[MPlayer-dev-eng] TODO (mplayer on PPC)

[Michael Niedermayer]

Hi

On Wednesday 13 February 2002 23:56, Mike Melanson wrote:
> On Wed, 13 Feb 2002, Michael Niedermayer wrote:
> > for (i)dct specific docs ... well i dunno many, mostly the different
> > versions from mplayer/ffmpeg, especially the older ones as they are more
> > readable but beware, the idct is pretty sensitive to rounding errors, a
> > few rounding errors at the wrong place and ppl will turn green after a
> > few seconds, or get stripes ;)
>
> 	I have a question about DCT code in general. Is there enough
> commonality between the transforms used in various algorithms to be able
> to separate the transforms out into a common library (say, libdct.a)? It
> seems to me that a lot of effort goes into optimizing the DCT of various
> algorithms for assorted platforms and I wonder if there's a lot of
> redundant effort that could be eliminated with such a library, internal to
> MPlayer. I don't know nearly enough about the DCT to know one way or
> another.
well, iam certainly no (i)dct expert either ...
but heres a list of the ones from ffmpeg and their special requirements ...
C IDCT			pre-permutation2
simple-IDCT C		nothing
simple-IDCT MMX		pre-permutation1
walken-IDCT MMX 	pre-permutation2
C DCT			post-scaleing
MMX DCT			nothing

the reason for this mess is that some of these extra operations can be 
performed for free, for example the dct needs to be scaled afterwards anyway 
(in the quantizer) and during decoding the zigzag permutation must be 
reversed anyway and it doesnt matter if its a more complex permutation 


>
> 	Also, is there any good place to obtain DCT test vectors?
dct-test.c from ffmpeg should be ok for testing (i)dct against the reference 
IEEE double? float thingy but just a good score here does not mean that its 
good enough ...

[...]

Michael