#
#asahi-dev · 2026-08-30 09:28
JamesCalligeros[m]that calculation does work for scanout buffers from the GPU (ive been running it for days) so im not sure what we're doing wrong here
#asahi-dev · 2026-08-30 09:27
JamesCalligeros[m]there are double the tiles in each dimension vs 32x32 tiles
#asahi-dev · 2026-08-30 09:27
JamesCalligeros[m]no with the larger one, the 16x16 tiles will result in a larger size being calculated as the size is number of tiles in x * number of tiles in y * bytes per tile
#asahi-dev · 2026-08-30 09:25
sofuswith the smaller metadata?
#asahi-dev · 2026-08-30 09:25
JamesCalligeros[m]hmm then i suspect the metadata buffer is probably being calculated as too big and dcp is trying to read past the end of the buffer, which will not be mapped by dart naturally
#asahi-dev · 2026-08-30 09:24
sofusand here is the crash, https://pastebin.com/PUd4HSTF
#asahi-dev · 2026-08-30 09:23
sofusmaybe the full log withouth the crash is also helpfull: https://pastebin.com/cG4AdUPF
#asahi-dev · 2026-08-30 09:21
sofusafter the first to with with the old changes it would have said apple-dcp 389c00000.dcp: RTKit: syslog message: UPPipeDCP_H13P.cpp:3981: IOMFB: Decomp fail: x=32 y=0 code=32 channel=15
#asahi-dev · 2026-08-30 09:20
sofusi got 2 times: IOMFB: Decomp fail: x=32 y=0 code=0 channel=4
#asahi-dev · 2026-08-30 09:20
JamesCalligeros[m]axi_rd_err is exactly what it sounds like so idk if it could be anything else
#asahi-dev · 2026-08-30 09:20
JamesCalligeros[m]did dcp log anything else before that?
#asahi-dev · 2026-08-30 09:19
sofusffmpeg -f lavfi -i "color=c=0xff0000" -s 3456x2234 -r 1 -c:v libx264 -crf 31 -t 1 -pix_fmt yuv420p test.h264
#asahi-dev · 2026-08-30 09:19
sofusthe size of the buffer should be exacly the same, including offset and stride
#asahi-dev · 2026-08-30 09:19
JamesCalligeros[m]like just a single frame of 235,0,0?
#asahi-dev · 2026-08-30 09:18
sofusred 3456x2234
#asahi-dev · 2026-08-30 09:17
JamesCalligeros[m]what is your test source?
#asahi-dev · 2026-08-30 09:17
sofusthis time i got most of the screen in blue and the rest in cyan
#asahi-dev · 2026-08-30 09:17
JamesCalligeros[m]perhaps the surface size calculated was bigger than the mapped object
#asahi-dev · 2026-08-30 09:16
JamesCalligeros[m]that looks like an iova fault, its tried to read some unmapped memory or something like this
#asahi-dev · 2026-08-30 09:15
sofusapple-dcp 389c00000.dcp: RTKit: Message (id=1): IOMFB int_handler_gated: failure: axi_rd_err [0x40310034]
#asahi-dev · 2026-08-30 09:15
sofusJamesCalligeros[m]: apple-dcp 389c00000.dcp: RTKit: co-processor has crashed
#asahi-dev · 2026-08-30 09:14
sofusjannau: im getting 77/135 on t6020, try adding -v to see the errors gstreamer gets
#asahi-dev · 2026-08-30 09:07
jannausofus: was GStreamer-H.264-V4L2SL always at 28/135? Seeing that after merging your PR on t6020
#asahi-dev · 2026-08-30 09:03
JamesCalligeros[m]maybe? i would expect green or magenta but blue is also plausible
#asahi-dev · 2026-08-30 09:02
sofusbtw is blue screen = no output?
#asahi-dev · 2026-08-30 08:59
JamesCalligeros[m]try with the change though, im curious
#asahi-dev · 2026-08-30 08:59
JamesCalligeros[m]what the heck
#asahi-dev · 2026-08-30 08:59
JamesCalligeros[m]oh so i was reading the format descriptor thingy right the first time
#asahi-dev · 2026-08-30 08:58
sofusJamesCalligeros[m]: luma metadata is 4 times bigger (at least avd's output)
#asahi-dev · 2026-08-30 08:57
sofusthis could actually explain the t8103 errors. I had completly forgotten about that write
#asahi-dev · 2026-08-30 08:43
chaos_princessyes, obviously, but it is a one line change
#asahi-dev · 2026-08-30 08:43
JamesCalligeros[m]it screams claude or codex to me
#asahi-dev · 2026-08-30 08:41
chaos_princesssofus: mind reviewing this thing https://github.com/AsahiLinux/linux/pull/585
#asahi-dev · 2026-08-30 08:32
JamesCalligeros[m]the compression subtiles are always 16x16 with 8 bytes of metadata per tile, the luma image tiles are 32x32 with 4 compression subtiles
#asahi-dev · 2026-08-30 08:31
JamesCalligeros[m]have a look at the latest dcp/interchange branch... i done goofed with the meta_tile_bytes thing
#asahi-dev · 2026-08-30 08:31
JamesCalligeros[m]ahhhhhh
#asahi-dev · 2026-08-30 08:30
sofusthis should be correct right? https://github.com/sofus13/linux/commit/4448c51a778594dff1f0fef9b6139bbcf2b5dceb
#asahi-dev · 2026-08-30 08:29
sofusJamesCalligeros[m]: looks sane, not sure if pitch[0]=5476 is correct
#asahi-dev · 2026-08-29 22:40
JamesCalligeros[m]sofus: can you check with drm_info if the video buffer being pushed to dcp is sane
#asahi-dev · 2026-08-29 20:27
svenjust rebased the tbt stuff on torvald's tree and wrote a cover letter. now all i need is -rc1 and i can submit this :)
#asahi-dev · 2026-08-29 19:46
sofusthis is what i did to gstreamer (hardcoded to 3456x2234), https://pastebin.com/WaEvCxg8
#asahi-dev · 2026-08-29 18:45
sofusas in dcp and avd calculate the same thing and metadata actually starts where they calculated
#asahi-dev · 2026-08-29 18:44
sofusnot really sure what im doing wrong... The offsets/sizes match
#asahi-dev · 2026-08-29 18:43
sofus[ 1628.537187] apple-dcp 389c00000.dcp: RTKit: syslog message: UPPipeDCP_H13P.cpp:3981: IOMFB: Decomp fail: x=32 y=0 code=0 channel=4
#asahi-dev · 2026-08-29 13:24
jannaunot sure how mesa's common yuv code handles hw specfic modifiers, it might just work as long as the drivers accept the the modifier
#asahi-dev · 2026-08-29 13:22
jannauuser space software needs to explicitly handle the v4l2 format to drm modifier mapping
#asahi-dev · 2026-08-29 13:20
chadmedif a client doesnt request a specific modifier it probably doesnt know about them at all and so we shouldnt assume interchange use is indicated. agx can deal with its own format fine though so we're allowed to use that if the requested buffer can be tiled and compressed
#asahi-dev · 2026-08-29 13:20
chadmedmesa currently only allows interchange if the drm client explicitly requests it to prevent it being used for agx-local framebuffers it shouldnt be used for
#asahi-dev · 2026-08-29 13:19
jannauyuka: the kernel always uses /chosen/bootargs unless the kernel config has a CMDLINE and CONFIG_CMDLINE_FORCE is defined
#asahi-dev · 2026-08-29 13:19
chadmedok maybe we dont need to change mesa then

plus ›