hevc_vaapi hwupload encode causes GPU reset and crash

[Miroslav Rovis]

Summary of the bug:
How to reproduce:

% ffmpeg -vaapi_device /dev/dri/renderD128 -i RTL-2023-02-11_19h39m.ts -vf format=p010,hwupload -c:v hevc_vaapi -b:v 1.2M -c:a aac -profile 2 -sn RTL-2023-02-11_19h39m.mkv

FFmpeg is the latest from git (version N-109851-ge506ea3ce1, built with gcc 12 (Debian 12.2.0-14)

No need to upload the media to videolan.org as the media is very short. That media file is attached now. It always makes the crash happen on my system.

[Miroslav Rovis]

This is the file that will make the crash.

[Miroslav Rovis]

the processor is AMD Ryzen 7 5700G, and this is the STDOUT I copied after the reset

[Miroslav Rovis]

I can use hevc_vaapi and it has some advantages, but it crashes my system too often.
I can produce more media that crash here if needed.
What causes this?
Thanks!

[Miroslav Rovis]

this is the verbose debug ffmpeg own log of the run today at 16:45 local

[Miroslav Rovis]

It's all mostly vague to me, but there's a line that looks suspicious:
​https://trac.ffmpeg.org/attachment/ticket/10191/RTL-2023-02-11_19h39m_230213_1645.log#L7411

	EOF in input file 0

after which it soon all terminates. The uploaded file is recorded by Raspberry Pi stock installed hevc-encoding program. I didn't mess with the complete media file (which is larger than this sample) in any way. I only fed it to FFmpeg, and once I kept getting this crash with it, I localazed where the crash is happening, and cut that much out for the uploaded sample.

And, as I said, I get more of files that do the GPU reset and crash, but randomly, not so very often. Unfortunately that often breaks my Wayland Sway compositor and I get to be left with just the console, all work lost.

What could be causing this?

[quinkblack]

I can't reproduce the issue with intel integrated graphics card, looks like it's an AMD GPU specific issue.

[Miroslav Rovis]

Replying to quinkblack:

I can't reproduce the issue with intel integrated graphics card, looks like it's an AMD GPU specific issue.

Thanks for looking into it. Also I have so far missed to provide the other requirements as per https://ffmpeg.org/bugreports.html
If it does not prove to be too hard for me I'll try and see if I can get any more info on this bug with gdb, but that remains to be seen if it isn't too hard for me.

Also I wonder if anybody else encounter similar issues.