Linux's New Compute Accelerator Framework Quickly Taking Shape
For years there has long been the Linux kernel debate over an "accel" subsystem with some drivers currently being tossed into the catch-all "char/misc" area of the kernel while the open-source graphics driver developers have wanted these drivers to go in via the DRM subsystem given the commonality with GPU drivers. What's now finally reached consensus and coming about with the work-in-progress patch series is building a compute accelerator framework/subsystem built atop the existing DRM infrastructure.
Oded Gabbay who led the work on the Habana Labs AI kernel driver has been the one leading the development of the compute accelerator framework. Besides working on a "dummy" driver, there are also plans for Intel-owned Habana Labs to eventually transition their AI driver from the char/misc area to this new framework. There are numerous other AI drivers relevant to this new framework like the Intel Meteor Lake Versatile Processing Unit, Toshiba DNN Visconti, NVIDIA NVDLA, Samsung Trinity NPU, Intel GNA, Qualcomm QAIC, and others.
Given the interest in this framework and it holding up the various other in-progress drivers, two weeks since the original patch series we are now up to the third iteration. The compute accelerator subsystem v3 patches address comments raised during earlier review, make various code changes based on testing thus far, some of the Intel VPU driver engineers have been looking at this framework, and other changes made.
Those interested in this forthcoming Linux compute accelerator subsystem can see the v3 patches on the kernel mailing list. Given the different drivers and vendors depending upon this framework, it's likely this long-awaited subsystem will premiere sooner rather than later.