GPU Hang During Multiple Hardware Decodes/Encodes
Performing parallel hardware operations (decode or encode) for HEVC or H.264 results in a GPU hang. Regardless of video format, frame rate, chrominance, or bitrate, two or more instances result in a "GPU Hang" and requires a power cycle to recover the unit.
Stream source type does not matter, both files and UDP transfers result in a "GPU Hang", but files reproduce the error faster assuming it is long enough to run in parallel for more than ten seconds. As long as both terminals are performing parallel decodes/encodes the issue is reproduced every time.
The issue has been reproduced on several setups: Ubuntu 20.04 with v5.14.0-1024-oem and v5.16.12-051612-generic, Ubuntu 21.10 with 5.13.0-19-generic and 5.17.0-051700rc7drmtip20220308-generic
Steps to reproduce
- Open two terminals
- Run the following command in both terminals at the same time, where 'file.trp' is a supported HEVC or H.264 video stream. ffmpeg -hwaccel qsv -c:v hevc_qsv -i file.trp -f null -
Requested information (dmesg, drm error, dmidecode, unit info) are in attached gpu_hang ZIP file: gpu_hang.zip