[FFmpeg-devel] [PATCH v2] swscale: add unscaled conversion from yuv420p to p010

[Ronald S. Bultje]

Hi,

On Sep 4, 2016 5:42 PM, "Timo Rothenpieler" <timo at rothenpieler.org> wrote:
>
> On 9/4/2016 4:06 PM, Carl Eugen Hoyos wrote:
> > 2016-09-04 16:02 GMT+02:00 Timo Rothenpieler <timo at rothenpieler.org>:
> >> The purpose of this patch is to make conversion from
> >> yuv420p (8 bit) to p010 (10 bit) fast.
> >
> > Do I understand you correctly that your patch is
> > faster without the change I suggested?
>
> With the &:
> [bench @ 0x600045b80] t:0.011178 avg:0.011172 max:0.018297 min:0.010505
>
> Without it:
> [bench @ 0x600045b80] t:0.008455 avg:0.008517 max:0.015815 min:0.007941
>
> So it is quite a bit faster.
>
> Tested with nvenc hevc10 encoding, and the output is visually identical,
> and the file size is also exactly the same.
> So it seems to cleanly ignore the unused bits.
>
> Also, given that at least microsoft argues with upcasting to 16 bit, the
> approach without zeroing the lsb would be more accurate, as
>
> t << 8 | t
>
> is how one would convert 8 bit to 16 bit.
>
>
> So I'd say going with the faster approach here should be fine.
> If at some point someone runs into something that chokes on the bits
> being non-zero, which I think is highly unlikely, it can be changed back.

Sounds good, thanks for testing and confirming.

Ronald