AI Tool: DRP-AI TVM

DRP-AI TVM^*1 is a tool that generates runtime executables for AI from trained AI models^*2 for Renesas' RZ/V Series MPUs.

We provide a complete development environment on GitHub, and the tutorial on the top page of GitHub includes a series of steps from environment construction to inference execution on a Renesas evaluation board. You can easily evaluate your AI model by following the steps in this tutorial.

This tool consists of the DRP-AI Translator embedded in the framework of the open-source deep learning compiler Apache TVM^*3. Apache TVM technology allows DRP-AI to operate on the layers of AI models that can be converted by the DRP-AI Translator, and the layers that cannot be converted can be assigned to the CPU. This makes it possible to support more AI models than ever before.

Features

DRP-AI TVM has the features listed at the end.

• Automatic generation of runtimes that enable DRP-AI and the CPU to work together
• Support for multiple AI frameworks (ONNX and PyTorch will be supported in v1.0 in September 2022, and Tensorflow in v1.1 in March 2023)
• Additional pre-processing/post-processing equivalent to DRP-AI Translator is possible (to be supported in v1.1 in March 2023)

图像

Configuration

The system configuration of DRP-AI TVM is shown in the figure below, where the DRP-AI Translator is integrated into the Apache TVM framework to achieve significantly higher performance across a wide range of AI models.

图像

^※1 DRP-AI TVM is powered by the Edgecortix MERA^TM Compiler Framework
^※2 Tensorflow will be supported in v1.1 (to be released in April 2023)
^※3 Please refer to https://tvm.apache.org for more information on Apache TVM