On Thu, 18 Oct 2001, Len Makin wrote:
> I haven't come across any apps which break because DP and REAL are the
> same size (yet;-).
I think I have come across such code in the past. Programs written by
sensible and knowledgeable programmers don't rely on details of storage
alignment, but many Fortran programs are written by scientists who don't
realise the potential problems of doing such things.
> Do you count machines with Alpha CPUs as supercomputers?
Here we're been running the software for some years on Alpha/Tru64 Unix
boxes; my present aim is to eliminate the VAX-isms and make the code
portable enough to run on say 80x86/Linux (such as a Beowulf cluster) and
maybe SGI Origin/Irix machines. These seem to much more popular than
Crays around these parts; I wonder if cost has anything to do with it?
> Do/will these machines use default 64bit? If not,why not?
> Will your code ever run on one or more? Probably (never say never :-).
> Even the GNU people are grappling with 64bitfitness. Why not you?
Good questions: I'm very much aware of the 64-bit issue (for a different
purpose we already have data files getting close to the 2 GB limit; and
I'm old enough to remember the trauma of the 16==>32-bit transition). It
was because I didn't know the answers to these questions that I started
this thread.
As Richard Maine pointed out, for new f90 code, it's hardly any extra
trouble to use selected_real_kind and append "_r8" or whatever to each
floating constant. But for existing code it's quite a lot of work (it
makes lots of lines longer, so you need to rewrap the lines but that then
upsets lots of nice alignments that the programmers already put in their
code). And it permanently prevents anyone compiling the code using an F77
compiler, which itself reduces portability by a small but perhaps
significant extent. Not everyone has an F95 compiler installed,
unfortunately.
--
Clive Page, /"\ ASCII ribbon
Dept of Physics & Astronomy, \ / campaign against
University of Leicester. X HTML mail
U.K. / \ and postings
|