[Mesa-dev] GL4.5 or bust...

[Ilia Mirkin]

On Sun, May 15, 2016 at 4:32 PM, Dave Airlie <airlied at gmail.com> wrote:
> So I said this on irc over the weekend and it seemed like we had some
> consensus on holding off 12.0 until we could announce 4.5 on some
> hardware. This assumes the FP64 stuff is going in at least.
>
> So I decided to roll out the proposal here, which is that we finish
> GL4.5 features off for at least Skylake I think.
>
> So what is needed/missing: please add as you see fit.
>
> a) robustness - radeonsi has some bits of this. We need to get
> KHR_robustness bits, that I think Kayden has patches started for, and
> i965 needs to ensure it uses robust buffer stuff. I don't think this
> one in unobtainable.
>
> b) cull_distance - I merged something, it broke, I'll fix it today, job done.
>
> c) enhanced_layouts - So tarceri has posted patches, we know that to
> do it properly we probably need to rip up attribute packing and
> rewrite it, however if Kayden thinks what tarceri has done is
> functional enough for now, we could merge the final pieces and work on
> perfection later.
>
> d) SIMD32 for i965 compute shaders - this is probably the most unknown
> to me, curro says he's got some patches, that need to rebase onto FP64
> when it lands, assuming he can do that, and reviewers can get on top
> of things, and we possibly only enable SIMD32 in the corner cases
> initially, it might be possible to get this landed.
>
> Have I missed anything? Should we go for it?

The bugs that get triggered when you expose GL 4.3+ to UE4 games. Some
are ours, some are theirs. Someone needs to sign up for this work.

Also, I'd like to mention that ES 3.2 is pretty close as well. But
probably not close enough to squeeze in here. Ian has started working
on the OES_shader_io_blocks bits of it (which IMO shouldn't be too bad
for someone who knows what all GLSL allows and what it doesn't), which
was the last remaining big chunk. I have preliminary patches for core
support of advanced blending, the rest should all be easy.

>
> For radeonsi, I think the only other missing bit is qbo and
> clear_texture, which may or may not make it in time.

I'm in favor of this plan. Nouveau should be ready for Fermi and
Kepler once Samuel's images patches for Fermi land (mostly reviewed,
had a couple of nits). Maxwell will be missing tess and images, and
it's unlikely that either of those will get done in a reasonable
period of time. I think we can just flip robustness on... probably not
meeting all the provisions of that spec, but ... meh.

That said, we should put a cap on this timewise - if e.g. it becomes
clear that SIMD32 will take a long time (I think the biggest potential
issue of the batch), we should just cut a release. Maybe a 1 month
cap?

  -ilia