09:39 DottorLeo: Hi!
15:06 fililip: is there something that can be done on the GPU driver side about lagging windows with high GPU usage on a Wayland compositor like KWin? FWIW, on Windows this works just fine, even with extreme GPU workloads and on the same hardware (RDNA2 dGPU)
15:47 MrCooper: fililip: in principle yes, see https://blogs.gnome.org/shell-dev/2023/03/30/ensuring-steady-frame-rates-with-gpu-intensive-clients/, though to work well this requires that the compositor's GPU context has high priority and can preempt lower-priority ones, which isn't the case with AMD GPU drivers yet
16:03 fililip: ah, ok, I've heard from Xaver that KWin already uses high priority contexts but I haven't seen it work
16:03 fililip: out of curiosity, does Intel work then?
16:09 glehmann: it can work on amd if you use a compositor that uses an async compute queue instead of the main graphics queue, like gamescope for example
16:12 fililip: hm, could the other pipe work too? on RDNA2 there is a second GFX ring (that I have disabled on my kernel due to it being less stable during GPU recovery) that I think is relevant here in conjuction with MCBP
16:12 fililip: but I'm unsure
16:18 agd5f: fililip, the second ring on RDNA2 is currently exposed as the high priority ring
16:22 fililip: interesting, with that ring enabled KWin is always smooth no matter what
16:23 fililip: does this work better or similar on RDNA4?
16:25 agd5f: fililip, only RDNA2. RDNA4 only has one gfx pipe.
16:26 fililip: but it doesn't have any kind of preemption functionality for this purpose?
16:29 agd5f: fililip, it does. we tried to enable it on gfx9, but we ran into compatibility issues with old versions of mesa and ultimately had to disable it
16:30 agd5f: with gfx11 and newer, it's handled per queue, so you need per process queues which is what we have been working to enable over the last few years
16:34 fililip: hm, just tested on gfx11.5 (Strix) and I don't see it working as smoothly with 100% GPU usage as with gfx10 with both MCBP on and off
16:36 fililip: so maybe I'm missing something
17:05 agd5f: fililip, user queues are not enabled yet
17:16 fililip: ok, that would make sense, is there a debug parameter in kernel/mesa I could use to test this?
17:16 fililip: (or a patch series)
17:49 fililip: after some research, from what I can see this also requires firmware changes, is that done yet?
17:49 agd5f: fililip, only for dGPUs at the moment
17:49 fililip: ok, thanks