Fixed: ffmpeg with Hardware acceletarion on Docker Windows Desktop

[ehnwebmaster]

Enable the WSL 2 Backend
This is the primary setting that unlocks GPU passthrough.

Open Docker Desktop.
Click the Settings (Gear Icon) in the top-right corner.
Stay on the General tab.
Ensure the box "Use the WSL 2 based engine" is checked.
If you had to check it, click "Apply & Restart".

Check if nvidia-smi is available

docker run --rm --gpus all nvidia/cuda:12.2.0-base-ubuntu22.04 nvidia-smi

Force the use of nvidia

-e VERTD_FORCE_GPU=nvidia

and

-e NVIDIA_DRIVER_CAPABILITIES=all

Errors:

cannot load libnvidia-encode.so.1

Cannot load libnvcuvid.so.1

hwaccel initialisation returned error.

or

ERROR: nvenc requested, but not all dependencies are satisfied: ffnvcodec

How to resolve inside the container:

ldconfig -p | grep nvidia

You should view:

libnvidia-encode.so and libnvcuvid.so

Those libraries are not found in the container so i copied manually from the docker root desktop at Windows

Open shell, cmd

Execute:

wsl

and

cd /usr/lib/wsl/lib

ls

Output

docker-desktop:/usr/lib/wsl/lib# ls
libcuda.so libnvcuvid.so.1 libnvidia-opticalflow.so
libcuda.so.1 libnvdxdlkernels.so libnvidia-opticalflow.so.1
libcuda.so.1.1 libnvidia-encode.so libnvoptix.so.1
libcudadebugger.so.1 libnvidia-encode.so.1 libnvoptix_loader.so.1
libd3d12.so libnvidia-gpucomp.so libnvwgf2umx.so
libd3d12core.so libnvidia-gpucomp.so.580.102.01 nvidia-ngx-updater
libdxcore.so libnvidia-ml.so.1 nvidia-smi
libnvcuvid.so libnvidia-ngx.so.1

Using docker cp copy all the libraries from docker-desktop to the container:

export LD_LIBRARY_PATH=/usr/lib/wsl/lib:$LD_LIBRARY_PATH
ldconfig

How to check:

env | grep NVIDIA

ffmpeg -encoders | grep nvenc

ldconfig -p | grep -E "libavcodec|libnvcuvid"
libnvcuvid.so.1 (libc6,x86-64) => /usr/lib/wsl/lib/libnvcuvid.so.1
libavcodec.so.60 (libc6,x86-64) => /lib/x86_64-linux-gnu/libavcodec.so.60

Docker logs container vertd

starting vertd
working w/ ffmpeg 6.1.1-3ubuntu5 and ffprobe 6.1.1-3ubuntu5
INFO vertd] using GPU from environment variable VERTD_FORCE_GPU: NVIDIA
INFO vertd] detected a NVIDIA GPU -- if this isn't your vendor, open an issue.

[azurejelly]

It's been a while since I last used Windows as a daily driver, but from my testing NVIDIA GPUs used to work just fine with little to no workarounds on Docker Desktop (I even mentioned it in #7). The only thing you might need to do is set VERTD_FORCE_GPU to nvidia to remove the big "vertd couldn't find a proper gpu" message every time the container starts.

-e NVIDIA_DRIVER_CAPABILITIES=all

The NVIDIA_DRIVER_CAPABILITIES environment variable is already set by the Dockerfile:

vertd/Dockerfile

Line 17 in d7c4aad

ENV NVIDIA_DRIVER_CAPABILITIES="all"

I don't know if Intel/AMD GPUs work and I don't have the hardware to test everything out. That's why I added a note saying that Docker Desktop is unsupported to the documentation.