[Scilab-Dev] Directed rounding, please
Eric WALTER
Eric.Walter at lss.supelec.fr
Wed Jun 26 18:52:45 CEST 2013
I would like Scilab to provide us with three command lines:
the first of them would order subsequent rounding to be carried out
towards +infinity,
the second one would order subsequent rounding to be carried out
towards -infinity
and the last one would restore standard rounding to the closest double.
This should not be too much work, as all IEEE-754 compliant processors
feature these types of rounding. It would allow Scilab users to get an
(admittedly rough) idea of the quality of what they are computing. If
the results with rounding towards +infinity are about the same as with
rounding towards -infinity, this of course does not prove that
everything is OK, but if the results differ widely, then for sure there
is some trouble brewing. Under some simplifying hypotheses, it is even
possible to infer an (again rough) estimate of the number of significant
digits in the results from the magnitude of the discrepancy.
By implementing this simple suggestion, Scilab would go a long way
towards helping its users assess the quality of their results, and
improve the way they evaluate critical results if need be. As Matlab
does not provide its users with such a facility, this would also give
Scilab another competitive advantage over its main competitor.
Regards,
Eric Walter
PS: Thanks for the nice ScilabTec 2013 Users Day
More information about the dev
mailing list