12:45 <tolszak> tomeu: Hi, following happened on 5.14 Odroid N2+: https://pastebin.com/eiZhyPyE, should I submit it or perhaps it is fixed in later releases?

13:06 <tomeu> tolszak: so it looks like the client closed their GPU device FD while some jobs were still in flight?

13:07 <tolszak> tomeu: Possible, but still userspace shouldn't crash kernel right?

13:09 <tolszak> tomeu: Yeah I confirm that it happened when the example program I created in other task (https://gitlab.freedesktop.org/mesa/mesa/-/issues/5605) was not closed cleanly

13:12 <tomeu> sure, it shouldn't happen, just trying to understand in what conditions this happened

13:28 <robmur01> looks like the same old thing that was fixed in 5.15

13:40 <tolszak> robmur01: Ok, thanks. Currently I don't have easy way to test it with 5.15 but will be back with the bug if it happens there too.

13:43 <robmur01> the key commit should be 7fdc48cc63a3 ("drm/panfrost: Make sure MMU context lifetime is not bound to panfrost_priv") - that fixes the worst of the raciness

13:52 <AreaScout> alyssa, thx for the MR/patch :), the blend errors are gone now, however I experience problems with shaders now, ChromeOS https://pastebin.com/raw/fyBAJKwH and PPSSPP https://pastebin.com/raw/XNmY4Gib can I somehow easily get around of this errors ?

15:07 <tolszak> robmur01: tomeu: Regarding this issue: https://gitlab.freedesktop.org/mesa/mesa/-/issues/5605 . I was able to determine that sampler works OK, the issue happens also without texture, for some reason when scale is > 15.170372 e.g. it starts with 15.170373. So when scale is > 15.170372 the glitch happens when it is <= 15.170372 it doesn't.

15:08 <tolszak> PAN_MESA_DEBUG=trace doesn't change between those 2 different scale values.

15:10 <tolszak> If I interpolate texture coords instead of changing vertex coords then it works alright. Seems like the issue is somewhere after output of vertex shader and before input to fragment shader

15:10 <tolszak> tomeu: robmur01: Any hint where to look more?

15:26 <robmur01> sorry, I'm just kernel guy; the actual graphics side of things is completely beyond me :)

15:36 <tolszak> robmur01: Thanks and sorry for bothering.

15:56 <tomeu> tolszak: hmm, that's already quite a bit

15:56 <tomeu> so scaling is something quite high level, can you reduce that to uniform values or something else at the driver or gpu level?

15:58 <macc24> imageine not knowing a bit about everything

15:58 <tolszak> I mean the scale changes vertex coord

15:59 <tolszak> tomeu: That's the scale I have in mind https://github.com/tolszak/gbm_es2_demo/blob/odroid_issue/demo/gbm_es2_demo.cpp#L310

15:59 <tolszak> tomeu: So effectively it just expands the rectangle by changing its coords

16:00 <tolszak> tomeu: Not really unless one can point me to some docs how to do that.

16:01 <tolszak> tomeu: ^ that was answer to "can you reduce that to uniform values or something else at the driver or gpu level?"

16:02 <macc24> that's a lot of pings

16:04 <tolszak> I'm sorry, it is not like I use IRC on regular basis. But I think I understand, thanks for tip.

16:04 <tomeu> tolszak: ok, so when you say scale, it's just vertex coords

16:04 <tolszak> yes it is only vertex coords

16:04 <tolszak> that changes

16:04 <tolszak> nothing else

16:04 <tomeu> would you be able to test your app under the DDK?

16:05 <macc24> tomeu: wouldn't it be easier to test with llvmpipe?

16:05 <tomeu> well, I was wondering if the DDK did something different

16:05 <macc24> hmm

16:05 <tomeu> my understanding is that it had been tested with llvmpipe, and the problem didn't show up

16:06 <macc24> well you never know

16:06 <tolszak> do I need to use odroid hardkernel image for that?

16:06 <tomeu> guess that would be easiest

16:06 <tomeu> but I don't have direct experience

16:06 <tolszak> this particular example can't be tested with softpipe as it uses some kind of gbm extension that doesn't appear on softpipe

16:07 <macc24> softpipe ≠ llvmpipe

16:07 <tolszak> I see

16:07 <tolszak> How can I use llvm

16:07 <tolszak> ?

16:07 <tolszak> I think I have it built

16:13 <tolszak> I will check it with llvmpipe, building it now, it was disabled in yocto by default

22:52 <alyssa> AreaScout_: Guess my fix broke more things. Thanks for testing 😅

23:02 <alyssa> Ah, just a little typo in the first patch I think. Will fix tomorrow (not at my Mesa computer right now)