A technique for translating clausal specifications of numerical methods into efficient programs

AbstractWe describe a technique for translating numerical algorithms specified as clauses into dataflow graphs. The graphs have the property that common subexpressions are computed only once. The purpose of this technique is to convert high-complexity (exponential) solutions derived from elegant clausal specifications into very efficient computations having low (linear or log) complexity. The translation is not a program transformation, but a compilation of a term deduced from a goal clause. The effect of this translation is demonstrated for an assortment of numerical algorithms, including the fast Fourier transform, solution of matrix equations, and series approximation