ARM 芯片的 Python + AI 算力优化 主讲人： 朱宏林 – 阿里云程序语言与编译器团队 简介 • 当今开发者们大量使用 Python 语言编写的 AI 程序。过去这些程序总跑在 GPU 或者 x86 架构的 CPU 上。然而综合考虑到功耗、成本、性能等因素，云厂商们开始建设 ARM 架构的服务平台，如 何整合 Python + AI 的相关软件并使其在该平台上发挥最高的性能成为了工程师们关注的焦点。 提升明显。目前，该成果已经被集成进 OpenBLAS 和 PyTorch 中。 • 本次演讲，将向大家介绍我们在倚天 710 ARM 芯片上开展的 Python + AI 优化工作，以及在 ARM 云平台上部署 Python + AI 任务的最佳实践。 深度学习 • 广泛使用的深度学习框架 • TensorFlow、PyTorch • 结合硬件（ARM 服务端芯片） • 倚天 710 •

Zynq® Systems on Chips (SoCs). • Using the Python language and libraries, designers can exploit the benefits of programmable logic and microprocessors in Zynq to build more capable and exciting embedded 26 Ø 目前赛灵思的AI开发工具支持 python 语言 输入，可以做 AI inference 的数据前处理 和后处理的工作。这部分工作可以在SOC 的ARM 里面实现。 Ø Xilinx AI Developer Resource Ø https://www.xilinx.com/products/ design-tools/ai-inference/ai- developer-hub Xilinx Edge AI Tutorial Ø https://github.com/Xilinx/Edge-AI- Platform-Tutorials Ø Xilinx ModelZoo Ø https://www.xilinx.com/bin/public /openDownload?filename=all_mo dels_20190528.zip Ø Xilinx AI Forums

Zynq® Systems on Chips (SoCs). • Using the Python language and libraries, designers can exploit the benefits of programmable logic and microprocessors in Zynq to build more capable and exciting embedded 26 Ø 目前赛灵思的AI开发工具支持 python 语言 输入，可以做 AI inference 的数据前处理和 后处理的工作。这部分工作可以在SOC的 ARM 里面实现。 Ø Xilinx AI Developer Resource Ø https://www.xilinx.com/products/desig n-tools/ai-inference/ai-developer- hub Xilinx Edge AI Tutorial Ø https://github.com/Xilinx/Edge-AI- Platform-Tutorials Ø Xilinx ModelZoo Ø https://www.xilinx.com/bin/public/ope nDownload?filename=all_models_2019 0528.zip Ø Xilinx AI Forums

Zynq® Systems on Chips (SoCs). • Using the Python language and libraries, designers can exploit the benefits of programmable logic and microprocessors in Zynq to build more capable and exciting embedded 26 ➢ 目前赛灵思的AI开发工具支持 python 语言 输入，可以做 AI inference 的数据前处理和 后处理的工作。这部分工作可以在SOC的 ARM 里面实现。 ➢ Xilinx AI Developer Resource ➢ https://www.xilinx.com/products/desig n-tools/ai-inference/ai-developer- hub Xilinx Edge AI Tutorial ➢ https://github.com/Xilinx/Edge-AI- Platform-Tutorials ➢ Xilinx ModelZoo ➢ https://www.xilinx.com/bin/public/ope nDownload?filename=all_models_2019 0528.zip ➢ Xilinx AI Forums

not a good idea to put any of this Python code within your Web server’s document root, because it risks the possibility that people may be able to view your code over the Web. That’s not good for security argument-order bugs – and you can reorder the arguments in your views’ function definitions. Of course, these benefits come at the cost of brevity; some developers find the named-group syntax ugly and too verbose. stored on disk and in memory, and how to customize the default behavior. Warning: There are security risks if you are accepting uploaded content from untrusted users! See the security guide’s topic on User-uploaded

not a good idea to put any of this Python code within your Web server’s document root, because it risks the possibility that people may be able to view your code over the Web. That’s not good for security argument-order bugs – and you can reorder the arguments in your views’ function definitions. Of course, these benefits come at the cost of brevity; some developers find the named-group syntax ugly and too verbose. stored on disk and in memory, and how to customize the default behavior. Warning There are security risks if you are accepting uploaded content from untrusted users! See the security guide’s topic on User-uploaded

not a good idea to put any of this Python code within your web server’s document root, because it risks the possibility that people may be able to view your code over the web. That’s not good for security async context. However, you will need to use an async server based on ASGI to get their performance benefits. Here’s an example of an async view: import datetime from django.http import HttpResponse async stored on disk and in memory, and how to customize the default behavior. Warning: There are security risks if you are accepting uploaded content from untrusted users! See the security guide’s topic on User-uploaded

not a good idea to put any of this Python code within your Web server’s document root, because it risks the possibility that people may be able to view your code over the Web. That’s not good for security async context. However, you will need to use an async server based on ASGI to get their performance benefits. Here’s an example of an async view: import datetime from django.http import HttpResponse async stored on disk and in memory, and how to customize the default behavior. Warning: There are security risks if you are accepting uploaded content from untrusted users! See the security guide’s topic on User-uploaded

not a good idea to put any of this Python code within your Web server’s document root, because it risks the possibility that people may be able to view your code over the Web. That’s not good for security async context. However, you will need to use an async server based on ASGI to get their performance benefits. Here’s an example of an async view: import datetime from django.http import HttpResponse async stored on disk and in memory, and how to customize the default behavior. Warning: There are security risks if you are accepting uploaded content from untrusted users! See the security guide’s topic on User-uploaded

not a good idea to put any of this Python code within your web server’s document root, because it risks the possibility that people may be able to view your code over the web. That’s not good for security async context. However, you will need to use an async server based on ASGI to get their performance benefits. Here’s an example of an async view: import datetime from django.http import HttpResponse async stored on disk and in memory, and how to customize the default behavior. Warning There are security risks if you are accepting uploaded content from untrusted users! See the security guide’s topic on User-uploaded