MJPEG Tools Svn

Completed initial experiments with gradient descent.
Solved problem with inaccurate fast integer quantisation
calculations (for now.... maybe) and completed porting over
of available MMX routines. There's really very little left in
the code that can benefit from this. One or two more routines plus
extending the fast MMX stuff to the other quantisation routines.

Basically, if it survives a couple of movie-length tests its
ready for beta.

O.k. now the bugs are out some polish. First cut a faster C routines
for quantization. The job is completely done yet - need to tidy up
parameter usage vs. globals and do intra as well and non_intra.
However, it works and its a lot quicker (mainly due to avoiding
the dreaded integer divide (positively glacial on the x86).

The code is now in good share for MMX-ification. I'll probably do
it when I merge the remaining MMX goodies from bbmpeg.

Fixed some nasty and damn subtle bugs causing corruption of the
fast motion compensation data. Took *ages* to pin down...

Found that wretched bug in the the assembler that was
causing weird, rare, random misbehaviour. Guess who
wasn't saving a modified register. Eeeeee-orrrr....

Annoyingly the (whatever it was) random value of an uninitialised
variable was giving *excellent* quantisation bit-allocation.
I'll have to experiment to find it...

Wasted effort - further tuning of motion compensation search. Instad
found my own and compiler bugs and gain .... wow ... and 8% speed-up
for a days work. Should have worked on SSE code for the quantisation
routines.

First compilation of movtar_unify.

Added hierarchical sub-sampling search. Now first search 4*4
and then uses the results of this to seed the 2*2 search.

Needless to say there's MMX/SSE to match.

This commit was manufactured by cvs2svn to create tag 'PreQuad'.

Introduced experimental coefficient magnitude based quantisation/
bit-rate control (so far so good, but early days).

Ported over MMX based contributed to bbmpeg by others
for dct routines for modest speed gain. The idct may be
sub-optimal - the one in the jpeg routines looks better...

*** empty log message ***