Nvidia-docker: How to link to nvidia driver library while building image?

[mythly]

I'm trying to compile OpenCV with NVCUVID in nvidia-docker image. The problem is that, libnvcuvid.so.1 is part of nvidia driver and only mounted into /usr/local/nvidia/lib64/ in runtime(nvidia-docker run ...). So OpenCV cannot find NVCUVID library while building image(docker build ...).

The trivial solution would be copying those libraries into the image in build-time, but losing the driver-agnostic advantages of nvidia-docker. Is there any better idea?

[3XX0]

Unfortunately we don't have prebuilt stubs for libnvcuvid (see here).
You can create a stub library for libnvcuvid in your container for linking purposes only.
libnvcudvid doesn't have many symbols so this should be pretty straightforward.

[flx42]

For nvcuvid, it looks like the official solution is now to use the files provided with the Video Codec SDK and that will take care of dlopen/dlsym of the cuvid symbols at runtime. So OpenCV might need to update their code.

[GudarJs]

Can you give to me a hint or guide on how to create the libnvcuvid stub library?