On 11.03.2016, at 19:09, Lou Logan <l...@lrcd.com> wrote: > On Thu, Mar 10, 2016, at 05:42 PM, Ganesh Ajjanagadde wrote: >> ffmpeg | branch: master | Ganesh Ajjanagadde <gajja...@gmail.com> | Tue >> Mar 8 22:08:29 2016 -0500| [66edd8656b851a0c85ba25ec293cc66192c363ae] | >> committer: Ganesh Ajjanagadde >> >> lavc/lpc: exploit even symmetry of window function > > kamedo2 has a comment about this on Twitter, although I can't quite > understand the translation: > > <https://twitter.com/kamedo2/status/708313239482806272>
Not sure, but if I guess: i <= len / 2 loop condition does one step too many if len is even. It is also possible to build a cos table with only one sin and cos. Depending on precision requirements only using cos[n] = cos[n/2]*cos[(n+1)/2]-sin[n/2]*sin[(n+1)/2] sin[n] = sin[n/2]*cos[(n+1)/2]+cos[n/2]*sin[(n+1)/2] I don't know what the speed or the maximum error of that would be though. The number of operations and thus rounding errors that go into a single value is limited by O(log2(len)) though, so it's likely if you use double precision everywhere the result would not be (much) worse off than cosf. The cos formula can for even n possibly optimize precision by doing 2*cos[n/2]*cos[n/2]-1. No guarentee I made no mistakes, but I think the idea is sound. If it's worth the effort is a separate question though. _______________________________________________ ffmpeg-cvslog mailing list ffmpeg-cvslog@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-cvslog
