20:08 saml2: I'm using 4k monitor on nv50 (Quadro FX 380). And I get blank monitor. TroubleShooting says "Raising the card performance mode might help". How do I do that? Or how can I confirm that my blank monitor issue is due to video memory bandwidth?
21:05 RSpliet: saml2: Quadro FX 380 is G96, it does not support 4K output.
21:13 karolherbst: yeah..
21:13 karolherbst: it shouldn't be supported actually
21:13 karolherbst: saml2: your desktop might be dump and add custom modes
21:14 karolherbst: *Adds
21:14 RSpliet: *dumb
21:14 RSpliet: although it may also be a dump :-D
21:15 RSpliet: But jokes aside, despite no support for 4K it should still succesfully try a lower resolution, unless it has a slightly outdated version of gnome shell/mutter.
21:15 karolherbst: yeah
21:15 karolherbst: that's my worry here
21:15 karolherbst: but we also need a patch
21:15 karolherbst: https://gitlab.freedesktop.org/drm/nouveau/-/merge_requests/3
21:15 karolherbst: should finish it...
21:16 RSpliet: 6bpc... is that the YUV420 stuff?
21:16 karolherbst: no
21:17 RSpliet: oh, no that's 6 bits per colour. ok
21:17 RSpliet: yeah I don't think that will make a difference on G96 though
21:17 karolherbst: yeah...
21:17 karolherbst: I just hit this on by 12 bpc display and turing and so
21:22 saml2: ah I'm running 1920x1080. With Option "NoAccel" "true", the 4k monitor displays fine. But dragging windows around or switch desktop or alt tab takes few seconds to render
21:22 saml2: is increasing VBIOS clock(?) an option?
21:23 saml2: If I set Option "NoAccel" "false", I get blank monitor sometimes. some times monitor displays fine but then xorg freezes.
21:28 RSpliet: saml2: I'm not sure that works on G96. I don't think I or anyone ever finished that implementation
21:29 saml2: ah so NoAccel is supposed be "true"? By default it becomes "false". And smaller monitor works fine with NoAccel false
21:30 RSpliet: no sorry, noaccel should be false
21:30 RSpliet: but changing the GPU clocks is probably a no-go
21:30 RSpliet: are you using the nouveau DDX, or modeset?
21:31 saml2: ah thanks. so, small monitor, noaccel true, everything is fast and fine. large monitor, noaccel true, it's fast but sometimes hangs or gets blank screen
21:31 saml2: let me google what those are (DDX or modeset)
21:31 RSpliet: can you find a paste website and just share the contents of your dmesg and your Xorg configuration?
21:33 saml2: alright i'll reproduce this blank screen and collect those
21:48 saml2: RSpliet, https://gist.github.com/saml/9c7f2c2d905528a67b237420d995dc77 I don't have /etc/X11/xorg.conf (I created /etc/X11/xorg.conf.d/01-graphics.conf with Option NoAccel true to avoid blank screen) so attached file is generated with Xorg -configure
21:51 RSpliet: ok, so yeah that's modeset
21:51 RSpliet: and I presume you have a reason for not using Wayland on Fedora 34
21:52 RSpliet: Can you share all of 01-graphics.conf?
21:54 saml2: RSpliet, added 01-graphics.conf https://gist.github.com/saml/9c7f2c2d905528a67b237420d995dc77#file-01-graphic-conf
21:55 RSpliet: Hmm ok. So... your 01-graphics.conf says it's using nouveau, but your Xorg logs are saying modeset
21:55 RSpliet: dnf install xorg-x11-drv-nouveau
21:56 saml2: Package xorg-x11-drv-nouveau-1:1.0.17-1.fc34.x86_64 is already installed.
21:57 karolherbst: RSpliet: fedora defaults to modeseting anyway
21:57 karolherbst: which is fine usually
21:57 karolherbst: and I would be surprised if the issue is caused by its usage
21:57 RSpliet: karolherbst: yes but he has a 01-graphics.conf with a Device section that says Driver "Nouveau"
21:58 RSpliet: saml2: actually, sorry... is "he" correct?
21:58 karolherbst: still the 4k bug?
21:58 karolherbst: ohh wait
21:58 karolherbst: ehh
21:58 karolherbst: what's the issue now?
21:59 RSpliet: either slow perf with noaccel (duh), or random flicker and slowness with accel
22:00 karolherbst: mhhh
22:01 RSpliet: oh, and firmware issues for vp and bsp, but that's just for video accel I think.
22:01 karolherbst: yeah
22:01 karolherbst: I mean... nouveaus performance is quite terrible if it comes to displaying anyway
22:01 karolherbst: without reclocking you can't really expect good perf
22:01 RSpliet: I'm sure we can do better if someone works out how that scanout cache works
22:02 karolherbst: no
22:02 RSpliet: yes
22:02 karolherbst: I mean... a bit, but it won't change the game
22:02 RSpliet: that's how caches work, they allow bursty traffic. Which is good for throughput
22:02 karolherbst: ehh sure... but only if you reuse shit
22:03 RSpliet: No
22:03 karolherbst: well.. a cahce can't magically make first time accesses faster
22:03 RSpliet: If you can transfer your FB in 16KiB chunks rather than in 64B chunks, you'll get better utilisation out of your DRAM
22:04 karolherbst: ohh sure
22:04 karolherbst: but that's not a cache then
22:04 RSpliet: Sure it is. You hide away your bits while they drip-feed into the display logic
22:05 karolherbst: but I am not convinced that the FV
22:05 karolherbst: FB is the issue here
22:05 karolherbst: but more compositing
22:05 RSpliet: Anyway, semantics, that things whose name I don't care about can help perf. Not drastic, but nouveau perf is death by 1000 papercuts
22:06 karolherbst: with plain X you get way better perf
22:06 karolherbst: sure
22:06 karolherbst: and that FB thing might help a little
22:06 karolherbst: I just don't expect it to turn everything smooth
22:07 RSpliet: If we have display underruns and actual flicker then I suspect it'll help, if it's just low frame rates, then probably only a tiny drop in the ocean
22:07 karolherbst: well
22:07 karolherbst: last time I tried 4K i got like 15 fps
22:08 karolherbst: moving a window around
22:08 karolherbst: so there is this
22:08 karolherbst: on a turing GPU
22:08 saml2:reads the irc log (it froze again because I had Option NoAccel "false")
22:08 RSpliet: I've got a GK107, 4K@30Hz, and it feels fine
22:08 karolherbst: I am sure it doesn't :p
22:08 karolherbst: I hate 30hz
22:09 RSpliet: saml2: I'm still a bit confused by how you seem to have the nouveau xorg module installed, yet xorg doesn't acknowledge its existence in its logs
22:09 karolherbst: RSpliet: but I guess you uplock?
22:09 RSpliet: karolherbst: nope, unless I want to watch full-screen 4K video
22:09 karolherbst: then you are lucky
22:09 RSpliet: I don't game though, but desktop is fine
22:09 karolherbst:doubts it
22:10 karolherbst: nerd desktop or normal one :p
22:11 RSpliet: Gnome-Shell, so... yeah can't pick a label for that one
22:11 karolherbst: :p
22:11 saml2: hrm interesting. I have Option NoAccel true now. And my Xorg.0.log shows NOUVEAU ... instead of modeset
22:11 karolherbst: saml2: huh...
22:12 karolherbst: maybe there was something weird going on in regards to ... whatever
22:12 karolherbst: but yeah.. I suspect the nouveau ddx to at least perform a little better
22:12 karolherbst: but you might want to enable DRI 3 while at it
22:12 karolherbst: "DRI" "3"
22:12 karolherbst: uhm
22:12 karolherbst: Option
22:12 saml2: https://gist.github.com/saml/4e45b079cbe5959b268f9c2d7787e364
22:12 karolherbst: otherwise you get high CPU usage and tearing
22:13 saml2:tries DRI 3
22:14 RSpliet: Hope it can be made to work with noaccel set to false. Assuming that's about EXI stuff, that should make a difference in perf.
22:15 RSpliet: EXA? I forgot
22:15 RSpliet: imirkin tends to be my crutch for this Xorg stuff :-P
22:17 karolherbst: yeah, we use EXA
22:17 karolherbst: but DRI 3 is a bit broken with EXA on... whatever desktop
22:17 karolherbst: none that matters anymore
22:17 karolherbst: I'd just use wayland to be honest
22:17 karolherbst: don't care a bit about X
22:17 karolherbst: it's broken and annoying
22:19 saml2: DRI 3 with NoAccel true still slow to render. should I try wayland?
22:19 karolherbst: mhhh
22:19 karolherbst: sitll using nouveau?
22:20 karolherbst: saml2: do you know if reclocking works for your GPU?
22:20 karolherbst: would be worth a shot
22:20 saml2: yeah nouveau
22:20 saml2: how do I reclock?
22:20 RSpliet: saml2: not going to lie, wayland has given me the best perf on my nouveau-driven GPU.
22:20 RSpliet: saml2: I don't think I or anyone finished implementing reclocking for G96
22:20 karolherbst: saml2: you have this "/sys/kernel/debug/dri/0/pstate" file
22:20 karolherbst: cat it
22:20 karolherbst: and pastebin the output
22:20 karolherbst: RSpliet: uhh
22:21 karolherbst: for all mem types?
22:21 karolherbst: sad :(
22:21 saml2: okay I installed https://github.com/envytools/envytools not sure if that helps. i'll paste that pstate
22:21 RSpliet: karolherbst: I recall doing work that works on a handful of NVA0, and that translated easily to G98, but not G96
22:21 karolherbst: mhhh
22:22 saml2: karolherbst, https://gist.github.com/saml/e4cab2de098afd26a729222d8cd8e631 this is the output
22:22 RSpliet: It's. Eh. Been a while since I last looked at it. Think I had a G96, just never got round to doing the leg work back then. And today... well... full time job, that old chestnut
22:22 karolherbst: mhhh
22:22 karolherbst: just one way to find out
22:22 karolherbst: saml2: echo 0xf into this file
22:22 RSpliet: Yeah, as I expected, not much to gain from reclocking anyway
22:22 karolherbst: but uff
22:23 karolherbst: 100 mhz...
22:23 RSpliet: memory is already at full speed
22:23 karolherbst: RSpliet: ehh...
22:23 karolherbst: ohhh
22:23 karolherbst: right
22:23 karolherbst: sad
22:23 karolherbst: much sad
22:23 karolherbst: I still think it's compositing though
22:23 RSpliet: should be plenty of BW for a full HD resolution
22:23 saml2: i should just buy a better GPU right?
22:23 RSpliet: saml2: if you do, go AMD
22:23 RSpliet: Not even joking
22:24 karolherbst: or intel
22:24 RSpliet: lol, yeah. But really, AMD
22:24 karolherbst: yeah.. without reclocking perf will suck
22:24 karolherbst: it's... annoying
22:24 RSpliet: Nothing against intel, but their desktop GPUs aren't exactly stellar
22:24 karolherbst: we can fix bugs, but that's about it
22:24 saml2: yeah i'll try. but NoAccel false gives me really good performance. except that it freezes or blacks out sometimes :P
22:24 karolherbst: RSpliet: good enough
22:24 karolherbst: saml2: mhhhhh
22:24 karolherbst: ohh
22:24 karolherbst: that's with NoAccel true
22:24 karolherbst: yeah
22:25 karolherbst: that will suck perf wise
22:25 karolherbst: saml2: do you have the error logs when it freezes?
22:25 karolherbst: and does it freeze with the nouveau ddx?
22:25 karolherbst: but I suspect it's just one of the random bugs
22:25 saml2: nouveaux ddx is when I don't see modeset in Xorg.0.log ? and error log means Xorg.0.log ?
22:25 karolherbst: yeah
22:30 RSpliet: karolherbst: I can't find a single Intel Iris Xe card anywhere...
22:31 karolherbst: :D
22:31 karolherbst: yeah.. dunno
22:33 saml2: I'm back to NoAccel false; I see no modeset in Xorg.0.log . Everything is fast and good. When I encounter freeze, I'll collect dmesg and Xorg.0.log :P
22:33 RSpliet: Thanks, that's very helpful
22:33 karolherbst: yeah
22:33 karolherbst: I wouldn't be surprisded if the nouveau DDX is more stable, because.. less GL
22:33 karolherbst: the nv50 GL driver really feels a bit buggy tbh
22:44 RSpliet: oh no
23:48 imirkin: nouveau ddx is way more stable. this is why i hate the redhat/debian patch to prefer modeset.
23:48 imirkin: nouveau ddx deals with errors, among other things
23:50 imirkin: saml2: how is the display connected? i'm guessing a DVI-D <-> HDMI adapter somewhere in there?
23:51 imirkin: if so, you can't have 4K -- you max out at 165mhz because of how DVI works
23:51 saml2: imirkin, yup exactly that. card had DVI and monitor has HDMI
23:51 imirkin: some people have successfully driven 4k on nv50 cards
23:52 imirkin: (4k@30)
23:52 imirkin: you can work around the clock speed limit, and just hope for the best
23:52 saml2: ah i'm just happy with 1920x1080 as long as it doesn't freeze. so far no freeze
23:52 imirkin: it'll be highly dependent on the physical wiring/etc
23:53 imirkin: since it would be way out of spec
23:53 imirkin: ok, cool
23:53 imirkin: i gather you've been instructed to remove NoAccel and force the nouveau ddx?
23:54 saml2: yeah. i'm now forcing nouveau ddx and set NoAccel false (I guess I can remove that Option line)
23:54 imirkin: cool
23:54 imirkin: that should be fairly stable
23:54 imirkin: note that the whole nv50 generation has general stability issues, so you will get freezes sometimes
23:55 imirkin: no real way around it (besides fixing the bug)
23:55 saml2: ah i see. thanks. i'm happy so far :)