#
#asahi-gpu · 2022-10-13 23:40
alyssaSo this might be an iter specific quirk
#asahi-gpu · 2022-10-13 23:40
alyssaI have a code sample from the blob doing writes-after-writes with ldcf
#asahi-gpu · 2022-10-13 23:29
alyssano idea about the uarch
#asahi-gpu · 2022-10-13 23:29
alyssaphire: yes, interpolates varyings
#asahi-gpu · 2022-10-13 23:28
alyssacr1901: yeah, in exceptionally rare cases
#asahi-gpu · 2022-10-13 23:28
phireis the divider not fully pipelined?
#asahi-gpu · 2022-10-13 23:26
phirewhat does the iter instruction do? based on the args, load and interpolate varyings?
#asahi-gpu · 2022-10-13 23:25
cr1901flake as in "draws garbage"?
#asahi-gpu · 2022-10-13 23:06
alyssaAt a uarch level I men
#asahi-gpu · 2022-10-13 23:06
alyssaWhat I don't know is why.
#asahi-gpu · 2022-10-13 23:06
alyssaMy guess is that there is a write-after-write dependency between the iter instructions
#asahi-gpu · 2022-10-13 23:06
alyssaExtraordinarily rare flake
#asahi-gpu · 2022-10-13 23:03
alyssathis program is flaking. it really should not be
#asahi-gpu · 2022-10-13 23:03
alyssahttps://rosenzweig.io/flake.txt
#asahi-gpu · 2022-10-13 23:02
alyssaughhh does AGX do magical dependency handling for iter instructions
#asahi-gpu · 2022-10-13 22:30
alyssait doesn't
#asahi-gpu · 2022-10-13 22:21
appleboy711[m]How well does the M2 GPU work right now for Asahi Linux? (I do have an M2 Macbook Pro 13", but I am too scared to download something experimental on it)
#asahi-gpu · 2022-10-13 17:03
eric_engestromhaha
#asahi-gpu · 2022-10-13 17:01
lina(I have whole-TLB invals spammed everywhere in the MMU code right now)
#asahi-gpu · 2022-10-13 17:01
linaI think I'm invalidating in so many codepaths I'm effectively already invalidating twice... ^^
#asahi-gpu · 2022-10-13 17:00
eric_engestromlina: did you try the "invalidate twice" trick? it's unlikely it's that bug, but it's probably a single line to copy (disclaimer: I haven't read your code yet) so it should be easy to try -> worth a shot :)
#asahi-gpu · 2022-10-13 16:05
karolherbstit's insance how much of the command stream stuff contains "invalidatre this and that" on hardware across the board
#asahi-gpu · 2022-10-13 16:04
linaYeah, I'm thinking either TLBs or virtually tagged caches clashing on remapped VAs or physically tagged caches clobbering unmapped memory with dirty data or both...
#asahi-gpu · 2022-10-13 16:02
karolherbstmaybe some random caches need to be invalidated or something or something something
#asahi-gpu · 2022-10-13 16:02
karolherbstthat does sound like some flushing issue
#asahi-gpu · 2022-10-13 16:02
karolherbstmhh yeah...
#asahi-gpu · 2022-10-13 15:59
linakarolherbst: The reason I think it's TLBs (or closely related caches) is all the problems go away if I never free/unmap anything, as far as I can tell
#asahi-gpu · 2022-10-13 15:18
karolherbstsome magic bit in the hw context init sequence *sigh*
#asahi-gpu · 2022-10-13 15:17
karolherbstwe'd never figured out what to do if nvidia wouldn't have told me what the solution is
#asahi-gpu · 2022-10-13 15:17
karolherbstand there is some magic privilged register you can poke to enable those loads
#asahi-gpu · 2022-10-13 15:17
karolherbstyeah.. I saw... those things are hard to figure out if no debugging tools are available :/ One super annoying thing we had to debug with nvidia was, that fragment shader helper invocation weren't doing any loads and loops in shaders were running infinitely...
#asahi-gpu · 2022-10-13 15:15
linaSo not a firmware issue
#asahi-gpu · 2022-10-13 15:15
linakarolherbst: The problems lately have been GPU timeouts, which happen when the GPU side buffers are bad (they aren't faults so it's not crazy pointers, but just "something that wedges the GPU")
#asahi-gpu · 2022-10-13 15:13
linaAnd now I crashed kmscube, okay, so it's just not as bad as before.
#asahi-gpu · 2022-10-13 15:12
karolherbstmaybe the firmware isn't thread safe
#asahi-gpu · 2022-10-13 15:11
alyssafearless concurrency
#asahi-gpu · 2022-10-13 15:11
karolherbstyeah... I don't have strong arguments here, just a gut feeling
#asahi-gpu · 2022-10-13 15:10
linaBecause previously I could reproduce the problem with kmscube (at least sometimes) and now I can't...
#asahi-gpu · 2022-10-13 15:08
linaI'm already doing that, all submissions are synchronous right now... but I've been messing with stuff today and now I'm wondering if I'm confusing myself again.
#asahi-gpu · 2022-10-13 15:03
karolherbstnormally you want all of that to be quite asynchronous to not starve either side
#asahi-gpu · 2022-10-13 15:02
karolherbstnot sure how that works on apple hardware, but e.g. on nvidia we submit and then have to do fencing and wait on those to be sure the GPU actually reached certain sync points in the command stream
#asahi-gpu · 2022-10-13 15:02
karolherbstwaiting until all command submissions are done
#asahi-gpu · 2022-10-13 15:01
linaI'm not sure what you mean with doing a WFI. I can't change ASC code so I can't force it to do that, and what does WFI mean for the GPU?
#asahi-gpu · 2022-10-13 15:01
karolherbstanyway, not sure that was Apple's reason for doing it, just reminded me of that reading your gist
#asahi-gpu · 2022-10-13 15:00
karolherbst40 isn't a lot
#asahi-gpu · 2022-10-13 15:00
karolherbstyeah...
#asahi-gpu · 2022-10-13 15:00
linaWell... sure, but that's the same as just using 40-bit VAs to begin with, which isn't great...
#asahi-gpu · 2022-10-13 14:59
karolherbstthe VA not matching isn't a problem though, in userspace you just reserve the entire area the GPU can't use and off you go
#asahi-gpu · 2022-10-13 14:59
karolherbstahh, might be
#asahi-gpu · 2022-10-13 14:59
linaIt's like the GPU team took a few steps in that direction but they're not there yet?

plus ›