[Mesa-dev] [PATCH 0/5] Enable up to 24 MRF registers in gen6

[Iago Toral]

On Mon, 2015-09-21 at 09:21 -0700, Matt Turner wrote:
> On Mon, Sep 21, 2015 at 8:00 AM, Iago Toral <itoral at igalia.com> wrote:
> > On Mon, 2015-09-21 at 07:49 -0700, Kenneth Graunke wrote:
> >> On Monday, September 21, 2015 09:46:24 AM Mark Janes wrote:
> >> > This series hits an assertion on ILK and G45:
> >> >
> >> > src/mesa/drivers/dri/i965/brw_eu_emit.c:150: brw_set_dest: Assertion
> >> > `dest.nr < (devinfo->gen == 6 ? 24 : 16)' failed.
> >> >
> >> > It triggers about 8k piglit assertions on those platforms.  I'm turning
> >> > off testing for G45 and ILK until it is resolved.
> >> >
> >> > https://bugs.freedesktop.org/show_bug.cgi?id=92066
> >> >
> >> > -Mark
> >>
> >> I've pushed a fix for this:
> >>
> >> commit c1070550c289d48ef389aeb8c564d1abd1123ad1
> >> Author: Kenneth Graunke <kenneth at whitecape.org>
> >> Date:   Mon Sep 21 07:42:27 2015 -0700
> >>
> >>     i965: Fix MRF register number assertions for compr4.
> >>
> >>     compr4 is represented by setting the high bit on the MRF number.
> >>     We need to mask it out before sanity checking the register number.
> >>
> >>     Fixes ~8000 assert fails on Ironlake and G45.
> >>
> >>     Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92066
> >>     Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
> >>
> >> Easy mistake...I always forget about compr4.  Hopefully we should be
> >> good now.
> >
> > that was fast, thanks Ken! I was scratching my head over this...
> >
> > BTW, I just noticed that the ILK docs also say that they have 24 MRFs...
> > (volume 4, part 2, 5.3.3 MRF Registers). Assuming that we don't find any
> > other issues, would we want to extend the fix to ILK too?
> 
> The ILK docs are notoriously bad and often contain more information
> about Sandybridge than Ironlake. I suspect that information is
> actually about SNB, though I suppose it couldn't hurt to try on ILK,
> though I'm doubtful.

For the record: I tested this and it does not seem to work in ILK, so
you're probably right.

In the process, however, I noticed that forcing spilling on the vec4
backend with current master in ILK breaks a lot of tests and also
produces at least one GPU hang, so there is something seriously broken
there. I filed this bug:
https://bugs.freedesktop.org/show_bug.cgi?id=92100

I'll try to see what is going on with that...

Iago