From Michel.Olagnon@ifremer.fr Wed May 7 11:23:32 1997 Received: from brest.ifremer.fr (brest.ifremer.fr [134.246.155.1]) by dkuug.dk (8.6.12/8.6.12) with ESMTP id LAA01978 for ; Wed, 7 May 1997 11:23:30 +0200 From: Michel.Olagnon@ifremer.fr Received: from marathon.ifremer.fr (marathon [134.246.149.16]) by brest.ifremer.fr (8.7.5/8.6.6) with SMTP id LAA14649; Wed, 7 May 1997 11:23:01 +0200 (MET DST) Received: by marathon.ifremer.fr (4.1) id AA03358; Wed, 7 May 97 11:23:00 +0200 Date: Wed, 7 May 97 11:23:00 +0200 Message-Id: <9705070923.AA03358@marathon.ifremer.fr> To: Michael.Metcalf@cern.ch, sc22wg5@dkuug.dk Subject: Floating-point arithmetic At 04:09 PM 5/6/97 +0200, Michael Metcalf wrote: > > There is a fascinating article, "On the need for predictable > floating-point arithmetic in the programmming languages Fortran 90 and > C/C++" in the March issue of SIGPLAN. It criticises both the language > standards and their (named) implementations (although C++ comes off better > than f90!). It applauds John Reid's efforts to improve the situation. > > Full details are at ftp://hwins.uia.ac.be/pub/CANT/SIGPLAN. > > My conclusion would be that, since this appears to be an area where C++ > beats f90 on its home ground, getting this right in F2000 deserves a > higher priority than burdening the language with interval arithmetic. > I strongly agree with Mike. I would appreciate if someone could point out an application, other than benchmarking purposes of vendors marketing teams, where standardization of interval arithmetic would be useful. On the opposite, predictability is essential in many actual applications. The authors of the above reference attribute a number of deficiencies in the F90 compilers that they tested to the lack of addressing of these points in the standard, and I support this idea. From my experience with stochastic arithmetic (which I believe is several orders of magnitude more useful in practice than pure interval arithmetic), I have heard only one reason put forward by potential users for not using it: It's no use, I know that my program gives good results, my application can't be of the numerically sensitive kind that you describe ! Standardization won't change that, nor make it easier for them to state that their published results of the last decade lay in the 0-5% confidence range, or, worse, that bridges may fail, planes crash, ships sink just because an algorithm had only _seemed_ to converge. I also fail to see how interval arithmetic could be efficiently standardized if the predictability deficiencies are not corrected beforehand. Michel Michel OLAGNON email: Michel.Olagnon@ifremer.fr