00:00imirkin: i don't think that's quite sound
00:01karolherbst: I couldn't tell
00:01karolherbst: and I am sure nobody else could
00:01karolherbst: the current calculation sounds like some heuristic anyway
00:01karolherbst: I am sure this would be too optimistic though
00:02karolherbst: and I am sure the proper fix would be to get rid of that nospill flag and handle all of that correctly
00:03imirkin: i think curro should know what this is all about
00:04imirkin: if you can track him down
00:05karolherbst: yeah.. I should be able to. He was reviewing all my clover stuff recently
00:08karolherbst: but I think I have a better understanding about all of that know... it's just always feels like the code is way more complex than it has to be, but maybe it actually has to be that complex.. dunno
00:10imirkin: he'll probably be able to tell you what the rel degree stuff is about
16:39imirkin_: skeggsb: purely anecdotal, but with an otherwise identical setup, GP107 is *way* more stable than GK208 with chrome with a particular website
16:39imirkin_: any theories on why this might be the case?
16:42imirkin_: (the website in question is lichess.org -- esp certain operations, like fetching a (new) computer analysis would hang X maybe 1/30 times)
16:43imirkin_: (as well as just general playing, although rarer)
16:44diogenes_:challenges imirkin_ for a chess game...
16:44imirkin_: i'm really bad.
16:44imirkin_: embarassingly so.
16:45imirkin_: below average, as it were.
16:45diogenes_: :) what rating?
16:45imirkin_: average is 1500
16:45imirkin_: so below that.
16:45imirkin_: [on lichess, that is]
16:46diogenes_: i see but still not bad.
16:46imirkin_: 0 is less than 1500 :)
16:46imirkin_: (one would have to work _hard_ to get that though...)
16:47imirkin_: i'm between 1300 and 1400 in both blitz (3+2) and bullet (1+0)
16:49diogenes_: imirkin_, this is mine: https://lichess.org/@/Adammah
16:49imirkin_: yeah, not even close.
16:50diogenes_: i don't play classical because there are many cheaters using engines.
16:50imirkin_: how do you have a provisional rating with 600 blitz games?
16:50karolherbst: maybe I should start to play chess again.. only played competetive offline ...
16:50imirkin_: careful - it's addictive :)
16:50diogenes_: no clue i mostly play on gameknot nowadays.
16:51imirkin_: diogenes_: i don't play classical because i don't have the patience
16:51karolherbst: imirkin_: well.. I used to play in tournaments and what not :p
16:51imirkin_: blitz is probably my sweet-spot
16:51imirkin_: karolherbst: so you know :)
16:52karolherbst: is lichess like the biggest community or are there others?
16:52imirkin_: chess.com is bigger, but it's paid and feels super-commercial
16:53imirkin_: lichess is a very modern-feeling UX, and they allow you to do by-move analysis against stockfish
16:53imirkin_: which is good for going back to see if you any better options at any particular point in the game
16:53diogenes_: on lichess you can even earn 500$ is you win a tournament.
16:53imirkin_: yeah, and there are tournaments for peopel rated under X
16:54imirkin_: (1500, 1800, etc)
16:54imirkin_: so you don't have to necessarily compete against GM's
16:54karolherbst: ahh, sounds cool
16:55imirkin_: anyways, this is me, in case you've been wondering what i've been doing instead of hacking on nouevau: https://lichess.org/@/ibm5000
16:55karolherbst: what a user name :D
16:56diogenes_: i'd have never guessed that name xD
16:56imirkin_: like i said. very sad.
16:59karolherbst: the AI level 1 is weak...
16:59imirkin_: it's very weak, and kinda random
17:00imirkin_: i think by level 4, it's fairly strong
17:01imirkin_: and you'll have a LOT of trouble at the highest level
17:01imirkin_: (i guess i don't know how good you were "back in the day")
17:02karolherbst: yeah.. and it's like 15 years ago
17:04diogenes_: learning chess has always been my dream since childhood, but i had no possibility, until i got a windows laptop with chess titans, it was my chess teacher, i learned the moves at age 25.
17:09imirkin_: diogenes_: and you did a much better job than i did, apparently
17:10imirkin_: with bullet, i just need to learn how to move my mouse faster and more accurately
17:10imirkin_: but with blitz, i get outplayed
17:20diogenes_: imirkin_, at the beginning my fav was 20+10s and it still wasn't enough, now 3min are quite alright and rules out the possibility that someone use an engine.
17:21imirkin_: i don't lose on time in blitz, i lose on position
17:22imirkin_: (i def lose on time in bullet... heh)
17:22diogenes_: well you have less time for chess because you do a lot of programming and that's why.
17:23imirkin_: lol, that's just an excuse.
17:23imirkin_: a bad one.
17:23imirkin_: anyways, my rating's going up. i think i can reach like 1600 with just practice.
17:24diogenes_: yes also watching some chess games on yt.
17:24imirkin_: been doing plenty of that
17:24imirkin_: the fried liver attack is always a fun one, for example
17:25karolherbst: ohh, I am so scarily out of practise
17:25imirkin_: karolherbst: are you playing against engine or people?
17:25imirkin_: engine is very hard to play against
17:25karolherbst: anyway, lvl 1 is behind me
17:25imirkin_: coz it's very strong
17:25karolherbst: lvl2 now :)
17:25imirkin_: and the lower levels just cut off some part of its cleverness
17:25imirkin_: but deep down inside it's still very strong
17:26karolherbst: yeah.. I noticed
17:26karolherbst: it did some nasty moves though
17:26imirkin_: so sometimes it doesn't notice some super-obvious thing
17:26imirkin_: and other times you'll go down a 10-deep forced mate path
17:26imirkin_: level 1 is *very* random
17:26imirkin_: so random that it's a bit hard to play
17:26imirkin_: level 2 is easier :)
17:26karolherbst: https://lichess.org/k4P9HDe71OoC this one was weird
17:27karolherbst: usually my play style doesn't help with the game being less random though...
17:27imirkin_: you can also look at the computer analysis
17:28imirkin_: which can be quite interesting
17:28imirkin_: (but hangs chrome on my GK208)
17:28imirkin_: (but works on GP107)
17:29karolherbst: the computer doesn't like my d3 move.. I disagree
17:29imirkin_: and that's why the computer's rating is what it is, and your is what it is.
17:29diogenes_: karolherbst, cool game.
17:30imirkin_: karolherbst: playing against people is a lot more fun
17:30imirkin_: it's a bit scary at first, but you get over it pretty quick
17:30karolherbst: I just need to get up to speed for now
17:30karolherbst: also I know why that move was dubious, it just didn't matter in the long run
17:30danvet: imirkin_, karolherbst https://patchwork.kernel.org/patch/11108207/
17:31danvet: what kind of bribes do you guys accept?
17:31imirkin_: danvet: for me, the best thing would be a well-timed reminder. like on friday.
17:31danvet: I also need some review/acks really
17:31imirkin_: although i'm also starting to feel bad
17:31imirkin_: so i might try to do it tonight.
17:32danvet: iirc you tested it already, but there was some unclarity on whether some flakiness was due to my patch or something else in the 4 kernel versions you upgraded
17:32imirkin_: i did
17:32imirkin_: with a nv5
17:32danvet: but technically it's really just stuck on the lack of rb/ack
17:32imirkin_: which is questionable to begin with
17:33imirkin_: i'd like to retest with something later, like nv3x or nv4x
17:33danvet: without rb/ack I can't push it any tree
17:33danvet: imirkin_, well atm rebasing, that's why I remembered
17:42danvet: imirkin_, https://patchwork.freedesktop.org/series/68959/ updated series
17:48imirkin_: karolherbst: https://bugs.freedesktop.org/show_bug.cgi?id=112197 - is this like what you've been investigating?
17:49karolherbst: imirkin_: it's runpm related I think
17:49karolherbst: but runpm is supposed to work on that machine
17:49imirkin_: he also mentions suspend
17:49imirkin_: not runpm
17:50karolherbst: yeah... it's still the same code more or less. Detecting outputs on the resume path
17:50karolherbst: Lyude was working on that stuff quite a lot
17:50karolherbst: and I think we have a x1 internally available
20:18Akronym: http://paste.debian.net/1113737/ <- broken hardware? GM107, debian sid, kernel 5.3 amd64
20:19imirkin_: broken software is more like it ...
20:19imirkin_: that looks like it tried executing a BS page of code
20:24Akronym: imirkin_, screen freezes when that happens. Sound keeps playing for a while. ssh into the box and reboot is as of now the only "solution"
20:26Akronym: imirkin_, broken software would be interesting. I mean, nouveau is used by many debian users and developers, but the bts looks "clean"
20:34karolherbst: looks like the multithreading issue
20:45imirkin_: could very well be
20:45imirkin_: or could be us just plain mismanaging the code page
20:45imirkin_: both have happened :)
20:46imirkin_: Akronym: oh, since you mentioned debian ... what version of mesa?
20:46imirkin_: if it's truly ancient, might not have the fixes around resizing the code page
20:47imirkin_: that means nothing to me
20:47karolherbst: it's essentially the rolling release debian
20:48karolherbst: seems to be in sid right now
20:48imirkin_: ok. so not truly ancient :)
20:48karolherbst: yeah.. sid is essentially always the newest stuff
20:48Akronym: imirkin_, not truly ancient, nice way to describe sid ;)
20:49karolherbst: sid is the only useable version of debian
20:49Akronym: karolherbst, nack, depends on what you want / need.
20:49karolherbst: if you want security, you need sid anyway...
20:49karolherbst: everything else is trash
20:49imirkin_: basically any time anyone mentions debian, i'm suspicious that it's something we've already fixed
20:50imirkin_: lots of time has been wasted that way
20:50Akronym: karolherbst, nack, but OT, so.
20:51Akronym: imirkin_, yeah, especially in the past where nouveau was... well, not where it is today and nouveau progress was moving fast, debian *stable* did lack eons behind.
20:52imirkin_: nowadays nouveau's pretty stably unstable
20:53imirkin_: there are a bunch of known issues, and no one is working to fix any of them
20:53Akronym: imirkin_, stably unstable?
20:53imirkin_: i.e. the instability is stable :)
20:53Akronym: ah :)
20:53Akronym: imirkin_, any one of them that could explain my issue?
20:54imirkin_: it's a lot more fun to work on new features than it is on redoing things to not die
20:54imirkin_: Akronym: karol's explanation of multithreading gremlins is fairly plausible
20:54karolherbst: doesn't help if those fixes just add a pile of work :/
20:56Akronym: imirkin_, I guess disableing all but one cpu core and limiting that core to ST is not what I would call a solution. ;)
20:56imirkin_: and also wouldn't help
20:57imirkin_: you can still have multiple threads in a single-cpu situation
20:57imirkin_: and a particular thread may be interrupted at any point
20:57imirkin_: so you still have all the same problems
20:59Akronym: which I wouldn't mid getting rid of btw. ;)
20:59Akronym: mind even
20:59imirkin_: then you move to a cooperative multi-tasking system
20:59imirkin_: like pre-OSX macos
20:59imirkin_: and win3
21:00imirkin_: not sure about OS/2
21:01Akronym: imirkin_, anyway, should this issue hit me with every nvidia card on this system?
21:02imirkin_: but as a bonus, some of the cards may get additional issues! :)
21:02Akronym: imirkin_, *lol*
21:02imirkin_: this is all based on what software you run btw
21:02karolherbst: imirkin_: maybe we should fail creating a second context... until we have a proper solution
21:02imirkin_: if you run non-GL-intensive software, things should fare better
21:03imirkin_: Akronym: oh wait
21:03imirkin_: debian has this #$@(& patch in xorg
21:03imirkin_: which makes it 100x more likely to hit this issue
21:04karolherbst: imirkin_: glamor only uses one context though afaik
21:04imirkin_: but it can also run into additional issues
21:04karolherbst: I don't see how glamor is relevant for the mluti context issue.. you are hitting the out of VRAM issue way faster.. yes
21:04imirkin_: Akronym: try forcing nouveau as your xorg driver
21:04karolherbst: and maybe other issues, right
21:06Akronym: imirkin_, I don't get you.... isn't it "forced" as long as it isn't blacklisted (e.g. while using the nvidia-blob)?
21:06imirkin_: Akronym: they have a special patch which disables it for post-nv50 gpu's
21:07Akronym: nv50? That really really ancient generation?
21:08imirkin_: that includes everything semi-modern
21:10Akronym: imirkin_, but nouveau IS used on debian by default (as long as it isn't blacklisted) for my card.
21:10imirkin_: check your xorg log
21:10Akronym: imirkin_, for what sould I keep looking?
21:11Akronym: should even
21:11imirkin_: NOUVEAU(0) vs modeset(0)
21:11imirkin_: this is the patch
21:12imirkin_: i really wish they'd drop it, people would have fewer problems
21:12imirkin_: looks like it was copied from redhat/fedora
21:12Akronym: imirkin_, ah, kms... so kernel param nomodeset or whatever it was called and I am good to go?
21:13imirkin_: where i also think it's a mistake, but ben works there and is the nouveau guy, so i can't really effect any change
21:13imirkin_: nomodeset = disable all modesetting. you don't want that.
21:13karolherbst: imirkin_: afaik the out of VRAM thing is the only issue really relevant here. Everything else seems to be glamor core stuff. I really don't know about other issues
21:14imirkin_: karolherbst: more GL usage = more opportunity for fail
21:14karolherbst: (obvously, the nouveau ddx has better perf, but yeah)
21:14imirkin_: definitely don't want that in the X server
21:14imirkin_: doubtful that it has better perf
21:14karolherbst: lower CPU overhead
21:14imirkin_: the stability is the real advantage.
21:17Akronym: [ 137.074] (II) modeset(0): using drv /dev/dri/card0 <-- imirkin_
21:17imirkin_: so you're using the generic modesetting driver
21:17imirkin_: which in turn uses GL to provide the X acceleration
21:18Lyude: it's stable
21:19imirkin_: Lyude: you're joking, right?
21:19imirkin_: nouveau GL driver has like zero error handling
21:20imirkin_: not to mention GLAMOR being a great test-case for a while to reproduce various failed rendering (which magically stopped, it seems ... i still have some old traces that render wrong)
21:20imirkin_: i'm not saying it's a bad idea, but for nouveau, the dedicated ddx is *way* better
21:21Akronym: imirkin_, so how does one force nouveau to be used instead of modeset?
21:21Lyude: imirkin_: I mean, I'm not saying there aren't issues. but the whole reason we went with the modesetting ddx is because it's substantially less work to maintain one ddx than multiple of them, and thus issues are easier to fix and it's easier to keep stable.
21:21imirkin_: Akronym: set the Driver explicitly
21:21Akronym: imirkin_, like, yorg.conf?
21:21imirkin_: Lyude: yeah, like i said, it's a nice idea
21:21Akronym: xorg even
21:21imirkin_: Akronym: xorg.conf preferably
21:22imirkin_: Lyude: in practice, nouveau ddx requires very little maintenance, and is super-well-tested
21:23imirkin_: nouveau GL driver is known to have tons of issues
21:23imirkin_: and also makes much more extensive use of the GPU than the ddx does
21:24Lyude: imirkin_: I mean, my workload dropped pretty significantly once we stopped using the other DDXs, and I've seen tons of various cargo cult bugs across the ddxs. that being said I don't think I ever really worked on nouveau's ddx for very long either (if I ever did?)
21:27imirkin_: can't really comment on the other DDX's
21:27imirkin_: and it did take a while to get MST goign, largely because it was a long while before i had access to such a setup
21:27imirkin_: (and it's hard to develop that stuff in the blind)
21:27karolherbst: imirkin_: ever looked into resource_from_user_memory?
21:27imirkin_: (even though most of it was copied)
21:27Lyude: imirkin_: yeah sorry about that :s, stuff had gotten extremely busy around when you had been asking me to test stuff
21:28imirkin_: yeah, no worries
21:28imirkin_: i suck at doing it for other people, i don't see why others would be any better at it :)
21:28karolherbst: imirkin_: my plan is to implement resource_from_user_memory on top of HMM and I was wondering how painful that might be
21:28Lyude: speaking of testing I am extremely looking forward to having igt running on nouveau once I finish up some work stuff this week :)
21:28Lyude: i bet we are going to find a lot of fun issues
21:29imirkin_: Lyude: yeah, that'll be great
21:29karolherbst: Lyude: I think the biggest issue with glamor right now is the high VRAM usage and on GPUs with low VRAM (512MB or lower) you can run into OOM situation glamor isn't handling all that well ;?
21:30karolherbst: I think there were multiple attempts on fixing this issue, but nobody I know of was even able to reproduce the original one
21:30imirkin_: in the end i managed to borrow a GPU with DP output from a friend, and then got it all going in a few evenings
21:30karolherbst: even I failed with 512MB cards
21:33Akronym: imirkin_, I dropped a snippet in /etc/X11/xorg.conf.d/ defining nouveau as the driver. ANything else?
21:33imirkin_: that is all
21:35Akronym: imirkin_, befor I reboot, care to explain what this modeset driver is? I mean, I have never seen it / read about it.
21:35imirkin_: it's a generic drier
21:35imirkin_: that uses GL to provide acceleration
21:36Akronym: imirkin_, GL as in OpenGL?
21:36imirkin_: the very same
21:36Akronym: imirkin_, any disadvantages I might have now?
21:36imirkin_: the nouveau ddx doesn't support DRI3 very well
21:37imirkin_: you can force-enable it if needed, but there are occasional issues
21:37imirkin_: this generally doesn't matter
21:37Akronym: imirkin_, what needs dri3?
21:39Akronym: anyway, brb, reboot
21:43Akronym: [ 16.474] (--) NOUVEAU(0): Chipset: "NVIDIA NV117"
21:46Akronym: imirkin_, weirdly so
21:47Akronym: imirkin_, how should I put it.... I wasn't even aware I wasn't using nouveau. Or to be more precise, I am still a bit confused about that modeset vs. nouveau stuff but I'll try to educate myself
21:48imirkin_: just different xorg backends
21:48imirkin_: for providing rendering/display services
21:48imirkin_: aka "ddx" - device-dependent ... X
21:48imirkin_: the nouveau one makes direct ioctl's to the nouveau kernel drier
21:48imirkin_: the modesetting one uses GL
21:49imirkin_: the benefit being that it can work on any GL impl
21:49imirkin_: and provide X acceleration
21:49imirkin_: so you don't have to make both a ddx *and* a GL driver
21:50imirkin_: and there can be just one correct thign to enhance and maintain
21:50imirkin_: in practice, the nouevau GL driver sucks
21:50imirkin_: so it's not a great base for something like this.
21:53Akronym: imirkin_, sucks in what way?
21:54imirkin_: error handling, correctness, dealing with adverse conditions
21:54Akronym: imirkin_, so I could run into other issues along the line?
21:58Akronym: imirkin_, let's see what happens in the next few days.
21:58karolherbst: heh.. mhh
21:59karolherbst: 50% fails for amd_pinned_memory :(
22:01HdkR: Oh, supporting pinned_memory?
22:02imirkin_: with the svm thing?
22:03HdkR: Dolphin supports pinned memory, so that could be a useful test case :P
23:54karolherbst: ahh, now all tests are crashing :)
23:54karolherbst: imirkin: yeah
23:54karolherbst: funny enough, we have the concept of nv04_resource objects using malloced memory
23:54karolherbst: nvc0 doesn't handle it all that well :)
23:55karolherbst: now I got a NULL bo added from that resource and libdrm doesn't like that at all
23:55karolherbst: imirkin: I actually need it for implementing the SVM stuff.... userptr support, for weird reasons, which are all annoying
23:56karolherbst: so I guess I just implement userptr for pascal+ then.. *sigh*
23:58karolherbst: ahh nvc0_add_resident...
23:58karolherbst: no point refing a resource with a user ptr
23:59karolherbst: "fifo: fault 00 [READ] at 0000000000000000 engine 00 [GR] client 02 [GPC1/PE_0] reason 00 [PDE] on channel 2 [00ffbaf000 amd_pinned_memo]" ehh