Floating-point multiply-and-accumulate unit with classes for alignment

Abstract

A floating point multiply-and-accumulate unit that performs an operation A*B.+-.C also determines an exponent difference (Ea+Eb)-Ec where Ea, Eb, and Ec are the exponents of values A, B, and C. The exponent difference indicates a class for the operation. In a first class, C is much greater than A*B and accumulation of a mantissa Mc of C with a mantissa Ma*Mb of A*B leaves mantissa Mc. In second and third classes, C and A*B are comparable and mantissas Mc and Ma*Mb overlap during accumulation. In a fourth class, A*B is much greater than C so that accumulation of mantissas Mc and Ma*Mb leaves Ma*Mb. The classes controls shift logic for alignment before accumulation or postnormalization after accumulation. For the first class, alignment or normalization are fixed according to Mc. For the second and third class, a fixed shift for alignment or normalization according to Ma*Mb is performed followed by a variable shift as indicated either by the exponent difference or cancellation detected during accumulation. For the fourth class, alignment or normalization are fixed according to Ma*Mb. Accordingly, alignment or normalization performs at most a fixed shift followed by are relatively small variable shift, and shift logic can be simpler than is required for arbitrary shifts across an entire shifter.