On Thu, Aug 20, 2015 at 3:01 PM, Dave Airlie <airl...@gmail.com> wrote: >> Shader-db results for vec4 on i965: > > reads like a Johnny Cash cover of NIN.
Agreed, so let me explain why we're enabling it now. The merge window closed yesterday for Mesa 11.0. vec4 NIR isn't quite where we'd like it to be in terms of code quality, but we want to avoid vec4 NIR getting no external testing for a whole release cycle, while still having to maintain both the NIR and non-NIR paths. I've been mostly against enabling NIR for the vec4 backend until its helped:hurt programs ratio dramatically improves, but I think that leaving it disabled for another three months is probably worse. From what I've found, the largest code-quality regression is from the combination vec4 backend's register coalescing optimization being inadequate (a known problem regardless of NIR) and NIR generating more instances of code that register coalescing cannot optimize. I'm working on a new vec4 register coalescing pass, and I think that there are some simple improvements we can make to improve the code the backend gets from NIR. I'm planning to have these things in the 11.0 release. If worse comes to worse, we can disable vec4 NIR before the release. _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
