Date: Mar 30, 2007 9:00 PM
Subject: Re: [Ffmpeg-devel] Motion Estimation in snow.c for Waevelet encoded frames (DWT)
Originally committed as revision 8579 to svn://svn.ffmpeg.org/ffmpeg/trunk
dia_size=512 + hexagon size (513 is the normal one used in h264)
large to small diamond search
dia_size=256 + diamond size
Originally committed as revision 7204 to svn://svn.ffmpeg.org/ffmpeg/trunk
this has pretty much no quality or speed effect except very small random changes
Originally committed as revision 7202 to svn://svn.ffmpeg.org/ffmpeg/trunk
mpegvideo.h:777: warning: ‘inline’ is not at beginning of declaration
motion_est_template.c:236: warning: ‘inline’ is not at beginning of declaration
Approved by Diego.
Originally committed as revision 6591 to svn://svn.ffmpeg.org/ffmpeg/trunk
different intra block coding (previous was just an ugly hack)
1.8% bitrate reduction -0.01PSNR (foreman@352x288 qscale=8)
1.5% bitrate reduction +0.05PSNR (foreman@352x288 qscale=1)
Originally committed as revision 3416 to svn://svn.ffmpeg.org/ffmpeg/trunk
wavelet based compare functions
make epzs_motion_search() more flexible so it can be used for a wider range of block sizes
make get_penalty_factor() independant of MpegEncContext
Originally committed as revision 3410 to svn://svn.ffmpeg.org/ffmpeg/trunk
be somewhat more tollerant on invalid input
return INT_MAX instead of -1 for unuseable mv/mb types as that ensures nicely that they arent used
initalize limits earlier for b frames
a few more asserts to check for out of picture vectors
Originally committed as revision 3213 to svn://svn.ffmpeg.org/ffmpeg/trunk
replace ugly macros by always_inline functions, that way its much more readable and flexible as always_inline can simply be removed while the macros couldnt be
about 0.5 % speedup with default parameters
Originally committed as revision 3037 to svn://svn.ffmpeg.org/ffmpeg/trunk
multithreaded/SMP encoding for MPEG1/MPEG2/MPEG4/H263
all pthread specific code is in pthread.c
to try it, run configure --enable-pthreads and ffmpeg ... -threads <num>
the internal thread API is a simple AVCodecContext.execute() callback which executes a given function pointer with different arguments and returns after finishing all, that way no mutexes or other thread-mess is needed outside pthread.c
Originally committed as revision 2772 to svn://svn.ffmpeg.org/ffmpeg/trunk
interlaced mpeg2 encoding
P & B frames
rate distored interlaced mb decission
alternate scantable support
4mv encoding fixes (thats also why the regression tests change)
passing height to most dsp functions
interlaced mpeg4 encoding (no direct mode MBs yet)
various related cleanups
disabled old motion estimaton algorithms (log, full, ...) they will either be fixed or removed
Originally committed as revision 2638 to svn://svn.ffmpeg.org/ffmpeg/trunk
more direct use of the new mb_type stuff instead of codec specific stuff
runtime mb_type debug output h264/h263 variants/mpeg1/2/4
error concealment /resilience for mpeg1/2
various minor optimizations
Originally committed as revision 1746 to svn://svn.ffmpeg.org/ffmpeg/trunk
user specified amount of MV predictors from the last frame
b frame MV predictor scaling fixed
Originally committed as revision 1384 to svn://svn.ffmpeg.org/ffmpeg/trunk
4mv+b frames encoding finally fixed
chroma ME
5 comparission functions for ME
b frame encoding speedup
wmv2 codec (unfinished)
user specified diamond size for EPZS
Originally committed as revision 1365 to svn://svn.ffmpeg.org/ffmpeg/trunk
