10:47karolherbst: okay... nice
10:48karolherbst: no regressions inside the CTS.. and glcts -n "*multithread*" also works :D yay
10:48karolherbst: I still suspect random multithreading state issues, but at least it shouldn't crash anymore
10:49karolherbst: let me verify I also don't regress those android tests.. those are quite a lot
10:52karolherbst: imirkin: soo.. if you got some time, 5 patches I think are ready to get merged: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10752
10:52gitlab-bot: Mesa issue (Merge request) 10752 in mesa "nouveau: multithreading fixes part 1" [Nouveau, Opened]
10:53karolherbst: not a lot of changes but still some :)
16:59imirkin: karolherbst: will try to look on sunday
17:45karolherbst: cool, thanks!
19:00karolherbst: mhh.. all aosp + GLES CTS + GL CTS tests: [110751/110751] Pass: 101871 Fail: 735 Skip: 8003 ExpectedFail: 0 UnexpectedPass: 0 Crash: 87 Timeout: 24 Missing: 0 Flake: 31 Duration: 5:41:14 Remaining: 0
19:00karolherbst: curious about those timeouts
19:01imirkin: wow, 5h? that seems like a lot.
19:01karolherbst: check the number of tests
19:01imirkin: that's also a lot of tests... yea
19:01karolherbst: I have a script which merges like all mustpass files together :D
19:01karolherbst: and removes dups
19:01karolherbst: and pushes it through the parallel deqp runner
19:01imirkin: which gen bw?
19:01karolherbst: turing
19:01imirkin: k
19:01karolherbst: but that's with my MT fixes
19:02karolherbst: but those timeouts also happen on main
19:02karolherbst: I set the timeout to one minute
19:03karolherbst: in 3 hours I have the result from the run on main :D
20:35karolherbst: cleaning up my other patches will be a little painful though :/
20:35karolherbst: imirkin: any opinions on reworking the state tracking we do in nvc0? I suspect that with my changes it will be bogus, due to the nature how submissions are made, but I don't think there is a pressing need to fix it atm as long as we don't necessarily hit bugs
21:22imirkin: "reworking"?
21:22karolherbst: well.. okay.. so here is the deal
21:22karolherbst: with my rework each context has it's own push buffer they are building up
21:22imirkin: there are a number of things it tries to optimize for
21:22imirkin: so it's important to maintain those
21:23karolherbst: so one context can build up its buffer base on one state, and the next on the previous one
21:23imirkin: per-context pushbufs are fine, but very tricky to actually implement.
21:23karolherbst: but
21:23karolherbst: the latter context could submit the pushbuffer before the previous one
21:23imirkin: that's why i have the screen-level graph state
21:23karolherbst: sure
21:23karolherbst: but we update the state when building the buffer
21:23karolherbst: and the order of building the buffer doesn't have to match the order of submission
21:23imirkin: yeah, but there's one pushbuf, so it all works out.
21:24karolherbst: sure.. atm... but it's also crashing :D
21:24karolherbst: so it might be fine on the GPU once you manage to submit it without damage from races
21:24karolherbst: which.. normally doesn't happen
21:24imirkin: anyways
21:24imirkin: i gtg now
21:24karolherbst: yeah
21:24karolherbst: have fun
21:24imirkin: but this is why per-context pushbufs are very tricky to do
21:24imirkin: and i was suggesting not going in that direction.
21:24karolherbst: I know
21:25karolherbst: I am thinking about the problem but don't have a good solution yet
21:25imirkin: (well, it's not THE reason why, but one of)
21:25karolherbst: and anyway.. doesn't matter
21:25karolherbst: today we crash
21:25karolherbst: with my changes there might be glitches
21:25karolherbst: which.. we had before anyway
21:25karolherbst: so....
21:25karolherbst: fixing this isn't exactly the highest priority atm
21:27karolherbst: I simply want to come up with a plan :D
22:26karolherbst: mhh.. on main: [110751/110751] Pass: 101855 Fail: 764 Skip: 8003 ExpectedFail: 0 UnexpectedPass: 0 Crash: 67 Timeout: 24 Missing: 0 Flake: 38 Duration: 5:46:02 Remaining: 0
22:26karolherbst: 20 crashes less..
22:26karolherbst: interesting
22:27karolherbst: probably something dump.. will look into it tomorrow
22:27RSpliet: it already is tomorrow!
22:27karolherbst: ehhh
22:27karolherbst: I shouldn't work on the weekend anyway
22:27RSpliet: (a paradox in four words :-D)
22:28karolherbst: and yesterday was today
22:29karolherbst: or the other way around? mhh