Construction of optimal space-time mappings for automatic parallelization of loop nests with static control flow

The article considers problems of affine scheduling and spatial distribution for programs composed of loop nests with static control flow as multi-objective optimization problems in context of data locality optimization. The presented methods, based on polyhedral model, construct space-time mappings to expose parallelism while minimizing data reuse latency in time domain and data reuse distance in space domain. Finding Pareto-optimal solution is reduced to solving of integer linear programming problem. The methods allow to specify subjective preferences more precisely in contrast to classic methods relying on lexicographic optimization. Application of the methods is illustrated with parallelization of LV-decomposition algorithm.