On Fri, Mar 25, 2016 at 9:32 AM, James Almer <jamr...@gmail.com> wrote: > On 3/24/2016 9:50 PM, Ganesh Ajjanagadde wrote: >> Useful for fast FFT computation: http://www.fftw.org/. >> >> Signed-off-by: Ganesh Ajjanagadde <gajja...@gmail.com> >> --- >> configure | 4 ++++ >> 1 file changed, 4 insertions(+) > > fftw brings nothing we don't already have. It's simply faster in some cases > and slower in others. > > Some projects use libavcodec for their FFT needs because of our > implementation.
Indeed, Firefox does in recent times as well: https://bugzilla.mozilla.org/show_bug.cgi?id=1157768. > In some cases they switched to another because, as pointed above, we're > lacking > assembly optimizations for some codepaths. See ticket #3921. > So instead of giving them more reason to not use our implementation, we should > address its few shortcomings. Depends if one views it as "few" or not, but whatever. At the very least, this showed something. For instance, it shows that for short lengths, avfft is reasonably ok, but performance degrades relative to alternatives for long lengths. It also showed the large amount of time spent permuting. And furthermore, regardless of whether we use FFTW or not, I view it as useful to keep in a working tree. That way, as avfft improves, we have a comparison point to test against. > > _______________________________________________ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
