Hi,

this is a bug report because some people get stuck at this point, but I'm not quite convinced it's your fault :)

Issue: cmake fails to find libnvcuvid when you try to build decord in an nvidia cuda container will all necessary libs linked in (-e NVIDIA_DRIVER_CAPABILITIES=all - this should provide the container with pretty much everything that's in CUDA, cuvid, cudnn and so on).

[edit]: Just to add, this is using the official nvidia pre-built cuda dev container nvidia/cuda:10.1-cudnn7-devel

That should obviously not happen. I've also successfully built ffmpeg with all the cuvid accelerations, dlib and other software - all successfully using cmake.

So why is this happening? The container has libnvcuvid here:

lrwxrwxrwx 1 root root   20 Oct 16 15:24 /usr/lib/x86_64-linux-gnu/libnvcuvid.so.1 -> libnvcuvid.so.450.57
-rw-r--r-- 1 root root 3.6M Jul  5 15:12 /usr/lib/x86_64-linux-gnu/libnvcuvid.so.450.57

This is kind of strange. We know that libnvcuvid is not part of cuda but part of the driver. The host has 450.57 installed, so that's why the lib is versioned and provided via a link. But!! Nvidia itself recommends against using *.so.1 to load dlibs - instead, one apparently should use libnvcuvid.so(no suffix). I guess that's what you guys do.

After linking /usr/lib/x86_64-linux-gnu/libnvcuvid.so.1 to libnvcuvid.so, everything works as expected.

So, maybe you want to add some safety check to your cmake config to catch this behavior? Or document it somewhere? (I've at least written up this report so people can find it via the issues).

BTW, thanks a lot for decord and your hard work!