#asahi · 2025-10-30 14:44
sveni just don't read phoronix because it's a waste of time tbh#asahi · 2025-10-30 14:44
sveni'd honestly be fine with enabling it inside expert mode or whatever once the basics work#asahi · 2025-10-30 14:43
svensure, but they're already claiming that for whatever reason right now and they'll never be happy *shrug*#asahi · 2025-10-30 14:42
svenyeah#asahi · 2025-10-30 14:42
svennow that being said, gpu acceleration is going to be the everest-sized one :D#asahi · 2025-10-30 14:41
sven+be#asahi · 2025-10-30 14:41
svenit's not gonna be great but i think it's gonna fine for a start #asahi · 2025-10-30 14:40
sveni also don't think that m3 with software acceleration will be all that bad tbh#asahi · 2025-10-30 14:39
svenbut the only way to learn is to get through that#asahi · 2025-10-30 14:39
svenyeah, and if you've never done those baby steps before they will feel quite huge#asahi · 2025-10-30 14:08
svenand then there's truly weird stuff like dwc3 on these machines: that sits behind *two* separate iommus and gadget mode and one half of xhci uses the first one and the other half of xhci the second one 🙃#asahi · 2025-10-30 12:10
svenon apple silicon: what j`ey said#asahi · 2025-10-30 12:10
svenon x86 they usually have their dedicated memory, yes#asahi · 2025-10-30 12:10
svenon apple silicon there's usually a single device in each group, on x86 you can have multiple devices (especially on the pci bus) in a single group#asahi · 2025-10-30 12:09
svenwhat you want to look for are iommu group assignments. each iommu group shares the same virtual address space#asahi · 2025-10-30 12:04
sveneach hardware usually has a separate iommu with a separate virtual address space for DMA#asahi · 2025-10-30 12:04
svenwhat do you mean with "same controller"?#asahi · 2025-10-30 12:03
svenlooks like a quick google search confirms that m1 actually only supports 4k and 16k pages#asahi · 2025-10-30 12:02
svenmost hardware is behind some iommu which also has pagetables for DMA#asahi · 2025-10-30 12:02
svenhm.. maybe i misremember, i thought m1+ allowed 64k pages#asahi · 2025-10-30 12:01
sveni thought it did#asahi · 2025-10-30 12:01
svenhuh, do i remember that wrong?#asahi · 2025-10-30 12:00
svenwhich effectively means that on linux only 16k and 64k will work#asahi · 2025-10-30 11:59
svenDART (iommu) and the gpu can only do 16k#asahi · 2025-10-30 11:59
sventhe cpu can do 4k, 16k and 64k#asahi · 2025-10-22 12:16
sventrue :(#asahi · 2025-10-22 12:15
svenunusual combination ;)#asahi · 2025-10-22 12:15
svenskydiving or IT :D#asahi · 2025-10-19 12:25
sven:D#asahi · 2025-10-19 12:21
sventhe co-processor state, especially for the gpu, is going to be very annoying#asahi · 2025-10-19 12:08
svenno psci support so no way to hand over the secondary cores. we also can't shutdown many of the co-processors to a sane state again#asahi · 2025-10-16 20:57
svenI’d rather get m3/4/5 support than improved haptics :p#asahi · 2025-10-16 07:50
svenoh… right. and that’s shared between all installations so you have a chance to screw up everything #asahi · 2025-10-16 06:00
svenit’s not impossible, just a lot of work #asahi · 2025-10-16 05:59
svennah, it just hasn’t been the focus of anyone interested and motivated enough #asahi · 2025-10-05 15:18
sventhough i guess at those speeds most of link training just has to happen in hardware :D#asahi · 2025-10-05 15:17
svenjust look at atcphy.c for how link training might otherwise looks like, and even that's already surprisingly simple and without much conditions or loops#asahi · 2025-10-05 15:14
svenelectrons travels very slow actually (something like milimeters per second iirc), what you care about are the fields and the interference caused by the high speed switching between 0/1#asahi · 2025-10-05 15:12
svenML for both branch prediction and link training also makes perfect sense. it's a very well specified problem and you immediately know if you made a mistake and the cost of a mistake isn't all that high. on the other hand, getting something correct can be verified immediately and you directly gain something#asahi · 2025-10-05 15:09
svenno, it really didn't#asahi · 2025-10-05 15:07
svenusually they exchange some well-known pattern at the beginning and tune all their parameters until the other side can recover that pattern successfully #asahi · 2025-10-05 15:05
svenit's usually clock recovery followed by a whole bunch of magic and adjusting filters and drive strengths and timings and whatever so that both sides can actually communicate at multi-Gbps speeds#asahi · 2025-10-05 15:03
sven:/#asahi · 2025-10-05 14:49
svenas painful as that rpc interface is I feel like implementing all that crap like link training for the various modes would be worse #asahi · 2025-10-05 12:42
svenyeah, that too. not impossible but unlikely#asahi · 2025-10-05 12:39
svenwhich i'd be happy to see ofc but i don't have high hopes someone signs up for that on the off-chance it unlocks something useful ;)#asahi · 2025-10-05 12:38
svenmaybe, but then we'd need someone to reverse engineer and re-implement all the stuff dcp does #asahi · 2025-10-05 12:36
svenwe can still access the display controller mmio from the main core as well, at least on m1/m2#asahi · 2025-10-05 12:36
svenif you look at the dcp <-> xnu firmware interface it very much looks like they cut their previous display controller driver in half#asahi · 2025-10-05 12:15
svenespecially for that always-on display