This is the mail archive of the
gsl-discuss@sources.redhat.com
mailing list for the GSL project.
Re: FAIL in make check /specfunc on FreeBSD4.3/i386
- To: Vladimir Savichev <vsavichev at mail dot wesleyan dot edu>
- Subject: Re: FAIL in make check /specfunc on FreeBSD4.3/i386
- From: Brian Gough <bjg at network-theory dot co dot uk>
- Date: Fri, 8 Jun 2001 12:21:38 +0100 (BST)
- Cc: gsl-discuss at sources dot redhat dot com
- References: <3B1F1BAF.4010201@wesleyan.edu>
- Reply-To: gsl-discuss at sources dot redhat dot com
Vladimir Savichev writes:
> after 891 we have
> sn = 0.96402778575700188 wanted 0.96402778575700186570 ~ 1.4 10^-17 <
> 2*GSL_DBL_EPSILON
> cn = 0.26580148285600663 0.26580148285600686381 ~ -2.3 10^-16 <
> 2*GSL_DBL_EPSILON
> dn = 0.26580323105264092 0.26580323105264131136 ~ -3.9 10^-16 <
> 2*GSL_DBL_EPSILON
>
> but in test_sf_frac_diff we return
> fabs((x1-x2)/(x1+x2)) so we get about 8 10^-16 at dn : HAS to FAIL !
> May be it is AMD issue or something particular about FreeBSD?
Thanks for reporting that... the tolerance in the test seems to be too
small to allow for the small variation between different
platforms/compilers/etc. We'll increase it before the next release.
> For example on entry in gsl_sf_elljac_e(u, m, &sn, &cn, &dn)
> I see like
> m=0.99999899999999997
> Shouldn't we see m=0.99999900000 for declared double m=0.999999 ?
It's ok, there is a conversion from decimal to binary so the result
does not always look exact in decimal.