#
✕ nick:bloom
#asahi-gpu · 2021-04-08 23:22
bloomindeed
#asahi-gpu · 2021-04-08 23:20
bloom(I think I see a mechanism to change the index / disable althoether for GL, but ugh)
#asahi-gpu · 2021-04-08 23:20
bloomso that will be fun to deal with
#asahi-gpu · 2021-04-08 23:19
bloomI learned that Metal forces primitive restart on, always
#asahi-gpu · 2021-04-08 23:19
bloomwell. yes
#asahi-gpu · 2021-04-08 23:17
bloom(power-of-two divisors just turn into a shift)
#asahi-gpu · 2021-04-08 23:17
bloomyeah, for non-power-of-two divisors
#asahi-gpu · 2021-04-08 23:15
bloomhttps://stackoverflow.com/questions/41183935/why-does-gcc-use-multiplication-by-a-strange-number-in-implementing-integer-divi <-- a variant of this
#asahi-gpu · 2021-04-08 23:15
bloomto avoid hardware division for instanced arrays, the driver calculates magic multipliers that work as an integer divide
#asahi-gpu · 2021-04-08 23:14
bloomalong with some seriously weird clever ideas (Mali instancing fits there, serious kudos to cwabbott for figuring that out!)
#asahi-gpu · 2021-04-08 23:14
bloomwell, yes... that's the *
#asahi-gpu · 2021-04-08 23:12
bloomso it's still a lot of work to match up the API with the command buffers... but there's little theoretical difficulty
#asahi-gpu · 2021-04-08 23:12
bloomCommand streams usually* don't... stuff tends to correspond pretty close to the API commands
#asahi-gpu · 2021-04-08 23:11
bloomthere's a lot of hard thinking to r/e an ISA correctly
#asahi-gpu · 2021-04-08 23:11
bloomInstruction sets have a lot of clever tricks
#asahi-gpu · 2021-04-08 23:04
bloomcmdstream r/e is usually "easier" than ISA r/e but also usually more tedious
#asahi-gpu · 2021-04-08 23:04
bloomsame way any r/e works
#asahi-gpu · 2021-04-08 22:55
bloomDarkShadow44: what figured out?
#asahi-gpu · 2021-04-08 22:45
bloomteapot is a bit further out still :v
#asahi-gpu · 2021-04-08 22:17
bloom(that's 2 of 7 known shaders here, but I digress)
#asahi-gpu · 2021-04-08 22:15
bloomit can now chew through enough of "hello, triangle" to dump the primary shaders
#asahi-gpu · 2021-04-08 03:46
bloomtomorrow will teach it to chew through graphics
#asahi-gpu · 2021-04-08 03:46
bloomok, decoder is far enough along that it can make sense of my compute shader test
#asahi-gpu · 2021-04-07 23:18
bloomuhhh
#asahi-gpu · 2021-04-07 23:15
bloomjn__: ;)
#asahi-gpu · 2021-04-07 23:13
bloomheh
#asahi-gpu · 2021-04-07 23:12
bloombut that's booring :-p
#asahi-gpu · 2021-04-07 23:12
bloom"ag_" for Asahi Graphics seems like the best bet
#asahi-gpu · 2021-04-07 23:11
bloomMostly looking for a pronounceable two-letter prefix
#asahi-gpu · 2021-04-07 23:11
bloomchrisf: Mm..
#asahi-gpu · 2021-04-07 23:07
bloomstill needs a name for $GPU_DRIVER
#asahi-gpu · 2021-04-07 23:05
bloomtries to smash together Panfrost code with Apple-isms and see what happens
#asahi-gpu · 2021-04-07 22:48
bloomWe've done some weird things to panfrost's since Mali is weird.
#asahi-gpu · 2021-04-07 22:48
bloomWondering if I'd rather base this on v3d's genxml instead of panfrost's
#asahi-gpu · 2021-04-07 22:47
bloomGuess I should start writing a cmdstream parser...
#asahi-gpu · 2021-04-06 02:01
bloomyeah, idk
#asahi-gpu · 2021-04-06 02:01
bloomoh, that..
#asahi-gpu · 2021-04-06 02:01
bloomwhich unk6?
#asahi-gpu · 2021-04-06 02:01
bloomhmm?
#asahi-gpu · 2021-04-06 01:55
bloomevidently hit her head capacity
#asahi-gpu · 2021-04-06 01:51
bloomPossibly, but there's still a lot magic for compute-only.
#asahi-gpu · 2021-04-06 01:50
bloombut not sure what it could be then.
#asahi-gpu · 2021-04-06 01:50
bloomcmd_* is conceptually ... not
#asahi-gpu · 2021-04-06 01:50
bloomin that case, unk2 is the command buffer
#asahi-gpu · 2021-04-06 01:33
bloomor something of that ilk
#asahi-gpu · 2021-04-06 01:32
bloomif I want to know the hw distinction I should just read the mtl docs :-p
#asahi-gpu · 2021-04-06 01:32
bloomin metal speak
#asahi-gpu · 2021-04-06 01:32
bloomcmd_* === the command buffer
#asahi-gpu · 2021-04-06 01:32
bloomunk2 === the encoder
#asahi-gpu · 2021-04-06 01:32
bloomD'OH!

plus ›