[x264-devel] Re: code analysis

[Loren Merritt]

On Mon, 19 Mar 2007, Alex Izvorski wrote:

> Guillaume: actually I've always wondered what the definitive reference
> for using Hadamard is.  There is stuff like JVT-O079 etc, and clearly
> Hadamard does work better than SAD, but: Hadamard is only one of a class
> of transforms, who can say whether another may be even better?  Possibly
> there is something in between the full IDCT and SATD which is also a
> good tradeoff?

H.264's dct is already an approximation of the real dct for faster
computation. I'm sure you could make something even more approximate, or
tweak it to eliminate the renormalization step (which is free when
combined with quantization, but not free in a satd-like context).

> Why 4x4?

satd is 4x4 because dct is 4x4. 8x8 satd is a better match for 8x8 dct.
And comparing 4x4 satd with 8x8 satd provides a decent estimate of
whether 4x4 or 8x8 dct will be better for a given block.

--Loren Merritt

-- 
This is the x264-devel mailing-list
To unsubscribe, go to: http://developers.videolan.org/lists.html