Unimodularity considered non-essential

Abstract

Loop transformations for parallelization are linear, invertible mappings from an iteration space to space-time; a further restriction on the mapping, unimodularity, simplifies code generation. In most previous work, non-unimodular transformations require a potentially large amount of overhead. We sketch a method for producing efficient code for non-unimodular loop transformations.