A path to upstream AI/ML accelerators

Currently, to support AI/ML hardware accelerators, SoC vendors develop their own drivers, reimplementing a lot of features common to many hardware accelerators. libAPU aims to provide a stack generic enough to support many AI/ML hardware accelerators:

libAPU relies on DRM to manage memory and schedule requests. It uses RPMsg to communicate with the hardware accelerator. It uses remoteproc to power up the hardware accelerator and load the firmware.

Alexandre Bailon will present the libAPU, talks about what has been already implemented, what remains to do and the upcoming challenges.

Alexandre Bailon, Baylibre