Efficient multiprecision floating point multiplication with optimal directional rounding

Abstract

An algorithm is described for multiplying multiprecision floating-point numbers. The algorithm can produce either the smallest floating-point number greater than or equal to the true product, or the greatest floating-point number smaller than or equal to the true product. Software implementations of multiprecision floating-point multiplication can reduce the computation time by a factor of two if they do not compute the low-order digits of the product of the two mantissas. However, these algorithms do not necessarily provide optimally rounded results. The algorithms described here is guaranteed to produce optimally rounded results and typically obtains the same savings.<>