#asahi-re · 2026-03-28 16:31
oliver"[08:01] <@chadmed> i am worried that quartz is just doing it in software and feeding dcp a ycc bt2020 framebuffer" wouldn't you see that in the dcp trace as three distinct planes and a different format? #asahi-re · 2026-03-28 16:27
oliverBut in the end we're happy they did, right? #asahi-re · 2026-03-28 06:28
oliverYea, and the names with & are just aliases in a big switch/case table, mapping to the fourcc codes from that descriptor array. #asahi-re · 2026-01-29 06:51
oliveryea, that was the first issue i had, but i adjusted the offsets. i also had to adjust mesa to put it on a page boundary. when i did all that i got the triangle to look like in my picture. before that it was even more garbage. but likei said, there is just not enough color data in the first 20mb of my framebuffer.#asahi-re · 2026-01-29 06:48
oliveryou then have to adjust the sizes a little. agx compression uses 64px "macro tiles", so the image is actually 192 16px tiles large, not only 189 as the interchange one, that uses 16px tiles directly.#asahi-re · 2026-01-29 06:47
oliverregarding the issue you saw: you need to set "unk2" in PlaneInfo (dcp/ipc.py) to 3 (kIOSurfaceAddressFormatTiled). the field is not yet part of the struct in the kernel.#asahi-re · 2026-01-29 06:47
oliverchadmed: i was using the m1n1 dcp experiment.py. i first got it to work in scanning out a dumped compressed framebuffer from macos to verify that i understand the numbers.