# PaddleLite使用百度XPU预测部署 Paddle Lite已支持百度XPU在x86和arm服务器(例如飞腾 FT-2000+/64)上进行预测部署。 目前支持Kernel和子图两种接入方式,其中子图接入方式与之前华为NPU类似,即加载并分析Paddle模型,将Paddle算子转成XTCL组网API进行网络构建,在线生成并执行模型。 ## 支持现状 ### 已支持的芯片 - 昆仑818-100(推理芯片) - 昆仑818-300(训练芯片) ### 已支持的设备 - K100/K200昆仑AI加速卡 ### 已支持的Paddle模型 - [ResNet50](https://paddlelite-demo.bj.bcebos.com/models/resnet50_fp32_224_fluid.tar.gz) - [BERT](https://paddlelite-demo.bj.bcebos.com/models/bert_fp32_fluid.tar.gz) - [ERNIE](https://paddlelite-demo.bj.bcebos.com/models/ernie_fp32_fluid.tar.gz) - YOLOv3 - Mask R-CNN - Faster R-CNN - UNet - SENet - SSD - 百度内部业务模型(由于涉密,不方便透露具体细节) ### 已支持(或部分支持)的Paddle算子(Kernel接入方式) - scale - relu - tanh - sigmoid - stack - matmul - pool2d - slice - lookup_table - elementwise_add - elementwise_sub - cast - batch_norm - mul - layer_norm - softmax - conv2d - io_copy - io_copy_once - __xpu__fc - __xpu__multi_encoder - __xpu__resnet50 - __xpu__embedding_with_eltwise_add ### 已支持(或部分支持)的Paddle算子(子图/XTCL接入方式) - relu - tanh - conv2d - depthwise_conv2d - elementwise_add - pool2d - softmax - mul - batch_norm - stack - gather - scale - lookup_table - slice - transpose - transpose2 - reshape - reshape2 - layer_norm - gelu - dropout - matmul - cast - yolo_box ## 参考示例演示 ### 测试设备(K100昆仑AI加速卡) ![baidu_xpu](https://paddlelite-demo.bj.bcebos.com/devices/baidu/baidu_xpu.jpg) ### 准备设备环境 - K100/200昆仑AI加速卡[规格说明书](https://paddlelite-demo.bj.bcebos.com/devices/baidu/K100_K200_spec.pdf),如需更详细的规格说明书或购买产品,请联系欧阳剑ouyangjian@baidu.com; - K100为全长半高PCI-E卡,K200为全长全高PCI-E卡,要求使用PCI-E x16插槽,且需要单独的8针供电线进行供电; - 安装K100/K200驱动,目前支持Ubuntu和CentOS系统,由于驱动依赖Linux kernel版本,请正确安装对应版本的驱动安装包。 ### 准备本地编译环境 - 为了保证编译环境一致,建议参考[源码编译](../user_guides/source_compile)中的Linux开发环境进行配置; - 由于编译示例程序需要依赖OpenCV和CMake 3.10.3,请执行如下命令进行安装; ```shell $ sudo apt-get update $ sudo apt-get install gcc g++ make wget unzip libopencv-dev pkg-config $ wget https://www.cmake.org/files/v3.10/cmake-3.10.3.tar.gz $ tar -zxvf cmake-3.10.3.tar.gz $ cd cmake-3.10.3 $ ./configure $ make $ sudo make install ``` ### 运行图像分类示例程序 - 从[https://paddlelite-demo.bj.bcebos.com/devices/baidu/PaddleLite-linux-demo.tar.gz](https://paddlelite-demo.bj.bcebos.com/devices/baidu/PaddleLite-linux-demo.tar.gz)下载示例程序,解压后清单如下: ```shell - PaddleLite-linux-demo - image_classification_demo - assets - images - tabby_cat.jpg # 测试图片 - labels - synset_words.txt # 1000分类label文件 - models - resnet50_fp32_224_fluid # Paddle fluid non-combined格式的resnet50 float32模型 - __model__ # Paddle fluid模型组网文件,可拖入https://lutzroeder.github.io/netron/进行可视化显示网络结构 - bn2a_branch1_mean # Paddle fluid模型参数文件 - bn2a_branch1_scale ... - shell - CMakeLists.txt # 示例程序CMake脚本 - build - image_classification_demo # 已编译好的,适用于amd64的示例程序 - image_classification_demo.cc # 示例程序源码 - build.sh # 示例程序编译脚本 - run.sh # 示例程序运行脚本 - libs - PaddleLite - amd64 - include # PaddleLite头文件 - lib - libiomp5.so # Intel OpenMP库 - libmklml_intel.so # Intel MKL库 - libxpuapi.so # XPU API库,提供设备管理和算子实现。 - llibxpurt.so # XPU runtime库 - libpaddle_full_api_shared.so # 预编译PaddleLite full api库 - arm64 - include # PaddleLite头文件 - lib - libxpuapi.so # XPU API库,提供设备管理和算子实现。 - llibxpurt.so # XPU runtime库 - libpaddle_full_api_shared.so # 预编译PaddleLite full api库 ``` - 进入PaddleLite-linux-demo/image_classification_demo/shell,直接执行./run.sh amd64即可; ```shell $ cd PaddleLite-linux-demo/image_classification_demo/shell $ ./run.sh amd64 # 默认已生成amd64版本的build/image_classification_demo,因此,无需重新编译示例程序就可以执行。 $ ./run.sh arm64 # 需要在arm64(FT-2000+/64)服务器上执行./build.sh arm64后才能执行该命令。 ... AUTOTUNE:(12758016, 16, 1, 2048, 7, 7, 512, 1, 1, 1, 1, 0, 0, 0) = 1by1_bsp(1, 32, 128, 128) Find Best Result in 150 choices, avg-conv-op-time = 40 us [INFO][XPUAPI][/home/qa_work/xpu_workspace/xpu_build_dailyjob/api_root/baidu/xpu/api/src/wrapper/conv.cpp:274] Start Tuning: (12758016, 16, 1, 512, 7, 7, 512, 3, 3, 1, 1, 1, 1, 0) AUTOTUNE:(12758016, 16, 1, 512, 7, 7, 512, 3, 3, 1, 1, 1, 1, 0) = wpinned_bsp(1, 171, 16, 128) Find Best Result in 144 choices, avg-conv-op-time = 79 us I0502 22:34:18.176113 15876 io_copy_compute.cc:75] xpu to host, copy size 4000 I0502 22:34:18.176406 15876 io_copy_compute.cc:36] host to xpu, copy size 602112 I0502 22:34:18.176697 15876 io_copy_compute.cc:75] xpu to host, copy size 4000 iter 0 cost: 2.116000 ms I0502 22:34:18.178530 15876 io_copy_compute.cc:36] host to xpu, copy size 602112 I0502 22:34:18.178792 15876 io_copy_compute.cc:75] xpu to host, copy size 4000 iter 1 cost: 2.101000 ms I0502 22:34:18.180634 15876 io_copy_compute.cc:36] host to xpu, copy size 602112 I0502 22:34:18.180881 15876 io_copy_compute.cc:75] xpu to host, copy size 4000 iter 2 cost: 2.089000 ms I0502 22:34:18.182726 15876 io_copy_compute.cc:36] host to xpu, copy size 602112 I0502 22:34:18.182976 15876 io_copy_compute.cc:75] xpu to host, copy size 4000 iter 3 cost: 2.085000 ms I0502 22:34:18.184814 15876 io_copy_compute.cc:36] host to xpu, copy size 602112 I0502 22:34:18.185068 15876 io_copy_compute.cc:75] xpu to host, copy size 4000 iter 4 cost: 2.101000 ms warmup: 1 repeat: 5, average: 2.098400 ms, max: 2.116000 ms, min: 2.085000 ms results: 3 Top0 tabby, tabby cat - 0.689418 Top1 tiger cat - 0.190557 Top2 Egyptian cat - 0.112354 Preprocess time: 1.553000 ms Prediction time: 2.098400 ms Postprocess time: 0.081000 ms ``` - 如果需要更改测试图片,可将图片拷贝到PaddleLite-linux-demo/image_classification_demo/assets/images目录下,然后将run.sh的IMAGE_NAME设置成指定文件名即可; - 如果需要重新编译示例程序,直接运行./build.sh amd64或./build.sh arm64即可。 ```shell $ cd PaddleLite-linux-demo/image_classification_demo/shell $ ./build.sh amd64 # For amd64 $ ./build.sh arm64 # For arm64(FT-2000+/64) ``` ### 更新模型 - 通过Paddle Fluid训练,或X2Paddle转换得到ResNet50 float32模型[resnet50_fp32_224_fluid](https://paddlelite-demo.bj.bcebos.com/models/resnet50_fp32_224_fluid.tar.gz); - 由于XPU一般部署在Server端,因此将使用PaddleLite的full api加载原始的Paddle Fluid模型进行预测,即采用CXXConfig配置相关参数。 ### 更新支持百度XPU的Paddle Lite库 - 下载PaddleLite源码; ```shell $ git clone https://github.com/PaddlePaddle/Paddle-Lite.git $ cd Paddle-Lite $ git checkout ``` - 下载xpu_toolchain for amd64 or arm64(FT-2000+/64); ```shell $ wget $ tar -xvf output.tar.gz $ mv output xpu_toolchain ``` - 编译full_publish for amd64 or arm64(FT-2000+/64); ```shell For amd64,如果报找不到cxx11::符号的编译错误,请将gcc切换到4.8版本。 $ ./lite/tools/build.sh --build_xpu=ON --xpu_sdk_root=./xpu_toolchain x86 For arm64(FT-2000+/64) $ ./lite/tools/build.sh --arm_os=armlinux --arm_abi=armv8 --arm_lang=gcc --build_extra=ON --build_xpu=ON --xpu_sdk_root=./xpu_toolchain --with_log=ON full_publish ``` - 将编译生成的build.lite.x86/inference_lite_lib/cxx/include替换PaddleLite-linux-demo/libs/PaddleLite/amd64/include目录; - 将编译生成的build.lite.x86/inference_lite_lib/cxx/include/lib/libpaddle_full_api_shared.so替换PaddleLite-linux-demo/libs/PaddleLite/amd64/lib/libpaddle_full_api_shared.so文件; - 将编译生成的build.lite.armlinux.armv8.gcc/inference_lite_lib.armlinux.armv8.xpu/cxx/include替换PaddleLite-linux-demo/libs/PaddleLite/arm64/include目录; - 将编译生成的build.lite.armlinux.armv8.gcc/inference_lite_lib.armlinux.armv8.xpu/cxx/lib/libpaddle_full_api_shared.so替换PaddleLite-linux-demo/libs/PaddleLite/arm64/lib/libpaddle_full_api_shared.so文件。 ## 其它说明 - 如需更进一步的了解相关产品的信息,请联系欧阳剑ouyangjian@baidu.com; - 百度昆仑的研发同学正在持续适配更多的Paddle算子,以便支持更多的Paddle模型。