#asahi-dev · 2023-03-04 14:59
alyssatry maybe vc4, just ignore the 3D bits#asahi-dev · 2023-03-04 14:58
alyssaChaosPrincess: 100%#asahi-dev · 2023-03-04 14:57
alyssayes#asahi-dev · 2023-03-04 14:56
alyssa(~:#asahi-dev · 2023-03-04 14:56
alyssai responded with what the convention already used for arm socs#asahi-dev · 2023-03-04 14:56
alyssawhen JANNAU and SVEN started writing the DCP driver and they asked me, a linux gfx veteran, how names should be chosen#asahi-dev · 2023-03-04 14:55
alyssai mean uhhhhhhhhh#asahi-dev · 2023-03-04 14:54
alyssafrankly when I started typing the drivers (DCP and Mesa) i had bigger problems than the names#asahi-dev · 2023-03-04 14:53
alyssaPossibly it would've made more sense to go with a BSD flavoured "appledcp" name#asahi-dev · 2023-03-04 14:53
alyssawhich gives apple/asahi as the split for display/3d#asahi-dev · 2023-03-04 14:53
alyssa3D drivers are spicy, branding matters a lot more, use your project name#asahi-dev · 2023-03-04 14:53
alyssadisplay controller drivers are boring, just use the vendor name#asahi-dev · 2023-03-04 14:52
alyssarockchip/panfrost#asahi-dev · 2023-03-04 14:52
alyssabasically just mirroring the split from my chromebook#asahi-dev · 2023-03-04 14:51
alyssachadmed: admittedly I'm responsible for BOTH of those naming choices ;p#asahi-dev · 2023-03-04 14:51
alyssa13:16 < chadmed> (also apple was taken by the dcp driver) #asahi-dev · 2023-03-04 14:51
alyssa13:16 < chadmed> and alyssa called the mesa driver asahi so it makes sense to keep the kernel driver nomenclature the same#asahi-dev · 2023-03-04 05:53
alyssa(and by "just works" I mean "I knew it was coming and architectured a chunk of the driver accordingly which was a PITA to get right but now that's there and indirect dispatch is nice and easy!")#asahi-dev · 2023-03-04 05:51
alyssait's nice working with decent hardware for a change :D#asahi-dev · 2023-03-04 05:51
alyssaworks like a charm#asahi-dev · 2023-03-04 05:51
alyssawired up indirect dispatch#asahi-dev · 2023-03-04 05:31
alyssaIDK if this is accurate#asahi-dev · 2023-03-04 05:31
alyssahttps://github.com/dougallj/applegpu/pull/25#issuecomment-1440976836#asahi-dev · 2023-03-04 05:31
alyssaTellowKrinkle: someone on github seems to think it's added in M2#asahi-dev · 2023-03-04 04:21
alyssait's not something you can emulate like that#asahi-dev · 2023-03-04 04:21
alyssa(t likely)#asahi-dev · 2023-03-04 04:21
alyssano#asahi-dev · 2023-03-04 04:15
alyssa(obviously can't prove a negative but yeah)#asahi-dev · 2023-03-04 03:59
alyssadisappointing that Apple didn't throw in an int64 cmpxchg instruction#asahi-dev · 2023-03-04 03:57
alyssanot needed for DXVK at least#asahi-dev · 2023-03-04 03:55
alyssaSM 6.6, bleh#asahi-dev · 2023-03-04 03:54
alyssayeah I don't feel good about emulating this. is shaderInt64Atomics actually needed? :(#asahi-dev · 2023-03-04 03:54
alyssafor full Vulkan shaderInt64 i'd have to check the spec#asahi-dev · 2023-03-04 03:54
alyssaMSL cheats here#asahi-dev · 2023-03-04 03:53
alyssaoh but ugh the other issue there is going to be concurrent non-atomic loads #asahi-dev · 2023-03-04 03:51
alyssaseems like there's prior art for doing that anyway#asahi-dev · 2023-03-04 03:50
alyssaOr, a bit more generously, hash the address and index into an array of locks accordingly#asahi-dev · 2023-03-04 03:50
alyssai.e. it would correct (if terribly inefficient) to have one big global lock to update any int64 atomic#asahi-dev · 2023-03-04 03:49
alyssa"ok" in the sense of "sound", at the cost of performance obviously#asahi-dev · 2023-03-04 03:49
alyssaalthough, again assuming the divergence rules work out so you don't deadlock, it should be ok to reuse locks for multiple pieces of memory#asahi-dev · 2023-03-04 03:48
alyssaand since in general you can do atomics on any memory you want.. you need a general mapping of memory to lock#asahi-dev · 2023-03-04 03:47
alyssauff, but you need to have a common bit of memory associated with the atomic memory to actually spin on#asahi-dev · 2023-03-04 02:09
alyssa(correctly implementing mutexes on GPUs is tricky -- it's really easy to get deadlocks due to the divergence handling -- but not impossible)#asahi-dev · 2023-03-04 02:08
alyssathis is presumably the logic Apple has in exposing float atomics on M1 as a convenience too#asahi-dev · 2023-03-04 02:08
alyssaquite frankly atomics can be expected to be so slow (everywhere) that if they're even slower ... meh ...#asahi-dev · 2023-03-04 02:03
alyssaif we do want to emulate.. I mean, do a mutex... it sucks, but#asahi-dev · 2023-03-04 02:03
alyssaand optional in VK#asahi-dev · 2023-03-04 02:03
alyssanot in GL, optional in CL, fwiw#asahi-dev · 2023-03-04 02:03
alyssaTellowKrinkle: that's definitely an area of silliness yes#asahi-dev · 2023-03-04 02:03
alyssaand yeah, even with silly things to get there, the games should run just fine :)