[FFmpeg-devel] [PATCH] Add NVENC encoder

[Philip Langdale]

On Sat, 29 Nov 2014 19:00:02 +0100
Timo Rothenpieler <timo at rothenpieler.org> wrote:

> > does supporting these additional features need the extra complexity
> > that the nvidia version has ?
> > or could these features be added into your version while keeping its
> > simplicity ? (note do not copy any code from the nvidia one as its
> > not redistriutable nor *GPL compatible with the current license
> > headers)
> 
> I haven't even looked at their code for quite exactly that reason.
> The primary problem with NVENC is, that there is only very poor
> documentation available, so I can only implement stuff based on
> reading the header and experimenting with the API and the available
> examples. NVIDIA employees clearly are in a better position for this,
> as they most likely have a much larger insight into how NVENC works
> internaly. I'm quite sure all of their features could be added to my
> version.

I'm already 'polluted' until they correct their licensing, but I can at
least describe what they have.

The main piece of complexity in their implementation is they have an
additional abstraction layer between the ffmpeg encoder implementation
and the nvenc API. So this layer 'insulates' the two APIs. From a
licensing point of view it's meaningless, but I think that's the reason
why it exists, based on a brief email exchange with them. If that was
squashed, it would make theirs more straight forward. There's also a
bunch of utility code that handles argument mapping between x264 args
and nvenc args. As I've written elsewhere - I think this is a very
useful thing to do, as people are, broadly, familiar with how to tell
x264 to do stuff.

And as I mentioned before, they have a directx linked initialisation
path as an alternative to the cuda one. I'm not sure what conditions
that would be more useful under, but it's documented in the SDK.

Hopefully, they will respond positively on Monday and will get engaged
here, and sort out their licensing so that we can all work on a single
implementation.

Thanks!

--phil