Habana Labs Posts Initial Thunk Library To Go With Their New Linux Kernel Driver
Back in January the startup Habana Labs posted an open-source Linux kernel driver for their Goya AI processor. That AI accelerator focused on speeding up deep learning inference workloads better than CPUs and GPUs will now see this new driver mainlined with Linux 5.1 while a user-space thunk driver has now been published.
The Habana Labs Thunk Library (HL-Thunk) is their basic open-source user-space library to interface with this kernel driver. At this stage it just stresses the minimal functionality of the exposed kernel driver IOCTLs and is able to run some basic user-space tests.
Habana Labs will be integrating this basic user-space library into their production software stack and use it as a layer rather than interfacing with their kernel IOCTLs directly. Oded Gabbay who has been leading this Linux driver effort commented that "this will be an integral part of our software stack going forward...I sincerely hope we can open source additional software parts in the future. I'll try pushing to this goal as much as I can inside the company."
This is great to hear with sadly a lot of AI / deep learning accelerator software being closed up. With now at least having this open-source user-space thunk library available, it eases the concerns about this driver going upstream but without any open-source user-space "clients" -- with generally upstream kernel developers frowning upon efforts to mainline kernel drivers that require closed-source user-space to function.
Details on HL-Thunk via the kernel mailing list while the code is on GitHub