[Ffmpeg-devel] diff_pixels_mmx: suggestion for speed improvement
Sun Nov 13 23:44:56 CET 2005
I noticed that diff_pixels_mmx is one of the biggest consumers of CPU time -
marginally under ff_fdct_sse2 on a P4.
diff_pixels_mmx only subtracts a 8x8 block from another 8x8. Do you think we
would get an improvement if there was also a 16x16 version? It might help with
MPEG1/2 and cases in MPEG4 when the MVs are for 16x16 macroblocks.
Perhaps an SSE2 version might also help.
Unfortunately I don't know any assembler so I can't write it myself.
More information about the ffmpeg-devel