[FFmpeg-devel] [PATCH] avcodec: add Actimagine VX video decoder

[Lynne]

Mar 15, 2021, 22:02 by fnouwt2 at gmail.com:

> It's actually closer to normal yuv than ycocg. If you look at the
> coefficients of normal yuv
> r = y + 1.14v
> g = y - 0.39u - 0.58v
> b = y + 2.03u
>
> ycocg
> r = y + co - cg
> g = y + cg
> b = y - co - cg
>
> the format used in actimagine
> r = y + 2v
> g = y - 0.5u - v
> b = y + 2u
>
> You can see it's more like yuv than ycocg. That's also why currently the
> decoded colors still look "alright". I think it wouldn't be a good idea to
> use converted ref frames and then convert back as it would likely introduce
> errors. But like you are saying, this coded is as far as I know, never used
> for large frame sizes, so it shouldn't really be an issue to have an extra
> frame and it prevents other problems.
>

Right, I couldn't remember the YUV formula offhand, and the shifts
made it look like YCoCg.
In that case, you can generate the magic constants to multiply, and
then shift the v and u components. For such small numbers you can
even do it by hand, for example 1.14/2 = 0.57 = (584 * v) >> 10.
You'll want to go as high as you can when you multiply without
overflowing, then shift down to decrease rounding error (it always
rounds to 0).