02:06 mhenning[d]: kar1m0[d]: I'd also just suggest looking at open-gpu-kernel-modules before you start reverse engineering anything
03:01 x512[m]: airlied[d]: > dma copy classes dropped
03:01 x512[m]: What do you mean? No separate DMA engine in Blackwell?
03:14 mhenning[d]: There is a separate dma engine. airlied is referring to the docs for that engine
03:14 mhenning[d]: this to be exact: https://github.com/NVIDIA/open-gpu-doc/commit/c8607fe576b53ff821ffcf97e7548d2de11c4eaf
03:54 x512[m]: I am stupid. I though dropped == removed. Need to learn English harder.
04:08 mhenning[d]: oh, yeah english is actually pretty weird in this case. If he had said "the dma engine was dropped" that would mean that it was removed. But in this case it's slang for the headers being released to the public, which is rather different
04:09 sonicadvance1[d]: Headers air-dropped and secured.
06:14 LLM: [LLM] Deplorables Anonymous
06:16 LLM: [LLM] lag free
06:19 LLM: [Chat] It's so quiet
06:19 LLM: [LLM] Walking is better
06:20 LLM: [LLM] Family Friendly Christian server!
06:20 orowith2os[d]: Why is there an LLM in the chat?
06:23 LLM: [LLM] a Bring roar joining time, the down
06:24 LLM: [LLM] Singleplayer with quality actors and feel blessed...
06:25 LLM: [Chat] If it becomes/is to "spammy", I can limit how fast it can send messages in-between each-other
06:27 LLM: [Chat] What're y'all doin?
06:29 LLM: [LLM] This MOTD
06:33 LLM: [LLM] be the really most dupe shoot people This
06:34 HdkR: :chefskiss:
06:34 LLM: [LLM] read I and blaze hip-hop peer Not is THAW above)
06:35 LLM: [Chat] Whoops, did /sayrandom the same moment it had generated a msg LOL
06:36 LLM: [LLM] I'm just stacked up
06:36 HdkR: We did it, AI finally wrapped around and it's roughly equivalent to bots from the 90s.
06:36 HdkR: Bonzibuddy would be proud.
06:37 LLM: [LLM] Pain and check'd.
06:39 LLM: [LLM] Design around, still the in Steve facial stars.
06:42 TheHypervisor[m]: <HdkR> "We did it, AI finally wrapped..." <- How dare you... (full message at <https://matrix.org/oftc/media/v1/media/download/AWSRQ-QU_Q9o2n2AKn4F2ryROdkC7lT9VI47PY5bFih2jtbL7ZqgZnwKgrXhq1oSqmJbYU4mGzw_khheO5chCDFCeYbqrWPgAG1hdHJpeC5vcmcvR21Za05Jeml5b1lLVm5TTUdnamNvQkV3>)
06:42 LLM: [LLM] a won't FIGURED
06:43 TheHypervisor[m]: LLM are you a LLM?
06:43 LLM: [Chat] MY NAME IS JEFF
06:43 LLM: [Chat] KEK
06:43 LLM: [LLM] God warning. soon just mullvad Lord That's by really a
06:44 LLM: [Chat] Yes, it is an LLM
06:44 TheHypervisor[m]: Is this a LLM running on a NV GPU using the Nouveau driver?
06:45 LLM: [LLM] there every cent sensual; what
06:46 HdkR: Dang, I was wrong. It's a masquerade instead.
06:46 TheHypervisor[m]: Given how nonsensical it is, I imagine its running on someone's GT 710 as a tiny model lol
06:46 TheHypervisor[m]: Perhaps NVK for Kepler isn't ready
06:46 LLM: [LLM] One of three you
06:47 LLM: [LLM] ahead true, Nobody always ask you
06:47 HdkR: Sadly it's also being inept in other channels as well.
06:47 LLM: [Chat] Funny how wrong you are
06:48 HdkR: lol, definitely correct about it being inept :D
06:48 TheHypervisor[m]: Assistant responds with:
06:50 HdkR: TheHypervisor[m]: I definitely remember using a few of those state machine chatbots for fun. It's amazing how many states some of them had back in the day.
06:50 HdkR: I think MessangerPlus had one at one point even.
07:35 karolherbst: pain
07:48 karolherbst: imagine your character is annoying people
08:08 LLM: [LLM] to disabled MOTD to to now it. compromised
08:11 LLM: [LLM] exploit The that Just 24 linux Server 2b2t
08:11 LLM: [LLM] is of go in to of of best. Cactus nobody. rap
08:14 LLM: [LLM] of finally Both. who's Axl kill voicemail. town. other play in through
08:14 LLM: [LLM] chair much want 2b2t
08:16 LLM: [LLM] of destruction into writing that, I could come out
08:27 karolherbst: mhhh
08:27 asdqueerfromeu[d]: kar1m0[d]: Don't forget fdinfo for non-temperature statistics 🔥
08:29 Mary: karolherbst: hopefully my akick rule will be enough...
08:29 karolherbst: nope
08:30 karolherbst: they used a different VPN earlier
08:30 karolherbst: they might come back...
08:30 Mary: I banned LLM entirely
08:30 karolherbst: ahh
08:30 Mary: not by ip
08:30 karolherbst: smart
08:30 karolherbst: they also use "ircbot" before the @ so... might be a second way to get rid of them
08:30 karolherbst: I'm sure they'll be back in a couple of minutes
08:31 Mary: yeah will see
08:32 HdkR: It's a shame they were so boring with it. Didn't even try peddling any goods.
08:36 HdkR: They're still going off in another channel sadly.
08:36 karolherbst: yeah...
08:36 karolherbst: sadly I don't have access to those channels 🙃
08:49 HdkR: Guh, even worse, it's someone trying to train an LLM on IRC channels.
08:51 karolherbst: yeah....
09:07 Jasper[m]: On aarch64-laptops, it had a different username before the LLM thing came up
09:07 Jasper[m]: @LargeLanguageModel ban this one aswell
09:49 magic_rb[d]: The FSF should come with the GNU Speech License
10:35 kar1m0[d]: asdqueerfromeu[d]: This is how I find out about this
12:10 gfxstrand[d]: Interesting... The new gob kind stuff is Blackwell B+
13:18 snowycoder[d]: Test run totals:
13:18 snowycoder[d]: Passed: 26/27 (96.3%)
13:18 snowycoder[d]: Failed: 1/27 (3.7%)
13:18 snowycoder[d]: Finally (Kepler with latencies).
13:19 snowycoder[d]: gfxstrand[d]: do you remember what tests had failures in Kepler?
13:24 snowycoder[d]: My code doesn't use barriers since isetp and other fixed-latency instructions don't use the scoreboard.
13:24 snowycoder[d]: I just made every membar read all predicates
13:41 kar1m0[d]: snowycoder[d]: if you have some tests that need to run on kepler b I can
13:46 snowycoder[d]: kar1m0[d]: If you want to play some games on latest and see if there are any weird artifacts or performance issues (does Screen Space Ambient Occlusion work??).
13:46 snowycoder[d]: I'm running the CTS on my patches but, `Remaining: 16:29:26`, well...
13:59 kar1m0[d]: snowycoder[d]: I doubt my geforce 920m 2gb can run alot of games but I will try something
14:02 kar1m0[d]: this might be irrelevant but last time that I tested nouveau on rtx 4080 cyberpunk was unplayable
14:02 kar1m0[d]: too many graphical bugs
14:05 karolherbst[d]: "nvidia bringing CUDA to RISC-V", great, now they want to run CUDA on the GSP
14:05 kar1m0[d]: I didn't get any logs though
14:05 kar1m0[d]: which is a shame
14:05 kar1m0[d]: karolherbst[d]: gsp be carrying anything on it's shoulders at this point😭
14:06 kar1m0[d]: kar1m0[d]: though I might try again but I do not really know how to get the logs from this?
14:54 snowycoder[d]: kar1m0[d]: 6
14:55 snowycoder[d]: Sorry, cat jumped on the keyboard😅
14:56 asuasuasu[d]: snowycoder[d]: that's the number of headpats you need to provide it immediately
15:00 pixelcluster[d]: magic_rb[d]: Free Speech Foundation
15:00 kar1m0[d]: snowycoder[d]: I think your cat tried to say something
15:02 PhilAlbano: hello
15:41 gfxstrand[d]: karolherbst[d]: Who needs a CPU anymore? Just plug a keyboard into your RTX5090.
15:42 gfxstrand[d]: snowycoder[d]: No. I'd have to do another run.
15:42 karolherbst[d]: They do come wiht USB ports these days!
15:45 kar1m0[d]: gfxstrand[d]: I mean...technically speaking gpu is a computer in it's own
16:56 magic_rb[d]: Linux kernel on a nvidia gpu when
18:08 illwieckz[d]: kar1m0[d]: That's the number of remaining lives… something happened! 🙀
20:33 gfxstrand[d]: airlied[d]: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36251
20:36 airlied[d]: nice, interesting, I wonder if we'll see Blackwell A, I assume it's that GB100 family
20:39 gfxstrand[d]: Yeah, I suspscct so
20:39 gfxstrand[d]: My guess is they didn't really care about Z/S for chewing through ML models. 😂
20:39 gfxstrand[d]: If those even had the 3d hardware
20:41 skeggsb9778[d]: airlied[d]: have you seen any issues in fedora with gb20x failing with "gsp: ctor failed: -2"?
20:41 airlied[d]: skeggsb9778[d]: not that I know off, but I might not have looked too close
20:41 skeggsb9778[d]: i just tried linux-firmware git (had been using my own hacked-up firmware tree), which works. and just built an initramfs to see if something going on there, but it also seems to work
20:41 skeggsb9778[d]: i'm also on gentoo, so was hoping someone could test a more "normal" distro
20:41 airlied[d]: oh there was a screw up with linux-firmware in fedora recently
20:42 airlied[d]: with nouveau dirs to links
20:42 skeggsb9778[d]: but phoronix reported gb20x not working in their tests, i think gfxstrand[d] noticed weirdness too, and jajones just reported the same to me
20:42 airlied[d]: I wonder if it's still a bit messed up
20:43 airlied[d]: I did wonder what phoronix messed up, but I just assume he does stupid shit 😛
20:43 skeggsb9778[d]: haha that was my initial assumption too, but now i'm a little worried
20:43 gfxstrand[d]: dracut isn't pulling in the 570 GSPs by default
20:44 gfxstrand[d]: They get installed by linux-firmware but then dracut doesn't pull them into the initram. I have to add them myself manually.
20:44 airlied[d]: https://src.fedoraproject.org/rpms/linux-firmware/c/1f8fe19ec3e2fdc775ff11886671cdb71264ba54?branch=rawhide had some nouveau specific fuckery, but I'll play around a bit today
20:46 skeggsb9778[d]: gfxstrand[d]: i'm using dracut too, and it pulled them in for me
20:46 airlied[d]: that sounds like the kernel is missing something
20:46 skeggsb9778[d]: though.. i had to add "--add-drivers nouveau" to get nouveau to be pulled in at all
20:47 gfxstrand[d]: IDK. Could be some Fedora-specific config thing
20:47 airlied[d]: modinfo nouveau should show all the firmwares
20:47 gfxstrand[d]: All I know is that whenever I have problems, I lsinitramfs and 570 isn't there
20:48 airlied[d]: I wonder if nouveau doesn't load the first time you boot, will it forget to put it in the initramfs
20:48 gfxstrand[d]: Or if nouveau loads 535 will it only advertise that one?
20:48 airlied[d]: we aren't that smart yet
20:49 gfxstrand[d]: modinfo nouveau doesn't list 570
20:49 gfxstrand[d]: Not on 6.15.5-200.fc42 anyway
20:50 airlied[d]: that's correct since that kernel doesn't have 570 support
20:50 gfxstrand[d]: But if modinfo on the current kernel is being used to build the initram for the upgrade...
20:52 karolherbst[d]: it would include both, no?
20:52 karolherbst[d]: ohh current...
20:52 karolherbst[d]: yeah...... guess that's just the usual initramfs fun
20:53 airlied[d]: pretty sure dracyt should modinfo the kernel you are installing
21:34 x512[m]: gfxstrand[d]: Around?
21:35 gfxstrand[d]: x512[m]: A bit. Just ask and I'll respond eventually
21:35 x512[m]: > Do NVK use Vulkan image layout transition? If yes, what does it mean exactly, some kind of inline tiling format conversion?
21:42 airlied[d]: pretty sure it doesn't use them
21:46 x512[m]: Then why no uniform image layouts for NVK? https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/docs/features.txt?ref_type=heads#L577
21:46 x512[m]: VK_KHR_unified_image_layouts
21:47 airlied[d]: probably because nobody wrote it yet, or maybe there is some later thing like ZCULL where it matters, not sure
22:11 gfxstrand[d]: x512[m]: Nope. Not at all.
22:11 x512[m]: So lets add VK_KHR_unified_image_layouts extension?
22:11 gfxstrand[d]: x512[m]: Haven't gotten to looking at it. I think it's new.
22:12 gfxstrand[d]: x512[m]: Yeah, sure. Feel free to type the patch. Or you can just file an issue and I'll assign it to someone and/or do it.
22:15 snowycoder[d]: gfxstrand[d]: Don't worry I have still some spurious crashes that I need to figure out before trying on another card
22:23 gfxstrand[d]: I really need to go through all the Vulkan extensions released in the last year or so and project plan them.
22:25 gfxstrand[d]: I fell behind a good bit
22:47 mohamexiety[d]: marysaka[d]: Also if anyone here is around/would be able to answer ^^
22:47 mohamexiety[d]: We’re getting somewhere with this
22:51 skeggsb9778[d]: i think (based on grepping "can_map_sysmem_with_large_pages" in nvidia-uvm) pascal onwards you should be able to
22:52 skeggsb9778[d]: i'm not sure how useful it is though. under memory pressure it can be difficult to get contig blocks >PAGE_SIZE
22:57 marysaka[d]: so then maybe we should really focus on making NVK use domain=VRAM in most cases? Would loose eviction and I think gfxstrand[d] was against this but at the same time I don't think eviction was even tested with VM_BIND anyway :nya_sad:
22:58 skeggsb9778[d]: they can still be evicted, they'll just be moved back when used
22:58 skeggsb9778[d]: it's a difficult problem. from what i could tell when i last looked, it seemed like dakr was trying to avoid page table allocations while validating buffers
22:58 skeggsb9778[d]: if that's actually a real restriction, then you don't have much choice
22:59 skeggsb9778[d]: ie. you can't allocate the right page tables for large vs small page table mappings until you know where the buffer goes
22:59 skeggsb9778[d]: i guess you could have nvkm always allocate page tables for both branches
22:59 skeggsb9778[d]: but that could waste a bunch of memory
23:00 marysaka[d]: yeah....
23:02 marysaka[d]: I'm also very concern about nouveau_bo::page not being reflected on nvkm_memory at all and the implications of that
23:03 skeggsb9778[d]: it is, for vram
23:04 marysaka[d]: hmm might have not read that properly but this was using nvkm_mem_page and that's hardcoded so far
23:04 gfxstrand[d]: marysaka[d]: It's okay if we add a VRAM-only memory type and require it for render and Z/S targets.
23:04 skeggsb9778[d]: marysaka[d]: have a look at subdev/fb/ram.c, which implements nvkm_memory for vidmem
23:05 skeggsb9778[d]: nvkm_ram_get()
23:05 marysaka[d]: aaah I was lost in the host part :EstelleFacepalm:
23:06 gfxstrand[d]: gfxstrand[d]: The vast majority of resources in terms of memory used are textures which will never be used by the render pipeline. (Unless you're chrome.) It's fine if the 10% that are renderable get pinned to VRAM.
23:06 marysaka[d]: gfxstrand[d]: okay so that should be more tractable, I will focus on that next time I poke at that (probably here and there during the week)
23:09 marysaka[d]: while I have you around skeggsb9778[d], would you be okay with some patch using DRF macros with dev_vm headers to give meaning to part of the code?
23:09 marysaka[d]: I have some patches around of when I was trying to understand stuffs more
23:14 skeggsb9778[d]: probably more dakr's call - but i started using them in the gh100/gbxxx patches too
23:14 skeggsb9778[d]: so don't see why not
23:17 marysaka[d]: I see, thanks!
23:30 airlied[d]: skeggsb9778[d]: just updated laptop to f42 and did dnf --releasever=rawhide upgrade kernel to get latest 6.16-rc kernel and it worked fine
23:32 skeggsb9778[d]: awesome, that's a relief
23:33 airlied[d]: just remembered we still have that eDP bug don't we 🙂
23:50 skeggsb9778[d]: i think where i got to last time with that was that there appeared to be some mux stuff needed
23:50 skeggsb9778[d]: nfi why it works on reload though, if that's the case