#
✕ nick:lina
#asahi-gpu · 2023-02-22 16:24
linaSo now with ASAHI_MESA_DEBUG=stats you get this:
#asahi-gpu · 2023-02-22 16:21
linaThat mystery didn't take up my whole day, so I still had time to implement render stats and fault feedback ^^
#asahi-gpu · 2023-02-22 16:21
linaglmark2 got another speed bump too, and I saw terrain at over 180 FPS now
#asahi-gpu · 2023-02-22 16:20
linaSo now we're back to passing the dEQPs, GLES2 and EGL, except now with no flakes and with gallium sync = flush, not sync ^^
#asahi-gpu · 2023-02-22 16:20
linaBut trying random stuff I found out that one bit in the VDM barrier command you found fixes everything, so now I always put one of those at the beginning of every batch.
#asahi-gpu · 2023-02-22 16:19
linaI don't know what the "official" way to fix this is, it's possible macOS just never hits it because it never actually pipelines GL commands into hardware queues? (It might always roundtrip through the CPU)
#asahi-gpu · 2023-02-22 16:17
linaMy conclusion was that the USC caches persist and are not invalidated when you queue things back to back on the GPU, so if the CPU writes some vertices to a VBO, submits a batch, writes some more, submits another batch, the first one could load a whole cache line before the second one's vertices were written, and then the second one would read those (usually just the first vertex) as zero from the cache.
#asahi-gpu · 2023-02-22 16:15
linaBut then there was a much subtler thing that had me stuck for a while... some dEQPs would flake, rarely. The symptom on at least some tests was that a vertex would be processed as all-0, rarely (but always the same one for any given test).
#asahi-gpu · 2023-02-22 16:13
linaThe actual big bug was really obvious, that only took an hour or so to find... I wasn't clearing ctx->batch so after a submission things kept getting piled onto the same batch.
#asahi-gpu · 2023-02-22 16:12
linaalyssa: Explicit sync should be fixed! ^^
#asahi-gpu · 2023-02-21 06:05
linaalyssa: I don't think we ever exhaustively proved that the kernel/UAPI handling for multisampling is correct... it's entirely possible that I missed something or that we need to expose one of the remaining unknowns (remember there's a bunch of bits in debug_flags you can flip globally)
#asahi-gpu · 2023-02-18 06:33
linaalyssa: I'll probably look at it tomorrow, busy today getting Luna packed up for her flight back ^^
#asahi-gpu · 2023-02-16 17:13
linaAnd now we know exactly what the depdendencies are!
#asahi-gpu · 2023-02-16 17:13
linaThat should make rebasing on newer kernels a lot easier going forward, no nasty merge with rust-for-linux/rust.
#asahi-gpu · 2023-02-16 17:12
linaThe driver is now rebased on top of mainline + cherry-picked dependencies! ^^
#asahi-gpu · 2023-02-15 17:38
linaflops
#asahi-gpu · 2023-02-15 17:38
linamood........
#asahi-gpu · 2023-02-15 17:37
linaI thought you had stuff to do sis, why are you still here? ^^;;
#asahi-gpu · 2023-02-15 17:36
linaAt least the batch tracking works well now, I'm sure alyssa can fix WSI for us!
#asahi-gpu · 2023-02-15 17:36
linaAnyway, this is clearly quite broken but I think I should get some sleep ^^;;
#asahi-gpu · 2023-02-15 17:34
linaI think the plan was the new-school way, the kernel driver only knows about out syncobjs (which it replaces the fence on)
#asahi-gpu · 2023-02-15 17:28
linaI should get some sleep... and other than the Rust meeting tomorrow night, I think I'm going to take a break until next week ^^
#asahi-gpu · 2023-02-15 17:26
linaThe function gets passed a pipe_fence_handle** and it replaces it with a new handle (that's from panfrost) which is a clone of the syncobj from the last batch flushed... is it supposed to do something different?
#asahi-gpu · 2023-02-15 17:20
lina*hides*
#asahi-gpu · 2023-02-15 17:20
linaThen again I stole the fence code from panfrost so who knows whether it works? ^^
#asahi-gpu · 2023-02-15 17:19
linaagx_flush does do the fence right, but it also didn't work when I didn't also force a full sync so clearly something is broken...
#asahi-gpu · 2023-02-15 17:15
linaSo I just winged it myself in the end ^^;;
#asahi-gpu · 2023-02-15 17:15
linaI wasn't sure of what driver to look at, couldn't find a good reference for what has proper explicit sync...
#asahi-gpu · 2023-02-15 17:11
linaalyssa: The code is unconditional there...
#asahi-gpu · 2023-02-15 17:11
linaalyssa: Yes, active used to mean both but it was more of a mess that way
#asahi-gpu · 2023-02-15 17:11
linaI just realized we're doing a CPU copy in there right now with no sync... I clearly glossed over it way too quickly today...
#asahi-gpu · 2023-02-15 17:08
linaI also have no idea what DCP does ^^;;
#asahi-gpu · 2023-02-15 17:08
linaI don't sync in flush_frontbuffer yet, I wasn't sure about that one, so maybe that's it...
#asahi-gpu · 2023-02-15 17:05
linaWayland
#asahi-gpu · 2023-02-15 17:05
linaPlus without that we leak memory
#asahi-gpu · 2023-02-15 17:05
linaWe do need the sync_all in context destruction if we don't want faults on clean exits
#asahi-gpu · 2023-02-15 17:04
linaWell KDE becomes a mess without it... so either kwin is broken or we're missing syncs in some other place...
#asahi-gpu · 2023-02-15 17:03
linaOkay then there's another problem somewhere else ^^;;
#asahi-gpu · 2023-02-15 17:01
linaBut I'm sure I missed something
#asahi-gpu · 2023-02-15 17:01
linaKDE was glitchy as heck until I remembered that a gallium flush should be a full sync...
#asahi-gpu · 2023-02-15 17:00
linaalyssa: You can probably just increase that constant to a silly value to work around it
#asahi-gpu · 2023-02-15 17:00
linamarcan: Can we please get rid of mailbox pretty please? ;;
#asahi-gpu · 2023-02-15 17:00
linaThis is mailbox.
#asahi-gpu · 2023-02-15 17:00
linaBut this isn't that
#asahi-gpu · 2023-02-15 17:00
linaAnother thing I added/fixed in this round is that GPU firmware crashes should actually cleanly abort all jobs (and mesa has an abort() in that path)
#asahi-gpu · 2023-02-15 16:59
linaOh mailbox. Oof. Yeah.
#asahi-gpu · 2023-02-15 16:40
linaYou are, it was just unexpected wwwww
#asahi-gpu · 2023-02-15 16:37
linaSis...
#asahi-gpu · 2023-02-15 16:37
linawwwwwwwwwwwwwwwwww
#asahi-gpu · 2023-02-15 16:37
lina"performance" is just "max" ^^

plus ›