[Mesa-dev] [PATCH v9 0/7] mesa/st: glsl_to_tgsi: refined register merge algorithm

[Michel Dänzer]

On 25/08/17 05:30 PM, Gert Wollny wrote:
> Am Freitag, den 25.08.2017, 10:20 +0900 schrieb Michel Dänzer:
>> On 25/08/17 02:38 AM, Gert Wollny wrote:
>>>
>>> The patch doesn't introduce piglit regression (I tested the shader
>>> subset). 
>>
>> I'd recommend testing at least the gpu profile, ideally running on X.
> 
> I've done that now running the gpu profile with clean upstream
> (f623e1742f) and my patches on top (twice each time). The only
> consistent change is fixing  
> 
>    vs-output-array-vec2-index-wr-before-gs 
> 
> A bit strange is spec/arb_sync/repeate_wait because it is skipped once
> with clean upstream but passes otherwise.
> 
> spec/nv_conditional_render failed once with my patch applied, but it
> doesn't do any register merging and when I run the test from the
> command line it always passes (also when explicitly cleaning the shader
> cache). I didn't try running it with a clean upstream more then twice
> though. 
> 
> All other changes versus a clean upstream are non-deterministic, i.e.
> fail and pass once for each code version. 
> 
> In detail these are 
> 
> glx/
>   - glx_arb_sync_control (all timing-* tests)
> 
>   - security
>       initialized-fbo       (warn/pass)
>       initialized-texmemory (warn/pass), 
> 
> spec/
>   - arb_tessellation_shader
>      tcs-input-read-array-interface
>      tcs-input-read-mat
>      tcs-input-read-nonconst-interface-builtin

Looks fine, thanks. Most of these tests are known to give non-stable
results with radeonsi as well.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer