The recommended development flow for Vitis™ AI is illustrated as the following figure. Vitis AI and Vitis IDE are needed for this flow which has three basic steps:
Figure 1. Vitis AI Flow
- The Vitis AI toolchain in the host machine is used to build the model. It takes the pre-trained floating models as the input and runs them through the AI Optimizer (optional).
- A custom hardware platform is built using the Vitis software platform based on the Vitis Target Platform. The generated hardware includes the DPU IP and other kernels. In the Vitis AI release package, pre-built SD card images (for ZCU102/104, KV260 and VCK190) and Alveo™ shells are included for quick start and application development. You can also use the Vivado® Design Suite to integrate the DPU and build the custom hardware to suit your need. For more information, see Integrating the DPU into Custom Platforms.
- You can build executable software which runs on the built hardware. You can write your applications with C++ or Python which calls the Vitis AI Runtime and Vitis AI Library to load and run the compiled model files.