HW Accelerated encoding

[pptudela]

Has anyone successfully configured ffmpeg with Quick Sync or Nvenc to use hardware acceleration for encoding the different quality copies?

I presume it's just a matter changing the ffmpeg commands on tasks.py and adequately compiling ffmpeg, but was wondering if someone has done this before.

[DidacValenciano]

Was hoping to find an answer to that, and just find your question. It's not only matter of tasks.py but helpers.py too.

I'm trying some combinations. Without luck. Here's a sample of the logs:

logs/debug.log:FileNotFoundError: [Errno 2] No such file or directory: 'ffmpeg -hwaccel qsv'

It will need more effort ...

[Anduin2017]

Hi, any updates?

I bought an Nvidia Tesla P40. However, seems I couldn't benefit from it for running MediaCMS.

Any solution to use hardware accelerated encoding?

[Anduin2017]

For example, to encode a video to HEVC with CPU, we can run:

ffmpeg -i input.mp4 -preset slow -codec:a copy -codec:v libx265 -crf 20

But with GPU, we need to run:

ffmpeg -i input.mp4 -preset slow -codec:a copy -codec:v hevc_nvenc -rc:v vbr -cq:v 20 -rc-lookahead 10 -profile:v main10

According to my test, for CPU E5-2696v4 16 cores and Nvidia Tesla P40, using GPU, P40 has 5 times performance advantage.

[mgogoulos]

Hi @Anduin2017 , thanks for sharing this interesting info!

One would have to edit files/helpers.py and add the parameters manually for now to ffmpeg, since this is not possible through a web interface.
Cheers

[MyKeyTeeNZ]

Hi @mgogoulos - thanks for that, do you have an example of what we need to change in the helpers.py file for this to work? My GPU is Nvidia, just not sure where to edit at this point? Thanks in advance!

[mgogoulos]

have a look on files.helpers.produce_ffmpeg_commands, this is calling files.helpers.get_base_ffmpeg_commands , and is providing the ffmpeg command that is run for a single encoding (or commands, because there is a two pass encoding in some cases)

there's a lot of logic there and it's not easy to decouple, however it's a start! if you know what extra arguments ffmpeg is going to need, or need to adjust existing ones, that's the place to look

[dmurphynj]

helpers-vaapi.txt

Here is an updated helpers.py that I hacked FFmpeg vaapi support into.

I'm not a coder, don't play one on TV, nor am I an ffmpeg expert, but seems to be working ... haven't tested it much but it's a starting point.

Run a diff against the existing helpers.py and you can see what was changed.

[BadVolf]

Hey guys, here you go:

Rename it to helpers.py
helpers-hvenc.txt

This is working with my the NVidia RTX 3070IT Super. you will see it tests if the media should be hardware encoded and if so, it has filters for h.264 and h265

if use_hardware_encoding: encoder = "h264_nvenc" if codec == "h264" else "hevc_nvenc"

Sorry, I don't know anything about github, just thought I would share this with you.

Let me know if you have any improvements.

[IllegalD]

Hey guys, here you go:

Rename it to helpers.py helpers-hvenc.txt

This is working with my the NVidia RTX 3070IT Super. you will see it tests if the media should be hardware encoded and if so, it has filters for h.264 and h265

if use_hardware_encoding: encoder = "h264_nvenc" if codec == "h264" else "hevc_nvenc"

Sorry, I don't know anything about github, just thought I would share this with you.

Let me know if you have any improvements.

Your two configs have been a godsend, thanks mate! For anyone else trying to get hardware transcoding working, I found it easiest to replace ffmpeg with jellyfin-ffmpeg. Have had it working great on vaapi and nvenc using these two configs. It also helps to bump up the bitrates too, as software transcoding still looks a little better.

[dmurphynj]

Your two configs have been a godsend, thanks mate! For anyone else trying to get hardware transcoding working, I found it easiest to replace ffmpeg with jellyfin-ffmpeg. Have had it working great on vaapi and nvenc using these two configs. It also helps to bump up the bitrates too, as software transcoding still looks a little better.

Glad it helps! I agree - using the jellyfin-ffmpeg package is a big, big help since it has everything compiled in.

I'm doing some interesting tricks here - my mediacms instance runs in a VM, under proxmox, on an Intel N100 CPU. It took quite a bit of work to get GPU passthrough to work, but I'm passing through the internal GPU in the N100 chip to the mediacms VM, so ffmpeg has direct access to it.

Complex config but working great - huge, huge improvement in transcode time vs. software-only.

The N100 CPU is ridiculously impressive for the cost... There's certainly better CPUs (and GPUs) out there, but for a 6-watt CPU? Insanely good.

[IllegalD]

I'm running it in an Ubuntu LXC an old Nuc with an i3-6100U, the vaapi performance is fantastic compared to CPU transcoding!

[mabruzzo3169]

Anyone have any interest in helping a new user replace ffmpeg with jellyfin-ffmpeg?

[Bewinxed]

How to make this work in the docker-compose image?

[mrjoshuap]

I've successfully added hardware encoding for Rockchip RK3588S arm64's, which is similar in modifications. I was however interested in thoughts on how to abstract the encoding profiles to make it a little easier to support different hardware encoders.

For my purposes, I had a custom build version of ffmpeg, bento4, and pass through transcoding hardware into the worker containers. If you're using a container deployment, make sure you pass through the devices into your docker-compose file. For instance, on RK3588s, you have to add the following devices to celery_worker:

  celery_worker:
    devices:
      - "/dev/dri:/dev/dri:rwm"
      - "/dev/dma_heap:/dev/dma_heap:rwm"
      - "/dev/rga:/dev/rga:rwm"
      - "/dev/mpp_service:/dev/mpp_service:rwm"
      - "/dev/video-dec0:/dev/video-dec0:rwm"
      - "/dev/video-enc0:/dev/video-enc0:rwm"

If you're using Nvidia or other hardware encoders, the devices will be different but will probably still be required...

It's also worth mentioning that when you are testing encoding profiles, grabbing the exact ffmpeg command being run is extremely important because a permissions problem on your devices usually ends up with a Segmentation Fault and basically no usable error messages in any of the logs.