00:46 DemiMarie: Venemo: Possibly! That person was probably me, since I work on it.
15:40 fililip: would it make more sense to change the default graphics timeout to 2-3s?
15:40 fililip: 10 seconds is a lot and likely makes some people reach for the reset button instead of letting their compositor recover
16:19 pixelcluster: no imo, 2-3s is a lot for a single submission but if you have one burst workload that exceeds that (especially an issue on small GPUs), the entire context will get killed and that's bad
16:44 fililip: what apps might make such submissions? and also, wouldn't they freeze the desktop anyway?
17:02 agd5f: fililip, compute shaders
21:12 DemiMarie: If AMDGPU switches to user queues and Mesa starts using them, will that make compute shaders preemptable?
21:13 agd5f: yes
21:13 agd5f: all user queues are preemptible
21:17 DemiMarie: I thought only compute could be preemptible.
21:17 DemiMarie: Is there a concrete plan to switch AMDGPU to user queues?
21:18 agd5f: yes, it's in progress right now
21:18 agd5f: KFD has always used them
21:19 DemiMarie: I ask because Qubes OS will expose AMDGPU at some point, but not KFD.
21:20 DemiMarie: Do you have a (very rough) ETA for when user queues will be the default?
21:22 agd5f: first half of next year I would guess
21:24 DemiMarie: That should be soon enough for Qubes OS to depend on it. I doubt that Qubes OS will ship GPU acceleration by default before then.
21:25 DemiMarie: "depend on it" is specifically for long running GPU compute — the plan is to use rusticl for OpenCL instead of ROCm
21:35 DemiMarie: Is firmware scheduling on consumer GPUs known to be buggy?
21:38 agd5f: no
21:44 DemiMarie: Is it used by Windows?
21:51 agd5f: yes
22:15 DemiMarie: Does AMD do fuzzing and other testing to ensure that the firmware cannot be compromised by malicious input?