[Mesa-dev] GfxBench & CSDof failures

[Eero Tamminen]

Hi,

On 28.03.2018 13:27, Eero Tamminen wrote:
> Mesa built from following (last evening) commit:
> ------------------------------------
> commit 76dfed8ae2d5c6c509eb2661389be3c6a25077df
> Author:     Rob Clark <robdclark at gmail.com>
> AuthorDate: Thu Mar 15 18:42:44 2018 -0400
> Commit:     Rob Clark <robdclark at gmail.com>
> CommitDate: Tue Mar 27 08:36:37 2018 -0400
> 
>      nir: mako all the intrinsics
> ------------------------------------
> 
> Fails with following GL 4.3 benchmarks I tested:
> * GfxBench Manhattan 3.1, CarChase and AztecRuins
> * SynMark CSDof
> 
> Issues are following:
> * CSDof & ActecRuins: compile takes so long that test is killed

Sorry, I should have read more of my mesa-dev 2k mail backlog,
not just check bugs.

It was the memory usage issue mentioned by Ian & Clayton,
not compile time issue.


> * Manhattan 3.1 shader linking failed:
> ---------------------------------
> error: declarations for uniform `depth_parameters` are inside block 
> `cameraConsts` and outside a block
> ---------------------------------

And this was also using all memory, so I assume issue was due to Mesa
shader compiler handling memory allocation failure badly (continuing
to a bogus compile error instead of reporting compile failure due to 
allocation failure).


> Commit from one day earlier, didn't have any problems.
> 
> Neither does latest Mesa git version.
> 
> Which commit fixed the regression, is it this one:

I assume it was:
---------------------------------
commit 5f21a7afe072f8a6e558ccc47407a0a94e0d1313
Author:     Jason Ekstrand <jason.ekstrand at intel.com>
AuthorDate: Tue Mar 27 16:12:16 2018 -0700
Commit:     Jason Ekstrand <jason.ekstrand at intel.com>
CommitDate: Tue Mar 27 18:18:26 2018 -0700

     nir/intrinsics: Don't report negative dest_components
---------------------------------


	- Eero