[FFmpeg-devel] [PATCH] libavcodec/nvenc.c Reduce initialization time for gpu id > 0

[Timo Rothenpieler]

Seems like a good idea to run the full set of capability checks only on
the device that's actually about to be used.
The only issue I see is that if a system has a bunch of different GPUs,
of which some don't support the required capabilities, one might get
unexpected results with the device indices.

Like, a system with 4 GPUs all with the required minimum nvenc version,
GPU 0 and 2 support capability X.
Now if you're working on GPU 0, with this patch applied, it will list
all 4 GPUs as supported.
So if you want to start another session, on GPU one, it will now skip
that GPU, and instead end up on GPU 2. Or even worse cases might happen,
the logic for this seems way overcomplicated at the moment.

This is not only an issue with this patch applied. At the moment it also
happens that depending on the capabilities required for an encode, the
device indices might mean something else, depending on which GPUs
support the required caps.
That seems bad to me, having universally unique indices on the same
machine no matter what settings are to be used, with some of them
failing if trying to use them, would be much better.

So what I propose instead of this patch:

Make the device index independent of which GPUs support the required caps.
Basically just means that
https://github.com/FFmpeg/FFmpeg/blob/f7e9275f83ec116fc859367d61998eae8af438fc/libavcodec/nvenc.c#L365
should compare against idx instead of the nvenc_device_count.

Add another check here:
https://github.com/FFmpeg/FFmpeg/blob/f7e9275f83ec116fc859367d61998eae8af438fc/libavcodec/nvenc.c#L340
That aborts early if the requested ctx->device does not match idx,
achieving the desired speedup.


Didn't fully plan this out yet, but that plus some cosmetics, like
adjusting error messages and so on, should work.