[Developers] [Patches] Make ADMConstraints GF public

David Rideout dprideout at gmail.com
Tue Sep 19 08:44:26 CDT 2006


On 9/19/06, Luca Baiotti <baiotti at aei.mpg.de> wrote:
> Hi,
>
> so, my character-comparison proposal is clearly unviable (...) and
> Jonathan's, for example, is far superior.
>
> However,
>
> Erik Schnetter wrote:
> > In that case, the numbers
> >
> >     0.000000000000123
> >     0.000000000000456
> >
> > would compare different. However, if these are constraint violations,
> > they should be treated as equal.
>
> I don't understand why the above numbers should be considered equal or
> their difference acceptable. Testsuites have, in my view, the aim of
> 1) checking whether a patch made to a code has accidentally changed
> something that the patch was not supposed to change (if the patch is
> supposed to change the testsuite output, then the testsuites should also
> be updated by the author of the patch);
> 2) checking whether running on different machines (or on different
> number of processors) produces different results, in which case the bug
> in the code or the problem with the machine should be addressed.
>
> Following this view, the numbers above should fail the testsuite.

Well no, I think they should pass.  (assuming that those values are of
order of the numerical error.)  This is the whole point.  Relative
error makes no sense at zero and absolute error makes no sense at
infinity.  So Erik's proposal of just demanding one of them pass is a
more permissive variant of my earlier proposal, which should work fine
I imagine.  Jonathan's proposal is essentially equivalent as well.

I'm surprised that things work as well as they do currently, when
requiring both absolute error and relative error to pass...

> > This idea is essentially a relative error, just in disguise.
>
> As said above, you are right.
>
> Is Jonathan's proposal being considered for implementation?
>
>
> Ciao,
>
> Luca.
>
> _______________________________________________
> Developers mailing list
> Developers at cactuscode.org
> http://www.cactuscode.org/mailman/listinfo/developers
>


More information about the Developers mailing list