On 10/04/2024 10:17, Bryan O'Donoghue wrote:
> On 09/04/2024 17:53, Marc Gonzalez wrote:
>
>> On 09/04/2024 13:27, Bryan O'Donoghue wrote:
>>
>>> Can you give it a try ?
>>
>> Random notes
>>
>> For easy reference, I've used this command to test:
>>
>> $ mpv --hwdec=v4l2m2m-copy --vo=tct --quiet --no-audio demo-480.webm
>>
>> And it displays the video directly in the terminal :)
>> (Rendering speed depends on terminal size)
>>
>> I'd never played the video to the end.
>> I notice I get:
>>
>> [ 397.410006] qcom-venus cc00000.video-codec: session error: event id:1001 (deadb000), session id:79d7c000
>> [ 397.410114] qcom-venus-decoder cc00000.video-codec:video-decoder: dec: event session error 1001
>>
>> How bad is that?
>>
>>
>> Sometimes, decoding simply fails immediately.
>> Must quit & restart.
>> Will have to script a 100 starts and check frequency of failures.
>>
>>
>> Will test with higher-resolution video.
>>
>> # time mpv --hwdec=v4l2m2m-copy --vo=null --quiet --no-audio --untimed demo-480.webm
>> (+) Video --vid=1 (*) (vp9 854x480 29.970fps)
>> Audio --aid=1 --alang=eng (*) (opus 2ch 48000Hz)
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> Using hardware decoding (v4l2m2m-copy).
>> VO: [null] 854x480 nv12
>> [ffmpeg/video] vp9_v4l2m2m: capture POLLERR
>> Exiting... (Quit)
>> /*** HANGS UNTIL CTRL-C ***/
>
> I think there are a number of different resolutions across SoCs that
> will exhibit this behavior, I've seen it with lower resolutions on 8250.
>
> Its a bug that we need to drill into in Venus but I don't think is a bug
> that is specific to your setup.
>
>>
>> real 0m21.467s
>> user 0m3.795s
>> sys 0m1.914s
>>
>>
>> # time mpv --hwdec=v4l2m2m-copy --vo=null --quiet --no-audio --untimed --length=30 demo-1440.webm
>> (+) Video --vid=1 (*) (vp9 2560x1440 59.940fps)
>> Audio --aid=1 --alang=eng (*) (opus 2ch 48000Hz)
>> [ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
>> Using hardware decoding (v4l2m2m-copy).
>> VO: [null] 2560x1440 nv12
>> Exiting... (End of file)
>>
>> real 0m16.433s
>> user 0m1.764s
>> sys 0m1.118s
>
> If this higher resolution is stable for you, I'd say this is about baseline.
>
> 1. The GDSC change should make no impact on playback or available resolution
>
> 2. Higher more "normal" use cases like 1080p should be fine.
>
> If so, then file "low resolution is broken" under a "known unknown" and
> scrub your patches for submission.
>
> If not, we need to do more 8998 specific debug.
FWIW, I get the same behavior with 854x480 and 2560x1440:
1) If mpv runs with '--length=N' (play only part of the file)
then mpv exits cleanly, with no kernel messages.
2) If mpv plays the entire file, then mpv hangs at the end
(needs CTRL-C to exit) and driver prints to kernel:
[68643.935888] qcom-venus cc00000.video-codec: session error: event id:1001 (deadb000), session id:79d42000
[68643.935995] qcom-venus-decoder cc00000.video-codec:video-decoder: dec: event session error 1001
# time mpv --hwdec=v4l2m2m-copy --vo=null --quiet --no-audio --untimed demo-1440.webm
(+) Video --vid=1 (*) (vp9 2560x1440 59.940fps)
Audio --aid=1 --alang=eng (*) (opus 2ch 48000Hz)
[ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
Using hardware decoding (v4l2m2m-copy).
VO: [null] 2560x1440 nv12
[ffmpeg/video] vp9_v4l2m2m: capture POLLERR
Exiting... (Quit)
real 2m43.292s
user 0m9.204s
sys 0m4.591s
# time mpv --hwdec=v4l2m2m-copy --vo=null --quiet --no-audio --untimed --length=300 demo-1440.webm
(+) Video --vid=1 (*) (vp9 2560x1440 59.940fps)
Audio --aid=1 --alang=eng (*) (opus 2ch 48000Hz)
[ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
[ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
[ffmpeg/video] vp9_v4l2m2m: VIDIOC_G_FMT ioctl
Using hardware decoding (v4l2m2m-copy).
VO: [null] 2560x1440 nv12
Exiting... (End of file)
real 2m32.412s
user 0m8.401s
sys 0m4.917s
300 seconds decoded in 152 seconds = 2x for 1440p.
Will send the patch series in a few hours.
Regards
