[discuss] RFC: Support IEEE 754R draft in the x86-64 psABI
matz at suse.de
Tue May 16 17:08:32 CEST 2006
On Tue, 16 May 2006, Menezes, Evandro wrote:
> > > Doesn't that happen anyway even with FP data between systems of
> > > different endianness? Isn't file format application-dependent and
> > > customer-defined?
> > It's similar to endianness (and indeed that's a problem). But
> > endianness applies to all data: binary integers, binary
> > floating-point, and decimal-floating-point. This can be handled
> > independently of the data content.
> Yet, it is something that the user has to care about. I agree that
> adding yet another thing to care about is a burden, but IEEE754 created
> this situation by endorsing two DFP formats.
Not if nobody uses the other format. I'm actually leaning towards the DPD
format for compatibility reasons with other platforms and for performance
reasons once hardware solutions occur. The performance numbers that
Marius showed are impressive, but it was an Apple/Orange comparison.
Someone should sit down, write the software routines to do only DPD math
(i.e. not using decNumber), and _then_ compare with a realistic data set
(i.e. input data in decimal formats). That's where those types will be
used, not for general HPC math.
More information about the discuss