float>(0),
static_cast<float>(1));
std::vector<int64_t> dims{1, 3, 300, 300};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 300, 300}, static_cast<float>(0),
static_cast<float>(1));
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
executor.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
paddle_mobile.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
}
return 0;
}
test/net/test_mobilenet.cpp
浏览文件 @ c67b701f
......@@ -17,28 +17,25 @@ limitations under the License. */
#include "../test_include.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
auto time1 = time();
auto program = loader.Load(g_mobilenet, true);
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, true);
std::vector<int64_t> dims{1, 3, 224, 224};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 224, 224}, static_cast<float>(0),
static_cast<float>(1));
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
auto vec_result = executor.Predict(input, dims);
float sum = 0;
for (const auto item : vec_result) {
sum += item;
if (paddle_mobile.Load(g_mobilenet, true)) {
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
std::vector<int64_t> dims{1, 3, 224, 224};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 224, 224}, static_cast<float>(0),
static_cast<float>(1));
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
auto vec_result = paddle_mobile.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
}
DLOG << "mobilenet output sum =" << sum;
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
return 0;
}
test/net/test_resnet.cpp
浏览文件 @ c67b701f
......@@ -17,23 +17,23 @@ limitations under the License. */
#include "../test_include.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
auto time1 = time();
auto program = loader.Load(g_resnet, false);
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, false);
if (paddle_mobile.Load(g_resnet, false)) {
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
std::vector<int64_t> dims{1, 3, 32, 32};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 32, 32}, static_cast<float>(0),
static_cast<float>(1));
std::vector<int64_t> dims{1, 3, 32, 32};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 32, 32}, static_cast<float>(0),
static_cast<float>(1));
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
paddle_mobile.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
}
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
executor.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
return 0;
}
test/net/test_squeezenet.cpp
浏览文件 @ c67b701f
......@@ -17,25 +17,25 @@ limitations under the License. */
#include "../test_include.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
// ../../../test/models/googlenet
// ../../../test/models/mobilenet
auto time1 = time();
auto program = loader.Load(g_squeezenet, false);
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, false);
if (paddle_mobile.Load(g_squeezenet, false)) {
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
std::vector<int64_t> dims{1, 3, 227, 227};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 227, 227}, static_cast<float>(0),
static_cast<float>(1));
std::vector<int64_t> dims{1, 3, 227, 227};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 227, 227}, static_cast<float>(0),
static_cast<float>(1));
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
paddle_mobile.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
}
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
executor.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
return 0;
}
test/net/test_yolo.cpp
浏览文件 @ c67b701f
......@@ -17,25 +17,25 @@ limitations under the License. */
#include "../test_include.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
// ../../../test/models/googlenet
// ../../../test/models/mobilenet
auto time1 = time();
auto program = loader.Load(g_yolo, false);
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
paddle_mobile::Executor<paddle_mobile::CPU> executor(program, 1, false);
if (paddle_mobile.Load(g_yolo, false)) {
auto time2 = time();
DLOG << "load cost :" << time_diff(time1, time1) << "ms";
std::vector<int64_t> dims{1, 3, 227, 227};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 227, 227}, static_cast<float>(0),
static_cast<float>(1));
std::vector<int64_t> dims{1, 3, 227, 227};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 227, 227}, static_cast<float>(0),
static_cast<float>(1));
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
executor.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
paddle_mobile.Predict(input, dims);
auto time4 = time();
DLOG << "predict cost :" << time_diff(time3, time4) << "ms";
}
return 0;
}
test/operators/test_im2sequence_op.cpp 0 → 100644
浏览文件 @ c67b701f
/* 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 "../executor_for_test.h"
#include "../test_include.h"
#include "operators/im2sequence_op.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
auto program = loader.Load(g_ocr_recg);
PADDLE_MOBILE_ENFORCE(program.originProgram != nullptr,
"program file read fail");
Executor4Test<paddle_mobile::CPU,
paddle_mobile::operators::ReluOp<paddle_mobile::CPU, float>>
executor(program, "im2sequence");
// 1. input_tensors;
vector<Tensor> input_tensors;
Tensor input1;
auto input1_data = CreateInput<float>(&input1, {2, 2, 3, 3}, -1, 1);
input_tensors.push_back(input1);
// 2. input_names
vector<string> input_names({
"conv2d_19.tmp_1",
});
// 3. output_names
vector<string> output_names({"im2sequence_0.tmp_0"});
// 4. out_dims;
vector<DDim> out_ddims;
auto out_ddim = paddle_mobile::framework::make_ddim({8, 9});
out_ddims.push_back(out_ddim);
auto output = executor.Predict<LoDTensor>(input_tensors, input_names,
output_names, out_ddims);
auto output0_data = output[0]->data<float>();
for (int j = 0; j < input_tensors[0].numel(); ++j) {
DLOG << " value of input: " << input1_data[j];
}
for (int j = 0; j < output[0]->numel(); ++j) {
DLOG << " value of output: " << output0_data[j];
}
return 0;
}
test/operators/test_sigmoid_op.cpp
浏览文件 @ c67b701f
......@@ -14,7 +14,7 @@ limitations under the License. */
#include "../../src/operators/kernel/sigmoid_kernel.h"
#include "../test_helper.h"
#include "io/io.h"
#include "io/executor.h"
int main() {
paddle_mobile::framework::Tensor input;
......
test/test_helper.h
浏览文件 @ c67b701f
......@@ -14,10 +14,10 @@ limitations under the License. */
#pragma once
#include <chrono>
#include <fstream>
#include <random>
#include "common/common.h"
#include "common/log.h"
#include "framework/ddim.h"
#include "framework/tensor.h"
......@@ -35,17 +35,6 @@ static const std::string g_test_image_1x3x224x224 =
using paddle_mobile::framework::DDim;
using paddle_mobile::framework::Tensor;
using Time = decltype(std::chrono::high_resolution_clock::now());
Time time() { return std::chrono::high_resolution_clock::now(); }
double time_diff(Time t1, Time t2) {
typedef std::chrono::microseconds ms;
auto diff = t2 - t1;
ms counter = std::chrono::duration_cast<ms>(diff);
return counter.count() / 1000.0;
}
template <typename T>
void SetupTensor(paddle_mobile::framework::Tensor *input,
paddle_mobile::framework::DDim dims, T lower, T upper) {
......
test/test_include.h
浏览文件 @ c67b701f
......@@ -30,4 +30,4 @@ limitations under the License. */
#include "framework/scope.h"
#include "framework/tensor.h"
#include "framework/variable.h"
#include "io/io.h"
#include "io/paddle_mobile.h"
tools/android-cmake/android.toolchain.cmake
浏览文件 @ c67b701f
......@@ -37,7 +37,7 @@
# ANDROID_DISABLE_FORMAT_STRING_CHECKS
# ANDROID_CCACHE
cmake_minimum_required(VERSION 3.6.0)
# cmake_minimum_required(VERSION 3.6.0)
# Inhibit all of CMake's own NDK handling code.
set(CMAKE_SYSTEM_VERSION 1)
......
tools/arm-platform.cmake 0 → 100644
浏览文件 @ c67b701f
set(ARCH "armv7-a")
set(FLOAT_ABI "softfp" CACHE STRING "-mfloat-api chosen")
set_property(CACHE FLOAT_ABI PROPERTY STRINGS "softfp" "soft" "hard")
set(FPU "neon")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=${ARCH} -mfloat-abi=${FLOAT_ABI} -mfpu=${FPU}")
tools/build.sh
浏览文件 @ c67b701f
......@@ -31,9 +31,9 @@ build_for_mac() {
}
build_for_android() {
rm -rf "../build"
if [ -z "${ANDROID_NDK}" ]; then
echo "ANDROID_NDK not found!"
#rm -rf "../build"
if [ -z "${NDK_ROOT}" ]; then
echo "NDK_ROOT not found!"
exit -1
fi
......@@ -60,7 +60,6 @@ build_for_android() {
TOOLCHAIN_FILE="./tools/android-cmake/android.toolchain.cmake"
ANDROID_ARM_MODE="arm"
if [ $# -eq 1 ]; then
NET=$1
cmake .. \
-B"../build/release/${PLATFORM}" \
-DANDROID_ABI="${ABI}" \
......@@ -70,7 +69,7 @@ build_for_android() {
-DCMAKE_CXX_FLAGS="${CXX_FLAGS}" \
-DANDROID_STL=c++_static \
-DANDROID=true \
-D"${NET}=true" \
-DNET=$1 \
-D"${ARM_PLATFORM}"=true
else
......@@ -90,7 +89,7 @@ build_for_android() {
}
build_for_ios() {
rm -rf "../build"
# rm -rf "../build"
PLATFORM="ios"
MODE="Release"
BUILD_DIR=../build/release/"${PLATFORM}"
......@@ -99,7 +98,6 @@ build_for_ios() {
CXX_FLAGS="-fobjc-abi-version=2 -fobjc-arc -std=gnu++14 -stdlib=libc++ -isysroot ${CMAKE_OSX_SYSROOT}"
mkdir -p "${BUILD_DIR}"
if [ $# -eq 1 ]; then
NET=$1
cmake .. \
-B"${BUILD_DIR}" \
-DCMAKE_BUILD_TYPE="${MODE}" \
......@@ -107,7 +105,7 @@ build_for_ios() {
-DIOS_PLATFORM=OS \
-DCMAKE_C_FLAGS="${C_FLAGS}" \
-DCMAKE_CXX_FLAGS="${CXX_FLAGS}" \
-D"${NET}"=true \
-DNET=$1 \
-DIS_IOS="true"
else
cmake .. \
......@@ -121,6 +119,9 @@ build_for_ios() {
fi
cd "${BUILD_DIR}"
make -j 8
cd ./build
# 生成符号表
ranlib *.a
}
build_error() {
......@@ -129,16 +130,12 @@ build_error() {
if [ $# -lt 1 ]; then
echo "error: target missing!"
echo "available targets: mac|linux|ios|android"
echo "sample usage: ./build.sh mac"
echo "available targets: ios|android"
echo "sample usage: ./build.sh android"
else
if [ $# -eq 2 ]; then
if [ $2 != "googlenet" -a $2 != "mobilenet" -a $2 != "yolo" -a $2 != "squeezenet" -a $2 != "resnet" ]; then
if [ $1 = "mac" ]; then
build_for_mac
elif [ $1 = "linux" ]; then
build_for_linux
elif [ $1 = "android" ]; then
if [ $1 = "android" ]; then
build_for_android
elif [ $1 = "ios" ]; then
build_for_ios
......@@ -146,11 +143,7 @@ else
build_error
fi
else
if [ $1 = "mac" ]; then
build_for_mac $2
elif [ $1 = "linux" ]; then
build_for_linux $2
elif [ $1 = "android" ]; then
if [ $1 = "android" ]; then
build_for_android $2
elif [ $1 = "ios" ]; then
build_for_ios $2
......@@ -159,11 +152,7 @@ else
fi
fi
else
if [ $1 = "mac" ]; then
build_for_mac
elif [ $1 = "linux" ]; then
build_for_linux
elif [ $1 = "android" ]; then
if [ $1 = "android" ]; then
build_for_android
elif [ $1 = "ios" ]; then
build_for_ios
......
tools/op.cmake 0 → 100644
浏览文件 @ c67b701f
if (NET STREQUAL "googlenet")
set(CONCAT_OP ON)
set(CONV_OP ON)
set(LRN_OP ON)
set(MUL_OP ON)
set(ELEMENTWISEADD_OP ON)
set(FUSION_FC_OP ON)
set(POOL_OP ON)
set(RELU_OP ON)
set(FUSION_CONVADD_OP ON)
set(FUSION_CONVADD_RELU_OP ON)
elseif (NET STREQUAL "mobilenet")
set(CONV_OP ON)
set(ELEMENTWISEADD_OP ON)
set(RELU_OP ON)
set(SOFTMAX_OP ON)
set(SOFTMAX_OP ON)
set(DEPTHWISECONV_OP ON)
set(BATCHNORM_OP ON)
set(POOL_OP ON)
set(RESHAPE_OP ON)
set(FUSION_CONVADDBNRELU_OP)
elseif (NET STREQUAL "yolo")
set(BATCHNORM_OP ON)
set(CONV_OP ON)
set(RELU_OP ON)
set(ELEMENTWISEADD_OP ON)
elseif (NET STREQUAL "squeezenet")
set(CONCAT_OP ON)
set(CONV_OP ON)
set(RELU_OP ON)
set(ELEMENTWISEADD_OP ON)
set(POOL_OP ON)
set(RESHAPE_OP ON)
set(SOFTMAX_OP ON)
elseif (NET STREQUAL "resnet")
set(CONV_OP ON)
set(BATCHNORM_OP ON)
set(ELEMENTWISEADD_OP ON)
set(SOFTMAX_OP ON)
set(MUL_OP ON)
set(POOL_OP ON)
set(RELU_OP ON)
else ()
set(BATCHNORM_OP ON)
set(BOXCODER_OP ON)
set(CONCAT_OP ON)
set(CONV_OP ON)
set(DEPTHWISECONV_OP ON)
set(ELEMENTWISEADD_OP ON)
set(FUSION_CONVADD_OP ON)
set(CONVADDRELU_OP ON)
set(FUSION_FC_OP ON)
set(LRN_OP ON)
set(MUL_OP ON)
set(MULTICLASSNMS_OP ON)
set(POOL_OP ON)
set(PRIORBOX_OP ON)
set(RELU_OP ON)
set(RESHAPE_OP ON)
set(SIGMOID_OP ON)
set(SOFTMAX_OP ON)
set(TRANSPOSE_OP ON)
set(FUSION_CONVADD_RELU_OP ON)
set(FUSION_CONVADDBNRELU_OP ON)
set(DROPOUT_OP ON)
set(IM2SEQUENCE_OP ON)
# option(BATCHNORM_OP "" ON)
# option(BOXCODER_OP "" ON)
# option(CONCAT_OP "" ON)
# option(CONV_OP "" ON)
# option(DEPTHWISECONV_OP "" ON)
# option(ELEMENTWISEADD_OP "" ON)
# option(FUSION_CONVADD_OP "" ON)
# option(CONVADDRELU_OP "" ON)
# option(FUSION_FC_OP "" ON)
# option(LRN_OP "" ON)
# option(MUL_OP "" ON)
# option(MULTICLASSNMS_OP "" ON)
# option(POOL_OP "" ON)
# option(PRIORBOX_OP "" ON)
# option(RELU_OP "" ON)
# option(RESHAPE_OP "" ON)
# option(SIGMOID_OP "" ON)
# option(SOFTMAX_OP "" ON)
# option(TRANSPOSE_OP "" ON)
# option(FUSION_CONVADD_RELU_OP "" ON)
endif ()
if (BATCHNORM_OP)
add_definitions(-DBATCHNORM_OP)
endif()
if (BOXCODER_OP)
add_definitions(-DBOXCODER_OP)
endif()
if (CONCAT_OP)
add_definitions(-DCONCAT_OP)
endif()
if (CONV_OP)
add_definitions(-DCONV_OP)
endif()
if (DEPTHWISECONV_OP)
add_definitions(-DDEPTHWISECONV_OP)
endif()
if (ELEMENTWISEADD_OP)
add_definitions(-DELEMENTWISEADD_OP)
endif()
if (FUSION_CONVADD_OP)
add_definitions(-DFUSION_CONVADD_OP)
endif()
if (CONVADDRELU_OP)
add_definitions(-DCONVADDRELU_OP)
endif()
if (FUSION_FC_OP)
add_definitions(-DFUSION_FC_OP)
endif()
if (LRN_OP)
add_definitions(-DLRN_OP)
endif()
if (MUL_OP)
add_definitions(-DMUL_OP)
endif()
if (MULTICLASSNMS_OP)
add_definitions(-DMULTICLASSNMS_OP)
endif()
if (POOL_OP)
add_definitions(-DPOOL_OP)
endif()
if (PRIORBOX_OP)
add_definitions(-DPRIORBOX_OP)
endif()
if (RELU_OP)
add_definitions(-DRELU_OP)
endif()
if (RESHAPE_OP)
add_definitions(-DRESHAPE_OP)
endif()
if (SIGMOID_OP)
add_definitions(-DSIGMOID_OP)
endif()
if (SOFTMAX_OP)
add_definitions(-DSOFTMAX_OP)
endif()
if (TRANSPOSE_OP)
add_definitions(-DTRANSPOSE_OP)
endif()
if (FUSION_CONVADD_RELU_OP)
add_definitions(-DFUSION_CONVADD_RELU_OP)
endif()
if (FUSION_CONVADDBNRELU_OP)
add_definitions(-DFUSION_CONVADDBNRELU_OP)
endif()
if (DROPOUT_OP)
add_definitions(-DDROPOUT_OP)
endif()
if (IM2SEQUENCE_OP)
add_definitions(-DIM2SEQUENCE_OP)
endif()
tools/pre-commit.hooks/clang-format.hook
浏览文件 @ c67b701f
......@@ -14,6 +14,10 @@ fi
# https://medicineyeh.wordpress.com/2017/07/13/clang-format-with-pragma/
shift
perl -i -pe 's|#pragma\s+omp|// <TRICKY-CLANG-FORMAT-PRAGMA-FIX> #pragma omp|' "$@"
clang-format -i $@
perl -i -pe 's|^\s+#pragma\s+omp|// <TRICKY-CLANG-FORMAT-PRAGMA-FIX> #pragma omp|' "$@"
(
# remove clang format ios_io folder
flist=$(echo "$@" | perl -pe 's|src/ios_io/[^ ]*||')
clang-format -i $flist
)
perl -i -pe 's|// <TRICKY-CLANG-FORMAT-PRAGMA-FIX> ||' "$@"
tools/toolchains/arm-android-neon.cmake 0 → 100644
浏览文件 @ c67b701f
set(ANDROID_ARM_NEON ON)
include("${CMAKE_CURRENT_LIST_DIR}/../android-cmake/android.toolchain.cmake")
\ No newline at end of file
tools/toolchains/arm-linux-gnueabi.cmake 0 → 100644
浏览文件 @ c67b701f
# CMake toolchain file for building ARM software on Linux environment
set(CMAKE_SYSTEM_NAME Linux)
set(CMAKE_SYSTEM_VERSION 1)
set(CMAKE_C_COMPILER /usr/bin/arm-linux-gnueabi-gcc)
set(CMAKE_CXX_COMPILER /usr/bin/arm-linux-gnueabi-g++)
set(CMAKE_STRIP /usr/bin/arm-linux-gnueabi-strip)
set(CMAKE_FIND_ROOT_PATH /usr/arm-linux-gnueabi)
set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
set(ARM_LINUX 1)
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