[Mesa-dev] soft fp64 on evergreen

Timothy Arceri tarceri at itsqueeze.com
Fri May 26 10:31:48 UTC 2017


On 26/05/17 20:16, Thomas Helland wrote:
> 2017-05-26 8:13 GMT+02:00 Dave Airlie <airlied at gmail.com>:
>> Hi Elie,
>>
>> Thanks for the fp64 work.
>>
>> I've pulled the latest branch from your repo and start running piglit
>> on a BARTS (r600g) GPU.
>>
>> I've noticed a few asserts with a debug build in ir_validate, do you
>> test with --enable-debug?
>>
>> I also notice some really large times in the GLSL->TGSI pass copy
>> propogation on mul/div tests with matrices, tests take a fair while to
>> complete.
>>
> 
> Unrelated to this I have noticed on some profiles when using nvc0
> that the GLSL->TGSI passes are quite painful on compile times.
> Especially copy propagation shows up as a big one.
> Not sure if there is some easy gains to be had here as I haven't
> looked at the pass at all.

Yeah it sucks. Copy prop is cleaning up the things that were lazily 
created by the glsl -> tgsi pass, and then dead_code is slow because its 
cleaning up after cop prop.

I made an attempt to speed it up but it seems the best way to fix it 
would be to avoid generating all the temps in the first place. There are 
code comments (TODOS?) about fixing this, but it seems those were copied 
from the mesa ir code years ago. It wasn't clear to me from my quick 
look how hard it would be to improve the glsl -> tgsi pass.

> 
>> Just wondering if you've seen anything like this?
>>
>> Dave.
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> 


More information about the mesa-dev mailing list