[MPlayer-dev-eng] Re: [LONG] new results & bugs report

[Romain Dolbeau]

Romain Dolbeau wrote:

> Of course the results need to be examined more carefully than
> with "sort".

Only considering the "vme=4 & mbd=2' cases (2352 combinations):

1) qpel always improve PSNR
2) mv0 nearly always improve PSNR (> 90% of the cases)
3) v4mv nearly always improve PSNR (> 95%), but the "bad" cases
    can be really bad.
4) qprd is mostly a loss (about 2/3 of the cases)
5) trell always improve PSNR
6) cbp is nearly always a loss (wrt trell alone)
7) *cmp=2 is marginally better than *cmp=3, while
    *cmp=6 is a little bit better than *cmp=2
8) *dia doesn't make much of a difference, with
    *dia=1 the "best" and *dia=-2 a close second,
    and *dia=-1 the worst.
9) vmax_b_frames is undecided (sometimes good, sometimes bad)

If we fix qpel,vm0 and trell, we have 384 combinations left:

3) v4mv become a sure winner (always improve PSNR)
4) qprd is still mostly bad
6) cbp become a sure loser (always diminish PSNR)
7) the *cmp order is inverted, with *cmp=3 taking
    the lead and *cmp=6 trailing !
8) *dia make even less of a difference
9) vmax_b_frames is still an open question, although
    the balance it's a bit in favor of 1 rather than 0

-- 
Romain Dolbeau