On Tue, Dec 1, 2009 at 1:02 AM, keith bierman <keith.bierman at quantum.com> wrote: > > > I. Szczesniak: > > All of Sun's SPARCV9 chips do not implement 128bit floating point in > hardware and use a software emulation for SPARCV9 ABI conformance. > > The ABI certainly permits that. > > Unfortunately this makes software like perl, python etc slower than > software running on competing hardware platforms. > > Which platforms have native 128bit IEEE hw? IBM used a rather perverse > "double double" last time I'd looked, which was fast but like most non-IEEE > schemes had some odd behaviors in corner cases (even IEEE has odd, but odd > and non-Standard conforming is worse).
80bit IEEE would be sufficient. 64bit IEEE is not. 128bit IEEE would be a major advantage but Sun failed to deliver a useful implementation. > Sun's 'novel solution' was to reduce the data type for float in perl > to double, giving Solaris an edge over its competition. > Unfortunately this 'novel solution' renders the /usr/bin/perl useless > for the majority of scientific applications while Sun's support > division keeps telling us that this is the fault of the scientific > applications (which is a terrific excuse to tell this a Sun customer > who is developing and selling such software together with Sun's high > end machines). > > I was previously unaware of scientific applications coded in Perl which just > means I need to get out more. Look out for bioinformatics applications. Each bit of precision counts there. Irek
