02:46nyef: Is falcon related to Tesla's shader processors, or is it an unrelated architecture?
02:46nyef:is expecting it to be unrelated, but willing to be pleasantly surprised.
02:53nyef: Okay, thanks.
02:53nyef: Yet another architecture to learn, then.
02:53imirkin: falcon came about to replace the xtensa processors embedded in the chips
02:54imirkin: which were used to control auxiliary engines
02:54nyef: Yeah, I just hit that part of envytools/docs/hw a minute or so ago.
02:55imirkin: nv98 and nva3+ all have falcons, all the others have xtensa (except nv50, which has nothing)
02:55nyef: I see that the Curie/Tesla context-switch microengine got swapped out for a falcon in Fermi, and presumably that hasn't changed in any later hardware revision?
02:55imirkin: correct. the actual falcon processor has gotten upgraded, but same idea.
03:02HdkR: imirkin: Do you have a custom emitter for the falcon?
03:06nyef:wonders if he needs to build a custom kernel in order to "free up" his NVAF for direct manipulation.
03:06imirkin: no, only envyas
03:07imirkin: but the context switching on tesla's is not falcon on any of them
03:07imirkin: the falcon's in nvaf control the video decoding engines, the copy engine (iirc), and the VIC
03:10nyef: ... Looks like I could get "full" coverage of the Teslas for a mere 10 cards and two all-in-ones?
03:12imirkin: for truly full coverage, you need pretty much every chip. i think there are exceptions for all kinds of junk
03:12imirkin: maybe only one of g84/g86
03:13imirkin: depends what you're trying to cover, of course
03:13nyef: Mmm. Right.
03:13imirkin: happy to send you a few, but i hardly have full coverage myself
03:14imirkin: i have g84, g92, g96, g98, gt215
03:15nyef: I currently have g94 and mcp89. I'm not really sure what I'm thinking of doing, or if I really want to go through with whatever it is.
03:17imirkin: most of those boards are about $10 each on ebay
03:17imirkin: with shipping
03:17nyef: I was just going to say, "$20-30 if you include shipping".
03:17imirkin: some might be as high as 20 with shipping, but most will be closer to 10
03:18nyef: But yeah, I could probably collect most of the set (less the MCPs) for $200-300 easily.
03:18nyef: Provided that I'm able to pick the models out correctly.
03:18imirkin: that's the real art
03:19imirkin: the quadro's are easiest to go for
03:19imirkin: there's no variation
03:19imirkin: and there's no premium. and they tend to be better built
03:19nyef: How do they tend to differ from the non-quadro versions?
03:20imirkin: e.g. a g84: https://www.ebay.com/itm/HP-NVIDIA-Quadro-FX-370-P588-Low-Profile-Graphics-Card/272940722077
03:20imirkin: i like how it says "low profile"
03:20imirkin: when it's clearly not
03:20imirkin: but you can see the "84" on the bios label, so you know it's a G84
03:21imirkin: the quadro's aren't materially different
03:21imirkin: i think some of them support genlock'd outputs
03:21nyef: The profile is pretty low when you lay it on its side. (-:
03:21imirkin: and various other fancy things
03:22imirkin: mostly they're single-slot width with generally good quality fans
03:22imirkin: which makes it convenient for sticking multiple ones on the same motherboard
03:23imirkin: even a quadro fx 3800 goes for barely over $20 - https://www.ebay.com/itm/Dell-Nvidia-Quadro-FX-3800-1GB-GDDR3-SDRAM-PCI-Express-x16-Video-Card-X9YDW/253760295362
03:28imirkin: g80 - $16.50 for a *whopping* 320mb of video memory. i can hardly believe so much can be put on one single board... https://www.ebay.com/itm/EVGA-NVIDIA-e-GeForce-8800-GTS-320MB-PCI-E-Graphics-Video-Card-320-P2-N815-AR/323350265367
03:36nyef: 320mb? I've had hard drives smaller than that!
03:37imirkin: full-height, no less
03:38nyef: Mmm. ST-412?
03:39imirkin: a bit before my time
03:39nyef: I definitely remember using half-height ST-225 drives.
03:39imirkin: ESDI is where my history starts
03:40nyef: Used to have a 10mb spinning rust that was built onto a full-length 8-bit ISA card.
03:40imirkin: and i don't think i've seen HD's under 20, or maybe even 40mb
03:42nyef: Even a full-height 10mb drive was an improvement over Apple ][ floppies. (-:
03:42imirkin: and for 1 million times more capacity....
03:43nyef: ... oh, wow.
03:44imirkin: although for now, you'll have to settle for 12TB for $400-500
03:45nyef: Right now I have two 1.8TB drives in my primary system, plus a 110GB flash module.
03:45nyef: One of the 1.8TB drives is literally only used for swap space.
03:45imirkin: lol. i swapped out my 2T drives for 8T a little while back
03:45mooch3: eh, i have a 4tb drive
03:46mooch3: which i got used for a steal, only to find out that it had been barely used at all!
03:46mooch3: at least, according to the smart data lol
03:46nyef: (A single 64GB swap partition. There's an LVM volume on the rest of it, but it's empty and not even associated to a VG.)
03:46mooch3: oh wow lol
03:46imirkin: if you're 1 TB into swap space... you might be in trouble
03:46imirkin: (even if you have 1TB of regular ram...)
03:47mooch3: lmfao that's true
03:47nyef: Nah, I've maxed out this system at 32GB of RAM.
03:47azaki: oh, dang. karolherbst pinged out. =o
03:47nyef: I considered maxing out the video card at the latest pre-GM20x card I could get, but it's just not worth it.
03:47azaki: i just sent an email with the next apitrace
03:49imirkin: nyef: gtx 780 ti? those are still expensive =/
03:51nyef: imirkin: Probably that. As an MXM card, no less. Costs more than replacing the rest of the computer!
03:51imirkin: ah yeah. for a regular PCIe card, it's like 150-200
03:52mooch3: i mean, i have a 750ti lol
03:52mooch3: got it for like, 70 bucks?
03:52nyef: I have a gtx 750 ti PCIe, but it doesn't fit the only PCIe machine I have.
03:53mooch3: runs doom 4 on medium at 50 fps somehow
03:53mooch3: oh :c
03:53mooch3: wait, how do you only have ONE pcie machine? lol
03:53nyef: Because everything else is either an AIO, rackmount, a laptop, or PCI.
03:53mooch3: nyef, oh weird o.o
03:54mooch3: i only have two pcie machines, but that's just because i'm poor, and one of them isn't even in use lol
03:54nyef: I mean, if I had a PCI nVidia card, I could drop it into my HPPA system and try to get it to run.
03:54mooch3: that one is also c2d era
03:54mooch3: nyef, uh, i'm pretty sure they didn't make ANY bioses for the hppa
03:55imirkin: most of the pci cards are pre-nv50
03:55mooch3: i know
03:55nyef: Or into my Origin 300, or Origin 350, or maybe an Octane, or a SPARC, or my AlphaServer DS10...
03:55mooch3: i assumed they meant a really old nvidia card lol
03:55imirkin: i have a pci nv34, and a pci nv4a
03:55mooch3: so unless the thing had a generic vga bios, i kinda doubt it lol
03:55nyef: And, sure, the BIOS wouldn't work, but that's what Linux is for. (-:
03:55mooch3: nyef, uh, how would it display ANYTHING then?
03:56nyef: Nouveau can POST a card from scratch, can't it?
03:56imirkin: mooch3: it wouldn't until linux loaded novueau
03:56mooch3: ya know, during the bios of the machine and shit
03:56mooch3: imirkin, lol
03:56nyef: Literally all of my rackmount kit uses serial console.
03:56mooch3: oh :/
03:56mooch3: i guess it isn't a problem then lol
03:56imirkin: pretty much all hardware has serial consoles...
03:56imirkin: except x86 consumer boxes
03:57imirkin: and even those only dropped it recently
03:57mooch3: i mean if that's the way they're using it
03:57nyef: To the point where I got a network-accessible 16-port serial hub, because USB-serial is obnoxious.
03:57mooch3: imirkin, i mean, i wouldn't call 2013 recent in terms of x86 but okay lol
03:57mooch3: my 2013 era pc doesn't have serial ports :/
03:57nyef: (Okay, some of them don't actually use serial console. They *have* it, but the network console is more convenient.)
03:57imirkin: mooch3: we were talking about 10MB hard drives earlier. 2013 is recent.
03:58mooch3: okay, fine lol :p
03:59mooch3: nyef, which serial terminal do you use? do you just use a hardware one, or do you plug the cable into your pc and use a software one?
03:59nyef: I have an EtherLite-16.
03:59nyef: (Okay, I have several. But only one PSU. The PSUs are stupidly rare for some reason.)
04:00nyef: Have to make custom cables for it, but whatever.
04:01mooch3: etherlite-16 is a network card, i'm pretty sure?
04:01nyef: Damnit, mixing things up am I?
04:01mooch3: ether in the name is short for ethernet
04:02mooch3: that's how you remember ;)
04:04nyef: Sure. It's a rebadged Digi EL-16 serial-port hub. Has 17 RJ-45 ports on it, one for ethernet and the other 16 for serial.
04:04nyef: I think I'm mixing it up with the EL-3(?) router, though.
04:04mooch3: AH, okay
04:05mooch3: that makes sense i guess :p
04:06mooch3: oh, btw, if you can, please consider dumping the entire flash chip on all of your nvidia cards. i'll archive the dumps so i can later use them for emulation purposes
04:06mooch3: keep in mind, later might mean several years in this case
04:06nyef: Heh. I'll consider it.
04:07nyef: Okay, it's an ERLite-3 router, apparently. So of course the case for that has gone missing, and most documentation for everything. /-:
09:34karolherbst: mupuf: what kind of fermi cards do you have?
11:41azaki_: it seems somehow my monitors have gone to sleep and aren't coming back, although everything seems to still be running.. i'm ssh'd in and i got the following errors from journalctl:
11:41azaki_: Jul 21 07:29:52 nekomimi kernel: nouveau 0000:01:00.0: DRM: base-0: timeout
11:41azaki_: Jul 21 07:29:54 nekomimi kernel: nouveau 0000:01:00.0: DRM: EVO timeout
11:42azaki_: there's a few dozen of those errors.
11:42karolherbst: ohh, uhm
11:43azaki_: not sure if there's a way to poke them to make them come back. or force a gpu reset (although i'm guessing that would crash X11?)
11:43karolherbst: you could stop X, unbind the vtconsole and reload nouveau
11:44karolherbst: I hit some suspend/resume issues on my laptop as well
11:44karolherbst: but normally skeggsb_ is the right person to poke about issues like that
11:45azaki_: hm. ok, i guess i'll just have to reboot. i was hoping there might've been a way without losing the X session.
11:45azaki_: karolherbst, by the way, i sent an email with some new apitraces, i dunno if you saw it.
11:45karolherbst: yeah, I saw
11:45karolherbst: I only download the small one :D
11:46azaki_: i actually tested them in qapitrace this time, so hopefully they work better than the last one.
11:47karolherbst: I can see the issue on the small one, right?
11:48karolherbst: then that's all I need
11:49azaki_: the main difference is the small file is me playing in fullscreen at 720x480, whereas the large file is me playing in windowed mode, but with the window sized really small.
11:49azaki_: the weird thing is that when doing glretrace, the one i ran fullscreen actually shows up as a large window, which is unexpected.. i would've expected it to be 720x480
11:57karolherbst: azaki_: maybe just the rendering resolution changed
11:57karolherbst: applications can do that
11:57karolherbst: instead of being silly and change the display resolution
11:58karolherbst: really, applications changing the display resolutions are the worst kind of applications
11:58azaki_: i think it did change my display resolution though, that's the weird thing.
11:58azaki_: i noticed my monitor say "invalid mode" for a moment
11:58azaki_: but then it started showing the game after a few seconds
11:59karolherbst: I don't really mind as the apitrace stays small enough to work with it
12:00karolherbst: ahhh, crap
12:00karolherbst: the trace doesn't fit into the cache :/
12:00karolherbst: ~200MB too big
12:02karolherbst: azaki_: mhh, the loading screen picture is high res, but the rendering is crappy
12:17azaki_: i really tried to get it as small as i could
12:17azaki_: the apitrace just seems to slow the game down so much, so it takes a long time to load
12:18azaki_: normally i'm in-game in about a minute or so. but when doing the apitrace it can be up to like 6 minutes before i can get in-game
12:18azaki_: and when i do finally get in, the fps is at like 1-2 fps
12:26karolherbst: azaki_: does it maybe have a benchmark mode or something?
12:32karolherbst: azaki_: also, you don't really need to stay in game soo long. I think the login screen alone would be good enough already, no?
12:33azaki: you mean up to the point where the character is displayed or what?
12:33azaki: i mean i figured you needed me to get in-game to see the issue with the ground textures
12:33karolherbst: ohh wait
12:33karolherbst: I mean, isn't the ground broken on character selection already?
12:34azaki: hm, yeah
12:34azaki: getting to that part takes the longest though.
12:34karolherbst: the trace size increases with each frame spend in game.
12:34karolherbst: all the frames before are not really big
12:35azaki: like about 80% of the time taken to load is to get to character select screen
12:35azaki: and then maybe the rest to get in-game
12:35karolherbst: yeah, but I am sure the trace will be significant smaller
12:36karolherbst: with the trace being cached it takes around 80 seconds to replay, it just takes super long if the file isn't cached anymore
12:40azaki: ok, i'm going to give this another try
13:09azaki_: nouveau 0000:01:00.0: systemd-logind: nv50cal_space: -16
13:13azaki_: hm, ssh'ing in and killing the game somehow brought the system back to life
13:37karolherbst: azaki_: uhh ohm. Yeah, that shouldn't happen
14:08nyef: Hrm. Can I have two nVidia cards in one system, and if so is it deterministic which one gets to be the initial console?
14:09vadimt: Hi! Is the GTX 1080 supported in kernel 4.17.6? I'm trying to load the module, but get absolutely nothing. No errors, no messages in the log, nothing at all. The module does load successfully though. I see nothing of interest in /sys/module/nouveau, just the very basic module stuff
14:10JodaZ: is power management really the hardest part about steering a nvida chip???
14:10nyef: Okay, the NV42 came up as BIOS console...
14:11nyef: ... and as fb0 and Linux console, and the G94 registered as fb1.
14:11JodaZ: i mean is it even hard to get a way to set clocks for the two states that would enable 90% usability? i.e. full on and idle 2d usage
14:18nyef: JodaZ: There are two cases here. One is for chipsets that we don't (yet) know how to power-manage. That's a reverse-engineering problem. The other is for chipsets that require signed firmware in order to power-manage. That's a political problem.
14:18nyef: Anything GM20x and newer falls under the "signed firmware" case.
14:19nyef: Anything else falls under the "reverse-engineering" case.
14:19azaki: nyef: one thing i've been wondering is how has nvidia managed to make their signed firmware 'unrippable' so to speak? i mean the firmware is there in the proprietary driver right?
14:20azaki: i'm aware this wouldn't exactly be a very orthodox solution, and would never be "out of the box", but even so, i'm just curious
14:20nyef: It's not that it's "unrippable", it's that it's copyright, so not *distributable*.
14:20azaki: because i recall with previous cards that you still need to extract firmware for the video decode stuff
14:21karolherbst: JodaZ: we can't set the clocks
14:21karolherbst: like literally
14:21nyef: ... oh, nuisance. I'm going to have to add video decoding to my possible projects list, aren't I?
14:22karolherbst: JodaZ: you have to upload signed code to make it work and the code isn't redistributable
14:22JodaZ: nyef, what about allowing people to place the blobs?
14:22karolherbst: and then we need to figure out the interfaces which could literally change every driver version
14:23karolherbst: and the problem with the secure boot stuff is more like not that the firmware loading fails, but the entire GPU goes down
14:24JodaZ: how do you not have a license to that code if you own the gpu?
14:24azaki: karolherbst: i just sent the email with the new apitraces. one is 22GB and just up to char select screen, the other is 27GB and goes all the way in-game.
14:24karolherbst: JodaZ: if you buy software, are you legally allowed to share the binaries?
14:24JodaZ: no, but i am allowed to use it
14:25nyef: It doesn't matter if you have a license or not if you can't usefully run the code.
14:25karolherbst: sure, but you can't share it
14:25karolherbst: anyway, we have some kind of extractor working
14:25JodaZ: that does not matter
14:25azaki: once again i tested them in qapitrace, so they should work.
14:25karolherbst: but still people have to figure out the interfaces
14:25JodaZ: karolherbst, exactly
14:25karolherbst: and then it is mainly a time issue
14:25karolherbst: well, do you have time to figure out the interfaces?
14:26karolherbst: azaki: thanks!
14:26JodaZ: nvidia isn't helping?
14:27karolherbst: there is no documentation about the interaces
14:27azaki: JodaZ: seriously? nvidia is probably the least helpful gpu vendor.. =\
14:27karolherbst: well we got the graph firmware at least
14:27karolherbst: so we can do acceleration
14:27azaki: i think even a few of the mobile ones have been a bit more helpful lately.
14:27karolherbst: I think nvidia just lacks behind AMD in terms of open source support
14:28azaki: and the mobile guys are notoriously horrible.
14:28karolherbst: that's all
14:28karolherbst: and may take 5 or 10 years to develop
14:28karolherbst: or maybe never
14:28karolherbst: we don't know
14:28karolherbst: azaki: they provide _some_ documentation
14:28karolherbst: and quite a lot for display related stuff
14:28karolherbst: which is helpful
14:29karolherbst: and getting the basics correct is kind of more important right now
14:29JodaZ: can't you just blindly capture whatever the official drivers send to the card for full throttle and idle mode and just use that?
14:29karolherbst: what matters performance if 50% of all users can't use their displays
14:29karolherbst: JodaZ: maybe?
14:29karolherbst: maybe not
14:29karolherbst: things differ on each GPU kind of
14:29karolherbst: and there is the vbios we have to parse most likely
14:29karolherbst: all that stuff isn't _that_ trivial
14:30karolherbst: reclocking on kepler is already quite insane
14:30JodaZ: can't use their display?
14:30azaki: i've noticed a pattern with stuff like cuda, gameworks, etc, it seems like nvidia really likes the strategy of providing value on the software side of things, as opposed to amd/intel which seem to care more about the hardware, and don't mind as much if there is open docs. nvidia seems to dislike anything that would allow other vendors to achieve software parity, so they don't seem to be fans of something like the gallium approach, where everyone shares
14:30azaki: a large chunk of userspace code.
14:31JodaZ: unless you have accel you won't get the enthusiast users which you actually want to test stuff
14:31azaki: sorry, i didn't know i had typed that much
14:31karolherbst: JodaZ: well, we have accesl
14:31karolherbst: just not as fast
14:31karolherbst: and performance is not the usual reason people work on open source drivers
14:32JodaZ: its the usual reason people use drivers
14:32karolherbst: like I didn't start on nouveau because the driver was fast or something, but because I want an open driver and there were bugs getting proper speed
14:32karolherbst: anyway, if somebody with quite some time and motivation would come around, that person could work on that
14:33karolherbst: I am kind of opcupied with other tasks
14:33karolherbst: which is sad, but I try to find time for smaller things
14:33vadimt: Say, does the project have a Patreon or something? I'd gladly contribute a bit
14:33azaki: JodaZ: if you really want open source drivers that're more vendor-supported, there's always amd and intel graphics.
14:33karolherbst: vadimt: not really. There are some RH paid developers though
14:34karolherbst: vadimt: hardware donations are always appreaciated though
14:34karolherbst: especially GPUs with rare bugs
14:34JodaZ: i am actually kinda impressed with how well intel fills the just good enough market
14:35karolherbst: the bigger problem is like that intel and AMD throw more than 10 times the devs on their drivers
14:35nyef: I didn't start on nouveau because the driver was fast, I started on nouveau because it flat-out didn't work right on my mac, and there was (still isn't, AFAIK) no PPC binary driver available. More recently, it's been things like working towards feature support that the hardware can do but the binary driver won't, or stability issues on specific hardware.
14:35karolherbst: so of course nouveau lacks behind ;)
14:35azaki: damn, ppc macs. oldschool
14:36nyef: I'll admit, I haven't done much with PPC macs and nouveau in a few years now, but I still have some.
14:36vadimt: karolherbst: well, that's the thing, new hardware is expensive. I do have some old stuff lying around that I'm about to throw out, what do you need? I will take a look
14:36JodaZ: so hows the current model for clocking and such?
14:36karolherbst: vadimt: mhh, hard to tell. Boils down to try nouveau on those GPUs and find broken stuff
14:37karolherbst: JodaZ: what do you mean by that?
14:37azaki: JodaZ: i dunno if you've kept up on the recent situation with amd, but they have full GL 4.5 now on radeonsi (GCN) cards, and performance is superior to the proprietary driver in pretty much all tests.
14:37JodaZ: if someone were to implement it for a card, would they implement some set_performance(int percent) function or what?
14:37vadimt: karolherbst: is the GTX 1080 supported right now, by the way?
14:37karolherbst: vadimt: it should "run"
14:37karolherbst: JodaZ: not really, you look more at the clocks
14:38karolherbst: anyway, the stuff is more complicated than just set a percentage
14:38vadimt: karolherbst: that's good to know, as of what kernel version? Because here absolutely nothing happens when I load the driver. No messages at all
14:38JodaZ: well, talking to the card is, but is the performance demand side, the scheduler so to speak of done?
14:39karolherbst: vadimt: uhm...
14:39karolherbst: JodaZ: it really ist more complicated than that
14:39karolherbst: even calcuating the load isn't as trivial
14:41karolherbst: I have some patches to do the dynamic reclocking bits, but they need a heavy cleanup and I don't have that much time to just do that right now
14:44vadimt: karolherbst: I'm on kernel 4.17.6 here, on Fedora. It modprobes fine, just absolutely nothing happens, not even "module loaded", or "I don't recognize this card". It's rather confusing.
14:44karolherbst: vadimt: mhh, weird
14:45karolherbst: vadimt: is maybe nvidia loaded?
14:45vadimt: karolherbst: no, completely uninstalled
14:45karolherbst: vadimt: dmesg
14:45karolherbst: but yeah, this is a bit weird
14:50JodaZ: karolherbst, but that piece of it seems to be kinda independant of any politics or re issues, no?
14:50vadimt: karolherbst: https://pastebin.com/CdfJ1zD3
14:51karolherbst: JodaZ: well, some stuff needs to be reverse engineered and we can't do it on hardware with signed firmware either
14:52karolherbst: vadimt: try to load nouveau via insmod
14:53karolherbst: vadimt: ohhh
14:53karolherbst: vadimt: you have nomodeset set
14:53karolherbst: you can't do that :)
14:53karolherbst: and why do you have "video=1024x768@60" set?
14:54vadimt: whoops. That makes sense. The nvidia driver was crashing on me, so I was trying to get back to nouveau
14:54vadimt: karolherbst: the video= thing was just because I was screwing around with stuff earlier seeing if anything made any kind of difference
14:56vadimt: well, this already looks better
15:00karolherbst: azaki: nice, the short trace replays now in a minute :)
15:07vadimt: karolherbst: thanks a lot, X is working now. By the way, maybe the driver could say something about nomodeset in dmesg? I was trying to switch back from nvidia and didn't notice it was there
15:08karolherbst: why do you have it set in the first place?
15:08karolherbst: evevn with nvidia you don't need it
15:08karolherbst: but yeah
15:08karolherbst: there was some discussions about adding an error or something
15:14vadimt: karolherbst: not really sure at this point, this install is something like half a decade old and has been upgraded since
17:33nyef: Hunh. Looks like I have a G86 as well. I don't think that that helps at all at the moment.
17:34karolherbst: azaki: mhh, there seems to be still some wine issues left
17:41imirkin: nyef: bios controls which video card is "primary" and receives vga mmio and ioport space
17:43karolherbst: azaki: you could ask some wine devs if they provide you with a patch or something. Anyway, I don't see more than 32 samplers being used allthough some shaders clearly require more
17:44nyef: imirkin: Right, which means that, unless the bios is demented (such as picking the newest card instead of the card in the lowest-numbered slot), it should be reliable.
17:46imirkin: nyef: yeah, it should be consistent. you can usually switch it up too, based on pci slot selection.
17:47imirkin: i.e. many bios's will have the option of selecting which card is the main one
17:52nyef: ... And the reason that my GM107 isn't particularly useful is that the external power connector fouls the shroud that covers the card slots in this case. Good to know.
17:53nyef: (I can use it, but I can't close the case with it installed.)
18:22ilmostro: kernel 4.17.8 exhibits interesting issue on AMD Ryzen system with nouveau driver. When xscreensaver starts, the GPU locks up with following messages in journal logs https://ptpb.pw/oxTA.png
18:23nyef: ilmostro: Card type?
18:24ilmostro: nyef: NVIDIA Corporation GM107 [GeForce GTX 750 Ti] (rev a2)
18:24nyef: ilmostro: And do you know if this is a regression or not?
18:25ilmostro: nyef: I believe it is, since I have no such issues with 4.17.6
18:25ilmostro: I haven't tested 4.17.7
18:25imirkin: does the log show any differences in nouveau?
18:26ilmostro: hmm, I'll have to dig around a bit; hold on :)
18:26imirkin: no nouveau stuff since 4.17.6
18:26imirkin: my guess is it's unrelated to the kernel
18:26ilmostro: not necessarily; e.g. recent issues with AMD-SEV leading to DMA (or DRM) issues
18:27imirkin: yeah, but with SEV you end up having to user a bounce buffer for a bunch of stuff
18:27imirkin: so... really not ideal
18:28nyef: There's some i2c stuff in 4.17.7, including tegra-related i2c, but that doesn't seem likely to cause a pfifo issue.
18:28ilmostro: imirkin: true; I'm just saying that there might be things that break nouveau-related stuff even if/when there are no changes in nouveau development :(
18:28imirkin: ilmostro: sure
18:29imirkin: but much more likely is that you didn't just update the kernel
18:30imirkin: but also other things. and one of those other things is tickling an issue in the nouveau dri driver
18:30ilmostro: So, for the time being, it should be good enough to make you guys aware until/unless other users show up with same problems. I'm not sure if it's specific to the AMD Ryzen platform or if it affects a wider set of users
18:30imirkin: which in turn causes the card to hang, like pretty much any problems do
18:31imirkin: ilmostro: ah ok, thanks. pretty everyone has problems with nouveau ... i'm thinking of starting to recommend distros don't ship nouveau_dri by default
18:31ilmostro: imirkin: I'm on gentoo; while user error is a feasible explanation, I'm quite sure that I did not make an error in not upgrading the kernel :)
18:32imirkin: ilmostro: my point is that you didn't *only* upgrade the kernel
18:32imirkin: i.e. there were other changes
18:32nyef: ... Are you running vanilla-sources, gentoo-sources, or something else?
18:32ilmostro: imirkin: ahh, I see; the nouveau driver has been at 1.0.15-r1 for quite some time. gentoo-sources
18:32ilmostro: not vanilla
18:33imirkin: ilmostro: what about user applications?
18:33imirkin: nowadays everyone thinks it's hilarious to use GL to accelerate the drawing of 1 pixel
18:33nyef: You might check for weird patches, then?
18:33ilmostro: imirkin: I'll dig around a bit for user applications. For the time being, I'll wait to see if other users report issues such as this
18:34imirkin: ilmostro: hangs with nouveau? been reported since the dawn of time.
18:34ilmostro: nyef: Gentoo patches are quite stable, AFAIK
18:34imirkin: since before nouveau existed ;)
18:35ilmostro: hmm...some reports are more substantial than others, though. Thanks for the tips, btw :)
18:35nyef: I didn't say that they aren't stable, I said "look for some". It could be that there's a gentoo patch to nouveau that we're not seeing, and that might be causing this issue. Or not. It's something to look for, at least.
18:35ilmostro: nyef: indeed; thank you
18:35imirkin: ilmostro: your issue is a very vanilla one ... copy engine tried to read page that wasn't mapped.
18:35imirkin: happens all the time
18:35imirkin: (well, not _all_ the time...)
18:36imirkin: we try to recover, and invariably fail
18:36imirkin: i almost prefer the days when there was NO recovery... at least things kinda limped along and if you killed the right thing fast enough, all would be well
18:36ilmostro: hmm, since it happens with the screensaver, I'll see if there's something to be done there on my system
18:37imirkin: ilmostro: stop using GL screensavers :)
18:37imirkin: i use xlockmore. no problems.
18:38ilmostro: imirkin: this is a recent phenomenon that I tried to adopt; in hopes of preventing lcd burn-in(?)
18:38ilmostro: I'll go back to the way I used to be :D
18:38imirkin: ilmostro: you can use screensavers. just don't use GL ones.
18:38ilmostro: imirkin: ahh, I see; I'll check out xlockmore and others
18:39imirkin: (actually xlockmore does include some gl ones, but you don't have to use them. i personally like the "wator" one)
18:39imirkin: (and of course "matrix" is an old favorite)
18:51ilmostro: hmm, let's see if I can find some in xscreensaver that don't use GL
19:32azaki: karolherbst: yeah, the wine bug i linked in the first email talks about how they have to implement ARB_bindless_texture to get that working.
19:33karolherbst: azaki: well
19:33karolherbst: azaki: this would be one solution
19:34karolherbst: but if the driver supports more than 32 samplers they wouldn't have to
19:36azaki: karolherbst: oh, hm. i thought i saw people saying this was an opengl limitation or something, and that bindless_texture was how you worked around it
19:36azaki: maybe i've totally misunderstood how it all works
19:38karolherbst: I am not 100% sure, but I think drivers just report 32 max. I dont' really know what the actual reason for this is
19:38karolherbst: imirkin gave me the impression that we might be able to advertise more than that, or could at least do that inside the driver
19:38karolherbst: maybe we could even have a rather small extension allowing that
19:39karolherbst: or maybe we could add an exception for wine or something
19:43azaki: karolherbst: so i should ask the wine guys if they have patches for their shader backend that can use more samplers, if the driver advertises more?
19:46karolherbst: yeah, something like that
19:47karolherbst: I mean, they can also just use bindless_textures, I don't really care either way, just a matter how much work that would be on both ends
20:00azaki: yeah based on what they said in the bug it seems using bindless_texture would be a huge rewrite of the shader backend
20:00azaki: anyways, i asked in #winehackers
20:01azaki: we'll see if i get a response
21:57azaki: karolherbst: looks like i might have to ask again on a weekday, someone responded that there wasn't a lot of people active in the channel on weekends, however, they also said this:
21:57azaki: <strfllw> azaki: but we'd probably use extra samplers if e.g. Nouveau advertised them
21:57azaki: <strfllw> still, in the longer term, bindless is probably the way to go
23:23karolherbst: imirkin: any idea why bumping the max samplers to 42 "works", but going above that causes some tests to fail?
23:24karolherbst: I think I forgot something stupid to increase as well
23:56karolherbst: ohhh, I see
23:57karolherbst: the test things there are 384 bindings available
23:58karolherbst: as GL_MAX_COMBINED_TEXTURE_IMAGE_UNITS is now 384
23:58karolherbst: and with 42 we are at 252