04:07airlied: skeggsb, jekstrand : any pointers to how the nvidia compressed kind stuff works?
15:52KungFuJesus: lol, well that's fun/new one. Font rendering is mirrored if X is started with modesetting instead of the nouveau DDX
15:53KungFuJesus: seemingly only in the terminal emulator?
15:55KungFuJesus: that's almost certain to be an endianness issue, right?
15:55KungFuJesus: looks like alien glyphs
15:58KungFuJesus: modesetting does seem to fix some of the weird vsync behavior where code couldn't reliably determine the swap interval
15:58KungFuJesus: you know what, I'll bet this is an issue with glamor
16:00KungFuJesus: yep, that's the trigger
16:01KungFuJesus: errr no setting AccelMethod to none seems to turn off all GPU acceleration? That's odd
16:07KungFuJesus: closest bug I can find that mimics the behavior: https://bugs.launchpad.net/stellarium/+bug/569701
16:24KungFuJesus: all the people whining about subpixel font rendering and smoothing have nothing on this issue, hah
16:25KungFuJesus: I think the nouveau GL doesn't do this specifically because accelerated GL was ripped out of it. Which is, I dunno kind of a convenient side effect
17:19KungFuJesus: I'm starting to think there's something wrong with the fencing around vertical sync: https://i.imgur.com/ycEm8TZ.png
17:23jekstrand: airlied: WDYM?
17:25KungFuJesus: I'm thinking though that that game code might be kind of racey in general when setting up the VBOs
18:39KungFuJesus: karolherbst: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10752
18:40KungFuJesus: could known issue #3 be something I'm hitting on nv30 as well
18:52KungFuJesus: I feel like there's something silly going on in this code somewhere and that's why I'm getting that NULL pushbuf and that assertion is firing
18:52KungFuJesus: perhaps suspect? https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10752/diffs#f1d971aa48b7aeefbc5871dc0936eb04ad9a3a23_211_203 that early return with NULL on the context creation?
18:54KungFuJesus: it seems like the error handling for this is a quite a bit more relaxed than nv50? https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10752/diffs#a807797068f6b76a0c8a072820fcb43064737202_422_427
18:57KungFuJesus: oh, nv30_context destroy kinda does the same amount of work
19:06KungFuJesus: karolherbst: so I'm like 90% sure the fencing code in nv30 has issues. It seems like the VBOs, before they are even fully uploaded, manage to be used. With lower framerates, this doesn't become apparent, but when vsync is enabled and it tries to hit a target 60fps it becomes _very_ apparent. I don't know if it's specific to this engine or not, but I don't think it's anything you did to the code
19:06KungFuJesus: that's causing this. In fact, from 2018 I see an early indicator that happened when framegrab managed to hurt the framerate
19:06KungFuJesus: e.g. https://youtu.be/TUi1cWcqJZk
19:08KungFuJesus: This: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/20926 might be a symptom of improper fences somewhere
19:10KungFuJesus: You may be the only one that help, hah. I have a GPU with an x86 VBIOS that I can try on an x86 machine to rule out endianness sorts of issues, I'll have to pick it up this weekend.
19:11KungFuJesus: The problem is I don't _really_ know what I'm doing or understand these GPU architectures to sus out how they're supposed to be used. It also sounds like nv50 might actually have a similar issue, too
20:03KungFuJesus: s/that help/that can help/
20:36KungFuJesus: what's interesting is that seemingly none of these issues rear their ugly head when built with the GL_LEGACY version of the engine, which doesn't use VBOs
20:36KungFuJesus: and, if vsync is disabled, 60fps is perfectly doable