MJPEG Tools Svn

Convert frame_i2f and frame_f2i to SSE, yields 20% performance improvement

use floats for intermediates (speeds things a little and changes 18% of the values!)

code reorg'd so there is one easy to vectorize function (convolveLine)

fix major bug in reversing vert/horiz filter processing order

Process vertical filter first (changes results but enables future optimizations)

Convolve Y completely before U, V instead of doing all horiz Y, U, V, then vert Y, U, V

13.5% faster for 4 elem filters on x86
(extract convolve code to convolveitem, manually increment pointers instead of using indexing)

More fixes for gcc-4 (I hope).

removed defined(NASM) check, is not needed anymore

fix comments