[MPlayer-dev-eng] fast SPP

[D Richard Felker III]

On Mon, Mar 21, 2005 at 04:19:53AM +0500, Nikolaj Poroshin wrote:
> Hello,
> 
> Probably you remember my post about a way to speed up the SPP filter.
> Now I want to add this code to the Mplayer project :)
> However, during passed time, I've made some improvements: thresholding
> now uses customizable matrix (1% slower against fixed), and some minor
> speed improvements.
> I've found out that normal threshold (as in the current Mplayer code,
> normal dct->threshold->normal idct) has better quality against my
> previous variant (aan dct w/o postscale->threshold->aan idct w/o
> prescale), in terms of ringing artefacts. Now, customizable threshold
> matrix allows for the same results, as current Mplayer code.
> (aan dct w/o postscale->threshold[x][y]->aan idct w/o prescale).
> 
> So, first question - create a new filter or patch existing?
> If new, then it means code duplication and probably some difficulties
> (I don't fully understand things like mpi, etc).
> If I patch existing, then it looses soft_threshold (softthresh_c)
> ability (I guess, forever). But you can play with threshold matrix
> coefficients instead. Also it looses non-mmx acceleration (ppc) (I
> think, not forever). But I guess ppc acceleration gains not more than
> 4x speedup(?), and my implementation in the C variant is about 4 times
> faster than current(also C). So, no real speed loss for ppc at first
> time (and later, someone can optimize it for ppc). And also, my
> implementation has only 4th and 5th processing levels, but I think
> this doesn't matter much (4th level can work at 512x272 resolution,
> 25 fps at Celeron 633).

Wow. I hope this gets committed in some form or another.

Rich