12:31 <Wizzup> enunes: following up, I found a way to reproduce the problem in apitrace, at least some of what I was seeing (I took a simpler case), and interestingly enough when replayed llvmpipe it looks fine, but not when I replay it on amdgpu, so perhaps the GL application is buggy, or there is some shared mesa bug

12:31 <Wizzup> I'm going bet on the former

12:31 <Wizzup> fwiw the traces are here and I also did a dump-images session: https://wizzup.org/dirlist/maemo-leste/lima/

12:31 <Wizzup> the trace is just me drag-selecting an empty terminal window

12:32 <Wizzup> of course the two traces aren't identical in the sense that I had to manually swipe and do some stuff, so the frame numbers per frame are different for dump images (I picked some at random)

12:42 <Wizzup> I do see gl errors in the lima trace that I do not see in the llvmpipe trace, even when replaying in amdgpu or llvmpipe

12:42 <enunes> Wizzup: hmm I played it but I'm still not sure what am I looking at

12:43 <enunes> best case would be if a single trace reproduces the bug when played on the lima device and does not reproduce it on e.g. intel

12:43 <Wizzup> hm, ok

12:44 <Wizzup> So the trace itself just shows swiping with a finger up and down on a terminal window, the black selection should follow the (to the replay) invisible finger

12:44 <Wizzup> this pattern is clearly incorrect: https://wizzup.org/dirlist/maemo-leste/lima/dump-images/llvmpipe-trace-amdgpu-redraw0000003542.png

12:45 <Wizzup> at this point I suspect there is some corner case in the application rather than mesa that makes it work with llvmpipe and some binary driver(s), but not with various mesa drivers

13:10 <Wizzup> enunes: btw, what about replaying maemo-summoner-hildon-desktop-llvmpipe.trace on lima?

13:11 <Wizzup> i.e. 'apitrace replay maemo-summoner-hildon-desktop-llvmpipe.trace' vs 'LIBGL_ALWAYS_SOFTWARE=1 apitrace replay maemo-summoner-hildon-desktop-llvmpipe.trace' shows the same problems for me on the pinephone

13:12 <Wizzup> (I tried to run apitrace on my powervr device where this works fine, but apitrace segfaulted, so I will have to try that another day.)

13:13 <Wizzup> in any case, don't break your head over it now (but if can confirm you're seeing the same as me would be nice), I'll try to see if I can find a bug in clutter 0.8's texture from pixmap code somehow

19:20 <Wizzup> I'm going to try to investigate more in the next few days but I found that both llvmpipe and nvidia render both traces correctly it looks like, but other mesa based renderers do not (amdgpu, lima)

19:22 <Wizzup> maybe it's some ambiguity in the spec, but it is suspect that other (non-mesa) implementations seem to mostly just work

19:22 <Wizzup> (in case it wasn't clear I added some images here https://wizzup.org/dirlist/maemo-leste/lima/dump-images/)

19:26 <Wizzup> frame 23 seems to be where the first visual difference are - glDrawArrays has wrong contents, glUniformMatrix4fv still has the same texture and framebuffer

19:26 <Wizzup> at least in the 'lima' trace: https://wizzup.org/check.png

19:26 <Wizzup> (one of those qapitrace instances is using llvmpipe)

23:22 <anarsoul> Wizzup: looks like an app bug to me if you get the same result with lima and amdgpu

23:22 <Wizzup> anarsoul: yes, bit llvmpipe, nvidia and powervr are all correct

23:22 <Wizzup> s/bit/but/

23:23 <Wizzup> still looking into it, was just a bit surprising

23:23 <Wizzup> (thoughts very welcome though)

23:23 <anarsoul> it could be hitting some undefined behavior that gets you the result you expect with llvmpipe

23:24 <Wizzup> yeah

23:28 <Wizzup> Just for completion sake I am not sure if I get the exact same with lima as I am with amdgpu but I'll verify that later, right now it's easier for me (testing wise) to focus on amdgpu

23:28 <anarsoul> Wizzup: or nvidia and llvmpipe doing some explicit flush somewhere

23:29 <anarsoul> you can try running it with LIMA_DEBUG=singlejob

23:30 <anarsoul> that will disable multijob optimization and will result in a flush whenever you switch render target

23:34 <Wizzup> just tried, didn't seem to help, from what I have seen in apitrace, at some point the framebuffer contents start to differ between llvmpipe and amdgpu, so I am not sure if that would relate to flushes as well, but it could be something like that or a fence