07:16 hakzsam: agd5f_: thanks for you GPUVM fault interface, I can now report them to applications, see https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25875 :)
07:51 mupuf: agd5f_: is it possible that amdgpu would fail cs submission when running of GTT space? We seem to be hitting a -ENOMEM on the deck when running glcts on zink on radv. By default, the deck has a BAR0 size of 256MB
07:53 mupuf: I'll try to enable resizeable bar to see if this fixes the problem, but I was wondering if amdgpu's behaviour here would be suboptimal and it should instead wait for more space to be available before submitting the command buffer
07:54 mupuf: Right now, zink will assert and kill the app/game because it doesn't know what to do
07:54 mupuf: It could wait and retry but that's also not ideal
08:03 mupuf: Would be good to add IGT tests for that, since I can get amdgpu to oops in a TTM buffer migration in this situation
12:57 agd5f: mupuf, the BAR size is not a limitation on APUs. "VRAM" is just system ram, so the driver just uses the physical addresses directly
12:58 agd5f: I suspect we just can make everything resident for the CS between GTT and "VRAM"
12:59 agd5f: hakzsam, woot!
12:59 agd5f: *can't
13:00 mupuf: agd5f: I assume "VRAM" can't be used for command submission though, right? Or if we could, then the kernel would have to use it for that
13:00 mupuf: but good to know the APUs can access the whole of system memory, without having to go through a BAR
13:02 agd5f: mupuf, buffers can be in VRAM or GTT, but it doesn't really matter which one is used from a perf perspective on APUs
13:03 mupuf: Ack, so it may be that radv is just forcing the use of GTT for command buffers
13:04 mupuf: and maybe one thing we could do is that, if amdgpu returns ENOMEM at command submission, we would try to use "VRAM" instead
13:05 mupuf:is done for the day and will check this out tomorrow
13:57 agd5f: mupuf, normally for APUs, it best to set VRAM|GTT for the domain
13:57 mupuf: oh, even better :)
14:21 Venemo: agd5f: do you mean that specifically for command buffers, or everything in general?
14:42 mareko: Venemo: everything
15:00 Venemo: is it that simple? I wonder why we don't already do it like that
15:04 MrCooper: FWIW, with older APUs / some time ago, VRAM was faster than GTT
15:52 mupuf: MrCooper: yeah, still the case on the deck
15:52 mupuf:spent a bit of time on that
16:47 Venemo: hmm
16:48 Venemo: so it is not that simple, after all...