VP8: 30% faster idct_mb
authorJason Garrett-Glaser <darkshikari@gmail.com>
Fri, 23 Jul 2010 02:58:27 +0000 (02:58 +0000)
committerJason Garrett-Glaser <darkshikari@gmail.com>
Fri, 23 Jul 2010 02:58:27 +0000 (02:58 +0000)
commit8a467b2d44d20c1a0b731dce9edeff772732a558
tree259b69473524a9db002d79b90009817f84af4113
parentef38842f0bc97ce5b158f51f3e65aae4164fc6a5
VP8: 30% faster idct_mb
Take shortcuts based on statistically common situations.
Add 4-at-a-time idct_dc function (mmx and sse2) since rows of 4 DC-only DCT
blocks are common.
TODO: tie this more directly into the MB mode, since the DC-level transform is
only used for non-splitmv blocks?

Originally committed as revision 24452 to svn://svn.ffmpeg.org/ffmpeg/trunk
libavcodec/vp8.c
libavcodec/vp8dsp.c
libavcodec/vp8dsp.h
libavcodec/x86/vp8dsp-init.c
libavcodec/x86/vp8dsp.asm