[Mesa-dev] radeonsi: NIR - Polaris triangle sprinkling running UH SOLVED - finally

[Timothy Arceri]

On 13/2/19 8:28 am, Dieter Nützel wrote:
> Hello Marek, Timo, Nicolai,
> 
> Timo SOLVED this long-standing NIR corruption on Polaris with his 'nir: 
> rewrite varying component packing' commit.
> 
> It was triggered with
> 
> commit 86b52d42368ac496fe24bc6674e754c323381635
> Author: Marek Olšák <marek.olsak at amd.com>
> Date:   Fri Jul 13 00:23:36 2018 -0400
> 
>      radeonsi: reduce LDS stalls by 40% for tessellation
> 
>      40% is the decrease in the LGKM counter (which includes SMEM too)
>      for the GFX9 LSHS stage.
> 
>      This will make the LDS size slightly larger, but I wasn't able to 
> increase
>      the patch stride without corruption, so I'm increasing the vertex 
> stride.
> 
> and now finally SOLVED with
> 
> commit 26aa460940f6222565ad5eb40a21c2377c59c3a6
> Author: Timothy Arceri <tarceri at itsqueeze.com>
> Date:   Mon Dec 10 10:23:51 2018 +1100
> 
>      nir: rewrite varying component packing
> 
>      There are a number of reasons for the rewrite.
> 
>      1. Adding support for packing tess patch varyings in a sane way.
> 
>      2. Making use of qsort allowing the code to be much easier to
>         follow.
> 
>      3. Fixes a bug where different interp types caused component
>         packing to be skipped for all varyings in some scenarios.
> 
>      4. Allows us to add a crude live range analysis for deciding
>         which components should be packed together. This support can
>         optionally be added in a future patch.
> 
>      Reviewed-by: Jason Ekstrand <jason at jlekstrand.net>
> 
> Maybe it should backported (Cc: <mesa-stable at lists.freedesktop.org>) 
> ) for 19.0?

I'd rather it didn't since NIR isn't default for radeonsi and this is 
use for other drivers like RADV. I'd rather have more testing in the dev 
branch.

This change really shouldn't fix anything, it could just be that the 
change avoids the bug. I'm really not sure.

One thing this change did fix was a bug with doubles, but I wasn't aware 
of anything actually using doubles so I doubt this is it.

> 
> I hope my bisect help to bring some more understanding for this Polaris 
> NIR bug.
> 
> Now, hunting for the (last) 19.0+ EQAA regression (DiRT Rally, black 
> squares like  radv/DXVK corruption, NOT NIR related) and 'meson' OpenCL 
> (Clover) build error.
> 
> Greetings,
> Dieter