Andreas Öman wrote: > Michael, > > I'm quite satisfied with the shape of the AAC-LC decoder now.
I've done various performance benchmarks at some interesting places and might as well comment on it right away. ========================== ivquant(): This is called quite often and therefore pow() is accelerated with a table: Increasing the table size wont gain much speed: table size 8192: 700 dezicycles in ivquant, 16777086 runs, 130 skips table size 512: 701 dezicycles in ivquant, 16777084 runs, 132 skips original: (256 entries) 702 dezicycles in ivquant, 16777089 runs, 127 skips table size 32: 707 dezicycles in ivquant, 16777088 runs, 128 skips table size 1: 1694 dezicycles in ivquant, 16776923 runs, 293 skips ========================== vector_fmul_add_add_add() There is a suggestion to move this into dsputil(), but i'll rather await your comment on the subject. It's only used by the short window transform (which are used relatively seldom) I wrote an pure SSE-version of it, and ... SSE: 247843 dezicycles in eight short sequence, 1024 runs, 0 skips current: 257722 dezicycles in eight short sequence, 1024 runs, 0 skips ... as you can see the speed gain is just a few percent. Since it's seldom used the speed gain isn't even measurable on a whole file (2-3 minutes of audio) _______________________________________________ FFmpeg-soc mailing list [email protected] https://lists.mplayerhq.hu/mailman/listinfo/ffmpeg-soc
