#asahi · 2023-12-16 14:02
jannegcolor matrix/eotf for avd is hopefully whatever the source is#asahi · 2023-12-16 13:57
eilnjanneg: fwiw isp/avd are both nv12 bt709, I expect other ones are too#asahi · 2023-12-16 13:48
eilnthe only thing it's worse than cpu at is what it can't do (e.g. 444)#asahi · 2023-12-16 13:46
j`eyI just meant youtube or something, Im mostly fine with 720p, not 1080p (although that may change as I ge used to the much nicer screen!)#asahi · 2023-12-16 13:45
j`eyyeah thats a very worst case#asahi · 2023-12-16 13:44
eilnj`ey: worse case single-frame 64x64 video, the memory allocations (we usually assume there's more than one frame, so we setup FIFO) and general init dance might take longer, but it shouldn't use more power#asahi · 2023-12-16 13:39
eilnalso need to adjust the HAL bitfields for different hw revisions (I don't even know how many there are), but those should be roughly the same across the board, at least the logic#asahi · 2023-12-16 13:38
eilnmpegs are mostly done but vp9 needs work, also av1 is wild wild west#asahi · 2023-12-16 13:38
eilnI've set up a framework for RE-ing it, lmk if you can get an m1n1 hypervisor setup (required to obtain source dumps we use the emulator to extract the instruction stream from)#asahi · 2023-12-16 13:37
eilnRetr0id: I got the boring IPC/job submission stuff down, left is the instruction stream part (which is kinda fun)#asahi · 2023-12-16 11:45
j`eyok fair, that sounds reasonable#asahi · 2023-12-16 11:44
jannegwhile reading text or code 10+ seconds shouldn't be that uncommon when you don't track the position with the mouse pointer#asahi · 2023-12-16 11:39
j`eyI guess if you left your computer to make a cuppa tea#asahi · 2023-12-16 11:38
ChaosPrincesstime only changes once per minute on most machines#asahi · 2023-12-16 11:38
j`eymost people have the time or something on screen at least#asahi · 2023-12-16 11:37
j`eythe thing that's confusing is how often do things not actually change on the screen#asahi · 2023-12-16 11:36
janneghuh, the power use is having the display hardware to continously blend the main and cursor plane even though both don't change. looks obvious to me that that would use more power than (direct) scanout of a single plane#asahi · 2023-12-16 11:29
j`eysome numbers would be good for sure#asahi · 2023-12-16 11:29
Nefsen402I'm sure emersion has talked with a few drm folks as he always does. He always does things for a reason#asahi · 2023-12-16 11:28
Nefsen402The MR is questionable to me, does the few bytes it takes to store the hw cursor actually improve power efficiency measurably?#asahi · 2023-12-16 11:27
j`eyah#asahi · 2023-12-16 11:27
Nefsen402but the api isn't available to backends#asahi · 2023-12-16 11:26
Nefsen402From the compositor's side, it can disable hw curosrs by "locking" a sw cursor#asahi · 2023-12-16 11:25
j`eythere's also https://gitlab.freedesktop.org/wlroots/wlroots/-/merge_requests/4204/diffs which means you can sometimes get sw cursors #asahi · 2023-12-16 11:21
j`eyta#asahi · 2023-12-16 11:21
Nefsen402https://gitlab.freedesktop.org/wlroots/wlroots/-/merge_requests/4032#asahi · 2023-12-16 11:20
j`eyNefsen402: I'm just interested in the code#asahi · 2023-12-16 11:20
Nefsen402(or at least not hooked up)#asahi · 2023-12-16 11:19
Nefsen402doesn't work though so doe't be getting any ideas here#asahi · 2023-12-16 11:19
Nefsen402there is a wip pr, yes#asahi · 2023-12-16 11:19
j`eyoh, is there a PR for it? is that part of the libliftoff stuff?#asahi · 2023-12-16 11:19
Nefsen402that is in the works#asahi · 2023-12-16 11:19
Nefsen402yep#asahi · 2023-12-16 11:19
j`eybreak the api ;)#asahi · 2023-12-16 11:18
j`eyoh :(#asahi · 2023-12-16 11:18
Nefsen402basically the api guarantees that by moving the cursor you'll still have a hw cursor#asahi · 2023-12-16 11:18
j`eypresumably support could be added?#asahi · 2023-12-16 11:18
j`eyI asked but was totally confused by the answer :)#asahi · 2023-12-16 11:17
Nefsen402I remember discussions about that. wlroots can't bail a hw cursor by just movement#asahi · 2023-12-16 11:15
j`eythe cursor plane is buggy if you move it near the edges#asahi · 2023-12-16 11:14
Nefsen402I figured, but I'm sure support will be added one day as the hw supports it#asahi · 2023-12-16 11:12
jannegmy comment was about the appledrm driver#asahi · 2023-12-16 11:12
j`eywill be neat if we can eventually use the cursor plane#asahi · 2023-12-16 11:10
Nefsen402well yeah, kms planes is still early days everywhere with most wayland compositors disabling it by default or just not having support at all and forget about x11#asahi · 2023-12-16 11:08
jannegat the moment neither can as there is no support for any yuv formats or planes besides the primary plane#asahi · 2023-12-16 11:08
Nefsen402in the context of video decoding#asahi · 2023-12-16 11:08
Nefsen402I guess it could be done in a unified archiceture, but nobody is crazy enough to teach the CPU to write into compressed buffers#asahi · 2023-12-16 11:07
j`eythe CPU can't do that?#asahi · 2023-12-16 11:07
j`eyNefsen402: well if the CPU is already on, might be more power efficient to just leave AVD off#asahi · 2023-12-16 11:07
Nefsen402Plus the decoder is able to write directly into a scanout capable buffer for use in a kms plane