02:31_lyude[d]: yeah that's kinda what I came up with too after thinking about it for a bit
03:13karolherbst[d]: phomes_[d]: maybe something gets messed up with the state... but I think a dump of the push buffers might show what's going wrong there, or it's just timestamps being weird
10:23phomes_[d]: karolherbst[d]: yes I am working a way to get just the push dump from that pass
10:31wanless[m]: Get in touch with this platform for greatness you’ll definitely thank me later ℹ️❤️ 👇🏻👇🏻👇🏻
10:31wanless[m]: https://t.me/+pa-CiKYv9-5lNWM8
12:02karolherbst[d]: phomes_[d]: is it the first one? But I think just the beginning of the dump would be helpful, it's probably just a big dump tho
12:02karolherbst[d]: like dunno.. maybe the first 3000 commands would be good enough
12:21soraldpayton: I got the neck injury and tendons injury after the impact of getting downed , so the chip has programmable effect on my neck, or was it just some bruise to the tendon/ligaments, i am unsure if i can turn this chip into my advantage , but the pain part is not so horrible anymore, there is no sensing in my legs
12:24soraldpayton: the scary part is i am religious in some way , certain things i do not allow myself to be done, such as accusing others when i am unsure, so mannerov asked if i am ever going to play something again, there is a technical opportunity that i could recover some strengths being somewhere 45 years of age
12:30karolherbst[d]: mhhh seems like the prepass with the better occupancy info does something well
12:30karolherbst[d]: but also...
12:30karolherbst[d]: I'm hitting that bug again 🙃
12:31karolherbst[d]: mhenning[d]: fyi, wiring up shared mem + block size for prepass:
12:31karolherbst[d]: Totals from 1096 (0.55% of 198421) affected shaders:
12:31karolherbst[d]: CodeSize: 21257840 -> 21166560 (-0.43%); split: -0.58%, +0.15%
12:31karolherbst[d]: Number of GPRs: 50448 -> 59760 (+18.46%); split: -0.02%, +18.47%
12:31karolherbst[d]: Static cycle count: 9694653 -> 9660710 (-0.35%); split: -1.23%, +0.88%
12:31karolherbst[d]: Spills to reg: 10022 -> 10166 (+1.44%); split: -0.23%, +1.67%
12:31karolherbst[d]: Fills from reg: 6013 -> 6143 (+2.16%); split: -0.23%, +2.39%
12:31karolherbst[d]: Max warps/SM: 33953 -> 33837 (-0.34%)
12:31karolherbst[d]: probably should figure out the bug where prepass hurts occupancy
13:28esdrastarsis[d]: marysaka[d]: Looks like the mmu fault isn't constant, you need to run the tests a few times
13:28esdrastarsis[d]: I can reproduce with `deqp-vk -n "dEQP-VK.api.buffer_view.access.dedicated_alloc.*"`
13:29esdrastarsis[d]: Running only the problematic test doesn't trigger the issue
13:32esdrastarsis[d]: I'll test with compression disabled in nvkmd and then I'll create the issue report
13:36marysaka[d]: it could be large page if it does trigger with compression disabled too
13:36marysaka[d]: as the kernel side will decide to use larger pages if the address align properly for it
19:48esdrastarsis[d]: marysaka[d]: The test works with `.has_compression = false,` :happy_gears:
20:31esdrastarsis[d]: marysaka[d]: mhenning[d] Done: https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15358
20:59mohamexiety[d]: 🙃
21:03mangodev[d]: esdrastarsis[d]: i do really hope the source of this issue can be found
21:03mangodev[d]: otherwise, Turing is toast on NVK (or at least NVK over Nouveau)
21:06mhenning[d]: I hope we fix it too
21:06mohamexiety[d]: might not be turing exclusive too, who knows. it wasnt caught by us on newer hw but there's another variable in that this is a laptop which means it has prime. no wsi makes it unlikely but also idk if prime factors in here as well some other way
21:09mohamexiety[d]: i was doing prime in the past with comp and it didn't have issues, but this was back before it got merged. was testing/running stuff on ada 4000 + screen connected to 5070
21:09esdrastarsis[d]: mangodev[d]: It''s already toasted without compute mme 💀 🤣
21:09airlied[d]: I don't think those tests will hit prime paths
21:10airlied[d]: I'm writing this on a tu117 laptop, and I have tu116 card in a desktop now
21:15mangodev[d]: esdrastarsis[d]: yeah, but we've had memory compression since… GeForce3?
21:15mangodev[d]: it'd be sending Turing (and all before it) back to the stone ages
21:17mhenning[d]: 6.19 doesn't have compression enabled. it's new in 7.0 for nouveau
21:17mangodev[d]: ahhh interesting
21:18mangodev[d]: what was the thing that was enabled a while back? large pages?
21:18mangodev[d]: i thought something was enabled in 6.17 or so
21:18mangodev[d]: or was it a delay due to nouveau-level faults
21:21karolherbst[d]: so I'm curious about the "buffer_view" part of this issue. Like when does compression get enabled for that?
21:22karolherbst[d]: like ampere can do compression just fine on buffers, but Turing can't
21:23mhenning[d]: karolherbst[d]: how do you know?
21:25karolherbst[d]: ehh wait.. it's Ada+
21:28karolherbst[d]: mhenning[d]: docs, but I don't know if that's publicly explained anywhere
21:40karolherbst[d]: https://patents.google.com/patent/DE112022001139T5
21:41karolherbst[d]: `post-L2 compression` is the term
21:42airlied: yeah buffer compression didn't really make sense before tensors became a thing
21:45airlied[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1499164689555525804/PXL_20260429_214325702.mp4?ex=69f3cd6b&is=69f27beb&hm=ec3dff3d45ed605168edc3f3eec88190f997e1962fa4d75c6548f1a8d1f26139&
21:45airlied[d]: Some weird artifacting, anyone seen it? Just Ubuntu 26 live cd on ampere
21:45karolherbst[d]: there are limitations prior to post-l2 compression, like only the 3D part can actually make use of it or something?
21:47karolherbst[d]: so yeah.. if using buffer views with compression is fine on later gens, but not on turing (do we know what's the status on ampere?) then this would be the explanation I figure
21:50mangodev[d]: would that mean that just disabling compression with buffer views would be the way to go? or is it more complicated than just "turn it off for feature X"
21:51mangodev[d]: *disabling it for Turing and before, of course
21:51karolherbst[d]: I don't know 🙂 but I'd say that turning it off for buffer views, if that simple, _might_ just solve this
21:57airlied[d]: Okay think I got the Ubuntu gnome Shell to crash on Ubuntu live, but on Turing not ampere
22:03mohamexiety[d]: karolherbst[d]: we dont enable plc anywhere (even on turing+)
22:04mohamexiety[d]: it's also available on ampere too from looking at openrm
22:11karolherbst[d]: mohamexiety[d]: mhhhhhhh
22:11karolherbst[d]: okay, but like it's still a bufer view
22:11karolherbst[d]: I'm just saying, if it's only broken on turing, this might be why
22:25mohamexiety[d]: i think the key lies in it being related to buffer views ye. even turing aside, we shouldnt have any comp on buffers
22:25mohamexiety[d]: unless if maybe something is passing in compressed pte kinds for texel buffer views? do buffers even take pte kinds?
22:26gfxstrand[d]: mhenning[d]: I think we’re in good shape now. If you want to give it another read, feel free.
23:28karolherbst[d]: mohamexiety[d]: well I'd argue that a buffer view probably does contain something well compressible.. but dunno 🙃
23:28karolherbst[d]: hard to tell
23:33jStefan: hello. I've opened this bug in launchpad: https://bugs.launchpad.net/ubuntu/+source/xserver-xorg-video-nouveau/+bug/2148792
23:34jStefan: should i keep with the 5.15 kernel to be able to switch to the proprietary driver, or just go with more recent software (ubuntu 26.04)
23:34jStefan: more recent doesn't fix the problem, but idk if it would help with debugging.