Optimal multiprocessor implementations from a serial algorithm specification

Techniques are presented for the automatic generation of optimal synchronous multiprocessor implementations from a serial specification of an algorithm. Although a total system is described, two elements of the system are of particular interest. These are series of graph operations which transform the algorithm to be implemented from its initial serial form into an optimal graph form and the mapping techniques which transform the resulting graph into a fully defined synchronous multiprocessor schedule. The final multiprocessor implementations are often maximally fast, perfectly efficient, minimum-delay realizations which can operate using a limited class of communications networks. The algorithms addressed are all members of a class of iteratively applied numeric techniques which are specified using a constrained subset of FORTRAN