On Sun, Nov 15, 2015 at 10:05 AM, Ganesh Ajjanagadde <gajja...@mit.edu> wrote: > On Sat, Nov 14, 2015 at 9:19 PM, Michael Niedermayer > <mich...@niedermayer.cc> wrote: >> On Sat, Nov 14, 2015 at 12:11:20PM -0500, Ganesh Ajjanagadde wrote: >>> This improves the mathematical behavior of hypotenuse computation. >>> >>> Signed-off-by: Ganesh Ajjanagadde <gajjanaga...@gmail.com> >>> --- >>> libavutil/eval.c | 2 +- >>> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> if a hypot() function fallback is added this is ok >> otherwise the code as it is before this should be fine too in practice >> doubles have a quite large exponent range, half of that is still quite >> large > > It is not just an issue of overflow, but also accuracy more generally. > I gave overflow as the most easily understood issue. > Concrete example: > 1e-1, 1e-8 > actual: 0.100000000000000499999999... > hypot: 0.100000000000000505151476 > naive: 0.100000000000000519029264 > > I am pretty sure that the error can be far worse, e.g when bad underflow > occurs. > > There is a fallback, albeit the naive one which Paul does not like, > though I myself don't see why in light of your comment above. > > Rewrote and posted updated version that although not as "perfect" as > GNU libm's, is better than a naive sqrt(x*x + y*y). > > I admit it is a minor thing, on the lines of the log10 stuff, but I > posted it as it actually has a greater significance. > avutil/eval was used as an illustrative example.
pushed since hypot fallback has been pushed, thanks. > >> >> [...] >> -- >> Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB >> >> The real ebay dictionary, page 1 >> "Used only once" - "Some unspecified defect prevented a second use" >> "In good condition" - "Can be repaird by experienced expert" >> "As is" - "You wouldnt want it even if you were payed for it, if you knew >> ..." >> >> _______________________________________________ >> 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
