[Support]: Need libmfxgen1 library installed to support Alder Lake CPU QSV transcoding

Describe the problem you are having

I have an Alder Lake gen12 CPU that I’m utilizing QuickSync for hardware transcoding.

In order to use the qsv device, the Docker image needs to have libmfxgen1 installed as per the Intel compatibility matrix

The stock Docker image will fail to initialize the hardware encoder/decoders using qsv if this package is not installed (see general Ubuntu bug)

After adding the[ Intel repository](deb [arch=amd64] https://repositories.intel.com/graphics/ubuntu focal main) and installing libmfxgen1 via apt-get results in the ffmpeg instance being able to correctly utilize qsv.

Version

0.10.1-83481AF

Frigate config file

ffmpeg:
        hwaccel_args:
                - -hwaccel
                - qsv
                - -qsv_device
                - /dev/dri/renderD128
        output_args:
          record: -f segment -segment_time 10 -segment_format mp4 -reset_timestamps 1 -strftime 1 -c:v h264_qsv -c:a copy -an
        input_args:
                - -hide_banner
                - -avoid_negative_ts
                - make_zero
                - -fflags
                - +genpts+discardcorrupt+igndts
                - -rtsp_transport
                - tcp
                - -stimeout
                - 5000000
                - -analyzeduration
                - 2147483647
                - -probesize
                - 2147483647

Relevant log output

Error initializing an internal MFX session: unsupported (-3)

FFprobe output from your camera

.

Frigate stats

No response

Operating system

Other Linux

Install method

Docker Compose

Coral version

M.2

Network connection

Wired

Camera make and model

.

Any other information that may be helpful

No response

[charlesmunger]

Ok, I got qsv hwaccel “working” on beta7 with this config:

  hwaccel_args:
    - -hwaccel
    - qsv
    - -qsv_device
    - /dev/dri/renderD128
    - -hwaccel_output_format
    - qsv
    - -c:v
    - h264_qsv
    - -gpu_copy
    - "on"
  output_args:
    detect: -vf vpp_qsv=format=yuv420p -f rawvideo -pix_fmt yuv420p

This does result in CPU improvement and I do see activity in intel_gpu_top. I think for Alder Lake on linux, qsv doesn’t support yuv420p as a hwdownload format (it does support nv12!), so we need a second pass to convert it to yuv420p in hardware, then use -gpu_copy for accelerated copy from the hardware memory to software. With this set up, each 640x480 7fps reolink detect steam uses 3% of one E-core on my 12700.

I also tried -vf hwdownload,format=nv12,format=yuv420p and it worked, but it it used slightly more CPU than the vpp_qsv configuration, plus I suspect it’s less portable since it’s not really documented which formats are supported for direct hwdownload with which drivers and hardware.

I think we should consider updating the docs since I suspect that users of that configuration are not actually getting hardware acceleration - if I didn’t specify -c:v h264_qsv then I got no errors… but it wasn’t actually using hardware acceleration.

I was also tripped up by the yaml converting on to True implicitly, so I had to add those quotes.

[jasonmadigan]

Update: nevermind, think I figured it out. Unraid currently (6.10.3) uses kernel 5.15.46. Knew it had a number of issues with the iGPU from various threads. Updated to a test build of Unraid, 6.11.0-beta9, which includes a new 5.18.12 kernel. QSV works again, with your config from above (and can also see load again via the GPU). Thanks @charlesmunger

Where does one find information on Unraid 6.11 beta? haven’t seen anything on it

It’s in test, Unraid don’t seem to publicise these as much as the RCs, but there were some links shared on their forums by someone.

Install Plugin > https://s3.amazonaws.com/dnld.lime-technology.com/test/unRAIDServer.plg