[Ffmpeg-devel] VP3/Theora Perfection

Michael Niedermayer michaelni
Wed May 18 15:07:10 CEST 2005


On Wednesday 18 May 2005 12:13, Diego Biurrun wrote:
> On Tue, May 17, 2005 at 05:43:08PM -0600, Mike Melanson wrote:
> > 	I just committed some code that uses VLC tables rather than the
> > tree-based get_bits1() or switch/case logic. I have replaced
> > get_fragment_run_length(), get_mode_code(), and
> > get_motion_vector_vlc() so far. Please test as I only have access to one
> > sample in a restricted dev environment right now.
> All the samples from mphq are working fine, but Michael disabled one
> optimization.

yes, md5 of the decoded output changed, i didnt investigate why ...

> > 	For those who want hard numbers for improvement, here they are:
> It's still between 50-100% slower than the binary decoder, which is a
> giant improvement from what it was a few days ago.

theres a lot of things left which need optimization, the way the coefficients 
are "collected" for each block is a absolute cache killer ill try to fix that 
one ...
then theres the loopfilter which shouldnt be done per picture, dequant and 
zeroing of coeffs unneccessarily, mmx optimization of the loop filter, that 
unpack_token() switch thing, ...


More information about the ffmpeg-devel mailing list