Use word-writing instead of dword-writing (with two cached but otherwise
authorRonald S. Bultje <rsbultje@gmail.com>
Sat, 31 Jul 2010 23:13:15 +0000 (23:13 +0000)
committerRonald S. Bultje <rsbultje@gmail.com>
Sat, 31 Jul 2010 23:13:15 +0000 (23:13 +0000)
commit6341838f3ca69c7850aa11b067165ef544cead95
tree7914c26ff9b26b9f544024e56b5032254d27f2b9
parentace7f813cd4b2bc092bd827f7e8257368781e9bb
Use word-writing instead of dword-writing (with two cached but otherwise
unchanged bytes) in the horizontal simple loopfilter. This makes the filter
quite a bit faster in itself (~30 cycles less on Core1), probably mostly
because we don't need a complex 4x4 transpose, but only a simple byte
interleave. Also allows using pextrw on SSE4, which speeds up even more
(e.g. 25% faster on Core i7).

Originally committed as revision 24638 to svn://svn.ffmpeg.org/ffmpeg/trunk
libavcodec/x86/vp8dsp-init.c
libavcodec/x86/vp8dsp.asm