Intel VPU Driver Adapted To Linux's New Accelerator Framework
Linux 6.2 is set to introduce the new "accel" accelerator subsystem and framework intended to be the new common area in the kernel for AI accelerators and similar devices. As outlined in prior articles, this accelerator subsystem/framework re-uses common elements from the Direct Rendering Manager (DRM) subsystem given all that is in common between today's accelerators and graphics processors. DMA-BUF and plenty of DRM interfaces fit in well into the accelerator scope and so this new accel area should work out nicely. It also works out better than tossing some drivers into "char/misc" as was one of the areas for accelerator drivers to this point.
The Intel Versatile Processing Unit (not to be confused with Video Processing Unit) is being introduced with Meteor Lake CPUs.
With the accel subsystem queued for introduction in Linux 6.2, Intel's Habana Labs driver is in the process of being adapted from char/misc to the new framework -- after all, Habana Labs kernel driver maintainer Oded Gabbay was instrumental in bringing together this new subsystem.
Now Intel's VPU driver for Meteor Lake has also been adapted to jive with the new accel code. The Intel VPU v4 patch series posted today has switched over to using the accel framework as well as making various other low-level code improvements.
The Versatile Processing Unit with Meteor Lake is a CPU-integrated inference accelerator to be used for Computer Vision (CV) and deep learning software.
The Intel VPU driver is still undergoing review and isn't expected to be merged until post-6.2 at some point likely in 2023. Those interested can see the v4 patches for the other latest changes to this new accelerator block for 14th Gen Core CPUs.