On Wed, 2008-05-07 at 11:08 +0100, David Woodhouse wrote:
> On Tue, 2008-05-06 at 14:17 -0600, Orion Poplawski wrote:
> > The hdf5 1.8.0 does not build on ppc64 (or it builds, but it fails its
> > tests). I've put in an ExcludeArch: ppc64 for now and filed a bug:
>
> All the test failures are precision/rounding errors with 'long double'.
> My first guess would be that it's assuming that 'long double' is 64-bit.
> That's not a valid assumption, in the general case.
Hm, seems like it's _expecting_ rounding errors. But only handles them
by ignoring the last few bytes of the mantissa. But in these cases, the
mantissa and the exponent are both quite different, although the total
error is small:
Testing soft long -> long double conversions *FAILED*
elmt 107:
src = 00 3f ff ff ff ff ff ff 18014398509481983
dst = 43 4f ff ff ff ff ff ff 80 00 00 00 00 00 00 00 18014398509481982.000000
ans = 43 50 00 00 00 00 00 00 bf f0 00 00 00 00 00 00 18014398509481983.000000
Here, I think that 'dst' is the result of the routine under test, while
'ans' is the result of letting the hardware/compiler do the conversion
with a simple cast.
I'm not sure if this is a real error which it intended to catch, or
whether it's considered acceptable lack of precision.
As ever, I can give accounts on suitable machines to more knowledgeable
people who want to investigate further. Just let me have a SSH public
key.
--
dwmw2
--
fedora-devel-list mailing list
fedora-devel-list@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/fedora-devel-list
