Bill Long wrote:
>
> Having a background in numerical analysis, I fully agree about what is a
> reasonable expectation when implementing algorithms using floating point
> arithmetic. However, you must realize that this has absolutely nothing
> to do with the design or implementation of a computer language like
> Fortran. I could recount many stories about compiler upgrades resulted
> in slightly different answers for codes, usually due to improved
> optimization of addition sequences that resulted in different orders of
> operation. Some customers immediately file critical bug reports if only
> one bit in the answer has changed. The possibility that the new answer
> might be more accurate than the old one is completely irrelevant. The
> certification procedures for their codes are so onerous that making any
> change (even if it is an improvement) is essentially impossible. The
> test for a new compiler is that the binary answers are exactly the same
> as the certification run outputs. These are the customers who define the
> constraints for the compiler vendor. Your standards for numerical
> literacy are laudable. Unfortunately, they are not relevant in the real
> world for a compiler vendor.
This kind of reasoning goes nowhere. If the rational argument that
the precision is better, thus the (new) product is also better
doesn't work in this case, how can you ever add new optimizations
to the language or, e.g., better precision FPU's?
The problem of changed iteration counts appears when you either
change the software or the hardware. That means that you should
either never change either one, or hide software change with
a hardware change going on at the same time.
This may be a public relations problem, but not an issue when
designing a programming language. Look at the success of Bill Gates
who every year sells millions of copies of software that is more
or less incompatible with it's previous releases, has totally
new bugs etc. I understand the clients of CRAY are sligthly
different than the clients of M$, but not from a different planet.
Regards,
----------------------------------------------------------------------
Artur Swietanowski mailto:[log in to unmask]
Institut fuer Statistik, Operations Research und Computerverfahren,
Universitaet Wien, Universitaetsstr. 5, A-1010 Wien, Austria
tel. +43 (1) 407 63 55 - 120 fax +43 (1) 406 41 59
----------------------------------------------------------------------
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|