#
#asahi-gpu · 2026-02-05 07:28
oliverno
#asahi-gpu · 2026-02-05 07:28
jannauyou don't have to use '&l1r' (fot ARGB2101010) as pixelformat?
#asahi-gpu · 2026-02-05 07:27
jannauoliver: great, congrats
#asahi-gpu · 2026-02-05 07:16
JamesCalligeros[m]Sick ill take a look in a bit
#asahi-gpu · 2026-02-05 07:15
oliverI think that tile size here is always 16x16px
#asahi-gpu · 2026-02-05 07:14
oliverhttps://gist.github.com/oliverbestmann/8b2d24c15450ebd3021f688ae70af47e
#asahi-gpu · 2026-02-05 07:13
oliverand as you said earlier, maybe the flip is now faster, due to it needing to transfer less data
#asahi-gpu · 2026-02-05 07:12
oliverit is not much more. i can share my hacky kernel diff in a sec
#asahi-gpu · 2026-02-05 07:11
chadmedbut if theres not a lot of logic involved above and beyond determining the tile size for a given bit depth then its probably fine
#asahi-gpu · 2026-02-05 07:11
chadmedi am slightly concerned about the overhead of calculating the compression metadata offsets etc for every plane update since we already have effectively zero slack to play with in the commit pipeline given how long the flip complete event takes to fire
#asahi-gpu · 2026-02-05 07:09
chadmedokay maybe this can work then
#asahi-gpu · 2026-02-05 07:09
chadmedoh i was thinking of the verify_surfaces error for agx format anyway lol
#asahi-gpu · 2026-02-05 07:00
oliverin that mode there is no z-curve twiddling, just compressed 16x16 tiles aligned in scanout order, i.e. from left to right, top to bottom. this is exactly the same mode that macos is using to scanout its framebuffers
#asahi-gpu · 2026-02-05 06:56
chadmedhuh interesting. is the address twiddling the same as agx format?
#asahi-gpu · 2026-02-05 06:53
oliverThe ail_initialize_interchange functions is mostly just copied.
#asahi-gpu · 2026-02-05 06:53
oliveralyssa: I got the gpu to read interchange compressed format just fine, it was just missing a little change in `agx_pack_texture` and `agx_batch_upload_pbe`. Have a look at the diff here https://gist.github.com/oliverbestmann/d5c4827374dfc03f85a0870074bfa49a#file-mesa-25-3-4-patch-L306
#asahi-gpu · 2026-02-03 19:53
Shizneat
#asahi-gpu · 2026-02-03 19:16
robertwendti will lurk on here for a bit but for more details you can also email me at robert@trycua.com
#asahi-gpu · 2026-02-03 19:15
robertwendthello o/ wanted to drop this off here https://macos-gpu-passthrough-bounty.vercel.app/ we are offering a 10k bounty to anyone who can improve gpu passthrough on macos vm guests. we will also donate an additional 5k to asahi linux if whoever cracks this is asahi affiliated
#asahi-gpu · 2026-02-02 16:23
EmberAstroidgepuhh
#asahi-gpu · 2026-02-01 13:27
oliverEven if Alyssa does not approve, I got the gpu to render a triangle in interchange compressed and could scan it out with that dcp. I am happy with that!
#asahi-gpu · 2026-01-31 15:52
karolherbstalyssa: funny.. nvidia uses the reverse order: 0xf0, 0xcc, 0xaa. Intel seems to be 0xaa, 0xcc, 0xf0
#asahi-gpu · 2026-01-31 15:49
alyssakarolherbst: src/util/lut.h
#asahi-gpu · 2026-01-31 15:37
karolherbstwondering if we should have a nir instruction for it, because writing opts in nir for it is simpler than in backends and every backend will need the same things... or maybe just have a util function, because in the end it's just three constants you apply the ops on and get a result which is your LUT
#asahi-gpu · 2026-01-31 14:07
alyssa((a & ~b) ^ c) in 1 instruction
#asahi-gpu · 2026-01-31 14:06
alyssaso you can do, like
#asahi-gpu · 2026-01-31 14:06
alyssanvidia & intel actually have a 3 src version
#asahi-gpu · 2026-01-31 14:06
alyssayes bitop already existed on m1, it's a standard instruction
#asahi-gpu · 2026-01-31 11:29
integralpilotthe existing applegpu testing method of running arbitary instructions on real hardware is broken, refactoring it to a new testing way i thought of that verifies our assumptions are right. hopefully it works.
#asahi-gpu · 2026-01-31 09:40
integralpilotthat's like actually kinda cool love that Apple
#asahi-gpu · 2026-01-31 09:40
integralpilotOMG there's a 10-byte long bitop instruction (just one opcode for this) that has a 4-bit truth table encoding so that ANY 2-input boolean function will run with the same opcode
#asahi-gpu · 2026-01-31 09:31
integralpilot(which i'm fixing now!)
#asahi-gpu · 2026-01-31 09:31
integralpilot*only some instructions are documented for M3 by the existing work on M3 branch, but not many, and none have an emulator impl
#asahi-gpu · 2026-01-31 09:29
integralpilotokay so instructions are documented (i.e. mov) but don't have a hardware testbed emulator yet so we can't test them (the testbed compares emulator vs real to make sure our expectations are right), so i'm going to work on expanding the emulator for M3
#asahi-gpu · 2026-01-31 09:28
integralpiloti figured it out, now the hardware testbed works!
#asahi-gpu · 2026-01-31 09:13
integralpilotokay seems to be working now but every 8th thread will only return 0s, finding out why now
#asahi-gpu · 2026-01-31 09:06
integralpilothave to fix up the hw testbed code (of applegpu) though as even in the M3 branch it doesn't seem to have been ported for M3 yet, doing that now.
#asahi-gpu · 2026-01-31 09:00
integralpilotstill decoding the other bytes
#asahi-gpu · 2026-01-31 09:00
integralpilotand byte 11 in the level of detail mode (0x00=implicit, 0x04=bias)
#asahi-gpu · 2026-01-31 08:59
integralpilotokay so now i know that byte 6 is the dimention (0x09 = 2D, 0x80=1D, 0x39=3D, 0x13=Cube)
#asahi-gpu · 2026-01-31 08:19
integralpilotokay looks like texture sampling has a new opcode, 0x8005 (was 0x31 in M1/M2), and it's 14 bytes long now (was 8-12 before)
#asahi-gpu · 2026-01-29 17:03
oliverThank you
#asahi-gpu · 2026-01-29 14:54
alyssao/
#asahi-gpu · 2026-01-29 14:54
alyssaanyway, I have work to do
#asahi-gpu · 2026-01-29 14:52
alyssaapparently they have shader code to decompress
#asahi-gpu · 2026-01-29 13:59
oliverAfter reading gpu part 5 I don't understand how macos can use the interchange format. I can configure a metal application to do direct scanout in fullscreen (I think they call it direct-to-display). I recently did that and observed the dcp traces. The way I understood you this should not work when you have partial renders, as the gpu might need to read parts of it back. Do you have any insights in that?
#asahi-gpu · 2026-01-29 13:36
alyssabeats me.
#asahi-gpu · 2026-01-29 13:36
alyssawhy don't they just add interchange sampling to AGX?
#asahi-gpu · 2026-01-29 13:35
alyssawhy don't they just scanout AGX fb's with DCP?
#asahi-gpu · 2026-01-29 13:35
alyssathey do, and their solutions suck.

plus ›