06:15Ironsmith: hello! sorry if this gets asked a lot: i have 2 video cards, GeForce GTX 570 / NVC8 (Fermi) and GeForce GTX 1060 / NV136 (Pascal) and according to the feature matrix they're virtually identical, but when i try wayland i get an error on the newer (latter) card. tried it with hyprland, sway, and river
06:18Ironsmith: i'm using guix and nouveau version is the latest i believe:
06:18Ironsmith: srcversion: AA677C6F8540F88B9855174
07:07Ironsmith: i'm looking at the troubleshooting and realising that nouveau itself is running and can display the gnome login manager just fine, it's after i log in and try to run a wayland compositor where the issue starts, so it's probably a wayland issue right?
07:30chikuwad[d]: pretty sure GDM also uses wayland by default
07:46Ironsmith: oh good insight! thanks for that. hummm i think i'll try creating a discussion topic in one of the compositors, and see if they're willing to dig into it
17:52snowycoder[d]: mhenning[d]: I'm back! what tests should I run?
19:01mhenning[d]: snowycoder[d]: Okay, the two things are:
19:01mhenning[d]: 1. Under the proprietary driver, do the same steps as last time, but with the code under the "vk_test_overlap_exec" directory in mhenning/re (directory is new, probably need to git pull)
19:03mhenning[d]: 2. Under nouveau, build https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37671 and run the `./main` executable from the vk_test_overlap_exec directory against that branch of nvk
19:27mhenning[d]: I suppose a blackwell cts run of https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37671 would also be useful if you're in the mood for it.
19:54mohamexiety[d]: phomes_[d]: hey when you tested compression last time, how was perf?
19:54mohamexiety[d]: was it identical before I fixed up the hacks?
19:55mohamexiety[d]: reason I am asking is I just noticed that xonotic wasn’t actually using compression after fixing up the old hacks, but not sure if it’s a general thing
19:55mohamexiety[d]: (Since xonotic uses zink so could have some interactions there blocking it off)
19:56phomes_[d]: yes it was identical
19:57mohamexiety[d]: hm interesting. thanks!
19:59mohamexiety[d]: this alignment stuff is actually really annoying
20:05mohamexiety[d]: Also I just noticed we have been aligning MEM_LOCAL (that’s GART | VRAM) images to 64K. I wonder why :thonk:
20:06mohamexiety[d]: Noticed that asserting larger alignment for compression only was blowing up and then I noticed this
20:44snowycoder[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1423772614945542305/nvidia_all.out?ex=68e18713&is=68e03593&hm=49c84b5149dfeae59685b872f2c64a5b0f88be80e578cce4a7a1e369f1908c14&
20:44snowycoder[d]: mhenning[d]: `vk_test_overlap_exec` with my RTX 5070ti using drivers 580.82.09
20:47snowycoder[d]: mhenning[d]: ```
20:47snowycoder[d]: ❯ run_with_mesa ./main
20:47snowycoder[d]: useBarrier = 0
20:47snowycoder[d]: dg dc dt
20:47snowycoder[d]: sg 00 ff ff
20:47snowycoder[d]: sc ff 00 ff
20:47snowycoder[d]: st ff ff ff
20:47snowycoder[d]: useBarrier = 1
20:47snowycoder[d]: dg dc dt
20:47snowycoder[d]: sg ff ff ff
20:47snowycoder[d]: sc ff ff ff
20:47snowycoder[d]: st ff ff ff
21:13snowycoder[d]: mhenning[d]: Running, there seem to be some failing tests :/
21:13snowycoder[d]: I'll see if they are regressions introduced by your patch
21:17snowycoder[d]: Also, interesting benchmark:
21:17snowycoder[d]: time alacritty -e /usr/bin/echo hi
21:17snowycoder[d]: On nvidia-prop it hovers around ~150ms while with ZINK+NVK it seems to be ~350/400ms.
21:17snowycoder[d]: Is it just me my system or do we have a big startup time problem? It makes the system feel a bit slow when opening programs
21:21mhenning[d]: snowycoder[d]: Thanks for running that! Looking through the results now
21:21mohamexiety[d]: we do have slower device creation time but idk if that's what you feel with e.g. opening programs
21:21mohamexiety[d]: your system might be defaulting to a lower screen refresh due to driver change which could contribute to your system feeling slower
21:22mhenning[d]: snowycoder[d]: That's definitely possible. I can't say we've optimized start up beyond some basic compiler benchmarking
21:26snowycoder[d]: mohamexiety[d]: It does (180 -> 165Hz) but what I feel is the time between opening alacritty and seeing a window, I don't think a 15Hz difference is what matters here
21:27mohamexiety[d]: ah yeah
21:36mhenning[d]: snowycoder[d]: If you want to play with it, you could try pointing a cpu profiler at your `alacritty -e /usr/bin/echo hi` example and see if anything stands out
21:37mhenning[d]: it's possible that could be informative, or it's possible that the overhead could be on the kernel or firmware side which will be harder to track
21:43snowycoder[d]: mhenning[d]: I'll do that, I added it to my TODO list!
22:25mhenning[d]: snowycoder[d]: I added a commit that may possibly fix something
22:26snowycoder[d]: mhenning[d]: A lot of tests fail in origin/main too, might not have run CTS in a bit
22:30mhenning[d]: hmm. cts passes on ampere for me on that branch, so if there are also fails on 7df6431bf727c21e10160cd0048442565b7cbbd6 (base of that branch) then they might be blackwell specific
22:30mhenning[d]: or maybe I'm on a different cts version
22:47mhenning[d]: karolherbst[d]: do your docs happen to mention what a NVB197 FORCE_HEAVYWEIGHT_METHOD_SYNC is? I'm wondering about its function in a trace from the proprietary driver
22:48karolherbst[d]: mhhhh that could mean anything 🙃
22:48mhenning[d]: i know, right? 😛
22:48karolherbst[d]: ohhh
22:48karolherbst[d]: actually...
22:49karolherbst[d]: mhhhh
22:50karolherbst[d]: nope, no idea
23:20marysaka[d]: there seems to be something about heavyweight and lightweight operations and if I extrapolate from `SET_TILED_CACHE_BUNDLE_CONTROL` it might have something to do with tile cache? now what that mean no idea
23:21marysaka[d]: so maybe it perform a flush in case of FORCE_HEAVYWEIGHT_METHOD_SYNC? but if a force method exist, other methods might have a "heavyweight" behavior isn't it
23:24mhenning[d]: huh, there's also a NVB197_SET_MUTABLE_METHOD_CONTROL_TREAT_MUTABLE_AS_HEAVYWEIGHT