00:46 airlied[d]: gfxstrand[d]: remaining comments on the latencies patch are probably for you
01:00 gfxstrand[d]: I'll look in detail tomorrow.
01:02 gfxstrand[d]: I'm not thrilled about having misc +N in there but I don't have a better plan. We need something more accurate than the shit Turing estimates. And it's that or I build a R/E framework for it.
01:02 airlied[d]: I think start from the numbers we get and twiddle until the tests pass is as good as we can expect right now
01:51 lightbulb: hello
01:52 lightbulb: i need help, i recently did a new Arch install on my computer and instead of choosimf
01:52 lightbulb: choosing* Nvidia proprietary drivers, i installed Nouveau
01:53 lightbulb: and now everytime i use my monitor's max refresh rate of 75hz, the screen flickers really badly
01:53 lightbulb: (i set it using xrandr)
01:54 lightbulb: and i looked at dmesg and there is a line that says "nouveau 0000:01:00.0: gr: intr 00000040"
01:55 lightbulb: GPU is an Nvidia 1050ti, latest zen kernel, running XFCE4 on X11
01:56 lightbulb: and the screen output seems to get darker and darker the long it is on 75hz, which does not happen in 60hz
02:30 lightbulb: nvm i:
02:30 lightbulb: i'm going back to proprietary
02:30 lightbulb: sorry for being a burden
02:30 lightbulb: bye
07:27 snowycoder[d]: gfxstrand[d]: How would a R/E framework for latencies look like?
07:46 airlied[d]: there would be two halves depending on how much input you'd know, first you have to figure out what the min delays for each instruction would be, then you'd have to create categories for them and group them into those, I think you could work out the min deps by brute forcing the hw, or maybe hammering on the vk or cuda drivers
08:34 JesusChrist: [JesusChrist] Corpus Loaded With 818 Messages
08:34 JesusChrist: [JesusChrist] Training Neural Network...
08:36 JesusChrist: [Chat] Sup
08:36 JesusChrist: [JesusChrist] Pretend its just to
08:37 JesusChrist: [JesusChrist] just full. wandering in it, New Server smh lying temporary
08:39 JesusChrist: [JesusChrist] Corpus Loaded With 820 Messages
08:39 JesusChrist: [JesusChrist] Training Neural Network...
08:40 JesusChrist: [Chat] Set the send message time to 3-5 Minutes
11:28 dakr: marysaka[d], I'm not a huge fan of the NVVAL() and NVDEF() stuff, but it's clearly better than magic values. :)
11:39 karolherbst: dakr: maybe reuse whatever we've come up with for nvk?
11:40 karolherbst: though that's not useful for mmio register programming but it might allow for a cleaner push buffer handling?
12:04 Mary: I guess I will clean up/continue those patches then
12:04 Mary: karolherbst: another thing is that dev_vm.h regs are offset and not the proper address, the rest is kind of normal (PMC, ect)
13:07 karolherbst: right...
14:36 PhilAlbano: could you tell me what is the best nvidia card fro use with nouveau? what I'm asking is wayland, hw video of at least h264 and fan control that it wont run on full speed all the time
14:36 PhilAlbano: OS is latest Fedora
14:39 PhilAlbano: NVK hardware support is described in little fuzzy way on their documentation
14:39 PhilAlbano: 'some GTX 700 and 800 series, most 900 series'
14:40 PhilAlbano: but what bout nouveau in it?
14:57 chikuwad[d]: hw video exists on proprietary, it's vaapi that doesn't
14:57 chikuwad[d]: unless you use the third party vaapi driver
14:57 chikuwad[d]: https://github.com/elFarto/nvidia-vaapi-driver
14:57 chikuwad[d]: it does? :doomthink:
14:58 chikuwad[d]: weird
14:58 chikuwad[d]: I've used it for months now and have had no issues really
14:59 PhilAlbano: then in context of vaapi I could use nouveau with this vaapi driver which as I see exists officialy on Fedora 41/42 and it should work?
15:04 chikuwad[d]: mangodev[d]: correct
15:06 PhilAlbano: I do not play games then fps performance is not the case here
15:06 PhilAlbano: rather power consumption
15:07 PhilAlbano: what bout CUDA support?
15:07 PhilAlbano: it is possible only with their prioprietary driver?
15:08 mhenning[d]: mangodev[d]: nouveau gl supported ada but not blackwell.
15:08 esdrastarsis[d]: chikuwad[d]: but it breaks on major driver updates
15:08 PhilAlbano: we are talking about sth around nvk supports
15:08 PhilAlbano: then GTX1050Ti for instance
15:09 chikuwad[d]: esdrastarsis[d]: somehow I haven't faced that either wtf
15:10 mhenning[d]: mangodev[d]: yeah, at this point the ada support is still present but deprecated in favor of nvk+zink
15:10 esdrastarsis[d]: chikuwad[d]: It broke in the update from 570 to 575: https://github.com/elFarto/nvidia-vaapi-driver/commit/0ef97c1849fac3e07cdf39ffe4cd9aa9f352896f
15:11 PhilAlbano: ada?
15:26 chikuwad[d]: I have a feeling they're looking to purchase one depending on support
16:43 PhilAlbano: chikuwad[d]: whi?
16:43 PhilAlbano: who?
16:48 gfxstrand[d]: PhilAlbano: Yeah, CUDA is an nvidia API. We don't support it with open drivers.
17:57 kar1m0[d]: I don't think it is legal to reverse engineer it either
17:57 kar1m0[d]: Since it's proprietary Nvidia technology
17:58 magic_rb[d]: Eeeehhhhhhh it depends on what jurisdiction and whats your budget
17:58 magic_rb[d]: Windows is a proprietary technology and wine exists
17:58 chikuwad[d]: wine is a clean room reimpl
17:59 magic_rb[d]: Well yes
17:59 chikuwad[d]: and https://github.com/vosen/ZLUDA exists
17:59 magic_rb[d]: But you can do cude the same way
18:00 mohamexiety[d]: there's no need for reverse engineering, the CUDA API is public
18:01 PhilAlbano: I have put 1050Ti, nouveau loaded, vulkaninfo shows:
18:01 PhilAlbano: https://pastebin.com/raw/hR6JTyG4
18:01 PhilAlbano: then looks like nvm is working?
18:01 chikuwad[d]: correct
18:01 PhilAlbano: what is strange glxgears with vblank_mode=0 on fullhd full screen reports only 154
18:01 PhilAlbano: fan on card does not speed up
18:03 kar1m0[d]: mohamexiety[d]: I didn't know that🙂
18:03 esdrastarsis[d]: PhilAlbano: reclocking is unsupported for pascal gpus
18:03 gfxstrand[d]: That's because there's no reclocking on those cards
18:03 PhilAlbano: all right
18:03 kar1m0[d]: gfxstrand[d]: So what cards do support reclocking?
18:04 kar1m0[d]: I think Ada Lovelace does
18:04 kar1m0[d]: But what others do?
18:04 mohamexiety[d]: Turing+
18:04 esdrastarsis[d]: kar1m0[d]: Kepler, maxwell v1, turing, ampere, ada and blackwell
18:05 PhilAlbano: but sth is wrong with libva
18:05 PhilAlbano: https://pastebin.com/raw/WPWRe8qm
18:05 kar1m0[d]: Kepler?
18:05 mohamexiety[d]: Kepler and maxwell v1 are manual iiuc relocking but not sure
18:05 PhilAlbano: no profiles listed
18:05 mohamexiety[d]: kar1m0[d]: GTX 7xx
18:05 mohamexiety[d]: (and 6xx)
18:05 kar1m0[d]: What about the 920m
18:05 kar1m0[d]: Laptop kepler
18:05 mohamexiety[d]: if it's kepler it should but not sure if it's automatic or manual
18:06 mohamexiety[d]: there were fermi 920ms too
18:06 esdrastarsis[d]: mohamexiety[d]: yes
18:06 chikuwad[d]: PhilAlbano: nvk/nouveau doesn't support VA-API at the moment
18:06 kar1m0[d]: mohamexiety[d]: 920m is Kepler b
18:06 kar1m0[d]: From what snowycoder[d] told me that is
18:06 esdrastarsis[d]: mohamexiety[d]: it's manual
18:07 esdrastarsis[d]: my friend tested Minecraft on zink/nvk yesterday
18:07 kar1m0[d]: esdrastarsis[d]: So how would I make it work
18:07 PhilAlbano: but chikuwad[d] said earlier today that it does :)
18:07 kar1m0[d]: If it's manual
18:07 PhilAlbano: take alook:
18:07 PhilAlbano: 16:59 PhilAlbano: then in context of vaapi I could use nouveau with this vaapi driver which as I see exists officialy on Fedora 41/42 and it should work?
18:07 PhilAlbano: 17:04 < chikuwad[d]> mangodev[d]: correct
18:08 PhilAlbano: or I misunderstand sth?
18:08 chikuwad[d]: the third party va-api driver is for the proprietary drivers
18:08 chikuwad[d]: not nouveau
18:08 chikuwad[d]: and my "correct" was for the driver being decode-only
18:09 PhilAlbano: decode-only?
18:09 esdrastarsis[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1396917015515500865/photo_5181419194544139807_y.jpg?ex=687fd3d9&is=687e8259&hm=f9a4659bd7c0b93e3e602482992f343f6beee0842559c11373f49367f47d57f1&
18:09 esdrastarsis[d]: esdrastarsis[d]: it's really an interesting test
18:09 kar1m0[d]: esdrastarsis[d]: Hwmon monitoring?
18:10 kar1m0[d]: Wait
18:10 chikuwad[d]: GT720
18:13 esdrastarsis[d]: kar1m0[d]: https://github.com/polkaulfield/nouveau-reclocking-guide
18:14 kar1m0[d]: esdrastarsis[d]: Thank you
18:19 esdrastarsis[d]: kar1m0[d]: Yes, pre-gsp gpus support hwmon on nouveau
18:31 mohamexiety[d]: mangodev[d]: no
19:52 mohamexiety[d]: mangodev[d]: yeah
20:19 mohamexiety[d]: whatever the HW supports
20:20 mohamexiety[d]: so 5 for blackwell, 4 for ampere and ada, 3 for turing till kepler and 2/1 for older