Connected Autonomous Vehicles

Abstract

Presentation of our EPOCHS application on collaborative autonomous vehicles and the use of our HPVM compiler and smart scheduler to execute this application on an heterogenous SoC with many accelerators. This work is being presented at the ISCA 2024 tutorial on Agile Software-Hardware Co-Design of AI-Centric Heterogeneous SoCs (EPOCHS). The EPOCHS project under the DARPA DSSoC program, a collaboration between IBM Research, UIUC, Harvard, and Columbia University. The project focused on building agile software and hardware methodologies for domain-specific SoCs. The chosen application was collaborative autonomous vehicles (CAVs). Our EPOCHS SoC had 9 different accelerators with RISC-V Ariane cores. This particular tutorial details different aspects of our software-stack and CAV application development. We detail the development of a real-time scheduler, HetSched, for highly heterogeneous SoCs. It also describes the development of a queue-based discrete event scheduling platform, STOMP, that allows you to develop plug-and-play scheduling policies. Based on the hierarchical scheduler, HetSched developed in STOMP, we built a scheduler-library. In collaboration with the HPVM team at UIUC, we created a backend for the scheduler-library. This scheduler-integrated HPVM compiler allows application developers to be agnostic of the scheduler-APIs and HW accelerator interfaces, yet accelerate their application on the DSSoC at hand. We demonstrated our software stack for the ERA CAV application on an FPGA showcased in a short video https://www.youtube.com/watch?v=NXzZOX5Ukiw&themeRefresh=1.