[Mesa-dev] [PATCH] radeonsi: add support for viewport array (v2)

[Dave Airlie]

On 26 June 2015 at 00:26, Marek Olšák <maraeo at gmail.com> wrote:
> Hi Dave,
>
> The change in si_shader_io_get_unique_index can be dropped. The
> function is only used for shaders before GS.
>
Ok okay I was hitting the assert in there for the layer/viewport index cases,
but if the patch you pushed to master helps I'll drop it.

> This looks good, but I've had a different plan for this feature:

Yeah I thought you might, I just wanted to hack something up to see it working
since it seems new Dota 2 uses this feature.

> I'd like the states to be converted into 2 atoms:
>
> 1 r600_atom for all 16 viewports
> 1 r600_atom for all 16 scissors
>
> Each atom should have a bitmask saying which "slots" are dirty. (the same
> idea as resource slots)
>
> The "emit" functions should only emit dirty viewports/scissors.
>
> Also, the "emit" functions shouldn't emit non-zero viewports/scissors
> if the viewport index isn't written by the hardware VS stage
> (si_get_vs_info(sctx)->...). This should keep the same level of
> effectiveness as before. When a shader that writes the viewport index
> is bound *and* there are any dirty viewports or scissors, that's the
> right time to mark the atoms as dirty again, so that non-zero dirty
> viewports/scissors are finally emitted.

I'm not sure if I'll get to doing more of it, I just had a day at home,
and wanted to get up to speed on radeonsi a bit, so I can just leave
that here, in case anyone wants to pick it up,

Dave.