On Mon, Nov 16, 2026 at 7:58 PM Joseph Telaak <[email protected]> wrote:
>
> On Mon, Nov 16, 2026 at 3:26 PM Patrik Jakobsson wrote:
> > On Mon, Nov 16, 2026 at 3:09 PM Gabriel Gama wrote:
> > > If I'm not mistaken, Joseph removed his kernel RFCs (This is in his
> > > GitHub repo)
> >
> > He offered to work together with you? Two drivers for the same
> > hardware is a waste of time and resources for everybody. Lets keep it
> > simple.
>
> Gabriel is right, I've withdrawn my RFCs.
>
> Alyssa Rosenzweig closed my Mesa MR [1]. Part of it was that I'd posted
> an early draft that was far from ready, but the core of it was the AI
> use itself: the proprietary SGX driver's source leaked widely enough
> that it's likely in the models' training data, so there's no way to
> show where LLM-written code for this GPU really came from. I wasn't
> aware of the leak until she pointed it out.

Thanks for the clarification. ImgTec have released an open source
driver and documentation for later versions of their chips. Perhaps a
way forward is to ask them to release the docs for SGX5xx?

>
> My driver is staying out of tree. I'm leaving it up, though, since
> several people are using it now, and fairly successfully.
>
> I do agree with you that the SGX cores belong together in one driver.
> powervr_sgx was built that way: a table of cores, with the SGX545
> in it and an SGX530 (TI OMAP3630) on the way. Unfortunately it's not
> going to happen with my driver.
>
> > > For Poulsbo I'll stick with the GMA500 drm, it makes much more sense;
> > > Once I finish reverse-engineering the SGX544MP2, I plan to develop a
> > > separate DRM driver for it.
> >
> > Again, why don't you join Joseph on this? Mesa is not going to accept
> > 3 different drivers either if the hardware is similar.
>
> Gabriel, on the AI question you raised earlier: from what you describe
> (most of reMESA generated with AI, built partly on earlier RE work that
> also used AI), I think you'll have a hard time with Mesa as well,
> whatever the PROVENANCE.md says. Check their rules on reverse
> engineering the vendor's binaries (xpsb-glx, psb_dri.so) too, before
> putting much more into the Mesa side. Here are some of the links Alyssa
> gave me:
>
> https://kusma.xyz/blog/2026/03/26/open-source-and-ai/
> https://asahilinux.org/llm-policy/
> https://nouveau.freedesktop.org/Humanity.html
> https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/docs/drivers/panfrost.rst#ai-policy
> https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/imagination/README.md#contribution-policy
>
> [1] 
> https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/46286#note_3783031
>
> Joseph

Reply via email to