building Docker image for hardware accel - ffmpeg binary not found

[WGandy]

Using docker-compose but having issues with ffmpeg binary not being found. It took me a bit till dawned that the nextcloud image being pulled doesn't have ffmpeg and other goodies installed in it and it would be best to build from a Dockerfile. So I put the Nextcloud "Full" Dockerfile in my root nextcloud docker compose folder and switched image to "Build" in the compose file. It seems to build correctly but I'm still getting the error.
I also see a line in the memories hardware acceleration page that suggests that a file called start.sh is needed? I copied the contents to a file with that name and put it in the root folder and used "docker compose --pull" and then "docker compose up" but I'm still having this same error. The path for ffmpeg in the admin page is simply "ffmpeg". I've tried changing it to usr/local/bin/ffmpeg etc but no fix to the error.
What am I missing, and if I'm missing something, then it should be documented?
Thanks,

[pulsejet]

Try running which ffmpeg inside the container. That should say the path if it's installed correctly.

[WGandy]

So, there should be ffmpeg running in the nextcloud container AND in the go-vod container?

[WGandy]

AHA! As expected, ffmpeg was installed in /usr/bin/. I tried placing that path manually on the admin page previously but it didn't find it. I just tried it again and walla, it found it. Not sure what I changed that would affect that.

But go-vod is now failing. cURL error 7: Failed to connect to go-vod port 47788
AND
VA-API device /dev/dri/renderD128 not found.

[pulsejet]

If you are running go-vod externally, then that container needs ffmpeg too. For the curl error, either go-vod isn't running or you didn't configure the "connect" url correctly.

[WGandy]

"connection address" in the admin page? Or something else? I do indeed have go-vod as a separate container in my docker compose file.

[pulsejet]

Connection address, yes. Check the hardware acceleration wiki page.

[WGandy]

removing this line from my config.php appears to have fixed the issue with vod failing.
'memories.vod.connect' => 'go-vod:47788',

in the admin page, I can see that the address is now '127.0.0.1:47788' instead of 'go-vod:47788'

But I'm still seeing the va-api error. I have both intel quicksync and an nvidia card in the server. I would like to experiment to see which works better.

[WGandy]

I'm not sure that I understand why go-vod has to bottom out at 127.0.0.1. Can't it be reached by its container name over a docker network?

[WGandy]

So, based on the admin page, it appears that I have ffmpeg and go-vod working. VA-API is still giving me the error and I don't know that nvidia has drivers installed or not.

But, regardless of whether I have go-vod enabled or not, videos are not playable in Memories. I get a message at the top "The media could not be loaded, either because the server or network failed or because the format is not supported." I do get a "preview" image of the video but it won't play. And most of the videos I'm testing are mp4 files (with an h.264 encoding, I believe) and should be quite playable. In fact, they play from the files tab of nextcloud but not in Memories.

[revr3nd]

Nvidia support still have to custom compile ffmpeg to include nvenc.
At some point, the Dockerfile.nvidia was removed, and the regular Dockerfile was changed to build on top of the ffmpeg docker image. The ffmpeg image does not include nvenc support.

Below is the Dockerfile.nvidia I have written to compile ffmpeg with nvenc support:

FROM nvidia/cuda:11.8.0-base-ubuntu20.04
ENV NVIDIA_VISIBLE_DEVICES=all
ENV NVIDIA_DRIVER_CAPABILITIES=all

WORKDIR /app
ENV DEBIAN_FRONTEND=noninteractive

#Install ffmpeg dependencies
RUN set -e; \
    apt-get update ;\
    apt-get install -y \
      sudo curl wget \
      autoconf libtool libdrm-dev xorg xorg-dev openbox \
      libx11-dev libgl1-mesa-glx libgl1-mesa-dev \
      xcb libxcb-xkb-dev x11-xkb-utils libx11-xcb-dev \
      libxkbcommon-x11-dev libxcb-dri3-dev \
      cmake git nasm build-essential \
      libx264-dev \
      libffmpeg-nvenc-dev clang

#Install codec headers
RUN set -e ;\
    git clone --branch sdk/11.1 https://git.videolan.org/git/ffmpeg/nv-codec-headers.git ;\
    cd /app/nv-codec-headers ;\
    make install

#Compile ffmpeg
RUN set -e ;\
    git clone --depth 1 --branch n5.1.3 https://github.com/FFmpeg/FFmpeg ;\
    cd /app/FFmpeg ;\
    ./configure \
        --enable-nonfree \
        --enable-gpl \
        --enable-libx264 \
        --enable-nvenc \
        --enable-ffnvcodec \
        --enable-cuda-llvm ;\
    make -j"$(nproc)" ;\
    make install ;\
    ldconfig

COPY run.sh /go-vod.sh

EXPOSE 47788

ENTRYPOINT ["/go-vod.sh"]

To build, you should use the following command:
docker build -t go-vod:nvidia -f Dockerfile.nvidia .

[pulsejet]

I don't have NVENC to test, but the image readme does say that it is supported https://hub.docker.com/r/linuxserver/ffmpeg

[revr3nd]

Doing a quick compare of build options. It looks like the lscr build does not include the CUDA nor NPP scalers, which is why it doesn't work.

the lscr ffmpeg image uses the following build options:

--disable-debug
--disable-doc
--disable-ffplay
--enable-cuvid
--enable-ffprobe
--enable-gpl
--enable-libaom
--enable-libass
--enable-libfdk_aac
--enable-libfreetype
--enable-libkvazaar
--enable-libmp3lame
--enable-libopencore-amrnb
--enable-libopencore-amrwb
--enable-libopenjpeg
--enable-libopus
--enable-libsvtav1
--enable-libtheora
--enable-libv4l2
--enable-libvidstab
--enable-libvmaf
--enable-libvorbis
--enable-libvpl
--enable-libvpx
--enable-libwebp
--enable-libx264
--enable-libx265
--enable-libxml2
--enable-libxvid
--enable-nonfree
--enable-nvdec
--enable-nvenc
--enable-opencl
--enable-openssl
--enable-stripping
--enable-vaapi
--enable-vdpau
--enable-version3

This is quite a bit more than what go-vod needs, but more critically go-vod requires either --enable-cuda-llvm or --enable-libnpp to work with the CUDA or NPP scalers. This is why it still requires custom compiling.

Recommendation would be to either pivot away from a hardware scalers when using nvenc and just use the software scale filter, or stick with the custom compile path.

[pulsejet]

Okay! I see two paths ahead:

1. We need to find out if there's a reason why the ffmpeg isn't built with either scaler. Does this have to do with licencing? In that case, is there another image which does ship with these / can we build and ship one legally?
2. (less ideal) If we can't, this Dockerfile should be put back to the repo and update the docs with instructions. This'll be more effort for those using NVENC, unfortunately.