#
#asahi-gpu · 2026-07-09 19:25
alyssaHonestly one of the biggest challenges was just geometry/tessellation emulation
#asahi-gpu · 2026-07-09 19:21
alyssaso I guess in a sense, StanfordAppliedCyber[m] noopwafel chaos_princess are all right in a way :)
#asahi-gpu · 2026-07-09 19:20
alyssaIf all the r/e was done, I expect the mesa side for m3 to be a few months to get something reasonbale, since you're building on top of years of my m1 work
#asahi-gpu · 2026-07-09 19:19
alyssait's not that gpu drivers are horribly impenetrable, it's just there's a metric ton of random details & rare features that you do need to support for conformance
#asahi-gpu · 2026-07-09 19:19
alyssaStanfordAppliedCyber[m]: writing enough Mesa code to spin cubes took me a few weeks while I was in school. writing enough to play D3D games at good speed took years
#asahi-gpu · 2026-07-09 19:18
alyssam1/m2 are basically the same, m3/m4 basically the same, m5 new but one would reasonably assume m6 will be close
#asahi-gpu · 2026-07-09 19:17
alyssaagx seems to follow a tick/tock cadence
#asahi-gpu · 2026-07-09 15:40
chaos_princessusb phy changed, usb-pd is now on spmi instead of i2c (pro/max only), internal mic decimator changed
#asahi-gpu · 2026-07-09 15:38
lacamar[m]It's good to see that the hardware is mostly unchanged. Is it just the GPU and interrupt controller that changed a lot from M2?
#asahi-gpu · 2026-07-09 14:41
noopwafelI'm kind of expecting 14.x->26.x to be more of a nightmare than M3->M4 :p
#asahi-gpu · 2026-07-09 14:25
tellowkrinkleBTW everything I've seen so far indicates that at least from an ISA perspective, M4 is almost identical to M3.
#asahi-gpu · 2026-07-09 14:11
pac85noopwafel: oh sure sounds good!
#asahi-gpu · 2026-07-09 14:10
noopwafelpac85: we should fold the different macos versions into that mess sometime :) I realized my local stuff has all this stuff I didn't push.. will try to get it organized and ping you
#asahi-gpu · 2026-07-09 14:05
noopwafeland a lot (most of?) this work is userspace, so you can do a bunch of it under macos.
#asahi-gpu · 2026-07-09 14:05
fl0_id_afkthat makes even more sense
#asahi-gpu · 2026-07-09 14:05
fl0_id_afkchaos_princess ah good to know. right I forgot it's basically a one chip
#asahi-gpu · 2026-07-09 13:59
chaos_princesseven x86 does not do pcie for igpus, they just pretend to have pcie as it is not just a bus but also an autoconfiguration system at this point
#asahi-gpu · 2026-07-09 13:58
chaos_princessthis is not x86, only wifi/bt/sd reader are pcie
#asahi-gpu · 2026-07-09 13:55
fl0_id_afkStanfordAppliedCyber[m] how can you work on gpu without pcie or a somewhat working system, isn't it connected through that?
#asahi-gpu · 2026-07-09 11:25
pac85Also, that branch is capable do tracing apple's driver. My personal approach to reing M4 would be to try tracing that and see where it fails and then find the differences. That said focusing on m3 for now probably makes more sense as M4 would be incremental work on top
#asahi-gpu · 2026-07-09 11:24
pac85StanfordAppliedCyber[m]: re how to develop Mesa witjouta kernel driver. We re-ed enough of the apple kernel interface to submit stuff through it. We have a standalone compute demo that runs on macos without metal. Also much of what's needed to draw a triangle is known at this point but I personally stopped working on the project (plan on coming back). https://gitlab.freedesktop.org/pac85/mesa/-/tree/suballocstuff_gfx?ref_type=heads
#asahi-gpu · 2026-07-09 11:08
Sobek[m]The M1/M2 work spent quite a while reverse engineering the whole stuff, getting proofs of concept working on top of a macOS kernel, (before the linux kernel driver ever existed), then it spent a fair amount of time on getting... (full message at <https://matrix.org/oftc/media/v1/media/download/Ac4s8JInSsPQoLO46LXlFzv81dK5TAaomqsZUSJtsGvDubUKJN_aykcCznYC6rqFr0oczhPEDboaQyqZzr_OqbpCefjqlGBQAG1hdHJpeC5vcmcvUVVFZFVabFhmYkdCdmtYSWducVdzbkV1>)
#asahi-gpu · 2026-07-09 11:07
noopwafelni$timeofday :)
#asahi-gpu · 2026-07-09 11:07
StanfordAppliedCyber[m]good {morning, afternoon, night} yall
#asahi-gpu · 2026-07-09 11:06
StanfordAppliedCyber[m]<Sobek[m]> "(PS, shouldn't you be getting..." <- yes
#asahi-gpu · 2026-07-09 11:04
noopwafelbut yes, I would also hope the M4 stuff is .. hopefully mostly incremental
#asahi-gpu · 2026-07-09 11:03
StanfordAppliedCyber[m]if i had to go from scratch the years timeline is very reasonable imo
#asahi-gpu · 2026-07-09 11:03
StanfordAppliedCyber[m]chaos_princess: oh yeah 100%, it's amazing how little time the m1 gpu driver took imo given its complexity
#asahi-gpu · 2026-07-09 11:03
StanfordAppliedCyber[m]that's fair, but i also feel to an extent it's if you get the capability working you get the capability working and you find a way to upstream that
#asahi-gpu · 2026-07-09 11:02
chaos_princess+ consider the fact that you will most likely be building on top of a pre-existing one
#asahi-gpu · 2026-07-09 11:02
chaos_princessyears - no, multiple months - more reasonable
#asahi-gpu · 2026-07-09 11:02
noopwafelit's kind of .. initial bringup vs something maintainable, especially on the mesa side
#asahi-gpu · 2026-07-09 11:02
StanfordAppliedCyber[m]if in a few weeks i go "holy shit this is way harder than i expected" i'm sure someone will tell me "i told you so"
#asahi-gpu · 2026-07-09 11:01
StanfordAppliedCyber[m]i've said it many times but I'm firmly unconvinced a gpu driver takes years
#asahi-gpu · 2026-07-09 11:01
StanfordAppliedCyber[m]probably a lot, but that's my business
#asahi-gpu · 2026-07-09 11:01
noopwafelthe m1n1/kernel stuff was all targeting 14.x, also. I don't know how much the initdata mess changed again in 26.x.
#asahi-gpu · 2026-07-09 11:00
noopwafelmore motivation for us to pick it up again I guess, because I imagine would be much better to build M4 support on top rather than trying independent reimplementations..
#asahi-gpu · 2026-07-09 10:59
StanfordAppliedCyber[m]fair enough
#asahi-gpu · 2026-07-09 10:59
noopwafelunfortunately too much of this is solely in our collective heads atm, and not written down anywhere other than kind of if you squint at the diffs of chaos.
#asahi-gpu · 2026-07-09 10:59
StanfordAppliedCyber[m]idk, i rationalize it by saying results are results, after all, i am the first person to get the hv working on m4 class SoCs
#asahi-gpu · 2026-07-09 10:59
Sobek[m]StanfordAppliedCyber[m]: *Opens a Salmon Can*
#asahi-gpu · 2026-07-09 10:58
StanfordAppliedCyber[m]Sobek[m]: if you knew me in real life you'd understand how generally true this is
#asahi-gpu · 2026-07-09 10:57
noopwafelM3+ generalized a bunch of things which used to be in the VDM struct etc, they're now constructed by what we call an 'entry shader' which also sets the 'real' shader
#asahi-gpu · 2026-07-09 10:56
Sobek[m](PS, shouldn't you be getting some sleep, if I remember your timezone correctly ?)
#asahi-gpu · 2026-07-09 10:56
Sobek[m](Stanford Applied Cyber, are you a cat ? Because clearly the Asahi lead seem to struggle herding you :D)
#asahi-gpu · 2026-07-09 10:52
StanfordAppliedCyber[m]it's the whole reason why i did this project and built the HV
#asahi-gpu · 2026-07-09 10:51
StanfordAppliedCyber[m]oh interesting, good to know, i'm VERY interested in the gpu to put it mildly
#asahi-gpu · 2026-07-09 10:49
StanfordAppliedCyber[m]what is this weird entry shader stuff?
#asahi-gpu · 2026-07-09 10:49
noopwafelbut yeah a few of us have been slowly working on this in the background, trying to get to The Blessed Triangle for m3 native
#asahi-gpu · 2026-07-09 10:49
noopwafelalthough I don't think it has the weird entry shader stuff which are the more interesting ones.

plus ›