(j3.2006) Liaison to IFIP WG 2.5
Tue Aug 21 14:44:15 EDT 2007
I think for once I am sort of agreeing with Bill, Mike and Alex, only sort of!
I do not see the need for intervals generally. There may be some critical steps in some very
specialist applications where it becomes essential that the true result is bracketed but most of the
time a well constructed FP calculation will do just fine.
Provided the system and the hardware supports the necessary directed roundings existing Fortran can,
via modules, support adequate interval arithmetic operations. Doing whole applications in intervals
probably indicates an inadequately analysed problem.
It is also the case that when intervals are crucial it is likely that fixed precision FP is not
adequate either. To this end such steps often need support for extended or even variable precision
arithmetic. I have not implemented it but a VPA module that also supported intervals would not be
that difficult. From experience, module implemented variable precision arithmetic is approx x100 on
hardware provided FP; useful when you really need it but not to be encouraged for lazy programmers
who use extra precision as a substitute for thinking about efficient solutions methods.
Intervals are interesting but without very careful analysis of the effects of error correlations
intervals almost always grow too wide to be useful. The hardware and the language need to support
the primitives to allow intervals to be implemented but the type interval as an intrinsic type in
the language is not desirable.
I would have to say that co-arrays are also not generally applicable. As an optional part/TR they
are obviously desirable but since they are clearly an architecturally specific efficiency aid, not
providing general problem solving expression support, I remain emphatically opposed to their
inclusion as a core part of Fortran in much the same way as I would not wish intervals to be
included as an intrinsic.
I remain convinced that F2010 with co-arrays as an integral part of the standard is unacceptable and
so would a core Fortran with intervals as an intrinsic type!
> -----Original Message-----
> From: j3-bounces at j3-fortran.org [mailto:j3-bounces at j3-fortran.org]On
> Behalf Of Aleksandar Donev
> Sent: 21 August 2007 19:13
> To: j3 at j3-fortran.org
> Subject: Re: (j3.2006) Liaison to IFIP WG 2.5
> On Tuesday 21 August 2007 08:43, Bill Long wrote:
> > 1) There would have to be significant customer demand (so far I've seen
> > exactly zero such demand), and
> Maybe, but the demand is not going to come unless there is something worthy to
> demand, i.e., there is potential for fast execution and there is a standard.
> As Mike said, it is not much different from co-arrays...it takes time...
> > 2) Hardware support needs to be widespread to make something beyond a
> > module worthwhile from a performance point of view. ?That means, at a
> > minimum, that IEEE 754r++ specify a standard.
> Yes, I see hardware support as a necessary precondition.
> Personally, I've wondered if it is possible to do interval FP in an additional
> off-processor unit (like GPUs or FPGAs or some other thing one can add to a
> processor)? Interval arithmetic is nice to have but IMO it is almost never
> necessary that ALL floating-point calculations be done with intervals.
> Certainly not at the cost of the usual FP calculations.
> J3 mailing list
> J3 at j3-fortran.org
More information about the J3