A Huawei engineer posted to the Mesa mailing list today over potentially upstreaming a Vulkan driver they have developed for an in-house GPU. However, at least in current form, this new in-house GPU Vulkan driver is likely dead in the water from being upstreamed into Mesa.

While it's interesting from a technical perspective that Huawei has developed a Vulkan driver using the Mesa Vulkan Runtime for an in-house GPU, it's a half-assed open-source attempt as they note that the shader compiler cannot be open-source. The Huawei engineer reports that their GPU ISA cannot be published and thus in turn they cannot publish their NIR-to-machine-code back-end for converting that common Mesa intermediate representation (IR) into the Huawei GPU machine code.

So the inquiry on the Mesa mailing list today was whether they could upstream their Huawei Vulkan driver while keeping that NIR-to-machine-code compiler as a run-time-loadable binary blob in Mesa. But that's likely to never fly for upstream Mesa given that the driver is non-functional without this binary blob.

To the mailing list post, Red Hat's David Airlie noted it would be very unlikely to upstream and Huawei should consider just opening up their GPU ISA. Especially in the era of AI/LLMs, that blob could be easily reverse engineered.

We'll see if Huawei decides to produce a fully open-source Vulkan driver for Mesa otherwise it's likely to remain a downstream, closed-source blob. The engineer did not elaborate on the in-house Huawei GPU design. Huawei currently doesn't produce any gaming or general purpose graphics cards but they do have the Ascend and Atlas AI accelerators.