regression in ttm

[Michel Dänzer]

On 2018-12-05 12:41 p.m., Koenig, Christian wrote:
> Am 05.12.18 um 12:33 schrieb Michel Dänzer:
>> On 2018-12-05 12:24 p.m., Koenig, Christian wrote:
>>> Am 05.12.18 um 11:29 schrieb Michel Dänzer:
>>>> On 2018-12-04 6:35 p.m., Koenig, Christian wrote:
>>>>> Thanks, going to take a look tomorrow.
>>>> I also hit this with Bonaire in my development system. I wonder why you
>>>> didn't? How are you running piglit?
>>> With more memory. The issue is in TTM and only happens when you start to
>>> evict something.
>> How much memory do you have, how many CPU cores, and how many piglit
>> tests running in parallel? 16G (1G VRAM) / 8/16 (SMT) / 16 here.
>>
>> Which piglit profile are you using? gpu here (with the
>> glx-multithread-texture test excluded, because that hangs due to a Mesa
>> bug with so many CPU cores).
>>
>> Are you running piglit with/out --process-isolation false? With it here.
> 
> I usually run piglit as last thing in the evening on my Vega/Ryzen 
> system. That box has 16GB VRAM, 32GB system.
> 
> No idea what the piglit settings actually are since I'm using the Ubuntu 
> package to update it. My script just runs "piglit run 
> /usr/lib/x86_64-linux-gnu/piglit/tests/gpu.py ~/Temp/".

Maybe you can try adding --process-isolation false. If nothing else, it
should reduce the time needed for the piglit run significantly.

Other than that, I guess the amount of VRAM is the most significant
difference, though I'd expect some of the piglit tests to exhaust that
anyway.


> I have to confess that I rarely look out for regressions in the actual 
> result, but rather look out for lockups or KASAN reports from the kernel 
> in dmesg.

That's fine, if there's nothing in dmesg, it's usually not the kernel's
fault. :)


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