On 7/19/20, Paul B Mahol <one...@gmail.com> wrote: > On 7/19/20, Paul B Mahol <one...@gmail.com> wrote: >> On 7/19/20, Steinar H. Gunderson <steinar+ffm...@gunderson.no> wrote: >>> On Sun, Jul 19, 2020 at 09:02:30PM +0200, Paul B Mahol wrote: >>>>> Yes, this is the non-recursive version, which is O(n) in the number of >>>>> samples. This is why I recommended the recursive version, which is >>>>> O(log >>>>> n) >>>>> in the number of samples and thus avoids the problem with big lengths. >>>>> It is fairly straightforward to convert one to the other. >>>> Really? How so? >>> >>> As previously mentioned, you do two samples, then save to a temporary >>> buffer, >>> two samples from that buffer with half the distance and so on until your >>> samples are only one pixel apart. >>> >>> You can also try doing three or four samples at a time instead of two; >>> experiment a bit and see what gives the best performance. >> >> This is very slow. >> > > Also I fixed silly center artifacts in rblur implementation. > > Speed here on historic CPU for hd720 is 12-13 frames per second for 23 > fps video with slice threading enabled on 4 core CPU. > > This could be made even faster at expense of more memory usage. > I really doubt there is faster algorithm that needs less memory but > with no quality drop. >
Gonna apply this soon, as I'm happy with overall performance and output quality. _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org https://ffmpeg.org/mailman/listinfo/ffmpeg-devel To unsubscribe, visit link above, or email ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
