ChanServ changed the topic of #asahi-gpu to: Asahi Linux: porting Linux to Apple Silicon macs | GPU / 3D graphics stack black-box RE and development (NO binary reversing) | Keep things on topic | GitHub: https://alx.sh/g | Wiki: https://alx.sh/w | Logs: https://alx.sh/l/asahi-gpu
<jannau>
I didn't stop the video, I spotted an atomic_disable
<alyssa>
jannau: even without agx in the loop, closing compositors (eg quitting sway to return to the tty) crashes dcp... I don't think that happened on my original prototype but IDK what would've changed. I think you mentioned you know about this issue?
<alyssa>
might that be the same bug as glmark2 is hitting?
<jannau>
alyssa: I think I've fixed that
<alyssa>
OK
<alyssa>
i haven't uprevved in a while, will need to for agx though ;)
<jannau>
and lina should have the fix
<alyssa>
hmm, ok
<jannau>
it makes sense to wait at least a few days, we're going to merge iommu mapping handling in m1n1 soon
<jannau>
that unfortunately requires devicetree changes compared to what lina merged so it might take a few days until everything is sorted out
<alyssa>
sure
<alyssa>
i've got macOS side work to do, not actually looking forward to fixing glamor anyway :p
<jannau>
I might have a workaround for the most glaring glamor bug. I hooked the cursor plane up and copy its framebuffer if it's too much off-screen
<alyssa>
I am missing a lot of context I think
<jannau>
not sure if I should publish that though
<alyssa>
I assumed glamor was blocked on mesa bugs
<alyssa>
What do cursor planes have to do with it?
<alyssa>
or is there a bug drawing the cursor in GL and using a plane offloads from AGX to DCP to workaround?
<alyssa>
(but of course cursors are differently broken in DCP hence the copy?)
<jannau>
the cursor triggered the most glaring glamor rendering bugs
<alyssa>
delightful
<lina>
I think it was drawing 512x512 squares for the cursor or something like that ^^
<alyssa>
what a delight
<boulabiar[m]>
So did you get over the gpu restart each frame issue today?
<lina>
Nope!
<lina>
Now it crashes even more! (Which is probably a good thing for debugging)
goldsoultheory has quit [Quit: My MacBook has gone to sleep. ZZZzzz…]
goldsoultheory has joined #asahi-gpu
goldsoultheory has quit []
goldsoultheory has joined #asahi-gpu
<eric_engestrom>
lina: I just randomly saw ARM64_ERRATUM_2441009 "Cortex-A510: Completion of affected memory accesses might not be guaranteed by completion of a TLBI"; might be relevant to your TLB issue?
<eric_engestrom>
(in kernel config)
<eric_engestrom>
(sorry if this is irrelevant, and/or if you've already seen it :)
<lina>
I doubt Apple's cores have Cortex errata... ^^;
<eric_engestrom>
no, but perhaps the same kind of thing applies?
<eric_engestrom>
I don't really know anything about all this so I'm really just guessing
<lina>
I actually don't think I have a TLB issue at this point, at least not with the GPU. As far as I can tell from the m1n1 experiment, TLBI actually works fine
<lina>
I might still have a TLB issue with the coprocessor, which was being hidden by the GPU shutdown thing beacuse I think the coprocessor shuts down shortly after the GPU (so it mostly worked in practice)
<lina>
But there's something else wonky going on with the GPU. If I force it to stay on on purpose, sometimes it crashes on the second frame, sometimes a bit later.
<lina>
For the coprocessor I'm just never unmapping/freeing anything for now, to eliminate variables (besides, I plan to do that in my allocator later on anyway, there's no good reason to thrash those mappings)
chipxxx has joined #asahi-gpu
chipxxx has quit [Remote host closed the connection]