#asahi-dev · 2026-06-27 08:29
sofusnot sure, is nv12 planar?#asahi-dev · 2026-06-27 08:25
jannausofus: can we support planar 4:2:0 surfaces?#asahi-dev · 2026-06-27 08:23
jannaugstreamer decodebin3 works out of the box on fedora asahi remix#asahi-dev · 2026-06-27 08:23
jannaulibva-v4l2-request has kinda the same issue that it requires magic configuration. you have to `export LIBVA_DRIVER_NAME=v4l2_request` to use it. That could be fixed with libva patches#asahi-dev · 2026-06-27 08:21
sofusthe offset between the first irq and dart irq is the same on all dt's i think#asahi-dev · 2026-06-27 08:19
sofusand its kinda weird because apple enables mbox 2/3#asahi-dev · 2026-06-27 08:18
sofusthis is for t6000#asahi-dev · 2026-06-27 08:18
sofushttps://github.com/ArcaneNibble/m1n1/blob/00f3db7fcde9c87277387e9c3d4a679326fd9817/proxyclient/experiments/avd.py#L606#asahi-dev · 2026-06-27 08:16
sofusyes#asahi-dev · 2026-06-27 08:16
JamesCalligeros[m]Not documented in the adt?#asahi-dev · 2026-06-27 08:16
sofusand btw i think avd has more than two interupt lines to the ap#asahi-dev · 2026-06-27 08:13
sofusi think your notes are in the right direction, theres still the question of generation the instruction stream on the cm3 vs ap#asahi-dev · 2026-06-27 08:12
sofusbut from what i could test, i think apple also resets the whole thing#asahi-dev · 2026-06-27 08:12
sofusthe firmware does a *lot* of random writes/reads #asahi-dev · 2026-06-27 08:11
sofusbut there could be a way to reset just one idk#asahi-dev · 2026-06-27 08:11
JamesCalligeros[m]Aaaaaaa#asahi-dev · 2026-06-27 08:10
sofusthe whole thing...#asahi-dev · 2026-06-27 08:10
JamesCalligeros[m]The whole thing or just the faulting vp?#asahi-dev · 2026-06-27 08:10
sofusso if we lose state and start another job on the same vp, we get another error#asahi-dev · 2026-06-27 08:09
sofusits also a matter of timeouts/errors if a vp has an error it usually needs a reset#asahi-dev · 2026-06-27 08:07
JamesCalligeros[m]If we pause a video and it shuts down we can just reprogram the stream can we not#asahi-dev · 2026-06-27 08:07
JamesCalligeros[m]If we’re storing anything that needs to survive a power cycle sure, but i dont think that would be the case right#asahi-dev · 2026-06-27 08:06
sofusbut i mean, we should just have a memory region that we copy on power down#asahi-dev · 2026-06-27 08:06
sofusapple is *super* quick to power down so i think the cost of power up/down is low, but i suppose that should be tested. Right now it powers down sometimes during decode#asahi-dev · 2026-06-27 08:03
JamesCalligeros[m]the kernel should not care if it gets shut down because if the kernel shuts it down then there are no more streams being decoded#asahi-dev · 2026-06-27 08:03
JamesCalligeros[m]and its still fundamentally stateless, the only state we care about is stuff that is in flight#asahi-dev · 2026-06-27 08:02
JamesCalligeros[m]i cant remember if apple does or not#asahi-dev · 2026-06-27 08:02
JamesCalligeros[m]also powering up and down and reinitialising at runtime is a design choice, we dont have to do that#asahi-dev · 2026-06-27 08:01
JamesCalligeros[m]forgor to push the notes, there should be some brainfart notes in there now#asahi-dev · 2026-06-27 08:00
JamesCalligeros[m]i dont think we need one right now but if we push more stuff into firmware it could be useful#asahi-dev · 2026-06-27 08:00
chaos_princessdo the same thing as apple is doing and log using message selectors, not real strings#asahi-dev · 2026-06-27 08:00
sofusyou think we need a logger?#asahi-dev · 2026-06-27 08:00
sofusi think the less state there is in the cm3 the better (its all lost on power down)#asahi-dev · 2026-06-27 07:59
sofusi agree#asahi-dev · 2026-06-27 07:58
JamesCalligeros[m]basically i just think reprogramming the decoders on every slice is cpu grunt work that we shouldnt really be waking up the big app cores to do#asahi-dev · 2026-06-27 07:58
JamesCalligeros[m]but the latter can just be iovas to something on host memory probably#asahi-dev · 2026-06-27 07:57
JamesCalligeros[m]the cm3 would only need to keep track of in-flight streams and the state of each decoder and maaaybe the dpb for each stream#asahi-dev · 2026-06-27 07:57
JamesCalligeros[m]https://github.com/chadmed/avd-fw/tree/work/interface i put some notes and the crap logger here, but im sure you have better ideas :P#asahi-dev · 2026-06-27 07:57
sofusi think it would be best not to have to much data in cm3 memory#asahi-dev · 2026-06-27 07:56
JamesCalligeros[m]yeah... it chokes easily and i think the cm3 is better poised to track streams in progress and dispatch arriving frames to the decoder units#asahi-dev · 2026-06-27 07:55
leioand other distros just need an appropriate dep, I wager. But yeah, that's a story when its ready to ship somewhere, to go together with that#asahi-dev · 2026-06-27 07:55
sofusJamesCalligeros[m]: for now the driver mostly ignores all cm3 stuff#asahi-dev · 2026-06-27 07:55
sofusJamesCalligeros[m]: we already support multiple decode streams at once, its just kinda bad#asahi-dev · 2026-06-27 07:54
leiook, then it needs gst-plugins-v4l2 patches to build two plugins like gst-plugins-opus#asahi-dev · 2026-06-27 07:54
JamesCalligeros[m]anyways i wrote a terrible janked up syslog thingy for the firmware and it "works" in the sense that it will write a log to a ringbuffer and kick the application core to read it, but its awful and idk where in sram we can put it without exploding stuff the driver is doing #asahi-dev · 2026-06-27 07:53
JamesCalligeros[m]they come from -bad #asahi-dev · 2026-06-27 07:53
JamesCalligeros[m]the stateless decoders dont come from that#asahi-dev · 2026-06-27 07:53
leiomedia-plugins/gst-plugins-v4l2 ?#asahi-dev · 2026-06-27 07:52
JamesCalligeros[m]most of the garbage has to move to firmware anyway so we can support multiple decode streams at once, so its going to happen regardless#asahi-dev · 2026-06-27 07:52
JamesCalligeros[m]including gentoo, which i guess could be fixed with a patch to gst-plugins-bad ebuild, but i think its one of those "do not build this it has external deps" ones