(j3.2006) Integration of co-arrays with the intrinsic shift functions
Thu Jul 12 18:10:58 EDT 2007
On Jul 12, 2007, at 3:14 PM, Andy Vaught wrote:
> I'm not saying that your proposal is without merit or shouldn't be
> adopted, but you're moving into an area in which calculations will be
> irreproducible from run to run.
What language in the current (or previous) standards can you point to
that requires reproducible results?
This isn't a new problem, but it certainly does get more obvious as
parallelism and heterogeneous computing becomes more common. But even
a uniprocessor with a single floating point unit (e.g. the 8087) can
exhibit it (admittedly that's pathologically rare; but I've seen that
sort of thing most of my life ... attached array processors, parallel
processors, different steppings of multiprocessors, etc.).
As best I can tell, no Standard from 66 onward ever *required*
reproducible results. Clearly, on conventional uniprocessors failure
to do so is unexpected and is a quality of implementation issue. But
on more exotic machines ... it's never been carved in stone.
More information about the J3