Better checks for ffmpeg nvenc version

[hmaarrfk]

Comment:

Apparently the versions are listed here:

FFmpeg/FFmpeg@n7.0.2...n7.1#diff-ba8b555333cbe9c4d768147d8a7abb152281257770aa314cc057a31f43460fecR247

so we could improve our version checking

[hmaarrfk]

We'll need to integrate newer ffnvcodecs conda-forge/ffnvcodec-headers-feedstock#26

[jakirkham]

Thanks Mark! 🙏

Also wonder if we can now drop this as we are building with newer CUDA 12.x's

ffmpeg-feedstock/recipe/conda_build_config.yaml

Lines 29 to 32 in 1705b1d

	cuda_version_for_ffnvcodec:
	- None
	- 12.2 # [linux64 or win]
	- 12.4 # [linux64 or win or aarch64]

Edit: Think we can also use this opportunity to add CUDA 12.8 ( conda-forge/conda-forge-pinning-feedstock#6980 )

[hmaarrfk]

no you can't since the encoding API doesn't match. thats why i had to build the matrix. i hit this problem myself.

[hmaarrfk]

For more info see these defines https://github.com/FFmpeg/FFmpeg/blob/43be8d07281caca2e88bfd8ee2333633e1fb1a13/libavcodec/nvenc.h#L96 which fix features depending on the compile time status. I think unless we package the equivalent of libnvidia-encode and libnvidia-decode we need to build out the matrix.

[hmaarrfk]

i was truthfully surprised that the CUDA version checks were not runtime checks.

I wonder how software even ships cuda powered FFMPEG? maybe just supporting a subset of features....