#asahi-dev · 2025-12-26 13:18
chadmedive been playing with the per-surface colour space/transfer function stuff in the dcp driver and it seems to also mostly Just Work on linux#asahi-dev · 2025-12-26 13:17
chadmedthe sdr content staying the same relative brightness is probably a mix of that, local dimming, and perhaps some fancy tone mapping/compensation on the sdr surface itself#asahi-dev · 2025-12-26 13:16
chadmedthe hdr content will be displayed on a different iosurface with its own colour primaries, xfer function, etc. #asahi-dev · 2025-12-26 12:30
chadmedi will test j416 tomorrow. still kind of crazy that page flips take so long...#asahi-dev · 2025-12-26 12:29
chadmedseems a little smoother but by and large no changes. 120hz is still dodgy on my j314.#asahi-dev · 2025-12-26 12:28
chadmedall atomic_flush does now is get a reference to the old state framebuffer, 1:1 put the plane state into the swap request then send it off#asahi-dev · 2025-12-26 12:27
chadmeda couple of experimental things are turned on and it needs cleaning up/authorship of commits fixed up but seems mostly sane and doesnt explode#asahi-dev · 2025-12-26 12:27
chadmedhttps://github.com/chadmed/linux/tree/dcp/scanout works on all laptops and tested with 144hz monitor#asahi-dev · 2025-12-26 10:57
chadmednot a big deal, ive just done a loooot of rebooting today and its a bit annoying #asahi-dev · 2025-12-26 10:56
chadmedyeah not sure whats up then. all my t602x machines take 20-30 seconds to decompress kernel/initramfs and u-boots idea of "1 second" does not match reality :P #asahi-dev · 2025-12-26 10:22
chadmedpre-linux environment is very slow and always has been#asahi-dev · 2025-12-26 10:22
chadmedare we spinning up the t602x ecores in a different pstate to every other soc (in m1n1) #asahi-dev · 2025-12-26 00:10
chadmedi just git-restored my entire branch. the thing i was doing was stupid #asahi-dev · 2025-12-26 00:10
chadmedyeah, and i spent like 3 hours wondering why the committed state was always one state ahead of our captured state :P#asahi-dev · 2025-12-25 23:53
chadmedoh you have to subclass state too#asahi-dev · 2025-12-25 14:45
chadmedits 1am though, problem for tomorrow#asahi-dev · 2025-12-25 14:45
chadmedso ive probably just got the order of operations wrong#asahi-dev · 2025-12-25 14:45
chadmedok only broken thing now is a null pointer deref when "update" = "mr dcp tear down this plane"#asahi-dev · 2025-12-25 14:33
chadmedso iomfb_flush is just filling in the swap from the captured state then flipping a switch on the crtc to tell plane_update to fill the other state and surface info#asahi-dev · 2025-12-25 14:32
chadmedmy version keeps all the guts in iomfb_template so that we dont have to worry about versioning and also tracks state with the surface so that we dont have surfaces drifting away from their "designated" atomic_state#asahi-dev · 2025-12-25 14:31
chadmedill take a look tomorrow and push something working so we can compare#asahi-dev · 2025-12-25 14:21
chadmedstill explodes sometimes but at least something gets drawn on the screen _before_ it explodes now! #asahi-dev · 2025-12-25 14:20
chadmedok its sorta working? keeping two states and updating one while scanning out the other seems to do the trick mostly#asahi-dev · 2025-12-24 13:46
chadmedstill not doing anything with them but ill figure it out next week now#asahi-dev · 2025-12-24 13:21
chadmeddidnt realise it was still gated for 6.5#asahi-dev · 2025-12-24 13:21
chadmedcrawls back into his hole#asahi-dev · 2025-12-24 13:20
chadmedat this point i think the speed limit is iomfb itself though like i really cant see anything we're doing that would cause such long scanout/flip times #asahi-dev · 2025-12-24 13:19
chadmedyeah once its correct we can make it fast #asahi-dev · 2025-12-24 13:18
chadmed(other than the cursor which still works fine of course) #asahi-dev · 2025-12-24 13:17
chadmedcurrent and hopefully less tricky problem to solve is why kwin absolutely refuses to ever scan out anything on any plane and is always compositing no matter what#asahi-dev · 2025-12-24 13:16
chadmedill play with it more after the food and booze wears off in a few days :P #asahi-dev · 2025-12-24 13:15
chadmedjannau: thats close to what i had but i had subclassed drm_plane in a dumb way so that couldve been the problem#asahi-dev · 2025-12-24 11:20
chadmedive given up on dcp for now for similar reasons, merry christmas :) #asahi-dev · 2025-12-24 07:06
chadmeda global shadow wont work because by the time we get to iomfb_flush we've already lost the state we need for the "current" commit#asahi-dev · 2025-12-24 07:05
chadmedyeah i cant figure out how to do this sanely given our firmware interface without a bunch of superfluous allocation and freeing#asahi-dev · 2025-12-24 04:33
chadmedsetting up the surfaces in atomic_plane_update is super unreliable/racy but im probably doing something wrong#asahi-dev · 2025-12-23 13:43
chadmedand definitely are not linear#asahi-dev · 2025-12-23 13:43
chadmedive traced a few different video formats being scanned out directly from avd and general desktop stuff and the not-avd surfaces definitely have different plane/compression info#asahi-dev · 2025-12-23 13:40
chadmedjannau: the compression info is a list of structs describing the compression format with a bunch of magic numbers. the "plane" info one is the same (see my dcp/scanout branch) but seems to do double duty for tiled formats with fields for tile size/stride/etc#asahi-dev · 2025-12-23 13:21
chadmedthe "compression" and tile format information#asahi-dev · 2025-12-23 13:15
chadmedand even so theres seriously no way a page flip takes 5ms... right?#asahi-dev · 2025-12-23 13:15
chadmedi dont think the timing hack is the issue here. i tried handling cases up to 5ms between swap_start and now and nothing really helped#asahi-dev · 2025-12-23 12:09
chadmedin more positive news it does seem like dcp supports the agx tiled formats after all, but i havent got any good ideas for setting the required metadata in a way that isnt awful #asahi-dev · 2025-12-23 11:49
chadmedyeah and kwin just reduced its own #asahi-dev · 2025-12-23 11:38
chadmedthis is uhhh#asahi-dev · 2025-12-23 11:38
chadmedoh yeah i see the thing in dcp.c #asahi-dev · 2025-12-23 11:28
chadmedwe never send a vblank unless a swap fails and i cant see anything else time-sensitive userspace should care about#asahi-dev · 2025-12-23 11:27
chadmedive made it asynchronous so that atomic_flush notifies that its done after putting the iomfb swap request on a workqueue but that didnt help#asahi-dev · 2025-12-23 10:08
chadmedthis does not happen on external monitors#asahi-dev · 2025-12-23 10:07
chadmedaquarium never goes above 60 and it takes ~10-15 seconds for the gears to get up to 120 fps