[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

IEEE754R



Ulrich Kulisch February 10, 2006

Dear colleagues

My thanks to everybody who sent me comments on my letter of January 28, 2006 to stds-754@xxxxxxxx and to Nelson Beebe. In the responses, the wordings accurate sum and dot product, exact sum and dot product, and faithfully rounded sum and dot product are used more or less synonymously. For me the meaning of the last differs from that of the others. An accurate or exact dot product means that the result is computed to the fullest possible accuracy. Not a single bit is lost.

Here is why I see the issue in this way:

A. The most natural way to accumulate numbers is fixed-point accumulation. It is simple, error free and fast. This is also true for the accumulation of floating-point numbers and of their products. If the result register is wide enough it can be done without exception. The result is exact. Not a single bit is lost. The arithmetic to achieve this is much the same as that of a conventional CPU. Fixed-point accumulation of floating-point sums and dot products can be realized in hardware at low cost. And it is very very fast. If supported by a vectorizing compiler it would boost both the speed of a computation and the accuracy of the result.

Fully accurate sums and dot products improve many applications. As a byproduct, multiple precision real and interval arithmetic can be done at very high speed. With operator overloading they are very easy to use. With a long precision interval arithmetic, for instance, highly accurate enclosures of orbits of dynamical systems have been obtained for considerable long durations. Iterated defect correction is another important class of applications. The method can be applied to compute enclosures of arithmetic expressions or of polynomials with very high accuracy. The result is an enclosure as a long precision interval. Verified solution of badly conditioned systems of linear equations by use of the Rump-operator is another large class of important applications. Finally, of course, a faithfully rounded result can be obtained. All these and other applications of an accurate dot product come with very high speed. They can be considered as top-down approaches of a fully accurate dot product.

B. In contrast to this the Rump-Ogita-Oishi method is a bottom-up approach. I mentioned in my mail of January 26, 2006 that I very much admire this methods. It achieves faithfully rounded sums and dot products just by using conventional floating-point arithmetic. The methods are fast in comparison with other software methods. This certainly is a great achievement of our field. Applications of these methods are inherently a subset of the applications of A.

I do not see any reason why we need the methods B. as justification for accurate sums and dot products in the next arithmetic standard. Fixed-point accumulation of sums and dot products is the additive equivalent of fast multiplication techniques, for instance by an adder tree. The advantages, the cost and gain in speed of both techniques are similar. I do not see any reason why mathematicians should hesitate to require this mode of operation from a new arithmetic standard. If we do not require it we will never get it. We would not have got floating-point operations with directed roundings if IEEE 754 hadnt required it.

With best regards
Ulrich Kulisch


--
*************************************************************************************
Ulrich Kulisch                                         Privat:
Institut für Angewandte Mathematik Im Eichbäumle 37 Universität Karlsruhe 76139 Karlsruhe 76128 Karlsruhe, Germany Te.: (+49) 721 686263 Tel.: (+49) 721 608 2680 oder (+49) 721 608 4202
*************************************************************************************

754 | revision | FAQ | references | list archive