[Mesa-dev] Change in Mesa triggers bug in Firefox Nightly with WebRender on old AMD hardware

[Dave Airlie]

On Thu, 2 May 2019 at 05:36, Viktor Jaegerskuepper
<viktor_jaegerskuepper at freenet.de> wrote:
>
> Dave Airlie:
> > On Wed, 1 May 2019 at 12:34, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> >>
> >> On Tue, Apr 30, 2019 at 10:42 AM Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> >>>
> >>> On Tue, Apr 30, 2019 at 2:48 AM Dave Airlie <airlied at gmail.com> wrote:
> >>>>
> >>>> On Tue, 30 Apr 2019 at 08:47, Dave Airlie <airlied at gmail.com> wrote:
> >>>>>
> >>>>> On Tue, 30 Apr 2019 at 08:42, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> >>>>>>
> >>>>>> Thing is, I think that would not be enough - with the "recent" (like
> >>>>>> past 2 years) CSO/state change detection changes, I think that you can
> >>>>>> end up with no sampler set for a buffer view. Perhaps someone with the
> >>>>>> hw can investigate what goes wrong?
> >>>>>
> >>>>> I'm trying to get a live r600/700 machine, I plugged in my rv635 and
> >>>>> rv740 gpus, and they both seemed toast, I've got an rv635 laptop I'll
> >>>>> try and get up to date.
> >>>>
> >>>> I've found a working rv770 and reproduced it here. This isn't an
> >>>> indication I'll get to fix it soon, but I will try.
> >>>
> >>> When I was debugging this issue on nouveau (and trying to nail down
> >>> specifically what didn't work and what was required to make it work),
> >>> I found that using a gallium test to debug this made the issues at
> >>> hand quite obvious, and I didn't have to worry about how to force the
> >>> st to assign mappings here or there for views or samplers. I'm not
> >>> sure if I still have those left over, but I'll have a look tonight.
> >>> They're also not terribly difficult to write.
> >>
> >> In case it's helpful:
> >> https://people.freedesktop.org/~imirkin/tbo-tex.c which fits together
> >> with the other gallium/tests/trivial ones.
> >>
> >> I think there were associated Makefile.am changes, but with the
> >> removal of autoconf, I probably lost the local changes.
> >>
> >> In its current form, it doesn't create a sampler, but you can see how
> >> trivial/quad-tex.c does it if you want to play around with having
> >> samplers around. You might also have to use PIPE_FORMAT_R8G8B8A8_UNORM
> >> instead of the BGRA variant for the view's format (iirc r600 has some
> >> limitations around swizzling and buffer textures).
> >>
> >
> > Thanks for test, but I ended up installing firefox and watching the
> > pain, then found it by accident.
> >
> > https://patchwork.freedesktop.org/patch/302675/
> >
> > For Viktor to test.
>
> Tested with RV770, the patch fixes the bug for Firefox Beta and Nightly.
>
> So was this really a bug in the old r600 code or is the patch a
> workaround for a Firefox/WebRender bug?

It's a bug in the old r600/rv770 specific codepaths that was uncovered
by the mesa state tracker change, but also relied on specific
webrender behaviour patterns to trigger, but it was a driver bug none
the less.

Dave.