Intel Habana Labs AI Driver Prepares For Move To Linux's New Accelerator Subsystem

With the in-development Linux 6.2 kernel there is the new compute accelerator "accel" subsystem/framework as part of the Direct Rendering Manager area. The creation of that new subsystem was pulled together by Intel (Habana Labs) and now their AI accelerator driver is preparing to move from the "char/misc" catch-all to this new subsystem.

The lead developer of the "habanalabs" accelerator driver is Oded Gabbay who also led the creation of the accel subsystem/framework. Going back years there has been discussions over having a hardware accelerator subsystem for the growing number of devices that aren't quite GPUs in the traditional sense and thus not necessarily fitting entirely within the DRM area but also "char/misc" continues to proliferate as the catch-all area for random drivers. With more open-source AI kernel drivers continuing to be volleyed, finally over the past several months this compute accelerator subsystem has come together. It also took some design agreements between upstream stakeholders with originally there being some differences whether it should be a subordinate of the DRM area and its interfaces. But as AI accelerators have a lot in common with GPUs and there already being a lot of common DRM code, it has made sense to pull in the accelerator framework as part of the DRM code.

With Linux 6.2 the compute accelerator subsystem/framework is being introduced while for Linux 6.3 will likely be the first in-tree drivers. Intel's in-development Versatile Processing Unit "VPU" driver for Meteor Lake has been adapted to the new framework and now the Habana Labs driver has patches out for moving homes from char/misc to the accel area.

Oded Gabbay today posted the patches moving the user-space API headers and then the driver itself from the "drivers/misc" directory over to "drivers/accel".

With this move, the files and changes to the build system are taken care of. However, future patches are still to come for properly registering into the accelerator subsystem as well as exposing the new /dev/accel* character files instead of the existing Habana Labs device char files. Given the timing, presumably this Habana Labs driver move will be squared away for the Linux 6.3 merge window come February.

It's good seeing this initial activity while there is also other open-source AI driver work for the Toshiba DNN Visconti, NVIDIA NVDLA, Samsung Trinity NPU, Intel GNA, Qualcomm QAIC, and a growing number of other devices/drivers that will likely target this subsystem. Intel / Habana Labs can be certainly praised for their continued work on being the leading AI accelerator hardware backed by open-source user-space and kernel-space software while being good upstream stewards to the kernel.