00:57airlied: TimurTabi: I've pushed your dma fix to drm-misc-fixes
03:04x512[m]: cubanismo[d]: What is z-order of NVKMS planes? Are overlay planes above or below main plane? Is it configurable?
03:06x512[m]: Layer in NVKMS terms.
08:19cubanismo[d]: I'm not familiar with that.
09:56karolherbst[d]: Do we have a quick way (quicker than 2 prmts) to swap high/low bits of two 32 bit ints? Like from `a_hi | a_lo` and `b_hi | bi_lo` to `a_hi | b_lo` and `b_hi | a_lo`
09:57karolherbst[d]: but I guess 2 prmts is the quickest we have here.. just hoped we'd have something where I could do it in place
09:58karolherbst[d]: not sure if my idea is worth it.. was thinking about doing 2 prmts + movm+ 2 prmts to transpose int32 matrices... but....
14:32gfxstrand[d]: For two 32-bit destinations, you need two instructions. There are very few actual 64-bit instructions.
14:55karolherbst[d]: I kinda dislike how those matrix instructions are never generally applicable.. movm operates on vec2s, and the halves end up in different threads 🥲
14:59karolherbst[d]: but maybe I need to think about it from a higher level.. like maybe I can change the load, so that after movm it's all correct
15:48w0rmi: Hello! I'm a newbie(very) that is trying to get into nouveau development(some day..) and I have some small questions that would help me a lot since i'm starting in this open-source world. Sorry in advance if my questions are too stup.., I'm very new on this.
15:48w0rmi: This is my situation; I was trying to set up a development environment for nouveau with my current pc, which is ubuntu(linux-kernel: 6.17.0-5-generic) + dgpu: rtx 5080.
15:48w0rmi: Later, I realized that I had an error trying to attach the nouveau driver to my dGPU. The error was "gsp ctor failed: -2", which tells me that it couldn't find the GSP firmware blob. I searched in the linux-firmware.git for the BLOB(which is gb203 for the RTX5080) and found that is not available yet. Then, I read this commit https://git.kernel.org/pub/scm/linux/kernel/git/firmware/linux-firmware.git/commit/?id=050f0a1e
15:48w0rmi: 3a76d671c073b59c813a4119b610b725 stating that the folder of the gb203 is linked to the gb202, and the BLOB from gb202 is linked to the BLOB of the ga102. Which is strange, because I don't see it in the main branch. After seeing that, I created the gb203 folder in my lib/firmware/nvidia and copied the blob(gsp-570.144.bin) from the ga102 into the gb203 folder. I reloaded the driver nouveau again into my dGPU and it
15:48w0rmi: worked! I was able to run vkcube, with many errors, but with my dgpu with nouveau.
15:48w0rmi: My question is: Is this the only way to make it to work for now, or is there a better way? If it's the only way, do you think that nvidia will release the official GSP BLOB for the rtx 5080 soon? Because with this setup, I think that I will have a lot of false positive bugs while devoping.
15:48w0rmi: Sorry for my very long message, and thank you a lot for reading all until the end. Thank you for your time.
15:49karolherbst: w0rmi: maybe ubuntu split it into a nouveau/nvidia specific firmware paackage?
15:49karolherbst: Fedora did something similarly recently
15:50chikuwad[d]: arch did as well
15:50karolherbst: how are they all failing to get the nvidia package installed ...
15:51karolherbst: w0rmi: apparently the package is called "nouveau-firmware"
15:55kwizart: It's nvidia-gpu-firmware in fedora, nouveau-firmware is likely the rpmfusion version of the old firmware for video hw
15:56kwizart: (but nvidia-gpu-firmware is from linux-firmware)
15:56chikuwad[d]: kwizart: we're troubleshooting for ubuntu
15:57karolherbst: ohh right...
15:58karolherbst: so what's the ubuntu package name then, because I can't find it out :)
15:59chikuwad[d]: nouveau-firmware
16:00karolherbst: no, that contains old files nobody needs
16:01karolherbst: seems like it's still linux-firmware
16:03chikuwad[d]: yeah, looks like it
16:04chikuwad[d]: but it's also just
16:04chikuwad[d]: out of date
16:04chikuwad[d]: `20250317`
16:22w0rmi: oh, I see. The part I'm unsure about is if the gsp-570.144.bin file(from the ga102 folder) is actually the correct one for my gpu. I wouldn't think it is, since the ga102 architecture is very different from the gb203(rtx 5080). But in the end, in the linux firmware commit that I sent shows they are linked, and it does work at the end.
16:25karolherbst: w0rmi: blackwell uses the same firmware
16:25karolherbst: at least the gsp one
16:26karolherbst: anyway... ubuntus firmware package should just install the needed files correctly
16:36w0rmi: oh, I see. I'll continue using it then. Thanks a lot Karol!
18:25phomes_[d]: I test on a variety of distros. Ubuntu is one of them. The gsp files are in linux-firmware. Things where not working untill I installed nvidia-firmware-580-580.95.05 too. I don't think it is needed but it conflicts with nvidia-firmware-580-server-580.95.05 and maybe that server package causes problem. I stopped caring when it started working
19:52kwizart: phomes_[d], this firmware comes from the nvidia driver package from the cuda repo, not ubuntu canonical repo, right ?
22:34karolherbst[d]: The current patches I have in https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37998 kinda make me think I need to rewrite the entire mess 🙃 I _wonder_ if I want to hook up `cmat_transpose`, and then lower matrices to be always row major on A/C, col majo on B, and then call into `nak_nir_lower_cmat`... but that also kinda requires having a good `cmat_transpose` impl that doesn't suck in
22:34karolherbst[d]: the worst case..
22:36karolherbst[d]: Tomorrow sounds like a "Karol draws stuff on 10 sheets of paper" day
22:39karolherbst[d]: one good thing is that ldsm + movm == ldsm.trans, which could be an opt pass...