TECHNIQUE OF REDUCTION TRANSFORMATIONS FOR MINIMIZATION OF ANALYZED NUMBER OF PARALLEL PROGRAM VARIANTS

Abstract

Formulated several important conclusions of the basic theorems and the main principles of the reduction transformations’ technique. In comparison with automatic parallelizing, the proposed principles provide a considerable decrease in the number of steps required for the applications’ adaptation to architectures of reconfigurable computer systems. At the first stage, the reduction coefficient is factorized, and the rational values of different types of reduction coefficients (by the number of basic subgraphs, by the number of computational operations and by the digit capacity of data) are chosen from the multipliers. At the second stage, the reduction of basic subgraphs’ number is performed according to the coefficient chosen at the previous stage. At the third stage, the number of computational operations and the digit capacity of processed data are reduced to the limit values. The application of the reduction transformations’ technique allows for the integration of multiple variants of parallel program in one group and for the shortening of time needed for its adaptation to the architecture and configuration of a computer system. According to the estimations, the number of reduction transformation steps needed for the calculations’ scaling in reconfigurable computer systems is 26 steps and considerably less than the one for automatic parallelizing compilers.